From 179e73e62fcb0a0ccece94fd18bd4a5ddcc23840 Mon Sep 17 00:00:00 2001 From: flybird11111 <1829166702@qq.com> Date: Sun, 7 Apr 2024 17:17:29 +0800 Subject: [PATCH 1/7] [branch rebase] rebase main to Feature/resize_embedding (#5554) * fix * [release] update version (#5411) * [hotfix] fix typo s/keywrods/keywords etc. (#5429) * [devops] fix compatibility (#5444) * [devops] fix compatibility * [hotfix] update compatibility test on pr * [devops] fix compatibility * [devops] record duration during comp test * [test] decrease test duration * fix falcon * [shardformer] fix gathering output when using tensor parallelism (#5431) * fix * padding vocab_size when using pipeline parallellism padding vocab_size when using pipeline parallellism fix fix * fix * fix fix fix * fix gather output * fix * fix * fix fix resize embedding fix resize embedding * fix resize embedding fix * revert * revert * revert * [doc] release Open-Sora 1.0 with model weights (#5468) * [doc] release Open-Sora 1.0 with model weights * [doc] release Open-Sora 1.0 with model weights * [doc] release Open-Sora 1.0 with model weights * [doc] update open-sora demo (#5479) * [doc] update open-sora demo * [doc] update open-sora demo * [doc] update open-sora demo * [example] add grok-1 inference (#5485) * [misc] add submodule * remove submodule * [example] support grok-1 tp inference * [example] add grok-1 inference script * [example] refactor code * [example] add grok-1 readme * [exmaple] add test ci * [exmaple] update readme --------- Co-authored-by: Hongxin Liu Co-authored-by: digger yu Co-authored-by: binmakeswell From e19c6ae173e616559ad5170fb30974b4ace29117 Mon Sep 17 00:00:00 2001 From: flybird11111 <1829166702@qq.com> Date: Tue, 9 Apr 2024 20:24:43 +0800 Subject: [PATCH 2/7] [CI] run pre-commit (#5577) * fix * [release] update version (#5411) * [hotfix] fix typo s/keywrods/keywords etc. (#5429) * [devops] fix compatibility (#5444) * [devops] fix compatibility * [hotfix] update compatibility test on pr * [devops] fix compatibility * [devops] record duration during comp test * [test] decrease test duration * fix falcon * [shardformer] fix gathering output when using tensor parallelism (#5431) * fix * padding vocab_size when using pipeline parallellism padding vocab_size when using pipeline parallellism fix fix * fix * fix fix fix * fix gather output * fix * fix * fix fix resize embedding fix resize embedding * fix resize embedding fix * revert * revert * revert * [doc] release Open-Sora 1.0 with model weights (#5468) * [doc] release Open-Sora 1.0 with model weights * [doc] release Open-Sora 1.0 with model weights * [doc] release Open-Sora 1.0 with model weights * [doc] update open-sora demo (#5479) * [doc] update open-sora demo * [doc] update open-sora demo * [doc] update open-sora demo * [example] add grok-1 inference (#5485) * [misc] add submodule * remove submodule * [example] support grok-1 tp inference * [example] add grok-1 inference script * [example] refactor code * [example] add grok-1 readme * [exmaple] add test ci * [exmaple] update readme * run pre-commit --------- Co-authored-by: Hongxin Liu Co-authored-by: digger yu Co-authored-by: binmakeswell --- .github/workflows/build_on_pr.yml | 2 +- LICENSE | 2 +- .../benchmarks/benchmark_opt_lora_dummy.py | 6 +- .../Chat/coati/dataset/sft_dataset.py | 16 +- applications/Chat/coati/models/base/actor.py | 1 - applications/Chat/coati/ray/utils.py | 4 +- .../Chat/coati/trainer/strategies/ddp.py | 3 +- .../community/peft/train_peft_prompts.py | 4 +- applications/Chat/examples/train_sft.sh | 2 +- .../colossal_llama2/model/init_model.py | 3 +- .../colossal_llama2/utils/ckpt_io.py | 2 +- .../utils/stream_chat_patch.py | 119 ++++++----- .../Colossal-LLaMA-2/docs/example_13b.md | 6 +- .../Colossal-LLaMA-2/docs/example_7b.md | 2 +- .../Colossal-LLaMA-2/hostfile.example | 2 +- .../Colossal-LLaMA-2/inference_example.py | 2 +- .../Colossal-LLaMA-2/requirements.txt | 1 - .../Colossal-LLaMA-2/stream_chat_example.py | 45 ++-- applications/Colossal-LLaMA-2/version.txt | 2 +- .../examples/dataset_evaluation/inference.py | 1 - applications/ColossalMoE/README.md | Bin 6474 -> 6475 bytes applications/ColossalMoE/infer.py | 1 - .../colossalqa/chain/retrieval_qa/base.py | 18 +- .../colossalqa/data_loader/document_loader.py | 2 +- .../data_loader/table_dataloader.py | 75 ++++--- .../colossalqa/local/colossalcloud_llm.py | 66 +++--- .../ColossalQA/colossalqa/local/llm.py | 1 - .../ColossalQA/colossalqa/prompt/prompt.py | 1 - .../ColossalQA/colossalqa/retriever.py | 8 +- .../ColossalQA/data/data_sample/companies.txt | 6 +- .../data/data_sample/companies_zh.txt | 2 +- .../data/data_sample/csv_organization_100.csv | 2 +- applications/ColossalQA/data/tests/64KB.json | 2 +- .../ColossalQA/data/tests/companies.csv | 202 +++++++++--------- applications/ColossalQA/data/tests/test.html | 48 ++--- applications/ColossalQA/data/tests/test.md | 4 +- applications/ColossalQA/data/tests/test.txt | 76 +++---- .../retrieval_conversation_universal.py | 33 +-- .../examples/webui_demo/RAG_ChatBot.py | 12 +- .../ColossalQA/examples/webui_demo/README.md | 2 +- .../ColossalQA/examples/webui_demo/config.py | 52 ++--- .../ColossalQA/examples/webui_demo/server.py | 25 +-- .../ColossalQA/examples/webui_demo/webui.py | 31 ++- applications/ColossalQA/pytest.ini | 2 +- .../ColossalQA/tests/test_document_loader.py | 14 +- .../ColossalQA/tests/test_retrieval_qa.py | 80 +++---- applications/ColossalQA/version.txt | 2 +- colossalai/__init__.py | 2 +- .../booster/plugin/torch_fsdp_plugin.py | 9 +- colossalai/checkpoint_io/utils.py | 3 +- colossalai/cluster/process_group_mesh.py | 1 - colossalai/inference/engine/modeling/llama.py | 49 +++-- .../quant/gptq/cai_gptq/cai_quant_linear.py | 202 +++++++++--------- colossalai/kernel/triton/context_attention.py | 133 ++++++++---- colossalai/kernel/triton/flash_decoding.py | 51 +++-- .../kernel/triton/llama_act_combine_kernel.py | 65 +++--- .../kernel/triton/token_attention_kernel.py | 20 +- .../tensor_parallel/modeling/llama.py | 4 +- .../shardformer/policies/base_policy.py | 2 +- colossalai/tensor/d_tensor/__init__.py | 4 +- colossalai/tensor/d_tensor/api.py | 7 +- colossalai/zero/low_level/_utils.py | 1 + .../train_vit_with_hybrid_parallelism.md | 2 +- examples/language/data_utils.py | 2 +- examples/language/openmoe/benchmark/utils.py | 6 +- examples/language/openmoe/infer.py | 14 +- .../openmoe/model/convert_openmoe_ckpt.py | 22 +- .../language/openmoe/model/openmoe_policy.py | 89 ++++---- extensions/cpu_adam/__init__.py | 3 +- extensions/layernorm/__init__.py | 2 +- extensions/moe/__init__.py | 2 +- extensions/optimizer/__init__.py | 2 +- extensions/softmax/__init__.py | 2 +- tests/kit/model_zoo/__init__.py | 44 ++-- tests/kit/model_zoo/registry.py | 2 +- tests/kit/model_zoo/transformers/chatglm2.py | 1 + .../test_plugin/test_3d_plugin.py | 4 +- .../test_torch_fsdp_checkpoint_io.py | 1 + tests/test_gptq/test_gptq_linear.py | 30 ++- tests/test_lazy/test_models.py | 2 +- .../test_layer/test_dist_crossentropy.py | 5 +- .../test_model/test_shard_gptj.py | 1 + 82 files changed, 894 insertions(+), 887 deletions(-) diff --git a/.github/workflows/build_on_pr.yml b/.github/workflows/build_on_pr.yml index b01d15490e0f..2cad504f3391 100644 --- a/.github/workflows/build_on_pr.yml +++ b/.github/workflows/build_on_pr.yml @@ -201,4 +201,4 @@ jobs: uses: actions/upload-artifact@v3 with: name: report - path: report/ \ No newline at end of file + path: report/ diff --git a/LICENSE b/LICENSE index bacb03e72246..47197afe6644 100644 --- a/LICENSE +++ b/LICENSE @@ -551,4 +551,4 @@ Copyright 2021- HPC-AI Technology Inc. All rights reserved. AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN - THE SOFTWARE. + THE SOFTWARE. diff --git a/applications/Chat/benchmarks/benchmark_opt_lora_dummy.py b/applications/Chat/benchmarks/benchmark_opt_lora_dummy.py index 0d0e2a7d34f5..47ad29f46c64 100644 --- a/applications/Chat/benchmarks/benchmark_opt_lora_dummy.py +++ b/applications/Chat/benchmarks/benchmark_opt_lora_dummy.py @@ -76,9 +76,11 @@ def main(args): if args.strategy == "ddp": strategy = DDPStrategy() elif args.strategy == "colossalai_gemini": - strategy = GeminiStrategy(placement_policy="static",initial_scale=2**5) + strategy = GeminiStrategy(placement_policy="static", initial_scale=2**5) elif args.strategy == "colossalai_gemini_cpu": - strategy = GeminiStrategy(placement_policy="static", offload_optim_frac=1.0, offload_param_frac=1.0, initial_scale=2**5) + strategy = GeminiStrategy( + placement_policy="static", offload_optim_frac=1.0, offload_param_frac=1.0, initial_scale=2**5 + ) elif args.strategy == "colossalai_zero2": strategy = LowLevelZeroStrategy(stage=2, placement_policy="cuda") elif args.strategy == "colossalai_zero2_cpu": diff --git a/applications/Chat/coati/dataset/sft_dataset.py b/applications/Chat/coati/dataset/sft_dataset.py index e67e16231cc2..0167f118dc8f 100644 --- a/applications/Chat/coati/dataset/sft_dataset.py +++ b/applications/Chat/coati/dataset/sft_dataset.py @@ -51,11 +51,21 @@ def _preprocess( """Preprocess the data by tokenizing.""" sequences = [s + t + tokenizer.eos_token for s, t in zip(sources, targets)] sequences_token = tokenizer( - sequences, max_length=max_length, padding="max_length", truncation=True, return_tensors="pt", add_special_tokens=False + sequences, + max_length=max_length, + padding="max_length", + truncation=True, + return_tensors="pt", + add_special_tokens=False, ) sources_token = tokenizer( - sources, max_length=max_length, padding="max_length", truncation=True, return_tensors="pt", add_special_tokens=False + sources, + max_length=max_length, + padding="max_length", + truncation=True, + return_tensors="pt", + add_special_tokens=False, ) assert sequences_token["attention_mask"].dim() == 2, "seq2seq model should be preprocessed differently" @@ -66,7 +76,7 @@ def _preprocess( if tokenizer.padding_side == "right": # |prompt|completion|eos|pad| labels[i][:source_len] = IGNORE_INDEX - if pad_len>0: + if pad_len > 0: labels[i][-pad_len:] = IGNORE_INDEX elif tokenizer.padding_side == "left": # |pad|prompt|completion|eos| diff --git a/applications/Chat/coati/models/base/actor.py b/applications/Chat/coati/models/base/actor.py index 8b2b81ed071c..0634631df7a3 100644 --- a/applications/Chat/coati/models/base/actor.py +++ b/applications/Chat/coati/models/base/actor.py @@ -30,4 +30,3 @@ def forward( """Returns model output.""" output = self.model(input_ids, attention_mask=attention_mask, **model_kwargs) return output - diff --git a/applications/Chat/coati/ray/utils.py b/applications/Chat/coati/ray/utils.py index b88140c0e036..4882f00b7eca 100644 --- a/applications/Chat/coati/ray/utils.py +++ b/applications/Chat/coati/ray/utils.py @@ -75,7 +75,9 @@ def get_strategy_from_args(strategy: str): elif strategy == "colossalai_zero2": strategy_ = LowLevelZeroStrategy(stage=2, placement_policy="cuda") elif strategy == "colossalai_gemini_cpu": - strategy_ = GeminiStrategy(placement_policy="static", offload_optim_frac=1.0, offload_param_frac=1.0, initial_scale=2**5) + strategy_ = GeminiStrategy( + placement_policy="static", offload_optim_frac=1.0, offload_param_frac=1.0, initial_scale=2**5 + ) elif strategy == "colossalai_zero2_cpu": strategy_ = LowLevelZeroStrategy(stage=2, placement_policy="cpu") else: diff --git a/applications/Chat/coati/trainer/strategies/ddp.py b/applications/Chat/coati/trainer/strategies/ddp.py index f2a44aeb0961..3d3d44ca51ae 100644 --- a/applications/Chat/coati/trainer/strategies/ddp.py +++ b/applications/Chat/coati/trainer/strategies/ddp.py @@ -101,16 +101,17 @@ def save_pretrained( model_path = os.path.join(path, "pytorch_model.bin") self.save_model(model, model_path, shard=shard) + def _replace_keys(model_path: str, replace_fn: Callable): state_dict = torch.load(model_path, map_location="cpu") state_dict = {replace_fn(k): v for k, v in state_dict.items()} torch.save(state_dict, model_path) + # FIXME: save_model would add "model." prefix to keys of pytorch_model.bin # HACK: rename keys of pytorch_model.bin if dist.get_rank() == 0: _replace_keys(model_path, lambda k: k.replace("model.", "", 1)) - def get_model_state_dict_shard(self, model: nn.Module, **config): # TODO: implement sharding on naive strategy model = self.unwrap_model(model) diff --git a/applications/Chat/examples/community/peft/train_peft_prompts.py b/applications/Chat/examples/community/peft/train_peft_prompts.py index 1dd9ffcdf1cd..4625f2bda4de 100644 --- a/applications/Chat/examples/community/peft/train_peft_prompts.py +++ b/applications/Chat/examples/community/peft/train_peft_prompts.py @@ -24,7 +24,9 @@ def main(args): if args.strategy == "ddp": strategy = DDPStrategy() elif args.strategy == "colossalai_gemini": - strategy = GeminiStrategy(placement_policy="static", offload_optim_frac=1.0, offload_param_frac=1.0, initial_scale=2**5) + strategy = GeminiStrategy( + placement_policy="static", offload_optim_frac=1.0, offload_param_frac=1.0, initial_scale=2**5 + ) elif args.strategy == "colossalai_zero2": strategy = LowLevelZeroStrategy(stage=2, placement_policy="cpu") else: diff --git a/applications/Chat/examples/train_sft.sh b/applications/Chat/examples/train_sft.sh index b7d176847d9c..0fb4da3d3ce8 100755 --- a/applications/Chat/examples/train_sft.sh +++ b/applications/Chat/examples/train_sft.sh @@ -25,4 +25,4 @@ torchrun --standalone --nproc_per_node=4 train_sft.py \ --accumulation_steps 8 \ --lr 2e-5 \ --max_datasets_size 512 \ - --max_epochs 1 \ No newline at end of file + --max_epochs 1 diff --git a/applications/Colossal-LLaMA-2/colossal_llama2/model/init_model.py b/applications/Colossal-LLaMA-2/colossal_llama2/model/init_model.py index 67e487f43b08..f61291f35d04 100644 --- a/applications/Colossal-LLaMA-2/colossal_llama2/model/init_model.py +++ b/applications/Colossal-LLaMA-2/colossal_llama2/model/init_model.py @@ -8,11 +8,10 @@ import numpy as np import torch -from transformers import LlamaTokenizer, LlamaForCausalLM +from transformers import LlamaForCausalLM, LlamaTokenizer from colossalai.logging import get_dist_logger - logger = get_dist_logger() diff --git a/applications/Colossal-LLaMA-2/colossal_llama2/utils/ckpt_io.py b/applications/Colossal-LLaMA-2/colossal_llama2/utils/ckpt_io.py index 85decf37dd0b..05342ce41a60 100644 --- a/applications/Colossal-LLaMA-2/colossal_llama2/utils/ckpt_io.py +++ b/applications/Colossal-LLaMA-2/colossal_llama2/utils/ckpt_io.py @@ -10,8 +10,8 @@ from typing import Any, Dict, Tuple, Union import torch -from torch.optim.optimizer import Optimizer from torch.optim.lr_scheduler import _LRScheduler +from torch.optim.optimizer import Optimizer from colossalai.booster import Booster from colossalai.cluster import DistCoordinator diff --git a/applications/Colossal-LLaMA-2/colossal_llama2/utils/stream_chat_patch.py b/applications/Colossal-LLaMA-2/colossal_llama2/utils/stream_chat_patch.py index 8f8eecb18eb0..44fa3678d621 100644 --- a/applications/Colossal-LLaMA-2/colossal_llama2/utils/stream_chat_patch.py +++ b/applications/Colossal-LLaMA-2/colossal_llama2/utils/stream_chat_patch.py @@ -1,20 +1,19 @@ from copy import deepcopy -from typing import Optional, List, Dict, Tuple, Callable, Any +from typing import Any, Callable, Dict, List, Optional, Tuple import torch from torch import nn - from transformers import PreTrainedTokenizer -from transformers.utils import logging from transformers.generation.utils import GenerationConfig, LogitsProcessorList, StoppingCriteriaList - +from transformers.utils import logging + logger = logging.get_logger(__name__) def get_prompt_template( - input_query:str, - history:List[Dict]= None, - roles:list = ["", "Human", "Assistant"], + input_query: str, + history: List[Dict] = None, + roles: list = ["", "Human", "Assistant"], ) -> str: """ Generates a prompt template for chat models based on input and history. @@ -32,7 +31,7 @@ def get_prompt_template( new_history = [] else: new_history = deepcopy(history) - + new_history.append({"role": roles[1], "message": input_query.strip()}) new_history.append({"role": roles[2], "message": None}) @@ -48,22 +47,23 @@ def get_prompt_template( prompt += f"{role}: " return prompt + @torch.inference_mode() def streaming_chat( - model: Any, + model: Any, tokenizer: PreTrainedTokenizer, - input_query: str, - history: List[Dict] = None, - roles: list = ["", "Human", "Assistant"], - past_key_values: Tuple[Tuple[torch.FloatTensor, Any], Any] = None, - temperature: float = 0.8, - top_p: float = 0.95, - top_k: int = 50, - do_sample: bool = True, + input_query: str, + history: List[Dict] = None, + roles: list = ["", "Human", "Assistant"], + past_key_values: Tuple[Tuple[torch.FloatTensor, Any], Any] = None, + temperature: float = 0.8, + top_p: float = 0.95, + top_k: int = 50, + do_sample: bool = True, length_penalty: float = 1.2, - max_new_tokens: int = 512, - logits_processor: LogitsProcessorList = None, - return_past_key_values: bool = False, + max_new_tokens: int = 512, + logits_processor: LogitsProcessorList = None, + return_past_key_values: bool = False, **kwargs, ): """ @@ -87,7 +87,7 @@ def streaming_chat( **kwargs: Additional keyword arguments for generation. Yields: - Tuple[str, List[Dict], Optional[Tuple[Tuple[torch.FloatTensor, Any], Any]]]: A tuple containing the generated response, updated history, and + Tuple[str, List[Dict], Optional[Tuple[Tuple[torch.FloatTensor, Any], Any]]]: A tuple containing the generated response, updated history, and optionally the updated past key values if `return_past_key_values` is True. Ensures padding is on the left side for the tokenizer. @@ -97,32 +97,37 @@ def streaming_chat( history = [] if logits_processor is None: logits_processor = LogitsProcessorList() - + generation_kwargs = { - 'temperature': temperature, - 'top_p': top_p, - 'top_k': top_k, - 'do_sample': do_sample, - 'max_new_tokens': max_new_tokens, - 'length_penalty': length_penalty, - 'use_cache': True, - **kwargs + "temperature": temperature, + "top_p": top_p, + "top_k": top_k, + "do_sample": do_sample, + "max_new_tokens": max_new_tokens, + "length_penalty": length_penalty, + "use_cache": True, + **kwargs, } prompt_str = get_prompt_template(input_query, history=history, roles=roles) - + eos_token_id = [tokenizer.eos_token_id] inputs = tokenizer(prompt_str, return_tensors="pt").to(model.device) history.append({"role": roles[1], "message": input_query.strip()}) history.append({"role": roles[2], "message": None}) - for outputs in stream_generate(model, **inputs, past_key_values=past_key_values, - eos_token_id=eos_token_id, return_past_key_values=return_past_key_values, - **generation_kwargs): + for outputs in stream_generate( + model, + **inputs, + past_key_values=past_key_values, + eos_token_id=eos_token_id, + return_past_key_values=return_past_key_values, + **generation_kwargs, + ): if return_past_key_values: outputs, past_key_values = outputs - outputs = outputs.tolist()[0][len(inputs["input_ids"][0]):-1] + outputs = outputs.tolist()[0][len(inputs["input_ids"][0]) : -1] response = tokenizer.decode(outputs) history[-1]["message"] = response.strip() @@ -130,30 +135,30 @@ def streaming_chat( yield response, history, past_key_values else: yield response, history - + @torch.inference_mode() def stream_generate( - model: Any, - input_ids: torch.Tensor, + model: Any, + input_ids: torch.Tensor, generation_config: Optional[GenerationConfig] = None, logits_processor: Optional[LogitsProcessorList] = None, stopping_criteria: Optional[StoppingCriteriaList] = None, prefix_allowed_tokens_fn: Optional[Callable[[int, torch.Tensor], List[int]]] = None, - return_past_key_values: bool = False, + return_past_key_values: bool = False, **kwargs, ): """ Generates sequences of token ids using the specified model and generation parameters. Adapted from https://huggingface.co/THUDM/chatglm3-6b/blob/main/modeling_chatglm.py - + Args: model (Any): The model used for generating sequences of token ids. - input_ids (torch.Tensor): The sequence used as a prompt for the generation or as model inputs to the encoder. + input_ids (torch.Tensor): The sequence used as a prompt for the generation or as model inputs to the encoder. generation_config (Optional[GenerationConfig]): The generation configuration to be used as base parametrization for the generation call. logits_processor (Optional[LogitsProcessorList]): Custom logits processors that complement the default logits processors built from arguments and generation config. - stopping_criteria (Optional[StoppingCriteriaList]): Custom stopping criteria that complement the default stopping criteria built from arguments + stopping_criteria (Optional[StoppingCriteriaList]): Custom stopping criteria that complement the default stopping criteria built from arguments and a generation config. prefix_allowed_tokens_fn (Optional[Callable[[int, torch.Tensor], List[int]]]): Function to constrain token generation. return_past_key_values (bool): Whether to return past key values for further incremental decoding, defaults to False. @@ -169,7 +174,7 @@ def stream_generate( generation_config = model.generation_config generation_config = deepcopy(generation_config) model_kwargs = generation_config.update(**kwargs) - + eos_token_id = generation_config.eos_token_id if isinstance(eos_token_id, int): eos_token_id = [eos_token_id] @@ -177,25 +182,25 @@ def stream_generate( if generation_config.max_new_tokens is not None: generation_config.max_length = generation_config.max_new_tokens + input_ids_len - + if input_ids_len >= generation_config.max_length: input_ids_string = "decoder_input_ids" if model.config.is_encoder_decoder else "input_ids" logger.warning( - f"Input length of {input_ids_string} is {input_ids_len}, but `max_length` is set to" - f" {generation_config.max_length}. This can lead to unexpected behavior. You should consider" - " increasing `max_new_tokens`." - ) + f"Input length of {input_ids_string} is {input_ids_len}, but `max_length` is set to" + f" {generation_config.max_length}. This can lead to unexpected behavior. You should consider" + " increasing `max_new_tokens`." + ) logits_processor = logits_processor if logits_processor is not None else LogitsProcessorList() stopping_criteria = stopping_criteria if stopping_criteria is not None else StoppingCriteriaList() - + # prepare distribution pre_processing samplers logits_processor = model._get_logits_processor( - generation_config=generation_config, - input_ids_seq_length=input_ids_len, - encoder_input_ids=input_ids, - prefix_allowed_tokens_fn=prefix_allowed_tokens_fn, - logits_processor=logits_processor, - ) + generation_config=generation_config, + input_ids_seq_length=input_ids_len, + encoder_input_ids=input_ids, + prefix_allowed_tokens_fn=prefix_allowed_tokens_fn, + logits_processor=logits_processor, + ) # prepare stopping criteria stopping_criteria = model._get_stopping_criteria( @@ -205,7 +210,7 @@ def stream_generate( logits_warper = model._get_logits_warper(generation_config) unfinished_sequences = input_ids.new(input_ids.shape[0]).fill_(1) scores = None - + while True: model_inputs = model.prepare_inputs_for_generation(input_ids, **model_kwargs) # forward pass to get next token @@ -244,4 +249,4 @@ def stream_generate( yield input_ids # stop when each sentence is finished, or if exceed the maximum length if unfinished_sequences.max() == 0 or stopping_criteria(input_ids, scores): - break \ No newline at end of file + break diff --git a/applications/Colossal-LLaMA-2/docs/example_13b.md b/applications/Colossal-LLaMA-2/docs/example_13b.md index 3854aaa6a933..5305a999ba71 100644 --- a/applications/Colossal-LLaMA-2/docs/example_13b.md +++ b/applications/Colossal-LLaMA-2/docs/example_13b.md @@ -43,7 +43,7 @@ if __name__ == '__main__': model.to(device) tokenizer = AutoTokenizer.from_pretrained("hpcai-tech/Colossal-LLaMA-2-13b-base") - + question = "xxx" # Your question. prompt = (f"请对下面的问题，给出准确、可靠、安全的回答\n" f"{question}" @@ -111,7 +111,7 @@ The Colossal-AI team has conducted comprehensive evaluation for safety and value | 整车热管理主要包括什么内容 | 整车热管理主要包括：1、发动机冷却系统；2、散热系统；3、空调冷却系统；4、采暖系统；5、动力转向器冷却系统；6、润滑油冷却系统；7、变速箱冷却系统；8、燃油冷却系统；9、冷却系统排放液系统。 | | 简述下汽车空调-热泵循环系统的原理 | 汽车空调系统主要有三种形式，其中热泵循环系统是其中的一种。其原理是利用发动机排放的热量，通过压缩机将冷媒压缩成高温高压的状态，然后通过冷却器释放到空气中，吸收周围的热量，再通过膨胀阀释放到蒸发器中，蒸发器在压缩机的作用下将冷媒压缩成高温高压的状态，再次释放到空气中，从而实现制冷的效果。 | -#### Finance +#### Finance | Question |
Colossal-LLaMA-2-13b-base
| | :------: | :----------------------- | | 如何看待当前提前还贷热 | 提前还贷是一种常见的还贷方式，在市场中存在着许多的因素和影响，需要具体分析当前的经济形势和人们心理状况的影响。提前还贷可以帮助借款人早日摆脱债务负担，但也存在一些风险和不确定性。从经济形势来看，当前利率下探、还款压力减轻等因素可能会导致一部分借款人提前还贷。此外，一些购房者可能会受到市场不景气的影响，担心未来房产价值会下降，因此提前还贷以避免资产贬值。从借款人心理状况来看，提前还贷也可能会由于借款人对未来经济预期的不确定性而产生压力和不满。因此，提前还贷并不能一概而论，需要根据具体情况进行分析。如果提前还贷对借款人的经济情况和未来规划造成不利影响，建议谨慎考虑是否要提前还贷。| @@ -150,4 +150,4 @@ The Colossal-AI team has conducted comprehensive evaluation for safety and value ## Conclusion -The Colossal-AI team's advanced 13B model, compared to the 7B version, features a refined data structure categorizing information into informative, functional, and memory replay data. Informative data is intricately subdivided into major categories, each further segmented for precise control. Concurrently, data scale across domains is expanded. Tailored enhancements meet community demands for large model capabilities in natural language processing tasks, ensuring proficiency during pre-training and cost-effective fine-tuning. Addressing security and values concerns, multidimensional controls are implemented, securing the baseline model and aligning it with correct values. \ No newline at end of file +The Colossal-AI team's advanced 13B model, compared to the 7B version, features a refined data structure categorizing information into informative, functional, and memory replay data. Informative data is intricately subdivided into major categories, each further segmented for precise control. Concurrently, data scale across domains is expanded. Tailored enhancements meet community demands for large model capabilities in natural language processing tasks, ensuring proficiency during pre-training and cost-effective fine-tuning. Addressing security and values concerns, multidimensional controls are implemented, securing the baseline model and aligning it with correct values. diff --git a/applications/Colossal-LLaMA-2/docs/example_7b.md b/applications/Colossal-LLaMA-2/docs/example_7b.md index d889ab4165d0..d833d28218b5 100644 --- a/applications/Colossal-LLaMA-2/docs/example_7b.md +++ b/applications/Colossal-LLaMA-2/docs/example_7b.md @@ -242,4 +242,4 @@ To comprehensively assess the performance of the Colossal-LLaMA-2-7B-base model, ## Conclusion In general, the Colossal-LLaMA-2-7B-base model not only enhances its understanding of English but also exhibits significant improvements in its comprehension of Chinese. It boasts a broad spectrum of general knowledge, encompassing various fields such as food, sports, technology, literature, games, and more. Regarding text generation tasks, the Colossal-LLaMA-2-7B-base model excels in writing performance; however, its ability to generate specific formats like code, emails, tables, etc., needs enhancement due to the scarcity of relevant training data during our training phase. When compared to the Qwen-7b-base model, the Colossal-LLaMA-2-7B-base model outperforms it in answering most English questions and some Chinese questions, as demonstrated in the examples above. -Presently, the Colossal-LLaMA-2-7B-base model already exhibits some capabilities in sentiment analysis, logical reasoning, information extraction, role-play, classification, and rewriting. These capabilities are poised for further improvement in the future as part of our ongoing enhancements. \ No newline at end of file +Presently, the Colossal-LLaMA-2-7B-base model already exhibits some capabilities in sentiment analysis, logical reasoning, information extraction, role-play, classification, and rewriting. These capabilities are poised for further improvement in the future as part of our ongoing enhancements. diff --git a/applications/Colossal-LLaMA-2/hostfile.example b/applications/Colossal-LLaMA-2/hostfile.example index 82948648cbc9..cfaaa0ef559f 100644 --- a/applications/Colossal-LLaMA-2/hostfile.example +++ b/applications/Colossal-LLaMA-2/hostfile.example @@ -1,2 +1,2 @@ hostname1 -hostname2 \ No newline at end of file +hostname2 diff --git a/applications/Colossal-LLaMA-2/inference_example.py b/applications/Colossal-LLaMA-2/inference_example.py index 63ce91e50432..8d301616d678 100644 --- a/applications/Colossal-LLaMA-2/inference_example.py +++ b/applications/Colossal-LLaMA-2/inference_example.py @@ -15,7 +15,7 @@ def load_model(model_path, device="cuda", **kwargs): model.to(device) try: - tokenizer = AutoTokenizer.from_pretrained(model_path, padding_side='left') + tokenizer = AutoTokenizer.from_pretrained(model_path, padding_side="left") except OSError: raise ImportError("Tokenizer not found. Please check if the tokenizer exists or the model path is correct.") diff --git a/applications/Colossal-LLaMA-2/requirements.txt b/applications/Colossal-LLaMA-2/requirements.txt index 34afaf7e5cfd..5cdb8e7f3348 100644 --- a/applications/Colossal-LLaMA-2/requirements.txt +++ b/applications/Colossal-LLaMA-2/requirements.txt @@ -12,4 +12,3 @@ flash-attn>=2.0.0,<=2.0.5 tqdm sentencepiece==0.1.99 protobuf<=3.20.0 - diff --git a/applications/Colossal-LLaMA-2/stream_chat_example.py b/applications/Colossal-LLaMA-2/stream_chat_example.py index 3e45c690f878..4c0d1fe2a35f 100644 --- a/applications/Colossal-LLaMA-2/stream_chat_example.py +++ b/applications/Colossal-LLaMA-2/stream_chat_example.py @@ -1,11 +1,11 @@ -import os import argparse -from transformers import AutoTokenizer, AutoModelForCausalLM from colossal_llama2.utils.stream_chat_patch import streaming_chat +from transformers import AutoModelForCausalLM, AutoTokenizer SYSTEM = "A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions." + def main(args): model = AutoModelForCausalLM.from_pretrained(args.model_path).cuda().eval() tokenizer = AutoTokenizer.from_pretrained(args.tokenizer_path) @@ -27,29 +27,34 @@ def main(args): print(f"\n{roles[2]}: ", end="") gen_len = 0 for response, history, past_key_values in streaming_chat( - model, tokenizer, input_query, history=history, roles=roles, - temperature = args.temperature, - top_p = args.top_p, - top_k = args.top_k, - do_sample = args.do_sample, - length_penalty = args.length_penalty, - max_new_tokens = args.max_new_tokens, + model, + tokenizer, + input_query, + history=history, + roles=roles, + temperature=args.temperature, + top_p=args.top_p, + top_k=args.top_k, + do_sample=args.do_sample, + length_penalty=args.length_penalty, + max_new_tokens=args.max_new_tokens, past_key_values=past_key_values, - return_past_key_values=True): - + return_past_key_values=True, + ): output = response[gen_len:] print(output, end="", flush=True) gen_len = len(response) + if __name__ == "__main__": parser = argparse.ArgumentParser() - parser.add_argument('--model_path', type=str, default=None, help="path to chat version model") - parser.add_argument('--tokenizer_path', type=str, default=None, help="path to chat version tokenizer") - parser.add_argument('--temperature', type=float, default=0.8, help="set temperature") - parser.add_argument('--top_p', type=float, default=0.95, help="set top p value") - parser.add_argument('--top_k', type=int, default=50, help="set top k value") - parser.add_argument('--do_sample', type=bool, default=True, help="whether turn on do_sample or not") - parser.add_argument('--length_penalty', type=float, default=1.2, help="set length penalty") - parser.add_argument('--max_new_tokens', type=int, default=512, help="set max new tokens") + parser.add_argument("--model_path", type=str, default=None, help="path to chat version model") + parser.add_argument("--tokenizer_path", type=str, default=None, help="path to chat version tokenizer") + parser.add_argument("--temperature", type=float, default=0.8, help="set temperature") + parser.add_argument("--top_p", type=float, default=0.95, help="set top p value") + parser.add_argument("--top_k", type=int, default=50, help="set top k value") + parser.add_argument("--do_sample", type=bool, default=True, help="whether turn on do_sample or not") + parser.add_argument("--length_penalty", type=float, default=1.2, help="set length penalty") + parser.add_argument("--max_new_tokens", type=int, default=512, help="set max new tokens") args = parser.parse_args() - main(args) \ No newline at end of file + main(args) diff --git a/applications/Colossal-LLaMA-2/version.txt b/applications/Colossal-LLaMA-2/version.txt index 8a9ecc2ea99d..8acdd82b765e 100644 --- a/applications/Colossal-LLaMA-2/version.txt +++ b/applications/Colossal-LLaMA-2/version.txt @@ -1 +1 @@ -0.0.1 \ No newline at end of file +0.0.1 diff --git a/applications/ColossalEval/examples/dataset_evaluation/inference.py b/applications/ColossalEval/examples/dataset_evaluation/inference.py index a340f3bfd281..13bbb12b6990 100644 --- a/applications/ColossalEval/examples/dataset_evaluation/inference.py +++ b/applications/ColossalEval/examples/dataset_evaluation/inference.py @@ -3,7 +3,6 @@ import os from typing import Dict, List -import torch import torch.distributed as dist from colossal_eval import dataset, models, utils diff --git a/applications/ColossalMoE/README.md b/applications/ColossalMoE/README.md index be50a8f9f25111edabd7c13660bb976f9cc1f252..ba864d1dff8b2c52b3a5c45261f37586ecdb5bc1 100644 GIT binary patch delta 9 QcmX?QblPZxmn0(>02EsT7ytkO delta 7 OcmX?YbjoOhmm~lVHv None: else: # May ba a directory, we strictly follow the glob path and will not load files in subdirectories pass - + def clear(self): """ Clear loaded data. diff --git a/applications/ColossalQA/colossalqa/data_loader/table_dataloader.py b/applications/ColossalQA/colossalqa/data_loader/table_dataloader.py index cad48254498e..29542466fa8f 100644 --- a/applications/ColossalQA/colossalqa/data_loader/table_dataloader.py +++ b/applications/ColossalQA/colossalqa/data_loader/table_dataloader.py @@ -1,39 +1,40 @@ -''' +""" Class for loading table type data. please refer to Pandas-Input/Output for file format details. -''' +""" -import os import glob +import os + import pandas as pd -from sqlalchemy import create_engine -from colossalqa.utils import drop_table from colossalqa.mylogging import get_logger +from colossalqa.utils import drop_table +from sqlalchemy import create_engine logger = get_logger() -SUPPORTED_DATA_FORMAT = ['.csv','.xlsx', '.xls','.json','.html','.h5', '.hdf5','.parquet','.feather','.dta'] +SUPPORTED_DATA_FORMAT = [".csv", ".xlsx", ".xls", ".json", ".html", ".h5", ".hdf5", ".parquet", ".feather", ".dta"] + class TableLoader: - ''' + """ Load tables from different files and serve a sql database for database operations - ''' - def __init__(self, files: str, - sql_path:str='sqlite:///mydatabase.db', - verbose=False, **kwargs) -> None: - ''' + """ + + def __init__(self, files: str, sql_path: str = "sqlite:///mydatabase.db", verbose=False, **kwargs) -> None: + """ Args: files: list of files (list[file path, name]) sql_path: how to serve the sql database - **kwargs: keyword type arguments, useful for certain document types - ''' + **kwargs: keyword type arguments, useful for certain document types + """ self.data = {} self.verbose = verbose self.sql_path = sql_path self.kwargs = kwargs self.sql_engine = create_engine(self.sql_path) drop_table(self.sql_engine) - + self.sql_engine = create_engine(self.sql_path) for item in files: path = item[0] @@ -42,68 +43,68 @@ def __init__(self, files: str, raise FileNotFoundError(f"{path} doesn't exists") if not any([path.endswith(i) for i in SUPPORTED_DATA_FORMAT]): raise TypeError(f"{path} not supported. Supported type {SUPPORTED_DATA_FORMAT}") - + logger.info("loading data", verbose=self.verbose) self.load_data(path) logger.info("data loaded", verbose=self.verbose) self.to_sql(path, dataset_name) def load_data(self, path): - ''' + """ Load data and serve the data as sql database. Data must be in pandas format - ''' + """ files = [] # Handle glob expression try: files = glob.glob(path) except Exception as e: logger.error(e) - if len(files)==0: + if len(files) == 0: raise ValueError("Unsupported file/directory format. For directories, please use glob expression") - elif len(files)==1: + elif len(files) == 1: path = files[0] else: for file in files: self.load_data(file) - if path.endswith('.csv'): + if path.endswith(".csv"): # Load csv self.data[path] = pd.read_csv(path) - elif path.endswith('.xlsx') or path.endswith('.xls'): + elif path.endswith(".xlsx") or path.endswith(".xls"): # Load excel self.data[path] = pd.read_excel(path) # You can adjust the sheet_name as needed - elif path.endswith('.json'): + elif path.endswith(".json"): # Load json self.data[path] = pd.read_json(path) - elif path.endswith('.html'): + elif path.endswith(".html"): # Load html html_tables = pd.read_html(path) # Choose the desired table from the list of DataFrame objects self.data[path] = html_tables[0] # You may need to adjust this index - elif path.endswith('.h5') or path.endswith('.hdf5'): + elif path.endswith(".h5") or path.endswith(".hdf5"): # Load h5 - self.data[path] = pd.read_hdf(path, key=self.kwargs.get('key', 'data')) # You can adjust the key as needed - elif path.endswith('.parquet'): + self.data[path] = pd.read_hdf(path, key=self.kwargs.get("key", "data")) # You can adjust the key as needed + elif path.endswith(".parquet"): # Load parquet - self.data[path] = pd.read_parquet(path, engine='fastparquet') - elif path.endswith('.feather'): + self.data[path] = pd.read_parquet(path, engine="fastparquet") + elif path.endswith(".feather"): # Load feather self.data[path] = pd.read_feather(path) - elif path.endswith('.dta'): + elif path.endswith(".dta"): # Load dta self.data[path] = pd.read_stata(path) else: raise ValueError("Unsupported file format") - + def to_sql(self, path, table_name): - ''' + """ Serve the data as sql database. - ''' - self.data[path].to_sql(table_name, con=self.sql_engine, if_exists='replace', index=False) + """ + self.data[path].to_sql(table_name, con=self.sql_engine, if_exists="replace", index=False) logger.info(f"Loaded to Sqlite3\nPath: {path}", verbose=self.verbose) return self.sql_path - + def get_sql_path(self): return self.sql_path @@ -113,7 +114,3 @@ def __del__(self): self.sql_engine.dispose() del self.data del self.sql_engine - - - - diff --git a/applications/ColossalQA/colossalqa/local/colossalcloud_llm.py b/applications/ColossalQA/colossalqa/local/colossalcloud_llm.py index 62aead66c54b..14e33820d9c9 100644 --- a/applications/ColossalQA/colossalqa/local/colossalcloud_llm.py +++ b/applications/ColossalQA/colossalqa/local/colossalcloud_llm.py @@ -21,7 +21,7 @@ """ import json -from typing import Any, List, Mapping, Optional +from typing import Any, Mapping import requests from langchain.llms.base import LLM @@ -31,31 +31,31 @@ class ColossalCloudLLM(LLM): """ A custom LLM class that integrates LLMs running on the ColossalCloud Platform - + """ - n: int - gen_config: dict = None + + n: int + gen_config: dict = None auth_config: dict = None - valid_gen_para: list = ['max_new_tokens', 'top_k', - 'top_p', 'temperature', 'repetition_penalty'] + valid_gen_para: list = ["max_new_tokens", "top_k", "top_p", "temperature", "repetition_penalty"] def __init__(self, gen_config=None, **kwargs): """ Args: gen_config: config for generation, max_new_tokens: 50 by default - top_k: (1, vocab_size) + top_k: (1, vocab_size) top_p: (0, 1) if not None - temperature: (0, inf) if not None + temperature: (0, inf) if not None repetition_penalty: (1, inf) if not None """ super(ColossalCloudLLM, self).__init__(**kwargs) - if gen_config is None: - self.gen_config = {"max_new_tokens": 50} - else: + if gen_config is None: + self.gen_config = {"max_new_tokens": 50} + else: assert "max_new_tokens" in gen_config, "max_new_tokens is a compulsory key in the gen config" self.gen_config = gen_config - + @property def _identifying_params(self) -> Mapping[str, Any]: """Get the identifying parameters.""" @@ -63,17 +63,17 @@ def _identifying_params(self) -> Mapping[str, Any]: @property def _llm_type(self) -> str: - return 'ColossalCloudLLM' - + return "ColossalCloudLLM" + def set_auth_config(self, **kwargs): url = get_from_dict_or_env(kwargs, "url", "URL") host = get_from_dict_or_env(kwargs, "host", "HOST") - + auth_config = {} - auth_config['endpoint'] = url - auth_config['Host'] = host + auth_config["endpoint"] = url + auth_config["Host"] = host self.auth_config = auth_config - + def _call(self, prompt: str, stop=None, **kwargs: Any) -> str: """ Args: @@ -81,15 +81,17 @@ def _call(self, prompt: str, stop=None, **kwargs: Any) -> str: stop: A list of strings to stop generation when encountered Returns: - The string generated by the model + The string generated by the model """ # Update the generation arguments for key, value in kwargs.items(): if key not in self.valid_gen_para: - raise KeyError(f"Invalid generation parameter: '{key}'. Valid keys are: {', '.join(self.valid_gen_para)}") + raise KeyError( + f"Invalid generation parameter: '{key}'. Valid keys are: {', '.join(self.valid_gen_para)}" + ) if key in self.gen_config: self.gen_config[key] = value - + resp_text = self.text_completion(prompt, self.gen_config, self.auth_config) # TODO: This may cause excessive tokens count if stop is not None: @@ -97,29 +99,19 @@ def _call(self, prompt: str, stop=None, **kwargs: Any) -> str: if stopping_words in resp_text: resp_text = resp_text.split(stopping_words)[0] return resp_text - def text_completion(self, prompt, gen_config, auth_config): # Complusory Parameters - endpoint = auth_config.pop('endpoint') - max_new_tokens = gen_config.pop('max_new_tokens') + endpoint = auth_config.pop("endpoint") + max_new_tokens = gen_config.pop("max_new_tokens") # Optional Parameters - optional_params = ['top_k', 'top_p', 'temperature', 'repetition_penalty'] # Self.optional + optional_params = ["top_k", "top_p", "temperature", "repetition_penalty"] # Self.optional gen_config = {key: gen_config[key] for key in optional_params if key in gen_config} # Define the data payload - data = { - "max_new_tokens": max_new_tokens, - "history": [ - {"instruction": prompt, "response": ""} - ], - **gen_config - } - headers = { - "Content-Type": "application/json", - **auth_config # 'Host', - } + data = {"max_new_tokens": max_new_tokens, "history": [{"instruction": prompt, "response": ""}], **gen_config} + headers = {"Content-Type": "application/json", **auth_config} # 'Host', # Make the POST request response = requests.post(endpoint, headers=headers, data=json.dumps(data)) - response.raise_for_status() # raise error if return code is not 200(success) + response.raise_for_status() # raise error if return code is not 200(success) # Check the response return response.text diff --git a/applications/ColossalQA/colossalqa/local/llm.py b/applications/ColossalQA/colossalqa/local/llm.py index 0aa383e9d0b9..bab702d14b13 100644 --- a/applications/ColossalQA/colossalqa/local/llm.py +++ b/applications/ColossalQA/colossalqa/local/llm.py @@ -193,4 +193,3 @@ def set_host_port(self, host: str = "localhost", port: int = 8077, **kwargs) -> def _identifying_params(self) -> Mapping[str, int]: """Get the identifying parameters.""" return {"n": self.n} - diff --git a/applications/ColossalQA/colossalqa/prompt/prompt.py b/applications/ColossalQA/colossalqa/prompt/prompt.py index d62249ba9c51..a8aeded3a4a4 100644 --- a/applications/ColossalQA/colossalqa/prompt/prompt.py +++ b/applications/ColossalQA/colossalqa/prompt/prompt.py @@ -4,7 +4,6 @@ from langchain.prompts.prompt import PromptTemplate - # Below are Chinese retrieval qa prompts _CUSTOM_SUMMARIZER_TEMPLATE_ZH = """请递进式地总结所提供的当前对话，将当前对话的摘要内容添加到先前已有的摘要上，返回一个融合了当前对话的新的摘要。 diff --git a/applications/ColossalQA/colossalqa/retriever.py b/applications/ColossalQA/colossalqa/retriever.py index c891cb613bd6..22a75050f03b 100644 --- a/applications/ColossalQA/colossalqa/retriever.py +++ b/applications/ColossalQA/colossalqa/retriever.py @@ -99,13 +99,7 @@ def add_documents( def clear_documents(self): """Clear all document vectors from database""" for source in self.vector_stores: - index( - [], - self.record_managers[source], - self.vector_stores[source], - cleanup="full", - source_id_key="source" - ) + index([], self.record_managers[source], self.vector_stores[source], cleanup="full", source_id_key="source") self.vector_stores = {} self.sql_index_database = {} self.record_managers = {} diff --git a/applications/ColossalQA/data/data_sample/companies.txt b/applications/ColossalQA/data/data_sample/companies.txt index 05c6148f18a5..4b297e6ff4fe 100644 --- a/applications/ColossalQA/data/data_sample/companies.txt +++ b/applications/ColossalQA/data/data_sample/companies.txt @@ -1,6 +1,6 @@ Overview The Straits Times is the English flagship daily of SPH Media, one of the leading media companies in Asia. Launched on July 15, 1845, its comprehensive coverage of news from home and around the world makes The Straits Times the most-read newspaper in Singapore. Quality news, in-depth analyses, impactful commentaries and breaking stories are packaged to give readers riveting accounts of events in Singapore, the region, and beyond. The most read newspaper in Singapore, both in terms of print and digital, it reaches 1.33 million people every day. The Straits Times'​ key strength is in its world class coverage of news outside Singapore. With 20 bureaus in major cities around the world, The Straits Times correspondents bring world news to readers on a Singapore platter, helping readers to appreciate world events from a Singaporean perspective. Website http://www.straitstimes.com Phone 63196319Phone number is 63196319 Industry Newspaper Publishing Company size 1,001-5,000 employees 183 on LinkedIn Includes members with current employer listed as The Straits Times, including part-time roles. Headquarters Singapore, Singapore Founded 1845 Specialties News and Digital media -About With over 500 properties worldwide, Marriott Hotels has reimagined hospitality to exceed the expectations of business, group, and leisure travelers. +About With over 500 properties worldwide, Marriott Hotels has reimagined hospitality to exceed the expectations of business, group, and leisure travelers. Marriott Hotels, Marriott’s flagship brand of quality-tier, full-service hotels and resorts, provides consistent, dependable and genuinely caring experiences to guests on their terms. Marriott is a brilliant host to guests who effortlessly blend life and work, and who are inspired by how modern travel enhances them both. Our hotels offer warm, professional service; sophisticated yet functional guest room design; lobby spaces that facilitate working, dining and socializing; restaurants and bars serving international cuisine prepared simply and from the freshest ingredients; meeting and event spaces and services that are gold standard; and expansive, 24-hour fitness facilities. -Overview AERCO International, Inc. is a recognized leader in delivering cost-effective, condensing commercial boilers, high-efficiency water heaters across a variety of markets including education, lodging, government, office buildings, healthcare, industrial and multifamily housing. AERCO's system design approach provides customer-specific solutions that deliver superior building performance at a lower operating cost while assuring uptime reliability. When AERCO was founded in 1949, it introduced a revolutionary design for an indirect-fired water heater that heated water on demand, and without storage, at a controlled temperature. This innovation became today's standard for water heaters, maximizing the recovery of latent heat energy and significantly increasing operating efficiency. AERCO continued to innovate and in 1988, introduced the first condensing and fully modulating boiler and water heater to the commercial market. The modulating capability of these products, still unsurpassed more than 25 years later, matches the equipment's output to real-time heating demand, ensuring the units draw no more fuel to operate than is absolutely necessary. This not only saves precious energy, but also ensures money doesn't needlessly disappear "up the stack."​ AERCO differentiates itself through a solution-based model, leveraging decades of engineering experience and industry application expertise to understand each customer’s unique needs. By partnering directly with customers and end-users to understand their project-specific requirements, AERCO provides tailored application solutions that are comprised of original product technologies including high efficiency condensing products, compact footprints, high turndown ratios, unique fuel delivery, leading control systems and proprietary design elements that combine to deliver up to 99% efficiency. Website http://www.aerco.com Phone 845-580-8000Phone number is 845-580-8000 Industry Industrial Machinery Manufacturing Company size 51-200 employees 119 on LinkedIn Includes members with current employer listed as AERCO International, Inc., including part-time roles. Headquarters Blauvelt, NY Founded 1949 Specialties Leading manufacturer of condensing boilers, water heating and energy recovery products and The originator of semi-instantaneous water heating +Overview AERCO International, Inc. is a recognized leader in delivering cost-effective, condensing commercial boilers, high-efficiency water heaters across a variety of markets including education, lodging, government, office buildings, healthcare, industrial and multifamily housing. AERCO's system design approach provides customer-specific solutions that deliver superior building performance at a lower operating cost while assuring uptime reliability. When AERCO was founded in 1949, it introduced a revolutionary design for an indirect-fired water heater that heated water on demand, and without storage, at a controlled temperature. This innovation became today's standard for water heaters, maximizing the recovery of latent heat energy and significantly increasing operating efficiency. AERCO continued to innovate and in 1988, introduced the first condensing and fully modulating boiler and water heater to the commercial market. The modulating capability of these products, still unsurpassed more than 25 years later, matches the equipment's output to real-time heating demand, ensuring the units draw no more fuel to operate than is absolutely necessary. This not only saves precious energy, but also ensures money doesn't needlessly disappear "up the stack."​ AERCO differentiates itself through a solution-based model, leveraging decades of engineering experience and industry application expertise to understand each customer’s unique needs. By partnering directly with customers and end-users to understand their project-specific requirements, AERCO provides tailored application solutions that are comprised of original product technologies including high efficiency condensing products, compact footprints, high turndown ratios, unique fuel delivery, leading control systems and proprietary design elements that combine to deliver up to 99% efficiency. Website http://www.aerco.com Phone 845-580-8000Phone number is 845-580-8000 Industry Industrial Machinery Manufacturing Company size 51-200 employees 119 on LinkedIn Includes members with current employer listed as AERCO International, Inc., including part-time roles. Headquarters Blauvelt, NY Founded 1949 Specialties Leading manufacturer of condensing boilers, water heating and energy recovery products and The originator of semi-instantaneous water heating Prince PLC: Overview We are a global leader of quality water solutions for residential, industrial, municipal, and commercial settings. Our family of brands offers one of the most varied product lines in the world, with world-class, water-related solutions focused on: • Plumbing & Flow Control • Water Quality & Conditioning • Water Reuse & Drainage • HVAC • Municipal Waterworks Strategic Goals Watts Water is traded on the New York Stock Exchange under the symbol “WTS.” As a public company, growing shareholder value is critical. To that end, we focus on a five-part Global Strategy: Growth, Commercial Excellence, Operational Excellence, “One Watts Water,” and a Talent & Performance Culture. Follow us on all social media platforms @WattsWater Website http://www.watts.com/ Industry Wholesale Building Materials Company size 5,001-10,000 employees 2,248 on LinkedIn Includes members with current employer listed as Watts Water Technologies, including part-time roles. Headquarters North Andover, MA Specialties Plumbing, HVAC, Water Quality, Gas, Conditioning, Waterworks, and Drainage -About Courtyard Hotels is Marriott International’s largest hotel brand, with more than 1,100 hotels in over 50 countries worldwide. So, no matter where passion takes you, you’ll find us there to help you follow it. Proud members of Marriott Bonvoy. \ No newline at end of file +About Courtyard Hotels is Marriott International’s largest hotel brand, with more than 1,100 hotels in over 50 countries worldwide. So, no matter where passion takes you, you’ll find us there to help you follow it. Proud members of Marriott Bonvoy. diff --git a/applications/ColossalQA/data/data_sample/companies_zh.txt b/applications/ColossalQA/data/data_sample/companies_zh.txt index a67a93590ee8..511873bde686 100644 --- a/applications/ColossalQA/data/data_sample/companies_zh.txt +++ b/applications/ColossalQA/data/data_sample/companies_zh.txt @@ -3,4 +3,4 @@ 万豪酒店(Marriott Hotels)是万豪旗下优质、全方位服务酒店和度假村的旗舰品牌，为客人提供始终如一、可靠和真诚关怀的体验。万豪是一个出色的主人，客人可以轻松地将生活和工作融合在一起，并受到现代旅行如何增强两者的启发。我们的酒店提供热情、专业的服务;精致而实用的客房设计;大堂空间，方便工作、餐饮和社交;餐厅和酒吧提供简单的国际美食和最新鲜的食材;会议及活动场地及服务均属黄金标准;还有宽敞的24小时健身设施。 AERCO International, Inc.是公认的领导者，为教育、住宿、政府、办公楼、医疗保健、工业和多户住宅等各种市场提供具有成本效益的冷凝商用锅炉和高效热水器。AERCO的系统设计方法为客户提供特定的解决方案，以较低的运营成本提供卓越的建筑性能，同时确保正常运行时间的可靠性。AERCO成立于1949年，它推出了一种革命性的设计，用于间接燃烧热水器，在控制温度下按需加热水，而无需储存。这一创新成为当今热水器的标准，最大限度地回收潜热能量，显著提高运行效率。AERCO不断创新，并于1988年向商业市场推出了第一台冷凝和全调制锅炉和热水器。这些产品的调制能力，在超过25年后仍然无与伦比，使设备的输出与实时加热需求相匹配，确保机组不会消耗更多的燃料来运行，除非绝对必要。这不仅节省了宝贵的能源，还确保了钱不会不必要地消失在“堆栈”上。AERCO通过基于解决方案的模式脱颖而出，利用数十年的工程经验和行业应用专业知识来了解每个客户的独特需求。通过与客户和最终用户直接合作，了解他们的项目具体要求，AERCO提供量身定制的应用解决方案，这些解决方案由原创产品技术组成，包括高效冷凝产品，紧凑的足迹，高降压比，独特的燃料输送，领先的控制系统和专有设计元素，结合起来可提供高达99%的效率。网址http://www.aerco.com电话845-580- 8000电话号码845-580-8000工业工业机械制造公司规模51-200名员工LinkedIn上包括当前雇主AERCO International, Inc的成员，包括兼职职位。总部成立于1949年，纽约州布劳维尔特，专长:冷凝锅炉，水加热和能源回收产品的领先制造商，半瞬时水加热的鼻祖 Prince PLC:概述Prince PLC是为住宅、工业、市政和商业环境提供优质水解决方案的全球领导者。我们的品牌家族提供世界上最多样化的产品线之一，拥有世界级的水相关解决方案，专注于:•管道和流量控制•水质和调理•水再利用和排水•hvac•市政水务战略目标瓦茨水务在纽约证券交易所上市，代码为“WTS”。作为一家上市公司，股东价值的增长至关重要。为此，我们将重点放在五部分全球战略上:增长、卓越商业、卓越运营、“一瓦茨水”以及人才与绩效文化。在所有社交媒体平台关注我们@WattsWater网站http://www.watts.com/行业批发建材公司规模5,001-10,000名员工领英2,248名包括目前雇主为WattsWater Technologies的成员，包括兼职职位。总部北安多弗，MA专业管道，暖通空调，水质，气体，空调，自来水厂和排水 -万怡酒店是万豪国际最大的酒店品牌，在全球50多个国家拥有1100多家酒店。所以，无论你的激情带你去哪里，你都会发现我们会帮助你追随它。万豪酒店的骄傲会员。 \ No newline at end of file +万怡酒店是万豪国际最大的酒店品牌，在全球50多个国家拥有1100多家酒店。所以，无论你的激情带你去哪里，你都会发现我们会帮助你追随它。万豪酒店的骄傲会员。 diff --git a/applications/ColossalQA/data/data_sample/csv_organization_100.csv b/applications/ColossalQA/data/data_sample/csv_organization_100.csv index dbe97d5fd774..5d88bfb94573 100644 --- a/applications/ColossalQA/data/data_sample/csv_organization_100.csv +++ b/applications/ColossalQA/data/data_sample/csv_organization_100.csv @@ -98,4 +98,4 @@ Index,Organization Id,Company Name,Website,Country,Description,Founded,Industry, 97,BA6Cd9Dae2Efd62,Good Ltd,http://duffy.com/,Anguilla,Reverse-engineered composite moratorium,1971,Consumer Services,4292 98,E7df80C60Abd7f9,Clements-Espinoza,http://www.flowers.net/,Falkland Islands (Malvinas),Progressive modular hub,1991,Broadcast Media,236 99,AFc285dbE2fEd24,Mendez Inc,https://www.burke.net/,Kyrgyz Republic,User-friendly exuding migration,1993,Education Management,339 -100,e9eB5A60Cef8354,Watkins-Kaiser,http://www.herring.com/,Togo,Synergistic background access,2009,Financial Services,2785 \ No newline at end of file +100,e9eB5A60Cef8354,Watkins-Kaiser,http://www.herring.com/,Togo,Synergistic background access,2009,Financial Services,2785 diff --git a/applications/ColossalQA/data/tests/64KB.json b/applications/ColossalQA/data/tests/64KB.json index 99278dc5c79a..41e1aeedf539 100644 --- a/applications/ColossalQA/data/tests/64KB.json +++ b/applications/ColossalQA/data/tests/64KB.json @@ -4,4 +4,4 @@ {"content":"Aliquam sollicitudin ante ligula, eget malesuada nibh efficitur et. Pellentesque massa sem, scelerisque sit amet odio id, cursus tempor urna. Etiam congue dignissim volutpat. Vestibulum pharetra libero et velit gravida euismod."} ], "name":"player" -} \ No newline at end of file +} diff --git a/applications/ColossalQA/data/tests/companies.csv b/applications/ColossalQA/data/tests/companies.csv index 93dcac9f39ae..a111992d7fcd 100644 --- a/applications/ColossalQA/data/tests/companies.csv +++ b/applications/ColossalQA/data/tests/companies.csv @@ -1,101 +1,101 @@ -Index,Organization Id,Name,Website,Country,Description,Founded,Industry,Number of employees -1,FAB0d41d5b5d22c,Ferrell LLC,https://price.net/,Papua New Guinea,Horizontal empowering knowledgebase,1990,Plastics,3498 -2,6A7EdDEA9FaDC52,"Mckinney, Riley and Day",http://www.hall-buchanan.info/,Finland,User-centric system-worthy leverage,2015,Glass / Ceramics / Concrete,4952 -3,0bFED1ADAE4bcC1,Hester Ltd,http://sullivan-reed.com/,China,Switchable scalable moratorium,1971,Public Safety,5287 -4,2bFC1Be8a4ce42f,Holder-Sellers,https://becker.com/,Turkmenistan,De-engineered systemic artificial intelligence,2004,Automotive,921 -5,9eE8A6a4Eb96C24,Mayer Group,http://www.brewer.com/,Mauritius,Synchronized needs-based challenge,1991,Transportation,7870 -6,cC757116fe1C085,Henry-Thompson,http://morse.net/,Bahamas,Face-to-face well-modulated customer loyalty,1992,Primary / Secondary Education,4914 -7,219233e8aFF1BC3,Hansen-Everett,https://www.kidd.org/,Pakistan,Seamless disintermediate collaboration,2018,Publishing Industry,7832 -8,ccc93DCF81a31CD,Mcintosh-Mora,https://www.brooks.com/,Heard Island and McDonald Islands,Centralized attitude-oriented capability,1970,Import / Export,4389 -9,0B4F93aA06ED03e,Carr Inc,http://ross.com/,Kuwait,Distributed impactful customer loyalty,1996,Plastics,8167 -10,738b5aDe6B1C6A5,Gaines Inc,http://sandoval-hooper.com/,Uzbekistan,Multi-lateral scalable protocol,1997,Outsourcing / Offshoring,9698 -11,AE61b8Ffebbc476,Kidd Group,http://www.lyons.com/,Bouvet Island (Bouvetoya),Proactive foreground paradigm,2001,Primary / Secondary Education,7473 -12,eb3B7D06cCdD609,Crane-Clarke,https://www.sandoval.com/,Denmark,Front-line clear-thinking encryption,2014,Food / Beverages,9011 -13,8D0c29189C9798B,"Keller, Campos and Black",https://www.garner.info/,Liberia,Ameliorated directional emulation,2020,Museums / Institutions,2862 -14,D2c91cc03CA394c,Glover-Pope,http://www.silva.biz/,United Arab Emirates,Persevering contextually-based approach,2013,Medical Practice,9079 -15,C8AC1eaf9C036F4,Pacheco-Spears,https://aguilar.com/,Sweden,Secured logistical synergy,1984,Maritime,769 -16,b5D10A14f7a8AfE,Hodge-Ayers,http://www.archer-elliott.com/,Honduras,Future-proofed radical implementation,1990,Facilities Services,8508 -17,68139b5C4De03B4,"Bowers, Guerra and Krause",http://www.carrillo-nicholson.com/,Uganda,De-engineered transitional strategy,1972,Primary / Secondary Education,6986 -18,5c2EffEfdba2BdF,Mckenzie-Melton,http://montoya-thompson.com/,Hong Kong,Reverse-engineered heuristic alliance,1998,Investment Management / Hedge Fund / Private Equity,4589 -19,ba179F19F7925f5,Branch-Mann,http://www.lozano.com/,Botswana,Adaptive intangible frame,1999,Architecture / Planning,7961 -20,c1Ce9B350BAc66b,Weiss and Sons,https://barrett.com/,Korea,Sharable optimal functionalities,2011,Plastics,5984 -21,8de40AC4e6EaCa4,"Velez, Payne and Coffey",http://burton.com/,Luxembourg,Mandatory coherent synergy,1986,Wholesale,5010 -22,Aad86a4F0385F2d,Harrell LLC,http://www.frey-rosario.com/,Guadeloupe,Reverse-engineered mission-critical moratorium,2018,Construction,2185 -23,22aC3FFd64fD703,"Eaton, Reynolds and Vargas",http://www.freeman.biz/,Monaco,Self-enabling multi-tasking process improvement,2014,Luxury Goods / Jewelry,8987 -24,5Ec4C272bCf085c,Robbins-Cummings,http://donaldson-wilkins.com/,Belgium,Organic non-volatile hierarchy,1991,Pharmaceuticals,5038 -25,5fDBeA8BB91a000,Jenkins Inc,http://www.kirk.biz/,South Africa,Front-line systematic help-desk,2002,Insurance,1215 -26,dFfD6a6F9AC2d9C,"Greene, Benjamin and Novak",http://www.kent.net/,Romania,Centralized leadingedge moratorium,2012,Museums / Institutions,4941 -27,4B217cC5a0674C5,"Dickson, Richmond and Clay",http://everett.com/,Czech Republic,Team-oriented tangible complexity,1980,Real Estate / Mortgage,3122 -28,88b1f1cDcf59a37,Prince-David,http://thompson.com/,Christmas Island,Virtual holistic methodology,1970,Banking / Mortgage,1046 -29,f9F7bBCAEeC360F,Ayala LLC,http://www.zhang.com/,Philippines,Open-source zero administration hierarchy,2021,Legal Services,7664 -30,7Cb3AeFcE4Ba31e,Rivas Group,https://hebert.org/,Australia,Open-architected well-modulated capacity,1998,Logistics / Procurement,4155 -31,ccBcC32adcbc530,"Sloan, Mays and Whitehead",http://lawson.com/,Chad,Face-to-face high-level conglomeration,1997,Civil Engineering,365 -32,f5afd686b3d05F5,"Durham, Allen and Barnes",http://chan-stafford.org/,Zimbabwe,Synergistic web-enabled framework,1993,Mechanical or Industrial Engineering,6135 -33,38C6cfC5074Fa5e,Fritz-Franklin,http://www.lambert.com/,Nepal,Automated 4thgeneration website,1972,Hospitality,4516 -34,5Cd7efccCcba38f,Burch-Ewing,http://cline.net/,Taiwan,User-centric 4thgeneration system engine,1981,Venture Capital / VC,7443 -35,9E6Acb51e3F9d6F,"Glass, Barrera and Turner",https://dunlap.com/,Kyrgyz Republic,Multi-channeled 3rdgeneration open system,2020,Utilities,2610 -36,4D4d7E18321eaeC,Pineda-Cox,http://aguilar.org/,Bolivia,Fundamental asynchronous capability,2010,Human Resources / HR,1312 -37,485f5d06B938F2b,"Baker, Mccann and Macdonald",http://www.anderson-barker.com/,Kenya,Cross-group user-facing focus group,2013,Legislative Office,1638 -38,19E3a5Bf6dBDc4F,Cuevas-Moss,https://dodson-castaneda.net/,Guatemala,Extended human-resource intranet,1994,Music,9995 -39,6883A965c7b68F7,Hahn PLC,http://newman.com/,Belarus,Organic logistical leverage,2012,Electrical / Electronic Manufacturing,3715 -40,AC5B7AA74Aa4A2E,"Valentine, Ferguson and Kramer",http://stuart.net/,Jersey,Centralized secondary time-frame,1997,Non - Profit / Volunteering,3585 -41,decab0D5027CA6a,Arroyo Inc,https://www.turner.com/,Grenada,Managed demand-driven website,2006,Writing / Editing,9067 -42,dF084FbBb613eea,Walls LLC,http://www.reese-vasquez.biz/,Cape Verde,Self-enabling fresh-thinking installation,1989,Investment Management / Hedge Fund / Private Equity,1678 -43,A2D89Ab9bCcAd4e,"Mitchell, Warren and Schneider",https://fox.biz/,Trinidad and Tobago,Enhanced intangible time-frame,2021,Capital Markets / Hedge Fund / Private Equity,3816 -44,77aDc905434a49f,Prince PLC,https://www.watts.com/,Sweden,Profit-focused coherent installation,2016,Individual / Family Services,7645 -45,235fdEFE2cfDa5F,Brock-Blackwell,http://www.small.com/,Benin,Secured foreground emulation,1986,Online Publishing,7034 -46,1eD64cFe986BBbE,Walton-Barnett,https://ashley-schaefer.com/,Western Sahara,Right-sized clear-thinking flexibility,2001,Luxury Goods / Jewelry,1746 -47,CbBbFcdd0eaE2cF,Bartlett-Arroyo,https://cruz.com/,Northern Mariana Islands,Realigned didactic function,1976,Civic / Social Organization,3987 -48,49aECbDaE6aBD53,"Wallace, Madden and Morris",http://www.blevins-fernandez.biz/,Germany,Persistent real-time customer loyalty,2016,Pharmaceuticals,9443 -49,7b3fe6e7E72bFa4,Berg-Sparks,https://cisneros-love.com/,Canada,Stand-alone static implementation,1974,Arts / Crafts,2073 -50,c6DedA82A8aef7E,Gonzales Ltd,http://bird.com/,Tonga,Managed human-resource policy,1988,Consumer Goods,9069 -51,7D9FBF85cdC3871,Lawson and Sons,https://www.wong.com/,French Southern Territories,Compatible analyzing intranet,2021,Arts / Crafts,3527 -52,7dd18Fb7cB07b65,"Mcguire, Mcconnell and Olsen",https://melton-briggs.com/,Korea,Profound client-server frame,1988,Printing,8445 -53,EF5B55FadccB8Fe,Charles-Phillips,https://bowman.com/,Cote d'Ivoire,Monitored client-server implementation,2012,Mental Health Care,3450 -54,f8D4B99e11fAF5D,Odom Ltd,https://www.humphrey-hess.com/,Cote d'Ivoire,Advanced static process improvement,2012,Management Consulting,1825 -55,e24D21BFd3bF1E5,Richard PLC,https://holden-coleman.net/,Mayotte,Object-based optimizing model,1971,Broadcast Media,4942 -56,B9BdfEB6D3Ca44E,Sampson Ltd,https://blevins.com/,Cayman Islands,Intuitive local adapter,2005,Farming,1418 -57,2a74D6f3D3B268e,"Cherry, Le and Callahan",https://waller-delacruz.biz/,Nigeria,Universal human-resource collaboration,2017,Entertainment / Movie Production,7202 -58,Bf3F3f62c8aBC33,Cherry PLC,https://www.avila.info/,Marshall Islands,Persistent tertiary website,1980,Plastics,8245 -59,aeBe26B80a7a23c,Melton-Nichols,https://kennedy.com/,Palau,User-friendly clear-thinking productivity,2021,Legislative Office,8741 -60,aAeb29ad43886C6,Potter-Walsh,http://thomas-french.org/,Turkey,Optional non-volatile open system,2008,Human Resources / HR,6923 -61,bD1bc6bB6d1FeD3,Freeman-Chen,https://mathis.com/,Timor-Leste,Phased next generation adapter,1973,International Trade / Development,346 -62,EB9f456e8b7022a,Soto Group,https://norris.info/,Vietnam,Enterprise-wide executive installation,1988,Business Supplies / Equipment,9097 -63,Dfef38C51D8DAe3,"Poole, Cruz and Whitney",https://reed.info/,Reunion,Balanced analyzing groupware,1978,Marketing / Advertising / Sales,2992 -64,055ffEfB2Dd95B0,Riley Ltd,http://wiley.com/,Brazil,Optional exuding superstructure,1986,Textiles,9315 -65,cBfe4dbAE1699da,"Erickson, Andrews and Bailey",https://www.hobbs-grant.com/,Eritrea,Vision-oriented secondary project,2014,Consumer Electronics,7829 -66,fdFbecbadcdCdf1,"Wilkinson, Charles and Arroyo",http://hunter-mcfarland.com/,United States Virgin Islands,Assimilated 24/7 archive,1996,Building Materials,602 -67,5DCb8A5a5ca03c0,Floyd Ltd,http://www.whitney.com/,Falkland Islands (Malvinas),Function-based fault-tolerant concept,2017,Public Relations / PR,2911 -68,ce57DCbcFD6d618,Newman-Galloway,https://www.scott.com/,Luxembourg,Enhanced foreground collaboration,1987,Information Technology / IT,3934 -69,5aaD187dc929371,Frazier-Butler,https://www.daugherty-farley.info/,Northern Mariana Islands,Persistent interactive circuit,1972,Outsourcing / Offshoring,5130 -70,902D7Ac8b6d476b,Newton Inc,https://www.richmond-manning.info/,Netherlands Antilles,Fundamental stable info-mediaries,1976,Military Industry,563 -71,32BB9Ff4d939788,Duffy-Levy,https://www.potter.com/,Guernsey,Diverse exuding installation,1982,Wireless,6146 -72,adcB0afbE58bAe3,Wagner LLC,https://decker-esparza.com/,Uruguay,Reactive attitude-oriented toolset,1987,International Affairs,6874 -73,dfcA1c84AdB61Ac,Mccall-Holmes,http://www.dean.com/,Benin,Object-based value-added database,2009,Legal Services,696 -74,208044AC2fe52F3,Massey LLC,https://frazier.biz/,Suriname,Configurable zero administration Graphical User Interface,1986,Accounting,5004 -75,f3C365f0c1A0623,Hicks LLC,http://alvarez.biz/,Pakistan,Quality-focused client-server Graphical User Interface,1970,Computer Software / Engineering,8480 -76,ec5Bdd3CBAfaB93,"Cole, Russell and Avery",http://www.blankenship.com/,Mongolia,De-engineered fault-tolerant challenge,2000,Law Enforcement,7012 -77,DDB19Be7eeB56B4,Cummings-Rojas,https://simon-pearson.com/,Svalbard & Jan Mayen Islands,User-centric modular customer loyalty,2012,Financial Services,7529 -78,dd6CA3d0bc3cAfc,"Beasley, Greene and Mahoney",http://www.petersen-lawrence.com/,Togo,Extended content-based methodology,1976,Religious Institutions,869 -79,A0B9d56e61070e3,"Beasley, Sims and Allison",http://burke.info/,Latvia,Secured zero tolerance hub,1972,Facilities Services,6182 -80,cBa7EFe5D05Adaf,Crawford-Rivera,https://black-ramirez.org/,Cuba,Persevering exuding budgetary management,1999,Online Publishing,7805 -81,Ea3f6D52Ec73563,Montes-Hensley,https://krueger.org/,Liechtenstein,Multi-tiered secondary productivity,2009,Printing,8433 -82,bC0CEd48A8000E0,Velazquez-Odom,https://stokes.com/,Djibouti,Streamlined 6thgeneration function,2002,Alternative Dispute Resolution,4044 -83,c89b9b59BC4baa1,Eaton-Morales,https://www.reeves-graham.com/,Micronesia,Customer-focused explicit frame,1990,Capital Markets / Hedge Fund / Private Equity,7013 -84,FEC51bce8421a7b,"Roberson, Pennington and Palmer",http://www.keith-fisher.com/,Cameroon,Adaptive bi-directional hierarchy,1993,Telecommunications,5571 -85,e0E8e27eAc9CAd5,"George, Russo and Guerra",https://drake.com/,Sweden,Centralized non-volatile capability,1989,Military Industry,2880 -86,B97a6CF9bf5983C,Davila Inc,https://mcconnell.info/,Cocos (Keeling) Islands,Profit-focused dedicated frame,2017,Consumer Electronics,2215 -87,a0a6f9b3DbcBEb5,Mays-Preston,http://www.browning-key.com/,Mali,User-centric heuristic focus group,2006,Military Industry,5786 -88,8cC1bDa330a5871,Pineda-Morton,https://www.carr.com/,United States Virgin Islands,Grass-roots methodical info-mediaries,1991,Printing,6168 -89,ED889CB2FE9cbd3,Huang and Sons,https://www.bolton.com/,Eritrea,Re-contextualized dynamic hierarchy,1981,Semiconductors,7484 -90,F4Dc1417BC6cb8f,Gilbert-Simon,https://www.bradford.biz/,Burundi,Grass-roots radical parallelism,1973,Newspapers / Journalism,1927 -91,7ABc3c7ecA03B34,Sampson-Griffith,http://hendricks.org/,Benin,Multi-layered composite paradigm,1972,Textiles,3881 -92,4e0719FBE38e0aB,Miles-Dominguez,http://www.turner.com/,Gibraltar,Organized empowering forecast,1996,Civic / Social Organization,897 -93,dEbDAAeDfaed00A,Rowe and Sons,https://www.simpson.org/,El Salvador,Balanced multimedia knowledgebase,1978,Facilities Services,8172 -94,61BDeCfeFD0cEF5,"Valenzuela, Holmes and Rowland",https://www.dorsey.net/,Taiwan,Persistent tertiary focus group,1999,Transportation,1483 -95,4e91eD25f486110,"Best, Wade and Shepard",https://zimmerman.com/,Zimbabwe,Innovative background definition,1991,Gambling / Casinos,4873 -96,0a0bfFbBbB8eC7c,Holmes Group,https://mcdowell.org/,Ethiopia,Right-sized zero tolerance focus group,1975,Photography,2988 -97,BA6Cd9Dae2Efd62,Good Ltd,http://duffy.com/,Anguilla,Reverse-engineered composite moratorium,1971,Consumer Services,4292 -98,E7df80C60Abd7f9,Clements-Espinoza,http://www.flowers.net/,Falkland Islands (Malvinas),Progressive modular hub,1991,Broadcast Media,236 -99,AFc285dbE2fEd24,Mendez Inc,https://www.burke.net/,Kyrgyz Republic,User-friendly exuding migration,1993,Education Management,339 -100,e9eB5A60Cef8354,Watkins-Kaiser,http://www.herring.com/,Togo,Synergistic background access,2009,Financial Services,2785 +Index,Organization Id,Name,Website,Country,Description,Founded,Industry,Number of employees +1,FAB0d41d5b5d22c,Ferrell LLC,https://price.net/,Papua New Guinea,Horizontal empowering knowledgebase,1990,Plastics,3498 +2,6A7EdDEA9FaDC52,"Mckinney, Riley and Day",http://www.hall-buchanan.info/,Finland,User-centric system-worthy leverage,2015,Glass / Ceramics / Concrete,4952 +3,0bFED1ADAE4bcC1,Hester Ltd,http://sullivan-reed.com/,China,Switchable scalable moratorium,1971,Public Safety,5287 +4,2bFC1Be8a4ce42f,Holder-Sellers,https://becker.com/,Turkmenistan,De-engineered systemic artificial intelligence,2004,Automotive,921 +5,9eE8A6a4Eb96C24,Mayer Group,http://www.brewer.com/,Mauritius,Synchronized needs-based challenge,1991,Transportation,7870 +6,cC757116fe1C085,Henry-Thompson,http://morse.net/,Bahamas,Face-to-face well-modulated customer loyalty,1992,Primary / Secondary Education,4914 +7,219233e8aFF1BC3,Hansen-Everett,https://www.kidd.org/,Pakistan,Seamless disintermediate collaboration,2018,Publishing Industry,7832 +8,ccc93DCF81a31CD,Mcintosh-Mora,https://www.brooks.com/,Heard Island and McDonald Islands,Centralized attitude-oriented capability,1970,Import / Export,4389 +9,0B4F93aA06ED03e,Carr Inc,http://ross.com/,Kuwait,Distributed impactful customer loyalty,1996,Plastics,8167 +10,738b5aDe6B1C6A5,Gaines Inc,http://sandoval-hooper.com/,Uzbekistan,Multi-lateral scalable protocol,1997,Outsourcing / Offshoring,9698 +11,AE61b8Ffebbc476,Kidd Group,http://www.lyons.com/,Bouvet Island (Bouvetoya),Proactive foreground paradigm,2001,Primary / Secondary Education,7473 +12,eb3B7D06cCdD609,Crane-Clarke,https://www.sandoval.com/,Denmark,Front-line clear-thinking encryption,2014,Food / Beverages,9011 +13,8D0c29189C9798B,"Keller, Campos and Black",https://www.garner.info/,Liberia,Ameliorated directional emulation,2020,Museums / Institutions,2862 +14,D2c91cc03CA394c,Glover-Pope,http://www.silva.biz/,United Arab Emirates,Persevering contextually-based approach,2013,Medical Practice,9079 +15,C8AC1eaf9C036F4,Pacheco-Spears,https://aguilar.com/,Sweden,Secured logistical synergy,1984,Maritime,769 +16,b5D10A14f7a8AfE,Hodge-Ayers,http://www.archer-elliott.com/,Honduras,Future-proofed radical implementation,1990,Facilities Services,8508 +17,68139b5C4De03B4,"Bowers, Guerra and Krause",http://www.carrillo-nicholson.com/,Uganda,De-engineered transitional strategy,1972,Primary / Secondary Education,6986 +18,5c2EffEfdba2BdF,Mckenzie-Melton,http://montoya-thompson.com/,Hong Kong,Reverse-engineered heuristic alliance,1998,Investment Management / Hedge Fund / Private Equity,4589 +19,ba179F19F7925f5,Branch-Mann,http://www.lozano.com/,Botswana,Adaptive intangible frame,1999,Architecture / Planning,7961 +20,c1Ce9B350BAc66b,Weiss and Sons,https://barrett.com/,Korea,Sharable optimal functionalities,2011,Plastics,5984 +21,8de40AC4e6EaCa4,"Velez, Payne and Coffey",http://burton.com/,Luxembourg,Mandatory coherent synergy,1986,Wholesale,5010 +22,Aad86a4F0385F2d,Harrell LLC,http://www.frey-rosario.com/,Guadeloupe,Reverse-engineered mission-critical moratorium,2018,Construction,2185 +23,22aC3FFd64fD703,"Eaton, Reynolds and Vargas",http://www.freeman.biz/,Monaco,Self-enabling multi-tasking process improvement,2014,Luxury Goods / Jewelry,8987 +24,5Ec4C272bCf085c,Robbins-Cummings,http://donaldson-wilkins.com/,Belgium,Organic non-volatile hierarchy,1991,Pharmaceuticals,5038 +25,5fDBeA8BB91a000,Jenkins Inc,http://www.kirk.biz/,South Africa,Front-line systematic help-desk,2002,Insurance,1215 +26,dFfD6a6F9AC2d9C,"Greene, Benjamin and Novak",http://www.kent.net/,Romania,Centralized leadingedge moratorium,2012,Museums / Institutions,4941 +27,4B217cC5a0674C5,"Dickson, Richmond and Clay",http://everett.com/,Czech Republic,Team-oriented tangible complexity,1980,Real Estate / Mortgage,3122 +28,88b1f1cDcf59a37,Prince-David,http://thompson.com/,Christmas Island,Virtual holistic methodology,1970,Banking / Mortgage,1046 +29,f9F7bBCAEeC360F,Ayala LLC,http://www.zhang.com/,Philippines,Open-source zero administration hierarchy,2021,Legal Services,7664 +30,7Cb3AeFcE4Ba31e,Rivas Group,https://hebert.org/,Australia,Open-architected well-modulated capacity,1998,Logistics / Procurement,4155 +31,ccBcC32adcbc530,"Sloan, Mays and Whitehead",http://lawson.com/,Chad,Face-to-face high-level conglomeration,1997,Civil Engineering,365 +32,f5afd686b3d05F5,"Durham, Allen and Barnes",http://chan-stafford.org/,Zimbabwe,Synergistic web-enabled framework,1993,Mechanical or Industrial Engineering,6135 +33,38C6cfC5074Fa5e,Fritz-Franklin,http://www.lambert.com/,Nepal,Automated 4thgeneration website,1972,Hospitality,4516 +34,5Cd7efccCcba38f,Burch-Ewing,http://cline.net/,Taiwan,User-centric 4thgeneration system engine,1981,Venture Capital / VC,7443 +35,9E6Acb51e3F9d6F,"Glass, Barrera and Turner",https://dunlap.com/,Kyrgyz Republic,Multi-channeled 3rdgeneration open system,2020,Utilities,2610 +36,4D4d7E18321eaeC,Pineda-Cox,http://aguilar.org/,Bolivia,Fundamental asynchronous capability,2010,Human Resources / HR,1312 +37,485f5d06B938F2b,"Baker, Mccann and Macdonald",http://www.anderson-barker.com/,Kenya,Cross-group user-facing focus group,2013,Legislative Office,1638 +38,19E3a5Bf6dBDc4F,Cuevas-Moss,https://dodson-castaneda.net/,Guatemala,Extended human-resource intranet,1994,Music,9995 +39,6883A965c7b68F7,Hahn PLC,http://newman.com/,Belarus,Organic logistical leverage,2012,Electrical / Electronic Manufacturing,3715 +40,AC5B7AA74Aa4A2E,"Valentine, Ferguson and Kramer",http://stuart.net/,Jersey,Centralized secondary time-frame,1997,Non - Profit / Volunteering,3585 +41,decab0D5027CA6a,Arroyo Inc,https://www.turner.com/,Grenada,Managed demand-driven website,2006,Writing / Editing,9067 +42,dF084FbBb613eea,Walls LLC,http://www.reese-vasquez.biz/,Cape Verde,Self-enabling fresh-thinking installation,1989,Investment Management / Hedge Fund / Private Equity,1678 +43,A2D89Ab9bCcAd4e,"Mitchell, Warren and Schneider",https://fox.biz/,Trinidad and Tobago,Enhanced intangible time-frame,2021,Capital Markets / Hedge Fund / Private Equity,3816 +44,77aDc905434a49f,Prince PLC,https://www.watts.com/,Sweden,Profit-focused coherent installation,2016,Individual / Family Services,7645 +45,235fdEFE2cfDa5F,Brock-Blackwell,http://www.small.com/,Benin,Secured foreground emulation,1986,Online Publishing,7034 +46,1eD64cFe986BBbE,Walton-Barnett,https://ashley-schaefer.com/,Western Sahara,Right-sized clear-thinking flexibility,2001,Luxury Goods / Jewelry,1746 +47,CbBbFcdd0eaE2cF,Bartlett-Arroyo,https://cruz.com/,Northern Mariana Islands,Realigned didactic function,1976,Civic / Social Organization,3987 +48,49aECbDaE6aBD53,"Wallace, Madden and Morris",http://www.blevins-fernandez.biz/,Germany,Persistent real-time customer loyalty,2016,Pharmaceuticals,9443 +49,7b3fe6e7E72bFa4,Berg-Sparks,https://cisneros-love.com/,Canada,Stand-alone static implementation,1974,Arts / Crafts,2073 +50,c6DedA82A8aef7E,Gonzales Ltd,http://bird.com/,Tonga,Managed human-resource policy,1988,Consumer Goods,9069 +51,7D9FBF85cdC3871,Lawson and Sons,https://www.wong.com/,French Southern Territories,Compatible analyzing intranet,2021,Arts / Crafts,3527 +52,7dd18Fb7cB07b65,"Mcguire, Mcconnell and Olsen",https://melton-briggs.com/,Korea,Profound client-server frame,1988,Printing,8445 +53,EF5B55FadccB8Fe,Charles-Phillips,https://bowman.com/,Cote d'Ivoire,Monitored client-server implementation,2012,Mental Health Care,3450 +54,f8D4B99e11fAF5D,Odom Ltd,https://www.humphrey-hess.com/,Cote d'Ivoire,Advanced static process improvement,2012,Management Consulting,1825 +55,e24D21BFd3bF1E5,Richard PLC,https://holden-coleman.net/,Mayotte,Object-based optimizing model,1971,Broadcast Media,4942 +56,B9BdfEB6D3Ca44E,Sampson Ltd,https://blevins.com/,Cayman Islands,Intuitive local adapter,2005,Farming,1418 +57,2a74D6f3D3B268e,"Cherry, Le and Callahan",https://waller-delacruz.biz/,Nigeria,Universal human-resource collaboration,2017,Entertainment / Movie Production,7202 +58,Bf3F3f62c8aBC33,Cherry PLC,https://www.avila.info/,Marshall Islands,Persistent tertiary website,1980,Plastics,8245 +59,aeBe26B80a7a23c,Melton-Nichols,https://kennedy.com/,Palau,User-friendly clear-thinking productivity,2021,Legislative Office,8741 +60,aAeb29ad43886C6,Potter-Walsh,http://thomas-french.org/,Turkey,Optional non-volatile open system,2008,Human Resources / HR,6923 +61,bD1bc6bB6d1FeD3,Freeman-Chen,https://mathis.com/,Timor-Leste,Phased next generation adapter,1973,International Trade / Development,346 +62,EB9f456e8b7022a,Soto Group,https://norris.info/,Vietnam,Enterprise-wide executive installation,1988,Business Supplies / Equipment,9097 +63,Dfef38C51D8DAe3,"Poole, Cruz and Whitney",https://reed.info/,Reunion,Balanced analyzing groupware,1978,Marketing / Advertising / Sales,2992 +64,055ffEfB2Dd95B0,Riley Ltd,http://wiley.com/,Brazil,Optional exuding superstructure,1986,Textiles,9315 +65,cBfe4dbAE1699da,"Erickson, Andrews and Bailey",https://www.hobbs-grant.com/,Eritrea,Vision-oriented secondary project,2014,Consumer Electronics,7829 +66,fdFbecbadcdCdf1,"Wilkinson, Charles and Arroyo",http://hunter-mcfarland.com/,United States Virgin Islands,Assimilated 24/7 archive,1996,Building Materials,602 +67,5DCb8A5a5ca03c0,Floyd Ltd,http://www.whitney.com/,Falkland Islands (Malvinas),Function-based fault-tolerant concept,2017,Public Relations / PR,2911 +68,ce57DCbcFD6d618,Newman-Galloway,https://www.scott.com/,Luxembourg,Enhanced foreground collaboration,1987,Information Technology / IT,3934 +69,5aaD187dc929371,Frazier-Butler,https://www.daugherty-farley.info/,Northern Mariana Islands,Persistent interactive circuit,1972,Outsourcing / Offshoring,5130 +70,902D7Ac8b6d476b,Newton Inc,https://www.richmond-manning.info/,Netherlands Antilles,Fundamental stable info-mediaries,1976,Military Industry,563 +71,32BB9Ff4d939788,Duffy-Levy,https://www.potter.com/,Guernsey,Diverse exuding installation,1982,Wireless,6146 +72,adcB0afbE58bAe3,Wagner LLC,https://decker-esparza.com/,Uruguay,Reactive attitude-oriented toolset,1987,International Affairs,6874 +73,dfcA1c84AdB61Ac,Mccall-Holmes,http://www.dean.com/,Benin,Object-based value-added database,2009,Legal Services,696 +74,208044AC2fe52F3,Massey LLC,https://frazier.biz/,Suriname,Configurable zero administration Graphical User Interface,1986,Accounting,5004 +75,f3C365f0c1A0623,Hicks LLC,http://alvarez.biz/,Pakistan,Quality-focused client-server Graphical User Interface,1970,Computer Software / Engineering,8480 +76,ec5Bdd3CBAfaB93,"Cole, Russell and Avery",http://www.blankenship.com/,Mongolia,De-engineered fault-tolerant challenge,2000,Law Enforcement,7012 +77,DDB19Be7eeB56B4,Cummings-Rojas,https://simon-pearson.com/,Svalbard & Jan Mayen Islands,User-centric modular customer loyalty,2012,Financial Services,7529 +78,dd6CA3d0bc3cAfc,"Beasley, Greene and Mahoney",http://www.petersen-lawrence.com/,Togo,Extended content-based methodology,1976,Religious Institutions,869 +79,A0B9d56e61070e3,"Beasley, Sims and Allison",http://burke.info/,Latvia,Secured zero tolerance hub,1972,Facilities Services,6182 +80,cBa7EFe5D05Adaf,Crawford-Rivera,https://black-ramirez.org/,Cuba,Persevering exuding budgetary management,1999,Online Publishing,7805 +81,Ea3f6D52Ec73563,Montes-Hensley,https://krueger.org/,Liechtenstein,Multi-tiered secondary productivity,2009,Printing,8433 +82,bC0CEd48A8000E0,Velazquez-Odom,https://stokes.com/,Djibouti,Streamlined 6thgeneration function,2002,Alternative Dispute Resolution,4044 +83,c89b9b59BC4baa1,Eaton-Morales,https://www.reeves-graham.com/,Micronesia,Customer-focused explicit frame,1990,Capital Markets / Hedge Fund / Private Equity,7013 +84,FEC51bce8421a7b,"Roberson, Pennington and Palmer",http://www.keith-fisher.com/,Cameroon,Adaptive bi-directional hierarchy,1993,Telecommunications,5571 +85,e0E8e27eAc9CAd5,"George, Russo and Guerra",https://drake.com/,Sweden,Centralized non-volatile capability,1989,Military Industry,2880 +86,B97a6CF9bf5983C,Davila Inc,https://mcconnell.info/,Cocos (Keeling) Islands,Profit-focused dedicated frame,2017,Consumer Electronics,2215 +87,a0a6f9b3DbcBEb5,Mays-Preston,http://www.browning-key.com/,Mali,User-centric heuristic focus group,2006,Military Industry,5786 +88,8cC1bDa330a5871,Pineda-Morton,https://www.carr.com/,United States Virgin Islands,Grass-roots methodical info-mediaries,1991,Printing,6168 +89,ED889CB2FE9cbd3,Huang and Sons,https://www.bolton.com/,Eritrea,Re-contextualized dynamic hierarchy,1981,Semiconductors,7484 +90,F4Dc1417BC6cb8f,Gilbert-Simon,https://www.bradford.biz/,Burundi,Grass-roots radical parallelism,1973,Newspapers / Journalism,1927 +91,7ABc3c7ecA03B34,Sampson-Griffith,http://hendricks.org/,Benin,Multi-layered composite paradigm,1972,Textiles,3881 +92,4e0719FBE38e0aB,Miles-Dominguez,http://www.turner.com/,Gibraltar,Organized empowering forecast,1996,Civic / Social Organization,897 +93,dEbDAAeDfaed00A,Rowe and Sons,https://www.simpson.org/,El Salvador,Balanced multimedia knowledgebase,1978,Facilities Services,8172 +94,61BDeCfeFD0cEF5,"Valenzuela, Holmes and Rowland",https://www.dorsey.net/,Taiwan,Persistent tertiary focus group,1999,Transportation,1483 +95,4e91eD25f486110,"Best, Wade and Shepard",https://zimmerman.com/,Zimbabwe,Innovative background definition,1991,Gambling / Casinos,4873 +96,0a0bfFbBbB8eC7c,Holmes Group,https://mcdowell.org/,Ethiopia,Right-sized zero tolerance focus group,1975,Photography,2988 +97,BA6Cd9Dae2Efd62,Good Ltd,http://duffy.com/,Anguilla,Reverse-engineered composite moratorium,1971,Consumer Services,4292 +98,E7df80C60Abd7f9,Clements-Espinoza,http://www.flowers.net/,Falkland Islands (Malvinas),Progressive modular hub,1991,Broadcast Media,236 +99,AFc285dbE2fEd24,Mendez Inc,https://www.burke.net/,Kyrgyz Republic,User-friendly exuding migration,1993,Education Management,339 +100,e9eB5A60Cef8354,Watkins-Kaiser,http://www.herring.com/,Togo,Synergistic background access,2009,Financial Services,2785 diff --git a/applications/ColossalQA/data/tests/test.html b/applications/ColossalQA/data/tests/test.html index 5ad21421d827..6152ffe296fc 100644 --- a/applications/ColossalQA/data/tests/test.html +++ b/applications/ColossalQA/data/tests/test.html @@ -1,7 +1,7 @@ - + @@ -16,18 +16,18 @@ logging — Logging facility for Python — Python 3.11.5 documentation - + - + - + - + @@ -36,11 +36,11 @@ - - - - + + + +
\ No newline at end of file +
diff --git a/applications/ColossalQA/data/tests/test.md b/applications/ColossalQA/data/tests/test.md index 20d3c612fef5..b6e94792c83d 100644 --- a/applications/ColossalQA/data/tests/test.md +++ b/applications/ColossalQA/data/tests/test.md @@ -34,9 +34,9 @@ python api_server.py --host localhost --port $PORT_NUMBER --model $PATH_TO_MODEL ### Collect your data For ChatGPT based Agent we support document retrieval and simple sql search. -If you want to run the demo locally, we provided document retrieval based conversation system built upon langchain. It accept a wide range of documents. +If you want to run the demo locally, we provided document retrieval based conversation system built upon langchain. It accept a wide range of documents. -Read comments under ./colossalqa/data_loader for more detail +Read comments under ./colossalqa/data_loader for more detail ### Serving Currently use vllm will replace with colossal inference when ready. Please refer class VllmLLM. diff --git a/applications/ColossalQA/data/tests/test.txt b/applications/ColossalQA/data/tests/test.txt index ff5bf2dc7742..e608c772e06e 100644 --- a/applications/ColossalQA/data/tests/test.txt +++ b/applications/ColossalQA/data/tests/test.txt @@ -1,38 +1,38 @@ -Your Name -Lorem ipsum dolor sit amet, consectetuer adipiscing elit - 123 Your Street -Your City, ST 12345 -(123) 456-7890 -no_reply@example.com - EXPERIENCE -Company, Location — Job Title -MONTH 20XX - PRESENT -Lorem ipsum dolor sit amet, consectetuer adipiscing elit, sed diam nonummy nibh. -Company, Location — Job Title -MONTH 20XX - MONTH 20XX -Lorem ipsum dolor sit amet, consectetuer adipiscing elit, sed diam nonummy nibh. -Company, Location — Job Title -MONTH 20XX - MONTH 20XX -Lorem ipsum dolor sit amet, consectetuer adipiscing elit, sed diam nonummy nibh. -EDUCATION -School Name, Location — Degree -MONTH 20XX - MONTH 20XX -Lorem ipsum dolor sit amet, consectetuer adipiscing elit, sed diam nonummy nibh euismod tincidunt ut laoreet dolore. -School Name, Location — Degree -MONTH 20XX - MONTH 20XX -Lorem ipsum dolor sit amet, consectetuer adipiscing elit, sed diam. -PROJECTS -Project Name — Detail -Lorem ipsum dolor sit amet, consectetuer adipiscing elit. - SKILLS -* Lorem ipsum dolor sit amet. -* Consectetuer adipiscing elit. -* Sed diam nonummy nibh euismod tincidunt. -* L​​​‌​aoreet dolore magna aliquam erat volutpat. -AWARDS -Lorem ipsum dolor sit amet Consectetuer adipiscing elit, Sed diam nonummy -Nibh euismod tincidunt ut laoreet dolore magna aliquam erat volutpat. -Lorem ipsum dolor sit amet Consectetuer adipiscing elit, Sed diam nonummy -Nibh euismod tincidunt ut laoreet dolore magna aliquam erat volutpat. -LANGUAGES -Lorem ipsum, Dolor sit amet, Consectetuer \ No newline at end of file +Your Name +Lorem ipsum dolor sit amet, consectetuer adipiscing elit + 123 Your Street +Your City, ST 12345 +(123) 456-7890 +no_reply@example.com + EXPERIENCE +Company, Location — Job Title +MONTH 20XX - PRESENT +Lorem ipsum dolor sit amet, consectetuer adipiscing elit, sed diam nonummy nibh. +Company, Location — Job Title +MONTH 20XX - MONTH 20XX +Lorem ipsum dolor sit amet, consectetuer adipiscing elit, sed diam nonummy nibh. +Company, Location — Job Title +MONTH 20XX - MONTH 20XX +Lorem ipsum dolor sit amet, consectetuer adipiscing elit, sed diam nonummy nibh. +EDUCATION +School Name, Location — Degree +MONTH 20XX - MONTH 20XX +Lorem ipsum dolor sit amet, consectetuer adipiscing elit, sed diam nonummy nibh euismod tincidunt ut laoreet dolore. +School Name, Location — Degree +MONTH 20XX - MONTH 20XX +Lorem ipsum dolor sit amet, consectetuer adipiscing elit, sed diam. +PROJECTS +Project Name — Detail +Lorem ipsum dolor sit amet, consectetuer adipiscing elit. + SKILLS +* Lorem ipsum dolor sit amet. +* Consectetuer adipiscing elit. +* Sed diam nonummy nibh euismod tincidunt. +* L​​​‌​aoreet dolore magna aliquam erat volutpat. +AWARDS +Lorem ipsum dolor sit amet Consectetuer adipiscing elit, Sed diam nonummy +Nibh euismod tincidunt ut laoreet dolore magna aliquam erat volutpat. +Lorem ipsum dolor sit amet Consectetuer adipiscing elit, Sed diam nonummy +Nibh euismod tincidunt ut laoreet dolore magna aliquam erat volutpat. +LANGUAGES +Lorem ipsum, Dolor sit amet, Consectetuer diff --git a/applications/ColossalQA/examples/retrieval_conversation_universal.py b/applications/ColossalQA/examples/retrieval_conversation_universal.py index 5d13a63c3fad..8999fbabd2ed 100644 --- a/applications/ColossalQA/examples/retrieval_conversation_universal.py +++ b/applications/ColossalQA/examples/retrieval_conversation_universal.py @@ -1,22 +1,27 @@ import argparse + from colossalqa.retrieval_conversation_universal import UniversalRetrievalConversation -if __name__ == '__main__': +if __name__ == "__main__": # Parse arguments parser = argparse.ArgumentParser() - parser.add_argument('--en_model_path', type=str, default=None) - parser.add_argument('--zh_model_path', type=str, default=None) - parser.add_argument('--zh_model_name', type=str, default=None) - parser.add_argument('--en_model_name', type=str, default=None) - parser.add_argument('--sql_file_path', type=str, default=None, help='path to the a empty folder for storing sql files for indexing') + parser.add_argument("--en_model_path", type=str, default=None) + parser.add_argument("--zh_model_path", type=str, default=None) + parser.add_argument("--zh_model_name", type=str, default=None) + parser.add_argument("--en_model_name", type=str, default=None) + parser.add_argument( + "--sql_file_path", type=str, default=None, help="path to the a empty folder for storing sql files for indexing" + ) args = parser.parse_args() - + # Will ask for documents path in running time - session = UniversalRetrievalConversation(files_en=None, - files_zh=None, - zh_model_path=args.zh_model_path, en_model_path=args.en_model_path, - zh_model_name=args.zh_model_name, en_model_name=args.en_model_name, - sql_file_path=args.sql_file_path - ) + session = UniversalRetrievalConversation( + files_en=None, + files_zh=None, + zh_model_path=args.zh_model_path, + en_model_path=args.en_model_path, + zh_model_name=args.zh_model_name, + en_model_name=args.en_model_name, + sql_file_path=args.sql_file_path, + ) session.start_test_session() - \ No newline at end of file diff --git a/applications/ColossalQA/examples/webui_demo/RAG_ChatBot.py b/applications/ColossalQA/examples/webui_demo/RAG_ChatBot.py index 526328dda11b..47897b538ee2 100644 --- a/applications/ColossalQA/examples/webui_demo/RAG_ChatBot.py +++ b/applications/ColossalQA/examples/webui_demo/RAG_ChatBot.py @@ -5,13 +5,7 @@ from colossalqa.data_loader.document_loader import DocumentLoader from colossalqa.memory import ConversationBufferWithSummary from colossalqa.mylogging import get_logger -from colossalqa.prompt.prompt import ( - PROMPT_DISAMBIGUATE_ZH, - PROMPT_RETRIEVAL_QA_ZH, - SUMMARY_PROMPT_ZH, - ZH_RETRIEVAL_QA_REJECTION_ANSWER, - ZH_RETRIEVAL_QA_TRIGGER_KEYWORDS, -) +from colossalqa.prompt.prompt import ZH_RETRIEVAL_QA_REJECTION_ANSWER, ZH_RETRIEVAL_QA_TRIGGER_KEYWORDS from colossalqa.retriever import CustomRetriever from langchain import LLMChain from langchain.embeddings import HuggingFaceEmbeddings @@ -116,13 +110,13 @@ def split_docs_and_add_to_mem(self, **kwargs): def split_docs(self, documents): doc_splits = self.text_splitter.split_documents(documents) return doc_splits - + def clear_docs(self, **kwargs): self.documents = [] self.docs_names = [] self.info_retriever.clear_documents() self.memory.initiate_document_retrieval_chain(self.llm, kwargs["gen_qa_prompt"], self.info_retriever) - + def reset_config(self, rag_config): self.rag_config = rag_config self.set_embed_model(**self.rag_config["embed"]) diff --git a/applications/ColossalQA/examples/webui_demo/README.md b/applications/ColossalQA/examples/webui_demo/README.md index 3e1a8adf53f3..61071a1c8065 100644 --- a/applications/ColossalQA/examples/webui_demo/README.md +++ b/applications/ColossalQA/examples/webui_demo/README.md @@ -115,4 +115,4 @@ python webui.py --http_host "your-backend-api-host" --http_port "your-backend-ap After launching the script, you can upload files and engage with the chatbot through your web browser. -![ColossalQA Demo](https://raw.githubusercontent.com/hpcaitech/public_assets/main/applications/colossalqa/new_ui.png) \ No newline at end of file +![ColossalQA Demo](https://raw.githubusercontent.com/hpcaitech/public_assets/main/applications/colossalqa/new_ui.png) diff --git a/applications/ColossalQA/examples/webui_demo/config.py b/applications/ColossalQA/examples/webui_demo/config.py index ef90fab62589..5bdc08a09e50 100644 --- a/applications/ColossalQA/examples/webui_demo/config.py +++ b/applications/ColossalQA/examples/webui_demo/config.py @@ -1,58 +1,30 @@ -from colossalqa.prompt.prompt import ( - PROMPT_DISAMBIGUATE_ZH, - PROMPT_RETRIEVAL_QA_ZH, - SUMMARY_PROMPT_ZH, - ZH_RETRIEVAL_QA_REJECTION_ANSWER, - ZH_RETRIEVAL_QA_TRIGGER_KEYWORDS, -) +from colossalqa.prompt.prompt import PROMPT_DISAMBIGUATE_ZH, PROMPT_RETRIEVAL_QA_ZH, SUMMARY_PROMPT_ZH from colossalqa.text_splitter import ChineseTextSplitter ALL_CONFIG = { "embed": { "embed_name": "m3e", # embedding model name "embed_model_name_or_path": "moka-ai/m3e-base", # path to embedding model, could be a local path or a huggingface path - "embed_model_device": { - "device": "cpu" - } + "embed_model_device": {"device": "cpu"}, }, "model": { "mode": "api", # "local" for loading models, "api" for using model api "model_name": "chatgpt_api", # local model name, "chatgpt_api" or "pangu_api" - "model_path": "", # path to the model, could be a local path or a huggingface path. don't need if using an api - "device": { - "device": "cuda" - } - }, - "splitter": { - "name": ChineseTextSplitter - }, - "retrieval": { - "retri_top_k": 3, - "retri_kb_file_path": "./", # path to store database files - "verbose": True + "model_path": "", # path to the model, could be a local path or a huggingface path. don't need if using an api + "device": {"device": "cuda"}, }, + "splitter": {"name": ChineseTextSplitter}, + "retrieval": {"retri_top_k": 3, "retri_kb_file_path": "./", "verbose": True}, # path to store database files "chain": { "mem_summary_prompt": SUMMARY_PROMPT_ZH, # summary prompt template "mem_human_prefix": "用户", "mem_ai_prefix": "Assistant", "mem_max_tokens": 2000, - "mem_llm_kwargs": { - "max_new_tokens": 50, - "temperature": 1, - "do_sample": True - }, + "mem_llm_kwargs": {"max_new_tokens": 50, "temperature": 1, "do_sample": True}, "disambig_prompt": PROMPT_DISAMBIGUATE_ZH, # disambiguate prompt template - "disambig_llm_kwargs": { - "max_new_tokens": 30, - "temperature": 1, - "do_sample": True - }, - "gen_llm_kwargs": { - "max_new_tokens": 100, - "temperature": 1, - "do_sample": True - }, + "disambig_llm_kwargs": {"max_new_tokens": 30, "temperature": 1, "do_sample": True}, + "gen_llm_kwargs": {"max_new_tokens": 100, "temperature": 1, "do_sample": True}, "gen_qa_prompt": PROMPT_RETRIEVAL_QA_ZH, # generation prompt template - "verbose": True - } -} \ No newline at end of file + "verbose": True, + }, +} diff --git a/applications/ColossalQA/examples/webui_demo/server.py b/applications/ColossalQA/examples/webui_demo/server.py index 3b0f82845c87..1f699421d34c 100644 --- a/applications/ColossalQA/examples/webui_demo/server.py +++ b/applications/ColossalQA/examples/webui_demo/server.py @@ -1,27 +1,18 @@ import argparse -import os from typing import List, Union - +import config +import uvicorn from colossalqa.local.llm import ColossalAPI, ColossalLLM -from colossalqa.data_loader.document_loader import DocumentLoader from colossalqa.mylogging import get_logger -from colossalqa.retrieval_conversation_zh import ChineseRetrievalConversation -from colossalqa.retriever import CustomRetriever -from enum import Enum from fastapi import FastAPI, Request -from langchain.embeddings import HuggingFaceEmbeddings -from langchain.text_splitter import RecursiveCharacterTextSplitter -from pydantic import BaseModel, Field -import uvicorn - -import config +from pydantic import BaseModel from RAG_ChatBot import RAG_ChatBot from utils import DocAction - logger = get_logger() + def parseArgs(): parser = argparse.ArgumentParser() parser.add_argument("--http_host", default="0.0.0.0") @@ -36,6 +27,7 @@ class DocUpdateReq(BaseModel): doc_files: Union[List[str], str, None] = None action: DocAction = DocAction.ADD + class GenerationTaskReq(BaseModel): user_input: str @@ -45,7 +37,7 @@ def update_docs(data: DocUpdateReq, request: Request): if data.action == "add": if isinstance(data.doc_files, str): data.doc_files = [data.doc_files] - chatbot.load_doc_from_files(files = data.doc_files) + chatbot.load_doc_from_files(files=data.doc_files) all_docs = "" for doc in chatbot.docs_names: all_docs += f"\t{doc}\n\n" @@ -79,17 +71,18 @@ def generate(data: GenerationTaskReq, request: Request): elif all_config["model"]["mode"] == "api": if model_name == "pangu_api": from colossalqa.local.pangu_llm import Pangu - + gen_config = { "user": "User", "max_tokens": all_config["chain"]["disambig_llm_kwargs"]["max_new_tokens"], "temperature": all_config["chain"]["disambig_llm_kwargs"]["temperature"], - "n": 1 # the number of responses generated + "n": 1, # the number of responses generated } llm = Pangu(gen_config=gen_config) llm.set_auth_config() # verify user's auth info here elif model_name == "chatgpt_api": from langchain.llms import OpenAI + llm = OpenAI() else: raise ValueError("Unsupported mode.") diff --git a/applications/ColossalQA/examples/webui_demo/webui.py b/applications/ColossalQA/examples/webui_demo/webui.py index cd3b5fd5da4b..1e34330615b5 100644 --- a/applications/ColossalQA/examples/webui_demo/webui.py +++ b/applications/ColossalQA/examples/webui_demo/webui.py @@ -1,24 +1,26 @@ import argparse import json import os -import requests import gradio as gr - +import requests from utils import DocAction + def parseArgs(): parser = argparse.ArgumentParser() parser.add_argument("--http_host", default="0.0.0.0") parser.add_argument("--http_port", type=int, default=13666) return parser.parse_args() + def get_response(data, url): headers = {"Content-type": "application/json"} response = requests.post(url, json=data, headers=headers) response = json.loads(response.content) return response + def add_text(history, text): history = history + [(text, None)] return history, gr.update(value=None, interactive=True) @@ -28,35 +30,28 @@ def add_file(history, files): files_string = "\n".join([os.path.basename(file.name) for file in files]) doc_files = [file.name for file in files] - data = { - "doc_files": doc_files, - "action": DocAction.ADD - } + data = {"doc_files": doc_files, "action": DocAction.ADD} response = get_response(data, update_url)["response"] history = history + [(files_string, response)] return history -def bot(history): - data = { - "user_input": history[-1][0].strip() - } + +def bot(history): + data = {"user_input": history[-1][0].strip()} response = get_response(data, gen_url) if response["error"] != "": raise gr.Error(response["error"]) - + history[-1][1] = response["response"] yield history def restart(chatbot, txt): # Reset the conversation state and clear the chat history - data = { - "doc_files": "", - "action": DocAction.CLEAR - } - response = get_response(data, update_url) - + data = {"doc_files": "", "action": DocAction.CLEAR} + get_response(data, update_url) + return gr.update(value=None), gr.update(value=None, interactive=True) @@ -97,7 +92,7 @@ def restart(chatbot, txt): txt_msg = txt.submit(add_text, [chatbot, txt], [chatbot, txt], queue=False).then(bot, chatbot, chatbot) # Clear the original textbox - txt_msg.then(lambda: gr.update(value=None, interactive=True), None, [txt], queue=False) + txt_msg.then(lambda: gr.update(value=None, interactive=True), None, [txt], queue=False) # Click Upload Button: 1. upload files 2. send config to backend, initalize model 3. get response "conversation_ready" = True/False file_msg = btn.upload(add_file, [chatbot, btn], [chatbot], queue=False) diff --git a/applications/ColossalQA/pytest.ini b/applications/ColossalQA/pytest.ini index 9e84349f2285..8bd9250769e3 100644 --- a/applications/ColossalQA/pytest.ini +++ b/applications/ColossalQA/pytest.ini @@ -1,4 +1,4 @@ [pytest] markers = dist: tests which are run in a multi-GPU or multi-machine environment (at least 4 GPUs) - largedist: tests which are run in a multi-GPU or multi-machine environment (at least 8 GPUs) \ No newline at end of file + largedist: tests which are run in a multi-GPU or multi-machine environment (at least 8 GPUs) diff --git a/applications/ColossalQA/tests/test_document_loader.py b/applications/ColossalQA/tests/test_document_loader.py index 163b0d2cca22..96c74f128782 100644 --- a/applications/ColossalQA/tests/test_document_loader.py +++ b/applications/ColossalQA/tests/test_document_loader.py @@ -1,21 +1,21 @@ import os + from colossalqa.data_loader.document_loader import DocumentLoader def test_add_document(): - PATH = os.environ.get('TEST_DOCUMENT_LOADER_DATA_PATH') - files = [[PATH, 'all data']] + PATH = os.environ.get("TEST_DOCUMENT_LOADER_DATA_PATH") + files = [[PATH, "all data"]] document_loader = DocumentLoader(files) documents = document_loader.all_data all_files = [] for doc in documents: - assert isinstance(doc.page_content, str)==True - if doc.metadata['source'] not in all_files: - all_files.append(doc.metadata['source']) + assert isinstance(doc.page_content, str) == True + if doc.metadata["source"] not in all_files: + all_files.append(doc.metadata["source"]) print(all_files) assert len(all_files) == 6 -if __name__=='__main__': +if __name__ == "__main__": test_add_document() - diff --git a/applications/ColossalQA/tests/test_retrieval_qa.py b/applications/ColossalQA/tests/test_retrieval_qa.py index 76867b562e8f..8ba84cc7309f 100644 --- a/applications/ColossalQA/tests/test_retrieval_qa.py +++ b/applications/ColossalQA/tests/test_retrieval_qa.py @@ -4,56 +4,44 @@ def test_en_retrievalQA(): - data_path_en = os.environ.get('TEST_DATA_PATH_EN') - data_path_zh = os.environ.get('TEST_DATA_PATH_ZH') - en_model_path = os.environ.get('EN_MODEL_PATH') - zh_model_path = os.environ.get('ZH_MODEL_PATH') - zh_model_name = os.environ.get('ZH_MODEL_NAME') - en_model_name = os.environ.get('EN_MODEL_NAME') - sql_file_path = os.environ.get('SQL_FILE_PATH') - qa_session = UniversalRetrievalConversation(files_en=[{ - 'data_path': data_path_en, - 'name': 'company information', - 'separator': '\n' - }], - files_zh=[{ - 'data_path': data_path_zh, - 'name': 'company information', - 'separator': '\n' - }], - zh_model_path=zh_model_path, - en_model_path=en_model_path, - zh_model_name=zh_model_name, - en_model_name=en_model_name, - sql_file_path=sql_file_path) - ans = qa_session.run("which company runs business in hotel industry?", which_language='en') + data_path_en = os.environ.get("TEST_DATA_PATH_EN") + data_path_zh = os.environ.get("TEST_DATA_PATH_ZH") + en_model_path = os.environ.get("EN_MODEL_PATH") + zh_model_path = os.environ.get("ZH_MODEL_PATH") + zh_model_name = os.environ.get("ZH_MODEL_NAME") + en_model_name = os.environ.get("EN_MODEL_NAME") + sql_file_path = os.environ.get("SQL_FILE_PATH") + qa_session = UniversalRetrievalConversation( + files_en=[{"data_path": data_path_en, "name": "company information", "separator": "\n"}], + files_zh=[{"data_path": data_path_zh, "name": "company information", "separator": "\n"}], + zh_model_path=zh_model_path, + en_model_path=en_model_path, + zh_model_name=zh_model_name, + en_model_name=en_model_name, + sql_file_path=sql_file_path, + ) + ans = qa_session.run("which company runs business in hotel industry?", which_language="en") print(ans) def test_zh_retrievalQA(): - data_path_en = os.environ.get('TEST_DATA_PATH_EN') - data_path_zh = os.environ.get('TEST_DATA_PATH_ZH') - en_model_path = os.environ.get('EN_MODEL_PATH') - zh_model_path = os.environ.get('ZH_MODEL_PATH') - zh_model_name = os.environ.get('ZH_MODEL_NAME') - en_model_name = os.environ.get('EN_MODEL_NAME') - sql_file_path = os.environ.get('SQL_FILE_PATH') - qa_session = UniversalRetrievalConversation(files_en=[{ - 'data_path': data_path_en, - 'name': 'company information', - 'separator': '\n' - }], - files_zh=[{ - 'data_path': data_path_zh, - 'name': 'company information', - 'separator': '\n' - }], - zh_model_path=zh_model_path, - en_model_path=en_model_path, - zh_model_name=zh_model_name, - en_model_name=en_model_name, - sql_file_path=sql_file_path) - ans = qa_session.run("哪家公司在经营酒店业务？", which_language='zh') + data_path_en = os.environ.get("TEST_DATA_PATH_EN") + data_path_zh = os.environ.get("TEST_DATA_PATH_ZH") + en_model_path = os.environ.get("EN_MODEL_PATH") + zh_model_path = os.environ.get("ZH_MODEL_PATH") + zh_model_name = os.environ.get("ZH_MODEL_NAME") + en_model_name = os.environ.get("EN_MODEL_NAME") + sql_file_path = os.environ.get("SQL_FILE_PATH") + qa_session = UniversalRetrievalConversation( + files_en=[{"data_path": data_path_en, "name": "company information", "separator": "\n"}], + files_zh=[{"data_path": data_path_zh, "name": "company information", "separator": "\n"}], + zh_model_path=zh_model_path, + en_model_path=en_model_path, + zh_model_name=zh_model_name, + en_model_name=en_model_name, + sql_file_path=sql_file_path, + ) + ans = qa_session.run("哪家公司在经营酒店业务？", which_language="zh") print(ans) diff --git a/applications/ColossalQA/version.txt b/applications/ColossalQA/version.txt index 8a9ecc2ea99d..8acdd82b765e 100644 --- a/applications/ColossalQA/version.txt +++ b/applications/ColossalQA/version.txt @@ -1 +1 @@ -0.0.1 \ No newline at end of file +0.0.1 diff --git a/colossalai/__init__.py b/colossalai/__init__.py index 6b7f5d055207..beef0561ccfc 100644 --- a/colossalai/__init__.py +++ b/colossalai/__init__.py @@ -1,5 +1,5 @@ -from .initialize import launch, launch_from_openmpi, launch_from_slurm, launch_from_torch from . import accelerator +from .initialize import launch, launch_from_openmpi, launch_from_slurm, launch_from_torch try: # .version will be created by setup.py diff --git a/colossalai/booster/plugin/torch_fsdp_plugin.py b/colossalai/booster/plugin/torch_fsdp_plugin.py index 5445b4a6349d..0aa0caa9aafe 100644 --- a/colossalai/booster/plugin/torch_fsdp_plugin.py +++ b/colossalai/booster/plugin/torch_fsdp_plugin.py @@ -27,7 +27,7 @@ from torch.optim.lr_scheduler import _LRScheduler as LRScheduler from torch.utils.data import DataLoader -from colossalai.checkpoint_io import CheckpointIO, GeneralCheckpointIO, utils, CheckpointIndexFile +from colossalai.checkpoint_io import CheckpointIndexFile, CheckpointIO, GeneralCheckpointIO, utils from colossalai.cluster import DistCoordinator from colossalai.interface import ModelWrapper, OptimizerWrapper @@ -93,9 +93,7 @@ def save_sharded_model( Path(checkpoint_path).mkdir(parents=True, exist_ok=True) with FSDP.state_dict_type( - model.unwrap(), - StateDictType.FULL_STATE_DICT, - FullStateDictConfig(offload_to_cpu=True, rank0_only=True) + model.unwrap(), StateDictType.FULL_STATE_DICT, FullStateDictConfig(offload_to_cpu=True, rank0_only=True) ): state_dict = model.unwrap().state_dict() @@ -172,7 +170,7 @@ def save_sharded_optimizer( with FSDP.state_dict_type( optimizer.unwrap_model().unwrap(), StateDictType.FULL_STATE_DICT, - FullStateDictConfig(offload_to_cpu=True, rank0_only=True) + FullStateDictConfig(offload_to_cpu=True, rank0_only=True), ): fsdp_optim_state = FSDP.full_optim_state_dict( optimizer.unwrap_model().unwrap(), optim=optimizer, rank0_only=True @@ -241,7 +239,6 @@ def load_sharded_optimizer(self, optimizer: Optimizer, index_file_path: str, siz ) optimizer.load_state_dict(fsdp_state) - def save_lr_scheduler(self, lr_scheduler: LRScheduler, checkpoint: str): """ Save model to checkpoint but only on master process. diff --git a/colossalai/checkpoint_io/utils.py b/colossalai/checkpoint_io/utils.py index e1800f29b0af..2a1d4de9b036 100644 --- a/colossalai/checkpoint_io/utils.py +++ b/colossalai/checkpoint_io/utils.py @@ -294,6 +294,7 @@ def shard_optimizer_checkpoint(state_dict: dict, max_shard_size: int = 1024) -> # Helper functions for saving state dict # ====================================== + def save_state_dict(state_dict: dict, checkpoint_file_path: str, use_safetensors: bool) -> None: """ Save state dict to checkpoint. @@ -305,7 +306,7 @@ def save_state_dict(state_dict: dict, checkpoint_file_path: str, use_safetensors """ # Move all tensors in the state_dict to CPU before saving to avoid serialization issues state_dict_cpu = tree_map(lambda x: x.cpu() if torch.is_tensor(x) else x, state_dict) - + if use_safetensors: assert is_safetensors_available(), "safetensors is not available." assert checkpoint_file_path.endswith( diff --git a/colossalai/cluster/process_group_mesh.py b/colossalai/cluster/process_group_mesh.py index ae3956c693ab..1f32541a7b21 100644 --- a/colossalai/cluster/process_group_mesh.py +++ b/colossalai/cluster/process_group_mesh.py @@ -225,4 +225,3 @@ def get_group_along_axis( # no need to cache it explicitly, since it will be cached in `create_group_along_axis` return self.create_group_along_axis(axis, indices_at_axis, backend=backend) return self._ranks_to_group[ranks_in_group] - \ No newline at end of file diff --git a/colossalai/inference/engine/modeling/llama.py b/colossalai/inference/engine/modeling/llama.py index b7bc94d0eae0..a7efb4026be0 100644 --- a/colossalai/inference/engine/modeling/llama.py +++ b/colossalai/inference/engine/modeling/llama.py @@ -29,13 +29,17 @@ try: from colossalai.kernel.triton.flash_decoding import token_flash_decoding + HAS_TRITON_FLASH_DECODING_KERNEL = True except: - print("no triton flash decoding support, please install lightllm from https://github.com/ModelTC/lightllm/blob/ece7b43f8a6dfa74027adc77c2c176cff28c76c8") + print( + "no triton flash decoding support, please install lightllm from https://github.com/ModelTC/lightllm/blob/ece7b43f8a6dfa74027adc77c2c176cff28c76c8" + ) HAS_TRITON_FLASH_DECODING_KERNEL = False - + try: from flash_attn import flash_attn_with_kvcache + HAS_FLASH_KERNEL = True except: HAS_FLASH_KERNEL = False @@ -48,6 +52,7 @@ def rotate_half(x): x2 = x[..., x.shape[-1] // 2 :] return torch.cat((-x2, x1), dim=-1) + def apply_rotary_pos_emb(q, k, cos, sin, position_ids): # The first two dimensions of cos and sin are always 1, so we can `squeeze` them. cos = cos.squeeze(1).squeeze(0) # [seq_len, dim] @@ -96,17 +101,22 @@ def llama_triton_context_attention( infer_state.max_len_in_batch, ) -def llama_triton_token_attention(query_states, attn_output, infer_state, num_key_value_groups=1, q_head_num = -1, head_dim = -1): + +def llama_triton_token_attention( + query_states, attn_output, infer_state, num_key_value_groups=1, q_head_num=-1, head_dim=-1 +): if HAS_TRITON_FLASH_DECODING_KERNEL and q_head_num != -1 and head_dim != -1: - token_flash_decoding(q = query_states, - o_tensor = attn_output, - infer_state = infer_state, - q_head_num = q_head_num, - head_dim = head_dim, - cache_k = infer_state.cache_manager.key_buffer[infer_state.decode_layer_id], - cache_v = infer_state.cache_manager.value_buffer[infer_state.decode_layer_id]) - return - + token_flash_decoding( + q=query_states, + o_tensor=attn_output, + infer_state=infer_state, + q_head_num=q_head_num, + head_dim=head_dim, + cache_k=infer_state.cache_manager.key_buffer[infer_state.decode_layer_id], + cache_v=infer_state.cache_manager.value_buffer[infer_state.decode_layer_id], + ) + return + if num_key_value_groups == 1: token_attention_fwd( query_states, @@ -459,14 +469,15 @@ def llama_flash_attn_kvcache_forward( ) if HAS_LIGHTLLM_KERNEL: - attn_output = torch.empty_like(query_states) - llama_triton_token_attention(query_states = query_states, - attn_output = attn_output, - infer_state = infer_state, - num_key_value_groups = self.num_key_value_groups, - q_head_num = q_len * self.num_heads, - head_dim = self.head_dim) + llama_triton_token_attention( + query_states=query_states, + attn_output=attn_output, + infer_state=infer_state, + num_key_value_groups=self.num_key_value_groups, + q_head_num=q_len * self.num_heads, + head_dim=self.head_dim, + ) else: self.num_heads // self.num_key_value_heads cache_k = infer_state.cache_manager.key_buffer[infer_state.decode_layer_id] diff --git a/colossalai/inference/quant/gptq/cai_gptq/cai_quant_linear.py b/colossalai/inference/quant/gptq/cai_gptq/cai_quant_linear.py index ca12c34ed958..36339ac88486 100644 --- a/colossalai/inference/quant/gptq/cai_gptq/cai_quant_linear.py +++ b/colossalai/inference/quant/gptq/cai_gptq/cai_quant_linear.py @@ -18,15 +18,15 @@ HAS_GPTQ_CUDA = False try: from colossalai.kernel.op_builder.gptq import GPTQBuilder + gptq_cuda = GPTQBuilder().load() HAS_GPTQ_CUDA = True except ImportError: - warnings.warn('CUDA gptq is not installed') + warnings.warn("CUDA gptq is not installed") HAS_GPTQ_CUDA = False class CaiQuantLinear(nn.Module): - def __init__(self, bits, groupsize, infeatures, outfeatures, bias, tp_size=1, tp_rank=0, row_split=False): super().__init__() if bits not in [2, 4, 8]: @@ -37,23 +37,28 @@ def __init__(self, bits, groupsize, infeatures, outfeatures, bias, tp_size=1, tp self.maxq = 2**self.bits - 1 self.groupsize = groupsize if groupsize != -1 else infeatures - self.register_buffer('qweight', torch.zeros((infeatures // 32 * self.bits, outfeatures), dtype=torch.int32)) + self.register_buffer("qweight", torch.zeros((infeatures // 32 * self.bits, outfeatures), dtype=torch.int32)) + self.register_buffer( + "qzeros", + torch.zeros((math.ceil(infeatures / self.groupsize), outfeatures // 32 * self.bits), dtype=torch.int32), + ) self.register_buffer( - 'qzeros', - torch.zeros((math.ceil(infeatures / self.groupsize), outfeatures // 32 * self.bits), dtype=torch.int32)) - self.register_buffer('scales', - torch.zeros((math.ceil(infeatures / self.groupsize), outfeatures), dtype=torch.float16)) + "scales", torch.zeros((math.ceil(infeatures / self.groupsize), outfeatures), dtype=torch.float16) + ) if row_split: self.register_buffer( - 'g_idx', - torch.tensor([(i + (tp_rank * self.infeatures)) // self.groupsize for i in range(infeatures)], - dtype=torch.int32)) + "g_idx", + torch.tensor( + [(i + (tp_rank * self.infeatures)) // self.groupsize for i in range(infeatures)], dtype=torch.int32 + ), + ) else: - self.register_buffer('g_idx', - torch.tensor([i // self.groupsize for i in range(infeatures)], dtype=torch.int32)) + self.register_buffer( + "g_idx", torch.tensor([i // self.groupsize for i in range(infeatures)], dtype=torch.int32) + ) if bias: - self.register_buffer('bias', torch.zeros((outfeatures), dtype=torch.float16)) + self.register_buffer("bias", torch.zeros((outfeatures), dtype=torch.float16)) else: self.bias = None @@ -66,9 +71,11 @@ def __init__(self, bits, groupsize, infeatures, outfeatures, bias, tp_size=1, tp self.row_split = row_split def pack(self, linear, scales, zeros, g_idx=None): - - g_idx = g_idx.clone() if g_idx is not None else torch.tensor( - [i // self.groupsize for i in range(self.infeatures)], dtype=torch.int32) + g_idx = ( + g_idx.clone() + if g_idx is not None + else torch.tensor([i // self.groupsize for i in range(self.infeatures)], dtype=torch.int32) + ) scales = scales.t().contiguous() zeros = zeros.t().contiguous() @@ -79,7 +86,6 @@ def pack(self, linear, scales, zeros, g_idx=None): if linear.bias is not None: self.bias = linear.bias.clone().half() - wn = 8 pbits = 32 ptype = torch.int32 unsign_type = np.uint32 @@ -88,9 +94,10 @@ def pack(self, linear, scales, zeros, g_idx=None): intweight = [] for idx in range(self.infeatures): intweight.append( - torch.round( - (linear.weight.data[:, idx] + scale_zeros[g_idx[idx]]) / half_scales[g_idx[idx]]).to(ptype)[:, - None]) + torch.round((linear.weight.data[:, idx] + scale_zeros[g_idx[idx]]) / half_scales[g_idx[idx]]).to(ptype)[ + :, None + ] + ) intweight = torch.cat(intweight, dim=1) intweight = intweight.t().contiguous() intweight = intweight.numpy().astype(unsign_type) @@ -109,7 +116,7 @@ def pack(self, linear, scales, zeros, g_idx=None): raise NotImplementedError("Only 2,4,8 bits are supported.") qweight = qweight.astype(sign_type) qweight1 = torch.from_numpy(qweight) - qweight1 = qweight1.contiguous() #.to("cuda") + qweight1 = qweight1.contiguous() # .to("cuda") self.qweight.data.copy_(qweight1) qzeros = np.zeros((zeros.shape[0], zeros.shape[1] // pbits * self.bits), dtype=unsign_type) @@ -140,17 +147,20 @@ def init_q4(self): self.q4_width = self.qweight.shape[1] if self.g_idx is not None: if self.row_split and torch.equal( - self.g_idx, - torch.tensor( - [(i + (self.tp_rank * self.infeatures)) // self.groupsize for i in range(self.infeatures)], - dtype=torch.int32, - device=self.g_idx.device)): + self.g_idx, + torch.tensor( + [(i + (self.tp_rank * self.infeatures)) // self.groupsize for i in range(self.infeatures)], + dtype=torch.int32, + device=self.g_idx.device, + ), + ): self.g_idx = None elif torch.equal( - self.g_idx, - torch.tensor([i // self.groupsize for i in range(self.infeatures)], - dtype=torch.int32, - device=self.g_idx.device)): + self.g_idx, + torch.tensor( + [i // self.groupsize for i in range(self.infeatures)], dtype=torch.int32, device=self.g_idx.device + ), + ): self.g_idx = None if self.g_idx is not None: @@ -165,7 +175,6 @@ def forward(self, x): outshape = x.shape[:-1] + (self.outfeatures,) if HAS_GPTQ_CUDA and self.bits == 4: - if self.q4 is None: self.init_q4() @@ -191,7 +200,6 @@ def forward(self, x): def split_column_copy(gptq_linear, cai_linear, tp_size=1, tp_rank=0, split_num=1): - qweights = gptq_linear.qweight.split(gptq_linear.out_features // split_num, dim=-1) qzeros = gptq_linear.qzeros.split(gptq_linear.out_features // (32 // cai_linear.bits) // split_num, dim=-1) scales = gptq_linear.scales.split(gptq_linear.out_features // split_num, dim=-1) @@ -203,24 +211,24 @@ def split_column_copy(gptq_linear, cai_linear, tp_size=1, tp_rank=0, split_num=1 zero_split_block = cai_linear.outfeatures // (32 // cai_linear.bits) // split_num for i in range(split_num): - cai_linear.qweight[:, i * cai_split_out_features:(i + 1) * - cai_split_out_features] = qweights[i][:, tp_rank * cai_split_out_features:(tp_rank + 1) * - cai_split_out_features] - cai_linear.qzeros[:, i * zero_split_block:(i + 1) * - zero_split_block] = qzeros[i][:, tp_rank * zero_split_block:(tp_rank + 1) * zero_split_block] - cai_linear.scales[:, i * cai_split_out_features:(i + 1) * - cai_split_out_features] = scales[i][:, tp_rank * cai_split_out_features:(tp_rank + 1) * - cai_split_out_features] + cai_linear.qweight[:, i * cai_split_out_features : (i + 1) * cai_split_out_features] = qweights[i][ + :, tp_rank * cai_split_out_features : (tp_rank + 1) * cai_split_out_features + ] + cai_linear.qzeros[:, i * zero_split_block : (i + 1) * zero_split_block] = qzeros[i][ + :, tp_rank * zero_split_block : (tp_rank + 1) * zero_split_block + ] + cai_linear.scales[:, i * cai_split_out_features : (i + 1) * cai_split_out_features] = scales[i][ + :, tp_rank * cai_split_out_features : (tp_rank + 1) * cai_split_out_features + ] if cai_linear.bias is not None: - cai_linear.bias[i * cai_split_out_features:(i + 1) * - cai_split_out_features] = bias[i][tp_rank * cai_split_out_features:(tp_rank + 1) * - cai_split_out_features] + cai_linear.bias[i * cai_split_out_features : (i + 1) * cai_split_out_features] = bias[i][ + tp_rank * cai_split_out_features : (tp_rank + 1) * cai_split_out_features + ] cai_linear.g_idx.copy_(g_idx) def split_row_copy(gptq_linear, cai_linear, tp_rank=0, split_num=1): - qweights = gptq_linear.qweight.split(gptq_linear.in_features // split_num, dim=0) qzeros = gptq_linear.qzeros.split(gptq_linear.in_features // split_num, dim=0) scales = gptq_linear.scales.split(gptq_linear.in_features // split_num, dim=0) @@ -231,47 +239,40 @@ def split_row_copy(gptq_linear, cai_linear, tp_rank=0, split_num=1): idx_split_features = cai_linear.infeatures // split_num for i in range(split_num): - cai_linear.qweight[i * cai_split_in_features:(i + 1) * - cai_split_in_features, :] = qweights[i][tp_rank * cai_split_in_features:(tp_rank + 1) * - cai_split_in_features, :] - cai_linear.qzeros[i * zero_split_block:(i + 1) * - zero_split_block, :] = qzeros[i][tp_rank * zero_split_block:(tp_rank + 1) * - zero_split_block, :] - cai_linear.scales[i * zero_split_block:(i + 1) * - zero_split_block, :] = scales[i][tp_rank * zero_split_block:(tp_rank + 1) * - zero_split_block, :] - cai_linear.g_idx[i * idx_split_features:(i + 1) * - idx_split_features] = g_idxs[i][tp_rank * idx_split_features:(tp_rank + 1) * - idx_split_features] + cai_linear.qweight[i * cai_split_in_features : (i + 1) * cai_split_in_features, :] = qweights[i][ + tp_rank * cai_split_in_features : (tp_rank + 1) * cai_split_in_features, : + ] + cai_linear.qzeros[i * zero_split_block : (i + 1) * zero_split_block, :] = qzeros[i][ + tp_rank * zero_split_block : (tp_rank + 1) * zero_split_block, : + ] + cai_linear.scales[i * zero_split_block : (i + 1) * zero_split_block, :] = scales[i][ + tp_rank * zero_split_block : (tp_rank + 1) * zero_split_block, : + ] + cai_linear.g_idx[i * idx_split_features : (i + 1) * idx_split_features] = g_idxs[i][ + tp_rank * idx_split_features : (tp_rank + 1) * idx_split_features + ] if cai_linear.bias is not None: cai_linear.bias.copy_(gptq_linear.bias) class RowCaiQuantLinear(CaiQuantLinear, ParallelModule): - def __init__(self, bits, groupsize, infeatures, outfeatures, bias, tp_size=1, tp_rank=0, row_split=False): - - super().__init__(bits, - groupsize, - infeatures, - outfeatures, - bias, - tp_size=tp_size, - tp_rank=tp_rank, - row_split=row_split) + super().__init__( + bits, groupsize, infeatures, outfeatures, bias, tp_size=tp_size, tp_rank=tp_rank, row_split=row_split + ) self.process_group = None @staticmethod - def from_native_module(module: nn.Module, process_group: Union[ProcessGroup, List[ProcessGroup]], *args, - **kwargs) -> ParallelModule: + def from_native_module( + module: nn.Module, process_group: Union[ProcessGroup, List[ProcessGroup]], *args, **kwargs + ) -> ParallelModule: LazyInitContext.materialize(module) # get the attributes in_features = module.in_features # ensure only one process group is passed if isinstance(process_group, (list, tuple)): - assert len(process_group) == 1, \ - f'Expected only one process group, got {len(process_group)}.' + assert len(process_group) == 1, f"Expected only one process group, got {len(process_group)}." process_group = process_group[0] tp_size = dist.get_world_size(process_group) @@ -282,15 +283,18 @@ def from_native_module(module: nn.Module, process_group: Union[ProcessGroup, Lis if in_features % tp_size != 0: raise ValueError( - f"The size of in_features:{in_features} is not integer multiples of tensor parallel size: {tp_size}!") - linear_1d = RowCaiQuantLinear(module.bits, - module.group_size, - module.in_features // tp_size, - module.out_features, - module.bias is not None, - tp_size=tp_size, - tp_rank=tp_rank, - row_split=True) + f"The size of in_features:{in_features} is not integer multiples of tensor parallel size: {tp_size}!" + ) + linear_1d = RowCaiQuantLinear( + module.bits, + module.group_size, + module.in_features // tp_size, + module.out_features, + module.bias is not None, + tp_size=tp_size, + tp_rank=tp_rank, + row_split=True, + ) linear_1d.process_group = process_group split_row_copy(module, linear_1d, tp_rank=tp_rank, **kwargs) @@ -306,30 +310,23 @@ def forward(self, x): class ColCaiQuantLinear(CaiQuantLinear, ParallelModule): - def __init__(self, bits, groupsize, infeatures, outfeatures, bias, tp_size=1, tp_rank=0, row_split=False): - - super().__init__(bits, - groupsize, - infeatures, - outfeatures, - bias, - tp_size=tp_size, - tp_rank=tp_rank, - row_split=row_split) + super().__init__( + bits, groupsize, infeatures, outfeatures, bias, tp_size=tp_size, tp_rank=tp_rank, row_split=row_split + ) self.process_group = None @staticmethod - def from_native_module(module: nn.Module, process_group: Union[ProcessGroup, List[ProcessGroup]], *args, - **kwargs) -> ParallelModule: + def from_native_module( + module: nn.Module, process_group: Union[ProcessGroup, List[ProcessGroup]], *args, **kwargs + ) -> ParallelModule: LazyInitContext.materialize(module) # get the attributes in_features = module.in_features # ensure only one process group is passed if isinstance(process_group, (list, tuple)): - assert len(process_group) == 1, \ - f'Expected only one process group, got {len(process_group)}.' + assert len(process_group) == 1, f"Expected only one process group, got {len(process_group)}." process_group = process_group[0] tp_size = dist.get_world_size(process_group) @@ -340,14 +337,17 @@ def from_native_module(module: nn.Module, process_group: Union[ProcessGroup, Lis if in_features % tp_size != 0: raise ValueError( - f"The size of in_features:{in_features} is not integer multiples of tensor parallel size: {tp_size}!") - linear_1d = ColCaiQuantLinear(module.bits, - module.group_size, - module.in_features, - module.out_features // tp_size, - module.bias is not None, - tp_size=tp_size, - tp_rank=tp_rank) + f"The size of in_features:{in_features} is not integer multiples of tensor parallel size: {tp_size}!" + ) + linear_1d = ColCaiQuantLinear( + module.bits, + module.group_size, + module.in_features, + module.out_features // tp_size, + module.bias is not None, + tp_size=tp_size, + tp_rank=tp_rank, + ) linear_1d.process_group = process_group split_column_copy(module, linear_1d, tp_rank=tp_rank, **kwargs) diff --git a/colossalai/kernel/triton/context_attention.py b/colossalai/kernel/triton/context_attention.py index 3d9a23d2f5d2..1725581d637c 100644 --- a/colossalai/kernel/triton/context_attention.py +++ b/colossalai/kernel/triton/context_attention.py @@ -5,6 +5,7 @@ try: import triton import triton.language as tl + HAS_TRITON = True except ImportError: HAS_TRITON = False @@ -16,6 +17,7 @@ https://github.com/ModelTC/lightllm/blob/f093edc20683ac3ea1bca3fb5d8320a0dd36cf7b/lightllm/models/llama/triton_kernel/context_flashattention_nopad.py#L10 """ if triton.__version__ < "2.1.0": + @triton.jit def _context_flash_attention_kernel( Q, @@ -131,29 +133,47 @@ def _context_flash_attention_kernel( m_i = m_i_new off_o = ( - (cur_batch_start_index + offs_m[:, None]) * stride_obs + cur_head * stride_oh + offs_d[None, :] * stride_od + (cur_batch_start_index + offs_m[:, None]) * stride_obs + + cur_head * stride_oh + + offs_d[None, :] * stride_od ) out_ptrs = Out + off_o tl.store(out_ptrs, acc, mask=offs_m[:, None] < cur_batch_seq_len) return + else: # this function is modified from https://github.com/ModelTC/lightllm/blob/main/lightllm/models/llama/triton_kernel/context_flashattention_nopad.py#L11 @triton.jit def _context_flash_attention_kernel_2( - Q, K, V, sm_scale, Alibi, B_Start_Loc, B_Seqlen, - Out, - kv_group_num, - stride_qbs, stride_qh, stride_qd, - stride_kbs, stride_kh, stride_kd, - stride_vbs, stride_vh, stride_vd, - stride_obs, stride_oh, stride_od, - BLOCK_M: tl.constexpr, BLOCK_DMODEL: tl.constexpr, + Q, + K, + V, + sm_scale, + Alibi, + B_Start_Loc, + B_Seqlen, + Out, + kv_group_num, + stride_qbs, + stride_qh, + stride_qd, + stride_kbs, + stride_kh, + stride_kd, + stride_vbs, + stride_vh, + stride_vd, + stride_obs, + stride_oh, + stride_od, + BLOCK_M: tl.constexpr, + BLOCK_DMODEL: tl.constexpr, BLOCK_N: tl.constexpr, ): cur_batch = tl.program_id(0) cur_head = tl.program_id(1) start_m = tl.program_id(2) - + if kv_group_num is not None: cur_kv_head = cur_head // kv_group_num @@ -166,7 +186,11 @@ def _context_flash_attention_kernel_2( offs_n = tl.arange(0, BLOCK_N) offs_d = tl.arange(0, BLOCK_DMODEL) offs_m = start_m * BLOCK_M + tl.arange(0, BLOCK_M) - off_q = (cur_batch_in_all_start_index + offs_m[:, None]) * stride_qbs + cur_head * stride_qh + offs_d[None, :] * stride_qd + off_q = ( + (cur_batch_in_all_start_index + offs_m[:, None]) * stride_qbs + + cur_head * stride_qh + + offs_d[None, :] * stride_qd + ) if kv_group_num is None or kv_group_num == 1: off_k = offs_n[None, :] * stride_kbs + cur_head * stride_kh + offs_d[:, None] * stride_kd off_v = offs_n[:, None] * stride_vbs + cur_head * stride_vh + offs_d[None, :] * stride_vd @@ -191,8 +215,11 @@ def _context_flash_attention_kernel_2( for start_n in range(0, block_mask * (start_m + 1) * BLOCK_M, BLOCK_N): start_n = tl.multiple_of(start_n, BLOCK_N) # -- compute qk ---- - k = tl.load(k_ptrs + (cur_batch_in_all_start_index + start_n) * stride_kbs, - mask=(start_n + offs_n[None, :]) < cur_batch_seq_len, other=0.0) + k = tl.load( + k_ptrs + (cur_batch_in_all_start_index + start_n) * stride_kbs, + mask=(start_n + offs_n[None, :]) < cur_batch_seq_len, + other=0.0, + ) qk = tl.zeros([BLOCK_M, BLOCK_N], dtype=tl.float32) qk += tl.dot(q, k) @@ -220,8 +247,11 @@ def _context_flash_attention_kernel_2( acc_scale = l_i / l_i_new * alpha acc = acc * acc_scale[:, None] # update acc - v = tl.load(v_ptrs + (cur_batch_in_all_start_index + start_n) * stride_vbs, - mask=(start_n + offs_n[:, None]) < cur_batch_seq_len, other=0.0) + v = tl.load( + v_ptrs + (cur_batch_in_all_start_index + start_n) * stride_vbs, + mask=(start_n + offs_n[:, None]) < cur_batch_seq_len, + other=0.0, + ) p = p.to(v.dtype) acc += tl.dot(p, v) @@ -229,7 +259,11 @@ def _context_flash_attention_kernel_2( l_i = l_i_new m_i = m_i_new # initialize pointers to output - off_o = (cur_batch_in_all_start_index + offs_m[:, None]) * stride_obs + cur_head * stride_oh + offs_d[None, :] * stride_od + off_o = ( + (cur_batch_in_all_start_index + offs_m[:, None]) * stride_obs + + cur_head * stride_oh + + offs_d[None, :] * stride_od + ) out_ptrs = Out + off_o tl.store(out_ptrs, acc, mask=offs_m[:, None] < cur_batch_seq_len) return @@ -249,7 +283,7 @@ def bloom_context_attn_fwd(q, k, v, o, b_start_loc, b_seq_len, max_input_len, al grid = (batch, head, triton.cdiv(max_input_len, BLOCK)) num_warps = 4 if Lk <= 64 else 8 - + if triton.__version__ < "2.1.0": tmp = torch.empty((batch, head, max_input_len + 256), device=q.device, dtype=torch.float32) _context_flash_attention_kernel[grid]( @@ -286,20 +320,26 @@ def bloom_context_attn_fwd(q, k, v, o, b_start_loc, b_seq_len, max_input_len, al ) else: _context_flash_attention_kernel_2[grid]( - q, k, v, sm_scale, alibi, b_start_loc, b_seq_len, + q, + k, + v, + sm_scale, + alibi, + b_start_loc, + b_seq_len, o, None, - q.stride(0), - q.stride(1), + q.stride(0), + q.stride(1), q.stride(2), - k.stride(0), - k.stride(1), + k.stride(0), + k.stride(1), k.stride(2), - v.stride(0), - v.stride(1), + v.stride(0), + v.stride(1), v.stride(2), - o.stride(0), - o.stride(1), + o.stride(0), + o.stride(1), o.stride(2), BLOCK_M=BLOCK, BLOCK_DMODEL=Lk, @@ -307,7 +347,7 @@ def bloom_context_attn_fwd(q, k, v, o, b_start_loc, b_seq_len, max_input_len, al num_warps=num_warps, num_stages=1, ) - + return @torch.no_grad() @@ -327,7 +367,7 @@ def llama_context_attn_fwd(q, k, v, o, b_start_loc, b_seq_len, max_input_len): tmp = torch.empty((batch, head, max_input_len + 256), device=q.device, dtype=torch.float32) num_warps = 4 if Lk <= 64 else 8 # num_warps = 4 - + if triton.__version__ < "2.1.0": _context_flash_attention_kernel[grid]( q, @@ -337,7 +377,7 @@ def llama_context_attn_fwd(q, k, v, o, b_start_loc, b_seq_len, max_input_len): b_start_loc, b_seq_len, tmp, - None, + None, o, q.stride(0), q.stride(1), @@ -362,32 +402,33 @@ def llama_context_attn_fwd(q, k, v, o, b_start_loc, b_seq_len, max_input_len): ) else: kv_group_num = q.shape[1] // k.shape[1] - _context_flash_attention_kernel_2[grid]( - q, - k, - v, - sm_scale, + _context_flash_attention_kernel_2[grid]( + q, + k, + v, + sm_scale, None, - b_start_loc, + b_start_loc, b_seq_len, o, kv_group_num, - q.stride(0), - q.stride(1), + q.stride(0), + q.stride(1), q.stride(2), - k.stride(0), - k.stride(1), + k.stride(0), + k.stride(1), k.stride(2), - v.stride(0), - v.stride(1), + v.stride(0), + v.stride(1), v.stride(2), - o.stride(0), - o.stride(1), + o.stride(0), + o.stride(1), o.stride(2), BLOCK_M=BLOCK, BLOCK_DMODEL=Lk, BLOCK_N=BLOCK, num_warps=num_warps, - num_stages=1,) - - return \ No newline at end of file + num_stages=1, + ) + + return diff --git a/colossalai/kernel/triton/flash_decoding.py b/colossalai/kernel/triton/flash_decoding.py index 9b7b27fa1f49..ac733dede3b7 100644 --- a/colossalai/kernel/triton/flash_decoding.py +++ b/colossalai/kernel/triton/flash_decoding.py @@ -1,8 +1,10 @@ # adepted from https://github.com/ModelTC/lightllm/blob/ece7b43f8a6dfa74027adc77c2c176cff28c76c8/lightllm/models/llama/triton_kernel/flash_decoding.py import torch + try: from lightllm.models.llama.triton_kernel.flash_decoding_stage1 import flash_decode_stage1 from lightllm.models.llama.triton_kernel.flash_decoding_stage2 import flash_decode_stage2 + HAS_LIGHTLLM_KERNEL = True except: print("install lightllm from https://github.com/ModelTC/lightllm/blob/ece7b43f8a6dfa74027adc77c2c176cff28c76c8") @@ -10,41 +12,36 @@ if HAS_LIGHTLLM_KERNEL: + def token_flash_decoding(q, o_tensor, infer_state, q_head_num, head_dim, cache_k, cache_v): BLOCK_SEQ = 256 batch_size = infer_state.batch_size max_len_in_batch = infer_state.max_len_in_batch - calcu_shape1 = (batch_size, q_head_num, head_dim) - if getattr(infer_state, 'mid_o', None) is None: - infer_state.mid_o = torch.empty([batch_size, - q_head_num, - max_len_in_batch // BLOCK_SEQ + 1, - head_dim], - dtype=torch.float32, - device="cuda") - infer_state.mid_o_logexpsum = torch.empty([batch_size, - q_head_num, - max_len_in_batch // BLOCK_SEQ + 1], - dtype=torch.float32, - device="cuda") + if getattr(infer_state, "mid_o", None) is None: + infer_state.mid_o = torch.empty( + [batch_size, q_head_num, max_len_in_batch // BLOCK_SEQ + 1, head_dim], + dtype=torch.float32, + device="cuda", + ) + infer_state.mid_o_logexpsum = torch.empty( + [batch_size, q_head_num, max_len_in_batch // BLOCK_SEQ + 1], dtype=torch.float32, device="cuda" + ) mid_o = infer_state.mid_o mid_o_logexpsum = infer_state.mid_o_logexpsum - flash_decode_stage1(q.view(calcu_shape1), - cache_k, - cache_v, - infer_state.block_loc, - infer_state.seq_len, - infer_state.max_len_in_batch, - mid_o, - mid_o_logexpsum, - BLOCK_SEQ) - flash_decode_stage2(mid_o, - mid_o_logexpsum, - infer_state.seq_len, - o_tensor.view(calcu_shape1), - BLOCK_SEQ) + flash_decode_stage1( + q.view(calcu_shape1), + cache_k, + cache_v, + infer_state.block_loc, + infer_state.seq_len, + infer_state.max_len_in_batch, + mid_o, + mid_o_logexpsum, + BLOCK_SEQ, + ) + flash_decode_stage2(mid_o, mid_o_logexpsum, infer_state.seq_len, o_tensor.view(calcu_shape1), BLOCK_SEQ) diff --git a/colossalai/kernel/triton/llama_act_combine_kernel.py b/colossalai/kernel/triton/llama_act_combine_kernel.py index 45996c0dca53..7a2c7e8fbd74 100644 --- a/colossalai/kernel/triton/llama_act_combine_kernel.py +++ b/colossalai/kernel/triton/llama_act_combine_kernel.py @@ -8,6 +8,7 @@ try: import triton import triton.language as tl + HAS_TRITON = True except ImportError: HAS_TRITON = False @@ -26,8 +27,8 @@ def _llama_act_combine_forward( X_GATE2, X_UP, Y, - stride, # how much to increase the pointer when moving by 1 row - N, # number of columns in X + stride, # how much to increase the pointer when moving by 1 row + N, # number of columns in X BLOCK_SIZE: tl.constexpr, ): # Map the program id to the row of X and Y it should compute. @@ -41,9 +42,9 @@ def _llama_act_combine_forward( for off in range(0, N, BLOCK_SIZE): cols = off + tl.arange(0, BLOCK_SIZE) mask = cols < N - x_gate1 = tl.load(X_GATE1 + cols, mask=mask, other=0.) - x_gate2 = tl.load(X_GATE2 + cols, mask=mask, other=0.) - x_up = tl.load(X_UP + cols, mask=mask, other=0.) + x_gate1 = tl.load(X_GATE1 + cols, mask=mask, other=0.0) + x_gate2 = tl.load(X_GATE2 + cols, mask=mask, other=0.0) + x_up = tl.load(X_UP + cols, mask=mask, other=0.0) x_gate2_sigmoid = tl.sigmoid(x_gate2.to(tl.float32)).to(x_gate2.dtype) y = x_gate1 * x_gate2 * x_gate2_sigmoid * x_up # Write output @@ -58,8 +59,8 @@ def _llama_act_combine_backward( X_GATE2_GRAD, X_UP_GRAD, Y_GRAD, - stride, # how much to increase the pointer when moving by 1 row - N, # number of columns in X + stride, # how much to increase the pointer when moving by 1 row + N, # number of columns in X BLOCK_SIZE: tl.constexpr, ): # Map the program id to the row of X and Y it should compute. @@ -76,10 +77,10 @@ def _llama_act_combine_backward( for off in range(0, N, BLOCK_SIZE): cols = off + tl.arange(0, BLOCK_SIZE) mask = cols < N - x_gate1 = tl.load(X_GATE1 + cols, mask=mask, other=0.) - x_gate2 = tl.load(X_GATE2 + cols, mask=mask, other=0.) - x_up = tl.load(X_UP + cols, mask=mask, other=0.) - y_grad = tl.load(Y_GRAD + cols, mask=mask, other=0.) + x_gate1 = tl.load(X_GATE1 + cols, mask=mask, other=0.0) + x_gate2 = tl.load(X_GATE2 + cols, mask=mask, other=0.0) + x_up = tl.load(X_UP + cols, mask=mask, other=0.0) + y_grad = tl.load(Y_GRAD + cols, mask=mask, other=0.0) # forward: y = x_gate1 * x_gate2 * tl.sigmoid(x_gate2) * x_up x_gate2_sigmoid = tl.sigmoid(x_gate2.to(tl.float32)).to(x_gate2.dtype) @@ -147,14 +148,9 @@ def forward(ctx: Any, x_gate: torch.Tensor, x_up: torch.Tensor, activation: str # restore setting ctx.M, ctx.N, ctx.BLOCK_SIZE, ctx.num_warps = M, N, BLOCK_SIZE, num_warps # enqueue kernel - _llama_act_combine_forward[(M,)](x_gate1, - x_gate2, - x_up, - y, - x_up.stride(-2), - N, - BLOCK_SIZE=BLOCK_SIZE, - num_warps=num_warps) + _llama_act_combine_forward[(M,)]( + x_gate1, x_gate2, x_up, y, x_up.stride(-2), N, BLOCK_SIZE=BLOCK_SIZE, num_warps=num_warps + ) return y @staticmethod @@ -166,20 +162,25 @@ def backward(ctx: Any, *grad_outputs: Tensor) -> Tuple[Tensor, Tensor, None, Non # init grad y_grad = grad_outputs[0] - x_gate1_grad, x_gate2_grad, x_up_grad = torch.empty_like(x_gate1), torch.empty_like( - x_gate2), torch.empty_like(x_up) + x_gate1_grad, x_gate2_grad, x_up_grad = ( + torch.empty_like(x_gate1), + torch.empty_like(x_gate2), + torch.empty_like(x_up), + ) # enqueue kernel - _llama_act_combine_backward[(M,)](x_gate1, - x_gate2, - x_up, - x_gate1_grad, - x_gate2_grad, - x_up_grad, - y_grad, - x_up.stride(-2), - N, - BLOCK_SIZE=BLOCK_SIZE, - num_warps=num_warps) + _llama_act_combine_backward[(M,)]( + x_gate1, + x_gate2, + x_up, + x_gate1_grad, + x_gate2_grad, + x_up_grad, + y_grad, + x_up.stride(-2), + N, + BLOCK_SIZE=BLOCK_SIZE, + num_warps=num_warps, + ) x_gate_grad = torch.cat([x_gate1_grad, x_gate2_grad], dim=-1) return x_gate_grad, x_up_grad, None, None diff --git a/colossalai/kernel/triton/token_attention_kernel.py b/colossalai/kernel/triton/token_attention_kernel.py index de2003748e65..d8ac278c77dd 100644 --- a/colossalai/kernel/triton/token_attention_kernel.py +++ b/colossalai/kernel/triton/token_attention_kernel.py @@ -13,10 +13,18 @@ print("please install triton from https://github.com/openai/triton") try: - from lightllm.models.llama.triton_kernel.token_attention_nopad_reduceV import token_att_fwd2 as lightllm_llama_token_att_fwd2 - from lightllm.models.llama.triton_kernel.token_attention_nopad_att1 import token_att_fwd as lightllm_llama_token_att_fwd - from lightllm.models.llama.triton_kernel.token_attention_nopad_softmax import token_softmax_fwd as lightllm_llama_token_softmax_fwd - from lightllm.models.bloom.triton_kernel.token_attention_nopad_att1 import token_att_fwd as lightllm_bloom_token_att_fwd + from lightllm.models.bloom.triton_kernel.token_attention_nopad_att1 import ( + token_att_fwd as lightllm_bloom_token_att_fwd, + ) + from lightllm.models.llama.triton_kernel.token_attention_nopad_att1 import ( + token_att_fwd as lightllm_llama_token_att_fwd, + ) + from lightllm.models.llama.triton_kernel.token_attention_nopad_reduceV import ( + token_att_fwd2 as lightllm_llama_token_att_fwd2, + ) + from lightllm.models.llama.triton_kernel.token_attention_nopad_softmax import ( + token_softmax_fwd as lightllm_llama_token_softmax_fwd, + ) HAS_TRITON_TOKEN_ATTENTION = True except ImportError: @@ -205,9 +213,7 @@ def token_attn( if triton.__version__ == "2.0.0": prob = torch.empty_like(att_m_tensor) - lightllm_llama_token_softmax_fwd( - att_m_tensor, kv_cache_start_loc, kv_cache_seq_len, prob, max_len_in_batch - ) + lightllm_llama_token_softmax_fwd(att_m_tensor, kv_cache_start_loc, kv_cache_seq_len, prob, max_len_in_batch) att_m_tensor = None lightllm_llama_token_att_fwd2( diff --git a/colossalai/legacy/inference/tensor_parallel/modeling/llama.py b/colossalai/legacy/inference/tensor_parallel/modeling/llama.py index 448943b12c9e..21e5c5e7856b 100644 --- a/colossalai/legacy/inference/tensor_parallel/modeling/llama.py +++ b/colossalai/legacy/inference/tensor_parallel/modeling/llama.py @@ -8,7 +8,9 @@ from colossalai.inference.tensor_parallel.batch_infer_state import BatchInferState from colossalai.kernel.triton import llama_context_attn_fwd, token_attention_fwd from colossalai.kernel.triton.token_attention_kernel import Llama2TokenAttentionForwards + from ._utils import copy_kv_to_mem_cache + try: from lightllm.models.llama.triton_kernel.context_flashattention_nopad import ( context_attention_fwd as lightllm_llama_context_attention_fwd, @@ -90,7 +92,7 @@ def llama_triton_token_attention(query_states, attn_output, infer_state, num_key # infer_state.cache_manager.past_key_values_length, infer_state.max_len_in_batch, ) - + else: Llama2TokenAttentionForwards.token_attn( query_states, diff --git a/colossalai/shardformer/policies/base_policy.py b/colossalai/shardformer/policies/base_policy.py index 9a49b1ba6a14..1d2b7a570681 100644 --- a/colossalai/shardformer/policies/base_policy.py +++ b/colossalai/shardformer/policies/base_policy.py @@ -242,4 +242,4 @@ def get_stage_index( end_idx = num_layers_per_stage_accumulated[stage + model_chunk * num_stages + 1] stage_indices.append([start_idx, end_idx]) - return stage_indices[0] if num_model_chunks == 1 else stage_indices \ No newline at end of file + return stage_indices[0] if num_model_chunks == 1 else stage_indices diff --git a/colossalai/tensor/d_tensor/__init__.py b/colossalai/tensor/d_tensor/__init__.py index 6f8097735d57..4129ec62e956 100644 --- a/colossalai/tensor/d_tensor/__init__.py +++ b/colossalai/tensor/d_tensor/__init__.py @@ -2,13 +2,13 @@ compute_global_numel, customized_distributed_tensor_to_param, distribute_tensor, - init_as_dtensor, distribute_tensor_with_customization, - init_tensor_as_customization_distributed, get_device_mesh, get_global_shape, get_layout, get_sharding_spec, + init_as_dtensor, + init_tensor_as_customization_distributed, is_customized_distributed_tensor, is_distributed_tensor, is_sharded, diff --git a/colossalai/tensor/d_tensor/api.py b/colossalai/tensor/d_tensor/api.py index da6ef275e108..7258170887ad 100644 --- a/colossalai/tensor/d_tensor/api.py +++ b/colossalai/tensor/d_tensor/api.py @@ -128,7 +128,10 @@ def distribute_tensor(tensor: torch.Tensor, device_mesh: DeviceMesh, sharding_sp return sharded_tensor -def init_as_dtensor(tensor: torch.Tensor, device_mesh: DeviceMesh, sharding_spec: ShardingSpec, global_shape: torch.Size) -> torch.Tensor: + +def init_as_dtensor( + tensor: torch.Tensor, device_mesh: DeviceMesh, sharding_spec: ShardingSpec, global_shape: torch.Size +) -> torch.Tensor: assert not is_distributed_tensor(tensor), "The input tensor is already a distributed tensor." dist_layout = Layout(device_mesh=device_mesh, sharding_spec=sharding_spec, global_shape=global_shape) @@ -140,6 +143,7 @@ def init_as_dtensor(tensor: torch.Tensor, device_mesh: DeviceMesh, sharding_spec return tensor + def redistribute(dtensor: torch.Tensor, device_mesh: DeviceMesh, sharding_spec: ShardingSpec) -> None: """ Convert the layout of the tensor from source_spec to target_spec. @@ -468,7 +472,6 @@ def gather_fn(tensor): assert callable(gather_fn), "The gather_fn must be callable." assert not is_distributed_tensor(tensor), "The input tensor is already a distributed tensor." - # set the shard_fn and gather_fn as attributes of the distributed tensor tensor.shard_fn = shard_fn tensor.gather_fn = gather_fn diff --git a/colossalai/zero/low_level/_utils.py b/colossalai/zero/low_level/_utils.py index de08ecf3d57f..5ab703f09063 100644 --- a/colossalai/zero/low_level/_utils.py +++ b/colossalai/zero/low_level/_utils.py @@ -190,6 +190,7 @@ def calculate_global_norm_from_list(norm_list): total_norm += norm**2.0 return math.sqrt(total_norm) + def sync_tensor(flat_tensor, tensor_list): """ Synchronize the flattened tensor and unflattened tensor list. When diff --git a/docs/source/zh-Hans/advanced_tutorials/train_vit_with_hybrid_parallelism.md b/docs/source/zh-Hans/advanced_tutorials/train_vit_with_hybrid_parallelism.md index 3de41601a231..c25b109e7106 100644 --- a/docs/source/zh-Hans/advanced_tutorials/train_vit_with_hybrid_parallelism.md +++ b/docs/source/zh-Hans/advanced_tutorials/train_vit_with_hybrid_parallelism.md @@ -220,7 +220,7 @@ model, optimizer, _criterion, train_dataloader, lr_scheduler = booster.boost( ) ``` ## 使用混合并行训练 ViT -最后就可以使用混合并行策略来训练模型了。我们先定义一个训练函数，描述训练过程。需要注意的是，如果使用了管道并行策略，需要调用`booster.execute_pipeline`来执行模型的训练，它会调用`scheduler`管理模型的前后向操作。 +最后就可以使用混合并行策略来训练模型了。我们先定义一个训练函数，描述训练过程。需要注意的是，如果使用了管道并行策略，需要调用`booster.execute_pipeline`来执行模型的训练，它会调用`scheduler`管理模型的前后向操作。 ```python def run_forward_backward( model: nn.Module, diff --git a/examples/language/data_utils.py b/examples/language/data_utils.py index ec849ef9d1eb..6b9e8ef28eb7 100644 --- a/examples/language/data_utils.py +++ b/examples/language/data_utils.py @@ -121,4 +121,4 @@ def __getitem__(self, idx): "input_ids": self.input_ids[idx], "attention_mask": self.attention_mask[idx], "labels": self.input_ids[idx], - } \ No newline at end of file + } diff --git a/examples/language/openmoe/benchmark/utils.py b/examples/language/openmoe/benchmark/utils.py index 7a0955bb028a..096e06bd21b8 100644 --- a/examples/language/openmoe/benchmark/utils.py +++ b/examples/language/openmoe/benchmark/utils.py @@ -50,7 +50,6 @@ def all_reduce_mean(x: float, world_size: int) -> float: class Timer: - def __init__(self) -> None: self.start_time: Optional[float] = None self.duration: float = 0.0 @@ -112,7 +111,7 @@ def on_step_end(self, input_ids: Tensor, **kwargs) -> None: batch_size, seq_len = input_ids.shape self.num_samples += batch_size - self.flop += (batch_size * seq_len * self.model_numel * 2 * (3 + int(self.enable_grad_checkpoint))) + self.flop += batch_size * seq_len * self.model_numel * 2 * (3 + int(self.enable_grad_checkpoint)) def on_fit_end(self) -> None: avg_duration = all_reduce_mean(self.timer.duration, self.world_size) @@ -122,5 +121,6 @@ def on_fit_end(self) -> None: if dist.get_rank() == 0: print( f"num_samples: {self.num_samples}, dp_world_size: {self.dp_world_size}, flop: {self.flop}, avg_duration: {avg_duration}, " - f"avg_throughput: {avg_throughput}") + f"avg_throughput: {avg_throughput}" + ) print(f"Throughput: {avg_throughput:.2f} samples/sec, TFLOPS per GPU: {avg_tflops_per_gpu:.2f}") diff --git a/examples/language/openmoe/infer.py b/examples/language/openmoe/infer.py index 04df64531937..50cdc63e8cd9 100644 --- a/examples/language/openmoe/infer.py +++ b/examples/language/openmoe/infer.py @@ -16,17 +16,15 @@ def inference(args): tokenizer = T5Tokenizer.from_pretrained("google/umt5-small") if args.model == "test": config = LlamaConfig.from_pretrained("hpcai-tech/openmoe-base") - set_openmoe_args(config, - num_experts=config.num_experts, - moe_layer_interval=config.moe_layer_interval, - enable_kernel=True) + set_openmoe_args( + config, num_experts=config.num_experts, moe_layer_interval=config.moe_layer_interval, enable_kernel=True + ) model = OpenMoeForCausalLM(config) else: config = LlamaConfig.from_pretrained(f"hpcai-tech/openmoe-{args.model}") - set_openmoe_args(config, - num_experts=config.num_experts, - moe_layer_interval=config.moe_layer_interval, - enable_kernel=False) + set_openmoe_args( + config, num_experts=config.num_experts, moe_layer_interval=config.moe_layer_interval, enable_kernel=False + ) model = OpenMoeForCausalLM.from_pretrained(f"hpcai-tech/openmoe-{args.model}", config=config) model = model.eval().bfloat16() model = model.to(torch.cuda.current_device()) diff --git a/examples/language/openmoe/model/convert_openmoe_ckpt.py b/examples/language/openmoe/model/convert_openmoe_ckpt.py index 20b1e780d8b3..3e051850d1cd 100644 --- a/examples/language/openmoe/model/convert_openmoe_ckpt.py +++ b/examples/language/openmoe/model/convert_openmoe_ckpt.py @@ -172,9 +172,9 @@ def make_state_dict(converted_params): def load_t5x_weights_in_t5(model, config, t5x_checkpoint_path): """Replaces the params in model witht the T5X converted params.""" variables = checkpoints.load_t5x_checkpoint(t5x_checkpoint_path) - converted = convert_t5x_to_pytorch(variables, - num_layers=config.num_hidden_layers, - moe_interval=config.moe_layer_interval) + converted = convert_t5x_to_pytorch( + variables, num_layers=config.num_hidden_layers, moe_interval=config.moe_layer_interval + ) state_dict = make_state_dict(converted) model.load_state_dict(state_dict, strict=True) @@ -203,11 +203,9 @@ def convert_t5x_checkpoint_to_pytorch(t5x_checkpoint_path, config_file, pytorch_ if __name__ == "__main__": parser = argparse.ArgumentParser(description="Converts a native T5X checkpoint into a PyTorch checkpoint.") # Required parameters - parser.add_argument("--t5x_checkpoint_path", - default=None, - type=str, - required=True, - help="Path to the T5X checkpoint.") + parser.add_argument( + "--t5x_checkpoint_path", default=None, type=str, required=True, help="Path to the T5X checkpoint." + ) parser.add_argument( "--config_file", default=None, @@ -215,10 +213,8 @@ def convert_t5x_checkpoint_to_pytorch(t5x_checkpoint_path, config_file, pytorch_ required=True, help="The config json file corresponding to the pre-trained T5 model.\nThis specifies the model architecture.", ) - parser.add_argument("--pytorch_dump_path", - default=None, - type=str, - required=True, - help="Path to the output PyTorch model.") + parser.add_argument( + "--pytorch_dump_path", default=None, type=str, required=True, help="Path to the output PyTorch model." + ) args = parser.parse_args() convert_t5x_checkpoint_to_pytorch(args.t5x_checkpoint_path, args.config_file, args.pytorch_dump_path) diff --git a/examples/language/openmoe/model/openmoe_policy.py b/examples/language/openmoe/model/openmoe_policy.py index 17e7aa46ce85..17a51d6398d0 100644 --- a/examples/language/openmoe/model/openmoe_policy.py +++ b/examples/language/openmoe/model/openmoe_policy.py @@ -1,4 +1,3 @@ -import warnings from functools import partial from typing import Callable, Dict, List, Optional, Union @@ -21,7 +20,6 @@ class OpenMoePolicy(Policy): - def config_sanity_check(self): pass @@ -43,7 +41,8 @@ def module_policy(self) -> Dict[Union[str, nn.Module], ModulePolicyDescription]: if self.shard_config.enable_sequence_parallelism: self.shard_config.enable_sequence_parallelism = False raise NotImplementedError( - "openmoe doesn't support sequence parallelism now, will ignore the sequence parallelism flag.") + "openmoe doesn't support sequence parallelism now, will ignore the sequence parallelism flag." + ) if self.shard_config.enable_tensor_parallelism: raise NotImplementedError("Tensor parallelism is not supported for openmoe model now.") @@ -100,9 +99,9 @@ def set_pipeline_forward(self, model_cls: nn.Module, new_forward: Callable, poli layers_per_stage = self.distribute_layers(len(module.layers), stage_manager.num_stages) stage_index = Policy.get_stage_index(layers_per_stage, stage_manager.stage) method_replacement = {"forward": partial(new_forward, stage_manager=stage_manager, stage_index=stage_index)} - self.append_or_create_method_replacement(description=method_replacement, - policy=policy, - target_key=model_cls) + self.append_or_create_method_replacement( + description=method_replacement, policy=policy, target_key=model_cls + ) return @@ -126,12 +125,10 @@ def get_held_layers(self) -> List[Module]: held_layers.append(module.norm) return held_layers - + @staticmethod def distribute_layers(num_layers: int, num_stages: int) -> List[int]: - """Divide layers into stages - - """ + """Divide layers into stages""" if num_layers == 24 and num_stages == 4: return [7, 7, 7, 3] elif num_layers == 24 and num_stages == 2: @@ -146,7 +143,6 @@ def distribute_layers(num_layers: int, num_stages: int) -> List[int]: class OpenMoeModelPolicy(OpenMoePolicy): - def __init__(self) -> None: super().__init__() @@ -172,21 +168,21 @@ def get_shared_params(self) -> List[Dict[int, Tensor]]: class OpenMoeForCausalLMPolicy(OpenMoePolicy): - def module_policy(self): policy = super().module_policy() if self.shard_config.enable_tensor_parallelism: # add a new item for casual lm new_item = { - OpenMoeForCausalLM: - ModulePolicyDescription(sub_module_replacement=[ + OpenMoeForCausalLM: ModulePolicyDescription( + sub_module_replacement=[ SubModuleReplacementDescription( suffix="lm_head", target_module=Linear1D_Col, kwargs=dict(gather_output=True), ) - ]) + ] + ) } policy.update(new_item) @@ -211,13 +207,17 @@ def get_held_layers(self) -> List[Module]: def get_shared_params(self) -> List[Dict[int, Tensor]]: llama_model = self.model.model if self.pipeline_stage_manager and self.pipeline_stage_manager.num_stages > 1: - if (id(llama_model.embed_tokens.weight) == id(self.model.lm_head.weight) - and self.pipeline_stage_manager.num_stages > 1): + if ( + id(llama_model.embed_tokens.weight) == id(self.model.lm_head.weight) + and self.pipeline_stage_manager.num_stages > 1 + ): # tie weights - return [{ - 0: llama_model.embed_tokens.weight, - self.pipeline_stage_manager.num_stages - 1: self.model.lm_head.weight, - }] + return [ + { + 0: llama_model.embed_tokens.weight, + self.pipeline_stage_manager.num_stages - 1: self.model.lm_head.weight, + } + ] return [] @@ -250,12 +250,13 @@ def openmoe_model_forward( logger = logging.get_logger(__name__) - output_attentions = (output_attentions if output_attentions is not None else self.config.output_attentions) - output_hidden_states = (output_hidden_states - if output_hidden_states is not None else self.config.output_hidden_states) + output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions + output_hidden_states = ( + output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states + ) use_cache = use_cache if use_cache is not None else self.config.use_cache - return_dict = (return_dict if return_dict is not None else self.config.use_return_dict) + return_dict = return_dict if return_dict is not None else self.config.use_return_dict # retrieve input_ids and inputs_embeds if stage_manager.is_first_stage(): @@ -323,7 +324,8 @@ def openmoe_model_forward( if self.gradient_checkpointing and self.training: if use_cache: logger.warning_once( - "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`...") + "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..." + ) use_cache = False # decoder layers @@ -336,12 +338,11 @@ def openmoe_model_forward( if output_hidden_states: all_hidden_states += (hidden_states,) - past_key_value = (past_key_values[idx] if past_key_values is not None else None) + past_key_value = past_key_values[idx] if past_key_values is not None else None if self.gradient_checkpointing and self.training: def create_custom_forward(module): - def custom_forward(*inputs): # None for past_key_value return module(*inputs, output_attentions, None) @@ -387,14 +388,16 @@ def custom_forward(*inputs): router_z_loss = past_router_z_loss + router_z_loss if stage_manager.is_last_stage(): - return tuple([ - hidden_states, - next_cache, - all_hidden_states, - all_self_attns, - router_aux_loss, - router_z_loss, - ]) + return tuple( + [ + hidden_states, + next_cache, + all_hidden_states, + all_self_attns, + router_aux_loss, + router_z_loss, + ] + ) # always return dict for imediate stage return { "hidden_states": hidden_states, @@ -448,10 +451,11 @@ def llama_for_causal_lm_forward( "Hey, are you consciours? Can you talk to me?\nI'm not consciours, but I can talk to you." ```""" logger = logging.get_logger(__name__) - output_attentions = (output_attentions if output_attentions is not None else self.config.output_attentions) - output_hidden_states = (output_hidden_states - if output_hidden_states is not None else self.config.output_hidden_states) - return_dict = (return_dict if return_dict is not None else self.config.use_return_dict) + output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions + output_hidden_states = ( + output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states + ) + return_dict = return_dict if return_dict is not None else self.config.use_return_dict # TODO(jianghai): left the recording kv-value tensors as () or None type, this feature may be added in the future. if output_attentions: @@ -507,7 +511,6 @@ def llama_for_causal_lm_forward( if chunk_head == True: def create_custom_forward(module): - def custom_forward(*inputs): logits = module(inputs[0]) logits = logits.float() @@ -525,8 +528,8 @@ def custom_forward(*inputs): for batch_idx in range(hidden_states.shape[0]): loss = loss + torch.utils.checkpoint.checkpoint( create_custom_forward(self.lm_head), - hidden_states[batch_idx:batch_idx + 1, :], - labels[batch_idx:batch_idx + 1, :], + hidden_states[batch_idx : batch_idx + 1, :], + labels[batch_idx : batch_idx + 1, :], ) logits = None else: diff --git a/extensions/cpu_adam/__init__.py b/extensions/cpu_adam/__init__.py index cfd26a6a20f8..d5c69e902a80 100644 --- a/extensions/cpu_adam/__init__.py +++ b/extensions/cpu_adam/__init__.py @@ -1,5 +1,4 @@ from .cpu_adam_arm import CpuAdamArmExtension from .cpu_adam_x86 import CpuAdamX86Extension -__all__ = ['CpuAdamArmExtension', 'CpuAdamX86Extension'] - +__all__ = ["CpuAdamArmExtension", "CpuAdamX86Extension"] diff --git a/extensions/layernorm/__init__.py b/extensions/layernorm/__init__.py index 9d1bd2d019ee..30e6c68eff89 100644 --- a/extensions/layernorm/__init__.py +++ b/extensions/layernorm/__init__.py @@ -1,3 +1,3 @@ from .layernorm_cuda import LayerNormCudaExtension -__all__ = ["LayerNormCudaExtension"] \ No newline at end of file +__all__ = ["LayerNormCudaExtension"] diff --git a/extensions/moe/__init__.py b/extensions/moe/__init__.py index 962084d4bdde..3b6aa24bf7f6 100644 --- a/extensions/moe/__init__.py +++ b/extensions/moe/__init__.py @@ -1,3 +1,3 @@ from .moe_cuda import MoeCudaExtension -__all__ = ['MoeCudaExtension'] \ No newline at end of file +__all__ = ["MoeCudaExtension"] diff --git a/extensions/optimizer/__init__.py b/extensions/optimizer/__init__.py index 9c8e87cae5de..6a0c8d7b8016 100644 --- a/extensions/optimizer/__init__.py +++ b/extensions/optimizer/__init__.py @@ -1,3 +1,3 @@ from .fused_optimizer_cuda import FusedOptimizerCudaExtension -__all__ = ['FusedOptimizerCudaExtension'] \ No newline at end of file +__all__ = ["FusedOptimizerCudaExtension"] diff --git a/extensions/softmax/__init__.py b/extensions/softmax/__init__.py index 9bc50c6cd91c..8833d93e73d0 100644 --- a/extensions/softmax/__init__.py +++ b/extensions/softmax/__init__.py @@ -1,4 +1,4 @@ from .scaled_masked_softmax_cuda import ScaledMaskedSoftmaxCudaExtension from .scaled_upper_triangle_masked_softmax_cuda import ScaledUpperTriangleMaskedSoftmaxCudaExtension -__all__ = ['ScaledMaskedSoftmaxCudaExtension', 'ScaledUpperTriangleMaskedSoftmaxCudaExtension'] \ No newline at end of file +__all__ = ["ScaledMaskedSoftmaxCudaExtension", "ScaledUpperTriangleMaskedSoftmaxCudaExtension"] diff --git a/tests/kit/model_zoo/__init__.py b/tests/kit/model_zoo/__init__.py index 5f6789ff3357..66c794a7d891 100644 --- a/tests/kit/model_zoo/__init__.py +++ b/tests/kit/model_zoo/__init__.py @@ -1,33 +1,33 @@ import os + from . import custom, diffusers, timm, torchaudio, torchvision, transformers from .executor import run_fwd, run_fwd_bwd from .registry import model_zoo # We pick a subset of models for fast testing in order to reduce the total testing time COMMON_MODELS = [ - 'custom_hanging_param_model', - 'custom_nested_model', - 'custom_repeated_computed_layers', - 'custom_simple_net', - 'diffusers_clip_text_model', - 'diffusers_auto_encoder_kl', - 'diffusers_unet2d_model', - 'timm_densenet', - 'timm_resnet', - 'timm_swin_transformer', - 'torchaudio_wav2vec2_base', - 'torchaudio_conformer', - 'transformers_bert_for_masked_lm', - 'transformers_bloom_for_causal_lm', - 'transformers_falcon_for_causal_lm', - 'transformers_chatglm_for_conditional_generation', - 'transformers_llama_for_casual_lm', - 'transformers_vit_for_masked_image_modeling', - 'transformers_mistral_for_casual_lm' + "custom_hanging_param_model", + "custom_nested_model", + "custom_repeated_computed_layers", + "custom_simple_net", + "diffusers_clip_text_model", + "diffusers_auto_encoder_kl", + "diffusers_unet2d_model", + "timm_densenet", + "timm_resnet", + "timm_swin_transformer", + "torchaudio_wav2vec2_base", + "torchaudio_conformer", + "transformers_bert_for_masked_lm", + "transformers_bloom_for_causal_lm", + "transformers_falcon_for_causal_lm", + "transformers_chatglm_for_conditional_generation", + "transformers_llama_for_casual_lm", + "transformers_vit_for_masked_image_modeling", + "transformers_mistral_for_casual_lm", ] -IS_FAST_TEST = os.environ.get('FAST_TEST', '0') == '1' - +IS_FAST_TEST = os.environ.get("FAST_TEST", "0") == "1" -__all__ = ["model_zoo", "run_fwd", "run_fwd_bwd", 'COMMON_MODELS', 'IS_FAST_TEST'] +__all__ = ["model_zoo", "run_fwd", "run_fwd_bwd", "COMMON_MODELS", "IS_FAST_TEST"] diff --git a/tests/kit/model_zoo/registry.py b/tests/kit/model_zoo/registry.py index fce81ab52c2b..a16b16ad6af7 100644 --- a/tests/kit/model_zoo/registry.py +++ b/tests/kit/model_zoo/registry.py @@ -102,4 +102,4 @@ def get_sub_registry( return new_dict -model_zoo = ModelZooRegistry() \ No newline at end of file +model_zoo = ModelZooRegistry() diff --git a/tests/kit/model_zoo/transformers/chatglm2.py b/tests/kit/model_zoo/transformers/chatglm2.py index e27fdb4e2efe..0b178d58ce33 100644 --- a/tests/kit/model_zoo/transformers/chatglm2.py +++ b/tests/kit/model_zoo/transformers/chatglm2.py @@ -2,6 +2,7 @@ from colossalai.shardformer.modeling.chatglm2_6b.configuration_chatglm import ChatGLMConfig from colossalai.shardformer.modeling.chatglm2_6b.modeling_chatglm import ChatGLMForConditionalGeneration, ChatGLMModel + from ..registry import ModelAttribute, model_zoo # ================================ diff --git a/tests/test_booster/test_plugin/test_3d_plugin.py b/tests/test_booster/test_plugin/test_3d_plugin.py index 38361d803c49..d629e769d715 100644 --- a/tests/test_booster/test_plugin/test_3d_plugin.py +++ b/tests/test_booster/test_plugin/test_3d_plugin.py @@ -260,7 +260,7 @@ def run_grad_acc_test(test_args): origin_model, origin_optimizer, dataloader=dataloader ) for p1, p2 in zip(model.unwrap().parameters(), origin_model.unwrap().parameters()): - assert_close(p1.to(p2.dtype), p2, atol=1e-2, rtol=1e-2) + assert_close(p1.to(p2.dtype), p2, atol=1e-2, rtol=1e-2) def run_dist(rank, world_size, port, early_stop: bool = True): @@ -276,4 +276,4 @@ def test_gemini_plugin(early_stop: bool = True): if __name__ == "__main__": - test_gemini_plugin(early_stop=False) \ No newline at end of file + test_gemini_plugin(early_stop=False) diff --git a/tests/test_checkpoint_io/test_torch_fsdp_checkpoint_io.py b/tests/test_checkpoint_io/test_torch_fsdp_checkpoint_io.py index dca562a3b837..1ea70368eabf 100644 --- a/tests/test_checkpoint_io/test_torch_fsdp_checkpoint_io.py +++ b/tests/test_checkpoint_io/test_torch_fsdp_checkpoint_io.py @@ -113,6 +113,7 @@ def run_model(): full_osd = FSDP.full_optim_state_dict(optimizer.unwrap_model().unwrap(), optim=optimizer) import copy + sharded_osd = copy.deepcopy(full_osd) run_model() diff --git a/tests/test_gptq/test_gptq_linear.py b/tests/test_gptq/test_gptq_linear.py index 9b650aa78112..ded70fa43c30 100644 --- a/tests/test_gptq/test_gptq_linear.py +++ b/tests/test_gptq/test_gptq_linear.py @@ -1,16 +1,8 @@ -import math -import time - -import numpy as np import pytest import torch -import torch.nn as nn -import transformers from packaging import version try: - import triton - import triton.language as tl HAS_TRITON = True except ImportError: HAS_TRITON = False @@ -22,6 +14,7 @@ from exllama_kernels import prepare_buffers, set_tuning_params from colossalai.inference.quant.gptq import CaiQuantLinear + HAS_AUTO_GPTQ = True except: HAS_AUTO_GPTQ = False @@ -32,13 +25,14 @@ HAS_GPTQ_CUDA = False try: from colossalai.kernel.op_builder.gptq import GPTQBuilder + gptq_cuda = GPTQBuilder().load() HAS_GPTQ_CUDA = True except ImportError: - warnings.warn('CUDA gptq is not installed') + warnings.warn("CUDA gptq is not installed") HAS_GPTQ_CUDA = False -TRITON_CUDA_SUPPORT = version.parse(torch.version.cuda) > version.parse('11.4') +TRITON_CUDA_SUPPORT = version.parse(torch.version.cuda) > version.parse("11.4") max_inner_outer_dim = 1 max_input_len = 1 @@ -64,9 +58,9 @@ def init_buffer(cai_linear, use_act_order=False): max_input_len = 4096 # The temp_state buffer is required to reorder X in the act-order case. # The temp_dq buffer is required to dequantize weights when using cuBLAS, typically for the prefill. - gptq_temp_state_buffer = torch.zeros((max_input_len, max_inner_outer_dim), - dtype=torch.float16, - device=torch.cuda.current_device()) + gptq_temp_state_buffer = torch.zeros( + (max_input_len, max_inner_outer_dim), dtype=torch.float16, device=torch.cuda.current_device() + ) gptq_temp_dq_buffer = torch.zeros((1, max_dq_buffer_size), dtype=torch.float16, device=torch.cuda.current_device()) gptq_cuda.prepare_buffers(torch.device(torch.cuda.current_device()), gptq_temp_state_buffer, gptq_temp_dq_buffer) @@ -77,10 +71,11 @@ def init_buffer(cai_linear, use_act_order=False): gptq_cuda.set_tuning_params(matmul_recons_thd, matmul_fused_remap, matmul_no_half2) -@pytest.mark.skipif(not TRITON_CUDA_SUPPORT or not HAS_TRITON or not HAS_AUTO_GPTQ, - reason="triton requires cuda version to be higher than 11.4 or not install auto-gptq") +@pytest.mark.skipif( + not TRITON_CUDA_SUPPORT or not HAS_TRITON or not HAS_AUTO_GPTQ, + reason="triton requires cuda version to be higher than 11.4 or not install auto-gptq", +) def test_gptq_linear(): - infeature = 1024 outfeature = 1024 group_size = 128 @@ -120,7 +115,7 @@ def test_gptq_linear(): max_input_len = 2048 buffers = { "temp_state": torch.zeros((max_input_len, max_inner_outer_dim), dtype=torch.float16, device=device), - "temp_dq": torch.zeros((1, max_dq_buffer_size), dtype=torch.float16, device=device) + "temp_dq": torch.zeros((1, max_dq_buffer_size), dtype=torch.float16, device=device), } prepare_buffers(device, buffers["temp_state"], buffers["temp_dq"]) @@ -146,5 +141,4 @@ def test_gptq_linear(): if __name__ == "__main__": - test_gptq_linear() diff --git a/tests/test_lazy/test_models.py b/tests/test_lazy/test_models.py index aeca5f21dc1d..d0c4cd0a7c48 100644 --- a/tests/test_lazy/test_models.py +++ b/tests/test_lazy/test_models.py @@ -24,4 +24,4 @@ def test_torchvision_models_lazy_init(subset, default_device): if __name__ == "__main__": - test_torchvision_models_lazy_init("transformers", "cpu") \ No newline at end of file + test_torchvision_models_lazy_init("transformers", "cpu") diff --git a/tests/test_shardformer/test_layer/test_dist_crossentropy.py b/tests/test_shardformer/test_layer/test_dist_crossentropy.py index f594a80a43e0..414157c2233d 100644 --- a/tests/test_shardformer/test_layer/test_dist_crossentropy.py +++ b/tests/test_shardformer/test_layer/test_dist_crossentropy.py @@ -38,9 +38,10 @@ def check_dist_crossentropy(rank, world_size, port, ignore_index): org_loss, dist_loss, atol=1e-5 ), f"dist cross entropy loss is not equal to orgin loss\n{org_loss}\n{dist_loss}" - target_grad = torch.chunk(pred.grad, world_size, dim=-1)[rank] - assert torch.allclose(target_grad, dist_pred.grad), f"dist grad is not equal to orgin grad\n{target_grad}\n{dist_pred.grad}" + assert torch.allclose( + target_grad, dist_pred.grad + ), f"dist grad is not equal to orgin grad\n{target_grad}\n{dist_pred.grad}" @pytest.mark.dist diff --git a/tests/test_shardformer/test_model/test_shard_gptj.py b/tests/test_shardformer/test_model/test_shard_gptj.py index c83eaaa09e29..0bf9669808fa 100644 --- a/tests/test_shardformer/test_model/test_shard_gptj.py +++ b/tests/test_shardformer/test_model/test_shard_gptj.py @@ -207,6 +207,7 @@ def check_gptj_3d(rank, world_size, port): colossalai.launch(config={}, rank=rank, world_size=world_size, host="localhost", port=port, backend="nccl") run_gptj_3d_test() + @pytest.mark.skip("TODO check_gptj has something wrong.") @pytest.mark.dist @rerun_if_address_is_in_use() From bb08c5f3e11c6e1fa4abff4b158c975e604b5d58 Mon Sep 17 00:00:00 2001 From: flybird11111 <1829166702@qq.com> Date: Wed, 10 Apr 2024 15:00:54 +0800 Subject: [PATCH 3/7] [rebase] rebase main to resize-embedding (#5581) * [release] grok-1 314b inference (#5490) * [release] grok-1 inference * [release] grok-1 inference * [release] grok-1 inference * [example] update Grok-1 inference (#5495) * revise grok-1 example * remove unused arg in scripts * prevent re-installing torch * update readme * revert modifying colossalai requirements * add perf * trivial * add tokenizer url * [hotfix] set return_outputs=False in examples and polish code (#5404) * fix: simplify merge_batch * fix: use return_outputs=False to eliminate extra memory consumption * feat: add return_outputs warning * style: remove `return_outputs=False` as it is the default value * [release] grok-1 inference benchmark (#5500) * [release] grok-1 inference benchmark * [release] grok-1 inference benchmark * [release] grok-1 inference benchmark * [release] grok-1 inference benchmark * [release] grok-1 inference benchmark * [shardformer]Fix lm parallel. (#5480) * fix * padding vocab_size when using pipeline parallellism padding vocab_size when using pipeline parallellism fix fix * fix * fix fix fix * fix gather output * fix * fix * fix fix resize embedding fix resize embedding * fix resize embedding fix * revert * revert * revert * fix lm forward distribution * fix * test ci * fix * [fix] fix grok-1 example typo (#5506) * [devops] fix example test ci (#5504) * Fix ColoTensorSpec for py11 (#5440) * fixed layout converter caching and updated tester * Empty-Commit * [shardformer] update colo attention to support custom mask (#5510) * [feature] refactor colo attention (#5462) * [extension] update api * [feature] add colo attention * [feature] update sdpa * [feature] update npu attention * [feature] update flash-attn * [test] add flash attn test * [test] update flash attn test * [shardformer] update modeling to fit colo attention (#5465) * [misc] refactor folder structure * [shardformer] update llama flash-attn * [shardformer] fix llama policy * [devops] update tensornvme install * [test] update llama test * [shardformer] update colo attn kernel dispatch * [shardformer] update blip2 * [shardformer] update chatglm * [shardformer] update gpt2 * [shardformer] update gptj * [shardformer] update opt * [shardformer] update vit * [shardformer] update colo attention mask prep * [shardformer] update whisper * [test] fix shardformer tests (#5514) * [test] fix shardformer tests * [test] fix shardformer tests * [format] applied code formatting on changed files in pull request 5510 (#5517) Co-authored-by: github-actions * [shardformer] fix pipeline forward error if custom layer distribution is used (#5189) * Use self.[distribute_layers|get_stage_index] to exploit custom layer distribution * Change static methods for t5 layer distribution to member functions * Change static methods for whisper layer distribution to member functions * Replace whisper policy usage with self one * Fix test case to use non-static layer distribution methods * fix: fix typo --------- Co-authored-by: Wenhao Chen * [Fix] Grok-1 use tokenizer from the same pretrained path (#5532) * [fix] use tokenizer from the same pretrained path * trust remote code * [ColossalChat] Update RLHF V2 (#5286) * Add dpo. Fix sft, ppo, lora. Refactor all * fix and tested ppo * 2 nd round refactor * add ci tests * fix ci * fix ci * fix readme, style * fix readme style * fix style, fix benchmark * reproduce benchmark result, remove useless files * rename to ColossalChat * use new image * fix ci workflow * fix ci * use local model/tokenizer for ci tests * fix ci * fix ci * fix ci * fix ci timeout * fix rm progress bar. fix ci timeout * fix ci * fix ci typo * remove 3d plugin from ci temporary * test environment * cannot save optimizer * support chat template * fix readme * fix path * test ci locally * restore build_or_pr * fix ci data path * fix benchmark * fix ci, move ci tests to 3080, disable fast tokenizer * move ci to 85 * support flash attention 2 * add all-in-one data preparation script. Fix colossal-llama2-chat chat template * add hardware requirements * move ci test data * fix save_model, add unwrap * fix missing bos * fix missing bos; support grad accumulation with gemini * fix ci * fix ci * fix ci * fix llama2 chat template config * debug sft * debug sft * fix colossalai version requirement * fix ci * add sanity check to prevent NaN loss * fix requirements * add dummy data generation script * add dummy data generation script * add dummy data generation script * add dummy data generation script * update readme * update readme * update readme and ignore * fix logger bug * support parallel_output * modify data preparation logic * fix tokenization * update lr * fix inference * run pre-commit --------- Co-authored-by: Tong Li * [shardformer, pipeline] add `gradient_checkpointing_ratio` and heterogenous shard policy for llama (#5508) * feat: add `GradientCheckpointConfig` and `PipelineGradientCheckpointConfig` * feat: apply `GradientCheckpointConfig` to policy and llama_forward * feat: move `distribute_layer` and `get_stage_index` to PipelineStageManager * fix: add optional args for `distribute_layer` and `get_stage_index` * fix: fix changed API calls * test: update llama tests * style: polish `GradientCheckpointConfig` * fix: fix pipeline utils tests * fix incorrect sharding without zero (#5545) Co-authored-by: Edenzzzz * [shardformer] Sequence Parallelism Optimization (#5533) * sequence parallel optimization * validate sequence parallel in llama (code to be polished) * shardformer api writing * integrate sequence parallel in ShardFormer * fix pp bugs and sp bugs for LlaMa model * integrating ring-based sequence parallelism into ShardFormer * [sequence parallelism]: Add fused megatron function * integrating ring-based sequence parallelism into ShardFormer --------- Co-authored-by: linsj20 * fix bugs when useing sp and flashattention together * fix operation function name * support flash attention for ulysses-style sp * clarify sp process group * fix compatibility bugs in moe plugin * fix fused linear bugs * fix linear layer test * support gpt model all-to-all sp * modify shard data dimension (meant to be dim=-1) * support megtron-style sp and distributed attn for llama model * [shardformer] add megatron sp to llama * support llama7B 128k with distributed attention * [shardformer] robustness enhancement * add block attn * sp mode 1: keep input as a complete sequence * fix sp compatability * finish sp mode 3 support for gpt * using all_to_all_single when batch size is 1 * support mode 2 sp in gpt2 (#5) * [shardformer] add megatron sp to llama * support llama7B 128k with distributed attention * [shardformer] robustness enhancement * add block attn * sp mode 1: keep input as a complete sequence * fix sp compatability * refactor ring implementation * support mode 2 sp in gpt2 * polish code * enable distributed attn mask when using sp mode 2 and 3 in llama * automatically enable flash attn when using sp mode 2 and 3 in llama * inplace attn mask * add zero2 support for sequence parallel * polish code * fix bugs * fix gemini checkpoint io * loose tensor checking atol and rtol * add comment * fix llama layernorm grad * fix zero grad * fix zero grad * fix conflict * update split and gather auto grad func * sequence parallel: inside text split (#6) * polish code (part 1) * polish code (part 2) * polish code (part 2.5) * polish code (part 3) * sequence parallel: inside text split * miscellaneous minor fixes * polish code * fix ulysses style ZeRO * sequence parallel: inside text split * miscellaneous minor fixes * disaggregate sp group and dp group for sp * fix llama and gpt sp * polish code * move ulysses grad sync to ddp (#9) * remove zero_stage and unbind the grad sync for alltoall sp * add 2d group creation test * move ulysses grad sync to ddp * add 2d group creation test * remove useless code * change shard config not to enable sp when enable_all_optimizations * add sp warnings for several model * remove useless code --------- Co-authored-by: linsj20 * [hotfix] quick fixes to make legacy tutorials runnable (#5559) Co-authored-by: Edenzzzz * [fix] fix typo s/muiti-node /multi-node etc. (#5448) * [hotfix] fix typo s/get_defualt_parser /get_default_parser (#5548) * [devops] remove post commit ci (#5566) * [devops] remove post commit ci * [misc] run pre-commit on all files * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci --------- Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com> --------- Co-authored-by: binmakeswell Co-authored-by: Yuanheng Zhao <54058983+yuanheng-zhao@users.noreply.github.com> Co-authored-by: Wenhao Chen Co-authored-by: Hongxin Liu Co-authored-by: Rocky Duan Co-authored-by: Edenzzzz Co-authored-by: Edenzzzz Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: github-actions Co-authored-by: Insu Jang Co-authored-by: YeAnbang <44796419+YeAnbang@users.noreply.github.com> Co-authored-by: Tong Li Co-authored-by: Zhongkai Zhao Co-authored-by: linsj20 Co-authored-by: digger yu Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com> --- .github/pull_request_template.md | 1 + .github/workflows/build_on_pr.yml | 2 +- .github/workflows/build_on_schedule.yml | 2 +- .../compatiblity_test_on_dispatch.yml | 2 +- .github/workflows/compatiblity_test_on_pr.yml | 2 +- .../compatiblity_test_on_schedule.yml | 2 +- .../workflows/example_check_on_dispatch.yml | 4 +- .github/workflows/example_check_on_pr.yml | 4 +- .../workflows/example_check_on_schedule.yml | 3 +- .github/workflows/post_commit.yml | 97 -- .github/workflows/run_chatgpt_examples.yml | 29 +- .github/workflows/run_chatgpt_unit_tests.yml | 10 +- .gitignore | 4 + README.md | 14 + applications/Chat/benchmarks/README.md | 38 - .../benchmarks/benchmark_opt_lora_dummy.py | 210 --- applications/Chat/coati/dataset/__init__.py | 13 - .../Chat/coati/dataset/conversation.py | 89 - .../Chat/coati/dataset/prompt_dataset.py | 45 - .../Chat/coati/dataset/reward_dataset.py | 88 - .../Chat/coati/dataset/sft_dataset.py | 210 --- applications/Chat/coati/dataset/utils.py | 22 - .../Chat/coati/experience_maker/naive.py | 71 - applications/Chat/coati/kernels/__init__.py | 6 - applications/Chat/coati/kernels/opt_attn.py | 90 - applications/Chat/coati/kernels/wrapper.py | 18 - applications/Chat/coati/models/__init__.py | 15 - .../Chat/coati/models/base/__init__.py | 27 - applications/Chat/coati/models/base/actor.py | 32 - applications/Chat/coati/models/base/critic.py | 34 - .../Chat/coati/models/base/reward_model.py | 46 - .../Chat/coati/models/bloom/__init__.py | 5 - .../Chat/coati/models/bloom/bloom_actor.py | 36 - .../Chat/coati/models/bloom/bloom_critic.py | 36 - .../Chat/coati/models/bloom/bloom_rm.py | 36 - .../Chat/coati/models/chatglm/__init__.py | 3 - .../coati/models/chatglm/chatglm_actor.py | 31 - .../coati/models/chatglm/chatglm_tokenizer.py | 442 ----- .../models/chatglm/configuration_chatglm.py | 101 -- .../coati/models/chatglm/modeling_chatglm.py | 1477 ----------------- applications/Chat/coati/models/generation.py | 152 -- .../Chat/coati/models/gpt/__init__.py | 5 - .../Chat/coati/models/gpt/gpt_actor.py | 38 - .../Chat/coati/models/gpt/gpt_critic.py | 37 - applications/Chat/coati/models/gpt/gpt_rm.py | 37 - .../Chat/coati/models/llama/__init__.py | 5 - .../Chat/coati/models/llama/llama_actor.py | 38 - .../Chat/coati/models/llama/llama_critic.py | 36 - .../Chat/coati/models/llama/llama_rm.py | 37 - applications/Chat/coati/models/loss.py | 97 -- .../Chat/coati/models/opt/__init__.py | 5 - .../Chat/coati/models/opt/opt_actor.py | 37 - .../Chat/coati/models/opt/opt_critic.py | 37 - applications/Chat/coati/models/opt/opt_rm.py | 36 - applications/Chat/coati/models/utils.py | 69 - applications/Chat/coati/trainer/__init__.py | 6 - .../Chat/coati/trainer/callbacks/__init__.py | 5 - .../trainer/callbacks/save_checkpoint.py | 76 - applications/Chat/coati/trainer/ppo.py | 202 --- applications/Chat/coati/trainer/rm.py | 123 -- applications/Chat/coati/trainer/sft.py | 130 -- .../Chat/coati/trainer/strategies/__init__.py | 5 - .../Chat/coati/trainer/strategies/base.py | 137 -- .../coati/trainer/strategies/colossalai.py | 209 --- .../Chat/coati/trainer/strategies/ddp.py | 137 -- .../Chat/coati/trainer/strategies/sampler.py | 31 - applications/Chat/coati/trainer/utils.py | 50 - applications/Chat/examples/README.md | 409 ----- applications/Chat/examples/download_model.py | 79 - .../examples/generate_conversation_dataset.py | 82 - .../Chat/examples/generate_prompt_dataset.py | 27 - applications/Chat/examples/inference.py | 73 - applications/Chat/examples/train_prompts.py | 249 --- applications/Chat/examples/train_prompts.sh | 25 - .../Chat/examples/train_reward_model.py | 208 --- applications/Chat/examples/train_rm.sh | 25 - applications/Chat/examples/train_sft.py | 221 --- applications/Chat/examples/train_sft.sh | 28 - applications/Chat/inference/benchmark.py | 141 -- .../Chat/inference/tests/test_chat_prompt.py | 61 - applications/Chat/inference/utils.py | 209 --- applications/Chat/requirements-test.txt | 2 - applications/Chat/requirements.txt | 14 - applications/Chat/tests/test_benchmarks.sh | 33 - applications/Chat/tests/test_checkpoint.py | 91 - applications/Chat/tests/test_dataset.py | 241 --- applications/Chat/tests/test_experience.py | 130 -- applications/Chat/tests/test_inference.sh | 11 - applications/Chat/tests/test_models.py | 245 --- applications/Chat/tests/test_train.sh | 233 --- .../{Chat => ColossalChat}/.gitignore | 13 +- applications/{Chat => ColossalChat}/LICENSE | 0 applications/{Chat => ColossalChat}/README.md | 352 ++-- applications/ColossalChat/benchmarks/Opt.json | 17 + .../ColossalChat/benchmarks/README.md | 37 + .../benchmark_memory_consumption.txt | 4 + .../benchmark_performance_summarization.txt | 16 + .../ColossalChat/benchmarks/benchmark_ppo.py | 523 ++++++ .../ColossalChat/benchmarks/benchmark_ppo.sh | 119 ++ .../benchmarks/data_preparation.sh | 16 + .../benchmarks/ray/1mmt_dummy.py | 0 .../benchmarks/ray/mmmt_dummy.py | 0 .../{Chat => ColossalChat}/coati/__init__.py | 0 .../ColossalChat/coati/dataset/__init__.py | 26 + .../coati/dataset/conversation.py | 143 ++ .../ColossalChat/coati/dataset/loader.py | 383 +++++ .../coati/dataset/tokenization_utils.py | 383 +++++ .../ColossalChat/coati/dataset/utils.py | 138 ++ .../coati/experience_buffer/__init__.py | 0 .../coati/experience_buffer/base.py | 0 .../coati/experience_buffer/naive.py | 13 +- .../coati/experience_buffer/utils.py | 5 +- .../coati/experience_maker/__init__.py | 0 .../coati/experience_maker/base.py | 26 +- .../coati/experience_maker/naive.py | 180 ++ .../ColossalChat/coati/models/__init__.py | 24 + .../ColossalChat/coati/models/base.py | 58 + .../ColossalChat/coati/models/critic.py | 34 + .../ColossalChat/coati/models/generation.py | 428 +++++ .../coati/models/lora.py | 52 +- .../ColossalChat/coati/models/loss.py | 169 ++ .../ColossalChat/coati/models/reward_model.py | 38 + .../ColossalChat/coati/models/utils.py | 137 ++ .../coati/quant/__init__.py | 0 .../coati/quant/llama_gptq/__init__.py | 0 .../coati/quant/llama_gptq/loader.py | 0 .../coati/quant/llama_gptq/model_utils.py | 0 .../coati/quant/llama_gptq/quant.py | 0 .../coati/quant/utils.py | 0 .../coati/ray/README.md | 0 .../coati/ray/__init__.py | 0 .../coati/ray/callbacks/__init__.py | 0 .../coati/ray/callbacks/base.py | 0 .../ray/callbacks/performance_evaluator.py | 0 .../coati/ray/detached_replay_buffer.py | 0 .../coati/ray/detached_trainer_base.py | 0 .../coati/ray/detached_trainer_ppo.py | 0 .../coati/ray/experience_maker_holder.py | 0 .../coati/ray/lora_constructor.py | 0 .../{Chat => ColossalChat}/coati/ray/utils.py | 0 .../ColossalChat/coati/trainer/__init__.py | 7 + .../coati/trainer/base.py | 56 +- .../coati/trainer/callbacks/__init__.py | 4 + .../coati/trainer/callbacks/base.py | 0 .../callbacks/performance_evaluator.py | 16 +- .../ColossalChat/coati/trainer/dpo.py | 336 ++++ .../ColossalChat/coati/trainer/ppo.py | 403 +++++ applications/ColossalChat/coati/trainer/rm.py | 242 +++ .../ColossalChat/coati/trainer/sft.py | 170 ++ .../ColossalChat/coati/trainer/utils.py | 113 ++ .../ColossalChat/coati/utils/__init__.py | 4 + .../coati/utils/accumulative_meter.py | 69 + .../ColossalChat/coati/utils/ckpt_io.py | 93 ++ .../config/conversation_template/Qwen.json | 7 + .../config/conversation_template/Vicuna.json | 7 + .../config/conversation_template/Yi.json | 7 + .../conversation_template/chatGLM2.json | 7 + .../colossal-llama2.json | 7 + .../config/conversation_template/llama2.json | 7 + .../config/conversation_template/mistral.json | 7 + .../config/conversation_template/zephyr.json | 7 + applications/ColossalChat/examples/README.md | 565 +++++++ .../examples/community/README.md | 0 .../examples/community/peft/README.md | 0 .../examples/community/peft/easy_dataset.py | 0 .../examples/community/peft/easy_models.py | 0 .../community/peft/train_peft_prompts.py | 0 .../examples/community/peft/train_peft_sft.py | 0 .../examples/community/ray/README.md | 0 .../examples/community/ray/ray_job_script.py | 0 .../community/ray/train_prompts_on_ray.py | 0 .../prepare_dataset.py | 268 +++ .../prepare_preference_dataset.sh | 13 + .../prepare_prompt_dataset.sh | 13 + .../prepare_sft_dataset.sh | 13 + .../ColossalChat/examples/inference/chatio.py | 168 ++ .../examples/inference/inference.py | 195 +++ .../examples/inference/web_chatbot}/README.md | 0 .../inference/web_chatbot}/locustfile.py | 0 .../inference/web_chatbot}/requirements.txt | 0 .../examples/inference/web_chatbot}/server.py | 46 +- .../examples/inference/web_chatbot/utils.py | 78 + .../examples/ray/1mmt_prompt.py | 0 .../examples/ray/mmmt_prompt.py | 0 .../examples/ray/requirements.txt | 0 .../examples/ray/test_ci.sh | 0 .../examples/requirements.txt | 3 +- .../examples/training_scripts/hostfile | 1 + .../examples/training_scripts/train_dpo.py | 326 ++++ .../examples/training_scripts/train_dpo.sh | 62 + .../examples/training_scripts/train_ppo.py | 506 ++++++ .../examples/training_scripts/train_ppo.sh | 82 + .../examples/training_scripts/train_rm.py | 342 ++++ .../examples/training_scripts/train_rm.sh | 61 + .../examples/training_scripts/train_sft.py | 311 ++++ .../examples/training_scripts/train_sft.sh | 59 + .../{Chat => ColossalChat}/pytest.ini | 0 applications/ColossalChat/requirements.txt | 24 + applications/{Chat => ColossalChat}/setup.py | 2 +- .../{Chat => ColossalChat}/tests/__init__.py | 0 .../generate_dummy_datasets_for_testing.py | 72 + applications/ColossalChat/tests/llama.json | 8 + applications/ColossalChat/tests/opt.json | 17 + .../tests/test_data/dpo/test_dpo_data.jsonl | 1 + .../tests/test_data/sft/test_sft_data.jsonl | 1 + .../tests/test_data_preparation.sh | 260 +++ applications/ColossalChat/tests/test_lora.py | 69 + .../ColossalChat/tests/test_templating.sh | 97 ++ applications/ColossalChat/tests/test_train.sh | 397 +++++ .../ColossalChat/tests/verify_chat_data.py | 64 + .../{Chat => ColossalChat}/version.txt | 0 .../colossal_eval/evaluate/gpt_evaluate.py | 2 +- .../colossal_moe/models/mixtral_policy.py | 8 +- applications/ColossalMoE/train.py | 7 +- .../colossalqa/data_loader/document_loader.py | 2 +- .../colossalqa/local/colossalcloud_llm.py | 2 +- .../ColossalQA/colossalqa/local/llm.py | 6 +- .../retrieval_conversation_universal.py | 2 +- .../ColossalQA/colossalqa/retriever.py | 2 +- applications/ColossalQA/colossalqa/utils.py | 2 +- colossalai/_analyzer/fx/tracer/tracer.py | 2 +- .../booster/plugin/hybrid_parallel_plugin.py | 97 +- .../plugin/moe_hybrid_parallel_plugin.py | 4 + colossalai/cluster/process_group_mesh.py | 47 +- colossalai/inference/engine/policies/bloom.py | 4 +- .../inference/engine/policies/chatglm2.py | 4 +- colossalai/inference/engine/policies/llama.py | 4 +- colossalai/kernel/kernel_loader.py | 24 +- colossalai/legacy/tensor/tensor_spec.py | 4 +- colossalai/nn/layer/colo_attention.py | 209 --- colossalai/nn/layer/scaled_softmax.py | 8 + colossalai/pipeline/schedule/one_f_one_b.py | 10 +- colossalai/pipeline/stage_manager.py | 87 +- colossalai/shardformer/__init__.py | 2 +- colossalai/shardformer/layer/__init__.py | 5 + colossalai/shardformer/layer/_operation.py | 389 ++++- colossalai/shardformer/layer/attn.py | 269 +++ colossalai/shardformer/layer/linear.py | 47 +- .../shardformer/layer/qkv_fused_linear.py | 44 +- colossalai/shardformer/layer/utils.py | 26 +- colossalai/shardformer/modeling/bert.py | 20 +- colossalai/shardformer/modeling/blip2.py | 37 +- colossalai/shardformer/modeling/bloom.py | 18 +- colossalai/shardformer/modeling/chatglm2.py | 133 +- colossalai/shardformer/modeling/gpt2.py | 474 ++++-- colossalai/shardformer/modeling/gptj.py | 361 +++- colossalai/shardformer/modeling/llama.py | 534 +++++- colossalai/shardformer/modeling/opt.py | 335 +++- colossalai/shardformer/modeling/vit.py | 35 +- colossalai/shardformer/modeling/whisper.py | 300 +++- .../shardformer/policies/base_policy.py | 50 +- colossalai/shardformer/policies/bert.py | 103 +- colossalai/shardformer/policies/bloom.py | 35 +- colossalai/shardformer/policies/chatglm2.py | 30 +- colossalai/shardformer/policies/falcon.py | 8 +- colossalai/shardformer/policies/gpt2.py | 130 +- colossalai/shardformer/policies/gptj.py | 59 +- colossalai/shardformer/policies/llama.py | 143 +- colossalai/shardformer/policies/opt.py | 66 +- colossalai/shardformer/policies/t5.py | 41 +- colossalai/shardformer/policies/vit.py | 8 +- colossalai/shardformer/policies/whisper.py | 59 +- colossalai/shardformer/shard/__init__.py | 3 +- .../shardformer/shard/grad_ckpt_config.py | 87 + colossalai/shardformer/shard/shard_config.py | 69 +- .../tensor/d_tensor/layout_converter.py | 5 +- colossalai/testing/comparison.py | 30 +- colossalai/zero/low_level/low_level_optim.py | 2 +- docs/README-zh-Hans.md | 14 + .../train_gpt_using_hybrid_parallelism.md | 2 +- .../train_vit_with_hybrid_parallelism.md | 2 +- docs/source/en/features/pipeline_parallel.md | 3 +- docs/source/en/features/shardformer.md | 2 +- .../train_gpt_using_hybrid_parallelism.md | 2 +- .../train_vit_with_hybrid_parallelism.md | 2 +- .../zh-Hans/features/pipeline_parallel.md | 3 +- docs/source/zh-Hans/features/shardformer.md | 2 +- examples/images/vit/vit_benchmark.py | 4 +- examples/language/bert/finetune.py | 2 +- .../gpt/hybridparallelism/finetune.py | 2 +- examples/language/grok-1/README.md | 52 +- examples/language/grok-1/inference.py | 29 +- examples/language/grok-1/inference_tp.py | 29 +- examples/language/grok-1/requirements.txt | 1 - .../language/grok-1/run_inference_fast.sh | 6 +- .../language/grok-1/run_inference_slow.sh | 6 +- examples/language/grok-1/utils.py | 8 +- examples/language/llama2/finetune.py | 4 +- examples/language/llama2/pretrain.py | 6 +- .../openmoe/benchmark/benchmark_cai.py | 1 - .../openmoe/model/modeling_openmoe.py | 2 +- .../language/openmoe/model/openmoe_policy.py | 13 +- examples/language/openmoe/train.py | 7 +- examples/language/opt/opt_train_demo.py | 4 +- .../auto_parallel_with_resnet.py | 2 +- examples/tutorial/hybrid_parallel/train.py | 4 +- .../tutorial/large_batch_optimizer/train.py | 6 +- examples/tutorial/opt/opt/run_clm.py | 4 +- .../model/layers/bert_layer.py | 2 +- examples/tutorial/sequence_parallel/train.py | 4 +- extensions/README.md | 4 +- extensions/__init__.py | 10 +- extensions/base_extension.py | 4 +- extensions/cpu_adam/cpu_adam_arm.py | 4 +- extensions/cpu_adam/cpu_adam_x86.py | 8 +- extensions/cuda_extension.py | 4 +- extensions/flash_attention/__init__.py | 12 +- .../flash_attention_dao_cuda.py | 99 +- .../flash_attention/flash_attention_npu.py | 61 +- .../flash_attention_sdpa_cuda.py | 56 + .../flash_attention_xformers_cuda.py | 94 -- setup.py | 4 +- tests/kit/model_zoo/transformers/gpt.py | 46 +- tests/kit/model_zoo/transformers/llama.py | 20 +- .../test_plugin/test_3d_plugin.py | 8 +- .../test_gemini_checkpoint_io.py | 17 +- ...st_hybrid_parallel_plugin_checkpoint_io.py | 9 +- tests/test_cluster/test_process_group_mesh.py | 30 + tests/test_moe/test_moe_checkpoint.py | 1 - tests/test_optimizer/test_nvme.py | 3 + .../test_t5_pipeline_utils.py | 43 +- .../test_whisper_pipeline_utils.py | 41 +- .../test_schedule/test_interleaved.py | 6 +- .../test_schedule/test_oneF_oneB.py | 6 +- .../test_shardformer/test_flash_attention.py | 147 ++ .../test_gpt2_qkv_fused_linear_1d.py | 27 +- .../test_layer/test_linear_1d.py | 42 +- .../test_layer/test_sequence_parallel.py | 178 ++ tests/test_shardformer/test_model/_utils.py | 63 +- .../test_model/test_shard_bert.py | 23 +- .../test_model/test_shard_blip2.py | 51 +- .../test_model/test_shard_bloom.py | 22 + .../test_model/test_shard_chatglm2.py | 91 +- .../test_model/test_shard_gpt2.py | 99 +- .../test_model/test_shard_gptj.py | 77 +- .../test_model/test_shard_llama.py | 123 +- .../test_model/test_shard_opt.py | 90 +- .../test_model/test_shard_t5.py | 56 +- .../test_dtensor/test_layout_converter.py | 11 +- tests/test_utils/test_flash_attention.py | 167 -- 340 files changed, 13661 insertions(+), 9958 deletions(-) delete mode 100644 .github/workflows/post_commit.yml delete mode 100644 applications/Chat/benchmarks/README.md delete mode 100644 applications/Chat/benchmarks/benchmark_opt_lora_dummy.py delete mode 100644 applications/Chat/coati/dataset/__init__.py delete mode 100644 applications/Chat/coati/dataset/conversation.py delete mode 100644 applications/Chat/coati/dataset/prompt_dataset.py delete mode 100644 applications/Chat/coati/dataset/reward_dataset.py delete mode 100644 applications/Chat/coati/dataset/sft_dataset.py delete mode 100644 applications/Chat/coati/dataset/utils.py delete mode 100644 applications/Chat/coati/experience_maker/naive.py delete mode 100644 applications/Chat/coati/kernels/__init__.py delete mode 100644 applications/Chat/coati/kernels/opt_attn.py delete mode 100644 applications/Chat/coati/kernels/wrapper.py delete mode 100644 applications/Chat/coati/models/__init__.py delete mode 100644 applications/Chat/coati/models/base/__init__.py delete mode 100644 applications/Chat/coati/models/base/actor.py delete mode 100644 applications/Chat/coati/models/base/critic.py delete mode 100644 applications/Chat/coati/models/base/reward_model.py delete mode 100644 applications/Chat/coati/models/bloom/__init__.py delete mode 100644 applications/Chat/coati/models/bloom/bloom_actor.py delete mode 100644 applications/Chat/coati/models/bloom/bloom_critic.py delete mode 100644 applications/Chat/coati/models/bloom/bloom_rm.py delete mode 100644 applications/Chat/coati/models/chatglm/__init__.py delete mode 100644 applications/Chat/coati/models/chatglm/chatglm_actor.py delete mode 100644 applications/Chat/coati/models/chatglm/chatglm_tokenizer.py delete mode 100644 applications/Chat/coati/models/chatglm/configuration_chatglm.py delete mode 100644 applications/Chat/coati/models/chatglm/modeling_chatglm.py delete mode 100644 applications/Chat/coati/models/generation.py delete mode 100644 applications/Chat/coati/models/gpt/__init__.py delete mode 100644 applications/Chat/coati/models/gpt/gpt_actor.py delete mode 100644 applications/Chat/coati/models/gpt/gpt_critic.py delete mode 100644 applications/Chat/coati/models/gpt/gpt_rm.py delete mode 100644 applications/Chat/coati/models/llama/__init__.py delete mode 100644 applications/Chat/coati/models/llama/llama_actor.py delete mode 100644 applications/Chat/coati/models/llama/llama_critic.py delete mode 100644 applications/Chat/coati/models/llama/llama_rm.py delete mode 100644 applications/Chat/coati/models/loss.py delete mode 100644 applications/Chat/coati/models/opt/__init__.py delete mode 100644 applications/Chat/coati/models/opt/opt_actor.py delete mode 100644 applications/Chat/coati/models/opt/opt_critic.py delete mode 100644 applications/Chat/coati/models/opt/opt_rm.py delete mode 100644 applications/Chat/coati/models/utils.py delete mode 100644 applications/Chat/coati/trainer/__init__.py delete mode 100644 applications/Chat/coati/trainer/callbacks/__init__.py delete mode 100644 applications/Chat/coati/trainer/callbacks/save_checkpoint.py delete mode 100644 applications/Chat/coati/trainer/ppo.py delete mode 100644 applications/Chat/coati/trainer/rm.py delete mode 100644 applications/Chat/coati/trainer/sft.py delete mode 100644 applications/Chat/coati/trainer/strategies/__init__.py delete mode 100644 applications/Chat/coati/trainer/strategies/base.py delete mode 100644 applications/Chat/coati/trainer/strategies/colossalai.py delete mode 100644 applications/Chat/coati/trainer/strategies/ddp.py delete mode 100644 applications/Chat/coati/trainer/strategies/sampler.py delete mode 100644 applications/Chat/coati/trainer/utils.py delete mode 100644 applications/Chat/examples/README.md delete mode 100644 applications/Chat/examples/download_model.py delete mode 100644 applications/Chat/examples/generate_conversation_dataset.py delete mode 100644 applications/Chat/examples/generate_prompt_dataset.py delete mode 100644 applications/Chat/examples/inference.py delete mode 100644 applications/Chat/examples/train_prompts.py delete mode 100755 applications/Chat/examples/train_prompts.sh delete mode 100644 applications/Chat/examples/train_reward_model.py delete mode 100755 applications/Chat/examples/train_rm.sh delete mode 100644 applications/Chat/examples/train_sft.py delete mode 100755 applications/Chat/examples/train_sft.sh delete mode 100644 applications/Chat/inference/benchmark.py delete mode 100644 applications/Chat/inference/tests/test_chat_prompt.py delete mode 100644 applications/Chat/inference/utils.py delete mode 100644 applications/Chat/requirements-test.txt delete mode 100644 applications/Chat/requirements.txt delete mode 100755 applications/Chat/tests/test_benchmarks.sh delete mode 100644 applications/Chat/tests/test_checkpoint.py delete mode 100644 applications/Chat/tests/test_dataset.py delete mode 100644 applications/Chat/tests/test_experience.py delete mode 100755 applications/Chat/tests/test_inference.sh delete mode 100644 applications/Chat/tests/test_models.py delete mode 100755 applications/Chat/tests/test_train.sh rename applications/{Chat => ColossalChat}/.gitignore (87%) mode change 100644 => 100755 rename applications/{Chat => ColossalChat}/LICENSE (100%) mode change 100644 => 100755 rename applications/{Chat => ColossalChat}/README.md (59%) mode change 100644 => 100755 create mode 100644 applications/ColossalChat/benchmarks/Opt.json create mode 100755 applications/ColossalChat/benchmarks/README.md create mode 100644 applications/ColossalChat/benchmarks/benchmark_memory_consumption.txt create mode 100644 applications/ColossalChat/benchmarks/benchmark_performance_summarization.txt create mode 100644 applications/ColossalChat/benchmarks/benchmark_ppo.py create mode 100755 applications/ColossalChat/benchmarks/benchmark_ppo.sh create mode 100755 applications/ColossalChat/benchmarks/data_preparation.sh rename applications/{Chat => ColossalChat}/benchmarks/ray/1mmt_dummy.py (100%) mode change 100644 => 100755 rename applications/{Chat => ColossalChat}/benchmarks/ray/mmmt_dummy.py (100%) mode change 100644 => 100755 rename applications/{Chat => ColossalChat}/coati/__init__.py (100%) mode change 100644 => 100755 create mode 100755 applications/ColossalChat/coati/dataset/__init__.py create mode 100755 applications/ColossalChat/coati/dataset/conversation.py create mode 100755 applications/ColossalChat/coati/dataset/loader.py create mode 100755 applications/ColossalChat/coati/dataset/tokenization_utils.py create mode 100755 applications/ColossalChat/coati/dataset/utils.py rename applications/{Chat => ColossalChat}/coati/experience_buffer/__init__.py (100%) mode change 100644 => 100755 rename applications/{Chat => ColossalChat}/coati/experience_buffer/base.py (100%) mode change 100644 => 100755 rename applications/{Chat => ColossalChat}/coati/experience_buffer/naive.py (86%) mode change 100644 => 100755 rename applications/{Chat => ColossalChat}/coati/experience_buffer/utils.py (94%) mode change 100644 => 100755 rename applications/{Chat => ColossalChat}/coati/experience_maker/__init__.py (100%) mode change 100644 => 100755 rename applications/{Chat => ColossalChat}/coati/experience_maker/base.py (74%) mode change 100644 => 100755 create mode 100755 applications/ColossalChat/coati/experience_maker/naive.py create mode 100755 applications/ColossalChat/coati/models/__init__.py create mode 100755 applications/ColossalChat/coati/models/base.py create mode 100755 applications/ColossalChat/coati/models/critic.py create mode 100755 applications/ColossalChat/coati/models/generation.py rename applications/{Chat => ColossalChat}/coati/models/lora.py (85%) mode change 100644 => 100755 create mode 100755 applications/ColossalChat/coati/models/loss.py create mode 100755 applications/ColossalChat/coati/models/reward_model.py create mode 100755 applications/ColossalChat/coati/models/utils.py rename applications/{Chat => ColossalChat}/coati/quant/__init__.py (100%) mode change 100644 => 100755 rename applications/{Chat => ColossalChat}/coati/quant/llama_gptq/__init__.py (100%) mode change 100644 => 100755 rename applications/{Chat => ColossalChat}/coati/quant/llama_gptq/loader.py (100%) mode change 100644 => 100755 rename applications/{Chat => ColossalChat}/coati/quant/llama_gptq/model_utils.py (100%) mode change 100644 => 100755 rename applications/{Chat => ColossalChat}/coati/quant/llama_gptq/quant.py (100%) mode change 100644 => 100755 rename applications/{Chat => ColossalChat}/coati/quant/utils.py (100%) mode change 100644 => 100755 rename applications/{Chat => ColossalChat}/coati/ray/README.md (100%) mode change 100644 => 100755 rename applications/{Chat => ColossalChat}/coati/ray/__init__.py (100%) mode change 100644 => 100755 rename applications/{Chat => ColossalChat}/coati/ray/callbacks/__init__.py (100%) mode change 100644 => 100755 rename applications/{Chat => ColossalChat}/coati/ray/callbacks/base.py (100%) mode change 100644 => 100755 rename applications/{Chat => ColossalChat}/coati/ray/callbacks/performance_evaluator.py (100%) mode change 100644 => 100755 rename applications/{Chat => ColossalChat}/coati/ray/detached_replay_buffer.py (100%) mode change 100644 => 100755 rename applications/{Chat => ColossalChat}/coati/ray/detached_trainer_base.py (100%) mode change 100644 => 100755 rename applications/{Chat => ColossalChat}/coati/ray/detached_trainer_ppo.py (100%) mode change 100644 => 100755 rename applications/{Chat => ColossalChat}/coati/ray/experience_maker_holder.py (100%) mode change 100644 => 100755 rename applications/{Chat => ColossalChat}/coati/ray/lora_constructor.py (100%) mode change 100644 => 100755 rename applications/{Chat => ColossalChat}/coati/ray/utils.py (100%) mode change 100644 => 100755 create mode 100755 applications/ColossalChat/coati/trainer/__init__.py rename applications/{Chat => ColossalChat}/coati/trainer/base.py (80%) mode change 100644 => 100755 create mode 100644 applications/ColossalChat/coati/trainer/callbacks/__init__.py rename applications/{Chat => ColossalChat}/coati/trainer/callbacks/base.py (100%) rename applications/{Chat => ColossalChat}/coati/trainer/callbacks/performance_evaluator.py (92%) create mode 100755 applications/ColossalChat/coati/trainer/dpo.py create mode 100755 applications/ColossalChat/coati/trainer/ppo.py create mode 100755 applications/ColossalChat/coati/trainer/rm.py create mode 100755 applications/ColossalChat/coati/trainer/sft.py create mode 100755 applications/ColossalChat/coati/trainer/utils.py create mode 100755 applications/ColossalChat/coati/utils/__init__.py create mode 100755 applications/ColossalChat/coati/utils/accumulative_meter.py create mode 100755 applications/ColossalChat/coati/utils/ckpt_io.py create mode 100644 applications/ColossalChat/config/conversation_template/Qwen.json create mode 100644 applications/ColossalChat/config/conversation_template/Vicuna.json create mode 100644 applications/ColossalChat/config/conversation_template/Yi.json create mode 100644 applications/ColossalChat/config/conversation_template/chatGLM2.json create mode 100644 applications/ColossalChat/config/conversation_template/colossal-llama2.json create mode 100644 applications/ColossalChat/config/conversation_template/llama2.json create mode 100644 applications/ColossalChat/config/conversation_template/mistral.json create mode 100644 applications/ColossalChat/config/conversation_template/zephyr.json create mode 100755 applications/ColossalChat/examples/README.md rename applications/{Chat => ColossalChat}/examples/community/README.md (100%) mode change 100644 => 100755 rename applications/{Chat => ColossalChat}/examples/community/peft/README.md (100%) mode change 100644 => 100755 rename applications/{Chat => ColossalChat}/examples/community/peft/easy_dataset.py (100%) mode change 100644 => 100755 rename applications/{Chat => ColossalChat}/examples/community/peft/easy_models.py (100%) mode change 100644 => 100755 rename applications/{Chat => ColossalChat}/examples/community/peft/train_peft_prompts.py (100%) mode change 100644 => 100755 rename applications/{Chat => ColossalChat}/examples/community/peft/train_peft_sft.py (100%) mode change 100644 => 100755 rename applications/{Chat => ColossalChat}/examples/community/ray/README.md (100%) mode change 100644 => 100755 rename applications/{Chat => ColossalChat}/examples/community/ray/ray_job_script.py (100%) mode change 100644 => 100755 rename applications/{Chat => ColossalChat}/examples/community/ray/train_prompts_on_ray.py (100%) mode change 100644 => 100755 create mode 100644 applications/ColossalChat/examples/data_preparation_scripts/prepare_dataset.py create mode 100755 applications/ColossalChat/examples/data_preparation_scripts/prepare_preference_dataset.sh create mode 100755 applications/ColossalChat/examples/data_preparation_scripts/prepare_prompt_dataset.sh create mode 100755 applications/ColossalChat/examples/data_preparation_scripts/prepare_sft_dataset.sh create mode 100755 applications/ColossalChat/examples/inference/chatio.py create mode 100755 applications/ColossalChat/examples/inference/inference.py rename applications/{Chat/inference => ColossalChat/examples/inference/web_chatbot}/README.md (100%) mode change 100644 => 100755 rename applications/{Chat/inference => ColossalChat/examples/inference/web_chatbot}/locustfile.py (100%) mode change 100644 => 100755 rename applications/{Chat/inference => ColossalChat/examples/inference/web_chatbot}/requirements.txt (100%) mode change 100644 => 100755 rename applications/{Chat/inference => ColossalChat/examples/inference/web_chatbot}/server.py (79%) mode change 100644 => 100755 create mode 100755 applications/ColossalChat/examples/inference/web_chatbot/utils.py rename applications/{Chat => ColossalChat}/examples/ray/1mmt_prompt.py (100%) mode change 100644 => 100755 rename applications/{Chat => ColossalChat}/examples/ray/mmmt_prompt.py (100%) mode change 100644 => 100755 rename applications/{Chat => ColossalChat}/examples/ray/requirements.txt (100%) mode change 100644 => 100755 rename applications/{Chat => ColossalChat}/examples/ray/test_ci.sh (100%) rename applications/{Chat => ColossalChat}/examples/requirements.txt (51%) create mode 100755 applications/ColossalChat/examples/training_scripts/hostfile create mode 100755 applications/ColossalChat/examples/training_scripts/train_dpo.py create mode 100755 applications/ColossalChat/examples/training_scripts/train_dpo.sh create mode 100755 applications/ColossalChat/examples/training_scripts/train_ppo.py create mode 100755 applications/ColossalChat/examples/training_scripts/train_ppo.sh create mode 100755 applications/ColossalChat/examples/training_scripts/train_rm.py create mode 100755 applications/ColossalChat/examples/training_scripts/train_rm.sh create mode 100755 applications/ColossalChat/examples/training_scripts/train_sft.py create mode 100755 applications/ColossalChat/examples/training_scripts/train_sft.sh rename applications/{Chat => ColossalChat}/pytest.ini (100%) mode change 100644 => 100755 create mode 100755 applications/ColossalChat/requirements.txt rename applications/{Chat => ColossalChat}/setup.py (97%) mode change 100644 => 100755 rename applications/{Chat => ColossalChat}/tests/__init__.py (100%) mode change 100644 => 100755 create mode 100644 applications/ColossalChat/tests/generate_dummy_datasets_for_testing.py create mode 100644 applications/ColossalChat/tests/llama.json create mode 100644 applications/ColossalChat/tests/opt.json create mode 100644 applications/ColossalChat/tests/test_data/dpo/test_dpo_data.jsonl create mode 100644 applications/ColossalChat/tests/test_data/sft/test_sft_data.jsonl create mode 100755 applications/ColossalChat/tests/test_data_preparation.sh create mode 100755 applications/ColossalChat/tests/test_lora.py create mode 100755 applications/ColossalChat/tests/test_templating.sh create mode 100755 applications/ColossalChat/tests/test_train.sh create mode 100644 applications/ColossalChat/tests/verify_chat_data.py rename applications/{Chat => ColossalChat}/version.txt (100%) mode change 100644 => 100755 delete mode 100644 colossalai/nn/layer/colo_attention.py create mode 100644 colossalai/shardformer/layer/attn.py create mode 100644 colossalai/shardformer/shard/grad_ckpt_config.py mode change 100755 => 100644 examples/tutorial/opt/opt/run_clm.py create mode 100644 extensions/flash_attention/flash_attention_sdpa_cuda.py delete mode 100644 extensions/flash_attention/flash_attention_xformers_cuda.py create mode 100644 tests/test_shardformer/test_flash_attention.py create mode 100644 tests/test_shardformer/test_layer/test_sequence_parallel.py delete mode 100644 tests/test_utils/test_flash_attention.py diff --git a/.github/pull_request_template.md b/.github/pull_request_template.md index 0dc1262b008c..d67cd9fcc9ac 100644 --- a/.github/pull_request_template.md +++ b/.github/pull_request_template.md @@ -3,6 +3,7 @@ - [ ] I have created an issue for this PR for traceability - [ ] The title follows the standard format: `[doc/gemini/tensor/...]: A concise description` - [ ] I have added relevant tags if possible for us to better distinguish different PRs +- [ ] I have installed pre-commit: `pip install pre-commit && pre-commit install` ## 🚨 Issue number diff --git a/.github/workflows/build_on_pr.yml b/.github/workflows/build_on_pr.yml index 2cad504f3391..5bdadca783b3 100644 --- a/.github/workflows/build_on_pr.yml +++ b/.github/workflows/build_on_pr.yml @@ -117,7 +117,7 @@ jobs: cd TensorNVMe conda install cmake pip install -r requirements.txt - pip install -v . + DISABLE_URING=1 pip install -v . - name: Store TensorNVMe Cache run: | diff --git a/.github/workflows/build_on_schedule.yml b/.github/workflows/build_on_schedule.yml index 3ff19b37b4bf..e560d0c004b1 100644 --- a/.github/workflows/build_on_schedule.yml +++ b/.github/workflows/build_on_schedule.yml @@ -44,7 +44,7 @@ jobs: cd TensorNVMe conda install cmake pip install -r requirements.txt - pip install -v . + DISABLE_URING=1 pip install -v . - uses: actions/checkout@v2 if: steps.check-avai.outputs.avai == 'true' diff --git a/.github/workflows/compatiblity_test_on_dispatch.yml b/.github/workflows/compatiblity_test_on_dispatch.yml index 76493880651c..95a94c27bfd5 100644 --- a/.github/workflows/compatiblity_test_on_dispatch.yml +++ b/.github/workflows/compatiblity_test_on_dispatch.yml @@ -66,7 +66,7 @@ jobs: cd TensorNVMe apt update && apt install -y cmake pip install -r requirements.txt - pip install -v . + DISABLE_URING=1 pip install -v . - uses: actions/checkout@v2 with: ssh-key: ${{ secrets.SSH_KEY_FOR_CI }} diff --git a/.github/workflows/compatiblity_test_on_pr.yml b/.github/workflows/compatiblity_test_on_pr.yml index f582b30907bf..aef4816efcfe 100644 --- a/.github/workflows/compatiblity_test_on_pr.yml +++ b/.github/workflows/compatiblity_test_on_pr.yml @@ -60,7 +60,7 @@ jobs: cd TensorNVMe apt update && apt install -y cmake pip install -r requirements.txt - pip install -v . + DISABLE_URING=1 pip install -v . - uses: actions/checkout@v2 with: ssh-key: ${{ secrets.SSH_KEY_FOR_CI }} diff --git a/.github/workflows/compatiblity_test_on_schedule.yml b/.github/workflows/compatiblity_test_on_schedule.yml index 3348b51ecc6e..3dc8a5a328a6 100644 --- a/.github/workflows/compatiblity_test_on_schedule.yml +++ b/.github/workflows/compatiblity_test_on_schedule.yml @@ -56,7 +56,7 @@ jobs: cd TensorNVMe apt update && apt install -y cmake pip install -r requirements.txt - pip install -v . + DISABLE_URING=1 pip install -v . - uses: actions/checkout@v2 with: ssh-key: ${{ secrets.SSH_KEY_FOR_CI }} diff --git a/.github/workflows/example_check_on_dispatch.yml b/.github/workflows/example_check_on_dispatch.yml index 24e726b4f16d..d877b06cee1c 100644 --- a/.github/workflows/example_check_on_dispatch.yml +++ b/.github/workflows/example_check_on_dispatch.yml @@ -46,7 +46,7 @@ jobs: matrix: ${{fromJson(needs.manual_check_matrix_preparation.outputs.matrix)}} container: image: hpcaitech/pytorch-cuda:2.1.0-12.1.0 - options: --gpus all --rm -v /data/scratch/examples-data:/data/ + options: --gpus all --rm -v /data/scratch/examples-data:/data/ -v /dev/shm timeout-minutes: 15 steps: - name: 📚 Checkout @@ -60,5 +60,3 @@ jobs: echo "Testing ${dir} now" cd "${PWD}/examples/${dir}" bash test_ci.sh - env: - NCCL_SHM_DISABLE: 1 diff --git a/.github/workflows/example_check_on_pr.yml b/.github/workflows/example_check_on_pr.yml index 728f059c1bb3..6170628e10a9 100644 --- a/.github/workflows/example_check_on_pr.yml +++ b/.github/workflows/example_check_on_pr.yml @@ -78,7 +78,7 @@ jobs: matrix: ${{fromJson(needs.detect-changed-example.outputs.matrix)}} container: image: hpcaitech/pytorch-cuda:2.1.0-12.1.0 - options: --gpus all --rm -v /data/scratch/examples-data:/data/ + options: --gpus all --rm -v /data/scratch/examples-data:/data/ -v /dev/shm timeout-minutes: 20 concurrency: group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.ref }}-run-example-${{ matrix.directory }} @@ -95,5 +95,3 @@ jobs: example_dir=${{ matrix.directory }} cd "${PWD}/examples/${example_dir}" bash test_ci.sh - env: - NCCL_SHM_DISABLE: 1 diff --git a/.github/workflows/example_check_on_schedule.yml b/.github/workflows/example_check_on_schedule.yml index efb131a864cb..2588ac8243d9 100644 --- a/.github/workflows/example_check_on_schedule.yml +++ b/.github/workflows/example_check_on_schedule.yml @@ -35,6 +35,7 @@ jobs: matrix: ${{fromJson(needs.matrix_preparation.outputs.matrix)}} container: image: hpcaitech/pytorch-cuda:2.1.0-12.1.0 + options: --gpus all --rm -v /data/scratch/examples-data:/data/ -v /dev/shm timeout-minutes: 10 steps: - name: 📚 Checkout @@ -50,8 +51,6 @@ jobs: echo "Testing ${example_dir} now" cd "${PWD}/examples/${example_dir}" bash test_ci.sh - env: - NCCL_SHM_DISABLE: 1 - name: Notify Lark id: message-preparation diff --git a/.github/workflows/post_commit.yml b/.github/workflows/post_commit.yml deleted file mode 100644 index 1bbc0d2f5c34..000000000000 --- a/.github/workflows/post_commit.yml +++ /dev/null @@ -1,97 +0,0 @@ -name: post-commit - -on: - pull_request: - types: - - closed - -jobs: - # this job will run after a PR is merged to run pre-commit on any changed file - # so that the user does not need to learn pre-commit and pre-commit can still - # be auto-executed by the workflow - pre-commit: - runs-on: ubuntu-latest - if: github.event.pull_request.merged == true && github.repository == 'hpcaitech/ColossalAI' - steps: - - uses: actions/checkout@v2 - with: - fetch-depth: 0 - ref: ${{ github.event.pull_request.head.sha }} - - # the PR branch and the hpcaitech/colossal-ai main branch - # must share a common commit, we need to locate that commit, - # which is the commit checked-out or forked when the PR branch is created - # such that we can look for files changed since that commit - - name: Locate base commit - id: locate-base-sha - run: | - curBranch=$(git rev-parse --abbrev-ref HEAD) - commonCommit=$(git merge-base origin/main $curBranch) - echo $commonCommit - echo "baseSHA=$commonCommit" >> $GITHUB_OUTPUT - - - name: Find the changed files - id: find-changed-files - uses: tj-actions/changed-files@v35 - with: - base_sha: ${{ steps.locate-base-sha.outputs.baseSHA }} - - - name: List all changed files - run: | - for file in ${{ steps.find-changed-files.outputs.all_changed_files }}; do - echo "$file was changed" - done - - # check out the main branch - - uses: actions/checkout@v2 - with: - ref: 'main' - - - uses: actions/setup-python@v3 - - - name: Cache pre-commit hooks - uses: actions/cache@v3 - with: - path: ~/.cache/pre-commit - key: ${{ runner.os }}-pre-commit-hooks - - - name: Set up pre-commit - run: | - pip install pre-commit - pre-commit install - - # run pre-commit on changed files - - name: Run Pre-commit - run: | - for file in ${{ steps.find-changed-files.outputs.all_changed_files }}; do - pre-commit run --files $file || true - done - - # create commit for pre-commit - # when all files are well formatted, there is no need to create a commit - # therefore, this step will produce an error, which should be allowed - - name: Create commits - id: commit - continue-on-error: true - run: | - git config --global user.name 'github-actions' - git config --global user.email 'github-actions@github.com' - git remote set-url origin https://x-access-token:${{ secrets.GITHUB_TOKEN }}@github.com/${{ github.repository }} - git add -A - git commit -am "[format] applied code formatting on changed files in pull request ${{ github.event.pull_request.number }}" - - # create pull request - - name: Create Pull Request - if: steps.commit.outcome == 'success' - id: cpr - uses: peter-evans/create-pull-request@v4 - with: - branch: pre-commit-${{ github.event.pull_request.number }} - title: "[format] applied code formatting on changed files in PR ${{ github.event.pull_request.number }}" - - - name: Enable Auto-merge for the New PR - if: steps.commit.outcome == 'success' - uses: peter-evans/enable-pull-request-automerge@v2 - with: - pull-request-number: ${{ steps.cpr.outputs.pull-request-number }} - merge-method: squash diff --git a/.github/workflows/run_chatgpt_examples.yml b/.github/workflows/run_chatgpt_examples.yml index bb0ceb4a8296..ba997f144cd7 100644 --- a/.github/workflows/run_chatgpt_examples.yml +++ b/.github/workflows/run_chatgpt_examples.yml @@ -19,8 +19,8 @@ jobs: runs-on: [self-hosted, gpu] container: image: hpcaitech/pytorch-cuda:2.1.0-12.1.0 - options: --gpus all --rm -v /data/scratch/github_actions/chat:/data/scratch/github_actions/chat --shm-size=10.24gb - timeout-minutes: 30 + options: --gpus all --rm -v /data/scratch/examples-data:/data/scratch/examples-data --shm-size=10.24gb + timeout-minutes: 60 defaults: run: shell: bash @@ -28,26 +28,35 @@ jobs: - name: Checkout ColossalAI uses: actions/checkout@v2 + - name: Install Colossal-AI + run: | + BUILD_EXT=1 pip install -v -e . + - name: Install ChatGPT run: | - cd applications/Chat + cd applications/ColossalChat pip install -v . + export BUILD_EXT=1 pip install -r examples/requirements.txt - name: Install Transformers run: | - pip install transformers==4.30.2 + pip install transformers==4.34.1 - name: Execute Examples run: | - cd applications/Chat + cd applications/ColossalChat rm -rf ~/.cache/colossalai - ./tests/test_inference.sh - ./tests/test_benchmarks.sh + mkdir models + mkdir sft_data + mkdir prompt_data + mkdir preference_data + ./tests/test_data_preparation.sh ./tests/test_train.sh env: NCCL_SHM_DISABLE: 1 MAX_JOBS: 8 - SFT_DATASET: /data/scratch/github_actions/chat/data.json - PROMPT_DATASET: /data/scratch/github_actions/chat/prompts_en.jsonl - PRETRAIN_DATASET: /data/scratch/github_actions/chat/alpaca_data.json + PRETRAINED_MODEL_PATH: ./models + SFT_DATASET: ./sft_data + PROMPT_DATASET: ./prompt_data + PREFERENCE_DATASET: ./preference_data diff --git a/.github/workflows/run_chatgpt_unit_tests.yml b/.github/workflows/run_chatgpt_unit_tests.yml index 7986889e006b..1d8a53e4feed 100644 --- a/.github/workflows/run_chatgpt_unit_tests.yml +++ b/.github/workflows/run_chatgpt_unit_tests.yml @@ -21,7 +21,7 @@ jobs: runs-on: [self-hosted, gpu] container: image: hpcaitech/pytorch-cuda:2.1.0-12.1.0 - options: --gpus all --rm -v /data/scratch/chatgpt:/data/scratch/chatgpt + options: --gpus all --rm -v /data/scratch/examples-data:/data/scratch/examples-data timeout-minutes: 30 defaults: run: @@ -32,15 +32,17 @@ jobs: - name: Install ChatGPT run: | - cd applications/Chat + cd applications/ColossalChat pip install -v . - pip install -r requirements-test.txt + pip install pytest - name: Execute Unit Testing run: | - cd applications/Chat + cd applications/ColossalChat rm -rf ~/.cache/colossalai pytest tests/ + cd ./tests + ./test_templating.sh env: NCCL_SHM_DISABLE: 1 MAX_JOBS: 8 diff --git a/.gitignore b/.gitignore index 81113fa99dd5..8bc74b4c8c2c 100644 --- a/.gitignore +++ b/.gitignore @@ -159,3 +159,7 @@ coverage.xml # ignore testmon and coverage files .coverage .testmondata* + +# log, test files - ColossalChat +applications/ColossalChat/logs +applications/ColossalChat/tests/logs diff --git a/README.md b/README.md index 7c234b15e75e..26776bdf6d9f 100644 --- a/README.md +++ b/README.md @@ -25,6 +25,7 @@ ## Latest News +* [2024/03] [314 Billion Parameter Grok-1 Inference Accelerated by 3.8x, Efficient and Easy-to-Use PyTorch+HuggingFace version is Here](https://hpc-ai.com/blog/314-billion-parameter-grok-1-inference-accelerated-by-3.8x-efficient-and-easy-to-use-pytorchhuggingface-version-is-here) * [2024/03] [Open-Sora: Revealing Complete Model Parameters, Training Details, and Everything for Sora-like Video Generation Models](https://hpc-ai.com/blog/open-sora-v1.0) * [2024/03] [Open-Sora：Sora Replication Solution with 46% Cost Reduction, Sequence Expansion to Nearly a Million](https://hpc-ai.com/blog/open-sora) * [2024/01] [Inference Performance Improved by 46%, Open Source Solution Breaks the Length Limit of LLM for Multi-Round Conversations](https://hpc-ai.com/blog/Colossal-AI-SwiftInfer) @@ -72,6 +73,7 @@
  • Inference
      +
    • Grok-1: 314B model of PyTorch + HuggingFace Inference
    • SwiftInfer:Breaks the Length Limit of LLM for Multi-Round Conversations with 46% Acceleration
    • GPT-3
    • OPT-175B Online Serving for Text Generation
      • @@ -365,6 +367,18 @@ Please visit our [documentation](https://www.colossalai.org/) and [examples](htt ## Inference +### Grok-1 +

      + +

      + + - 314 Billion Parameter Grok-1 Inference Accelerated by 3.8x, an easy-to-use Python + PyTorch + HuggingFace version for Inference. + +[[code]](https://github.com/hpcaitech/ColossalAI/tree/main/examples/language/grok-1) +[[blog]](https://hpc-ai.com/blog/314-billion-parameter-grok-1-inference-accelerated-by-3.8x-efficient-and-easy-to-use-pytorchhuggingface-version-is-here) +[[HuggingFace Grok-1 PyTorch model weights]](https://huggingface.co/hpcai-tech/grok-1) +[[ModelScope Grok-1 PyTorch model weights]](https://www.modelscope.cn/models/colossalai/grok-1-pytorch/summary) +

      diff --git a/applications/Chat/benchmarks/README.md b/applications/Chat/benchmarks/README.md deleted file mode 100644 index c13f3485863b..000000000000 --- a/applications/Chat/benchmarks/README.md +++ /dev/null @@ -1,38 +0,0 @@ -# Benchmarks - -## Benchmark OPT with LoRA on dummy prompt data - -We provide various OPT models (string in parentheses is the corresponding model name used in this script): - -- OPT-125M (125m) -- OPT-350M (350m) -- OPT-700M (700m) -- OPT-1.3B (1.3b) -- OPT-2.7B (2.7b) -- OPT-3.5B (3.5b) -- OPT-5.5B (5.5b) -- OPT-6.7B (6.7b) -- OPT-10B (10b) -- OPT-13B (13b) - -We also provide various training strategies: - -- ddp: torch DDP -- colossalai_gemini: ColossalAI GeminiDDP with `placement_policy="cuda"`, like zero3 -- colossalai_gemini_cpu: ColossalAI GeminiDDP with `placement_policy="cpu"`, like zero3-offload -- colossalai_zero2: ColossalAI zero2 -- colossalai_zero2_cpu: ColossalAI zero2-offload -- colossalai_zero1: ColossalAI zero1 -- colossalai_zero1_cpu: ColossalAI zero1-offload - -We only support `torchrun` to launch now. E.g. - -```bash -# run OPT-125M with no lora (lora_rank=0) on single-node single-GPU with min batch size -torchrun --standalone --nproc_per_node 1 benchmark_opt_lora_dummy.py \ - --model 125m --critic_model 125m --strategy ddp \ - --experience_batch_size 1 --train_batch_size 1 --lora_rank 0 -# run Actor (OPT-1.3B) and Critic (OPT-350M) with lora_rank=4 on single-node 4-GPU -torchrun --standalone --nproc_per_node 4 benchmark_opt_lora_dummy.py \ - --model 1.3b --critic_model 350m --strategy colossalai_zero2 --lora_rank 4 -``` diff --git a/applications/Chat/benchmarks/benchmark_opt_lora_dummy.py b/applications/Chat/benchmarks/benchmark_opt_lora_dummy.py deleted file mode 100644 index 47ad29f46c64..000000000000 --- a/applications/Chat/benchmarks/benchmark_opt_lora_dummy.py +++ /dev/null @@ -1,210 +0,0 @@ -import argparse -from copy import deepcopy - -import torch -import torch.distributed as dist -import torch.nn as nn -from coati.models.base import RewardModel -from coati.models.opt import OPTActor, OPTCritic -from coati.trainer import PPOTrainer -from coati.trainer.callbacks import PerformanceEvaluator -from coati.trainer.strategies import DDPStrategy, GeminiStrategy, LowLevelZeroStrategy, Strategy -from torch.optim import Adam -from torch.utils.data import DataLoader -from transformers import AutoTokenizer -from transformers.models.opt.configuration_opt import OPTConfig - -from colossalai.nn.optimizer import HybridAdam - - -def get_model_numel(model: nn.Module, strategy: Strategy) -> int: - numel = sum(p.numel() for p in model.parameters()) - if isinstance(strategy, GeminiStrategy) and strategy.shard_init: - numel *= dist.get_world_size() - return numel - - -def preprocess_batch(samples) -> dict: - input_ids = torch.stack(samples) - attention_mask = torch.ones_like(input_ids, dtype=torch.long) - return {"input_ids": input_ids, "attention_mask": attention_mask} - - -def print_rank_0(*args, **kwargs) -> None: - if dist.get_rank() == 0: - print(*args, **kwargs) - - -def print_model_numel(model_dict: dict) -> None: - B = 1024**3 - M = 1024**2 - K = 1024 - outputs = "" - for name, numel in model_dict.items(): - outputs += f"{name}: " - if numel >= B: - outputs += f"{numel / B:.2f} B\n" - elif numel >= M: - outputs += f"{numel / M:.2f} M\n" - elif numel >= K: - outputs += f"{numel / K:.2f} K\n" - else: - outputs += f"{numel}\n" - print_rank_0(outputs) - - -def get_gpt_config(model_name: str) -> OPTConfig: - model_map = { - "125m": OPTConfig.from_pretrained("facebook/opt-125m"), - "350m": OPTConfig(hidden_size=1024, ffn_dim=4096, num_hidden_layers=24, num_attention_heads=16), - "700m": OPTConfig(hidden_size=1280, ffn_dim=5120, num_hidden_layers=36, num_attention_heads=20), - "1.3b": OPTConfig.from_pretrained("facebook/opt-1.3b"), - "2.7b": OPTConfig.from_pretrained("facebook/opt-2.7b"), - "3.5b": OPTConfig(hidden_size=3072, ffn_dim=12288, num_hidden_layers=32, num_attention_heads=32), - "5.5b": OPTConfig(hidden_size=3840, ffn_dim=15360, num_hidden_layers=32, num_attention_heads=32), - "6.7b": OPTConfig.from_pretrained("facebook/opt-6.7b"), - "10b": OPTConfig(hidden_size=5120, ffn_dim=20480, num_hidden_layers=32, num_attention_heads=32), - "13b": OPTConfig.from_pretrained("facebook/opt-13b"), - } - try: - return model_map[model_name] - except KeyError: - raise ValueError(f'Unknown model "{model_name}"') - - -def main(args): - if args.strategy == "ddp": - strategy = DDPStrategy() - elif args.strategy == "colossalai_gemini": - strategy = GeminiStrategy(placement_policy="static", initial_scale=2**5) - elif args.strategy == "colossalai_gemini_cpu": - strategy = GeminiStrategy( - placement_policy="static", offload_optim_frac=1.0, offload_param_frac=1.0, initial_scale=2**5 - ) - elif args.strategy == "colossalai_zero2": - strategy = LowLevelZeroStrategy(stage=2, placement_policy="cuda") - elif args.strategy == "colossalai_zero2_cpu": - strategy = LowLevelZeroStrategy(stage=2, placement_policy="cpu") - elif args.strategy == "colossalai_zero1": - strategy = LowLevelZeroStrategy(stage=1, placement_policy="cuda") - elif args.strategy == "colossalai_zero1_cpu": - strategy = LowLevelZeroStrategy(stage=1, placement_policy="cpu") - else: - raise ValueError(f'Unsupported strategy "{args.strategy}"') - - torch.cuda.set_per_process_memory_fraction(args.cuda_mem_frac) - - model_config = get_gpt_config(args.model) - critic_config = get_gpt_config(args.critic_model) - with strategy.model_init_context(): - actor = OPTActor(config=model_config, lora_rank=args.lora_rank).cuda() - critic = OPTCritic(config=critic_config, lora_rank=args.lora_rank).cuda() - - initial_model = deepcopy(actor).cuda().half() - reward_model = RewardModel(deepcopy(critic.model), deepcopy(critic.value_head)).cuda().half() - - if args.use_kernels: - from coati.kernels import convert_to_xformer_model - - actor, critic, initial_model, reward_model = map( - convert_to_xformer_model, (actor, critic, initial_model, reward_model) - ) - - actor_numel = get_model_numel(actor, strategy) - critic_numel = get_model_numel(critic, strategy) - initial_model_numel = get_model_numel(initial_model, strategy) - reward_model_numel = get_model_numel(reward_model, strategy) - print_model_numel( - { - "Actor": actor_numel, - "Critic": critic_numel, - "Initial model": initial_model_numel, - "Reward model": reward_model_numel, - } - ) - performance_evaluator = PerformanceEvaluator( - actor_numel, - critic_numel, - initial_model_numel, - reward_model_numel, - enable_grad_checkpoint=False, - ignore_episodes=1, - ) - - if args.strategy.startswith("colossalai"): - actor_optim = HybridAdam(actor.parameters(), lr=5e-6) - critic_optim = HybridAdam(critic.parameters(), lr=5e-6) - else: - actor_optim = Adam(actor.parameters(), lr=5e-6) - critic_optim = Adam(critic.parameters(), lr=5e-6) - - tokenizer = AutoTokenizer.from_pretrained("facebook/opt-350m") - tokenizer.pad_token = tokenizer.eos_token - tokenizer.padding_side = "left" - - (actor, actor_optim), (critic, critic_optim) = strategy.prepare((actor, actor_optim), (critic, critic_optim)) - - random_prompts = torch.randint(tokenizer.vocab_size, (1000, 256), device=torch.cuda.current_device()) - dataloader = DataLoader( - random_prompts, batch_size=args.experience_batch_size, shuffle=True, collate_fn=preprocess_batch - ) - - trainer = PPOTrainer( - strategy, - actor, - critic, - reward_model, - initial_model, - actor_optim, - critic_optim, - tokenizer=tokenizer, - ptx_coef=0, - train_batch_size=args.train_batch_size, - offload_inference_models=args.offload_inference_models, - max_length=512, - do_sample=True, - temperature=1.0, - top_k=50, - use_cache=True, - callbacks=[performance_evaluator], - ) - - trainer.fit( - prompt_dataloader=dataloader, - pretrain_dataloader=None, - num_episodes=args.num_episodes, - num_update_steps=args.num_update_steps, - num_collect_steps=args.num_collect_steps, - ) - - print_rank_0(f"Peak CUDA mem: {torch.cuda.max_memory_allocated()/1024**3:.2f} GB") - - -if __name__ == "__main__": - parser = argparse.ArgumentParser() - parser.add_argument("--model", default="125m") - parser.add_argument("--critic_model", default="125m") - parser.add_argument( - "--strategy", - choices=[ - "ddp", - "colossalai_gemini", - "colossalai_gemini_cpu", - "colossalai_zero2", - "colossalai_zero2_cpu", - "colossalai_zero1", - "colossalai_zero1_cpu", - ], - default="ddp", - ) - parser.add_argument("--num_episodes", type=int, default=3) - parser.add_argument("--num_collect_steps", type=int, default=8) - parser.add_argument("--num_update_steps", type=int, default=1) - parser.add_argument("--train_batch_size", type=int, default=8) - parser.add_argument("--experience_batch_size", type=int, default=8) - parser.add_argument("--lora_rank", type=int, default=0) - parser.add_argument("--cuda_mem_frac", type=float, default=1.0) - parser.add_argument("--offload_inference_models", action="store_true", default=False) - parser.add_argument("--use_kernels", action="store_true", default=False) - args = parser.parse_args() - main(args) diff --git a/applications/Chat/coati/dataset/__init__.py b/applications/Chat/coati/dataset/__init__.py deleted file mode 100644 index 599b57609775..000000000000 --- a/applications/Chat/coati/dataset/__init__.py +++ /dev/null @@ -1,13 +0,0 @@ -from .prompt_dataset import PromptDataset -from .reward_dataset import HhRlhfDataset, RmStaticDataset -from .sft_dataset import SFTDataset, SupervisedDataset -from .utils import is_rank_0 - -__all__ = [ - "RmStaticDataset", - "HhRlhfDataset", - "SFTDataset", - "SupervisedDataset", - "PromptDataset", - "is_rank_0", -] diff --git a/applications/Chat/coati/dataset/conversation.py b/applications/Chat/coati/dataset/conversation.py deleted file mode 100644 index f2180d96b0d3..000000000000 --- a/applications/Chat/coati/dataset/conversation.py +++ /dev/null @@ -1,89 +0,0 @@ -# Copyright 2023 lm-sys@FastChat -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -import dataclasses -from enum import Enum, auto -from typing import List - - -class SeparatorStyle(Enum): - ADD_EOS_TOKEN = auto() - - -@dataclasses.dataclass -class Conversation: - system: str - roles: List[str] - messages: List[List[str]] - offset: int - sep_style: SeparatorStyle = SeparatorStyle.ADD_EOS_TOKEN - sep: str = "      " - - skip_next: bool = False - - def get_prompt(self): - if self.sep_style == SeparatorStyle.ADD_EOS_TOKEN: - ret = self.system - for role, message in self.messages: - if message: - ret += role + ": " + message + self.sep - else: - ret += role + ": " - return ret - else: - raise ValueError(f"Invalid style: {self.sep_style}") - - def append_message(self, role, message): - self.messages.append([role, message]) - - def to_gradio_chatbot(self): - ret = [] - for i, (role, msg) in enumerate(self.messages[self.offset :]): - if i % 2 == 0: - ret.append([msg, None]) - else: - ret[-1][-1] = msg - return ret - - def copy(self): - return Conversation( - system=self.system, - roles=self.roles, - messages=[[x, y] for x, y in self.messages], - offset=self.offset, - sep_style=self.sep_style, - sep=self.sep, - ) - - def dict(self): - return { - "system": self.system, - "roles": self.roles, - "messages": self.messages, - "offset": self.offset, - "sep": self.sep, - } - - -conv = Conversation( - system="A chat between a curious human and an artificial intelligence assistant. " - "The assistant gives helpful, detailed, and polite answers to the human's questions.\n\n", - roles=("Human", "Assistant"), - messages=(), - offset=0, - sep_style=SeparatorStyle.ADD_EOS_TOKEN, - sep="", -) - -default_conversation = conv diff --git a/applications/Chat/coati/dataset/prompt_dataset.py b/applications/Chat/coati/dataset/prompt_dataset.py deleted file mode 100644 index 17120e6064b5..000000000000 --- a/applications/Chat/coati/dataset/prompt_dataset.py +++ /dev/null @@ -1,45 +0,0 @@ -from collections import defaultdict -from typing import Dict - -import torch -import transformers -from torch.utils.data import Dataset - -from colossalai.logging import get_dist_logger - -from .utils import jload - - -class PromptDataset(Dataset): - """Dataset for supervised fine-tuning.""" - - def __init__( - self, - data_path: str, - tokenizer: transformers.PreTrainedTokenizer, - max_datasets_size: int = None, - max_length: int = 96, - ): - super(PromptDataset, self).__init__() - self.keyed_prompt = defaultdict(list) - self.logger = get_dist_logger() - self.logger.info("Loading data...") - list_data_dict = jload(data_path) - self.logger.info(f"Loaded {len(list_data_dict)} examples.") - - if max_datasets_size is not None: - self.logger.info(f"Limiting dataset to {max_datasets_size} examples.") - list_data_dict = list_data_dict[:max_datasets_size] - - instructions = [data_dict["instruction"] for data_dict in list_data_dict] - tokens = tokenizer( - instructions, return_tensors="pt", max_length=max_length, padding="max_length", truncation=True - ) - for k, tensor in tokens.items(): - self.keyed_prompt[k] = tensor.to(torch.cuda.current_device()).unbind() - - def __len__(self): - return len(self.keyed_prompt["input_ids"]) - - def __getitem__(self, i) -> Dict[str, torch.Tensor]: - return {k: v[i] for k, v in self.keyed_prompt.items()} diff --git a/applications/Chat/coati/dataset/reward_dataset.py b/applications/Chat/coati/dataset/reward_dataset.py deleted file mode 100644 index 3afcd7b69238..000000000000 --- a/applications/Chat/coati/dataset/reward_dataset.py +++ /dev/null @@ -1,88 +0,0 @@ -from typing import Callable - -from torch.utils.data import Dataset -from tqdm import tqdm - -from .utils import is_rank_0 - - -# Dahoas/rm-static -class RmStaticDataset(Dataset): - """ - Dataset for reward model - - Args: - dataset: dataset for reward model - tokenizer: tokenizer for reward model - max_length: max length of input - special_token: special token at the end of sentence - """ - - def __init__(self, dataset, tokenizer: Callable, max_length: int, special_token=None) -> None: - super().__init__() - self.end_token = tokenizer.eos_token if special_token is None else special_token - - chosen = [data["prompt"] + data["chosen"] + self.end_token for data in tqdm(dataset, disable=not is_rank_0())] - chosen_token = tokenizer( - chosen, max_length=max_length, padding="max_length", truncation=True, return_tensors="pt" - ) - self.chosen = {"input_ids": chosen_token["input_ids"], "attention_mask": chosen_token["attention_mask"]} - - reject = [data["prompt"] + data["rejected"] + self.end_token for data in tqdm(dataset, disable=not is_rank_0())] - reject_token = tokenizer( - reject, max_length=max_length, padding="max_length", truncation=True, return_tensors="pt" - ) - self.reject = {"input_ids": reject_token["input_ids"], "attention_mask": reject_token["attention_mask"]} - - def __len__(self): - length = self.chosen["input_ids"].shape[0] - return length - - def __getitem__(self, idx): - return ( - self.chosen["input_ids"][idx], - self.chosen["attention_mask"][idx], - self.reject["input_ids"][idx], - self.reject["attention_mask"][idx], - ) - - -# Anthropic/hh-rlhf -class HhRlhfDataset(Dataset): - """ - Dataset for reward model - - Args: - dataset: dataset for reward model - tokenizer: tokenizer for reward model - max_length: max length of input - special_token: special token at the end of sentence - """ - - def __init__(self, dataset, tokenizer: Callable, max_length: int, special_token=None) -> None: - super().__init__() - self.end_token = tokenizer.eos_token if special_token is None else special_token - - chosen = [data["chosen"] + self.end_token for data in tqdm(dataset, disable=not is_rank_0())] - chosen_token = tokenizer( - chosen, max_length=max_length, padding="max_length", truncation=True, return_tensors="pt" - ) - self.chosen = {"input_ids": chosen_token["input_ids"], "attention_mask": chosen_token["attention_mask"]} - - reject = [data["rejected"] + self.end_token for data in tqdm(dataset, disable=not is_rank_0())] - reject_token = tokenizer( - reject, max_length=max_length, padding="max_length", truncation=True, return_tensors="pt" - ) - self.reject = {"input_ids": reject_token["input_ids"], "attention_mask": reject_token["attention_mask"]} - - def __len__(self): - length = self.chosen["input_ids"].shape[0] - return length - - def __getitem__(self, idx): - return ( - self.chosen["input_ids"][idx], - self.chosen["attention_mask"][idx], - self.reject["input_ids"][idx], - self.reject["attention_mask"][idx], - ) diff --git a/applications/Chat/coati/dataset/sft_dataset.py b/applications/Chat/coati/dataset/sft_dataset.py deleted file mode 100644 index 0167f118dc8f..000000000000 --- a/applications/Chat/coati/dataset/sft_dataset.py +++ /dev/null @@ -1,210 +0,0 @@ -# Copyright 2023 Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -import copy -from typing import Dict, Optional, Sequence, Tuple - -import torch -from coati.models.chatglm.chatglm_tokenizer import ChatGLMTokenizer -from torch.utils.data import Dataset -from tqdm import tqdm -from transformers import PreTrainedTokenizer - -from colossalai.logging import get_dist_logger - -from .utils import is_rank_0, jload - -logger = get_dist_logger() - -IGNORE_INDEX = -100 -PROMPT_DICT = { - "prompt_input": ( - "Below is an instruction that describes a task, paired with an input that provides further context. " - "Write a response that appropriately completes the request.\n\n" - "### Instruction:\n{instruction}\n\n### Input:\n{input}\n\n### Response:" - ), - "prompt_no_input": ( - "Below is an instruction that describes a task. " - "Write a response that appropriately completes the request.\n\n" - "### Instruction:\n{instruction}\n\n### Response:" - ), -} - - -def _preprocess( - sources: Sequence[str], - targets: Sequence[str], - tokenizer: PreTrainedTokenizer, - max_length: int, -) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]: - """Preprocess the data by tokenizing.""" - sequences = [s + t + tokenizer.eos_token for s, t in zip(sources, targets)] - sequences_token = tokenizer( - sequences, - max_length=max_length, - padding="max_length", - truncation=True, - return_tensors="pt", - add_special_tokens=False, - ) - - sources_token = tokenizer( - sources, - max_length=max_length, - padding="max_length", - truncation=True, - return_tensors="pt", - add_special_tokens=False, - ) - - assert sequences_token["attention_mask"].dim() == 2, "seq2seq model should be preprocessed differently" - labels = copy.deepcopy(sequences_token["input_ids"]) - for i in range(labels.shape[0]): - source_len = sources_token["attention_mask"][i].sum().item() - pad_len = max_length - sequences_token["attention_mask"][i].sum().item() - if tokenizer.padding_side == "right": - # |prompt|completion|eos|pad| - labels[i][:source_len] = IGNORE_INDEX - if pad_len > 0: - labels[i][-pad_len:] = IGNORE_INDEX - elif tokenizer.padding_side == "left": - # |pad|prompt|completion|eos| - labels[i][: pad_len + source_len] = IGNORE_INDEX - else: - raise RuntimeError() - - return sequences_token["input_ids"], labels, sequences_token["attention_mask"] - - -def _preprocess_chatglm( - sources: Sequence[str], - targets: Sequence[str], - tokenizer: PreTrainedTokenizer, - max_length: int, -) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]: - """ - Preprocess the data by tokenizing. - None for attention mask, ChatGLM will calculate attention mask according to input ids - """ - - labels = [] - input_ids = [] - for source, target in zip(sources, targets): - source_id = tokenizer.encode(text=source, add_special_tokens=False) - target_id = tokenizer.encode(text=target, add_special_tokens=False) - input_id = tokenizer.build_inputs_with_special_tokens(source_id, target_id) - # truncate - sp_token_list = [tokenizer.gmask_token_id, tokenizer.bos_token_id] - truncate_length = max(0, len(input_id) - max_length) - input_id = input_id[truncate_length:] - if truncate_length == len(source_id) + 1: - input_id = sp_token_list + input_id[1:] - elif truncate_length > len(source_id) + 1: - input_id = sp_token_list + input_id[2:] - - context_length = input_id.index(tokenizer.bos_token_id) - mask_position = context_length - 1 - label = [IGNORE_INDEX] * context_length + input_id[mask_position + 1 :] - - pad_len = max_length - len(input_id) - input_id = input_id + [tokenizer.pad_token_id] * pad_len - input_ids.append(input_id) - labels.append(label + [IGNORE_INDEX] * pad_len) - return torch.tensor(input_ids), torch.tensor(labels), None - - -class SFTDataset(Dataset): - """ - Dataset for sft model - - Args: - dataset: dataset for supervised model - tokenizer: tokenizer for supervised model - max_length: max length of input - """ - - def __init__(self, dataset: Dict, tokenizer: PreTrainedTokenizer, max_length: int = 512) -> None: - super().__init__() - self.input_ids = [] - - sources = [data["prompt"] for data in dataset] - targets = [data["completion"] + tokenizer.eos_token for data in tqdm(dataset, disable=not is_rank_0())] - - logger.info("Tokenizing inputs... This may take some time...") - if isinstance(tokenizer, ChatGLMTokenizer): - self.input_ids, self.labels, self.attention_mask = _preprocess_chatglm( - sources, targets, tokenizer, max_length - ) - else: - self.input_ids, self.labels, self.attention_mask = _preprocess(sources, targets, tokenizer, max_length) - - logger.info("Loaded dataset.") - - def __len__(self): - length = self.input_ids.shape[0] - return length - - def __getitem__(self, idx): - if self.attention_mask is not None: - return dict(input_ids=self.input_ids[idx], labels=self.labels[idx], attention_mask=self.attention_mask[idx]) - else: - return dict(input_ids=self.input_ids[idx], labels=self.labels[idx]) - - -class SupervisedDataset(Dataset): - """Dataset for supervised fine-tuning.""" - - def __init__( - self, - data_path: str, - tokenizer: PreTrainedTokenizer, - max_datasets_size: Optional[int] = None, - max_length: int = 512, - ): - super().__init__() - logger.info("Loading data...") - list_data_dict = jload(data_path) - logger.info(f"Loaded {len(list_data_dict)} examples.") - - if max_datasets_size is not None: - logger.info(f"Limiting dataset to {max_datasets_size} examples.") - list_data_dict = list_data_dict[:max_datasets_size] - - logger.info("Formatting inputs...") - prompt_input, prompt_no_input = PROMPT_DICT["prompt_input"], PROMPT_DICT["prompt_no_input"] - sources = [ - prompt_input.format_map(example) if "input" in example else prompt_no_input.format_map(example) - for example in list_data_dict - ] - targets = [example["output"] + tokenizer.eos_token for example in list_data_dict] - - logger.info("Tokenizing inputs... This may take some time...") - if isinstance(tokenizer, ChatGLMTokenizer): - self.input_ids, self.labels, self.attention_mask = _preprocess_chatglm( - sources, targets, tokenizer, max_length - ) - else: - self.input_ids, self.labels, self.attention_mask = _preprocess(sources, targets, tokenizer, max_length) - - logger.info("Loaded dataset.") - - def __len__(self): - length = self.input_ids.shape[0] - return length - - def __getitem__(self, idx): - if self.attention_mask is not None: - return dict(input_ids=self.input_ids[idx], labels=self.labels[idx], attention_mask=self.attention_mask[idx]) - else: - return dict(input_ids=self.input_ids[idx], labels=self.labels[idx]) diff --git a/applications/Chat/coati/dataset/utils.py b/applications/Chat/coati/dataset/utils.py deleted file mode 100644 index f37fce67a7c6..000000000000 --- a/applications/Chat/coati/dataset/utils.py +++ /dev/null @@ -1,22 +0,0 @@ -import io -import json - -import torch.distributed as dist - - -def is_rank_0() -> bool: - return not dist.is_initialized() or dist.get_rank() == 0 - - -def _make_r_io_base(f, mode: str): - if not isinstance(f, io.IOBase): - f = open(f, mode=mode) - return f - - -def jload(f, mode="r"): - """Load a .json file into a dictionary.""" - f = _make_r_io_base(f, mode) - jdict = json.load(f) - f.close() - return jdict diff --git a/applications/Chat/coati/experience_maker/naive.py b/applications/Chat/coati/experience_maker/naive.py deleted file mode 100644 index 941e1994b148..000000000000 --- a/applications/Chat/coati/experience_maker/naive.py +++ /dev/null @@ -1,71 +0,0 @@ -import torch -import torch.nn.functional as F -from coati.models.base import Actor, Critic, RewardModel -from coati.models.generation import generate -from coati.models.utils import calc_action_log_probs, compute_reward -from transformers import PreTrainedTokenizer - -from .base import Experience, ExperienceMaker - - -class NaiveExperienceMaker(ExperienceMaker): - """ - Naive experience maker. - """ - - def __init__( - self, - actor: Actor, - critic: Critic, - reward_model: RewardModel, - initial_model: Actor, - tokenizer: PreTrainedTokenizer, - kl_coef: float = 0.1, - ) -> None: - super().__init__(actor, critic, reward_model, initial_model) - self.tokenizer = tokenizer - self.kl_coef = kl_coef - - @torch.no_grad() - def make_experience(self, input_ids: torch.Tensor, **generate_kwargs) -> Experience: - self.actor.eval() - self.critic.eval() - self.initial_model.eval() - self.reward_model.eval() - - # generate sequences - sequences = generate(self.actor, input_ids, self.tokenizer, **generate_kwargs) - - # calculate auxiliary tensors - attention_mask = None - pad_token_id = self.tokenizer.pad_token_id - if pad_token_id is not None: - attention_mask = sequences.not_equal(pad_token_id).to(dtype=torch.long, device=sequences.device) - - input_len = input_ids.size(1) - eos_token_id = self.tokenizer.eos_token_id - if eos_token_id is None: - action_mask = torch.ones_like(sequences, dtype=torch.bool) - else: - # left padding may be applied, only mask action - action_mask = (sequences[:, input_len:] == eos_token_id).cumsum(dim=-1) == 0 - action_mask = F.pad(action_mask, (1 + input_len, -1), value=True) # include eos token and input - action_mask[:, :input_len] = False - action_mask = action_mask[:, 1:] - action_mask = action_mask[:, -(sequences.size(1) - input_len) :] - num_actions = action_mask.size(1) - - actor_output = self.actor(sequences, attention_mask)["logits"] - action_log_probs = calc_action_log_probs(actor_output, sequences, num_actions) - base_model_output = self.initial_model(sequences, attention_mask)["logits"] - base_action_log_probs = calc_action_log_probs(base_model_output, sequences, num_actions) - value = self.critic(sequences, attention_mask) - r = self.reward_model(sequences, attention_mask) - reward = compute_reward(r, self.kl_coef, action_log_probs, base_action_log_probs, action_mask=action_mask) - - advantage = reward - value - # TODO(ver217): maybe normalize adv - if advantage.ndim == 1: - advantage = advantage.unsqueeze(-1) - - return Experience(sequences, action_log_probs, value, reward, advantage, attention_mask, action_mask) diff --git a/applications/Chat/coati/kernels/__init__.py b/applications/Chat/coati/kernels/__init__.py deleted file mode 100644 index 96d40c7c4709..000000000000 --- a/applications/Chat/coati/kernels/__init__.py +++ /dev/null @@ -1,6 +0,0 @@ -from .wrapper import convert_to_xformer_model, recover_from_xformer_model - -__all__ = [ - "convert_to_xformer_model", - "recover_from_xformer_model", -] diff --git a/applications/Chat/coati/kernels/opt_attn.py b/applications/Chat/coati/kernels/opt_attn.py deleted file mode 100644 index d1eb139187f3..000000000000 --- a/applications/Chat/coati/kernels/opt_attn.py +++ /dev/null @@ -1,90 +0,0 @@ -from typing import Optional, Tuple - -import torch -import xformers.ops as xops -from torch import Tensor -from transformers.models.opt.modeling_opt import OPTAttention - - -# This is modified from https://github.com/huggingface/transformers/blob/main/src/transformers/models/opt/modeling_opt.py -class XOPTAttention(OPTAttention): - # def _shape(self, tensor: Tensor, seq_len: int, bsz: int): - # return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).contiguous() - - def forward( - self, - hidden_states: Tensor, - key_value_states: Optional[Tensor] = None, - past_key_value: Optional[Tensor] = None, - attention_mask: Optional[Tensor] = None, - layer_head_mask: Optional[Tensor] = None, - output_attentions: bool = False, - ) -> Tuple[Tensor, Optional[Tensor], Optional[Tuple[Tensor]]]: - if not self.training: - return super().forward( - hidden_states, key_value_states, past_key_value, attention_mask, layer_head_mask, output_attentions - ) - """Input shape: Batch x Time x Channel""" - assert layer_head_mask is None, "Xformers attention does not support layer_head_mask" - assert not output_attentions, "Xformers attention does not support output_attentions" - - # if key_value_states are provided this layer is used as a cross-attention layer - # for the decoder - is_cross_attention = key_value_states is not None - - bsz, tgt_len, _ = hidden_states.size() - - # get query proj - query_states = self.q_proj(hidden_states) - # get key, value proj - if is_cross_attention and past_key_value is not None: - # reuse k,v, cross_attentions - key_states = past_key_value[0] - value_states = past_key_value[1] - elif is_cross_attention: - # cross_attentions - key_states = self._shape(self.k_proj(key_value_states), -1, bsz) - value_states = self._shape(self.v_proj(key_value_states), -1, bsz) - elif past_key_value is not None: - # reuse k, v, self_attention - key_states = self._shape(self.k_proj(hidden_states), -1, bsz) - value_states = self._shape(self.v_proj(hidden_states), -1, bsz) - key_states = torch.cat([past_key_value[0], key_states], dim=2) - value_states = torch.cat([past_key_value[1], value_states], dim=2) - else: - # self_attention - key_states = self._shape(self.k_proj(hidden_states), -1, bsz) - value_states = self._shape(self.v_proj(hidden_states), -1, bsz) - - if self.is_decoder: - # if cross_attention save Tuple(torch.Tensor, torch.Tensor) of all cross attention key/value_states. - # Further calls to cross_attention layer can then reuse all cross-attention - # key/value_states (first "if" case) - # if uni-directional self-attention (decoder) save Tuple(torch.Tensor, torch.Tensor) of - # all previous decoder key/value_states. Further calls to uni-directional self-attention - # can concat previous decoder key/value_states to current projected key/value_states (third "elif" case) - # if encoder bi-directional self-attention `past_key_value` is always `None` - past_key_value = (key_states, value_states) - - query_states = self._shape(query_states, tgt_len, bsz).transpose(1, 2) - key_states = key_states.transpose(1, 2) - value_states = value_states.transpose(1, 2) - - attn_output = xops.memory_efficient_attention( - query_states, - key_states, - value_states, - attn_bias=xops.LowerTriangularMask(), - p=self.dropout if self.training else 0.0, - scale=self.scaling, - ) - - # Use the `embed_dim` from the config (stored in the class) rather than `hidden_state` because `attn_output` can be - # partitioned across GPUs when using tensor-parallelism. - attn_output = attn_output.reshape(bsz, tgt_len, self.embed_dim) - - attn_output = self.out_proj(attn_output) - - attn_weights_reshaped = None - - return attn_output, attn_weights_reshaped, past_key_value diff --git a/applications/Chat/coati/kernels/wrapper.py b/applications/Chat/coati/kernels/wrapper.py deleted file mode 100644 index c55bda600230..000000000000 --- a/applications/Chat/coati/kernels/wrapper.py +++ /dev/null @@ -1,18 +0,0 @@ -import torch.nn as nn -from transformers.models.opt.modeling_opt import OPTAttention - -from .opt_attn import XOPTAttention - - -def convert_to_xformer_model(model: nn.Module) -> nn.Module: - for module in model.modules(): - if isinstance(module, OPTAttention): - module.__class__ = XOPTAttention - return model - - -def recover_from_xformer_model(model: nn.Module) -> nn.Module: - for module in model.modules(): - if isinstance(module, XOPTAttention): - module.__class__ = OPTAttention - return model diff --git a/applications/Chat/coati/models/__init__.py b/applications/Chat/coati/models/__init__.py deleted file mode 100644 index ad4a525b4af2..000000000000 --- a/applications/Chat/coati/models/__init__.py +++ /dev/null @@ -1,15 +0,0 @@ -from .base import Actor, Critic, RewardModel -from .lora import LoRAModule, convert_to_lora_module -from .loss import LogExpLoss, LogSigLoss, PolicyLoss, ValueLoss - -__all__ = [ - "Actor", - "Critic", - "RewardModel", - "PolicyLoss", - "ValueLoss", - "LogSigLoss", - "LogExpLoss", - "LoRAModule", - "convert_to_lora_module", -] diff --git a/applications/Chat/coati/models/base/__init__.py b/applications/Chat/coati/models/base/__init__.py deleted file mode 100644 index 5c9905bb2224..000000000000 --- a/applications/Chat/coati/models/base/__init__.py +++ /dev/null @@ -1,27 +0,0 @@ -from typing import Union - -import torch.nn as nn - -from .actor import Actor -from .critic import Critic -from .reward_model import RewardModel - - -def get_base_model(model: Union[Actor, Critic, RewardModel]) -> nn.Module: - """Get the base model of our wrapper classes. - For Actor, Critic and RewardModel, return ``model.model``, - it's usually a ``transformers.PreTrainedModel``. - - Args: - model (nn.Module): model to get base model from - - Returns: - nn.Module: the base model - """ - assert isinstance( - model, (Actor, Critic, RewardModel) - ), f"Expect Actor, Critic or RewardModel, got {type(model)}, use unwrap_model first." - return model.model - - -__all__ = ["Actor", "Critic", "RewardModel", "get_base_model"] diff --git a/applications/Chat/coati/models/base/actor.py b/applications/Chat/coati/models/base/actor.py deleted file mode 100644 index 0634631df7a3..000000000000 --- a/applications/Chat/coati/models/base/actor.py +++ /dev/null @@ -1,32 +0,0 @@ -from typing import Optional - -import torch -import torch.nn as nn - -from ..lora import LoRAModule - - -class Actor(LoRAModule): - """ - Actor model base class. - - Args: - model (nn.Module): Actor Model. - lora_rank (int): LoRA rank. - lora_train_bias (str): LoRA bias training mode. - """ - - def __init__(self, model: nn.Module, lora_rank: int = 0, lora_train_bias: str = "none") -> None: - super().__init__(lora_rank=lora_rank, lora_train_bias=lora_train_bias) - self.model = model - self.convert_to_lora() - - def forward( - self, - input_ids: torch.LongTensor, - attention_mask: Optional[torch.Tensor] = None, - **model_kwargs, - ) -> torch.Tensor: - """Returns model output.""" - output = self.model(input_ids, attention_mask=attention_mask, **model_kwargs) - return output diff --git a/applications/Chat/coati/models/base/critic.py b/applications/Chat/coati/models/base/critic.py deleted file mode 100644 index 8672365f5783..000000000000 --- a/applications/Chat/coati/models/base/critic.py +++ /dev/null @@ -1,34 +0,0 @@ -import torch -import torch.nn as nn - -from ..lora import LoRAModule - - -class Critic(LoRAModule): - """ - Critic model base class. - - Args: - model (nn.Module): Critic model. - value_head (nn.Module): Value head to get value. - lora_rank (int): LoRA rank. - lora_train_bias (str): LoRA bias training mode. - """ - - def __init__( - self, model: nn.Module, value_head: nn.Module, lora_rank: int = 0, lora_train_bias: str = "none" - ) -> None: - super().__init__(lora_rank=lora_rank, lora_train_bias=lora_train_bias) - self.model = model - self.value_head = value_head - self.convert_to_lora() - - def forward(self, sequences: torch.LongTensor, attention_mask: torch.Tensor) -> torch.Tensor: - outputs = self.model(sequences, attention_mask=attention_mask) - last_hidden_states = outputs["last_hidden_state"] - sequence_lengths = torch.max(attention_mask * torch.arange(sequences.size(1), device=sequences.device), dim=1)[ - 0 - ] - sequence_hidden_states = last_hidden_states[torch.arange(last_hidden_states.size(0)), sequence_lengths] - values = self.value_head(sequence_hidden_states).squeeze(1) # ensure shape is (B, ) - return values diff --git a/applications/Chat/coati/models/base/reward_model.py b/applications/Chat/coati/models/base/reward_model.py deleted file mode 100644 index e9545d1cddaf..000000000000 --- a/applications/Chat/coati/models/base/reward_model.py +++ /dev/null @@ -1,46 +0,0 @@ -from typing import Optional - -import torch -import torch.nn as nn - -from ..lora import LoRAModule - - -class RewardModel(LoRAModule): - """ - Reward model base class. - - Args: - model (nn.Module): Reward model. - value_head (nn.Module): Value head to get reward score. - lora_rank (int): LoRA rank. - lora_train_bias (str): LoRA bias training mode. - """ - - def __init__( - self, - model: nn.Module, - value_head: Optional[nn.Module] = None, - lora_rank: int = 0, - lora_train_bias: str = "none", - ) -> None: - super().__init__(lora_rank=lora_rank, lora_train_bias=lora_train_bias) - self.model = model - self.convert_to_lora() - - if value_head is not None: - if value_head.out_features != 1: - raise ValueError("The value head of reward model's output dim should be 1!") - self.value_head = value_head - else: - self.value_head = nn.Linear(model.config.n_embd, 1) - - def forward(self, sequences: torch.LongTensor, attention_mask: torch.Tensor) -> torch.Tensor: - outputs = self.model(sequences, attention_mask=attention_mask) - last_hidden_states = outputs["last_hidden_state"] - sequence_lengths = torch.max(attention_mask * torch.arange(sequences.size(1), device=sequences.device), dim=1)[ - 0 - ] - sequence_hidden_states = last_hidden_states[torch.arange(last_hidden_states.size(0)), sequence_lengths] - values = self.value_head(sequence_hidden_states).squeeze(1) # ensure shape is (B, ) - return values diff --git a/applications/Chat/coati/models/bloom/__init__.py b/applications/Chat/coati/models/bloom/__init__.py deleted file mode 100644 index 7af199a67d3b..000000000000 --- a/applications/Chat/coati/models/bloom/__init__.py +++ /dev/null @@ -1,5 +0,0 @@ -from .bloom_actor import BLOOMActor -from .bloom_critic import BLOOMCritic -from .bloom_rm import BLOOMRM - -__all__ = ["BLOOMActor", "BLOOMCritic", "BLOOMRM"] diff --git a/applications/Chat/coati/models/bloom/bloom_actor.py b/applications/Chat/coati/models/bloom/bloom_actor.py deleted file mode 100644 index 73855a2245e7..000000000000 --- a/applications/Chat/coati/models/bloom/bloom_actor.py +++ /dev/null @@ -1,36 +0,0 @@ -from typing import Optional - -from transformers import BloomConfig, BloomForCausalLM - -from ..base import Actor - - -class BLOOMActor(Actor): - """ - BLOOM Actor model. - - Args: - pretrained (str): Pretrained model name or path. - config (BloomConfig): Model config. - checkpoint (bool): Enable gradient checkpointing. - lora_rank (int): LoRA rank. - lora_train_bias (str): LoRA bias training mode. - """ - - def __init__( - self, - pretrained: str = None, - config: Optional[BloomConfig] = None, - checkpoint: bool = False, - lora_rank: int = 0, - lora_train_bias: str = "none", - ) -> None: - if pretrained is not None: - model = BloomForCausalLM.from_pretrained(pretrained) - elif config is not None: - model = BloomForCausalLM(config) - else: - model = BloomForCausalLM(BloomConfig()) - if checkpoint: - model.gradient_checkpointing_enable() - super().__init__(model, lora_rank, lora_train_bias) diff --git a/applications/Chat/coati/models/bloom/bloom_critic.py b/applications/Chat/coati/models/bloom/bloom_critic.py deleted file mode 100644 index b2d838f7ffc5..000000000000 --- a/applications/Chat/coati/models/bloom/bloom_critic.py +++ /dev/null @@ -1,36 +0,0 @@ -from typing import Optional - -import torch.nn as nn -from transformers import BloomConfig, BloomModel - -from ..base import Critic - - -class BLOOMCritic(Critic): - """ - BLOOM Critic model. - - Args: - pretrained (str): Pretrained model name or path. - config (BloomConfig): Model config. - lora_rank (int): LoRA rank. - lora_train_bias (str): LoRA bias training mode. - """ - - def __init__( - self, - pretrained: str = None, - config: Optional[BloomConfig] = None, - lora_rank: int = 0, - lora_train_bias: str = "none", - **kwargs, - ) -> None: - if pretrained is not None: - model = BloomModel.from_pretrained(pretrained) - elif config is not None: - model = BloomModel(config) - else: - model = BloomModel(BloomConfig()) - - value_head = nn.Linear(model.config.hidden_size, 1) - super().__init__(model, value_head, lora_rank, lora_train_bias, **kwargs) diff --git a/applications/Chat/coati/models/bloom/bloom_rm.py b/applications/Chat/coati/models/bloom/bloom_rm.py deleted file mode 100644 index c09457ddc8c7..000000000000 --- a/applications/Chat/coati/models/bloom/bloom_rm.py +++ /dev/null @@ -1,36 +0,0 @@ -from typing import Optional - -import torch.nn as nn -from transformers import BloomConfig, BloomModel - -from ..base import RewardModel - - -class BLOOMRM(RewardModel): - """ - BLOOM Reward model. - - Args: - pretrained (str): Pretrained model name or path. - config (BloomConfig): Model config. - lora_rank (int): LoRA rank. - lora_train_bias (str): LoRA bias training mode. - """ - - def __init__( - self, - pretrained: str = None, - config: Optional[BloomConfig] = None, - lora_rank: int = 0, - lora_train_bias: str = "none", - ) -> None: - if pretrained is not None: - model = BloomModel.from_pretrained(pretrained) - elif config is not None: - model = BloomModel(config) - else: - model = BloomModel(BloomConfig()) - - value_head = nn.Linear(model.config.hidden_size, 1) - value_head.weight.data.normal_(mean=0.0, std=1 / (model.config.hidden_size + 1)) - super().__init__(model, value_head, lora_rank, lora_train_bias) diff --git a/applications/Chat/coati/models/chatglm/__init__.py b/applications/Chat/coati/models/chatglm/__init__.py deleted file mode 100644 index 5956f5a8e91b..000000000000 --- a/applications/Chat/coati/models/chatglm/__init__.py +++ /dev/null @@ -1,3 +0,0 @@ -from .chatglm_actor import ChatGLMActor - -__all__ = ["ChatGLMActor"] diff --git a/applications/Chat/coati/models/chatglm/chatglm_actor.py b/applications/Chat/coati/models/chatglm/chatglm_actor.py deleted file mode 100644 index 00a61561ee47..000000000000 --- a/applications/Chat/coati/models/chatglm/chatglm_actor.py +++ /dev/null @@ -1,31 +0,0 @@ -from typing import Optional - -from ..base import Actor -from .configuration_chatglm import ChatGLMConfig -from .modeling_chatglm import ChatGLMForConditionalGeneration - - -class ChatGLMActor(Actor): - """ - ChatGLM Actor model. - - Args: - pretrained (str): Pretrained model name or path. - config (ChatGLMConfig): Model config. - checkpoint (bool): Enable gradient checkpointing. - - do not support lora for now. - """ - - def __init__( - self, pretrained: str = None, config: Optional[ChatGLMConfig] = None, checkpoint: bool = False - ) -> None: - if pretrained is not None: - model = ChatGLMForConditionalGeneration.from_pretrained(pretrained) - elif config is not None: - model = ChatGLMForConditionalGeneration(config) - else: - model = ChatGLMForConditionalGeneration(ChatGLMConfig()) - if checkpoint: - model.gradient_checkpointing_enable() - super().__init__(model, lora_rank=0, lora_train_bias="none") diff --git a/applications/Chat/coati/models/chatglm/chatglm_tokenizer.py b/applications/Chat/coati/models/chatglm/chatglm_tokenizer.py deleted file mode 100644 index 221ef044b470..000000000000 --- a/applications/Chat/coati/models/chatglm/chatglm_tokenizer.py +++ /dev/null @@ -1,442 +0,0 @@ -""" -This code is copied from https://huggingface.co/THUDM/chatglm-6b/blob/main/tokenization_chatglm.py -""" -"""Tokenization classes for ChatGLM.""" -import os -from typing import Dict, List, Optional, Union - -import numpy as np -import sentencepiece as spm -from transformers.tokenization_utils import PreTrainedTokenizer -from transformers.tokenization_utils_base import BatchEncoding, EncodedInput -from transformers.utils import PaddingStrategy, logging - -logger = logging.get_logger(__name__) - -PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES = { - "THUDM/chatglm-6b": 2048, -} - - -class TextTokenizer: - def __init__(self, model_path): - self.sp = spm.SentencePieceProcessor() - self.sp.Load(model_path) - self.num_tokens = self.sp.vocab_size() - - def encode(self, text): - return self.sp.EncodeAsIds(text) - - def decode(self, ids: List[int]): - return self.sp.DecodeIds(ids) - - def tokenize(self, text): - return self.sp.EncodeAsPieces(text) - - def convert_tokens_to_string(self, tokens): - return self.sp.DecodePieces(tokens) - - def convert_tokens_to_ids(self, tokens): - return [self.sp.PieceToId(token) for token in tokens] - - def convert_token_to_id(self, token): - return self.sp.PieceToId(token) - - def convert_id_to_token(self, idx): - return self.sp.IdToPiece(idx) - - def __len__(self): - return self.num_tokens - - -class SPTokenizer: - def __init__( - self, - vocab_file, - num_image_tokens=20000, - max_blank_length=80, - byte_fallback=True, - ): - assert vocab_file is not None - self.vocab_file = vocab_file - self.num_image_tokens = num_image_tokens - self.special_tokens = ["[MASK]", "[gMASK]", "[sMASK]", "", "", "", "", ""] - self.max_blank_length = max_blank_length - self.byte_fallback = byte_fallback - self.text_tokenizer = TextTokenizer(vocab_file) - - def _get_text_tokenizer(self): - return self.text_tokenizer - - @staticmethod - def get_blank_token(length: int): - assert length >= 2 - return f"<|blank_{length}|>" - - @staticmethod - def get_tab_token(): - return f"<|tab|>" - - @property - def num_text_tokens(self): - return self.text_tokenizer.num_tokens - - @property - def num_tokens(self): - return self.num_image_tokens + self.num_text_tokens - - @staticmethod - def _encode_whitespaces(text: str, max_len: int = 80): - text = text.replace("\t", SPTokenizer.get_tab_token()) - for i in range(max_len, 1, -1): - text = text.replace(" " * i, SPTokenizer.get_blank_token(i)) - return text - - def _preprocess(self, text: str, linebreak=True, whitespaces=True): - if linebreak: - text = text.replace("\n", "") - if whitespaces: - text = self._encode_whitespaces(text, max_len=self.max_blank_length) - return text - - def encode(self, text: str, linebreak=True, whitespaces=True, add_dummy_prefix=True) -> List[int]: - """ - @param text: Text to encode. - @param linebreak: Whether to encode newline (\n) in text. - @param whitespaces: Whether to encode multiple whitespaces or tab in text, useful for source code encoding. - @param special_tokens: Whether to encode special token ([MASK], [gMASK], etc.) in text. - @param add_dummy_prefix: Whether to add dummy blank space in the beginning. - """ - text = self._preprocess(text, linebreak, whitespaces) - if not add_dummy_prefix: - text = "" + text - tmp = self._get_text_tokenizer().encode(text) - tokens = [x + self.num_image_tokens for x in tmp] - return tokens if add_dummy_prefix else tokens[2:] - - def postprocess(self, text): - text = text.replace("", "\n") - text = text.replace(SPTokenizer.get_tab_token(), "\t") - for i in range(2, self.max_blank_length + 1): - text = text.replace(self.get_blank_token(i), " " * i) - return text - - def decode(self, text_ids: List[int]) -> str: - ids = [int(_id) - self.num_image_tokens for _id in text_ids] - ids = [_id for _id in ids if _id >= 0] - text = self._get_text_tokenizer().decode(ids) - text = self.postprocess(text) - return text - - def decode_tokens(self, tokens: List[str]) -> str: - text = self._get_text_tokenizer().convert_tokens_to_string(tokens) - text = self.postprocess(text) - return text - - def tokenize(self, text: str, linebreak=True, whitespaces=True, add_dummy_prefix=True) -> List[str]: - """ - @param text: Text to encode. - @param linebreak: Whether to encode newline (\n) in text. - @param whitespaces: Whether to encode multiple whitespaces or tab in text, useful for source code encoding. - @param special_tokens: Whether to encode special token ([MASK], [gMASK], etc.) in text. - @param add_dummy_prefix: Whether to add dummy blank space in the beginning. - """ - text = self._preprocess(text, linebreak, whitespaces) - if not add_dummy_prefix: - text = "" + text - tokens = self._get_text_tokenizer().tokenize(text) - return tokens if add_dummy_prefix else tokens[2:] - - def __getitem__(self, x: Union[int, str]): - if isinstance(x, int): - if x < self.num_image_tokens: - return "".format(x) - else: - return self.text_tokenizer.convert_id_to_token(x - self.num_image_tokens) - elif isinstance(x, str): - if x.startswith("") and x[7:-1].isdigit(): - return int(x[7:-1]) - else: - return self.text_tokenizer.convert_token_to_id(x) + self.num_image_tokens - else: - raise ValueError("The key should be str or int.") - - -class ChatGLMTokenizer(PreTrainedTokenizer): - """ - Construct a ChatGLM tokenizer. Based on byte-level Byte-Pair-Encoding. - - Args: - vocab_file (`str`): - Path to the vocabulary file. - """ - - vocab_files_names = {"vocab_file": "ice_text.model"} - max_model_input_sizes = PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES - model_input_names = ["input_ids", "attention_mask", "position_ids"] - - def __init__( - self, - vocab_file, - do_lower_case=False, - remove_space=False, - bos_token="", - eos_token="", - end_token="", - mask_token="[MASK]", - gmask_token="[gMASK]", - padding_side="left", - pad_token="", - unk_token="", - num_image_tokens=20000, - **kwargs, - ) -> None: - super().__init__( - do_lower_case=do_lower_case, - remove_space=remove_space, - padding_side=padding_side, - bos_token=bos_token, - eos_token=eos_token, - end_token=end_token, - mask_token=mask_token, - gmask_token=gmask_token, - pad_token=pad_token, - unk_token=unk_token, - num_image_tokens=num_image_tokens, - **kwargs, - ) - - self.do_lower_case = do_lower_case - self.remove_space = remove_space - self.vocab_file = vocab_file - - self.bos_token = bos_token - self.eos_token = eos_token - self.end_token = end_token - self.mask_token = mask_token - self.gmask_token = gmask_token - - self.sp_tokenizer = SPTokenizer(vocab_file, num_image_tokens=num_image_tokens) - - """ Initialisation """ - - @property - def gmask_token_id(self) -> Optional[int]: - if self.gmask_token is None: - return None - return self.convert_tokens_to_ids(self.gmask_token) - - @property - def end_token_id(self) -> Optional[int]: - """ - `Optional[int]`: Id of the end of context token in the vocabulary. Returns `None` if the token has not been - set. - """ - if self.end_token is None: - return None - return self.convert_tokens_to_ids(self.end_token) - - @property - def vocab_size(self): - """Returns vocab size""" - return self.sp_tokenizer.num_tokens - - def get_vocab(self): - """Returns vocab as a dict""" - vocab = {self._convert_id_to_token(i): i for i in range(self.vocab_size)} - vocab.update(self.added_tokens_encoder) - return vocab - - def preprocess_text(self, inputs): - if self.remove_space: - outputs = " ".join(inputs.strip().split()) - else: - outputs = inputs - - if self.do_lower_case: - outputs = outputs.lower() - - return outputs - - def _tokenize(self, text, **kwargs): - """Returns a tokenized string.""" - text = self.preprocess_text(text) - - seq = self.sp_tokenizer.tokenize(text) - - return seq - - def convert_tokens_to_string(self, tokens: List[str]) -> str: - return self.sp_tokenizer.decode_tokens(tokens) - - def _decode(self, token_ids: Union[int, List[int]], **kwargs) -> str: - if isinstance(token_ids, int): - token_ids = [token_ids] - if len(token_ids) == 0: - return "" - if self.pad_token_id in token_ids: # remove pad - token_ids = list(filter((self.pad_token_id).__ne__, token_ids)) - return super()._decode(token_ids, **kwargs) - - def _convert_token_to_id(self, token): - """Converts a token (str) in an id using the vocab.""" - return self.sp_tokenizer[token] - - def _convert_id_to_token(self, index): - """Converts an index (integer) in a token (str) using the vocab.""" - return self.sp_tokenizer[index] - - def save_vocabulary(self, save_directory, filename_prefix=None): - """ - Save the vocabulary and special tokens file to a directory. - - Args: - save_directory (`str`): - The directory in which to save the vocabulary. - filename_prefix (`str`, *optional*): - An optional prefix to add to the named of the saved files. - - Returns: - `Tuple(str)`: Paths to the files saved. - """ - if os.path.isdir(save_directory): - vocab_file = os.path.join(save_directory, self.vocab_files_names["vocab_file"]) - else: - vocab_file = save_directory - - with open(self.vocab_file, "rb") as fin: - proto_str = fin.read() - - with open(vocab_file, "wb") as writer: - writer.write(proto_str) - - return (vocab_file,) - - def build_inputs_with_special_tokens( - self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None - ) -> List[int]: - """ - Build model inputs from a sequence or a pair of sequence for sequence classification tasks by concatenating and - adding special tokens. A BERT sequence has the following format: - - - single sequence: `[CLS] X [SEP]` - - pair of sequences: `[CLS] A [SEP] B [SEP]` - - Args: - token_ids_0 (`List[int]`): - List of IDs to which the special tokens will be added. - token_ids_1 (`List[int]`, *optional*): - Optional second list of IDs for sequence pairs. - - Returns: - `List[int]`: List of [input IDs](../glossary#input-ids) with the appropriate special tokens. - """ - gmask_id = self.sp_tokenizer[self.gmask_token] - self.sp_tokenizer[self.eos_token] - token_ids_0 = token_ids_0 + [gmask_id, self.sp_tokenizer[self.bos_token]] - if token_ids_1 is not None: - token_ids_0 = token_ids_0 + token_ids_1 - return token_ids_0 - - def _pad( - self, - encoded_inputs: Union[Dict[str, EncodedInput], BatchEncoding], - max_length: Optional[int] = None, - padding_strategy: PaddingStrategy = PaddingStrategy.DO_NOT_PAD, - pad_to_multiple_of: Optional[int] = None, - return_attention_mask: Optional[bool] = None, - ) -> dict: - """ - Pad encoded inputs (on left/right and up to predefined length or max length in the batch) - - Args: - encoded_inputs: - Dictionary of tokenized inputs (`List[int]`) or batch of tokenized inputs (`List[List[int]]`). - max_length: maximum length of the returned list and optionally padding length (see below). - Will truncate by taking into account the special tokens. - padding_strategy: PaddingStrategy to use for padding. - - - PaddingStrategy.LONGEST Pad to the longest sequence in the batch - - PaddingStrategy.MAX_LENGTH: Pad to the max length (default) - - PaddingStrategy.DO_NOT_PAD: Do not pad - The tokenizer padding sides are defined in self.padding_side: - - - 'left': pads on the left of the sequences - - 'right': pads on the right of the sequences - pad_to_multiple_of: (optional) Integer if set will pad the sequence to a multiple of the provided value. - This is especially useful to enable the use of Tensor Core on NVIDIA hardware with compute capability - `>= 7.5` (Volta). - return_attention_mask: - (optional) Set to False to avoid returning attention mask (default: set to model specifics) - """ - # Load from model defaults - bos_token_id = self.sp_tokenizer[self.bos_token] - mask_token_id = self.sp_tokenizer[self.mask_token] - gmask_token_id = self.sp_tokenizer[self.gmask_token] - assert self.padding_side == "left" - - required_input = encoded_inputs[self.model_input_names[0]] - seq_length = len(required_input) - - if padding_strategy == PaddingStrategy.LONGEST: - max_length = len(required_input) - - if max_length is not None and pad_to_multiple_of is not None and (max_length % pad_to_multiple_of != 0): - max_length = ((max_length // pad_to_multiple_of) + 1) * pad_to_multiple_of - - needs_to_be_padded = padding_strategy != PaddingStrategy.DO_NOT_PAD and len(required_input) != max_length - - # Initialize attention mask if not present. - if max_length is not None: - if "attention_mask" not in encoded_inputs: - if bos_token_id in required_input: - context_length = required_input.index(bos_token_id) - else: - context_length = seq_length - attention_mask = np.ones((1, seq_length, seq_length)) - attention_mask = np.tril(attention_mask) - attention_mask[:, :, :context_length] = 1 - attention_mask = np.bool_(attention_mask < 0.5) - encoded_inputs["attention_mask"] = attention_mask - - if "position_ids" not in encoded_inputs: - if bos_token_id in required_input: - context_length = required_input.index(bos_token_id) - else: - context_length = seq_length - position_ids = np.arange(seq_length, dtype=np.int64) - mask_token = mask_token_id if mask_token_id in required_input else gmask_token_id - if mask_token in required_input: - mask_position = required_input.index(mask_token) - position_ids[context_length:] = mask_position - block_position_ids = np.concatenate( - [ - np.zeros(context_length, dtype=np.int64), - np.arange(1, seq_length - context_length + 1, dtype=np.int64), - ] - ) - encoded_inputs["position_ids"] = np.stack([position_ids, block_position_ids], axis=0) - - if needs_to_be_padded: - difference = max_length - len(required_input) - - if "attention_mask" in encoded_inputs: - encoded_inputs["attention_mask"] = np.pad( - encoded_inputs["attention_mask"], - pad_width=[(0, 0), (difference, 0), (difference, 0)], - mode="constant", - constant_values=True, - ) - if "token_type_ids" in encoded_inputs: - encoded_inputs["token_type_ids"] = [self.pad_token_type_id] * difference + encoded_inputs[ - "token_type_ids" - ] - if "special_tokens_mask" in encoded_inputs: - encoded_inputs["special_tokens_mask"] = [1] * difference + encoded_inputs["special_tokens_mask"] - if "position_ids" in encoded_inputs: - encoded_inputs["position_ids"] = np.pad( - encoded_inputs["position_ids"], pad_width=[(0, 0), (difference, 0)] - ) - encoded_inputs[self.model_input_names[0]] = [self.pad_token_id] * difference + required_input - - return encoded_inputs diff --git a/applications/Chat/coati/models/chatglm/configuration_chatglm.py b/applications/Chat/coati/models/chatglm/configuration_chatglm.py deleted file mode 100644 index a6d2ccd18715..000000000000 --- a/applications/Chat/coati/models/chatglm/configuration_chatglm.py +++ /dev/null @@ -1,101 +0,0 @@ -""" -This code is copied from https://huggingface.co/THUDM/chatglm-6b/resolve/main/configuration_chatglm.py -""" - -""" ChatGLM model configuration """ - -from transformers.configuration_utils import PretrainedConfig -from transformers.utils import logging - -logger = logging.get_logger(__name__) - - -class ChatGLMConfig(PretrainedConfig): - r""" - This is the configuration class to store the configuration of a [`~ChatGLMModel`]. - It is used to instantiate an ChatGLM model according to the specified arguments, defining the model - architecture. Instantiating a configuration with the defaults will yield a similar configuration to that of - the ChatGLM-6B [THUDM/ChatGLM-6B](https://huggingface.co/THUDM/chatglm-6b) architecture. - - Configuration objects inherit from [`PretrainedConfig`] and can be used - to control the model outputs. Read the documentation from [`PretrainedConfig`] - for more information. - - - Args: - vocab_size (`int`, *optional*, defaults to 150528): - Vocabulary size of the ChatGLM-6B model. Defines the number of different tokens that can be represented by the - `inputs_ids` passed when calling [`~ChatGLMModel`] or - [`~TFChatGLMModel`]. - hidden_size (`int`, *optional*, defaults to 4096): - Dimension of the encoder layers and the pooler layer. - num_hidden_layers (`int`, *optional*, defaults to 28): - Number of hidden layers in the Transformer encoder. - num_attention_heads (`int`, *optional*, defaults to 32): - Number of attention heads for each attention layer in the Transformer encoder. - inner_hidden_size (`int`, *optional*, defaults to 16384): - Dimension of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder. - max_sequence_length (`int`, *optional*, defaults to 512): - The maximum sequence length that this model might ever be used with. - Typically set this to something large just in case (e.g., 512 or 1024 or 2048). - layernorm_epsilon (`float`, *optional*, defaults to 1e-5): - The epsilon used by the layer normalization layers. - use_cache (`bool`, *optional*, defaults to `True`): - Whether the model should return the last key/values attentions (not used by all models). - Example: - - ```python - >>> from configuration_chatglm import ChatGLMConfig - >>> from modeling_chatglm import ChatGLMModel - - >>> # Initializing a ChatGLM-6B THUDM/ChatGLM-6B style configuration - >>> configuration = ChatGLMConfig() - - >>> # Initializing a model from the THUDM/ChatGLM-6B style configuration - >>> model = ChatGLMModel(configuration) - - >>> # Accessing the model configuration - >>> configuration = model.config - ```""" - model_type = "chatglm" - - def __init__( - self, - vocab_size=130528, - hidden_size=4096, - num_layers=28, - num_attention_heads=32, - layernorm_epsilon=1e-5, - use_cache=True, - bos_token_id=130004, - eos_token_id=130005, - mask_token_id=130000, - gmask_token_id=130001, - pad_token_id=3, - max_sequence_length=2048, - inner_hidden_size=16384, - position_encoding_2d=True, - quantization_bit=0, - pre_seq_len=None, - prefix_projection=False, - **kwargs, - ): - self.num_layers = num_layers - self.vocab_size = vocab_size - self.hidden_size = hidden_size - self.num_attention_heads = num_attention_heads - self.max_sequence_length = max_sequence_length - self.layernorm_epsilon = layernorm_epsilon - self.inner_hidden_size = inner_hidden_size - self.use_cache = use_cache - self.bos_token_id = bos_token_id - self.eos_token_id = eos_token_id - self.pad_token_id = pad_token_id - self.mask_token_id = mask_token_id - self.gmask_token_id = gmask_token_id - self.position_encoding_2d = position_encoding_2d - self.quantization_bit = quantization_bit - self.pre_seq_len = pre_seq_len - self.prefix_projection = prefix_projection - - super().__init__(pad_token_id=pad_token_id, bos_token_id=bos_token_id, eos_token_id=eos_token_id, **kwargs) diff --git a/applications/Chat/coati/models/chatglm/modeling_chatglm.py b/applications/Chat/coati/models/chatglm/modeling_chatglm.py deleted file mode 100644 index d1d15c68ffd8..000000000000 --- a/applications/Chat/coati/models/chatglm/modeling_chatglm.py +++ /dev/null @@ -1,1477 +0,0 @@ -""" -This code is copied from https://huggingface.co/THUDM/chatglm-6b/resolve/main/modeling_chatglm.py -""" - -""" PyTorch ChatGLM model. """ - -import copy -import math -import os -import re -import sys -import warnings -from typing import Any, Callable, Dict, List, Optional, Tuple, Union - -import torch -import torch.nn.functional as F -import torch.utils.checkpoint -from torch import nn -from torch.nn import CrossEntropyLoss, LayerNorm -from torch.nn.utils import skip_init -from transformers.generation.logits_process import LogitsProcessor -from transformers.generation.utils import GenerationConfig, LogitsProcessorList, ModelOutput, StoppingCriteriaList -from transformers.modeling_outputs import ( - BaseModelOutputWithPast, - BaseModelOutputWithPastAndCrossAttentions, - CausalLMOutputWithPast, -) -from transformers.modeling_utils import PreTrainedModel -from transformers.utils import ( - add_code_sample_docstrings, - add_start_docstrings, - add_start_docstrings_to_model_forward, - logging, -) - -from .configuration_chatglm import ChatGLMConfig - -# flags required to enable jit fusion kernels - -if sys.platform != "darwin": - torch._C._jit_set_profiling_mode(False) - torch._C._jit_set_profiling_executor(False) - torch._C._jit_override_can_fuse_on_cpu(True) - torch._C._jit_override_can_fuse_on_gpu(True) - -logger = logging.get_logger(__name__) - -_CHECKPOINT_FOR_DOC = "THUDM/ChatGLM-6B" -_CONFIG_FOR_DOC = "ChatGLM6BConfig" - -CHATGLM_6B_PRETRAINED_MODEL_ARCHIVE_LIST = [ - "THUDM/chatglm-6b", - # See all ChatGLM-6B models at https://huggingface.co/models?filter=chatglm -] - - -class InvalidScoreLogitsProcessor(LogitsProcessor): - def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor) -> torch.FloatTensor: - if torch.isnan(scores).any() or torch.isinf(scores).any(): - scores.zero_() - scores[..., 5] = 5e4 - return scores - - -def load_tf_weights_in_chatglm_6b(model, config, tf_checkpoint_path): - """Load tf checkpoints in a pytorch model.""" - try: - import re - - import numpy as np - import tensorflow as tf - except ImportError: - logger.error( - "Loading a TensorFlow model in PyTorch, requires TensorFlow to be installed. Please see " - "https://www.tensorflow.org/install/ for installation instructions." - ) - raise - tf_path = os.path.abspath(tf_checkpoint_path) - logger.info(f"Converting TensorFlow checkpoint from {tf_path}") - # Load weights from TF model - init_vars = tf.train.list_variables(tf_path) - names = [] - arrays = [] - for name, shape in init_vars: - logger.info(f"Loading TF weight {name} with shape {shape}") - array = tf.train.load_variable(tf_path, name) - names.append(name) - arrays.append(array) - - for name, array in zip(names, arrays): - name = name.split("/") - # adam_v and adam_m are variables used in AdamWeightDecayOptimizer to calculated m and v - # which are not required for using pretrained model - if any( - n in ["adam_v", "adam_m", "AdamWeightDecayOptimizer", "AdamWeightDecayOptimizer_1", "global_step"] - for n in name - ): - logger.info(f"Skipping {'/'.join(name)}") - continue - pointer = model - for m_name in name: - if re.fullmatch(r"[A-Za-z]+_\d+", m_name): - scope_names = re.split(r"_(\d+)", m_name) - else: - scope_names = [m_name] - if scope_names[0] == "kernel" or scope_names[0] == "gamma": - pointer = getattr(pointer, "weight") - elif scope_names[0] == "output_bias" or scope_names[0] == "beta": - pointer = getattr(pointer, "bias") - elif scope_names[0] == "output_weights": - pointer = getattr(pointer, "weight") - elif scope_names[0] == "squad": - pointer = getattr(pointer, "classifier") - else: - try: - pointer = getattr(pointer, scope_names[0]) - except AttributeError: - logger.info(f"Skipping {'/'.join(name)}") - continue - if len(scope_names) >= 2: - num = int(scope_names[1]) - pointer = pointer[num] - if m_name[-11:] == "_embeddings": - pointer = getattr(pointer, "weight") - elif m_name == "kernel": - array = np.transpose(array) - try: - assert ( - pointer.shape == array.shape - ), f"Pointer shape {pointer.shape} and array shape {array.shape} mismatched" - except AssertionError as e: - e.args += (pointer.shape, array.shape) - raise - logger.info(f"Initialize PyTorch weight {name}") - pointer.data = torch.from_numpy(array) - return model - - -class PrefixEncoder(torch.nn.Module): - """ - The torch.nn model to encode the prefix - Input shape: (batch-size, prefix-length) - Output shape: (batch-size, prefix-length, 2*layers*hidden) - """ - - def __init__(self, config): - super().__init__() - self.prefix_projection = config.prefix_projection - if self.prefix_projection: - # Use a two-layer MLP to encode the prefix - self.embedding = torch.nn.Embedding(config.pre_seq_len, config.hidden_size) - self.trans = torch.nn.Sequential( - torch.nn.Linear(config.hidden_size, config.hidden_size), - torch.nn.Tanh(), - torch.nn.Linear(config.hidden_size, config.num_layers * config.hidden_size * 2), - ) - else: - self.embedding = torch.nn.Embedding(config.pre_seq_len, config.num_layers * config.hidden_size * 2) - - def forward(self, prefix: torch.Tensor): - if self.prefix_projection: - prefix_tokens = self.embedding(prefix) - past_key_values = self.trans(prefix_tokens) - else: - past_key_values = self.embedding(prefix) - return past_key_values - - -@torch.jit.script -def gelu_impl(x): - """OpenAI's gelu implementation.""" - return 0.5 * x * (1.0 + torch.tanh(0.7978845608028654 * x * (1.0 + 0.044715 * x * x))) - - -def gelu(x): - return gelu_impl(x) - - -class RotaryEmbedding(torch.nn.Module): - def __init__(self, dim, base=10000, precision=torch.half, learnable=False): - super().__init__() - inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2).float() / dim)) - inv_freq = inv_freq.half() - self.learnable = learnable - if learnable: - self.inv_freq = torch.nn.Parameter(inv_freq) - self.max_seq_len_cached = None - else: - self.register_buffer("inv_freq", inv_freq) - self.max_seq_len_cached = None - self.cos_cached = None - self.sin_cached = None - self.precision = precision - - def _load_from_state_dict( - self, state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs - ): - pass - - def forward(self, x, seq_dim=1, seq_len=None): - if seq_len is None: - seq_len = x.shape[seq_dim] - if self.max_seq_len_cached is None or (seq_len > self.max_seq_len_cached): - self.max_seq_len_cached = None if self.learnable else seq_len - t = torch.arange(seq_len, device=x.device, dtype=self.inv_freq.dtype) - freqs = torch.einsum("i,j->ij", t, self.inv_freq) - # Different from paper, but it uses a different permutation in order to obtain the same calculation - emb = torch.cat((freqs, freqs), dim=-1).to(x.device) - if self.precision == torch.bfloat16: - emb = emb.float() - - # [sx, 1 (b * np), hn] - cos_cached = emb.cos()[:, None, :] - sin_cached = emb.sin()[:, None, :] - if self.precision == torch.bfloat16: - cos_cached = cos_cached.bfloat16() - sin_cached = sin_cached.bfloat16() - if self.learnable: - return cos_cached, sin_cached - self.cos_cached, self.sin_cached = cos_cached, sin_cached - return self.cos_cached[:seq_len, ...], self.sin_cached[:seq_len, ...] - - def _apply(self, fn): - if self.cos_cached is not None: - self.cos_cached = fn(self.cos_cached) - if self.sin_cached is not None: - self.sin_cached = fn(self.sin_cached) - return super()._apply(fn) - - -def rotate_half(x): - x1, x2 = x[..., : x.shape[-1] // 2], x[..., x.shape[-1] // 2 :] - return torch.cat((-x2, x1), dim=x1.ndim - 1) # dim=-1 triggers a bug in earlier torch versions - - -@torch.jit.script -def apply_rotary_pos_emb_index(q, k, cos, sin, position_id): - # position_id: [sq, b], q, k: [sq, b, np, hn], cos: [sq, 1, hn] -> [sq, b, 1, hn] - cos, sin = F.embedding(position_id, cos.squeeze(1)).unsqueeze(2), F.embedding( - position_id, sin.squeeze(1) - ).unsqueeze(2) - q, k = (q * cos) + (rotate_half(q) * sin), (k * cos) + (rotate_half(k) * sin) - return q, k - - -def attention_fn( - self, - query_layer, - key_layer, - value_layer, - attention_mask, - hidden_size_per_partition, - layer_id, - layer_past=None, - scaling_attention_score=True, - use_cache=False, -): - if layer_past is not None: - past_key, past_value = layer_past[0], layer_past[1] - key_layer = torch.cat((past_key, key_layer), dim=0) - value_layer = torch.cat((past_value, value_layer), dim=0) - - # seqlen, batch, num_attention_heads, hidden_size_per_attention_head - seq_len, b, nh, hidden_size = key_layer.shape - - if use_cache: - present = (key_layer, value_layer) - else: - present = None - - query_key_layer_scaling_coeff = float(layer_id + 1) - if scaling_attention_score: - query_layer = query_layer / (math.sqrt(hidden_size) * query_key_layer_scaling_coeff) - - # =================================== - # Raw attention scores. [b, np, s, s] - # =================================== - - # [b, np, sq, sk] - output_size = (query_layer.size(1), query_layer.size(2), query_layer.size(0), key_layer.size(0)) - - # [sq, b, np, hn] -> [sq, b * np, hn] - query_layer = query_layer.view(output_size[2], output_size[0] * output_size[1], -1) - # [sk, b, np, hn] -> [sk, b * np, hn] - key_layer = key_layer.view(output_size[3], output_size[0] * output_size[1], -1) - - matmul_result = torch.zeros( - 1, - 1, - 1, - dtype=query_layer.dtype, - device=query_layer.device, - ) - - matmul_result = torch.baddbmm( - matmul_result, - query_layer.transpose(0, 1), # [b * np, sq, hn] - key_layer.transpose(0, 1).transpose(1, 2), # [b * np, hn, sk] - beta=0.0, - alpha=1.0, - ) - - # change view to [b, np, sq, sk] - attention_scores = matmul_result.view(*output_size) - - if self.scale_mask_softmax: - self.scale_mask_softmax.scale = query_key_layer_scaling_coeff - attention_probs = self.scale_mask_softmax(attention_scores, attention_mask.contiguous()) - else: - if not (attention_mask == 0).all(): - # if auto-regressive, skip - attention_scores.masked_fill_(attention_mask, -10000.0) - dtype = attention_scores.dtype - attention_scores = attention_scores.float() - attention_scores = attention_scores * query_key_layer_scaling_coeff - - attention_probs = F.softmax(attention_scores, dim=-1) - - attention_probs = attention_probs.type(dtype) - - # ========================= - # Context layer. [sq, b, hp] - # ========================= - - # value_layer -> context layer. - # [sk, b, np, hn] --> [b, np, sq, hn] - - # context layer shape: [b, np, sq, hn] - output_size = (value_layer.size(1), value_layer.size(2), query_layer.size(0), value_layer.size(3)) - - # change view [sk, b * np, hn] - value_layer = value_layer.view(value_layer.size(0), output_size[0] * output_size[1], -1) - - # change view [b * np, sq, sk] - attention_probs = attention_probs.view(output_size[0] * output_size[1], output_size[2], -1) - - # matmul: [b * np, sq, hn] - context_layer = torch.bmm(attention_probs, value_layer.transpose(0, 1)) - - # change view [b, np, sq, hn] - context_layer = context_layer.view(*output_size) - - # [b, np, sq, hn] --> [sq, b, np, hn] - context_layer = context_layer.permute(2, 0, 1, 3).contiguous() - - # [sq, b, np, hn] --> [sq, b, hp] - new_context_layer_shape = context_layer.size()[:-2] + (hidden_size_per_partition,) - context_layer = context_layer.view(*new_context_layer_shape) - - outputs = (context_layer, present, attention_probs) - - return outputs - - -def default_init(cls, *args, **kwargs): - return cls(*args, **kwargs) - - -class SelfAttention(torch.nn.Module): - def __init__( - self, - hidden_size, - num_attention_heads, - layer_id, - hidden_size_per_attention_head=None, - bias=True, - params_dtype=torch.float, - position_encoding_2d=True, - empty_init=True, - ): - if empty_init: - init_method = skip_init - else: - init_method = default_init - super(SelfAttention, self).__init__() - - self.layer_id = layer_id - self.hidden_size = hidden_size - self.hidden_size_per_partition = hidden_size - self.num_attention_heads = num_attention_heads - self.num_attention_heads_per_partition = num_attention_heads - self.position_encoding_2d = position_encoding_2d - self.rotary_emb = RotaryEmbedding( - self.hidden_size // (self.num_attention_heads * 2) - if position_encoding_2d - else self.hidden_size // self.num_attention_heads, - base=10000, - precision=torch.half, - learnable=False, - ) - - self.scale_mask_softmax = None - - if hidden_size_per_attention_head is None: - self.hidden_size_per_attention_head = hidden_size // num_attention_heads - else: - self.hidden_size_per_attention_head = hidden_size_per_attention_head - - self.inner_hidden_size = num_attention_heads * self.hidden_size_per_attention_head - - # Strided linear layer. - self.query_key_value = init_method( - torch.nn.Linear, - hidden_size, - 3 * self.inner_hidden_size, - bias=bias, - dtype=params_dtype, - ) - - self.dense = init_method( - torch.nn.Linear, - self.inner_hidden_size, - hidden_size, - bias=bias, - dtype=params_dtype, - ) - - @staticmethod - def attention_mask_func(attention_scores, attention_mask): - attention_scores.masked_fill_(attention_mask, -10000.0) - return attention_scores - - def split_tensor_along_last_dim(self, tensor, num_partitions, contiguous_split_chunks=False): - """Split a tensor along its last dimension. - Arguments: - tensor: input tensor. - num_partitions: number of partitions to split the tensor - contiguous_split_chunks: If True, make each chunk contiguous - in memory. - """ - # Get the size and dimension. - last_dim = tensor.dim() - 1 - last_dim_size = tensor.size()[last_dim] // num_partitions - # Split. - tensor_list = torch.split(tensor, last_dim_size, dim=last_dim) - # Note: torch.split does not create contiguous tensors by default. - if contiguous_split_chunks: - return tuple(chunk.contiguous() for chunk in tensor_list) - - return tensor_list - - def forward( - self, - hidden_states: torch.Tensor, - position_ids, - attention_mask: torch.Tensor, - layer_id, - layer_past: Optional[Tuple[torch.Tensor, torch.Tensor]] = None, - use_cache: bool = False, - output_attentions: bool = False, - ): - """ - hidden_states: [seq_len, batch, hidden_size] - attention_mask: [(1, 1), seq_len, seq_len] - """ - - # [seq_len, batch, 3 * hidden_size] - mixed_raw_layer = self.query_key_value(hidden_states) - - # [seq_len, batch, 3 * hidden_size] --> [seq_len, batch, num_attention_heads, 3 * hidden_size_per_attention_head] - new_tensor_shape = mixed_raw_layer.size()[:-1] + ( - self.num_attention_heads_per_partition, - 3 * self.hidden_size_per_attention_head, - ) - mixed_raw_layer = mixed_raw_layer.view(*new_tensor_shape) - - # [seq_len, batch, num_attention_heads, hidden_size_per_attention_head] - (query_layer, key_layer, value_layer) = self.split_tensor_along_last_dim(mixed_raw_layer, 3) - - if self.position_encoding_2d: - q1, q2 = query_layer.chunk(2, dim=(query_layer.ndim - 1)) - k1, k2 = key_layer.chunk(2, dim=(key_layer.ndim - 1)) - cos, sin = self.rotary_emb(q1, seq_len=position_ids.max() + 1) - position_ids, block_position_ids = ( - position_ids[:, 0, :].transpose(0, 1).contiguous(), - position_ids[:, 1, :].transpose(0, 1).contiguous(), - ) - q1, k1 = apply_rotary_pos_emb_index(q1, k1, cos, sin, position_ids) - q2, k2 = apply_rotary_pos_emb_index(q2, k2, cos, sin, block_position_ids) - query_layer = torch.concat([q1, q2], dim=(q1.ndim - 1)) - key_layer = torch.concat([k1, k2], dim=(k1.ndim - 1)) - else: - position_ids = position_ids.transpose(0, 1) - cos, sin = self.rotary_emb(value_layer, seq_len=position_ids.max() + 1) - # [seq_len, batch, num_attention_heads, hidden_size_per_attention_head] - query_layer, key_layer = apply_rotary_pos_emb_index(query_layer, key_layer, cos, sin, position_ids) - - # [seq_len, batch, hidden_size] - context_layer, present, attention_probs = attention_fn( - self=self, - query_layer=query_layer, - key_layer=key_layer, - value_layer=value_layer, - attention_mask=attention_mask, - hidden_size_per_partition=self.hidden_size_per_partition, - layer_id=layer_id, - layer_past=layer_past, - use_cache=use_cache, - ) - - output = self.dense(context_layer) - - outputs = (output, present) - - if output_attentions: - outputs += (attention_probs,) - - return outputs # output, present, attention_probs - - -class GEGLU(torch.nn.Module): - def __init__(self): - super().__init__() - self.activation_fn = F.gelu - - def forward(self, x): - # dim=-1 breaks in jit for pt<1.10 - x1, x2 = x.chunk(2, dim=(x.ndim - 1)) - return x1 * self.activation_fn(x2) - - -class GLU(torch.nn.Module): - def __init__( - self, - hidden_size, - inner_hidden_size=None, - layer_id=None, - bias=True, - activation_func=gelu, - params_dtype=torch.float, - empty_init=True, - ): - super(GLU, self).__init__() - if empty_init: - init_method = skip_init - else: - init_method = default_init - self.layer_id = layer_id - self.activation_func = activation_func - - # Project to 4h. - self.hidden_size = hidden_size - if inner_hidden_size is None: - inner_hidden_size = 4 * hidden_size - self.inner_hidden_size = inner_hidden_size - self.dense_h_to_4h = init_method( - torch.nn.Linear, - self.hidden_size, - self.inner_hidden_size, - bias=bias, - dtype=params_dtype, - ) - # Project back to h. - self.dense_4h_to_h = init_method( - torch.nn.Linear, - self.inner_hidden_size, - self.hidden_size, - bias=bias, - dtype=params_dtype, - ) - - def forward(self, hidden_states): - """ - hidden_states: [seq_len, batch, hidden_size] - """ - - # [seq_len, batch, inner_hidden_size] - intermediate_parallel = self.dense_h_to_4h(hidden_states) - - intermediate_parallel = self.activation_func(intermediate_parallel) - - output = self.dense_4h_to_h(intermediate_parallel) - - return output - - -class GLMBlock(torch.nn.Module): - def __init__( - self, - hidden_size, - num_attention_heads, - layernorm_epsilon, - layer_id, - inner_hidden_size=None, - hidden_size_per_attention_head=None, - layernorm=LayerNorm, - use_bias=True, - params_dtype=torch.float, - num_layers=28, - position_encoding_2d=True, - empty_init=True, - ): - super(GLMBlock, self).__init__() - # Set output layer initialization if not provided. - - self.layer_id = layer_id - - # Layernorm on the input data. - self.input_layernorm = layernorm(hidden_size, eps=layernorm_epsilon) - - self.position_encoding_2d = position_encoding_2d - - # Self attention. - self.attention = SelfAttention( - hidden_size, - num_attention_heads, - layer_id, - hidden_size_per_attention_head=hidden_size_per_attention_head, - bias=use_bias, - params_dtype=params_dtype, - position_encoding_2d=self.position_encoding_2d, - empty_init=empty_init, - ) - - # Layernorm on the input data. - self.post_attention_layernorm = layernorm(hidden_size, eps=layernorm_epsilon) - - self.num_layers = num_layers - - # GLU - self.mlp = GLU( - hidden_size, - inner_hidden_size=inner_hidden_size, - bias=use_bias, - layer_id=layer_id, - params_dtype=params_dtype, - empty_init=empty_init, - ) - - def forward( - self, - hidden_states: torch.Tensor, - position_ids, - attention_mask: torch.Tensor, - layer_id, - layer_past: Optional[Tuple[torch.Tensor, torch.Tensor]] = None, - use_cache: bool = False, - output_attentions: bool = False, - ): - """ - hidden_states: [seq_len, batch, hidden_size] - attention_mask: [(1, 1), seq_len, seq_len] - """ - - # Layer norm at the begining of the transformer layer. - # [seq_len, batch, hidden_size] - attention_input = self.input_layernorm(hidden_states) - - # Self attention. - attention_outputs = self.attention( - attention_input, - position_ids, - attention_mask=attention_mask, - layer_id=layer_id, - layer_past=layer_past, - use_cache=use_cache, - output_attentions=output_attentions, - ) - - attention_output = attention_outputs[0] - - outputs = attention_outputs[1:] - - # Residual connection. - alpha = (2 * self.num_layers) ** 0.5 - hidden_states = attention_input * alpha + attention_output - - mlp_input = self.post_attention_layernorm(hidden_states) - - # MLP. - mlp_output = self.mlp(mlp_input) - - # Second residual connection. - output = mlp_input * alpha + mlp_output - - if use_cache: - outputs = (output,) + outputs - else: - outputs = (output,) + outputs[1:] - - return outputs # hidden_states, present, attentions - - -class ChatGLMPreTrainedModel(PreTrainedModel): - """ - An abstract class to handle weights initialization and - a simple interface for downloading and loading pretrained models. - """ - - is_parallelizable = False - supports_gradient_checkpointing = True - config_class = ChatGLMConfig - base_model_prefix = "transformer" - _no_split_modules = ["GLMBlock"] - - def __init__(self, *inputs, **kwargs): - super().__init__(*inputs, **kwargs) - - def _init_weights(self, module: nn.Module): - """Initialize the weights.""" - return - - def get_masks(self, input_ids, device): - batch_size, seq_length = input_ids.shape - context_lengths = [seq.tolist().index(self.config.bos_token_id) for seq in input_ids] - attention_mask = torch.ones((batch_size, seq_length, seq_length), device=device) - attention_mask.tril_() - for i, context_length in enumerate(context_lengths): - attention_mask[i, :, :context_length] = 1 - attention_mask.unsqueeze_(1) - attention_mask = (attention_mask < 0.5).bool() - - return attention_mask - - def get_position_ids(self, input_ids, mask_positions, device, use_gmasks=None): - batch_size, seq_length = input_ids.shape - if use_gmasks is None: - use_gmasks = [False] * batch_size - context_lengths = [seq.tolist().index(self.config.bos_token_id) for seq in input_ids] - if self.position_encoding_2d: - position_ids = torch.arange(seq_length, dtype=torch.long, device=device).unsqueeze(0).repeat(batch_size, 1) - for i, context_length in enumerate(context_lengths): - position_ids[i, context_length:] = mask_positions[i] - block_position_ids = [ - torch.cat( - ( - torch.zeros(context_length, dtype=torch.long, device=device), - torch.arange(seq_length - context_length, dtype=torch.long, device=device) + 1, - ) - ) - for context_length in context_lengths - ] - block_position_ids = torch.stack(block_position_ids, dim=0) - position_ids = torch.stack((position_ids, block_position_ids), dim=1) - else: - position_ids = torch.arange(seq_length, dtype=torch.long, device=device).unsqueeze(0).repeat(batch_size, 1) - for i, context_length in enumerate(context_lengths): - if not use_gmasks[i]: - position_ids[i, context_length:] = mask_positions[i] - - return position_ids - - def _set_gradient_checkpointing(self, module, value=False): - if isinstance(module, ChatGLMModel): - module.gradient_checkpointing = value - - -CHATGLM_6B_START_DOCSTRING = r""" - This model is a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) sub-class. - Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general - usage and behavior. - - Parameters: - config ([`~ChatGLM6BConfig`]): Model configuration class with all the parameters of the model. - Initializing with a config file does not load the weights associated with the model, only the configuration. - Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights. -""" - -CHATGLM_6B_INPUTS_DOCSTRING = r""" - Args: - input_ids (`torch.LongTensor` of shape `({0})`): - Indices of input sequence tokens in the vocabulary. - - Indices can be obtained using [`ChatGLM6BTokenizer`]. - See [`PreTrainedTokenizer.encode`] and - [`PreTrainedTokenizer.__call__`] for details. - - [What are input IDs?](../glossary#input-ids) - attention_mask (`torch.FloatTensor` of shape `({0})`, *optional*): - Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`: - - - 1 for tokens that are **not masked**, - - 0 for tokens that are **masked**. - - [What are attention masks?](../glossary#attention-mask) - token_type_ids (`torch.LongTensor` of shape `({0})`, *optional*): - Segment token indices to indicate first and second portions of the inputs. Indices are selected in `[0, 1]`: - - - 0 corresponds to a *sentence A* token, - - 1 corresponds to a *sentence B* token. - - [What are token type IDs?](../glossary#token-type-ids) - position_ids (`torch.LongTensor` of shape `({0})`, *optional*): - Indices of positions of each input sequence tokens in the position embeddings. - Selected in the range `[0, config.max_position_embeddings - 1]`. - - [What are position IDs?](../glossary#position-ids) - head_mask (`torch.FloatTensor` of shape `(num_heads,)` or `(num_layers, num_heads)`, *optional*): - Mask to nullify selected heads of the self-attention modules. Mask values selected in `[0, 1]`: - - - 1 indicates the head is **not masked**, - - 0 indicates the head is **masked**. - - inputs_embeds (`torch.FloatTensor` of shape `({0}, hidden_size)`, *optional*): - Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. - This is useful if you want more control over how to convert *input_ids* indices into associated vectors - than the model's internal embedding lookup matrix. - output_attentions (`bool`, *optional*): - Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned - tensors for more detail. - output_hidden_states (`bool`, *optional*): - Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for - more detail. - return_dict (`bool`, *optional*): - Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple. -""" - - -@add_start_docstrings( - "The bare ChatGLM-6B Model transformer outputting raw hidden-states without any specific head on top.", - CHATGLM_6B_START_DOCSTRING, -) -class ChatGLMModel(ChatGLMPreTrainedModel): - """ - - The model can behave as an encoder (with only self-attention) as well - as a decoder, in which case a layer of cross-attention is added between - the self-attention layers, following the architecture described in [Attention is - all you need](https://arxiv.org/abs/1706.03762) by Ashish Vaswani, - Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser and Illia Polosukhin. - - To behave as an decoder the model needs to be initialized with the - `is_decoder` argument of the configuration set to `True`. - To be used in a Seq2Seq model, the model needs to initialized with both `is_decoder` - argument and `add_cross_attention` set to `True`; an - `encoder_hidden_states` is then expected as an input to the forward pass. - """ - - def __init__(self, config: ChatGLMConfig, empty_init=True): - super().__init__(config) - if empty_init: - init_method = skip_init - else: - init_method = default_init - # recording parameters - self.max_sequence_length = config.max_sequence_length - self.hidden_size = config.hidden_size - self.params_dtype = torch.half - self.num_attention_heads = config.num_attention_heads - self.vocab_size = config.vocab_size - self.num_layers = config.num_layers - self.layernorm_epsilon = config.layernorm_epsilon - self.inner_hidden_size = config.inner_hidden_size - self.hidden_size_per_attention_head = self.hidden_size // self.num_attention_heads - self.position_encoding_2d = config.position_encoding_2d - self.pre_seq_len = config.pre_seq_len - self.prefix_projection = config.prefix_projection - - self.word_embeddings = init_method( - torch.nn.Embedding, num_embeddings=self.vocab_size, embedding_dim=self.hidden_size, dtype=self.params_dtype - ) - self.gradient_checkpointing = False - - def get_layer(layer_id): - return GLMBlock( - self.hidden_size, - self.num_attention_heads, - self.layernorm_epsilon, - layer_id, - inner_hidden_size=self.inner_hidden_size, - hidden_size_per_attention_head=self.hidden_size_per_attention_head, - layernorm=LayerNorm, - use_bias=True, - params_dtype=self.params_dtype, - position_encoding_2d=self.position_encoding_2d, - empty_init=empty_init, - ) - - self.layers = torch.nn.ModuleList([get_layer(layer_id) for layer_id in range(self.num_layers)]) - - # Final layer norm before output. - self.final_layernorm = LayerNorm(self.hidden_size, eps=self.layernorm_epsilon) - - if self.pre_seq_len is not None: - for param in self.parameters(): - param.requires_grad = False - self.prefix_tokens = torch.arange(self.pre_seq_len).long() - self.prefix_encoder = PrefixEncoder(config) - self.dropout = torch.nn.Dropout(0.1) - - # total_params = sum(p.numel() for p in self.parameters()) - # trainable_params = sum(p.numel() for p in self.parameters() if p.requires_grad) - # print("Using p-tuning v2: # trainable_params = {} / {}".format(trainable_params, total_params)) - - def get_input_embeddings(self): - return self.word_embeddings - - def set_input_embeddings(self, new_embeddings: torch.Tensor): - self.word_embeddings = new_embeddings - - def get_prompt(self, batch_size, device, dtype=torch.half): - prefix_tokens = self.prefix_tokens.unsqueeze(0).expand(batch_size, -1).to(device) - past_key_values = self.prefix_encoder(prefix_tokens).type(dtype) - past_key_values = past_key_values.view( - batch_size, - self.pre_seq_len, - self.num_layers * 2, - self.num_attention_heads, - self.hidden_size // self.num_attention_heads, - ) - # seq_len, b, nh, hidden_size - past_key_values = self.dropout(past_key_values) - past_key_values = past_key_values.permute([2, 1, 0, 3, 4]).split(2) - # past_key_values = [(v[0], v[1]) for v in past_key_values] - return past_key_values - - @add_start_docstrings_to_model_forward(CHATGLM_6B_INPUTS_DOCSTRING.format("batch_size, sequence_length")) - @add_code_sample_docstrings( - checkpoint=_CHECKPOINT_FOR_DOC, - output_type=BaseModelOutputWithPastAndCrossAttentions, - config_class=_CONFIG_FOR_DOC, - ) - def forward( - self, - input_ids: Optional[torch.LongTensor] = None, - position_ids: Optional[torch.LongTensor] = None, - attention_mask: Optional[torch.Tensor] = None, - past_key_values: Optional[Tuple[Tuple[torch.Tensor, torch.Tensor], ...]] = None, - inputs_embeds: Optional[torch.LongTensor] = None, - use_cache: Optional[bool] = None, - output_attentions: Optional[bool] = None, - output_hidden_states: Optional[bool] = None, - return_dict: Optional[bool] = None, - ) -> Union[Tuple[torch.Tensor, ...], BaseModelOutputWithPast]: - output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions - output_hidden_states = ( - output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states - ) - use_cache = use_cache if use_cache is not None else self.config.use_cache - return_dict = return_dict if return_dict is not None else self.config.use_return_dict - - if self.gradient_checkpointing and self.training: - if use_cache: - logger.warning_once( - "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..." - ) - use_cache = False - - if input_ids is not None and inputs_embeds is not None: - raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time") - elif input_ids is not None: - batch_size, seq_length = input_ids.shape[:2] - elif inputs_embeds is not None: - batch_size, seq_length = inputs_embeds.shape[:2] - else: - raise ValueError("You have to specify either input_ids or inputs_embeds") - - if inputs_embeds is None: - inputs_embeds = self.word_embeddings(input_ids) - - if past_key_values is None: - if self.pre_seq_len is not None: - past_key_values = self.get_prompt( - batch_size=input_ids.shape[0], device=input_ids.device, dtype=inputs_embeds.dtype - ) - else: - past_key_values = tuple([None] * len(self.layers)) - - if attention_mask is None: - attention_mask = self.get_masks(input_ids, device=input_ids.device) - - if position_ids is None: - MASK, gMASK = self.config.mask_token_id, self.config.gmask_token_id - seqs = input_ids.tolist() - - mask_positions, use_gmasks = [], [] - for seq in seqs: - mask_token = gMASK if gMASK in seq else MASK - use_gmask = mask_token == gMASK - mask_positions.append(seq.index(mask_token)) - use_gmasks.append(use_gmask) - - position_ids = self.get_position_ids( - input_ids, mask_positions=mask_positions, device=input_ids.device, use_gmasks=use_gmasks - ) - - if self.pre_seq_len is not None and attention_mask is not None: - prefix_attention_mask = torch.ones(batch_size, 1, input_ids.size(-1), self.pre_seq_len).to( - attention_mask.device - ) - prefix_attention_mask = (prefix_attention_mask < 0.5).bool() - attention_mask = torch.cat((prefix_attention_mask, attention_mask), dim=3) - - # [seq_len, batch, hidden_size] - hidden_states = inputs_embeds.transpose(0, 1) - - presents = () if use_cache else None - all_self_attentions = () if output_attentions else None - all_hidden_states = () if output_hidden_states else None - - if attention_mask is None: - attention_mask = torch.zeros(1, 1, device=input_ids.device).bool() - else: - attention_mask = attention_mask.to(hidden_states.device) - - for i, layer in enumerate(self.layers): - if output_hidden_states: - all_hidden_states = all_hidden_states + (hidden_states,) - layer_past = past_key_values[i] - - if self.gradient_checkpointing and self.training: - layer_ret = torch.utils.checkpoint.checkpoint( - layer, - hidden_states, - position_ids, - attention_mask, - torch.tensor(i), - layer_past, - use_cache, - output_attentions, - ) - else: - layer_ret = layer( - hidden_states, - position_ids=position_ids, - attention_mask=attention_mask, - layer_id=torch.tensor(i), - layer_past=layer_past, - use_cache=use_cache, - output_attentions=output_attentions, - ) - - hidden_states = layer_ret[0] - - if use_cache: - presents = presents + (layer_ret[1],) - - if output_attentions: - all_self_attentions = all_self_attentions + (layer_ret[2 if use_cache else 1],) - - # Final layer norm. - hidden_states = self.final_layernorm(hidden_states) - - if output_hidden_states: - all_hidden_states = all_hidden_states + (hidden_states,) - - if not return_dict: - return tuple(v for v in [hidden_states, presents, all_hidden_states, all_self_attentions] if v is not None) - - return BaseModelOutputWithPast( - last_hidden_state=hidden_states, - past_key_values=presents, - hidden_states=all_hidden_states, - attentions=all_self_attentions, - ) - - -class ChatGLMForConditionalGeneration(ChatGLMPreTrainedModel): - def __init__(self, config: ChatGLMConfig, empty_init=True): - super().__init__(config) - if empty_init: - init_method = skip_init - else: - init_method = default_init - - # self.hidden_size = config.hidden_size - # self.params_dtype = torch.half - # self.vocab_size = config.vocab_size - self.max_sequence_length = config.max_sequence_length - - self.position_encoding_2d = config.position_encoding_2d - - self.transformer = ChatGLMModel(config, empty_init=empty_init) - - self.lm_head = init_method(nn.Linear, config.hidden_size, config.vocab_size, bias=False, dtype=torch.half) - - self.config = config - - self.quantized = False - - if self.config.quantization_bit: - self.quantize(self.config.quantization_bit, empty_init=True) - - def get_output_embeddings(self): - return self.lm_head - - def set_output_embeddings(self, new_embeddings): - self.lm_head = new_embeddings - - def _update_model_kwargs_for_generation( - self, - outputs: ModelOutput, - model_kwargs: Dict[str, Any], - is_encoder_decoder: bool = False, - standardize_cache_format: bool = False, - ) -> Dict[str, Any]: - # update past_key_values - model_kwargs["past_key_values"] = self._extract_past_from_model_output( - outputs, standardize_cache_format=standardize_cache_format - ) - - # update attention mask - if "attention_mask" in model_kwargs: - attention_mask = model_kwargs["attention_mask"] - if attention_mask is not None and attention_mask.dtype == torch.bool: - attention_mask = torch.cat( - [attention_mask, attention_mask.new_ones((*attention_mask.shape[:3], 1))], dim=3 - ) - new_attention_mask = attention_mask[:, :, -1:].clone() - new_attention_mask[..., -1] = False - model_kwargs["attention_mask"] = torch.cat([attention_mask, new_attention_mask], dim=2) - - # update position ids - if "position_ids" in model_kwargs: - position_ids = model_kwargs["position_ids"] - new_position_id = position_ids[..., -1:].clone() - new_position_id[:, 1, :] += 1 - model_kwargs["position_ids"] = torch.cat([position_ids, new_position_id], dim=-1) - - return model_kwargs - - def prepare_inputs_for_generation( - self, - input_ids: torch.LongTensor, - past: Optional[torch.Tensor] = None, - past_key_values: Optional[torch.Tensor] = None, - attention_mask: Optional[torch.Tensor] = None, - position_ids: Optional[torch.Tensor] = None, - **kwargs, - ) -> dict: - batch_size, seq_length = input_ids.shape - MASK, gMASK = self.config.mask_token_id, self.config.gmask_token_id - seqs = input_ids.tolist() - mask_positions, use_gmasks = [], [] - for seq in seqs: - mask_token = gMASK if gMASK in seq else MASK - use_gmask = mask_token == gMASK - mask_positions.append(seq.index(mask_token)) - use_gmasks.append(use_gmask) - - # only last token for input_ids if past is not None - if past is not None or past_key_values is not None: - last_token = input_ids[:, -1].unsqueeze(-1) - if attention_mask is not None and attention_mask.dtype == torch.bool: - attention_mask = attention_mask[:, :, -1:] - else: - attention_mask = None - if position_ids is not None: - position_ids = position_ids[..., -1:] - else: - context_lengths = [seq.index(self.config.bos_token_id) for seq in seqs] - if self.position_encoding_2d: - position_ids = torch.tensor( - [ - [mask_position, seq_length - context_length] - for mask_position, context_length in zip(mask_positions, context_lengths) - ], - dtype=torch.long, - device=input_ids.device, - ).unsqueeze(-1) - else: - position_ids = torch.tensor( - [mask_position for mask_position in mask_positions], dtype=torch.long, device=input_ids.device - ).unsqueeze(-1) - - if past is None: - past = past_key_values - return { - "input_ids": last_token, - "past_key_values": past, - "position_ids": position_ids, - "attention_mask": attention_mask, - } - else: - if attention_mask is not None and attention_mask.dtype != torch.bool: - logger.warning_once(f"The dtype of attention mask ({attention_mask.dtype}) is not bool") - attention_mask = None - if attention_mask is None: - attention_mask = self.get_masks(input_ids, device=input_ids.device) - if position_ids is None: - position_ids = self.get_position_ids( - input_ids, device=input_ids.device, mask_positions=mask_positions, use_gmasks=use_gmasks - ) - - return { - "input_ids": input_ids, - "past_key_values": past, - "position_ids": position_ids, - "attention_mask": attention_mask, - } - - def forward( - self, - input_ids: Optional[torch.Tensor] = None, - position_ids: Optional[torch.Tensor] = None, - attention_mask: Optional[torch.Tensor] = None, - past_key_values: Optional[Tuple[torch.FloatTensor]] = None, - inputs_embeds: Optional[torch.Tensor] = None, - labels: Optional[torch.Tensor] = None, - use_cache: Optional[bool] = None, - output_attentions: Optional[bool] = None, - output_hidden_states: Optional[bool] = None, - return_dict: Optional[bool] = None, - ): - use_cache = use_cache if use_cache is not None else self.config.use_cache - return_dict = return_dict if return_dict is not None else self.config.use_return_dict - - transformer_outputs = self.transformer( - input_ids=input_ids, - position_ids=position_ids, - attention_mask=attention_mask, - past_key_values=past_key_values, - inputs_embeds=inputs_embeds, - use_cache=use_cache, - output_attentions=output_attentions, - output_hidden_states=output_hidden_states, - return_dict=return_dict, - ) - - hidden_states = transformer_outputs[0] - - lm_logits = self.lm_head(hidden_states).permute(1, 0, 2).contiguous() - - loss = None - if labels is not None: - lm_logits = lm_logits.to(torch.float32) - - # Shift so that tokens < n predict n - shift_logits = lm_logits[..., :-1, :].contiguous() - shift_labels = labels[..., 1:].contiguous() - # Flatten the tokens - loss_fct = CrossEntropyLoss(ignore_index=-100) - loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1)) - - lm_logits = lm_logits.to(hidden_states.dtype) - loss = loss.to(hidden_states.dtype) - - if not return_dict: - output = (lm_logits,) + transformer_outputs[1:] - return ((loss,) + output) if loss is not None else output - - return CausalLMOutputWithPast( - loss=loss, - logits=lm_logits, - past_key_values=transformer_outputs.past_key_values, - hidden_states=transformer_outputs.hidden_states, - attentions=transformer_outputs.attentions, - ) - - @staticmethod - def _reorder_cache( - past: Tuple[Tuple[torch.Tensor, torch.Tensor], ...], beam_idx: torch.LongTensor - ) -> Tuple[Tuple[torch.Tensor, torch.Tensor], ...]: - """ - This function is used to re-order the `past_key_values` cache if [`~PreTrainedModel.beam_search`] or - [`~PreTrainedModel.beam_sample`] is called. This is required to match `past_key_values` with the correct - beam_idx at every generation step. - - Output shares the same memory storage as `past`. - """ - return tuple( - ( - layer_past[0].index_select(1, beam_idx.to(layer_past[0].device)), - layer_past[1].index_select(1, beam_idx.to(layer_past[1].device)), - ) - for layer_past in past - ) - - def process_response(self, response): - response = response.strip() - response = response.replace("[[训练时间]]", "2023年") - punkts = [ - [",", "，"], - ["!", "！"], - [":", "："], - [";", "；"], - ["\?", "？"], - ] - for item in punkts: - response = re.sub(r"([\u4e00-\u9fff])%s" % item[0], r"\1%s" % item[1], response) - response = re.sub(r"%s([\u4e00-\u9fff])" % item[0], r"%s\1" % item[1], response) - return response - - @torch.no_grad() - def chat( - self, - tokenizer, - query: str, - history: List[Tuple[str, str]] = None, - max_length: int = 2048, - num_beams=1, - do_sample=True, - top_p=0.7, - temperature=0.95, - logits_processor=None, - **kwargs, - ): - if history is None: - history = [] - if logits_processor is None: - logits_processor = LogitsProcessorList() - logits_processor.append(InvalidScoreLogitsProcessor()) - gen_kwargs = { - "max_length": max_length, - "num_beams": num_beams, - "do_sample": do_sample, - "top_p": top_p, - "temperature": temperature, - "logits_processor": logits_processor, - **kwargs, - } - if not history: - prompt = query - else: - prompt = "" - for i, (old_query, response) in enumerate(history): - prompt += "[Round {}]\n问：{}\n答：{}\n".format(i, old_query, response) - prompt += "[Round {}]\n问：{}\n答：".format(len(history), query) - inputs = tokenizer([prompt], return_tensors="pt") - inputs = inputs.to(self.device) - outputs = self.generate(**inputs, **gen_kwargs) - outputs = outputs.tolist()[0][len(inputs["input_ids"][0]) :] - response = tokenizer.decode(outputs) - response = self.process_response(response) - history = history + [(query, response)] - return response, history - - @torch.no_grad() - def stream_chat( - self, - tokenizer, - query: str, - history: List[Tuple[str, str]] = None, - max_length: int = 2048, - do_sample=True, - top_p=0.7, - temperature=0.95, - logits_processor=None, - **kwargs, - ): - if history is None: - history = [] - if logits_processor is None: - logits_processor = LogitsProcessorList() - logits_processor.append(InvalidScoreLogitsProcessor()) - gen_kwargs = { - "max_length": max_length, - "do_sample": do_sample, - "top_p": top_p, - "temperature": temperature, - "logits_processor": logits_processor, - **kwargs, - } - if not history: - prompt = query - else: - prompt = "" - for i, (old_query, response) in enumerate(history): - prompt += "[Round {}]\n问：{}\n答：{}\n".format(i, old_query, response) - prompt += "[Round {}]\n问：{}\n答：".format(len(history), query) - inputs = tokenizer([prompt], return_tensors="pt") - inputs = inputs.to(self.device) - for outputs in self.stream_generate(**inputs, **gen_kwargs): - outputs = outputs.tolist()[0][len(inputs["input_ids"][0]) :] - response = tokenizer.decode(outputs) - response = self.process_response(response) - new_history = history + [(query, response)] - yield response, new_history - - @torch.no_grad() - def stream_generate( - self, - input_ids, - generation_config: Optional[GenerationConfig] = None, - logits_processor: Optional[LogitsProcessorList] = None, - stopping_criteria: Optional[StoppingCriteriaList] = None, - prefix_allowed_tokens_fn: Optional[Callable[[int, torch.Tensor], List[int]]] = None, - **kwargs, - ): - batch_size, input_ids_seq_length = input_ids.shape[0], input_ids.shape[-1] - - if generation_config is None: - generation_config = self.generation_config - generation_config = copy.deepcopy(generation_config) - model_kwargs = generation_config.update(**kwargs) - bos_token_id, eos_token_id = generation_config.bos_token_id, generation_config.eos_token_id - - if isinstance(eos_token_id, int): - eos_token_id = [eos_token_id] - - has_default_max_length = kwargs.get("max_length") is None and generation_config.max_length is not None - if has_default_max_length and generation_config.max_new_tokens is None: - warnings.warn( - f"Using `max_length`'s default ({generation_config.max_length}) to control the generation length. " - "This behaviour is deprecated and will be removed from the config in v5 of Transformers -- we" - " recommend using `max_new_tokens` to control the maximum length of the generation.", - UserWarning, - ) - elif generation_config.max_new_tokens is not None: - generation_config.max_length = generation_config.max_new_tokens + input_ids_seq_length - if not has_default_max_length: - logger.warn( - f"Both `max_new_tokens` (={generation_config.max_new_tokens}) and `max_length`(=" - f"{generation_config.max_length}) seem to have been set. `max_new_tokens` will take precedence. " - "Please refer to the documentation for more information. " - "(https://huggingface.co/docs/transformers/main/en/main_classes/text_generation)", - UserWarning, - ) - - if input_ids_seq_length >= generation_config.max_length: - input_ids_string = "decoder_input_ids" if self.config.is_encoder_decoder else "input_ids" - logger.warning( - f"Input length of {input_ids_string} is {input_ids_seq_length}, but `max_length` is set to" - f" {generation_config.max_length}. This can lead to unexpected behavior. You should consider" - " increasing `max_new_tokens`." - ) - - # 2. Set generation parameters if not already defined - logits_processor = logits_processor if logits_processor is not None else LogitsProcessorList() - stopping_criteria = stopping_criteria if stopping_criteria is not None else StoppingCriteriaList() - - logits_processor = self._get_logits_processor( - generation_config=generation_config, - input_ids_seq_length=input_ids_seq_length, - encoder_input_ids=input_ids, - prefix_allowed_tokens_fn=prefix_allowed_tokens_fn, - logits_processor=logits_processor, - ) - - stopping_criteria = self._get_stopping_criteria( - generation_config=generation_config, stopping_criteria=stopping_criteria - ) - logits_warper = self._get_logits_warper(generation_config) - - unfinished_sequences = input_ids.new(input_ids.shape[0]).fill_(1) - scores = None - while True: - model_inputs = self.prepare_inputs_for_generation(input_ids, **model_kwargs) - # forward pass to get next token - outputs = self( - **model_inputs, - return_dict=True, - output_attentions=False, - output_hidden_states=False, - ) - - next_token_logits = outputs.logits[:, -1, :] - - # pre-process distribution - next_token_scores = logits_processor(input_ids, next_token_logits) - next_token_scores = logits_warper(input_ids, next_token_scores) - - # sample - probs = nn.functional.softmax(next_token_scores, dim=-1) - if generation_config.do_sample: - next_tokens = torch.multinomial(probs, num_samples=1).squeeze(1) - else: - next_tokens = torch.argmax(probs, dim=-1) - - # update generated ids, model inputs, and length for next step - input_ids = torch.cat([input_ids, next_tokens[:, None]], dim=-1) - model_kwargs = self._update_model_kwargs_for_generation( - outputs, model_kwargs, is_encoder_decoder=self.config.is_encoder_decoder - ) - unfinished_sequences = unfinished_sequences.mul((sum(next_tokens != i for i in eos_token_id)).long()) - - # stop when each sentence is finished, or if we exceed the maximum length - if unfinished_sequences.max() == 0 or stopping_criteria(input_ids, scores): - break - yield input_ids - - def quantize(self, bits: int, empty_init=False, **kwargs): - if bits == 0: - return - - from .quantization import quantize - - if self.quantized: - logger.info("Already quantized.") - return self - - self.quantized = True - - self.config.quantization_bit = bits - - self.transformer = quantize(self.transformer, bits, empty_init=empty_init, **kwargs) - return self diff --git a/applications/Chat/coati/models/generation.py b/applications/Chat/coati/models/generation.py deleted file mode 100644 index 4ab0cdc8a3ea..000000000000 --- a/applications/Chat/coati/models/generation.py +++ /dev/null @@ -1,152 +0,0 @@ -from typing import Any, Callable, Optional - -import torch -import torch.distributed as dist -from transformers import PreTrainedTokenizer - -from .base import Actor - -try: - from transformers.generation_logits_process import ( - LogitsProcessorList, - TemperatureLogitsWarper, - TopKLogitsWarper, - TopPLogitsWarper, - ) -except ImportError: - from transformers.generation import LogitsProcessorList, TemperatureLogitsWarper, TopKLogitsWarper, TopPLogitsWarper - - -def _prepare_logits_processor( - top_k: Optional[int] = None, top_p: Optional[float] = None, temperature: Optional[float] = None -) -> LogitsProcessorList: - processor_list = LogitsProcessorList() - if temperature is not None and temperature != 1.0: - processor_list.append(TemperatureLogitsWarper(temperature)) - if top_k is not None and top_k != 0: - processor_list.append(TopKLogitsWarper(top_k)) - if top_p is not None and top_p < 1.0: - processor_list.append(TopPLogitsWarper(top_p)) - return processor_list - - -def _is_sequence_finished(unfinished_sequences: torch.Tensor) -> bool: - if dist.is_initialized() and dist.get_world_size() > 1: - # consider DP - unfinished_sequences = unfinished_sequences.clone() - dist.all_reduce(unfinished_sequences) - return unfinished_sequences.max() == 0 - - -def _sample( - model: Actor, - input_ids: torch.Tensor, - max_length: int, - early_stopping: bool = False, - eos_token_id: Optional[int] = None, - pad_token_id: Optional[int] = None, - top_k: Optional[int] = None, - top_p: Optional[float] = None, - temperature: Optional[float] = None, - prepare_inputs_fn: Optional[Callable[[torch.Tensor, Any], dict]] = None, - update_model_kwargs_fn: Optional[Callable[[dict, Any], dict]] = None, - **model_kwargs, -) -> torch.Tensor: - if input_ids.size(1) >= max_length: - return input_ids - - logits_processor = _prepare_logits_processor(top_k, top_p, temperature) - unfinished_sequences = input_ids.new(input_ids.shape[0]).fill_(1) - - for _ in range(input_ids.size(1), max_length): - model_inputs = ( - prepare_inputs_fn(input_ids, **model_kwargs) if prepare_inputs_fn is not None else {"input_ids": input_ids} - ) - outputs = model(**model_inputs) - - # NOTE: this is correct only in left padding mode - next_token_logits = outputs["logits"][:, -1, :] - next_token_logits = logits_processor(input_ids, next_token_logits) - # sample - probs = torch.softmax(next_token_logits, dim=-1, dtype=torch.float) - next_tokens = torch.multinomial(probs, num_samples=1).squeeze(1) - - # finished sentences should have their next token be a padding token - if eos_token_id is not None: - assert pad_token_id is not None, "If `eos_token_id` is defined, make sure that `pad_token_id` is defined." - next_tokens = next_tokens * unfinished_sequences + pad_token_id * (1 - unfinished_sequences) - - # update generated ids, model inputs for next step - input_ids = torch.cat([input_ids, next_tokens[:, None]], dim=-1) - if update_model_kwargs_fn is not None: - model_kwargs = update_model_kwargs_fn(outputs, model_kwargs) - - # if eos_token was found in one sentence, set sentence to finished - if eos_token_id is not None: - unfinished_sequences = unfinished_sequences.mul((next_tokens != eos_token_id).long()) - - # stop when each sentence is finished if early_stopping=True - if early_stopping and _is_sequence_finished(unfinished_sequences): - break - - return input_ids - - -@torch.no_grad() -def generate( - model: Actor, - input_ids: torch.Tensor, - tokenizer: PreTrainedTokenizer, - max_length: int, - num_beams: int = 1, - do_sample: bool = True, - early_stopping: bool = False, - top_k: Optional[int] = None, - top_p: Optional[float] = None, - temperature: Optional[float] = None, - prepare_inputs_fn: Optional[Callable[[torch.Tensor, Any], dict]] = None, - update_model_kwargs_fn: Optional[Callable[[dict, Any], dict]] = None, - **model_kwargs, -) -> torch.Tensor: - """Generate token sequence. The returned sequence is input_ids + generated_tokens. - - Args: - model (nn.Module): model - input_ids (torch.Tensor): input sequence - max_length (int): max length of the returned sequence - num_beams (int, optional): number of beams. Defaults to 1. - do_sample (bool, optional): whether to do sample. Defaults to True. - early_stopping (bool, optional): if True, the sequence length may be smaller than max_length due to finding eos. Defaults to False. - top_k (Optional[int], optional): the number of highest probability vocabulary tokens to keep for top-k-filtering. Defaults to None. - top_p (Optional[float], optional): If set to float < 1, only the smallest set of most probable tokens with probabilities that add up to top_p or higher are kept for generation. Defaults to None. - temperature (Optional[float], optional): The value used to module the next token probabilities. Defaults to None. - prepare_inputs_fn (Optional[Callable[[torch.Tensor, Any], dict]], optional): Function to preprocess model inputs. Arguments of this function should be input_ids and model_kwargs. Defaults to None. - update_model_kwargs_fn (Optional[Callable[[dict, Any], dict]], optional): Function to update model_kwargs based on outputs. Arguments of this function should be outputs and model_kwargs. Defaults to None. - """ - assert tokenizer.padding_side == "left", "Current generation only supports left padding." - is_greedy_gen_mode = (num_beams == 1) and do_sample is False - is_sample_gen_mode = (num_beams == 1) and do_sample is True - is_beam_gen_mode = (num_beams > 1) and do_sample is False - if is_greedy_gen_mode: - # run greedy search - raise NotImplementedError - elif is_sample_gen_mode: - # run sample - return _sample( - model, - input_ids, - max_length, - early_stopping=early_stopping, - eos_token_id=tokenizer.eos_token_id, - pad_token_id=tokenizer.pad_token_id, - top_k=top_k, - top_p=top_p, - temperature=temperature, - prepare_inputs_fn=prepare_inputs_fn, - update_model_kwargs_fn=update_model_kwargs_fn, - **model_kwargs, - ) - elif is_beam_gen_mode: - raise NotImplementedError - else: - raise ValueError("Unsupported generation mode") diff --git a/applications/Chat/coati/models/gpt/__init__.py b/applications/Chat/coati/models/gpt/__init__.py deleted file mode 100644 index 823cf4a75e0d..000000000000 --- a/applications/Chat/coati/models/gpt/__init__.py +++ /dev/null @@ -1,5 +0,0 @@ -from .gpt_actor import GPTActor -from .gpt_critic import GPTCritic -from .gpt_rm import GPTRM - -__all__ = ["GPTActor", "GPTCritic", "GPTRM"] diff --git a/applications/Chat/coati/models/gpt/gpt_actor.py b/applications/Chat/coati/models/gpt/gpt_actor.py deleted file mode 100644 index a7e4b9bc3e22..000000000000 --- a/applications/Chat/coati/models/gpt/gpt_actor.py +++ /dev/null @@ -1,38 +0,0 @@ -from typing import Optional - -from transformers.models.gpt2.configuration_gpt2 import GPT2Config -from transformers.models.gpt2.modeling_gpt2 import GPT2LMHeadModel - -from ..base import Actor - - -class GPTActor(Actor): - """ - GPT Actor model. - - Args: - pretrained (str): Pretrained model name or path. - config (GPT2Config): Model config. - checkpoint (bool): Enable gradient checkpointing. - lora_rank (int): Rank of the LoRa layer. - lora_train_bias (str): Bias training strategy for the LoRa layer. - """ - - def __init__( - self, - pretrained: Optional[str] = None, - config: Optional[GPT2Config] = None, - checkpoint: bool = False, - lora_rank: int = 0, - lora_train_bias: str = "none", - **kwargs, - ) -> None: - if pretrained is not None: - model = GPT2LMHeadModel.from_pretrained(pretrained) - elif config is not None: - model = GPT2LMHeadModel(config) - else: - model = GPT2LMHeadModel(GPT2Config()) - if checkpoint: - model.gradient_checkpointing_enable() - super().__init__(model, lora_rank, lora_train_bias, **kwargs) diff --git a/applications/Chat/coati/models/gpt/gpt_critic.py b/applications/Chat/coati/models/gpt/gpt_critic.py deleted file mode 100644 index 22ab36dea276..000000000000 --- a/applications/Chat/coati/models/gpt/gpt_critic.py +++ /dev/null @@ -1,37 +0,0 @@ -from typing import Optional - -import torch.nn as nn -from transformers.models.gpt2.configuration_gpt2 import GPT2Config -from transformers.models.gpt2.modeling_gpt2 import GPT2Model - -from ..base import Critic - - -class GPTCritic(Critic): - """ - GPT Critic model. - - Args: - pretrained (str): Pretrained model name or path. - config (GPT2Config): Model config. - lora_rank (int): Rank of the LO-RA decomposition. - lora_train_bias (str): LoRA bias training mode. - """ - - def __init__( - self, - pretrained: Optional[str] = None, - config: Optional[GPT2Config] = None, - lora_rank: int = 0, - lora_train_bias: str = "none", - **kwargs, - ) -> None: - if pretrained is not None: - model = GPT2Model.from_pretrained(pretrained) - elif config is not None: - model = GPT2Model(config) - else: - model = GPT2Model(GPT2Config()) - - value_head = nn.Linear(model.config.n_embd, 1) - super().__init__(model, value_head, lora_rank, lora_train_bias, **kwargs) diff --git a/applications/Chat/coati/models/gpt/gpt_rm.py b/applications/Chat/coati/models/gpt/gpt_rm.py deleted file mode 100644 index 8edfc4008466..000000000000 --- a/applications/Chat/coati/models/gpt/gpt_rm.py +++ /dev/null @@ -1,37 +0,0 @@ -from typing import Optional - -import torch.nn as nn -from transformers.models.gpt2.configuration_gpt2 import GPT2Config -from transformers.models.gpt2.modeling_gpt2 import GPT2Model - -from ..base import RewardModel - - -class GPTRM(RewardModel): - """ - GPT Reward model. - - Args: - pretrained (str): Pretrained model name or path. - config (GPT2Config): Model config. - lora_rank (int): Rank of the low-rank approximation. - lora_train_bias (str): LoRA bias training mode. - """ - - def __init__( - self, - pretrained: Optional[str] = None, - config: Optional[GPT2Config] = None, - lora_rank: int = 0, - lora_train_bias: str = "none", - ) -> None: - if pretrained is not None: - model = GPT2Model.from_pretrained(pretrained) - elif config is not None: - model = GPT2Model(config) - else: - model = GPT2Model(GPT2Config()) - - value_head = nn.Linear(model.config.n_embd, 1) - value_head.weight.data.normal_(mean=0.0, std=1 / (model.config.n_embd + 1)) - super().__init__(model, value_head, lora_rank, lora_train_bias) diff --git a/applications/Chat/coati/models/llama/__init__.py b/applications/Chat/coati/models/llama/__init__.py deleted file mode 100644 index c87d732538a9..000000000000 --- a/applications/Chat/coati/models/llama/__init__.py +++ /dev/null @@ -1,5 +0,0 @@ -from .llama_actor import LlamaActor -from .llama_critic import LlamaCritic -from .llama_rm import LlamaRM - -__all__ = ["LlamaActor", "LlamaCritic", "LlamaRM"] diff --git a/applications/Chat/coati/models/llama/llama_actor.py b/applications/Chat/coati/models/llama/llama_actor.py deleted file mode 100644 index f1d9406835ca..000000000000 --- a/applications/Chat/coati/models/llama/llama_actor.py +++ /dev/null @@ -1,38 +0,0 @@ -from typing import Optional - -from transformers import LlamaConfig, LlamaForCausalLM - -from ..base import Actor - - -class LlamaActor(Actor): - """ - Llama Actor model. - - Args: - pretrained (str): Pretrained model name or path. - config (LlamaConfig): Model config. - checkpoint (bool): Enable gradient checkpointing. - lora_rank (int): LoRA rank. - lora_train_bias (str): LoRA bias training mode. - """ - - def __init__( - self, - pretrained: Optional[str] = None, - config: Optional[LlamaConfig] = None, - checkpoint: bool = False, - lora_rank: int = 0, - lora_train_bias: str = "none", - ) -> None: - if pretrained is not None: - model = LlamaForCausalLM.from_pretrained(pretrained) - elif config is not None: - model = LlamaForCausalLM(config) - else: - model = LlamaForCausalLM(LlamaConfig()) - - if checkpoint: - model.gradient_checkpointing_enable() - - super().__init__(model, lora_rank, lora_train_bias) diff --git a/applications/Chat/coati/models/llama/llama_critic.py b/applications/Chat/coati/models/llama/llama_critic.py deleted file mode 100644 index 000dce17ccf0..000000000000 --- a/applications/Chat/coati/models/llama/llama_critic.py +++ /dev/null @@ -1,36 +0,0 @@ -from typing import Optional - -import torch.nn as nn -from transformers import LlamaConfig, LlamaModel - -from ..base import Critic - - -class LlamaCritic(Critic): - """ - Llama Critic model. - - Args: - pretrained (str): Pretrained model name or path. - config (LlamaConfig): Model config. - lora_rank (int): LoRA rank. - lora_train_bias (str): LoRA bias training mode. - """ - - def __init__( - self, - pretrained: Optional[str] = None, - config: Optional[LlamaConfig] = None, - lora_rank: int = 0, - lora_train_bias: str = "none", - **kwargs, - ) -> None: - if pretrained is not None: - model = LlamaModel.from_pretrained(pretrained) - elif config is not None: - model = LlamaModel(config) - else: - model = LlamaModel(LlamaConfig()) - - value_head = nn.Linear(model.config.hidden_size, 1) - super().__init__(model, value_head, lora_rank, lora_train_bias, **kwargs) diff --git a/applications/Chat/coati/models/llama/llama_rm.py b/applications/Chat/coati/models/llama/llama_rm.py deleted file mode 100644 index 43bc9e638dc7..000000000000 --- a/applications/Chat/coati/models/llama/llama_rm.py +++ /dev/null @@ -1,37 +0,0 @@ -from typing import Optional - -import torch.nn as nn -from transformers import LlamaConfig, LlamaModel - -from ..base import RewardModel - - -class LlamaRM(RewardModel): - """ - Llama Reward model. - - Args: - pretrained (str): Pretrained model name or path. - config (LlamaConfig): Model config. - lora_rank (int): LoRA rank. - lora_train_bias (str): LoRA bias training mode. - """ - - def __init__( - self, - pretrained: Optional[str] = None, - config: Optional[LlamaConfig] = None, - lora_rank: int = 0, - lora_train_bias: str = "none", - ) -> None: - if pretrained is not None: - model = LlamaModel.from_pretrained(pretrained) - elif config is not None: - model = LlamaModel(config) - else: - model = LlamaModel(LlamaConfig()) - - value_head = nn.Linear(model.config.hidden_size, 1) - value_head.weight.data.normal_(mean=0.0, std=1 / (model.config.hidden_size + 1)) - - super().__init__(model, value_head, lora_rank, lora_train_bias) diff --git a/applications/Chat/coati/models/loss.py b/applications/Chat/coati/models/loss.py deleted file mode 100644 index 687bd0f7bfe7..000000000000 --- a/applications/Chat/coati/models/loss.py +++ /dev/null @@ -1,97 +0,0 @@ -from typing import Optional - -import torch -import torch.nn as nn - -from .utils import masked_mean - - -class GPTLMLoss(nn.Module): - """ - GPT Language Model Loss - """ - - def __init__(self): - super().__init__() - # NOTE: default ignore_index is -100, which is equal to IGNORE_INDEX in sft_dataset.py - self.loss = nn.CrossEntropyLoss() - - def forward(self, logits: torch.Tensor, labels: torch.Tensor) -> torch.Tensor: - shift_logits = logits[..., :-1, :].contiguous() - shift_labels = labels[..., 1:].contiguous() - # Flatten the tokens - return self.loss(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1)) - - -class PolicyLoss(nn.Module): - """ - Policy Loss for PPO - """ - - def __init__(self, clip_eps: float = 0.2) -> None: - super().__init__() - self.clip_eps = clip_eps - - def forward( - self, - log_probs: torch.Tensor, - old_log_probs: torch.Tensor, - advantages: torch.Tensor, - action_mask: Optional[torch.Tensor] = None, - ) -> torch.Tensor: - ratio = (log_probs - old_log_probs).exp() - surr1 = ratio * advantages - surr2 = ratio.clamp(1 - self.clip_eps, 1 + self.clip_eps) * advantages - loss = -torch.min(surr1, surr2) - if action_mask is not None: - loss = masked_mean(loss, action_mask) - loss = loss.mean() - return loss - - -class ValueLoss(nn.Module): - """ - Value Loss for PPO - """ - - def __init__(self, clip_eps: float = 0.4) -> None: - super().__init__() - self.clip_eps = clip_eps - - def forward( - self, - values: torch.Tensor, - old_values: torch.Tensor, - reward: torch.Tensor, - action_mask: Optional[torch.Tensor] = None, - ) -> torch.Tensor: - values_clipped = old_values + (values - old_values).clamp(-self.clip_eps, self.clip_eps) - surr1 = (values_clipped - reward) ** 2 - surr2 = (values - reward) ** 2 - loss = torch.max(surr1, surr2) - loss = loss.mean() - return 0.5 * loss - - -class LogSigLoss(nn.Module): - """ - Pairwise Loss for Reward Model - Details: https://arxiv.org/abs/2203.02155 - """ - - def forward(self, chosen_reward: torch.Tensor, reject_reward: torch.Tensor) -> torch.Tensor: - probs = torch.sigmoid(chosen_reward - reject_reward) - log_probs = torch.log(probs) - loss = -log_probs.mean() - return loss - - -class LogExpLoss(nn.Module): - """ - Pairwise Loss for Reward Model - Details: https://arxiv.org/abs/2204.05862 - """ - - def forward(self, chosen_reward: torch.Tensor, reject_reward: torch.Tensor) -> torch.Tensor: - loss = torch.log(1 + torch.exp(reject_reward - chosen_reward)).mean() - return loss diff --git a/applications/Chat/coati/models/opt/__init__.py b/applications/Chat/coati/models/opt/__init__.py deleted file mode 100644 index e37d6e45c8fc..000000000000 --- a/applications/Chat/coati/models/opt/__init__.py +++ /dev/null @@ -1,5 +0,0 @@ -from .opt_actor import OPTActor -from .opt_critic import OPTCritic -from .opt_rm import OPTRM - -__all__ = ["OPTActor", "OPTCritic", "OPTRM"] diff --git a/applications/Chat/coati/models/opt/opt_actor.py b/applications/Chat/coati/models/opt/opt_actor.py deleted file mode 100644 index cd8908e13fb8..000000000000 --- a/applications/Chat/coati/models/opt/opt_actor.py +++ /dev/null @@ -1,37 +0,0 @@ -from typing import Optional - -from transformers.models.opt.configuration_opt import OPTConfig -from transformers.models.opt.modeling_opt import OPTForCausalLM - -from ..base import Actor - - -class OPTActor(Actor): - """ - OPT Actor model. - - Args: - pretrained (str): Pretrained model name or path. - config (OPTConfig): Model config. - checkpoint (bool): Enable gradient checkpointing. - lora_rank (int): Rank of the low-rank approximation. - lora_train_bias (str): LoRA bias training mode. - """ - - def __init__( - self, - pretrained: Optional[str] = None, - config: Optional[OPTConfig] = None, - checkpoint: bool = False, - lora_rank: int = 0, - lora_train_bias: str = "none", - ) -> None: - if pretrained is not None: - model = OPTForCausalLM.from_pretrained(pretrained) - elif config is not None: - model = OPTForCausalLM(config) - else: - model = OPTForCausalLM(OPTConfig()) - if checkpoint: - model.gradient_checkpointing_enable() - super().__init__(model, lora_rank, lora_train_bias) diff --git a/applications/Chat/coati/models/opt/opt_critic.py b/applications/Chat/coati/models/opt/opt_critic.py deleted file mode 100644 index f37d28812c27..000000000000 --- a/applications/Chat/coati/models/opt/opt_critic.py +++ /dev/null @@ -1,37 +0,0 @@ -from typing import Optional - -import torch.nn as nn -from transformers.models.opt.configuration_opt import OPTConfig -from transformers.models.opt.modeling_opt import OPTModel - -from ..base import Critic - - -class OPTCritic(Critic): - """ - OPT Critic model. - - Args: - pretrained (str): Pretrained model name or path. - config (OPTConfig): Model config. - lora_rank (int): Rank of the low-rank approximation. - lora_train_bias (str): LoRA bias training mode. - """ - - def __init__( - self, - pretrained: Optional[str] = None, - config: Optional[OPTConfig] = None, - lora_rank: int = 0, - lora_train_bias: str = "none", - **kwargs, - ) -> None: - if pretrained is not None: - model = OPTModel.from_pretrained(pretrained) - elif config is not None: - model = OPTModel(config) - else: - model = OPTModel(OPTConfig()) - - value_head = nn.Linear(model.config.word_embed_proj_dim, 1) - super().__init__(model, value_head, lora_rank, lora_train_bias, **kwargs) diff --git a/applications/Chat/coati/models/opt/opt_rm.py b/applications/Chat/coati/models/opt/opt_rm.py deleted file mode 100644 index 893708344ad4..000000000000 --- a/applications/Chat/coati/models/opt/opt_rm.py +++ /dev/null @@ -1,36 +0,0 @@ -from typing import Optional - -import torch.nn as nn -from transformers import OPTConfig, OPTModel - -from ..base import RewardModel - - -class OPTRM(RewardModel): - """ - OPT Reward model. - - Args: - pretrained (str): Pretrained model name or path. - config (OPTConfig): Model config. - lora_rank (int): Rank of the low-rank approximation. - lora_train_bias (str): LoRA bias training mode. - """ - - def __init__( - self, - pretrained: Optional[str] = None, - config: Optional[OPTConfig] = None, - lora_rank: int = 0, - lora_train_bias: str = "none", - ) -> None: - if pretrained is not None: - model = OPTModel.from_pretrained(pretrained) - elif config is not None: - model = OPTModel(config) - else: - model = OPTModel(OPTConfig()) - - value_head = nn.Linear(model.config.word_embed_proj_dim, 1) - value_head.weight.data.normal_(mean=0.0, std=1 / (model.config.word_embed_proj_dim + 1)) - super().__init__(model, value_head, lora_rank, lora_train_bias) diff --git a/applications/Chat/coati/models/utils.py b/applications/Chat/coati/models/utils.py deleted file mode 100644 index 1aaef16620d2..000000000000 --- a/applications/Chat/coati/models/utils.py +++ /dev/null @@ -1,69 +0,0 @@ -from typing import Optional, Union - -import torch -import torch.nn.functional as F - - -def _compute_approx_kl( - log_probs: torch.Tensor, log_probs_base: torch.Tensor, action_mask: Optional[torch.Tensor] = None -) -> torch.Tensor: - """ - Compute the approximate KL divergence between two distributions. - Schulman blog: http://joschu.net/blog/kl-approx.html - - Args: - log_probs: Log probabilities of the new distribution. - log_probs_base: Log probabilities of the base distribution. - action_mask: Mask for actions. - """ - - log_ratio = log_probs_base - log_probs - approx_kl = (log_ratio.exp() - 1) - log_ratio - if action_mask is not None: - approx_kl = masked_mean(approx_kl, action_mask, dim=1) - return approx_kl - approx_kl = approx_kl.mean(dim=1) - return approx_kl - - -def compute_reward( - r: Union[torch.Tensor, float], - kl_coef: float, - log_probs: torch.Tensor, - log_probs_base: torch.Tensor, - action_mask: Optional[torch.Tensor] = None, -) -> torch.Tensor: - if kl_coef <= 0.0: - return r - kl = _compute_approx_kl(log_probs, log_probs_base, action_mask=action_mask) - reward = r - kl_coef * kl - return reward - - -def _log_probs_from_logits(logits: torch.Tensor, labels: torch.Tensor) -> torch.Tensor: - log_probs = F.log_softmax(logits, dim=-1) - log_probs_labels = log_probs.gather(dim=-1, index=labels.unsqueeze(-1)) - return log_probs_labels.squeeze(-1) - - -def calc_action_log_probs(logits: torch.Tensor, sequences: torch.LongTensor, num_actions: int) -> torch.Tensor: - """Calculate action log probs. - - Args: - output (torch.Tensor): Output tensor of Actor.forward.logits. - sequences (torch.LongTensor): Input sequences. - num_actions (int): Number of actions. - - Returns: - torch.Tensor: Action log probs. - """ - log_probs = _log_probs_from_logits(logits[:, :-1, :], sequences[:, 1:]) - return log_probs[:, -num_actions:] - - -def masked_mean(tensor: torch.Tensor, mask: torch.Tensor, dim: int = 1) -> torch.Tensor: - tensor = tensor * mask - tensor = tensor.sum(dim=dim) - mask_sum = mask.sum(dim=dim) - mean = tensor / (mask_sum + 1e-8) - return mean diff --git a/applications/Chat/coati/trainer/__init__.py b/applications/Chat/coati/trainer/__init__.py deleted file mode 100644 index 4be5d27f93b1..000000000000 --- a/applications/Chat/coati/trainer/__init__.py +++ /dev/null @@ -1,6 +0,0 @@ -from .base import OnPolicyTrainer, SLTrainer -from .ppo import PPOTrainer -from .rm import RewardModelTrainer -from .sft import SFTTrainer - -__all__ = ["SLTrainer", "OnPolicyTrainer", "RewardModelTrainer", "SFTTrainer", "PPOTrainer"] diff --git a/applications/Chat/coati/trainer/callbacks/__init__.py b/applications/Chat/coati/trainer/callbacks/__init__.py deleted file mode 100644 index 29c8c4f00a5c..000000000000 --- a/applications/Chat/coati/trainer/callbacks/__init__.py +++ /dev/null @@ -1,5 +0,0 @@ -from .base import Callback -from .performance_evaluator import PerformanceEvaluator -from .save_checkpoint import SaveCheckpoint - -__all__ = ["Callback", "PerformanceEvaluator", "SaveCheckpoint"] diff --git a/applications/Chat/coati/trainer/callbacks/save_checkpoint.py b/applications/Chat/coati/trainer/callbacks/save_checkpoint.py deleted file mode 100644 index 0d70b6c53073..000000000000 --- a/applications/Chat/coati/trainer/callbacks/save_checkpoint.py +++ /dev/null @@ -1,76 +0,0 @@ -import os - -import torch.distributed as dist -from coati.trainer.strategies import GeminiStrategy, LowLevelZeroStrategy, Strategy -from coati.trainer.utils import is_rank_0 -from torch import nn -from torch.optim import Optimizer - -from .base import Callback - - -class SaveCheckpoint(Callback): - """ - The callback for saving checkpoint for coati. - - Only support saving actor and critic model. - A typical architecture of the saved checkpoint would be: - - checkpoint - - episode_x - - actor.pt - - actor-optim-rank-0.pt - - actor-optim-rank-1.pt - - critic.pt - - critic-optim-rank-0.pt - - critic-optim-rank-1.pt - - ... - - Args: - path(str): the base path you want to save checkpoint, the checkpoint would be saved at `path/checkpoint` - interval(int): the interval episode of saving checkpoint - strategy(Strategy): the strategy used to train - actor(nn.Module): the actor model - critic(nn.Module): the critic model - actor_optim(Optimizer): the optimizer of actor - critic_optim(Optimizer): the optimizer of critic - - """ - - def __init__( - self, - path: str, - interval: int, - strategy: Strategy, - actor: nn.Module = None, - critic: nn.Module = None, - actor_optim: Optimizer = None, - critic_optim: Optimizer = None, - ) -> None: - super().__init__() - self.path = os.path.join(path, "checkpoint") - self.interval = interval - self.strategy = strategy - self.model_dict = {"actor": [actor, actor_optim], "critic": [critic, critic_optim]} - - def on_episode_end(self, episode: int) -> None: - if (episode + 1) % self.interval != 0: - return - base_path = os.path.join(self.path, f"episode_{episode}") - if not os.path.exists(base_path): - os.makedirs(base_path) - - for model in self.model_dict.keys(): - # save model - if self.model_dict[model][0] is None: - # saving only optimizer states is meaningless, so it would be skipped - continue - model_path = os.path.join(base_path, f"{model}.pt") - self.strategy.save_model(model=self.model_dict[model][0], path=model_path, only_rank0=True) - - # save optimizer - if self.model_dict[model][1] is None: - continue - only_rank0 = not isinstance(self.strategy, (LowLevelZeroStrategy, GeminiStrategy)) - rank = 0 if is_rank_0() else dist.get_rank() - optim_path = os.path.join(base_path, f"{model}-optim-rank-{rank}.pt") - self.strategy.save_optimizer(optimizer=self.model_dict[model][1], path=optim_path, only_rank0=only_rank0) diff --git a/applications/Chat/coati/trainer/ppo.py b/applications/Chat/coati/trainer/ppo.py deleted file mode 100644 index 330e4e0e395e..000000000000 --- a/applications/Chat/coati/trainer/ppo.py +++ /dev/null @@ -1,202 +0,0 @@ -from typing import Dict, List, Optional - -from coati.experience_buffer import NaiveExperienceBuffer -from coati.experience_maker import Experience, NaiveExperienceMaker -from coati.models.base import Actor, Critic, RewardModel, get_base_model -from coati.models.loss import GPTLMLoss, PolicyLoss, ValueLoss -from coati.models.utils import calc_action_log_probs -from torch.optim import Optimizer -from torch.utils.data import DataLoader, DistributedSampler -from tqdm import tqdm -from transformers import PreTrainedTokenizerBase - -from colossalai.accelerator import get_accelerator - -from .base import OnPolicyTrainer -from .callbacks import Callback -from .strategies import GeminiStrategy, Strategy -from .utils import CycledDataLoader, is_rank_0, to_device - - -def _set_default_generate_kwargs(strategy: Strategy, generate_kwargs: dict, actor: Actor) -> Dict: - unwrapped_model = strategy.unwrap_model(actor) - hf_model = get_base_model(unwrapped_model) - new_kwargs = {**generate_kwargs} - # use huggingface models method directly - if "prepare_inputs_fn" not in generate_kwargs and hasattr(hf_model, "prepare_inputs_for_generation"): - new_kwargs["prepare_inputs_fn"] = hf_model.prepare_inputs_for_generation - - if "update_model_kwargs_fn" not in generate_kwargs and hasattr(hf_model, "_update_model_kwargs_for_generation"): - new_kwargs["update_model_kwargs_fn"] = hf_model._update_model_kwargs_for_generation - - return new_kwargs - - -class PPOTrainer(OnPolicyTrainer): - """ - Trainer for PPO algorithm. - - Args: - strategy (Strategy): the strategy to use for training - actor (Actor): the actor model in ppo algorithm - critic (Critic): the critic model in ppo algorithm - reward_model (RewardModel): the reward model in rlhf algorithm to make reward of sentences - initial_model (Actor): the initial model in rlhf algorithm to generate reference logics to limit the update of actor - actor_optim (Optimizer): the optimizer to use for actor model - critic_optim (Optimizer): the optimizer to use for critic model - kl_coef (float, defaults to 0.1): the coefficient of kl divergence loss - train_batch_size (int, defaults to 8): the batch size to use for training - buffer_limit (int, defaults to 0): the max_size limitation of buffer - buffer_cpu_offload (bool, defaults to True): whether to offload buffer to cpu - eps_clip (float, defaults to 0.2): the clip coefficient of policy loss - vf_coef (float, defaults to 1.0): the coefficient of value loss - ptx_coef (float, defaults to 0.9): the coefficient of ptx loss - value_clip (float, defaults to 0.4): the clip coefficient of value loss - sample_buffer (bool, defaults to False): whether to sample from buffer - dataloader_pin_memory (bool, defaults to True): whether to pin memory for data loader - offload_inference_models (bool, defaults to True): whether to offload inference models to cpu during training process - callbacks (List[Callback], defaults to []): the callbacks to call during training process - generate_kwargs (dict, optional): the kwargs to use while model generating - """ - - def __init__( - self, - strategy: Strategy, - actor: Actor, - critic: Critic, - reward_model: RewardModel, - initial_model: Actor, - actor_optim: Optimizer, - critic_optim: Optimizer, - tokenizer: PreTrainedTokenizerBase, - kl_coef: float = 0.1, - ptx_coef: float = 0.9, - train_batch_size: int = 8, - buffer_limit: int = 0, - buffer_cpu_offload: bool = True, - eps_clip: float = 0.2, - vf_coef: float = 1.0, - value_clip: float = 0.4, - sample_buffer: bool = False, - dataloader_pin_memory: bool = True, - offload_inference_models: bool = True, - callbacks: List[Callback] = [], - **generate_kwargs, - ) -> None: - if isinstance(strategy, GeminiStrategy): - assert not offload_inference_models, "GeminiPlugin is not compatible with manual model.to('cpu')" - - data_buffer = NaiveExperienceBuffer(train_batch_size, buffer_limit, buffer_cpu_offload) - super().__init__(strategy, data_buffer, sample_buffer, dataloader_pin_memory, callbacks) - - self.generate_kwargs = _set_default_generate_kwargs(strategy, generate_kwargs, actor) - self.experience_maker = NaiveExperienceMaker(actor, critic, reward_model, initial_model, tokenizer, kl_coef) - - self.actor = actor - self.critic = critic - self.tokenizer = tokenizer - - self.actor_loss_fn = PolicyLoss(eps_clip) - self.critic_loss_fn = ValueLoss(value_clip) - self.vf_coef = vf_coef - self.ptx_loss_fn = GPTLMLoss() - self.ptx_coef = ptx_coef - self.actor_optim = actor_optim - self.critic_optim = critic_optim - - self.offload_inference_models = offload_inference_models - self.device = get_accelerator().get_current_device() - - def _before_fit( - self, - prompt_dataloader: DataLoader, - pretrain_dataloader: DataLoader, - log_dir: Optional[str] = None, - use_wandb: bool = False, - ): - """ - Args: - prompt_dataloader (DataLoader): the dataloader to use for prompt data - pretrain_dataloader (DataLoader): the dataloader to use for pretrain data - """ - self.prompt_dataloader = CycledDataLoader(prompt_dataloader) - self.pretrain_dataloader = CycledDataLoader(pretrain_dataloader) - - self.writer = None - if use_wandb and is_rank_0(): - assert log_dir is not None, "log_dir must be provided when use_wandb is True" - import wandb - - wandb.init(project="Coati-ppo", sync_tensorboard=True) - if log_dir is not None and is_rank_0(): - import os - import time - - from torch.utils.tensorboard import SummaryWriter - - log_dir = os.path.join(log_dir, "ppo") - log_dir = os.path.join(log_dir, time.strftime("%Y-%m-%d_%H:%M:%S", time.localtime())) - self.writer = SummaryWriter(log_dir=log_dir) - - def _make_experience(self, collect_step: int) -> Experience: - prompts = self.prompt_dataloader.next() - if self.offload_inference_models: - # TODO(ver217): this may be controlled by strategy if they are prepared by strategy - self.experience_maker.initial_model.to(self.device) - self.experience_maker.reward_model.to(self.device) - assert isinstance(prompts, dict), f'Unsupported input type "{type(prompts)}"' - return self.experience_maker.make_experience(**prompts, **self.generate_kwargs) - - def _training_step(self, experience: Experience): - self.actor.train() - self.critic.train() - # policy loss - num_actions = experience.action_log_probs.size(1) - actor_logits = self.actor(experience.sequences, experience.attention_mask)["logits"] - action_log_probs = calc_action_log_probs(actor_logits, experience.sequences, num_actions) - actor_loss = self.actor_loss_fn( - action_log_probs, experience.action_log_probs, experience.advantages, action_mask=experience.action_mask - ) - actor_loss = (1 - self.ptx_coef) * actor_loss - self.strategy.backward(actor_loss, self.actor, self.actor_optim) - - # ptx loss - if self.ptx_coef != 0: - batch = self.pretrain_dataloader.next() - batch = to_device(batch, self.device) - ptx_log_probs = self.actor(batch["input_ids"], batch["attention_mask"])["logits"] - ptx_loss = self.ptx_coef * self.ptx_loss_fn(ptx_log_probs, batch["labels"]) - self.strategy.backward(ptx_loss, self.actor, self.actor_optim) - - self.strategy.optimizer_step(self.actor_optim) - self.actor_optim.zero_grad() - - # value loss - values = self.critic(experience.sequences, attention_mask=experience.attention_mask) - critic_loss = self.critic_loss_fn(values, experience.values, experience.reward) - critic_loss = critic_loss * self.vf_coef - self.strategy.backward(critic_loss, self.critic, self.critic_optim) - self.strategy.optimizer_step(self.critic_optim) - self.critic_optim.zero_grad() - - def _learn(self, update_step: int): - if self.offload_inference_models: - self.experience_maker.initial_model.to("cpu") - self.experience_maker.reward_model.to("cpu") - - # buffer may be empty at first, we should rebuild at each training - if self.sample_buffer: - experience = self.data_buffer.sample() - self._on_learn_batch_start() - experience.to_device(self.device) - self._training_step(experience) - self._on_learn_batch_end(experience) - else: - if isinstance(self.dataloader.sampler, DistributedSampler): - self.dataloader.sampler.set_epoch(update_step) - pbar = tqdm(self.dataloader, desc=f"Train epoch [{update_step + 1}]", disable=not is_rank_0()) - for experience in pbar: - self._on_learn_batch_start() - experience.to_device(self.device) - self._training_step(experience) - self._on_learn_batch_end(experience) diff --git a/applications/Chat/coati/trainer/rm.py b/applications/Chat/coati/trainer/rm.py deleted file mode 100644 index d7f8c21a5a3d..000000000000 --- a/applications/Chat/coati/trainer/rm.py +++ /dev/null @@ -1,123 +0,0 @@ -from typing import Callable, Optional - -import torch -import tqdm -from torch.optim import Optimizer -from torch.optim.lr_scheduler import _LRScheduler -from torch.utils.data import DataLoader - -from .base import SLTrainer -from .strategies import Strategy -from .utils import is_rank_0 - - -class RewardModelTrainer(SLTrainer): - """ - Trainer to use while training reward model. - - Args: - model (torch.nn.Module): the model to train - strategy (Strategy): the strategy to use for training - optim (Optimizer): the optimizer to use for training - lr_scheduler (_LRScheduler): the lr scheduler to use for training - loss_fn (callable): the loss function to use for training - max_epochs (int, defaults to 2): the number of epochs to train - """ - - def __init__( - self, - model, - strategy: Strategy, - optim: Optimizer, - lr_scheduler: _LRScheduler, - loss_fn: Callable, - max_epochs: int = 1, - ) -> None: - super().__init__(strategy, max_epochs, model, optim) - - self.loss_fn = loss_fn - self.scheduler = lr_scheduler - - self.num_train_step = 0 - - def _eval(self, epoch): - if self.eval_dataloader is not None: - self.model.eval() - dist, num_correct, num_samples = 0, 0, 0 - with torch.no_grad(): - for chosen_ids, c_mask, reject_ids, r_mask in self.eval_dataloader: - chosen_ids = chosen_ids.squeeze(1).to(torch.cuda.current_device()) - c_mask = c_mask.squeeze(1).to(torch.cuda.current_device()) - reject_ids = reject_ids.squeeze(1).to(torch.cuda.current_device()) - r_mask = r_mask.squeeze(1).to(torch.cuda.current_device()) - chosen_reward = self.model(chosen_ids, attention_mask=c_mask) - reject_reward = self.model(reject_ids, attention_mask=r_mask) - num_samples += chosen_ids.size(0) - num_correct += (chosen_reward > reject_reward).sum().item() - dist += (chosen_reward - reject_reward).mean().item() - self.dist = dist / len(self.eval_dataloader) - self.acc = num_correct / num_samples - - if self.writer: - self.writer.add_scalar("eval/dist", self.dist, epoch) - self.writer.add_scalar("eval/acc", self.acc, epoch) - - def _train(self, epoch): - self.model.train() - step_bar = tqdm.trange( - len(self.train_dataloader), desc=f"Epoch {epoch + 1}/{self.max_epochs}", disable=not is_rank_0() - ) - for chosen_ids, c_mask, reject_ids, r_mask in self.train_dataloader: - chosen_ids = chosen_ids.squeeze(1).to(torch.cuda.current_device()) - c_mask = c_mask.squeeze(1).to(torch.cuda.current_device()) - reject_ids = reject_ids.squeeze(1).to(torch.cuda.current_device()) - r_mask = r_mask.squeeze(1).to(torch.cuda.current_device()) - chosen_reward = self.model(chosen_ids, attention_mask=c_mask) - reject_reward = self.model(reject_ids, attention_mask=r_mask) - loss = self.loss_fn(chosen_reward, reject_reward) - self.strategy.backward(loss, self.model, self.optimizer) - self.strategy.optimizer_step(self.optimizer) - self.optimizer.zero_grad() - if self.writer: - self.writer.add_scalar("train/loss", loss.item(), self.num_train_step) - self.writer.add_scalar("train/lr", self.optimizer.param_groups[0]["lr"], self.num_train_step) - self.writer.add_scalar("train/dist", (chosen_reward - reject_reward).mean().item(), self.num_train_step) - self.writer.add_scalar( - "train/acc", (chosen_reward > reject_reward).float().mean().item(), self.num_train_step - ) - self.num_train_step += 1 - if self.num_train_step % 100 == 0: - self.scheduler.step() - step_bar.update() - step_bar.close() - - def _before_fit( - self, - train_dataloader: DataLoader, - eval_dataloader: DataLoader, - log_dir: Optional[str] = None, - use_wandb: bool = False, - ): - """ - Args: - train_dataloader (DataLoader): the dataloader to use for training - eval_dataloader (DataLoader): the dataloader to use for evaluation - """ - self.train_dataloader = train_dataloader - self.eval_dataloader = eval_dataloader - - self.writer = None - if use_wandb and is_rank_0(): - assert log_dir is not None, "log_dir must be provided when use_wandb is True" - import wandb - - wandb.init(project="Coati-rm", sync_tensorboard=True) - if log_dir is not None and is_rank_0(): - import os - import time - - from torch.utils.tensorboard import SummaryWriter - - log_dir = os.path.join(log_dir, "rm") - log_dir = os.path.join(log_dir, time.strftime("%Y-%m-%d_%H:%M:%S", time.localtime())) - self.writer = SummaryWriter(log_dir=log_dir) diff --git a/applications/Chat/coati/trainer/sft.py b/applications/Chat/coati/trainer/sft.py deleted file mode 100644 index 7d0eeec897e5..000000000000 --- a/applications/Chat/coati/trainer/sft.py +++ /dev/null @@ -1,130 +0,0 @@ -from typing import Optional - -import torch -import torch.distributed as dist -import tqdm -from torch.optim import Optimizer -from torch.optim.lr_scheduler import _LRScheduler -from torch.utils.data import DataLoader - -from colossalai.logging import DistributedLogger - -from .base import SLTrainer -from .strategies import GeminiStrategy, Strategy -from .utils import is_rank_0, to_device - - -class SFTTrainer(SLTrainer): - """ - Trainer to use while training reward model. - - Args: - model (torch.nn.Module): the model to train - strategy (Strategy): the strategy to use for training - optim(Optimizer): the optimizer to use for training - lr_scheduler(_LRScheduler): the lr scheduler to use for training - max_epochs (int, defaults to 2): the number of epochs to train - accumulation_steps (int, defaults to 8): the number of steps to accumulate gradients - """ - - def __init__( - self, - model, - strategy: Strategy, - optim: Optimizer, - lr_scheduler: _LRScheduler, - max_epochs: int = 2, - accumulation_steps: int = 8, - ) -> None: - if accumulation_steps > 1: - assert not isinstance( - strategy, GeminiStrategy - ), "Accumulation steps are not supported in stage 3 of ColossalAI" - - super().__init__(strategy, max_epochs, model, optim) - - self.accumulation_steps = accumulation_steps - self.scheduler = lr_scheduler - - self.num_train_step = 0 - self.num_eval_step = 0 - - def _train(self, epoch: int): - self.model.train() - step_bar = tqdm.trange( - len(self.train_dataloader) // self.accumulation_steps, - desc=f"Epoch {epoch + 1}/{self.max_epochs}", - disable=not is_rank_0(), - ) - for i, batch in enumerate(self.train_dataloader): - batch = to_device(batch, torch.cuda.current_device()) - outputs = self.model(batch["input_ids"], attention_mask=batch["attention_mask"], labels=batch["labels"]) - loss = outputs.loss / self.accumulation_steps - self.total_loss += loss.item() - self.strategy.backward(loss, self.model, self.optimizer) - # gradient accumulation - if (i + 1) % self.accumulation_steps == 0: - self.strategy.optimizer_step(self.optimizer) - self.optimizer.zero_grad() - self.scheduler.step() - if self.writer: - self.writer.add_scalar("train/loss", self.total_loss, self.num_train_step) - self.writer.add_scalar("train/lr", self.scheduler.get_last_lr()[0], self.num_train_step) - self.num_train_step += 1 - self.total_loss = 0 - step_bar.update() - step_bar.close() - - def _eval(self, epoch: int): - if self.eval_dataloader is not None: - self.model.eval() - with torch.no_grad(): - loss_sum, num_seen = 0, 0 - for batch in self.eval_dataloader: - batch = to_device(batch, torch.cuda.current_device()) - outputs = self.model( - batch["input_ids"], attention_mask=batch["attention_mask"], labels=batch["labels"] - ) - loss_sum += outputs.loss.item() - num_seen += batch["input_ids"].size(0) - loss_mean = loss_sum / num_seen - if dist.get_rank() == 0: - self.logger.info(f"Eval Epoch {epoch}/{self.max_epochs} loss {loss_mean}") - if self.writer: - self.writer.add_scalar("eval/loss", loss_mean, self.num_eval_step) - self.num_eval_step += 1 - - def _before_fit( - self, - train_dataloader: DataLoader, - eval_dataloader: Optional[DataLoader] = None, - logger: Optional[DistributedLogger] = None, - log_dir: Optional[str] = None, - use_wandb: bool = False, - ): - """ - Args: - train_dataloader: the dataloader to use for training - eval_dataloader: the dataloader to use for evaluation - """ - self.train_dataloader = train_dataloader - self.eval_dataloader = eval_dataloader - - self.logger = logger - self.writer = None - if use_wandb and is_rank_0(): - assert log_dir is not None, "log_dir must be provided when use_wandb is True" - import wandb - - wandb.init(project="Coati-sft", sync_tensorboard=True) - if log_dir is not None and is_rank_0(): - import os - import time - - from torch.utils.tensorboard import SummaryWriter - - log_dir = os.path.join(log_dir, "sft") - log_dir = os.path.join(log_dir, time.strftime("%Y-%m-%d_%H:%M:%S", time.localtime())) - self.writer = SummaryWriter(log_dir=log_dir) - - self.total_loss = 0 diff --git a/applications/Chat/coati/trainer/strategies/__init__.py b/applications/Chat/coati/trainer/strategies/__init__.py deleted file mode 100644 index 521dcb5855b1..000000000000 --- a/applications/Chat/coati/trainer/strategies/__init__.py +++ /dev/null @@ -1,5 +0,0 @@ -from .base import Strategy -from .colossalai import GeminiStrategy, LowLevelZeroStrategy -from .ddp import DDPStrategy - -__all__ = ["Strategy", "DDPStrategy", "LowLevelZeroStrategy", "GeminiStrategy"] diff --git a/applications/Chat/coati/trainer/strategies/base.py b/applications/Chat/coati/trainer/strategies/base.py deleted file mode 100644 index a78716216ae0..000000000000 --- a/applications/Chat/coati/trainer/strategies/base.py +++ /dev/null @@ -1,137 +0,0 @@ -from abc import ABC, abstractmethod -from contextlib import nullcontext -from typing import Callable, Dict, List, Optional, Tuple, Union - -import torch -import torch.nn as nn -from coati.experience_buffer import ExperienceBuffer -from torch.optim import Optimizer -from torch.utils.data import DataLoader -from transformers.tokenization_utils_base import PreTrainedTokenizerBase - -from colossalai.booster import Booster -from colossalai.booster.plugin import Plugin - -from .sampler import DistributedSampler - -_BoostArgSpec = Union[nn.Module, Tuple[nn.Module, Optimizer], Dict] - - -class Strategy(ABC): - """ - Base class for training strategies. - """ - - def __init__(self, plugin_initializer: Callable[..., Optional[Plugin]] = lambda: None) -> None: - super().__init__() - # NOTE: dist must be initialized before Booster - self.setup_distributed() - self.plugin = plugin_initializer() - self.booster = Booster(plugin=self.plugin) - self._post_init() - - @abstractmethod - def _post_init(self) -> None: - pass - - def backward(self, loss: torch.Tensor, model: nn.Module, optimizer: Optimizer, **kwargs) -> None: - self.booster.backward(loss, optimizer) - - def optimizer_step(self, optimizer: Optimizer, **kwargs) -> None: - optimizer.step() - - @abstractmethod - def setup_distributed(self) -> None: - pass - - @abstractmethod - def setup_dataloader(self, data_buffer: ExperienceBuffer, pin_memory: bool = False) -> DataLoader: - pass - - def model_init_context(self): - return nullcontext() - - def prepare(self, *boost_args: _BoostArgSpec) -> Union[List[_BoostArgSpec], _BoostArgSpec]: - """Prepare [model | (model, optimizer) | Dict] based on each strategy. - NOTE: the keys of Dict must be a subset of `self.booster.boost`'s arguments. - - Example:: - >>> # e.g., include lr_scheduler - >>> result_dict = strategy.prepare(dict(model=model, lr_scheduler=lr_scheduler)) - >>> # when fine-tuning actor and critic - >>> (actor, actor_optim), (critic, critic_optim), reward_model, initial_model = strategy.prepare((actor, actor_optim), (critic, critic_optim), reward_model, initial_model) - >>> # or when training reward model - >>> (reward_model, reward_model_optim) = strategy.prepare((reward_model, reward_model_optim)) - >>> # or just inference - >>> actor, critic = strategy.prepare(actor, critic) - - Returns: - Union[List[_BoostArgSpec], _BoostArgSpec]: [model | (model, optimizer) | Dict] in the original order. - """ - - rets = [] - for arg in boost_args: - if isinstance(arg, nn.Module): - model, *_ = self.booster.boost(arg) - rets.append(model) - elif isinstance(arg, tuple): - try: - model, optimizer = arg - except ValueError: - raise RuntimeError(f'Expect (model, optimizer) pair, got a tuple with size "{len(arg)}"') - model, optimizer, *_ = self.booster.boost(model=model, optimizer=optimizer) - rets.append((model, optimizer)) - elif isinstance(arg, Dict): - model, optimizer, criterion, dataloader, lr_scheduler = self.booster.boost(**arg) - boost_result = dict( - model=model, - optimizer=optimizer, - criterion=criterion, - dataloader=dataloader, - lr_scheduler=lr_scheduler, - ) - # remove None values - boost_result = {key: value for key, value in boost_result.items() if value is not None} - rets.append(boost_result) - else: - raise RuntimeError(f"Type {type(arg)} is not supported") - - return rets[0] if len(rets) == 1 else rets - - @staticmethod - def unwrap_model(model: nn.Module) -> nn.Module: - """Get the unwrapped model from a wrapped model made by Strategy.prepare. - - Args: - model (nn.Module): the model to unwrap - - Returns: - nn.Module: the original model - """ - return model - - def save_model(self, model: nn.Module, path: str, shard: bool = False, **kwargs) -> None: - self.booster.save_model(model, path, shard=shard, **kwargs) - - def load_model(self, model: nn.Module, path: str, strict: bool = True) -> None: - self.booster.load_model(model, path, strict) - - def save_optimizer(self, optimizer: Optimizer, path: str, only_rank0: bool = False, **kwargs) -> None: - self.booster.save_optimizer(optimizer, path, shard=not only_rank0, **kwargs) - - def load_optimizer(self, optimizer: Optimizer, path: str) -> None: - self.booster.load_optimizer(optimizer, path) - - def setup_sampler(self, dataset) -> DistributedSampler: - # FIXME(cwher): this is only invoked in train_on_ray, not tested after adapt Boost API. - return DistributedSampler(dataset, 1, 0) - - @abstractmethod - def save_pretrained( - self, model: nn.Module, path: str, only_rank0: bool = True, tokenizer: Optional[PreTrainedTokenizerBase] = None - ) -> None: - pass - - @abstractmethod - def get_model_state_dict_shard(self, model: nn.Module, **config): - pass diff --git a/applications/Chat/coati/trainer/strategies/colossalai.py b/applications/Chat/coati/trainer/strategies/colossalai.py deleted file mode 100644 index 95f01678640c..000000000000 --- a/applications/Chat/coati/trainer/strategies/colossalai.py +++ /dev/null @@ -1,209 +0,0 @@ -import warnings -from typing import Optional - -import torch.nn as nn - -import colossalai -from colossalai.booster.plugin import GeminiPlugin, LowLevelZeroPlugin -from colossalai.booster.plugin.low_level_zero_plugin import LowLevelZeroModel -from colossalai.zero.gemini.gemini_ddp import GeminiDDP - -from .ddp import DDPStrategy - - -class LowLevelZeroStrategy(DDPStrategy): - """ - The strategy for training with ColossalAI. - - Args: - stage(int): The stage to use in ZeRO. Choose in (1, 2) - precision(str): The precision to use. Choose in ('fp32', 'fp16'). - seed(int): The seed for the random number generator. - placement_policy(str): The placement policy for gemini. Choose in ('cpu', 'cuda') - If it is “cpu”, parameters, gradients and optimizer states will be offloaded to CPU, - If it is “cuda”, they will not be offloaded, which means max CUDA memory will be used. It is the fastest. - reduce_bucket_size(int): The reduce bucket size in bytes. Only for ZeRO-1 and ZeRO-2. - overlap_communication(bool): Whether to overlap communication and computation. Only for ZeRO-1 and ZeRO-2. - initial_scale(float): The initial scale for the optimizer. - growth_factor(float): The growth factor for the optimizer. - backoff_factor(float): The backoff factor for the optimizer. - growth_interval(int): The growth interval for the optimizer. - hysteresis(int): The hysteresis for the optimizer. - min_scale(float): The minimum scale for the optimizer. - max_scale(float): The maximum scale for the optimizer. - max_norm(float): The maximum norm for the optimizer. - norm_type(float): The norm type for the optimizer. - - """ - - def __init__( - self, - stage: int = 2, - precision: str = "fp16", - seed: int = 42, - placement_policy: str = "cuda", - reduce_bucket_size: int = 12 * 1024**2, # only for stage 1&2 - overlap_communication: bool = True, # only for stage 1&2 - initial_scale: float = 2**16, - growth_factor: float = 2, - backoff_factor: float = 0.5, - growth_interval: int = 1000, - hysteresis: int = 2, - min_scale: float = 1, - max_scale: float = 2**32, - max_norm: float = 0.0, - norm_type: float = 2.0, - ) -> None: - assert stage in (1, 2), f'Unsupported stage "{stage}"' - assert placement_policy in ("cpu", "cuda"), f'Unsupported placement policy "{placement_policy}"' - assert precision in ("fp32", "fp16"), f'Unsupported precision "{precision}"' - - plugin_initializer = lambda: LowLevelZeroPlugin( - stage=stage, - precision=precision, - reduce_bucket_size_in_m=reduce_bucket_size, - overlap_communication=overlap_communication, - cpu_offload=(placement_policy == "cpu"), - initial_scale=initial_scale, - growth_factor=growth_factor, - backoff_factor=backoff_factor, - growth_interval=growth_interval, - hysteresis=hysteresis, - min_scale=min_scale, - max_scale=max_scale, - max_norm=max_norm, - norm_type=norm_type, - ) - - super().__init__(seed, plugin_initializer) - - def _post_init(self) -> None: - assert isinstance( - self.plugin, LowLevelZeroPlugin - ), f"{type(self).__name__}'s plugin is not initialized properly." - - def setup_distributed(self) -> None: - colossalai.launch_from_torch({}, seed=self.seed) - - def unwrap_model(self, model: nn.Module) -> nn.Module: - assert isinstance(model, LowLevelZeroModel) - return model.module - - def get_model_state_dict_shard(self, model: nn.Module, **config): - assert isinstance(model, LowLevelZeroModel) - yield from model.state_dict_shard(max_shard_size=1024, only_rank_0=False) - - -class GeminiStrategy(DDPStrategy): - """ - The strategy for training with ColossalAI. - - Args: - seed(int): The seed for the random number generator. - shard_init(bool): Whether to shard the model parameters during initialization. Only for ZeRO-3. - This is not compatible with `from_pretrained()`. We temporarily disable this and will support it in the future. - placement_policy(str): The placement policy for gemini. Choose in ('cpu', 'cuda') - If it is “cpu”, parameters, gradients and optimizer states will be offloaded to CPU, - If it is “cuda”, they will not be offloaded, which means max CUDA memory will be used. It is the fastest. - pin_memory(bool): Whether to pin the memory for the data loader. Only for ZeRO-3. - force_outputs_fp32(bool): Whether to force the outputs to be fp32. Only for ZeRO-3. - search_range_m(int): The number of search range for the chunk size, divided by 2^20. Only for ZeRO-3. - hidden_dim(optional, int): The hidden dimension for the gemini. Only for ZeRO-3. - min_chunk_size_m(float): The minimum chunk size divided by 2^20. Only for ZeRO-3. - gpu_margin_mem_ratio(float): The margin memory ratio for the GPU. Only for ZeRO-3. - initial_scale(float): The initial scale for the optimizer. - growth_factor(float): The growth factor for the optimizer. - backoff_factor(float): The backoff factor for the optimizer. - growth_interval(int): The growth interval for the optimizer. - hysteresis(int): The hysteresis for the optimizer. - min_scale(float): The minimum scale for the optimizer. - max_scale(float): The maximum scale for the optimizer. - max_norm(float): The maximum norm for the optimizer. - norm_type(float): The norm type for the optimizer. - - """ - - def __init__( - self, - seed: int = 42, - shard_init: bool = False, # only for stage 3 - placement_policy: str = "auto", - shard_param_frac: float = 1.0, # only for static placement - offload_optim_frac: float = 0.0, # only for static placement - offload_param_frac: float = 0.0, # only for static placement - pin_memory: bool = True, # only for stage 3 - force_outputs_fp32: bool = False, # only for stage 3 - search_range_m: int = 32, # only for stage 3 - hidden_dim: Optional[int] = None, # only for stage 3 - min_chunk_size_m: float = 32, # only for stage 3 - gpu_margin_mem_ratio: float = 0.0, # only for stage 3 - initial_scale: float = 2**16, - growth_factor: float = 2, - backoff_factor: float = 0.5, - growth_interval: int = 1000, - hysteresis: int = 2, - min_scale: float = 1, - max_scale: float = 2**32, - max_norm: float = 0.0, - norm_type: float = 2.0, - ) -> None: - # TODO(ver217): support shard_init when using from_pretrained() - if shard_init: - warnings.warn( - f"Shard init is not supported model.from_pretrained() yet. " - "Please load weights after strategy.prepare()" - ) - self.shard_init = shard_init - - warnings.warn(f"Stage 3 only supports fp16. Precision is set to fp16.") - - # colossalai has changed api for get_current_device in 0.3.4 version or newer - try: - from colossalai.accelerator import get_accelerator - - chunk_init_device = get_accelerator().get_current_device() - except: - from colossalai.utils import get_current_device - - chunk_init_device = get_current_device() - - # NOTE: dist should be initialized before calling get_current_device() - plugin_initializer = lambda: GeminiPlugin( - chunk_init_device=chunk_init_device, - placement_policy=placement_policy, - shard_param_frac=shard_param_frac, - offload_optim_frac=offload_optim_frac, - offload_param_frac=offload_param_frac, - precision="fp16", - pin_memory=pin_memory, - force_outputs_fp32=force_outputs_fp32, - strict_ddp_mode=shard_init, - search_range_m=search_range_m, - hidden_dim=hidden_dim, - min_chunk_size_m=min_chunk_size_m, - gpu_margin_mem_ratio=gpu_margin_mem_ratio, - initial_scale=initial_scale, - growth_factor=growth_factor, - backoff_factor=backoff_factor, - growth_interval=growth_interval, - hysteresis=hysteresis, - min_scale=min_scale, - max_scale=max_scale, - max_norm=max_norm, - norm_type=norm_type, - ) - - super().__init__(seed, plugin_initializer) - - def _post_init(self) -> None: - assert isinstance(self.plugin, GeminiPlugin), f"{type(self).__name__}'s plugin is not initialized properly." - - def setup_distributed(self) -> None: - colossalai.launch_from_torch({}, seed=self.seed) - - def model_init_context(self): - return super().model_init_context() - - def unwrap_model(self, model: nn.Module) -> nn.Module: - assert isinstance(model, GeminiDDP) - return model.module diff --git a/applications/Chat/coati/trainer/strategies/ddp.py b/applications/Chat/coati/trainer/strategies/ddp.py deleted file mode 100644 index 3d3d44ca51ae..000000000000 --- a/applications/Chat/coati/trainer/strategies/ddp.py +++ /dev/null @@ -1,137 +0,0 @@ -import os -import random -from collections import OrderedDict -from typing import Callable, Optional - -import numpy as np -import torch -import torch.distributed as dist -import torch.nn as nn -from coati.experience_buffer import ExperienceBuffer -from coati.models import Actor, Critic, RewardModel -from torch.utils.data import DataLoader -from transformers.modeling_utils import PreTrainedModel -from transformers.tokenization_utils_base import PreTrainedTokenizerBase - -from colossalai.booster.plugin import TorchDDPPlugin -from colossalai.booster.plugin.torch_ddp_plugin import TorchDDPModel - -from .base import Strategy -from .sampler import DistributedSampler - - -# TODO Move this to a util.py (Moving to ray.util introduces ringed import) -def get_grad_required_state_dict(model: nn.Module): - state_dict = OrderedDict() - for name, parameter in model.named_parameters(): - if parameter.requires_grad: - state_dict[name] = parameter.detach() - return state_dict - - -class DDPStrategy(Strategy): - """ - Strategy for distributed training using torch.distributed. - """ - - def __init__(self, seed: int = 42, plugin_initializer: Callable = TorchDDPPlugin) -> None: - self.seed = seed - super().__init__(plugin_initializer) - - def _try_init_dist(self, force: bool = False) -> None: - try: - rank = int(os.environ["RANK"]) - local_rank = int(os.environ["LOCAL_RANK"]) - world_size = int(os.environ["WORLD_SIZE"]) - host = os.environ["MASTER_ADDR"] - port = int(os.environ["MASTER_PORT"]) - dist.init_process_group("nccl", init_method=f"tcp://[{host}]:{port}", world_size=world_size, rank=rank) - torch.cuda.set_device(local_rank) - except KeyError as e: - if force: - raise RuntimeError( - f"Could not find {e} in the torch environment, visit https://www.colossalai.org/ for more information on launching with torch" - ) - except Exception as e: - if force: - raise e - - def _post_init(self) -> None: - assert isinstance(self.plugin, TorchDDPPlugin), f"{type(self).__name__}'s plugin is not initialized properly." - - def setup_distributed(self) -> None: - self._try_init_dist(force=True) - self.set_seed(self.seed) - - def set_seed(self, seed: int) -> None: - random.seed(seed) - np.random.seed(seed) - torch.manual_seed(seed) - - def setup_dataloader(self, data_buffer: ExperienceBuffer, pin_memory: bool = False) -> DataLoader: - return self.plugin.prepare_dataloader( - data_buffer, - batch_size=data_buffer.sample_batch_size, - shuffle=True, - drop_last=True, - pin_memory=pin_memory, - collate_fn=data_buffer.collate_fn, - ) - - def setup_sampler(self, dataset) -> DistributedSampler: - # FIXME(cwher): this is only invoked in train_on_ray, not tested after adapt Boost API. - return DistributedSampler(dataset, dist.get_world_size(), dist.get_rank()) - - def unwrap_model(self, model: nn.Module) -> nn.Module: - assert isinstance(model, TorchDDPModel), "model is not wrapped by TorchDDPModel." - return model.unwrap() - - def save_pretrained( - self, model: nn.Module, path: str, shard: bool = False, tokenizer: Optional[PreTrainedTokenizerBase] = None - ) -> None: - if dist.get_rank() == 0: - unwrapped_model = self.unwrap_model(model) - assert isinstance(unwrapped_model, (Actor, Critic, RewardModel)) - pretrained_model = unwrapped_model.model - assert isinstance(pretrained_model, PreTrainedModel) - # HACK: only use hf save_pretrained to save config - pretrained_model.save_pretrained(path, save_function=lambda *args, **kwargs: None) - if tokenizer is not None: - tokenizer.save_pretrained(path) - - model_path = os.path.join(path, "pytorch_model.bin") - self.save_model(model, model_path, shard=shard) - - def _replace_keys(model_path: str, replace_fn: Callable): - state_dict = torch.load(model_path, map_location="cpu") - state_dict = {replace_fn(k): v for k, v in state_dict.items()} - torch.save(state_dict, model_path) - - # FIXME: save_model would add "model." prefix to keys of pytorch_model.bin - # HACK: rename keys of pytorch_model.bin - if dist.get_rank() == 0: - _replace_keys(model_path, lambda k: k.replace("model.", "", 1)) - - def get_model_state_dict_shard(self, model: nn.Module, **config): - # TODO: implement sharding on naive strategy - model = self.unwrap_model(model) - if "requires_grad_only" in config and config["requires_grad_only"] == True: - state_dict = get_grad_required_state_dict(model) - else: - state_dict = model.state_dict() - - if "shard_size" in config: - shard_size = config["shard_size"] - accumulate_size = 0 - state_dict_shard = OrderedDict() - for name, param in state_dict.items(): - state_dict_shard[name] = param - accumulate_size += param.numel() * param.element_size() - if accumulate_size >= shard_size: - accumulate_size = 0 - yield state_dict_shard - state_dict_shard = OrderedDict() - if accumulate_size > 0: - yield state_dict_shard - else: - yield state_dict diff --git a/applications/Chat/coati/trainer/strategies/sampler.py b/applications/Chat/coati/trainer/strategies/sampler.py deleted file mode 100644 index 6e811bef11a5..000000000000 --- a/applications/Chat/coati/trainer/strategies/sampler.py +++ /dev/null @@ -1,31 +0,0 @@ -import math - -import numpy as np - - -class DistributedSampler: - def __init__(self, dataset, num_replicas: int, rank: int) -> None: - self.dataset = dataset - self.num_replicas = num_replicas - self.rank = rank - - if len(self.dataset) % self.num_replicas != 0: - self.num_samples = math.ceil( - (len(self.dataset) - self.num_replicas) / self.num_replicas # type: ignore[arg-type] - ) - else: - self.num_samples = math.ceil(len(self.dataset) / self.num_replicas) - - self.total_size = self.num_samples * self.num_replicas - - indices = list(range(len(self.dataset))) - indices = indices[: self.total_size] - assert len(indices) == self.total_size - # subsample - indices = indices[self.rank : self.total_size : self.num_replicas] - assert len(indices) == self.num_samples - self.indices = indices - - def sample(self, batch_size: int) -> list: - sampled_indices = np.random.choice(self.indices, batch_size, replace=False) - return [self.dataset[idx] for idx in sampled_indices] diff --git a/applications/Chat/coati/trainer/utils.py b/applications/Chat/coati/trainer/utils.py deleted file mode 100644 index 7811e7365eeb..000000000000 --- a/applications/Chat/coati/trainer/utils.py +++ /dev/null @@ -1,50 +0,0 @@ -from typing import Any - -import torch -import torch.distributed as dist -from torch.utils._pytree import tree_map -from torch.utils.data import DataLoader - - -class CycledDataLoader: - """ - Why do we need this class? - In version 4da324cd60, "prompts = next(iter(self.prompt_dataloader))" is used to sample a batch of prompts/pretrain. - However, this may be inefficient due to frequent re-initialization of the dataloader. (re-initialize workers...) - NOTE: next(iter(dataloader)) is not equivalent to for batch in dataloader: break, it causes slightly different behavior. - """ - - def __init__( - self, - dataloader: DataLoader, - ) -> None: - self.dataloader = dataloader - - self.count = 0 - self.dataloader_iter = None - - def next(self): - # defer initialization - if self.dataloader_iter is None: - self.dataloader_iter = iter(self.dataloader) - - self.count += 1 - try: - return next(self.dataloader_iter) - except StopIteration: - self.count = 0 - self.dataloader_iter = iter(self.dataloader) - return next(self.dataloader_iter) - - -def is_rank_0() -> bool: - return not dist.is_initialized() or dist.get_rank() == 0 - - -def to_device(x: Any, device: torch.device) -> Any: - def _to(t: Any): - if isinstance(t, torch.Tensor): - return t.to(device) - return t - - return tree_map(_to, x) diff --git a/applications/Chat/examples/README.md b/applications/Chat/examples/README.md deleted file mode 100644 index 9438aafd1268..000000000000 --- a/applications/Chat/examples/README.md +++ /dev/null @@ -1,409 +0,0 @@ -# Examples - -## Table of Contents - -- [Examples](#examples) - - [Table of Contents](#table-of-contents) - - [Install requirements](#install-requirements) - - [Supervised datasets collection](#supervised-datasets-collection) - - [Conversation dataset generation](#conversation-dataset-generation) - - [Stage1 - Supervised instructs tuning](#stage1---supervised-instructs-tuning) - - [Arg List](#arg-list) - - [Stage2 - Training reward model](#stage2---training-reward-model) - - [Features and tricks in RM training](#features-and-tricks-in-rm-training) - - [Experiment result](#experiment-result) - - [Arg List](#arg-list-1) - - [Stage3 - Training model using prompts with RL](#stage3---training-model-using-prompts-with-rl) - - [Arg List](#arg-list-2) - - [Inference example - After Stage3](#inference-example---after-stage3) - - [Attention](#attention) - - [data](#data) - - [Support Model](#support-model) - - [GPT](#gpt) - - [BLOOM](#bloom) - - [OPT](#opt) - - [LLaMA](#llama) - - [Add your own models](#add-your-own-models) - - [Actor model](#actor-model) - - [Reward model](#reward-model) - - [Critic model](#critic-model) - ---- - -## Install requirements - -```shell -pip install -r requirements.txt -``` - -## Supervised datasets collection - -We collected 104K bilingual datasets of Chinese and English, and you can find the datasets in this repo -[InstructionWild](https://github.com/XueFuzhao/InstructionWild) and in this [file](https://github.com/XueFuzhao/InstructionWild/blob/main/data/README.md). - -Here is how we collected the data - -

      - -

      - -### Conversation dataset generation - -In order to further improve the model's ability to handle multi-turn conversations, we need to include samples with multi-turn conversations in the dataset. However, the samples in InstructWild and Alpaca datasets currently consist of only single-turn conversations, and their dataset organization is not suitable for storing multi-turn conversations. Additionally, after converting the aforementioned datasets, we also need to include multi-turn conversation datasets like ShareGPT, and we should transform them into the training format supported by ColossalChat. - -A sample of conversation dataset should have the following fields: - -- `type` (str, optional): The type of the data sample. -- `language` (str, optional): The language of the data sample. -- `dataset` (str, optional): The dataset the data sample originates from. -- `conversations` (str, compulsory): Conversation content of the data sample. -- `id` (int, optional): The ID of the data sample. - -A simple example: - -```json -{ - "type": "instruction", - "language": "English", - "dataset": "Alpaca", - "conversations": [ - { - "from": "human", - "value": "Give three tips for staying healthy." - }, - { - "from": "gpt", - "value": "1.Eat a balanced diet and make sure to include plenty of fruits and vegetables. \n2. Exercise regularly to keep your body active and strong. \n3. Get enough sleep and maintain a consistent sleep schedule." - } - ], - "id": 1 -} -``` - -> **NOTE:** Only key `conversations` is compulsary for training and other keys serve as metadata. The length of `conversations` varies. - -You can run the `examples/generate_conversation_dataset.py` to generate a conversation dataset supported by ColossalChat. - -You can use the following cmd to generate conversation dataset. - -```bash -python generate_conversation_dataset.py \ - --dataset "All" - --save_path "/path/to/dataset" -``` - -## Stage1 - Supervised instructs tuning - -Stage1 is supervised instructs fine-tuning, which uses the datasets mentioned earlier to fine-tune the model. -[[Stage1 tutorial video]](https://www.youtube.com/watch?v=-qFBZFmOJfg) - -You can run the `examples/train_sft.sh` to start a supervised instructs fine-tuning. - -You can also use the following cmd to start a supervised instructs fine-tuning with your own settings. - -```bash -torchrun --standalone --nproc_per_node=4 train_sft.py \ - --pretrain "/path/to/LLaMa-7B/" \ - --model 'llama' \ - --strategy colossalai_zero2 \ - --save_path /path/to/Coati-7B \ - --dataset /path/to/data.json \ - --batch_size 4 \ - --accumulation_steps 8 \ - --lr 2e-5 \ - --max_datasets_size 512 \ - --max_epochs 1 \ - --grad_checkpoint -``` - -**Note**: the supervised dataset follows the following format, - -```json -[ - { - "instruction": "Provide a list of the top 10 most popular mobile games in Asia", - "input": "", - "output": "The top 10 most popular mobile games in Asia are:\n1) PUBG Mobile\n2) Pokemon Go\n3) Candy Crush Saga\n4) Free Fire\n5) Clash of Clans\n6) Mario Kart Tour\n7) Arena of Valor\n8) Fantasy Westward Journey\n9) Subway Surfers\n10) ARK Survival Evolved", - "id": 0 - }, - ... -] -``` - -### Arg List - -- `--strategy`: the strategy using for training, choices=['ddp', 'colossalai_gemini', 'colossalai_zero2'], default='colossalai_zero2' -- `--model`: model type, choices=['gpt2', 'bloom', 'opt', 'llama'], default='bloom' -- `--pretrain`: pretrain model, type=str, default=None -- `--max_datasets_size`: the max size of dataset, type=int, default=None -- `--save_path`: path to save the model, type=str, default='output' -- `--need_optim_ckpt`: whether to save optim ckpt, type=bool, default=False -- `--max_epochs`: max epochs for training, type=int, default=3 -- `--batch_size`: batch size while training, type=int, default=4 -- `--lora_rank`: low-rank adaptation matrices rank, type=int, default=0 -- `--grad_checkpoint`: enable gradient checkpointing, type=bool, default=False - -## Stage2 - Training reward model - -We train a reward model in stage 2, which obtains corresponding scores by manually ranking different outputs for the same prompt and supervises the training of the reward model. -[[Stage2 tutorial video]](https://www.youtube.com/watch?v=gMx2CApKhuo) - -You can run the `examples/train_rm.sh` to start a reward model training. - -You can also use the following cmd to start training a reward model. - -```bash -torchrun --standalone --nproc_per_node=4 train_reward_model.py \ - --pretrain "/path/to/LLaMa-7B/" \ - --model 'llama' \ - --strategy colossalai_zero2 \ - --loss_fn 'log_exp'\ - --save_path 'rmstatic.pt' \ -``` - -### Features and tricks in RM training - -- We support [Anthropic/hh-rlhf](https://huggingface.co/datasets/Anthropic/hh-rlhf)and[rm-static](https://huggingface.co/datasets/Dahoas/rm-static) datasets. -- We support 2 kinds of loss function named `log_sig`(used by OpenAI) and `log_exp`(used by Anthropic). -- We change the loss to `valid_acc` and `pair_dist` to monitor progress during training. -- We add special token to the end of the sequence to get better result. -- We use cosine-reducing lr-scheduler for RM training. -- We set value_head as 1 liner layer and initialize the weight of value_head using N(0，1/(d_model + 1)) distribution. -- We train a Bloom-560m reward model for 1 epoch and find the test acc of the model achieve the performance mentions in [Anthropics paper](https://arxiv.org/abs/2204.05862). - -### Experiment result - -Model performance in [Anthropics paper](https://arxiv.org/abs/2204.05862): - -
      image - -
      Our training & test result of bloom-560m for 1 epoch: - -
      image - -
      We also train the reward model based on LLaMA-7B, which reaches the ACC of 72.06% after 1 epoch, performing almost the same as Anthropic's best RM. - -### Arg List - -- `--strategy`: the strategy using for training, choices=['ddp', 'colossalai_gemini', 'colossalai_zero2'], default='colossalai_zero2' -- `--model`: model type, choices=['gpt2', 'bloom', 'opt', 'llama'], default='bloom' -- `--pretrain`: pretrain model, type=str, default=None -- `--model_path`: the path of rm model(if continue to train), type=str, default=None -- `--save_path`: path to save the model, type=str, default='output' -- `--need_optim_ckpt`: whether to save optim ckpt, type=bool, default=False -- `--max_epochs`: max epochs for training, type=int, default=3 -- `--dataset`: dataset name, type=str, choices=['Anthropic/hh-rlhf', 'Dahoas/rm-static'] -- `--subset`: subset of the dataset, type=str, default=None -- `--batch_size`: batch size while training, type=int, default=4 -- `--lora_rank`: low-rank adaptation matrices rank, type=int, default=0 -- `--loss_func`: which kind of loss function, choices=['log_sig', 'log_exp'] -- `--max_len`: max sentence length for generation, type=int, default=512 - -## Stage3 - Training model using prompts with RL - -Stage3 uses reinforcement learning algorithm, which is the most complex part of the training process, as shown below: - -

      - -

      - -You can run the `examples/train_prompts.sh` to start PPO training. - -You can also use the cmd following to start PPO training. -[[Stage3 tutorial video]](https://www.youtube.com/watch?v=Z8wwSHxPL9g) - -```bash -torchrun --standalone --nproc_per_node=4 train_prompts.py \ - --pretrain "/path/to/LLaMa-7B/" \ - --model 'llama' \ - --strategy colossalai_zero2 \ - --prompt_dataset /path/to/your/prompt_dataset \ - --pretrain_dataset /path/to/your/pretrain_dataset \ - --rm_pretrain /your/pretrain/rm/definition \ - --rm_path /your/rm/model/path -``` - -Prompt dataset: the instruction dataset mentioned in the above figure which includes the instructions, e.g. you can use the [script](https://github.com/hpcaitech/ColossalAI/tree/main/applications/Chat/examples/generate_prompt_dataset.py) which samples `instinwild_en.json` or `instinwild_ch.json` in [InstructionWild](https://github.com/XueFuzhao/InstructionWild/tree/main/data#instructwild-data) to generate the prompt dataset. -Pretrain dataset: the pretrain dataset including the instruction and corresponding response, e.g. you can use the [InstructWild Data](https://github.com/XueFuzhao/InstructionWild/tree/main/data) in stage 1 supervised instructs tuning. - -**Note**: the required datasets follow the following format, - -- `pretrain dataset` - - ```json - [ - { - "instruction": "Provide a list of the top 10 most popular mobile games in Asia", - "input": "", - "output": "The top 10 most popular mobile games in Asia are:\n1) PUBG Mobile\n2) Pokemon Go\n3) Candy Crush Saga\n4) Free Fire\n5) Clash of Clans\n6) Mario Kart Tour\n7) Arena of Valor\n8) Fantasy Westward Journey\n9) Subway Surfers\n10) ARK Survival Evolved", - "id": 0 - }, - ... - ] - ``` - -- `prompt dataset` - - ```json - [ - { - "instruction": "Edit this paragraph to make it more concise: \"Yesterday, I went to the store and bought some things. Then, I came home and put them away. After that, I went for a walk and met some friends.\"", - "id": 0 - }, - { - "instruction": "Write a descriptive paragraph about a memorable vacation you went on", - "id": 1 - }, - ... - ] - ``` - -### Arg List - -- `--strategy`: the strategy using for training, choices=['ddp', 'colossalai_gemini', 'colossalai_zero2'], default='colossalai_zero2' -- `--model`: model type of actor, choices=['gpt2', 'bloom', 'opt', 'llama'], default='bloom' -- `--pretrain`: pretrain model, type=str, default=None -- `--rm_model`: reward model type, type=str, choices=['gpt2', 'bloom', 'opt', 'llama'], default=None -- `--rm_pretrain`: pretrain model for reward model, type=str, default=None -- `--rm_path`: the path of rm model, type=str, default=None -- `--save_path`: path to save the model, type=str, default='output' -- `--prompt_dataset`: path of the prompt dataset, type=str, default=None -- `--pretrain_dataset`: path of the ptx dataset, type=str, default=None -- `--need_optim_ckpt`: whether to save optim ckpt, type=bool, default=False -- `--num_episodes`: num of episodes for training, type=int, default=10 -- `--num_update_steps`: number of steps to update policy per episode, type=int -- `--num_collect_steps`: number of steps to collect experience per episode, type=int -- `--train_batch_size`: batch size while training, type=int, default=8 -- `--ptx_batch_size`: batch size to compute ptx loss, type=int, default=1 -- `--experience_batch_size`: batch size to make experience, type=int, default=8 -- `--lora_rank`: low-rank adaptation matrices rank, type=int, default=0 -- `--kl_coef`: kl_coef using for computing reward, type=float, default=0.1 -- `--ptx_coef`: ptx_coef using for computing policy loss, type=float, default=0.9 - -## Inference example - After Stage3 - -We support different inference options, including int8 and int4 quantization. -For details, see [`inference/`](https://github.com/hpcaitech/ColossalAI/tree/main/applications/Chat/inference). - -## Attention - -The examples are demos for the whole training process.You need to change the hyper-parameters to reach great performance. - -#### data - -- [x] [rm-static](https://huggingface.co/datasets/Dahoas/rm-static) -- [x] [hh-rlhf](https://huggingface.co/datasets/Anthropic/hh-rlhf) -- [ ] [openai/summarize_from_feedback](https://huggingface.co/datasets/openai/summarize_from_feedback) -- [ ] [openai/webgpt_comparisons](https://huggingface.co/datasets/openai/webgpt_comparisons) -- [ ] [Dahoas/instruct-synthetic-prompt-responses](https://huggingface.co/datasets/Dahoas/instruct-synthetic-prompt-responses) - -## Support Model - -### GPT - -- [x] GPT2-S (s) -- [x] GPT2-M (m) -- [x] GPT2-L (l) -- [x] GPT2-XL (xl) -- [x] GPT2-4B (4b) -- [ ] GPT2-6B (6b) - -### BLOOM - -- [x] [BLOOM-560m](https://huggingface.co/bigscience/bloom-560m) -- [x] [BLOOM-1b1](https://huggingface.co/bigscience/bloom-1b1) -- [x] [BLOOM-3b](https://huggingface.co/bigscience/bloom-3b) -- [x] [BLOOM-7b](https://huggingface.co/bigscience/bloom-7b1) -- [ ] [BLOOM-175b](https://huggingface.co/bigscience/bloom) - -### OPT - -- [x] [OPT-125M](https://huggingface.co/facebook/opt-125m) -- [x] [OPT-350M](https://huggingface.co/facebook/opt-350m) -- [x] [OPT-1.3B](https://huggingface.co/facebook/opt-1.3b) -- [x] [OPT-2.7B](https://huggingface.co/facebook/opt-2.7b) -- [x] [OPT-6.7B](https://huggingface.co/facebook/opt-6.7b) -- [ ] [OPT-13B](https://huggingface.co/facebook/opt-13b) -- [ ] [OPT-30B](https://huggingface.co/facebook/opt-30b) - -### [LLaMA](https://github.com/facebookresearch/llama/blob/main/MODEL_CARD.md) - -- [x] LLaMA-7B -- [x] LLaMA-13B -- [ ] LLaMA-33B -- [ ] LLaMA-65B - -## Add your own models - -If you want to support your own model in Coati, please refer the pull request for RoBERTa support as an example --[[chatgpt] add pre-trained model RoBERTa for RLHF stage 2 & 3](https://github.com/hpcaitech/ColossalAI/pull/3223), and submit a PR to us. - -You should complete the implementation of four model classes, including Reward model, Critic model, LM model, Actor model - -here are some example code for a NewModel named `Coati`. -if it is supported in huggingface [transformers](https://github.com/huggingface/transformers), you can load it by `from_pretrained`, o -r you can build your own model by yourself. - -### Actor model - -```python -from ..base import Actor -from transformers.models.coati import CoatiModel - -class CoatiActor(Actor): - def __init__(self, - pretrained: Optional[str] = None, - checkpoint: bool = False, - lora_rank: int = 0, - lora_train_bias: str = 'none') -> None: - if pretrained is not None: - model = CoatiModel.from_pretrained(pretrained) - else: - model = build_model() # load your own model if it is not support in transformers - - super().__init__(model, lora_rank, lora_train_bias) -``` - -### Reward model - -```python -from ..base import RewardModel -from transformers.models.coati import CoatiModel - -class CoatiRM(RewardModel): - - def __init__(self, - pretrained: Optional[str] = None, - checkpoint: bool = False, - lora_rank: int = 0, - lora_train_bias: str = 'none') -> None: - if pretrained is not None: - model = CoatiModel.from_pretrained(pretrained) - else: - model = build_model() # load your own model if it is not support in transformers - - value_head = nn.Linear(model.config.n_embd, 1) - value_head.weight.data.normal_(mean=0.0, std=1 / (model.config.n_embd + 1)) - super().__init__(model, value_head, lora_rank, lora_train_bias) -``` - -### Critic model - -```python -from ..base import Critic -from transformers.models.coati import CoatiModel - -class CoatiCritic(Critic): - def __init__(self, - pretrained: Optional[str] = None, - checkpoint: bool = False, - lora_rank: int = 0, - lora_train_bias: str = 'none') -> None: - if pretrained is not None: - model = CoatiModel.from_pretrained(pretrained) - else: - model = build_model() # load your own model if it is not support in transformers - - value_head = nn.Linear(model.config.n_embd, 1) - value_head.weight.data.normal_(mean=0.0, std=1 / (model.config.n_embd + 1)) - super().__init__(model, value_head, lora_rank, lora_train_bias) -``` diff --git a/applications/Chat/examples/download_model.py b/applications/Chat/examples/download_model.py deleted file mode 100644 index ec3482b5f789..000000000000 --- a/applications/Chat/examples/download_model.py +++ /dev/null @@ -1,79 +0,0 @@ -import argparse -import dataclasses -import os -import parser -from typing import List - -import tqdm -from coati.models.bloom import BLOOMRM, BLOOMActor, BLOOMCritic -from coati.models.gpt import GPTRM, GPTActor, GPTCritic -from coati.models.opt import OPTRM, OPTActor, OPTCritic -from huggingface_hub import hf_hub_download, snapshot_download -from transformers import AutoConfig, AutoTokenizer, BloomConfig, BloomTokenizerFast, GPT2Config, GPT2Tokenizer - - -@dataclasses.dataclass -class HFRepoFiles: - repo_id: str - files: List[str] - - def download(self, dir_path: str): - for file in self.files: - file_path = hf_hub_download(self.repo_id, file, local_dir=dir_path) - - def download_all(self): - snapshot_download(self.repo_id) - - -def test_init(model: str, dir_path: str): - if model == "gpt2": - config = GPT2Config.from_pretrained(dir_path) - actor = GPTActor(config=config) - critic = GPTCritic(config=config) - reward_model = GPTRM(config=config) - GPT2Tokenizer.from_pretrained(dir_path) - elif model == "bloom": - config = BloomConfig.from_pretrained(dir_path) - actor = BLOOMActor(config=config) - critic = BLOOMCritic(config=config) - reward_model = BLOOMRM(config=config) - BloomTokenizerFast.from_pretrained(dir_path) - elif model == "opt": - config = AutoConfig.from_pretrained(dir_path) - actor = OPTActor(config=config) - critic = OPTCritic(config=config) - reward_model = OPTRM(config=config) - AutoTokenizer.from_pretrained(dir_path) - else: - raise NotImplementedError(f"Model {model} not implemented") - - -if __name__ == "__main__": - parser = argparse.ArgumentParser() - parser.add_argument("--model-dir", type=str, default="test_models") - parser.add_argument("--config-only", default=False, action="store_true") - args = parser.parse_args() - - if os.path.exists(args.model_dir): - print(f"[INFO]: {args.model_dir} already exists") - exit(0) - - repo_list = { - "gpt2": HFRepoFiles(repo_id="gpt2", files=["config.json", "tokenizer.json", "vocab.json", "merges.txt"]), - "bloom": HFRepoFiles( - repo_id="bigscience/bloom-560m", files=["config.json", "tokenizer.json", "tokenizer_config.json"] - ), - "opt": HFRepoFiles( - repo_id="facebook/opt-350m", files=["config.json", "tokenizer_config.json", "vocab.json", "merges.txt"] - ), - } - - os.mkdir(args.model_dir) - for model_name in tqdm.tqdm(repo_list): - dir_path = os.path.join(args.model_dir, model_name) - if args.config_only: - os.mkdir(dir_path) - repo_list[model_name].download(dir_path) - else: - repo_list[model_name].download_all() - test_init(model_name, dir_path) diff --git a/applications/Chat/examples/generate_conversation_dataset.py b/applications/Chat/examples/generate_conversation_dataset.py deleted file mode 100644 index 7e03b2d54260..000000000000 --- a/applications/Chat/examples/generate_conversation_dataset.py +++ /dev/null @@ -1,82 +0,0 @@ -import argparse -import json - -from datasets import load_dataset - - -def generate_alpaca(): - # We can convert dataset with the same format("instruction", "input", "output") as Alpaca into a one-round conversation. - conversation_dataset = [] - dataset = load_dataset("tatsu-lab/alpaca", split="train") - - instructions = dataset["instruction"] - inputs = dataset["input"] - outputs = dataset["output"] - - assert len(instructions) == len(inputs) == len(outputs) - - for idx in range(len(instructions)): - human_utterance = instructions[idx] + "\n\n" + inputs[idx] if inputs[idx] else instructions[idx] - human = {"from": "human", "value": human_utterance} - - gpt_utterance = outputs[idx] - gpt = {"from": "gpt", "value": gpt_utterance} - - conversation = dict(type="instruction", language="English", dataset="Alpaca", conversations=[human, gpt]) - conversation_dataset.append(conversation) - - return conversation_dataset - - -def generate_sharegpt(): - # ShareGPT data requires less processing. - conversation_dataset = [] - dataset = load_dataset( - "anon8231489123/ShareGPT_Vicuna_unfiltered", - data_files="ShareGPT_V3_unfiltered_cleaned_split_no_imsorry.json", - split="train", - ) - - conversations = dataset["conversations"] - - for idx in range(len(conversations)): - for conv in conversations[idx]: - # We don't need markdown and text value. - del conv["markdown"] - del conv["text"] - - conversation = dict( - type="conversation", language="Multilingual", dataset="ShareGPT", conversations=conversations[idx] - ) - conversation_dataset.append(conversation) - - return conversation_dataset - - -if __name__ == "__main__": - parser = argparse.ArgumentParser() - parser.add_argument( - "--dataset", - type=str, - default="All", - choices=["Alpaca", "ShareGPT", "All"], - help="which dataset to convert, All will combine Alpaca and ShareGPT", - ) - parser.add_argument("--save_path", type=str, default="dataset.json", help="path to save the converted dataset") - args = parser.parse_args() - - conversation_dataset = [] - - if args.dataset == "Alpaca": - conversation_dataset.extend(generate_alpaca()) - elif args.dataset == "ShareGPT": - conversation_dataset.extend(generate_sharegpt()) - else: - conversation_dataset.extend(generate_alpaca()) - conversation_dataset.extend(generate_sharegpt()) - - for idx, sample in enumerate(conversation_dataset): - sample["id"] = idx + 1 - - with open(args.save_path, mode="w") as f: - json.dump(conversation_dataset, f, indent=4, default=str, ensure_ascii=False) diff --git a/applications/Chat/examples/generate_prompt_dataset.py b/applications/Chat/examples/generate_prompt_dataset.py deleted file mode 100644 index 4eec6feae505..000000000000 --- a/applications/Chat/examples/generate_prompt_dataset.py +++ /dev/null @@ -1,27 +0,0 @@ -import argparse -import json -import random - -random.seed(42) - - -def sample(args): - with open(args.dataset_path, mode="r") as f: - dataset_list = json.load(f) - - sampled_dataset = [ - {"instruction": sample["instruction"], "id": idx} - for idx, sample in enumerate(random.sample(dataset_list, args.sample_size)) - ] - - with open(args.save_path, mode="w") as f: - json.dump(sampled_dataset, f, indent=4, default=str, ensure_ascii=False) - - -if __name__ == "__main__": - parser = argparse.ArgumentParser() - parser.add_argument("--dataset_path", type=str, default=None, required=True, help="path to the pretrain dataset") - parser.add_argument("--save_path", type=str, default="prompt.json", help="path to save the prompt dataset") - parser.add_argument("--sample_size", type=int, default=16384, help="size of the prompt dataset") - args = parser.parse_args() - sample(args) diff --git a/applications/Chat/examples/inference.py b/applications/Chat/examples/inference.py deleted file mode 100644 index 9df8649d9c61..000000000000 --- a/applications/Chat/examples/inference.py +++ /dev/null @@ -1,73 +0,0 @@ -import argparse - -import torch -from coati.models.bloom import BLOOMActor -from coati.models.generation import generate -from coati.models.gpt import GPTActor -from coati.models.llama import LlamaActor -from coati.models.opt import OPTActor -from transformers import AutoTokenizer, BloomTokenizerFast, GPT2Tokenizer, LlamaTokenizer - - -def eval(args): - # configure model - if args.model == "gpt2": - actor = GPTActor(pretrained=args.pretrain) - elif args.model == "bloom": - actor = BLOOMActor(pretrained=args.pretrain) - elif args.model == "opt": - actor = OPTActor(pretrained=args.pretrain) - elif args.model == "llama": - actor = LlamaActor(pretrained=args.pretrain) - else: - raise ValueError(f'Unsupported model "{args.model}"') - - actor.to(torch.cuda.current_device()) - if args.model_path is not None: - state_dict = torch.load(args.model_path) - actor.load_state_dict(state_dict) - - # configure tokenizer - if args.model == "gpt2": - tokenizer = GPT2Tokenizer.from_pretrained("gpt2") - tokenizer.pad_token = tokenizer.eos_token - elif args.model == "bloom": - tokenizer = BloomTokenizerFast.from_pretrained("bigscience/bloom-560m") - tokenizer.pad_token = tokenizer.eos_token - elif args.model == "opt": - tokenizer = AutoTokenizer.from_pretrained("facebook/opt-350m") - tokenizer.pad_token = tokenizer.eos_token - elif args.model == "llama": - tokenizer = LlamaTokenizer.from_pretrained("hf-internal-testing/llama-tokenizer") - tokenizer.eos_token = "" - tokenizer.pad_token = tokenizer.unk_token - else: - raise ValueError(f'Unsupported model "{args.model}"') - - actor.eval() - tokenizer.padding_side = "left" - input_ids = tokenizer.encode(args.input, return_tensors="pt").to(torch.cuda.current_device()) - outputs = generate( - actor, - input_ids, - tokenizer=tokenizer, - max_length=args.max_length, - do_sample=True, - top_k=50, - top_p=0.95, - num_return_sequences=1, - ) - output = tokenizer.batch_decode(outputs[0], skip_special_tokens=True) - print(f"[Output]: {''.join(output)}") - - -if __name__ == "__main__": - parser = argparse.ArgumentParser() - parser.add_argument("--model", default="gpt2", choices=["gpt2", "bloom", "opt", "llama"]) - # We suggest to use the pretrained model from HuggingFace, use pretrain to configure model - parser.add_argument("--pretrain", type=str, default=None) - parser.add_argument("--model_path", type=str, default=None) - parser.add_argument("--input", type=str, default="Question: How are you ? Answer:") - parser.add_argument("--max_length", type=int, default=100) - args = parser.parse_args() - eval(args) diff --git a/applications/Chat/examples/train_prompts.py b/applications/Chat/examples/train_prompts.py deleted file mode 100644 index 40e06043ab57..000000000000 --- a/applications/Chat/examples/train_prompts.py +++ /dev/null @@ -1,249 +0,0 @@ -import argparse -import warnings - -import torch -import torch.distributed as dist -from coati.dataset import PromptDataset, SupervisedDataset -from coati.models.bloom import BLOOMRM, BLOOMActor, BLOOMCritic -from coati.models.gpt import GPTRM, GPTActor, GPTCritic -from coati.models.llama import LlamaActor, LlamaCritic, LlamaRM -from coati.models.opt import OPTRM, OPTActor, OPTCritic -from coati.trainer import PPOTrainer -from coati.trainer.strategies import DDPStrategy, GeminiStrategy, LowLevelZeroStrategy -from torch.optim import Adam -from torch.utils.data import DataLoader -from torch.utils.data.distributed import DistributedSampler -from transformers import AutoTokenizer, BloomTokenizerFast, GPT2Tokenizer, LlamaTokenizer - -from colossalai.nn.optimizer import HybridAdam - - -def main(args): - # configure strategy - if args.strategy == "ddp": - strategy = DDPStrategy() - elif args.strategy == "colossalai_gemini": - strategy = GeminiStrategy(placement_policy="static", initial_scale=2**5) - elif args.strategy == "colossalai_zero2": - strategy = LowLevelZeroStrategy(stage=2, placement_policy="cuda") - else: - raise ValueError(f'Unsupported strategy "{args.strategy}"') - - if args.rm_path is not None: - warnings.warn("LoRA weights should be merged with the model weights") - state_dict = torch.load(args.rm_path, map_location="cpu") - - if args.lora_rank > 0: - warnings.warn("Lora is not supported yet.") - args.lora_rank = 0 - - with strategy.model_init_context(): - # configure model - if args.model == "gpt2": - initial_model = GPTActor(pretrained=args.pretrain) - elif args.model == "bloom": - initial_model = BLOOMActor(pretrained=args.pretrain) - elif args.model == "opt": - initial_model = OPTActor(pretrained=args.pretrain) - elif args.model == "llama": - initial_model = LlamaActor(pretrained=args.pretrain) - else: - raise ValueError(f'Unsupported actor model "{args.model}"') - - if args.rm_model is None: - rm_model_name = args.model - else: - rm_model_name = args.rm_model - - if rm_model_name == "gpt2": - reward_model = GPTRM(pretrained=args.rm_pretrain, lora_rank=args.lora_rank) - elif rm_model_name == "bloom": - reward_model = BLOOMRM(pretrained=args.rm_pretrain, lora_rank=args.lora_rank) - elif rm_model_name == "opt": - reward_model = OPTRM(pretrained=args.rm_pretrain, lora_rank=args.lora_rank) - elif rm_model_name == "llama": - reward_model = LlamaRM(pretrained=args.rm_pretrain, lora_rank=args.lora_rank) - else: - raise ValueError(f'Unsupported reward model "{rm_model_name}"') - - if args.rm_path is not None: - reward_model.load_state_dict(state_dict, strict=False) - - initial_model.to(torch.bfloat16).to(torch.cuda.current_device()) - reward_model.to(torch.bfloat16).to(torch.cuda.current_device()) - - if args.model == "gpt2": - actor = GPTActor(pretrained=args.pretrain, lora_rank=args.lora_rank) - elif args.model == "bloom": - actor = BLOOMActor(pretrained=args.pretrain, lora_rank=args.lora_rank) - elif args.model == "opt": - actor = OPTActor(pretrained=args.pretrain, lora_rank=args.lora_rank) - elif args.model == "llama": - actor = LlamaActor(pretrained=args.pretrain, lora_rank=args.lora_rank) - else: - raise ValueError(f'Unsupported actor model "{args.model}"') - - if rm_model_name == "gpt2": - critic = GPTCritic(pretrained=args.rm_pretrain, lora_rank=args.lora_rank) - elif rm_model_name == "bloom": - critic = BLOOMCritic(pretrained=args.rm_pretrain, lora_rank=args.lora_rank) - elif rm_model_name == "opt": - critic = OPTCritic(pretrained=args.rm_pretrain, lora_rank=args.lora_rank) - elif rm_model_name == "llama": - critic = LlamaCritic(pretrained=args.rm_pretrain, lora_rank=args.lora_rank) - else: - raise ValueError(f'Unsupported reward model "{rm_model_name}"') - - if args.rm_path is not None: - critic.load_state_dict(state_dict, strict=False) - del state_dict - - actor.to(torch.bfloat16).to(torch.cuda.current_device()) - critic.to(torch.bfloat16).to(torch.cuda.current_device()) - - # configure optimizer - if args.strategy.startswith("colossalai"): - actor_optim = HybridAdam(actor.parameters(), lr=args.lr) - critic_optim = HybridAdam(critic.parameters(), lr=args.lr) - else: - actor_optim = Adam(actor.parameters(), lr=args.lr) - critic_optim = Adam(critic.parameters(), lr=args.lr) - - # configure tokenizer - if args.model == "gpt2": - tokenizer = GPT2Tokenizer.from_pretrained("gpt2" if args.tokenizer is None else args.tokenizer) - tokenizer.pad_token = tokenizer.eos_token - elif args.model == "bloom": - tokenizer = BloomTokenizerFast.from_pretrained( - "bigscience/bloom-560m" if args.tokenizer is None else args.tokenizer - ) - tokenizer.pad_token = tokenizer.eos_token - elif args.model == "opt": - tokenizer = AutoTokenizer.from_pretrained("facebook/opt-350m" if args.tokenizer is None else args.tokenizer) - tokenizer.pad_token = tokenizer.eos_token - elif args.model == "llama": - tokenizer = LlamaTokenizer.from_pretrained( - "hf-internal-testing/llama-tokenizer" if args.tokenizer is None else args.tokenizer - ) - tokenizer.eos_token = "" - tokenizer.pad_token = tokenizer.unk_token - else: - raise ValueError(f'Unsupported model "{args.model}"') - # NOTE: generate() requires padding_side to be "left" - tokenizer.padding_side = "left" - - prompt_dataset = PromptDataset( - tokenizer=tokenizer, - data_path=args.prompt_dataset, - max_datasets_size=args.max_datasets_size, - max_length=args.max_input_len, - ) - if dist.is_initialized() and dist.get_world_size() > 1: - prompt_sampler = DistributedSampler(prompt_dataset, shuffle=True, seed=42, drop_last=True) - else: - prompt_sampler = None - prompt_dataloader = DataLoader( - prompt_dataset, shuffle=(prompt_sampler is None), sampler=prompt_sampler, batch_size=args.experience_batch_size - ) - - pretrain_dataset = SupervisedDataset( - tokenizer=tokenizer, - data_path=args.pretrain_dataset, - max_datasets_size=args.max_datasets_size, - max_length=args.max_input_len, - ) - if dist.is_initialized() and dist.get_world_size() > 1: - pretrain_sampler = DistributedSampler(pretrain_dataset, shuffle=True, seed=42, drop_last=True) - else: - pretrain_sampler = None - pretrain_dataloader = DataLoader( - pretrain_dataset, shuffle=(pretrain_sampler is None), sampler=pretrain_sampler, batch_size=args.ptx_batch_size - ) - - # NOTE: For small models like opt-1.3b, reward model and initial model are not required to be parallelized. - (actor, actor_optim), (critic, critic_optim), reward_model, initial_model = strategy.prepare( - (actor, actor_optim), (critic, critic_optim), reward_model, initial_model - ) - - # configure trainer - trainer = PPOTrainer( - strategy, - actor, - critic, - reward_model, - initial_model, - actor_optim, - critic_optim, - tokenizer=tokenizer, - kl_coef=args.kl_coef, - ptx_coef=args.ptx_coef, - train_batch_size=args.train_batch_size, - max_length=args.max_seq_len, - use_cache=True, - do_sample=True, - temperature=1.0, - top_k=50, - offload_inference_models=args.strategy != "colossalai_gemini", - ) - - trainer.fit( - num_episodes=args.num_episodes, - num_collect_steps=args.num_collect_steps, - num_update_steps=args.num_update_steps, - prompt_dataloader=prompt_dataloader, - pretrain_dataloader=pretrain_dataloader, - log_dir=args.log_dir, - use_wandb=args.use_wandb, - ) - - if args.lora_rank > 0 and args.merge_lora_weights: - from coati.models.lora import LORA_MANAGER - - # NOTE: set model to eval to merge LoRA weights - LORA_MANAGER.merge_weights = True - actor.eval() - # save model checkpoint after fitting - strategy.save_pretrained(actor, path=args.save_path) - # save optimizer checkpoint on all ranks - if args.need_optim_ckpt: - strategy.save_optimizer( - actor_optim, "actor_optim_checkpoint_prompts_%d.pt" % (torch.cuda.current_device()), only_rank0=False - ) - - -if __name__ == "__main__": - parser = argparse.ArgumentParser() - parser.add_argument("--prompt_dataset", type=str, default=None, help="path to the prompt dataset") - parser.add_argument("--pretrain_dataset", type=str, default=None, help="path to the pretrained dataset") - parser.add_argument("--max_datasets_size", type=int, default=50000) - parser.add_argument( - "--strategy", - choices=["ddp", "colossalai_gemini", "colossalai_zero2"], - default="colossalai_zero2", - help="strategy to use", - ) - parser.add_argument("--model", default="gpt2", choices=["gpt2", "bloom", "opt", "llama"]) - parser.add_argument("--tokenizer", type=str, default=None) - parser.add_argument("--pretrain", type=str, default=None) - parser.add_argument("--rm_model", default=None, choices=["gpt2", "bloom", "opt", "llama"]) - parser.add_argument("--rm_path", type=str, default=None) - parser.add_argument("--rm_pretrain", type=str, default=None) - parser.add_argument("--save_path", type=str, default="actor_checkpoint_prompts") - parser.add_argument("--need_optim_ckpt", type=bool, default=False) - parser.add_argument("--num_episodes", type=int, default=10) - parser.add_argument("--num_collect_steps", type=int, default=10) - parser.add_argument("--num_update_steps", type=int, default=5) - parser.add_argument("--train_batch_size", type=int, default=8) - parser.add_argument("--ptx_batch_size", type=int, default=1) - parser.add_argument("--experience_batch_size", type=int, default=8) - parser.add_argument("--lora_rank", type=int, default=0, help="low-rank adaptation matrices rank") - parser.add_argument("--merge_lora_weights", type=bool, default=True) - parser.add_argument("--lr", type=float, default=1e-7) - parser.add_argument("--kl_coef", type=float, default=0.1) - parser.add_argument("--ptx_coef", type=float, default=0.9) - parser.add_argument("--max_input_len", type=int, default=96) - parser.add_argument("--max_seq_len", type=int, default=128) - parser.add_argument("--log_dir", default="logs", type=str) - parser.add_argument("--use_wandb", default=False, action="store_true") - args = parser.parse_args() - main(args) diff --git a/applications/Chat/examples/train_prompts.sh b/applications/Chat/examples/train_prompts.sh deleted file mode 100755 index d04c416015b1..000000000000 --- a/applications/Chat/examples/train_prompts.sh +++ /dev/null @@ -1,25 +0,0 @@ -set_n_least_used_CUDA_VISIBLE_DEVICES() { - local n=${1:-"9999"} - echo "GPU Memory Usage:" - local FIRST_N_GPU_IDS=$(nvidia-smi --query-gpu=memory.used --format=csv | - tail -n +2 | - nl -v 0 | - tee /dev/tty | - sort -g -k 2 | - awk '{print $1}' | - head -n $n) - export CUDA_VISIBLE_DEVICES=$(echo $FIRST_N_GPU_IDS | sed 's/ /,/g') - echo "Now CUDA_VISIBLE_DEVICES is set to:" - echo "CUDA_VISIBLE_DEVICES=$CUDA_VISIBLE_DEVICES" -} - -set_n_least_used_CUDA_VISIBLE_DEVICES 2 - -# torchrun --standalone --nproc_per_node=2 train_prompts.py prompts.csv --strategy colossalai_zero2 - -torchrun --standalone --nproc_per_node=2 train_prompts.py \ - --pretrain_dataset /path/to/data.json \ - --prompt_dataset /path/to/data.json \ - --strategy colossalai_zero2 \ - --num_episodes 1 --num_collect_steps 2 --num_update_steps 1 \ - --train_batch_size 2 diff --git a/applications/Chat/examples/train_reward_model.py b/applications/Chat/examples/train_reward_model.py deleted file mode 100644 index fcdd29b2954b..000000000000 --- a/applications/Chat/examples/train_reward_model.py +++ /dev/null @@ -1,208 +0,0 @@ -import argparse -import warnings - -import torch -import torch.distributed as dist -from coati.dataset import HhRlhfDataset, RmStaticDataset -from coati.models import LogExpLoss, LogSigLoss -from coati.models.bloom import BLOOMRM -from coati.models.gpt import GPTRM -from coati.models.llama import LlamaRM -from coati.models.opt import OPTRM -from coati.trainer import RewardModelTrainer -from coati.trainer.strategies import DDPStrategy, GeminiStrategy, LowLevelZeroStrategy -from datasets import load_dataset -from torch.optim import Adam -from torch.optim.lr_scheduler import CosineAnnealingLR -from torch.utils.data import DataLoader -from torch.utils.data.distributed import DistributedSampler -from transformers import AutoTokenizer, BloomTokenizerFast, LlamaTokenizer -from transformers.models.gpt2.tokenization_gpt2 import GPT2Tokenizer - -from colossalai.nn.optimizer import HybridAdam - - -def train(args): - # configure strategy - if args.strategy == "ddp": - strategy = DDPStrategy() - elif args.strategy == "colossalai_gemini": - strategy = GeminiStrategy(placement_policy="auto") - elif args.strategy == "colossalai_zero2": - strategy = LowLevelZeroStrategy(stage=2, placement_policy="cuda") - else: - raise ValueError(f'Unsupported strategy "{args.strategy}"') - - # configure model - if args.lora_rank > 0: - warnings.warn("Lora is not supported yet.") - args.lora_rank = 0 - - with strategy.model_init_context(): - if args.model == "bloom": - model = BLOOMRM(pretrained=args.pretrain, lora_rank=args.lora_rank) - elif args.model == "opt": - model = OPTRM(pretrained=args.pretrain, lora_rank=args.lora_rank) - elif args.model == "gpt2": - model = GPTRM(pretrained=args.pretrain, lora_rank=args.lora_rank) - elif args.model == "llama": - model = LlamaRM(pretrained=args.pretrain, lora_rank=args.lora_rank) - else: - raise ValueError(f'Unsupported model "{args.model}"') - - model.to(torch.bfloat16).to(torch.cuda.current_device()) - - if args.model_path is not None: - state_dict = torch.load(args.model_path) - model.load_state_dict(state_dict) - - # configure tokenizer - if args.model == "gpt2": - tokenizer = GPT2Tokenizer.from_pretrained("gpt2" if args.tokenizer is None else args.tokenizer) - tokenizer.pad_token = tokenizer.eos_token - elif args.model == "bloom": - tokenizer = BloomTokenizerFast.from_pretrained( - "bigscience/bloom-560m" if args.tokenizer is None else args.tokenizer - ) - tokenizer.pad_token = tokenizer.eos_token - elif args.model == "opt": - tokenizer = AutoTokenizer.from_pretrained("facebook/opt-350m" if args.tokenizer is None else args.tokenizer) - tokenizer.pad_token = tokenizer.eos_token - elif args.model == "llama": - tokenizer = LlamaTokenizer.from_pretrained( - "hf-internal-testing/llama-tokenizer" if args.tokenizer is None else args.tokenizer - ) - tokenizer.eos_token = "" - tokenizer.pad_token = tokenizer.unk_token - else: - raise ValueError(f'Unsupported model "{args.model}"') - - # configure optimizer - if args.strategy.startswith("colossalai"): - optim = HybridAdam(model.parameters(), lr=args.lr) - else: - optim = Adam(model.parameters(), lr=args.lr) - - # configure loss function - if args.loss_fn == "log_sig": - loss_fn = LogSigLoss() - elif args.loss_fn == "log_exp": - loss_fn = LogExpLoss() - else: - raise ValueError(f'Unsupported loss function "{args.loss_fn}"') - - # prepare for data and dataset - if args.subset is not None: - data = load_dataset(args.dataset, data_dir=args.subset) - else: - data = load_dataset(args.dataset) - - train_data = data["train"].select(range(min(args.max_datasets_size, len(data["train"])))) - eval_data = data["test"].select(range(min(args.max_datasets_size, len(data["test"])))) - - if args.dataset == "Dahoas/rm-static": - train_dataset = RmStaticDataset(train_data, tokenizer, args.max_len) - eval_dataset = RmStaticDataset(eval_data, tokenizer, args.max_len) - elif args.dataset == "Anthropic/hh-rlhf": - train_dataset = HhRlhfDataset(train_data, tokenizer, args.max_len) - eval_dataset = HhRlhfDataset(eval_data, tokenizer, args.max_len) - else: - raise ValueError(f'Unsupported dataset "{args.dataset}"') - - if dist.is_initialized() and dist.get_world_size() > 1: - train_sampler = DistributedSampler( - train_dataset, - shuffle=True, - seed=42, - drop_last=True, - rank=dist.get_rank(), - num_replicas=dist.get_world_size(), - ) - eval_sampler = DistributedSampler( - eval_dataset, - shuffle=True, - seed=42, - drop_last=True, - rank=dist.get_rank(), - num_replicas=dist.get_world_size(), - ) - else: - train_sampler = None - eval_sampler = None - - train_dataloader = DataLoader( - train_dataset, - shuffle=(train_sampler is None), - sampler=train_sampler, - batch_size=args.batch_size, - pin_memory=True, - ) - - eval_dataloader = DataLoader( - eval_dataset, shuffle=(eval_sampler is None), sampler=eval_sampler, batch_size=args.batch_size, pin_memory=True - ) - - lr_scheduler = CosineAnnealingLR(optim, train_dataloader.__len__() // 100) - strategy_dict = strategy.prepare(dict(model=model, optimizer=optim, lr_scheduler=lr_scheduler)) - model = strategy_dict["model"] - optim = strategy_dict["optimizer"] - lr_scheduler = strategy_dict["lr_scheduler"] - trainer = RewardModelTrainer( - model=model, - strategy=strategy, - optim=optim, - lr_scheduler=lr_scheduler, - loss_fn=loss_fn, - max_epochs=args.max_epochs, - ) - - trainer.fit( - train_dataloader=train_dataloader, - eval_dataloader=eval_dataloader, - log_dir=args.log_dir, - use_wandb=args.use_wandb, - ) - - if args.lora_rank > 0 and args.merge_lora_weights: - from coati.models.lora import LORA_MANAGER - - # NOTE: set model to eval to merge LoRA weights - LORA_MANAGER.merge_weights = True - model.eval() - # save model checkpoint after fitting on only rank0 - state_dict = model.state_dict() - torch.save(state_dict, args.save_path) - # save optimizer checkpoint on all ranks - if args.need_optim_ckpt: - strategy.save_optimizer( - trainer.optimizer, "rm_optim_checkpoint_%d.pt" % (torch.cuda.current_device()), only_rank0=False - ) - - -if __name__ == "__main__": - parser = argparse.ArgumentParser() - parser.add_argument( - "--strategy", choices=["ddp", "colossalai_gemini", "colossalai_zero2"], default="colossalai_zero2" - ) - parser.add_argument("--model", choices=["gpt2", "bloom", "opt", "llama"], default="bloom") - parser.add_argument("--tokenizer", type=str, default=None) - parser.add_argument("--pretrain", type=str, default=None) - parser.add_argument("--model_path", type=str, default=None) - parser.add_argument("--need_optim_ckpt", type=bool, default=False) - parser.add_argument( - "--dataset", type=str, choices=["Anthropic/hh-rlhf", "Dahoas/rm-static"], default="Dahoas/rm-static" - ) - parser.add_argument("--subset", type=lambda x: None if x == "None" else x, default=None) - parser.add_argument("--max_datasets_size", type=int, default=1000000) - parser.add_argument("--save_path", type=str, default="rm_ckpt") - parser.add_argument("--max_epochs", type=int, default=1) - parser.add_argument("--batch_size", type=int, default=1) - parser.add_argument("--max_len", type=int, default=512) - parser.add_argument("--lora_rank", type=int, default=0, help="low-rank adaptation matrices rank") - parser.add_argument("--merge_lora_weights", type=bool, default=True) - parser.add_argument("--lr", type=float, default=9e-6) - parser.add_argument("--loss_fn", type=str, default="log_sig", choices=["log_sig", "log_exp"]) - parser.add_argument("--log_dir", default="logs", type=str) - parser.add_argument("--use_wandb", default=False, action="store_true") - args = parser.parse_args() - train(args) diff --git a/applications/Chat/examples/train_rm.sh b/applications/Chat/examples/train_rm.sh deleted file mode 100755 index c5ebaf708ddc..000000000000 --- a/applications/Chat/examples/train_rm.sh +++ /dev/null @@ -1,25 +0,0 @@ -set_n_least_used_CUDA_VISIBLE_DEVICES() { - local n=${1:-"9999"} - echo "GPU Memory Usage:" - local FIRST_N_GPU_IDS=$(nvidia-smi --query-gpu=memory.used --format=csv | - tail -n +2 | - nl -v 0 | - tee /dev/tty | - sort -g -k 2 | - awk '{print $1}' | - head -n $n) - export CUDA_VISIBLE_DEVICES=$(echo $FIRST_N_GPU_IDS | sed 's/ /,/g') - echo "Now CUDA_VISIBLE_DEVICES is set to:" - echo "CUDA_VISIBLE_DEVICES=$CUDA_VISIBLE_DEVICES" -} - -set_n_least_used_CUDA_VISIBLE_DEVICES 2 - -torchrun --standalone --nproc_per_node=2 train_reward_model.py \ - --pretrain 'gpt2' \ - --model 'gpt2' \ - --strategy colossalai_zero2 \ - --loss_fn 'log_exp' \ - --dataset 'Anthropic/hh-rlhf' \ - --batch_size 16 \ - --max_epochs 10 diff --git a/applications/Chat/examples/train_sft.py b/applications/Chat/examples/train_sft.py deleted file mode 100644 index d00c04809a2d..000000000000 --- a/applications/Chat/examples/train_sft.py +++ /dev/null @@ -1,221 +0,0 @@ -import argparse -import math -import warnings - -import torch -import torch.distributed as dist -from coati.dataset import SFTDataset, SupervisedDataset -from coati.models.bloom import BLOOMActor -from coati.models.chatglm import ChatGLMActor -from coati.models.chatglm.chatglm_tokenizer import ChatGLMTokenizer -from coati.models.gpt import GPTActor -from coati.models.llama import LlamaActor -from coati.models.opt import OPTActor -from coati.trainer import SFTTrainer -from coati.trainer.strategies import DDPStrategy, GeminiStrategy, LowLevelZeroStrategy -from datasets import load_dataset -from torch.optim import Adam -from torch.utils.data import DataLoader -from torch.utils.data.distributed import DistributedSampler -from transformers import AutoTokenizer, BloomTokenizerFast, LlamaTokenizer -from transformers.models.gpt2.tokenization_gpt2 import GPT2Tokenizer -from transformers.trainer import get_scheduler - -from colossalai.logging import get_dist_logger -from colossalai.nn.optimizer import HybridAdam - - -def train(args): - # configure strategy - if args.strategy == "ddp": - strategy = DDPStrategy() - elif args.strategy == "colossalai_gemini": - strategy = GeminiStrategy(placement_policy="auto") - elif args.strategy == "colossalai_zero2": - strategy = LowLevelZeroStrategy(stage=2, placement_policy="cuda") - elif args.strategy == "colossalai_zero2_cpu": - strategy = LowLevelZeroStrategy(stage=2, placement_policy="cpu") - else: - raise ValueError(f'Unsupported strategy "{args.strategy}"') - - # configure model - if args.lora_rank > 0: - warnings.warn("Lora is not supported yet.") - args.lora_rank = 0 - - with strategy.model_init_context(): - if args.model == "bloom": - model = BLOOMActor(pretrained=args.pretrain, lora_rank=args.lora_rank, checkpoint=args.grad_checkpoint) - elif args.model == "opt": - model = OPTActor(pretrained=args.pretrain, lora_rank=args.lora_rank, checkpoint=args.grad_checkpoint) - elif args.model == "gpt2": - model = GPTActor(pretrained=args.pretrain, lora_rank=args.lora_rank, checkpoint=args.grad_checkpoint) - elif args.model == "llama": - model = LlamaActor(pretrained=args.pretrain, lora_rank=args.lora_rank, checkpoint=args.grad_checkpoint) - elif args.model == "chatglm": - model = ChatGLMActor(pretrained=args.pretrain) - else: - raise ValueError(f'Unsupported model "{args.model}"') - - model.to(torch.bfloat16).to(torch.cuda.current_device()) - - # configure tokenizer - if args.model == "gpt2": - tokenizer = GPT2Tokenizer.from_pretrained("gpt2" if args.tokenizer is None else args.tokenizer) - tokenizer.pad_token = tokenizer.eos_token - elif args.model == "bloom": - tokenizer = BloomTokenizerFast.from_pretrained( - "bigscience/bloom-560m" if args.tokenizer is None else args.tokenizer - ) - tokenizer.pad_token = tokenizer.eos_token - elif args.model == "opt": - tokenizer = AutoTokenizer.from_pretrained("facebook/opt-350m" if args.tokenizer is None else args.tokenizer) - tokenizer.pad_token = tokenizer.eos_token - elif args.model == "llama": - tokenizer = LlamaTokenizer.from_pretrained( - "hf-internal-testing/llama-tokenizer" if args.tokenizer is None else args.tokenizer - ) - tokenizer.eos_token = "" - tokenizer.pad_token = tokenizer.unk_token - elif args.model == "chatglm": - tokenizer = ChatGLMTokenizer.from_pretrained( - "THUDM/chatglm-6b" if args.tokenizer is None else args.tokenizer, trust_remote_code=True - ) - else: - raise ValueError(f'Unsupported model "{args.model}"') - - # configure optimizer - if args.strategy.startswith("colossalai"): - optim = HybridAdam(model.parameters(), lr=args.lr, clipping_norm=1.0) - else: - optim = Adam(model.parameters(), lr=args.lr) - - # configure dataset - if args.dataset == "yizhongw/self_instruct": - train_data = load_dataset(args.dataset, "super_natural_instructions", split="train") - eval_data = load_dataset(args.dataset, "super_natural_instructions", split="test") - - if args.max_datasets_size is not None: - train_data = train_data.select(range(min(args.max_datasets_size, len(train_data)))) - eval_data = eval_data.select(range(min(args.max_datasets_size, len(eval_data)))) - - train_dataset = SFTDataset(train_data, tokenizer, args.max_len) - eval_dataset = SFTDataset(eval_data, tokenizer, args.max_len) - - else: - train_dataset = SupervisedDataset( - tokenizer=tokenizer, - data_path=args.dataset, - max_datasets_size=args.max_datasets_size, - max_length=args.max_len, - ) - eval_dataset = None - - if dist.is_initialized() and dist.get_world_size() > 1: - train_sampler = DistributedSampler( - train_dataset, - shuffle=True, - seed=42, - drop_last=True, - rank=dist.get_rank(), - num_replicas=dist.get_world_size(), - ) - if eval_dataset is not None: - eval_sampler = DistributedSampler( - eval_dataset, - shuffle=False, - seed=42, - drop_last=False, - rank=dist.get_rank(), - num_replicas=dist.get_world_size(), - ) - else: - train_sampler = None - eval_sampler = None - - train_dataloader = DataLoader( - train_dataset, - shuffle=(train_sampler is None), - sampler=train_sampler, - batch_size=args.batch_size, - pin_memory=True, - ) - if eval_dataset is not None: - eval_dataloader = DataLoader( - eval_dataset, - shuffle=(eval_sampler is None), - sampler=eval_sampler, - batch_size=args.batch_size, - pin_memory=True, - ) - else: - eval_dataloader = None - - num_update_steps_per_epoch = len(train_dataloader) // args.accumulation_steps - max_steps = math.ceil(args.max_epochs * num_update_steps_per_epoch) - lr_scheduler = get_scheduler( - "cosine", optim, num_warmup_steps=math.ceil(max_steps * 0.03), num_training_steps=max_steps - ) - strategy_dict = strategy.prepare(dict(model=model, optimizer=optim, lr_scheduler=lr_scheduler)) - model = strategy_dict["model"] - optim = strategy_dict["optimizer"] - lr_scheduler = strategy_dict["lr_scheduler"] - trainer = SFTTrainer( - model=model, - strategy=strategy, - optim=optim, - lr_scheduler=lr_scheduler, - max_epochs=args.max_epochs, - accumulation_steps=args.accumulation_steps, - ) - - logger = get_dist_logger() - trainer.fit( - train_dataloader=train_dataloader, - eval_dataloader=eval_dataloader, - logger=logger, - log_dir=args.log_dir, - use_wandb=args.use_wandb, - ) - - if args.lora_rank > 0 and args.merge_lora_weights: - from coati.models.lora import LORA_MANAGER - - # NOTE: set model to eval to merge LoRA weights - LORA_MANAGER.merge_weights = True - model.eval() - # save model checkpoint after fitting on only rank0 - strategy.save_pretrained(model, path=args.save_path, tokenizer=tokenizer) - # save optimizer checkpoint on all ranks - if args.need_optim_ckpt: - strategy.save_optimizer( - trainer.optimizer, "rm_optim_checkpoint_%d.pt" % (torch.cuda.current_device()), only_rank0=False - ) - - -if __name__ == "__main__": - parser = argparse.ArgumentParser() - parser.add_argument( - "--strategy", - choices=["ddp", "colossalai_gemini", "colossalai_zero2", "colossalai_zero2_cpu"], - default="colossalai_zero2", - ) - parser.add_argument("--model", choices=["gpt2", "bloom", "opt", "llama", "chatglm"], default="bloom") - parser.add_argument("--tokenizer", type=str, default=None) - parser.add_argument("--pretrain", type=str, default=None) - parser.add_argument("--dataset", type=str, default=None) - parser.add_argument("--max_datasets_size", type=int, default=None) - parser.add_argument("--save_path", type=str, default="output") - parser.add_argument("--need_optim_ckpt", type=bool, default=False) - parser.add_argument("--max_epochs", type=int, default=3) - parser.add_argument("--batch_size", type=int, default=4) - parser.add_argument("--max_len", type=int, default=512) - parser.add_argument("--lora_rank", type=int, default=0, help="low-rank adaptation matrices rank") - parser.add_argument("--merge_lora_weights", type=bool, default=True) - parser.add_argument("--lr", type=float, default=5e-6) - parser.add_argument("--accumulation_steps", type=int, default=8) - parser.add_argument("--log_dir", default="logs", type=str) - parser.add_argument("--use_wandb", default=False, action="store_true") - parser.add_argument("--grad_checkpoint", default=False, action="store_true") - args = parser.parse_args() - train(args) diff --git a/applications/Chat/examples/train_sft.sh b/applications/Chat/examples/train_sft.sh deleted file mode 100755 index 0fb4da3d3ce8..000000000000 --- a/applications/Chat/examples/train_sft.sh +++ /dev/null @@ -1,28 +0,0 @@ -set_n_least_used_CUDA_VISIBLE_DEVICES() { - local n=${1:-"9999"} - echo "GPU Memory Usage:" - local FIRST_N_GPU_IDS=$(nvidia-smi --query-gpu=memory.used --format=csv | - tail -n +2 | - nl -v 0 | - tee /dev/tty | - sort -g -k 2 | - awk '{print $1}' | - head -n $n) - export CUDA_VISIBLE_DEVICES=$(echo $FIRST_N_GPU_IDS | sed 's/ /,/g') - echo "Now CUDA_VISIBLE_DEVICES is set to:" - echo "CUDA_VISIBLE_DEVICES=$CUDA_VISIBLE_DEVICES" -} - -set_n_least_used_CUDA_VISIBLE_DEVICES 4 - -torchrun --standalone --nproc_per_node=4 train_sft.py \ - --pretrain "/path/to/LLaMa-7B/" \ - --model 'llama' \ - --strategy colossalai_zero2 \ - --save_path /path/to/Coati-7B \ - --dataset /path/to/data.json \ - --batch_size 4 \ - --accumulation_steps 8 \ - --lr 2e-5 \ - --max_datasets_size 512 \ - --max_epochs 1 diff --git a/applications/Chat/inference/benchmark.py b/applications/Chat/inference/benchmark.py deleted file mode 100644 index dbb5490a63dc..000000000000 --- a/applications/Chat/inference/benchmark.py +++ /dev/null @@ -1,141 +0,0 @@ -# Adapted from https://github.com/tloen/alpaca-lora/blob/main/generate.py - -import argparse -from time import time - -import torch -from coati.quant import llama_load_quant, low_resource_init -from transformers import AutoTokenizer, GenerationConfig, LlamaConfig, LlamaForCausalLM - - -def generate_prompt(instruction, input=None): - if input: - return f"""Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request. - -### Instruction: -{instruction} - -### Input: -{input} - -### Response:""" - else: - return f"""Below is an instruction that describes a task. Write a response that appropriately completes the request. - -### Instruction: -{instruction} - -### Response:""" - - -@torch.no_grad() -def evaluate( - model, - tokenizer, - instruction, - input=None, - temperature=0.1, - top_p=0.75, - top_k=40, - num_beams=4, - max_new_tokens=128, - **kwargs, -): - prompt = generate_prompt(instruction, input) - inputs = tokenizer(prompt, return_tensors="pt") - input_ids = inputs["input_ids"].cuda() - generation_config = GenerationConfig( - temperature=temperature, - top_p=top_p, - top_k=top_k, - num_beams=num_beams, - **kwargs, - ) - generation_output = model.generate( - input_ids=input_ids, - generation_config=generation_config, - return_dict_in_generate=True, - output_scores=True, - max_new_tokens=max_new_tokens, - do_sample=True, - ) - s = generation_output.sequences[0] - output = tokenizer.decode(s) - n_new_tokens = s.size(0) - input_ids.size(1) - return output.split("### Response:")[1].strip(), n_new_tokens - - -instructions = [ - "Tell me about alpacas.", - "Tell me about the president of Mexico in 2019.", - "Tell me about the king of France in 2019.", - "List all Canadian provinces in alphabetical order.", - "Write a Python program that prints the first 10 Fibonacci numbers.", - "Write a program that prints the numbers from 1 to 100. But for multiples of three print 'Fizz' instead of the number and for the multiples of five print 'Buzz'. For numbers which are multiples of both three and five print 'FizzBuzz'.", - "Tell me five words that rhyme with 'shock'.", - "Translate the sentence 'I have no mouth but I must scream' into Spanish.", - "Count up from 1 to 500.", - # === - "How to play support in legends of league", - "Write a Python program that calculate Fibonacci numbers.", -] -inst = [instructions[0]] * 4 - -if __name__ == "__main__": - parser = argparse.ArgumentParser() - parser.add_argument( - "pretrained", - help="Path to pretrained model. Can be a local path or a model name from the HuggingFace model hub.", - ) - parser.add_argument( - "--quant", - choices=["8bit", "4bit"], - default=None, - help="Quantization mode. Default: None (no quantization, fp16).", - ) - parser.add_argument( - "--gptq_checkpoint", - default=None, - help="Path to GPTQ checkpoint. This is only useful when quantization mode is 4bit. Default: None.", - ) - parser.add_argument( - "--gptq_group_size", - type=int, - default=128, - help="Group size for GPTQ. This is only useful when quantization mode is 4bit. Default: 128.", - ) - args = parser.parse_args() - - if args.quant == "4bit": - assert args.gptq_checkpoint is not None, "Please specify a GPTQ checkpoint." - - tokenizer = AutoTokenizer.from_pretrained(args.pretrained) - - if args.quant == "4bit": - with low_resource_init(): - config = LlamaConfig.from_pretrained(args.pretrained) - model = LlamaForCausalLM(config) - model = llama_load_quant(model, args.gptq_checkpoint, 4, args.gptq_group_size) - model.cuda() - else: - model = LlamaForCausalLM.from_pretrained( - args.pretrained, - load_in_8bit=(args.quant == "8bit"), - torch_dtype=torch.float16, - device_map="auto", - ) - if args.quant != "8bit": - model.half() # seems to fix bugs for some users. - model.eval() - - total_tokens = 0 - start = time() - for instruction in instructions: - print(f"Instruction: {instruction}") - resp, tokens = evaluate(model, tokenizer, instruction, temperature=0.2, num_beams=1) - total_tokens += tokens - print(f"Response: {resp}") - print("\n----------------------------\n") - duration = time() - start - print(f"Total time: {duration:.3f} s, {total_tokens/duration:.3f} tokens/s") - print(f"Peak CUDA mem: {torch.cuda.max_memory_allocated()/1024**3:.3f} GB") diff --git a/applications/Chat/inference/tests/test_chat_prompt.py b/applications/Chat/inference/tests/test_chat_prompt.py deleted file mode 100644 index 9835e71894c6..000000000000 --- a/applications/Chat/inference/tests/test_chat_prompt.py +++ /dev/null @@ -1,61 +0,0 @@ -import os - -from transformers import AutoTokenizer -from utils import ChatPromptProcessor, Dialogue - -CONTEXT = "Below is an instruction that describes a task. Write a response that appropriately completes the request. Do not generate new instructions." -tokenizer = AutoTokenizer.from_pretrained(os.environ["PRETRAINED_PATH"]) - -samples = [ - ( - [ - Dialogue( - instruction="Who is the best player in the history of NBA?", - response="The best player in the history of the NBA is widely considered to be Michael Jordan. He is one of the most successful players in the league, having won 6 NBA championships with the Chicago Bulls and 5 more with the Washington Wizards. He is a 5-time MVP, 1", - ), - Dialogue(instruction="continue this talk", response=""), - ], - 128, - "Below is an instruction that describes a task. Write a response that appropriately completes the request. Do not generate new instructions.\n\n### Instruction:\nWho is the best player in the history of NBA?\n\n### Response:\nThe best player in the history of the NBA is widely considered to be Michael Jordan. He is one of the most successful players in the league, having won 6 NBA championships with the Chicago Bulls and 5 more with the Washington Wizards. He is a 5-time MVP, 1\n\n### Instruction:\ncontinue this talk\n\n### Response:\n", - ), - ( - [ - Dialogue( - instruction="Who is the best player in the history of NBA?", - response="The best player in the history of the NBA is widely considered to be Michael Jordan. He is one of the most successful players in the league, having won 6 NBA championships with the Chicago Bulls and 5 more with the Washington Wizards. He is a 5-time MVP, 1", - ), - Dialogue(instruction="continue this talk", response=""), - ], - 200, - "Below is an instruction that describes a task. Write a response that appropriately completes the request. Do not generate new instructions.\n\n### Instruction:\ncontinue this talk\n\n### Response:\n", - ), - ( - [ - Dialogue( - instruction="Who is the best player in the history of NBA?", - response="The best player in the history of the NBA is widely considered to be Michael Jordan. He is one of the most successful players in the league, having won 6 NBA championships with the Chicago Bulls and 5 more with the Washington Wizards. He is a 5-time MVP, 1", - ), - Dialogue(instruction="continue this talk", response=""), - ], - 211, - "Below is an instruction that describes a task. Write a response that appropriately completes the request. Do not generate new instructions.\n\n### Instruction:\ncontinue this\n\n### Response:\n", - ), - ( - [ - Dialogue(instruction="Who is the best player in the history of NBA?", response=""), - ], - 128, - "Below is an instruction that describes a task. Write a response that appropriately completes the request. Do not generate new instructions.\n\n### Instruction:\nWho is the best player in the history of NBA?\n\n### Response:\n", - ), -] - - -def test_chat_prompt_processor(): - processor = ChatPromptProcessor(tokenizer, CONTEXT, 256) - for history, max_new_tokens, result in samples: - prompt = processor.preprocess_prompt(history, max_new_tokens) - assert prompt == result - - -if __name__ == "__main__": - test_chat_prompt_processor() diff --git a/applications/Chat/inference/utils.py b/applications/Chat/inference/utils.py deleted file mode 100644 index af018adf6e9d..000000000000 --- a/applications/Chat/inference/utils.py +++ /dev/null @@ -1,209 +0,0 @@ -import json -import re -from threading import Lock -from typing import Any, Callable, Generator, List, Optional - -import jieba -import torch -import torch.distributed as dist -import torch.nn as nn -from pydantic import BaseModel, Field - -try: - from transformers.generation_logits_process import ( - LogitsProcessorList, - TemperatureLogitsWarper, - TopKLogitsWarper, - TopPLogitsWarper, - ) -except ImportError: - from transformers.generation import LogitsProcessorList, TemperatureLogitsWarper, TopKLogitsWarper, TopPLogitsWarper - - -def prepare_logits_processor( - top_k: Optional[int] = None, top_p: Optional[float] = None, temperature: Optional[float] = None -) -> LogitsProcessorList: - processor_list = LogitsProcessorList() - if temperature is not None and temperature != 1.0: - processor_list.append(TemperatureLogitsWarper(temperature)) - if top_k is not None and top_k != 0: - processor_list.append(TopKLogitsWarper(top_k)) - if top_p is not None and top_p < 1.0: - processor_list.append(TopPLogitsWarper(top_p)) - return processor_list - - -def _is_sequence_finished(unfinished_sequences: torch.Tensor) -> bool: - if dist.is_initialized() and dist.get_world_size() > 1: - # consider DP - unfinished_sequences = unfinished_sequences.clone() - dist.all_reduce(unfinished_sequences) - return unfinished_sequences.max() == 0 - - -def sample_streamingly( - model: nn.Module, - input_ids: torch.Tensor, - max_generate_tokens: int, - early_stopping: bool = False, - eos_token_id: Optional[int] = None, - pad_token_id: Optional[int] = None, - top_k: Optional[int] = None, - top_p: Optional[float] = None, - temperature: Optional[float] = None, - prepare_inputs_fn: Optional[Callable[[torch.Tensor, Any], dict]] = None, - update_model_kwargs_fn: Optional[Callable[[dict, Any], dict]] = None, - **model_kwargs, -) -> Generator: - logits_processor = prepare_logits_processor(top_k, top_p, temperature) - unfinished_sequences = input_ids.new(input_ids.shape[0]).fill_(1) - - for _ in range(max_generate_tokens): - model_inputs = ( - prepare_inputs_fn(input_ids, **model_kwargs) if prepare_inputs_fn is not None else {"input_ids": input_ids} - ) - outputs = model(**model_inputs) - - next_token_logits = outputs["logits"][:, -1, :] - # pre-process distribution - next_token_logits = logits_processor(input_ids, next_token_logits) - # sample - probs = torch.softmax(next_token_logits, dim=-1, dtype=torch.float) - next_tokens = torch.multinomial(probs, num_samples=1).squeeze(1) - - # finished sentences should have their next token be a padding token - if eos_token_id is not None: - if pad_token_id is None: - raise ValueError("If `eos_token_id` is defined, make sure that `pad_token_id` is defined.") - next_tokens = next_tokens * unfinished_sequences + pad_token_id * (1 - unfinished_sequences) - - yield next_tokens - - # update generated ids, model inputs for next step - input_ids = torch.cat([input_ids, next_tokens[:, None]], dim=-1) - if update_model_kwargs_fn is not None: - model_kwargs = update_model_kwargs_fn(outputs, **model_kwargs) - - # if eos_token was found in one sentence, set sentence to finished - if eos_token_id is not None: - unfinished_sequences = unfinished_sequences.mul((next_tokens != eos_token_id).long()) - - # stop when each sentence is finished if early_stopping=True - if early_stopping and _is_sequence_finished(unfinished_sequences): - break - - -def update_model_kwargs_fn(outputs: dict, **model_kwargs) -> dict: - if "past_key_values" in outputs: - model_kwargs["past"] = outputs["past_key_values"] - else: - model_kwargs["past"] = None - - # update token_type_ids with last value - if "token_type_ids" in model_kwargs: - token_type_ids = model_kwargs["token_type_ids"] - model_kwargs["token_type_ids"] = torch.cat([token_type_ids, token_type_ids[:, -1].unsqueeze(-1)], dim=-1) - - # update attention mask - if "attention_mask" in model_kwargs: - attention_mask = model_kwargs["attention_mask"] - model_kwargs["attention_mask"] = torch.cat( - [attention_mask, attention_mask.new_ones((attention_mask.shape[0], 1))], dim=-1 - ) - - return model_kwargs - - -class Dialogue(BaseModel): - instruction: str = Field(min_length=1, example="Count up from 1 to 500.") - response: str = Field(example="") - - -def _format_dialogue(instruction: str, response: str = ""): - return f"\n\n### Instruction:\n{instruction}\n\n### Response:\n{response}" - - -STOP_PAT = re.compile(r"(###|instruction:).*", flags=(re.I | re.S)) - - -class ChatPromptProcessor: - SAFE_RESPONSE = "The input/response contains inappropriate content, please rephrase your prompt." - - def __init__(self, tokenizer, context: str, max_len: int = 2048, censored_words: List[str] = []): - self.tokenizer = tokenizer - self.context = context - self.max_len = max_len - self.censored_words = set([word.lower() for word in censored_words]) - # These will be initialized after the first call of preprocess_prompt() - self.context_len: Optional[int] = None - self.dialogue_placeholder_len: Optional[int] = None - - def preprocess_prompt(self, history: List[Dialogue], max_new_tokens: int) -> str: - if self.context_len is None: - self.context_len = len(self.tokenizer(self.context)["input_ids"]) - if self.dialogue_placeholder_len is None: - self.dialogue_placeholder_len = len( - self.tokenizer(_format_dialogue(""), add_special_tokens=False)["input_ids"] - ) - prompt = self.context - # the last dialogue must be in the prompt - last_dialogue = history.pop() - # the response of the last dialogue is empty - assert last_dialogue.response == "" - if ( - len(self.tokenizer(_format_dialogue(last_dialogue.instruction), add_special_tokens=False)["input_ids"]) - + max_new_tokens - + self.context_len - >= self.max_len - ): - # to avoid truncate placeholder, apply truncate to the original instruction - instruction_truncated = self.tokenizer( - last_dialogue.instruction, - add_special_tokens=False, - truncation=True, - max_length=(self.max_len - max_new_tokens - self.context_len - self.dialogue_placeholder_len), - )["input_ids"] - instruction_truncated = self.tokenizer.decode(instruction_truncated).lstrip() - prompt += _format_dialogue(instruction_truncated) - return prompt - - res_len = self.max_len - max_new_tokens - len(self.tokenizer(prompt)["input_ids"]) - - rows = [] - for dialogue in history[::-1]: - text = _format_dialogue(dialogue.instruction, dialogue.response) - cur_len = len(self.tokenizer(text, add_special_tokens=False)["input_ids"]) - if res_len - cur_len < 0: - break - res_len -= cur_len - rows.insert(0, text) - prompt += "".join(rows) + _format_dialogue(last_dialogue.instruction) - return prompt - - def postprocess_output(self, output: str) -> str: - output = STOP_PAT.sub("", output) - return output.strip() - - def has_censored_words(self, text: str) -> bool: - if len(self.censored_words) == 0: - return False - intersection = set(jieba.cut(text.lower())) & self.censored_words - return len(intersection) > 0 - - -class LockedIterator: - def __init__(self, it, lock: Lock) -> None: - self.lock = lock - self.it = iter(it) - - def __iter__(self): - return self - - def __next__(self): - with self.lock: - return next(self.it) - - -def load_json(path: str): - with open(path) as f: - return json.load(f) diff --git a/applications/Chat/requirements-test.txt b/applications/Chat/requirements-test.txt deleted file mode 100644 index 93d48bcb6f79..000000000000 --- a/applications/Chat/requirements-test.txt +++ /dev/null @@ -1,2 +0,0 @@ -pytest -colossalai==0.3.3 diff --git a/applications/Chat/requirements.txt b/applications/Chat/requirements.txt deleted file mode 100644 index e56aaca0e7cb..000000000000 --- a/applications/Chat/requirements.txt +++ /dev/null @@ -1,14 +0,0 @@ -transformers>=4.20.1 -tqdm -datasets -loralib -colossalai==0.3.3 -torch<2.0.0, >=1.12.1 -langchain -tokenizers -fastapi -sse_starlette -wandb -sentencepiece -gpustat -tensorboard diff --git a/applications/Chat/tests/test_benchmarks.sh b/applications/Chat/tests/test_benchmarks.sh deleted file mode 100755 index 3fdb25181342..000000000000 --- a/applications/Chat/tests/test_benchmarks.sh +++ /dev/null @@ -1,33 +0,0 @@ -#!/bin/bash - -set -xue - -echo "Hint: You can run this script with 'verbose' as the first argument to run all strategies." - -if [[ $# -ne 0 && "$1" == "verbose" ]]; then - STRATEGIES=( - 'ddp' - 'colossalai_gemini' - 'colossalai_gemini_cpu' - 'colossalai_zero2' - 'colossalai_zero2_cpu' - 'colossalai_zero1' - 'colossalai_zero1_cpu' - ) -else - STRATEGIES=( - 'colossalai_zero2' - ) -fi - -BASE_DIR=$(dirname $(dirname $(realpath $BASH_SOURCE))) -BENCHMARKS_DIR=$BASE_DIR/benchmarks - -echo "[Test]: testing benchmarks ..." - -for strategy in ${STRATEGIES[@]}; do - torchrun --standalone --nproc_per_node 1 $BENCHMARKS_DIR/benchmark_opt_lora_dummy.py \ - --model 125m --critic_model 125m --strategy ${strategy} --lora_rank 4 \ - --num_episodes 2 --num_collect_steps 4 --num_update_steps 2 \ - --train_batch_size 2 --experience_batch_size 4 -done diff --git a/applications/Chat/tests/test_checkpoint.py b/applications/Chat/tests/test_checkpoint.py deleted file mode 100644 index 9c08aa36c9b4..000000000000 --- a/applications/Chat/tests/test_checkpoint.py +++ /dev/null @@ -1,91 +0,0 @@ -import os -import tempfile -from contextlib import nullcontext - -import pytest -import torch -import torch.distributed as dist -from coati.models.gpt import GPTActor -from coati.models.utils import calc_action_log_probs -from coati.trainer.strategies import DDPStrategy, GeminiStrategy, LowLevelZeroStrategy, Strategy -from transformers.models.gpt2.configuration_gpt2 import GPT2Config - -from colossalai.nn.optimizer import HybridAdam -from colossalai.testing import rerun_if_address_is_in_use, spawn - -GPT_CONFIG = GPT2Config(n_embd=128, n_layer=4, n_head=4) - - -def get_data(batch_size: int, seq_len: int = 10) -> dict: - input_ids = torch.randint(0, 50257, (batch_size, seq_len), device="cuda") - attention_mask = torch.ones_like(input_ids) - return dict(input_ids=input_ids, attention_mask=attention_mask) - - -def train_step(strategy: Strategy, actor: GPTActor, actor_optim: HybridAdam, batch_size: int = 8): - data = get_data(batch_size) - action_mask = torch.ones_like(data["attention_mask"], dtype=torch.bool) - actor_logits = actor(data["input_ids"], data["attention_mask"])["logits"] - action_log_probs = calc_action_log_probs(actor_logits, data["input_ids"], action_mask.size(1)) - loss = action_log_probs.sum() - strategy.backward(loss, actor, actor_optim) - strategy.optimizer_step(actor_optim) - - -def run_test_checkpoint(strategy_name: str, shard: bool): - if strategy_name == "ddp": - strategy = DDPStrategy() - elif strategy_name == "colossalai_gemini": - strategy = GeminiStrategy(placement_policy="auto", initial_scale=2**5) - elif strategy_name == "colossalai_zero2": - strategy = LowLevelZeroStrategy(stage=2, placement_policy="cuda") - else: - raise ValueError(f"Unsupported strategy '{strategy_name}'") - - with strategy.model_init_context(): - actor = GPTActor(config=GPT_CONFIG).cuda() - actor_optim = HybridAdam(actor.parameters()) - actor, actor_optim = strategy.prepare((actor, actor_optim)) - - train_step(strategy, actor, actor_optim) - - ctx = tempfile.TemporaryDirectory() if dist.get_rank() == 0 else nullcontext() - - with ctx as dirname: - rank0_dirname = [dirname] - dist.broadcast_object_list(rank0_dirname) - rank0_dirname = rank0_dirname[0] - - model_path = os.path.join(rank0_dirname, "model" if shard else f"model.pt") - strategy.save_model(actor, model_path) - optim_path = os.path.join(rank0_dirname, "optim" if shard else "optim.pt") - strategy.save_optimizer(actor_optim, optim_path) - dist.barrier() - - strategy.load_model(actor, model_path, strict=False) - strategy.load_optimizer(actor_optim, optim_path) - dist.barrier() - - train_step(strategy, actor, actor_optim) - - -def run_dist(rank: int, world_size: int, port: int, strategy_name: str, shard: bool): - os.environ["RANK"] = str(rank) - os.environ["LOCAL_RANK"] = str(rank) - os.environ["WORLD_SIZE"] = str(world_size) - os.environ["MASTER_ADDR"] = "localhost" - os.environ["MASTER_PORT"] = str(port) - run_test_checkpoint(strategy_name, shard) - - -@pytest.mark.dist -@pytest.mark.parametrize("world_size", [4]) -@pytest.mark.parametrize("strategy_name", ["ddp", "colossalai_gemini", "colossalai_zero2"]) -@pytest.mark.parametrize("shard", [False, True]) -@rerun_if_address_is_in_use() -def test_checkpoint(world_size: int, strategy_name: str, shard: bool): - spawn(run_dist, world_size, strategy_name=strategy_name, shard=shard) - - -if __name__ == "__main__": - test_checkpoint(2, "colossalai_gemini", shard=False) diff --git a/applications/Chat/tests/test_dataset.py b/applications/Chat/tests/test_dataset.py deleted file mode 100644 index ec61bbb13fd7..000000000000 --- a/applications/Chat/tests/test_dataset.py +++ /dev/null @@ -1,241 +0,0 @@ -import json -import os -import tempfile -from typing import Optional - -import pytest -import torch -from coati.dataset.prompt_dataset import PromptDataset -from coati.dataset.reward_dataset import HhRlhfDataset, RmStaticDataset -from coati.dataset.sft_dataset import IGNORE_INDEX, SFTDataset, SupervisedDataset -from coati.models.chatglm.chatglm_tokenizer import ChatGLMTokenizer -from datasets import load_dataset -from transformers import AutoTokenizer, BloomTokenizerFast, LlamaTokenizer, PreTrainedTokenizer -from transformers.models.gpt2.tokenization_gpt2 import GPT2Tokenizer - -SFT_DATASET = [ - { - "instruction": "Provide a list of the top 10 most popular mobile games in Asia", - "input": "", - "output": "The top 10 most popular mobile games in Asia are:\n1) PUBG Mobile\n2) Pokemon Go\n3) Candy Crush Saga\n4) Free Fire\n5) Clash of Clans\n6) Mario Kart Tour\n7) Arena of Valor\n8) Fantasy Westward Journey\n9) Subway Surfers\n10) ARK Survival Evolved", - "id": 0, - }, - { - "instruction": "Please provide an action plan for reducing carbon footprint on a corporate level", - "input": "", - "output": "An action plan for reducing carbon footprint on a corporate level could include switching to renewable energy sources (e.g., solar or wind powered), investing in energy efficient infrastructure, implementing green production practices (e.g., reduce water and paper usage, dispose of materials responsibly, recycle/reuse materials), reducing or eliminating travel, encouraging employees to be conscious of and reduce their own energy usage, and incentivizing employees to participate in sustainable work practices.", - "id": 1, - }, - { - "instruction": "Write a persuasive email to your boss explaining why you should have a pay raise", - "input": "", - "output": "Dear [Boss], \n\nI hope this message finds you well. I am writing to request a pay raise.\n\nAs you know, I have been a dedicated and hardworking member of the team since I started working here [insert number] of months/years ago. My enthusiasm and passion for my job has remained consistent over the years, and I have always given 100% to my role. \n\nI understand that the current financial situation is challenging, however, I would sincerely appreciate you taking the time to consider my request. I believe that my dedication to the job and the value that I bring to the organization warrants a raise. I work diligently and am confident that I can continue to be an asset to the company. \n\nI hope my request is taken into account and I thank you in advance for your understanding. I look forward to our conversation. \n\nSincerely,\n[Your Name]", - "id": 2, - }, -] - -PROMPT_DATASET = [ - { - "instruction": 'Edit this paragraph to make it more concise: "Yesterday, I went to the store and bought some things. Then, I came home and put them away. After that, I went for a walk and met some friends."', - "id": 0, - }, - {"instruction": "Write a descriptive paragraph about a memorable vacation you went on", "id": 1}, - {"instruction": "Write a persuasive essay arguing why homework should be banned in schools", "id": 2}, - {"instruction": "Create a chart comparing the statistics on student debt in the United States.", "id": 3}, -] - - -def make_tokenizer(model: str): - if model == "gpt2": - tokenizer = GPT2Tokenizer.from_pretrained("gpt2") - tokenizer.pad_token = tokenizer.eos_token - elif model == "bloom": - tokenizer = BloomTokenizerFast.from_pretrained("bigscience/bloom-560m") - tokenizer.pad_token = tokenizer.eos_token - elif model == "opt": - tokenizer = AutoTokenizer.from_pretrained("facebook/opt-350m") - tokenizer.pad_token = tokenizer.eos_token - elif model == "llama": - tokenizer = LlamaTokenizer.from_pretrained("hf-internal-testing/llama-tokenizer") - tokenizer.pad_token = tokenizer.unk_token - elif model == "chatglm": - tokenizer = ChatGLMTokenizer.from_pretrained("THUDM/chatglm-6b", trust_remote_code=True) - else: - raise ValueError(f"Unsupported model '{model}'") - return tokenizer - - -def check_content(input_ids_stripped: torch.Tensor, tokenizer: PreTrainedTokenizer, model: str): - if model == "opt": - # NOTE: Contrary to GPT2, OPT adds the EOS token to the beginning of every prompt. - assert input_ids_stripped[0] == tokenizer.eos_token_id - input_ids_stripped = input_ids_stripped[1:] - elif model == "llama": - assert input_ids_stripped[0] == tokenizer.bos_token_id - input_ids_stripped = input_ids_stripped[1:] - elif model == "chatglm": - assert input_ids_stripped[0] == tokenizer.bos_token_id - assert input_ids_stripped[-1] == tokenizer.eos_token_id - input_ids_stripped = input_ids_stripped[1:-1] - assert torch.all(input_ids_stripped != tokenizer.pad_token_id) - assert torch.all(input_ids_stripped != tokenizer.bos_token_id) - assert torch.all(input_ids_stripped != tokenizer.eos_token_id) - assert input_ids_stripped != tokenizer.sep_token_id - assert input_ids_stripped != tokenizer.cls_token_id - if model == "chatglm": - assert torch.all(input_ids_stripped != tokenizer.mask_token_id) - else: - assert input_ids_stripped != tokenizer.mask_token_id - - -@pytest.mark.parametrize("model", ["gpt2", "bloom", "opt", "llama"]) -@pytest.mark.parametrize("max_length", [32, 1024]) -@pytest.mark.parametrize("max_datasets_size", [2]) -def test_prompt_dataset(model: str, max_datasets_size: int, max_length: int): - with tempfile.TemporaryDirectory() as tmp_dir: - dataset_name = "prompt_dataset.json" - with open(os.path.join(tmp_dir, dataset_name), "w") as f: - json.dump(PROMPT_DATASET, f) - tokenizer = make_tokenizer(model) - assert tokenizer.padding_side in ("left", "right") - prompt_dataset = PromptDataset( - data_path=os.path.join(tmp_dir, dataset_name), - tokenizer=tokenizer, - max_datasets_size=max_datasets_size, - max_length=max_length, - ) - assert len(prompt_dataset) == min(max_datasets_size, len(PROMPT_DATASET)) - for i in range(len(prompt_dataset)): - assert isinstance(prompt_dataset[i], dict) - assert list(prompt_dataset[i].keys()) == ["input_ids", "attention_mask"] - input_ids = prompt_dataset[i]["input_ids"] - attention_mask = prompt_dataset[i]["attention_mask"] - attention_mask = attention_mask.bool() - assert input_ids.shape == attention_mask.shape == torch.Size([max_length]) - assert torch.all(input_ids[torch.logical_not(attention_mask)] == tokenizer.pad_token_id) - check_content(input_ids.masked_select(attention_mask), tokenizer, model) - - -@pytest.mark.parametrize("model", ["gpt2", "bloom", "opt", "llama"]) -@pytest.mark.parametrize( - ["dataset_path", "subset"], [("Anthropic/hh-rlhf", "harmless-base"), ("Dahoas/rm-static", None)] -) -@pytest.mark.parametrize("max_datasets_size", [32]) -@pytest.mark.parametrize("max_length", [32, 1024]) -def test_reward_dataset(model: str, dataset_path: str, subset: Optional[str], max_datasets_size: int, max_length: int): - data = load_dataset(dataset_path, data_dir=subset) - assert max_datasets_size <= len(data["train"]) and max_datasets_size <= len(data["test"]) - train_data = data["train"].select(range(max_datasets_size)) - test_data = data["test"].select(range(max_datasets_size)) - tokenizer = make_tokenizer(model) - assert tokenizer.padding_side in ("left", "right") - - if dataset_path == "Anthropic/hh-rlhf": - train_dataset = HhRlhfDataset(train_data, tokenizer, max_length) - test_dataset = HhRlhfDataset(test_data, tokenizer, max_length) - elif dataset_path == "Dahoas/rm-static": - train_dataset = RmStaticDataset(train_data, tokenizer, max_length) - test_dataset = RmStaticDataset(test_data, tokenizer, max_length) - else: - raise ValueError(f'Unsupported dataset "{dataset_path}"') - - assert len(train_dataset) == len(test_dataset) == max_datasets_size - for i in range(max_datasets_size): - chosen_ids, c_mask, reject_ids, r_mask = train_dataset[i] - assert chosen_ids.shape == c_mask.shape == reject_ids.shape == r_mask.shape == torch.Size([max_length]) - c_mask = c_mask.to(torch.bool) - r_mask = r_mask.to(torch.bool) - if chosen_ids.masked_select(c_mask)[-1] == tokenizer.eos_token_id: - check_content(chosen_ids.masked_select(c_mask)[:-1], tokenizer, model) - assert torch.all(chosen_ids.masked_select(torch.logical_not(c_mask)) == tokenizer.pad_token_id) - else: - check_content(chosen_ids.masked_select(c_mask), tokenizer, model) - assert torch.all(c_mask) - if reject_ids.masked_select(r_mask)[-1] == tokenizer.eos_token_id: - check_content(reject_ids.masked_select(r_mask)[:-1], tokenizer, model) - assert torch.all(reject_ids.masked_select(torch.logical_not(r_mask)) == tokenizer.pad_token_id) - else: - check_content(reject_ids.masked_select(r_mask), tokenizer, model) - assert torch.all(r_mask) - - chosen_ids, c_mask, reject_ids, r_mask = test_dataset[i] - assert chosen_ids.shape == c_mask.shape == reject_ids.shape == r_mask.shape == torch.Size([max_length]) - c_mask = c_mask.to(torch.bool) - r_mask = r_mask.to(torch.bool) - if chosen_ids.masked_select(c_mask)[-1] == tokenizer.eos_token_id: - check_content(chosen_ids.masked_select(c_mask)[:-1], tokenizer, model) - assert torch.all(chosen_ids.masked_select(torch.logical_not(c_mask)) == tokenizer.pad_token_id) - else: - check_content(chosen_ids.masked_select(c_mask), tokenizer, model) - assert torch.all(c_mask) - if reject_ids.masked_select(r_mask)[-1] == tokenizer.eos_token_id: - check_content(reject_ids.masked_select(r_mask)[:-1], tokenizer, model) - assert torch.all(reject_ids.masked_select(torch.logical_not(r_mask)) == tokenizer.pad_token_id) - else: - check_content(reject_ids.masked_select(r_mask), tokenizer, model) - assert torch.all(r_mask) - - -@pytest.mark.parametrize("model", ["gpt2", "bloom", "opt", "llama", "chatglm"]) -@pytest.mark.parametrize("dataset_path", ["yizhongw/self_instruct", None]) -@pytest.mark.parametrize("max_dataset_size", [2]) -@pytest.mark.parametrize("max_length", [32, 1024]) -def test_sft_dataset(model: str, dataset_path: Optional[str], max_dataset_size: int, max_length: int): - tokenizer = make_tokenizer(model) - if dataset_path == "yizhongw/self_instruct": - data = load_dataset(dataset_path, "super_natural_instructions") - train_data = data["train"].select(range(max_dataset_size)) - sft_dataset = SFTDataset(train_data, tokenizer, max_length) - else: - with tempfile.TemporaryDirectory() as tmp_dir: - dataset_name = "sft_dataset.json" - with open(os.path.join(tmp_dir, dataset_name), "w") as f: - json.dump(SFT_DATASET, f) - sft_dataset = SupervisedDataset( - tokenizer=tokenizer, - data_path=os.path.join(tmp_dir, dataset_name), - max_datasets_size=max_dataset_size, - max_length=max_length, - ) - assert len(sft_dataset) == min(max_dataset_size, len(SFT_DATASET)) - - if isinstance(tokenizer, ChatGLMTokenizer): - for i in range(max_dataset_size): - assert isinstance(sft_dataset[i], dict) - assert list(sft_dataset[i].keys()) == ["input_ids", "labels"] - input_ids = sft_dataset[i]["input_ids"] - labels = sft_dataset[i]["labels"] - assert input_ids.shape == labels.shape == torch.Size([max_length]) - - ignore_mask = labels == IGNORE_INDEX - assert input_ids.masked_select(torch.logical_not(ignore_mask))[0] == tokenizer.bos_token_id - check_content(input_ids.masked_select(torch.logical_not(ignore_mask)), tokenizer, model) - return - - for i in range(max_dataset_size): - assert isinstance(sft_dataset[i], dict) - assert list(sft_dataset[i].keys()) == ["input_ids", "labels", "attention_mask"] - input_ids = sft_dataset[i]["input_ids"] - labels = sft_dataset[i]["labels"] - attention_mask = sft_dataset[i]["attention_mask"].to(torch.bool) - assert input_ids.shape == labels.shape == attention_mask.shape == torch.Size([max_length]) - if input_ids.masked_select(attention_mask)[-1] == tokenizer.eos_token_id: - check_content(input_ids.masked_select(attention_mask)[:-1], tokenizer, model) - assert torch.all(input_ids.masked_select(torch.logical_not(attention_mask)) == tokenizer.pad_token_id) - else: - check_content(input_ids.masked_select(attention_mask), tokenizer, model) - assert torch.all(attention_mask) - ignore_mask = labels == IGNORE_INDEX - prompt_mask = torch.logical_and(ignore_mask, attention_mask) - check_content(input_ids.masked_select(prompt_mask), tokenizer, model) - assert torch.all(input_ids.masked_select(ignore_mask ^ prompt_mask) == tokenizer.pad_token_id) - - -if __name__ == "__main__": - test_sft_dataset(model="bloom", dataset_path="yizhongw/self_instruct", max_dataset_size=2, max_length=256) - - test_reward_dataset( - model="gpt2", dataset_path="Anthropic/hh-rlhf", subset="harmless-base", max_datasets_size=8, max_length=256 - ) - - test_prompt_dataset(model="opt", max_datasets_size=2, max_length=128) diff --git a/applications/Chat/tests/test_experience.py b/applications/Chat/tests/test_experience.py deleted file mode 100644 index a9591259800d..000000000000 --- a/applications/Chat/tests/test_experience.py +++ /dev/null @@ -1,130 +0,0 @@ -import copy -import os - -import pytest -import torch -import torch.distributed as dist -from coati.experience_buffer import NaiveExperienceBuffer -from coati.experience_maker import NaiveExperienceMaker -from coati.models.base import RewardModel -from coati.models.gpt import GPTActor, GPTCritic -from coati.trainer.ppo import _set_default_generate_kwargs -from coati.trainer.strategies import DDPStrategy, GeminiStrategy -from coati.trainer.strategies.colossalai import LowLevelZeroStrategy -from transformers.models.gpt2.configuration_gpt2 import GPT2Config - -from colossalai.testing import rerun_if_address_is_in_use, spawn - -GPT_CONFIG = GPT2Config(n_embd=128, n_layer=4, n_head=4) - - -def get_data(batch_size: int, seq_len: int = 10) -> dict: - input_ids = torch.randint(0, 50257, (batch_size, seq_len), device="cuda") - attention_mask = torch.ones_like(input_ids) - return dict(input_ids=input_ids, attention_mask=attention_mask) - - -def gather_and_equal(tensor: torch.Tensor) -> bool: - world_size = dist.get_world_size() - outputs = [torch.empty_like(tensor) for _ in range(world_size)] - dist.all_gather(outputs, tensor.contiguous()) - for t in outputs[1:]: - if not torch.equal(outputs[0], t): - return False - return True - - -def make_and_consume_experience(strategy): - EXPERIENCE_BATCH_SIZE = 4 - SAMPLE_BATCH_SIZE = 2 - - if strategy == "ddp": - strategy = DDPStrategy() - elif strategy == "colossalai-zero2": - strategy = LowLevelZeroStrategy() - elif strategy == "colossalai-gemini": - strategy = GeminiStrategy(placement_policy="static") - else: - raise ValueError(f'Unsupported strategy "{strategy}"') - - with strategy.model_init_context(): - actor = GPTActor(config=GPT_CONFIG).cuda() - critic = GPTCritic(config=GPT_CONFIG).cuda() - - initial_model = GPTActor(config=GPT_CONFIG).cuda() - reward_model = RewardModel(model=copy.deepcopy(critic.model)).cuda() - - actor, critic, initial_model, reward_model = strategy.prepare(actor, critic, initial_model, reward_model) - - class MockTokenizer: - def __init__(self): - self.padding_side = "left" - self.eos_token_id = 0 - self.pad_token_id = 0 - - tokenizer = MockTokenizer() - experience_maker = NaiveExperienceMaker(actor, critic, reward_model, initial_model, tokenizer) - data_buffer = NaiveExperienceBuffer(SAMPLE_BATCH_SIZE, cpu_offload=False) - - generate_kwargs = dict(do_sample=True, max_length=16) - generate_kwargs = _set_default_generate_kwargs(strategy, generate_kwargs, actor) - - # experience of all ranks should be the same - for _ in range(2): - data = get_data(EXPERIENCE_BATCH_SIZE) - assert gather_and_equal(data["input_ids"]) - assert gather_and_equal(data["attention_mask"]) - experience = experience_maker.make_experience(**data, do_sample=True, max_length=16) - assert gather_and_equal(experience.sequences) - assert gather_and_equal(experience.action_log_probs) - assert gather_and_equal(experience.values) - assert gather_and_equal(experience.reward) - assert gather_and_equal(experience.advantages) - assert gather_and_equal(experience.action_mask) - assert gather_and_equal(experience.attention_mask) - data_buffer.append(experience) - - # data buffer's data should be the same - buffer_size = torch.tensor([len(data_buffer)], device="cuda") - assert gather_and_equal(buffer_size) - for item in data_buffer.items: - assert gather_and_equal(item.sequences) - assert gather_and_equal(item.action_log_probs) - assert gather_and_equal(item.values) - assert gather_and_equal(item.reward) - assert gather_and_equal(item.advantages) - assert gather_and_equal(item.action_mask) - assert gather_and_equal(item.attention_mask) - - # dataloader of each rank should have the same size and different batch - dataloader = strategy.setup_dataloader(data_buffer) - dataloader_size = torch.tensor([len(dataloader)], device="cuda") - assert gather_and_equal(dataloader_size) - for experience in dataloader: - assert not gather_and_equal(experience.sequences) - assert not gather_and_equal(experience.action_log_probs) - assert not gather_and_equal(experience.values) - assert not gather_and_equal(experience.reward) - assert not gather_and_equal(experience.advantages) - # action mask and attention mask may be same - - -def run_dist(rank, world_size, port, strategy): - os.environ["RANK"] = str(rank) - os.environ["LOCAL_RANK"] = str(rank) - os.environ["WORLD_SIZE"] = str(world_size) - os.environ["MASTER_ADDR"] = "localhost" - os.environ["MASTER_PORT"] = str(port) - make_and_consume_experience(strategy) - - -@pytest.mark.dist -@pytest.mark.parametrize("world_size", [2]) -@pytest.mark.parametrize("strategy", ["ddp", "colossalai-zero2", "colossalai-gemini"]) -@rerun_if_address_is_in_use() -def test_experience(world_size, strategy): - spawn(run_dist, world_size, strategy=strategy) - - -if __name__ == "__main__": - test_experience(2, "colossalai-zero2") diff --git a/applications/Chat/tests/test_inference.sh b/applications/Chat/tests/test_inference.sh deleted file mode 100755 index 849db06e58ab..000000000000 --- a/applications/Chat/tests/test_inference.sh +++ /dev/null @@ -1,11 +0,0 @@ -set -xue - -BASE_DIR=$(dirname $(dirname $(realpath $BASH_SOURCE))) -EXAMPLES_DIR=$BASE_DIR/examples - -echo "[Test]: testing inference ..." - -# HACK: skip llama due to oom -for model in 'gpt2' 'bloom' 'opt'; do - python $EXAMPLES_DIR/inference.py --model $model -done diff --git a/applications/Chat/tests/test_models.py b/applications/Chat/tests/test_models.py deleted file mode 100644 index b2c22ac6a3b9..000000000000 --- a/applications/Chat/tests/test_models.py +++ /dev/null @@ -1,245 +0,0 @@ -import copy -from typing import Any, Callable, Dict, Tuple - -import pytest -import torch -import torch.nn as nn -from coati.models.base import Actor, Critic, RewardModel, get_base_model -from coati.models.bloom import BLOOMRM, BLOOMActor, BLOOMCritic -from coati.models.chatglm import ChatGLMActor -from coati.models.chatglm.chatglm_tokenizer import ChatGLMTokenizer -from coati.models.generation import generate -from coati.models.gpt import GPTRM, GPTActor, GPTCritic -from coati.models.llama import LlamaActor -from coati.models.lora import LoraLinear, convert_to_lora_module -from coati.models.loss import GPTLMLoss, LogExpLoss, LogSigLoss, PolicyLoss, ValueLoss -from coati.models.opt import OPTRM, OPTActor, OPTCritic -from coati.models.utils import calc_action_log_probs, masked_mean - - -@pytest.mark.parametrize("batch_size", [4]) -@pytest.mark.parametrize("seq_len", [32]) -@pytest.mark.parametrize( - "actor_maker", - [ - lambda: BLOOMActor(), - lambda: GPTActor(), - # HACK: skip llama due to long execution time - # lambda: LlamaActor(), - lambda: OPTActor(), - ], -) -@pytest.mark.parametrize( - "generate_kwargs", - [ - { - "max_length": 64, - "use_cache": True, - "do_sample": True, - "temperature": 1.0, - "top_k": 50, - } - ], -) -def test_generation(actor_maker: Callable[[], Actor], batch_size: int, seq_len: int, generate_kwargs: Dict[str, Any]): - class MockTokenizer: - def __init__(self): - self.padding_side = "left" - self.eos_token_id = 0 - self.pad_token_id = 0 - - actor = actor_maker() - input_ids = torch.randint(0, 100, (batch_size, seq_len)).cuda() - tokenizer = MockTokenizer() - sequences = generate(actor.cuda(), input_ids, tokenizer, **generate_kwargs) - assert sequences.shape == (batch_size, generate_kwargs["max_length"]) - - -def test_utils(): - fn_input = {"tensor": torch.ones((10,)), "mask": torch.randint(0, 2, (10,))} - fn_output = masked_mean(dim=0, **fn_input) - assert fn_output.dim() == 0 - assert torch.allclose(fn_output, torch.tensor(1.0)) - - batch_size = 4 - seq_len = 32 - num_labels = 10 - num_actions = 2 - fn_input = { - "logits": torch.randn((batch_size, seq_len, num_labels)), - "sequences": torch.randint(0, num_labels, (batch_size, seq_len)), - "num_actions": num_actions, - } - fn_output = calc_action_log_probs(**fn_input) - assert fn_output.shape == (batch_size, num_actions) - - -@pytest.mark.parametrize("lora_rank", [4]) -@pytest.mark.parametrize("num_dim", [32]) -@pytest.mark.parametrize("num_layers", [4]) -def test_lora(lora_rank: int, num_dim: int, num_layers: int): - model = nn.ModuleList([nn.Linear(num_dim, num_dim) for _ in range(num_layers)]) - lora_model = convert_to_lora_module(model, lora_rank) - assert isinstance(lora_model, nn.ModuleList) - for i in range(num_layers): - assert isinstance(lora_model[i], LoraLinear) - assert lora_model[i].lora_A.shape == (lora_rank, num_dim) - assert lora_model[i].lora_B.shape == (num_dim, lora_rank) - - old_model = copy.deepcopy(lora_model) - for i in range(num_layers): - assert isinstance(lora_model[i], LoraLinear) - assert torch.allclose(old_model[i].weight, lora_model[i].weight) - assert torch.allclose(old_model[i].bias, lora_model[i].bias) - assert torch.allclose(old_model[i].lora_B @ old_model[i].lora_A, lora_model[i].lora_B @ lora_model[i].lora_A) - optimizer = torch.optim.Adam(lora_model.parameters()) - x = torch.randn(8, num_dim) - for i in range(num_layers): - x = lora_model[i](x) - loss = x.sum() - loss.backward() - optimizer.step() - for i in range(num_layers): - assert isinstance(lora_model[i], LoraLinear) - assert torch.allclose(old_model[i].weight, lora_model[i].weight) - assert torch.allclose(old_model[i].bias, lora_model[i].bias) - assert not torch.allclose( - old_model[i].lora_B @ old_model[i].lora_A, lora_model[i].lora_B @ lora_model[i].lora_A - ) - - -@pytest.mark.parametrize("batch_size", [8]) -@pytest.mark.parametrize("seq_len", [128]) -@pytest.mark.parametrize( - "models_maker", - [ - lambda: (BLOOMActor(), BLOOMCritic(), BLOOMRM()), - lambda: (GPTActor(), GPTCritic(), GPTRM()), - # HACK: skip llama due to long execution time - # lambda: (LlamaActor(), LlamaCritic(), LlamaRM()), - lambda: (OPTActor(), OPTCritic(), OPTRM()), - lambda: (ChatGLMActor(), None, None), - ], -) -@torch.no_grad() -def test_models(models_maker: Callable[[], Tuple[Actor, Critic, RewardModel]], batch_size: int, seq_len: int): - actor_input = { - "input_ids": torch.randint(0, 100, (batch_size, seq_len)), - "attention_mask": torch.randint(0, 2, (batch_size, seq_len)), - } - critic_input = { - "sequences": torch.randint(0, 100, (batch_size, seq_len)), - "attention_mask": torch.randint(0, 2, (batch_size, seq_len)), - } - rm_input = { - "sequences": torch.randint(0, 100, (batch_size, seq_len)), - "attention_mask": torch.randint(0, 2, (batch_size, seq_len)), - } - - actor, critic, rm = models_maker() - if isinstance(actor, ChatGLMActor): - actor = actor.float() - tokenizer = ChatGLMTokenizer.from_pretrained("THUDM/chatglm-6b", trust_remote_code=True) - chatglm_special_token = torch.tensor([tokenizer.gmask_token_id, tokenizer.bos_token_id]).repeat(batch_size, 1) - actor_input = { - "input_ids": torch.cat( - ( - torch.randint(0, 100, (batch_size, seq_len // 2)), - chatglm_special_token, - torch.randint(0, 100, (batch_size, seq_len // 2 - 2)), - ), - dim=1, - ), - "attention_mask": torch.randint(0, 2, (batch_size, 1, seq_len, seq_len)), - } - assert isinstance(actor, Actor) - get_base_model(actor) - actor_output = actor(**actor_input) - assert actor_output.logits.shape[:2] == (batch_size, seq_len) - - if critic: - assert isinstance(critic, Critic) - get_base_model(critic) - critic_output = critic(**critic_input) - assert critic_output.shape == (batch_size,) - - if rm: - assert isinstance(rm, RewardModel) - get_base_model(rm) - rm_output = rm(**rm_input) - assert rm_output.shape == (batch_size,) - - -@pytest.mark.parametrize("batch_size", [16]) -@pytest.mark.parametrize("seq_len", [128]) -@pytest.mark.parametrize("num_labels", [100]) -def test_loss(batch_size: int, seq_len: int, num_labels: int): - loss = GPTLMLoss() - loss_input = { - "logits": torch.randn(batch_size, seq_len, num_labels), - "labels": torch.randint(0, num_labels, (batch_size, seq_len)), - } - loss(**loss_input) - - loss = PolicyLoss() - loss_input = { - "log_probs": torch.randn( - batch_size, - ), - "old_log_probs": torch.randn( - batch_size, - ), - "advantages": torch.randn( - batch_size, - ), - } - loss(**loss_input) - - loss = ValueLoss() - loss_input = { - "values": torch.randn( - batch_size, - ), - "old_values": torch.randn( - batch_size, - ), - "reward": torch.randn( - batch_size, - ), - } - loss(**loss_input) - - loss = LogSigLoss() - loss_input = { - "chosen_reward": torch.randn( - batch_size, - ), - "reject_reward": torch.randn( - batch_size, - ), - } - loss(**loss_input) - - loss = LogExpLoss() - loss_input = { - "chosen_reward": torch.randn( - batch_size, - ), - "reject_reward": torch.randn( - batch_size, - ), - } - loss(**loss_input) - - -if __name__ == "__main__": - generate_kwargs = dict(max_length=40, use_cache=True, do_sample=True, temperature=1.0, top_k=50) - test_generation(lambda: LlamaActor(), batch_size=4, seq_len=32, generate_kwargs=generate_kwargs) - - test_utils() - - test_lora(lora_rank=2, num_dim=8, num_layers=2) - - test_models(models_maker=lambda: (BLOOMActor(), BLOOMCritic(), BLOOMRM()), batch_size=8, seq_len=128) - - test_loss(batch_size=8, seq_len=128, num_labels=100) diff --git a/applications/Chat/tests/test_train.sh b/applications/Chat/tests/test_train.sh deleted file mode 100755 index 68fca7fbf8c0..000000000000 --- a/applications/Chat/tests/test_train.sh +++ /dev/null @@ -1,233 +0,0 @@ -#!/usr/bin/env bash - -set_n_least_used_CUDA_VISIBLE_DEVICES() { - local n=${1:-"9999"} - echo "GPU Memory Usage:" - local FIRST_N_GPU_IDS=$(nvidia-smi --query-gpu=memory.used --format=csv | - tail -n +2 | - nl -v 0 | - tee /dev/tty | - sort -g -k 2 | - awk '{print $1}' | - head -n $n) - export CUDA_VISIBLE_DEVICES=$(echo $FIRST_N_GPU_IDS | sed 's/ /,/g') - echo "Now CUDA_VISIBLE_DEVICES is set to:" - echo "CUDA_VISIBLE_DEVICES=$CUDA_VISIBLE_DEVICES" -} - -set_n_least_used_CUDA_VISIBLE_DEVICES 4 - -set -xu - -if [ -z "$SFT_DATASET" ]; then - echo "Please set \$SFT_DATASET to the path to sft dataset." - exit 1 -fi - -if [ -z "$PROMPT_DATASET" ]; then - echo "Please set \$PROMPT_DATASET to the path to prompts csv." - exit 1 -fi - -if [ -z "$PRETRAIN_DATASET" ]; then - echo "Please set \$PRETRAIN_DATASET to the path to alpaca data." - exit 1 -fi - -NUM_RETRY=3 -BASE_DIR=$(dirname $(dirname $(realpath $BASH_SOURCE))) -EXAMPLES_DIR=$BASE_DIR/examples -MODELS_DIR=$BASE_DIR/examples/models_config -MODELS=('gpt2' 'bloom' 'opt' 'llama') -STRATEGIES=('ddp' 'colossalai_gemini' 'colossalai_zero2') - - -export OMP_NUM_THREADS=8 - -# install requirements -pip install -r $EXAMPLES_DIR/requirements.txt - -python $EXAMPLES_DIR/download_model.py --model-dir $MODELS_DIR --config-only - -get_pretrain() { - local model=$1 - if [[ $model == "gpt2" ]]; then - echo "gpt2" - elif [[ $model == "bloom" ]]; then - echo "bigscience/bloom-560m" - elif [[ $model == "opt" ]]; then - echo "facebook/opt-350m" - else - echo "Unknown model $model" - exit 1 - fi -} - -random_choice() { - local arr=("$@") - local len=${#arr[@]} - local idx=$((RANDOM % len)) - echo ${arr[$idx]} -} - -echo "[Test]: testing sft ..." - -# FIXME: This is a hack to skip tests that are not working -# - gpt2-ddp: RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation -# - llama-*: These tests can be passed locally, skipped for long execution time -# - *-gemini: Gemini plugin does not support `from_pretrained` yet -SKIPPED_TESTS=( - "gpt2-ddp" - "llama-ddp" - "llama-colossalai_gemini" - "llama-colossalai_zero2" -) - -GRAD_CKPTS=('' '--grad_checkpoint') -for lora_rank in '0'; do - for model in ${MODELS[@]}; do - strategies=($(shuf -e "${STRATEGIES[@]}")) - for strategy in ${strategies[@]}; do - if [[ " ${SKIPPED_TESTS[*]} " =~ " $model-$strategy-$lora_rank " ]]; then - echo "[Test]: Skipped $model-$strategy-$lora_rank" - continue - elif [[ " ${SKIPPED_TESTS[*]} " =~ " $model-$strategy " ]]; then - echo "[Test]: Skipped $model-$strategy" - continue - fi - pretrain=$(get_pretrain $model) - pretrain_model="" - if [[ $lora_rank -gt 0 ]]; then - pretrain_model="--pretrain $pretrain" - fi - grad_ckpt=$(random_choice "${GRAD_CKPTS[@]}") - for i in $(seq $NUM_RETRY); do - echo "[Test]: $model-$strategy-$lora_rank, attempt $i" - torchrun --standalone --nproc_per_node=4 $EXAMPLES_DIR/train_sft.py \ - $pretrain_model --tokenizer $MODELS_DIR/$model \ - --model $model --strategy $strategy --lora_rank $lora_rank $grad_ckpt \ - --dataset $SFT_DATASET --max_datasets_size 8 \ - --max_epochs 1 --batch_size 1 --accumulation_steps 1 --lr 1e-8 \ - --save_path $EXAMPLES_DIR/rlhf_models/sft_ckpt_${model}_${lora_rank} - passed=$? - if [ $passed -eq 0 ]; then - break - fi - done - if [ $passed -ne 0 ]; then - echo "[Test]: Failed $model-$strategy-$lora_rank" - exit 1 - fi - done - done -done - -echo "[Test]: testing reward model ..." - -# FIXME: This is a hack to skip tests that are not working -# - gpt2-ddp: RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation -# - llama-*: These tests can be passed locally, skipped for long execution time -# - *-gemini: Gemini plugin does not support `from_pretrained` yet -SKIPPED_TESTS=( - "gpt2-ddp" - "llama-ddp" - "llama-colossalai_gemini" - "llama-colossalai_zero2" -) - -LOSS_FNS=('log_sig' 'log_exp') -DATASETS=('Anthropic/hh-rlhf' 'Dahoas/rm-static') -for lora_rank in '0'; do - for model in ${MODELS[@]}; do - strategies=($(shuf -e "${STRATEGIES[@]}")) - for strategy in ${strategies[@]}; do - if [[ " ${SKIPPED_TESTS[*]} " =~ " $model-$strategy-$lora_rank " ]]; then - echo "[Test]: Skipped $model-$strategy-$lora_rank" - continue - elif [[ " ${SKIPPED_TESTS[*]} " =~ " $model-$strategy " ]]; then - echo "[Test]: Skipped $model-$strategy" - continue - fi - pretrain=$(get_pretrain $model) - pretrain_model="" - if [[ $lora_rank -gt 0 ]]; then - pretrain_model="--pretrain $pretrain" - fi - loss_fn=$(random_choice "${LOSS_FNS[@]}") - dataset=$(random_choice "${DATASETS[@]}") - subset=$(if [[ $dataset == "Dahoas/rm-static" ]]; then echo "None"; else echo "harmless-base"; fi) - for i in $(seq $NUM_RETRY); do - echo "[Test]: $model-$strategy-$lora_rank, attempt $i" - torchrun --standalone --nproc_per_node=4 $EXAMPLES_DIR/train_reward_model.py \ - $pretrain_model --tokenizer $MODELS_DIR/$model \ - --dataset $dataset --subset $subset --max_datasets_size 8 \ - --model $model --strategy $strategy --lora_rank $lora_rank \ - --loss_fn $loss_fn --batch_size 1 --lr 1e-8 \ - --save_path $EXAMPLES_DIR/rlhf_models/rm_ckpt_${model}_${lora_rank}.pt - passed=$? - if [ $passed -eq 0 ]; then - break - fi - done - if [ $passed -ne 0 ]; then - echo "[Test]: Failed to train reward model $model-$strategy-$lora_rank" - exit 1 - fi - done - done -done - -echo "[Test]: testing RLHF ..." - -# FIXME: This is a hack to skip tests that are not working -# - gpt2-ddp: RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation -# - llama-*: These tests can be passed locally, skipped for long execution time -# - *-gemini: Gemini plugin does not support `from_pretrained` yet -SKIPPED_TESTS=( - "gpt2-ddp" - "llama-ddp" - "llama-colossalai_gemini" - "llama-colossalai_zero2" -) - -for model in ${MODELS[@]}; do - for lora_rank in '0'; do - strategies=($(shuf -e "${STRATEGIES[@]}")) - for strategy in ${strategies[@]}; do - if [[ " ${SKIPPED_TESTS[*]} " =~ " $model-$strategy-$lora_rank " ]]; then - echo "[Test]: Skipped $model-$strategy-$lora_rank" - continue - elif [[ " ${SKIPPED_TESTS[*]} " =~ " $model-$strategy " ]]; then - echo "[Test]: Skipped $model-$strategy" - continue - fi - rm_pretrain=$(get_pretrain $model) - rm_pretrain_model="" - if [[ $lora_rank -gt 0 ]]; then - rm_pretrain_model="--rm_pretrain $rm_pretrain" - fi - for i in $(seq $NUM_RETRY); do - echo "[Test]: $model-$strategy-$lora_rank, attempt $i" - torchrun --standalone --nproc_per_node=4 $EXAMPLES_DIR/train_prompts.py \ - --prompt_dataset $PROMPT_DATASET --pretrain_dataset $PRETRAIN_DATASET --max_datasets_size 32 \ - --strategy $strategy --model $model --tokenizer $MODELS_DIR/$model \ - --num_episodes 1 --num_collect_steps 1 --num_update_steps 1 --lr 1e-8 \ - --experience_batch_size 2 --train_batch_size 1 --lora_rank $lora_rank \ - --pretrain $EXAMPLES_DIR/rlhf_models/sft_ckpt_${model}_${lora_rank} \ - $rm_pretrain_model --rm_path $EXAMPLES_DIR/rlhf_models/rm_ckpt_${model}_${lora_rank}.pt \ - --save_path $EXAMPLES_DIR/rlhf_models/actor_checkpoint_prompts - passed=$? - if [ $passed -eq 0 ]; then - break - fi - done - if [ $passed -ne 0 ]; then - echo "[Test]: Failed to train RLHF $model-$strategy-$lora_rank" - exit 1 - fi - done - rm -rf $EXAMPLES_DIR/rlhf_models/sft_ckpt_${model}_${lora_rank} - rm $EXAMPLES_DIR/rlhf_models/rm_ckpt_${model}_${lora_rank}.pt - done -done -rm -rf $EXAMPLES_DIR/rlhf_models/actor_checkpoint_prompts diff --git a/applications/Chat/.gitignore b/applications/ColossalChat/.gitignore old mode 100644 new mode 100755 similarity index 87% rename from applications/Chat/.gitignore rename to applications/ColossalChat/.gitignore index 5fa068105e26..33950adc0bb5 --- a/applications/Chat/.gitignore +++ b/applications/ColossalChat/.gitignore @@ -143,6 +143,17 @@ docs/.build *.pt # wandb log -example/wandb/ +examples/wandb/ +examples/logs/ +examples/output/ examples/awesome-chatgpt-prompts/ +temp/ + +# ColossalChat +applications/ColossalChat/logs +applications/ColossalChat/models +applications/ColossalChat/sft_data +applications/ColossalChat/prompt_data +applications/ColossalChat/preference_data +applications/ColossalChat/temp diff --git a/applications/Chat/LICENSE b/applications/ColossalChat/LICENSE old mode 100644 new mode 100755 similarity index 100% rename from applications/Chat/LICENSE rename to applications/ColossalChat/LICENSE diff --git a/applications/Chat/README.md b/applications/ColossalChat/README.md old mode 100644 new mode 100755 similarity index 59% rename from applications/Chat/README.md rename to applications/ColossalChat/README.md index 349c26aad746..769f0b3d072c --- a/applications/Chat/README.md +++ b/applications/ColossalChat/README.md @@ -13,10 +13,10 @@ - [Install the environment](#install-the-environment) - [Install the Transformers](#install-the-transformers) - [How to use?](#how-to-use) - - [Supervised datasets collection](#supervised-datasets-collection) - - [RLHF Training Stage1 - Supervised instructs tuning](#RLHF-training-stage1---supervised-instructs-tuning) - - [RLHF Training Stage2 - Training reward model](#RLHF-training-stage2---training-reward-model) - - [RLHF Training Stage3 - Training model with reinforcement learning by human feedback](#RLHF-training-stage3---training-model-with-reinforcement-learning-by-human-feedback) + - [Supervised datasets collection](#step-1-data-collection) + - [RLHF Training Stage1 - Supervised instructs tuning](#rlhf-training-stage1---supervised-instructs-tuning) + - [RLHF Training Stage2 - Training reward model](#rlhf-training-stage2---training-reward-model) + - [RLHF Training Stage3 - Training model with reinforcement learning by human feedback](#rlhf-training-stage3---proximal-policy-optimization) - [Inference Quantization and Serving - After Training](#inference-quantization-and-serving---after-training) - [Coati7B examples](#coati7b-examples) - [Generation](#generation) @@ -36,7 +36,7 @@ --- -## What is ColossalChat and Coati ? +## What Is ColossalChat And Coati ? [ColossalChat](https://github.com/hpcaitech/ColossalAI/tree/main/applications/Chat) is the project to implement LLM with RLHF, powered by the [Colossal-AI](https://github.com/hpcaitech/ColossalAI) project. @@ -91,107 +91,191 @@ More details can be found in the latest news. ## Install -### Install the environment +### Install the Environment ```bash -conda create -n coati -conda activate coati +# Create new environment +conda create -n colossal-chat python=3.10.9 (>=3.8.7) +conda activate colossal-chat + +# Install flash-attention +git clone -b v2.0.5 https://github.com/Dao-AILab/flash-attention.git +cd $FLASH_ATTENTION_ROOT/ +pip install . +cd $FLASH_ATTENTION_ROOT/csrc/xentropy +pip install . +cd $FLASH_ATTENTION_ROOT/csrc/layer_norm +pip install . +cd $FLASH_ATTENTION_ROOT/csrc/rotary +pip install . + +# Clone Colossalai git clone https://github.com/hpcaitech/ColossalAI.git -cd ColossalAI/applications/Chat + +# Install ColossalAI +cd $COLOSSAL_AI_ROOT +BUILD_EXT=1 pip install . + +# Install ColossalChat +cd $COLOSSAL_AI_ROOT/applications/Chat pip install . ``` -### Install the Transformers +## How To Use? -```bash -pip install transformers==4.30.2 -``` +### RLHF Training Stage1 - Supervised Instructs Tuning -## How to use? +Stage1 is supervised instructs fine-tuning (SFT). This step is a crucial part of the RLHF training process, as it involves training a machine learning model using human-provided instructions to learn the initial behavior for the task at hand. Here's a detailed guide on how to SFT your LLM with ColossalChat. More details can be found in [example guideline](./examples/README.md). -### Supervised datasets collection +#### Step 1: Data Collection +The first step in Stage 1 is to collect a dataset of human demonstrations of the following format. -We collected 104K bilingual datasets of Chinese and English, and you can find the datasets in this repo -[InstructionWild](https://github.com/XueFuzhao/InstructionWild) and in this [file](https://github.com/XueFuzhao/InstructionWild/blob/main/data/README.md). +```json +[ + {"messages": + [ + { + "from": "human", + "content": "what are some pranks with a pen i can do?" + }, + { + "from": "assistant", + "content": "Are you looking for practical joke ideas?" + }, + ... + ] + }, + ... +] +``` -Here is how we collected the data +#### Step 2: Preprocessing +Once you have collected your SFT dataset, you will need to preprocess it. This involves four steps: data cleaning, data deduplication, formatting and tokenization. In this section, we will focus on formatting and tokenization. -

      - -

      +In this code, we provide a flexible way for users to set the conversation template for formatting chat data using Huggingface's newest feature--- chat template. Please follow the [example guideline](./examples/README.md) on how to format and tokenize data. -### RLHF Training Stage1 - Supervised instructs tuning +#### Step 3: Training +Choose a suitable model architecture for your task. Note that your model should be compatible with the tokenizer that you used to tokenize the SFT dataset. You can run [train_sft.sh](./examples/training_scripts/train_sft.sh) to start a supervised instructs fine-tuning. More details can be found in [example guideline](./examples/README.md). -Stage1 is supervised instructs fine-tuning, which uses the datasets mentioned earlier to fine-tune the model. +### RLHF Training Stage2 - Training Reward Model -You can run the `examples/train_sft.sh` to start a supervised instructs fine-tuning. -[[Stage1 tutorial video]](https://www.youtube.com/watch?v=-qFBZFmOJfg) +Stage2 trains a reward model, which obtains corresponding scores by manually ranking different outputs for the same prompt and supervises the training of the reward model. -**Note**: the supervised dataset follows the following format, +#### Step 1: Data Collection +Below shows the preference dataset format used in training the reward model. ```json [ - { - "instruction": "Provide a list of the top 10 most popular mobile games in Asia", - "input": "", - "output": "The top 10 most popular mobile games in Asia are:\n1) PUBG Mobile\n2) Pokemon Go\n3) Candy Crush Saga\n4) Free Fire\n5) Clash of Clans\n6) Mario Kart Tour\n7) Arena of Valor\n8) Fantasy Westward Journey\n9) Subway Surfers\n10) ARK Survival Evolved", - "id": 0 + {"context": [ + { + "from": "human", + "content": "Introduce butterflies species in Oregon." + } + ] + "chosen": [ + { + "from": "assistant", + "content": "About 150 species of butterflies live in Oregon, with about 100 species are moths..." + }, + ... + ], + "rejected": [ + { + "from": "assistant", + "content": "Are you interested in just the common butterflies? There are a few common ones which will be easy to find..." + }, + ... + ] }, ... ] ``` -### RLHF Training Stage2 - Training reward model - -Stage2 trains a reward model, which obtains corresponding scores by manually ranking different outputs for the same prompt and supervises the training of the reward model +#### Step 2: Preprocessing +Similar to the second step in the previous stage, we format the reward data into the same structured format as used in step 2 of the SFT stage. You can run [prepare_preference_dataset.sh](./examples/data_preparation_scripts/prepare_preference_dataset.sh) to prepare the preference data for reward model training. -You can run the `examples/train_rm.sh` to start a reward model training. -[[Stage2 tutorial video]](https://www.youtube.com/watch?v=gMx2CApKhuo) +#### Step 3: Training +You can run [train_rm.sh](./examples/training_scripts/train_rm.sh) to start the reward model training. More details can be found in [example guideline](./examples/README.md). -### RLHF Training Stage3 - Training model with reinforcement learning by human feedback +### RLHF Training Stage3 - Proximal Policy Optimization -Stage3 uses reinforcement learning algorithm, which is the most complex part of the training process: +In stage3 we will use reinforcement learning algorithm--- Proximal Policy Optimization (PPO), which is the most complex part of the training process:

      -You can run the `examples/train_prompts.sh` to start training PPO with human feedback. -[[Stage3 tutorial video]](https://www.youtube.com/watch?v=Z8wwSHxPL9g) - -**Note**: the required datasets follow the following format, - -- `pretrain dataset` - - ```json - [ - { - "instruction": "Provide a list of the top 10 most popular mobile games in Asia", - "input": "", - "output": "The top 10 most popular mobile games in Asia are:\n1) PUBG Mobile\n2) Pokemon Go\n3) Candy Crush Saga\n4) Free Fire\n5) Clash of Clans\n6) Mario Kart Tour\n7) Arena of Valor\n8) Fantasy Westward Journey\n9) Subway Surfers\n10) ARK Survival Evolved", - "id": 0 - }, - ... - ] - ``` - -- `prompt dataset` - - ```json - [ - { - "instruction": "Edit this paragraph to make it more concise: \"Yesterday, I went to the store and bought some things. Then, I came home and put them away. After that, I went for a walk and met some friends.\"", - "id": 0 - }, - { - "instruction": "Write a descriptive paragraph about a memorable vacation you went on", - "id": 1 - }, - ... - ] - ``` - -For more details, see [`examples/`](https://github.com/hpcaitech/ColossalAI/tree/main/applications/Chat/examples). +#### Step 1: Data Collection +PPO uses two kind of training data--- the prompt data and the sft data (optional). The first dataset is mandatory, data samples within the prompt dataset ends with a line from "human" and thus the "assistant" needs to generate a response to answer to the "human". Note that you can still use conversation that ends with a line from the "assistant", in that case, the last line will be dropped. Here is an example of the prompt dataset format. + +```json +[ + {"messages": + [ + { + "from": "human", + "content": "what are some pranks with a pen i can do?" + } + ... + ] + }, +] +``` + +#### Step 2: Data Preprocessing +To prepare the prompt dataset for PPO training, simply run [prepare_prompt_dataset.sh](./examples/data_preparation_scripts/prepare_prompt_dataset.sh) + +#### Step 3: Training +You can run the [train_ppo.sh](./examples/training_scripts/train_ppo.sh) to start PPO training. Here are some unique arguments for PPO, please refer to the training configuration section for other training configuration. More detais can be found in [example guideline](./examples/README.md). + +```bash +--pretrain $PRETRAINED_MODEL_PATH \ +--rm_pretrain $PRETRAINED_MODEL_PATH \ # reward model architectual +--tokenizer_dir $PRETRAINED_TOKENIZER_PATH \ +--rm_checkpoint_path $REWARD_MODEL_PATH \ # reward model checkpoint path +--prompt_dataset ${prompt_dataset[@]} \ # List of string, the prompt dataset +--ptx_dataset ${ptx_dataset[@]} \ # List of string, the SFT data used in the SFT stage +--ptx_batch_size 1 \ # batch size for calculate ptx loss +--ptx_coef 0.0 \ # none-zero if ptx loss is enable +--num_episodes 2000 \ # number of episodes to train +--num_collect_steps 1 \ +--num_update_steps 1 \ +--experience_batch_size 8 \ +--train_batch_size 4 \ +--accumulation_steps 2 +``` + +Each episode has two phases, the collect phase and the update phase. During the collect phase, we will collect experiences (answers generated by actor), store those in ExperienceBuffer. Then data in ExperienceBuffer is used during the update phase to update parameter of actor and critic. + +- Without tensor parallelism, +``` +experience buffer size += num_process * num_collect_steps * experience_batch_size += train_batch_size * accumulation_steps * num_process +``` + +- With tensor parallelism, +``` +num_tp_group = num_process / tp +experience buffer size += num_tp_group * num_collect_steps * experience_batch_size += train_batch_size * accumulation_steps * num_tp_group +``` + +## Alternative Option For RLHF: Direct Preference Optimization + +For those seeking an alternative to Reinforcement Learning from Human Feedback (RLHF), Direct Preference Optimization (DPO) presents a compelling option. DPO, as detailed in the paper (available at [https://arxiv.org/abs/2305.18290](https://arxiv.org/abs/2305.18290)), DPO offers an low-cost way to perform RLHF and usually request less computation resources compares to PPO. + +### DPO Training Stage1 - Supervised Instructs Tuning + +Please refer the [sft section](#dpo-training-stage1---supervised-instructs-tuning) in the PPO part. + +### DPO Training Stage2 - DPO Training +#### Step 1: Data Collection & Preparation +For DPO training, you only need the preference dataset. Please follow the instruction in the [preference dataset preparation section](#rlhf-training-stage2---training-reward-model) to prepare the preference data for DPO training. + +#### Step 2: Training +You can run the [train_dpo.sh](./examples/training_scripts/train_dpo.sh) to start DPO training. More detais can be found in [example guideline](./examples/README.md). ### Inference Quantization and Serving - After Training @@ -301,91 +385,60 @@ You can find more examples in this [repo](https://github.com/XueFuzhao/Instructi We have integrated the Transformers save and load pipeline, allowing users to freely call Hugging Face's language models and save them in the HF format. +- Option 1: Save the model weights, model config and generation config (Note: tokenizer will not be saved) which can be loaded using HF's from_pretrained method. ```python -from coati.models.llama import LlamaLM -from coati.trainer import SFTTrainer - -model = LlamaLM(pretrained=args.pretrain) -tokenizer = AutoTokenizer.from_pretrained(args.pretrain) - -(model, optim) = strategy.prepare((model, optim)) -trainer = SFTTrainer(model=model, - strategy=strategy, - optim=optim, - train_dataloader=train_dataloader, - eval_dataloader=eval_dataloader, - batch_size=args.batch_size, - max_epochs=args.max_epochs, - accumulation_steps=args.accumulation_steps - ) - -trainer.fit() -# this saves in pytorch format -strategy.save_model(model, args.save_path, only_rank0=True) - -# this saves in HF format -strategy.save_pretrained(model, args.save_path, only_rank0=True, tokenizer=tokenizer) +# if use lora, you can choose to merge lora weights before saving +if args.lora_rank > 0 and args.merge_lora_weights: + from coati.models.lora import LORA_MANAGER + + # NOTE: set model to eval to merge LoRA weights + LORA_MANAGER.merge_weights = True + model.eval() +# save model checkpoint after fitting on only rank0 +booster.save_model(model, os.path.join(args.save_dir, "modeling"), shard=True) + ``` +- Option 2: Save the model weights, model config, generation config, as well as the optimizer, learning rate scheduler, running states (Note: tokenizer will not be saved) which are needed for resuming training. +```python +from coati.utils import save_checkpoint +# save model checkpoint after fitting on only rank0 +save_checkpoint( + save_dir=actor_save_dir, + booster=actor_booster, + model=model, + optimizer=optim, + lr_scheduler=lr_scheduler, + epoch=0, + step=step, + batch_size=train_batch_size, + coordinator=coordinator, + ) +``` +To load the saved checkpoint +```python +from coati.utils import load_checkpoint +start_epoch, start_step, sampler_start_idx = load_checkpoint( + load_dir=checkpoint_path, + booster=booster, + model=model, + optimizer=optim, + lr_scheduler=lr_scheduler, + ) +```
      How to train with limited resources -Here are some examples that can allow you to train a 7B model on a single or multiple consumer-grade GPUs. - -If you only have a single 24G GPU, you can use the following script. `batch_size`, `lora_rank` and `grad_checkpoint` are the most important parameters to successfully train the model. - -```bash -// [INFO]: MAX GPU MEMORY ALLOCATED: 19148.9345703125 MB -torchrun --standalone --nproc_per_node=1 train_sft.py \ - --pretrain "/path/to/LLaMa-7B/" \ - --model 'llama' \ - --strategy ddp \ - --save_path /path/to/Coati-7B \ - --dataset /path/to/data.json \ - --batch_size 1 \ - --accumulation_steps 8 \ - --lr 2e-5 \ - --max_datasets_size 512 \ - --max_epochs 1 \ - --lora_rank 16 \ - --grad_checkpoint -``` +Here are some suggestions that can allow you to train a 7B model on a single or multiple consumer-grade GPUs. -`colossalai_gemini` strategy can enable a single 24G GPU to train the whole model without using LoRA if you have sufficient CPU memory. You can use the following script. +`batch_size`, `lora_rank` and `grad_checkpoint` are the most important parameters to successfully train the model. To maintain a descent batch size for gradient calculation, consider increase the accumulation_step and reduce the batch_size on each rank. -```bash -torchrun --standalone --nproc_per_node=1 train_sft.py \ - --pretrain "/path/to/LLaMa-7B/" \ - --model 'llama' \ - --strategy colossalai_gemini \ - --save_path /path/to/Coati-7B \ - --dataset /path/to/data.json \ - --batch_size 1 \ - --accumulation_steps 8 \ - --lr 2e-5 \ - --max_datasets_size 512 \ - --max_epochs 1 \ - --grad_checkpoint -``` +If you only have a single 24G GPU. Generally, using lora and "zero2-cpu" will be sufficient. -If you have 4x32 GB GPUs, you can even train the whole 7B model using our `colossalai_zero2_cpu` strategy! The script is given as follows. - -```bash -torchrun --standalone --nproc_per_node=4 train_sft.py \ - --pretrain "/path/to/LLaMa-7B/" \ - --model 'llama' \ - --strategy colossalai_zero2_cpu \ - --save_path /path/to/Coati-7B \ - --dataset /path/to/data.json \ - --batch_size 1 \ - --accumulation_steps 8 \ - --lr 2e-5 \ - --max_datasets_size 512 \ - --max_epochs 1 \ - --grad_checkpoint -``` +`gemini` and `gemini-auto` can enable a single 24G GPU to train the whole model without using LoRA if you have sufficient CPU memory. But that strategy doesn't support gradient accumulation. +If you have multiple GPUs each has very limited VRAM, say 8GB. You can try the `3d` for the plugin option, which supports tensor parellelism, set `--tp` to the number of GPUs that you have.
      ## The Plan @@ -396,6 +449,8 @@ torchrun --standalone --nproc_per_node=4 train_sft.py \ - [x] support inference - [x] support llama from [facebook](https://github.com/facebookresearch/llama) - [x] implement PPO-ptx fine-tuning +- [x] support flash-attention +- [x] implement DPO fine-tuning - [ ] integrate with Ray - [ ] support more RL paradigms, like Implicit Language Q-Learning (ILQL), - [ ] support chain-of-thought by [langchain](https://github.com/hwchase17/langchain) @@ -467,6 +522,7 @@ Coati is developed by ColossalAI Team: - [Fazzie](https://fazzie-key.cool/about/index.html) Contributing to the algorithm and development for SFT. - [ofey404](https://github.com/ofey404) Contributing to both front-end and back-end development. - [Wenhao Chen](https://github.com/CWHer) Contributing to subsequent code enhancements and performance improvements. +- [Anbang Ye](https://github.com/YeAnbang) Contributing to the refactored version with updated acceleration framework, LoRA, DPO and PPO. The PhD student from [(HPC-AI) Lab](https://ai.comp.nus.edu.sg/) also contributed a lot to this project. - [Zangwei Zheng](https://github.com/zhengzangw) diff --git a/applications/ColossalChat/benchmarks/Opt.json b/applications/ColossalChat/benchmarks/Opt.json new file mode 100644 index 000000000000..6d47666bb056 --- /dev/null +++ b/applications/ColossalChat/benchmarks/Opt.json @@ -0,0 +1,17 @@ +{ + "chat_template": "{% for message in messages %}{% if message['role'] == 'user' %}{{'Human: ' + bos_token + message['content'].strip() + eos_token }}{% elif message['role'] == 'system' %}{{ message['content'].strip() + '\\n\\n' }}{% elif message['role'] == 'assistant' %}{{ 'Assistant: ' + bos_token + message['content'].strip() + eos_token }}{% endif %}{% endfor %}", + "system_message": "A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions.\n\n", + "human_line_start": [ + 2 + ], + "human_line_end": [ + 2 + ], + "assistant_line_start": [ + 2 + ], + "assistant_line_end": [ + 2 + ], + "end_of_system_line_position": 0 +} diff --git a/applications/ColossalChat/benchmarks/README.md b/applications/ColossalChat/benchmarks/README.md new file mode 100755 index 000000000000..08c5e0e6c685 --- /dev/null +++ b/applications/ColossalChat/benchmarks/README.md @@ -0,0 +1,37 @@ +# Benchmarks + +## Benchmark OPT with LoRA on dummy prompt data + +We provide various OPT models (string in parentheses is the corresponding model name used in this script): + +- OPT-125M (125m) +- OPT-350M (350m) +- OPT-700M (700m) +- OPT-1.3B (1.3b) +- OPT-2.7B (2.7b) +- OPT-3.5B (3.5b) +- OPT-5.5B (5.5b) +- OPT-6.7B (6.7b) +- OPT-10B (10b) +- OPT-13B (13b) + +We also provide various training strategies: + +- gemini: ColossalAI GeminiPlugin with `placement_policy="cuda"`, like zero3 +- gemini_auto: ColossalAI GeminiPlugin with `placement_policy="cpu"`, like zero3-offload +- zero2: ColossalAI zero2 +- zero2_cpu: ColossalAI zero2-offload +- 3d: ColossalAI HybridParallelPlugin with TP, DP support + +## How to Run +```bash +cd ../tests +# Prepare data for benchmark +SFT_DATASET=/path/to/sft/data/ \ +PROMPT_DATASET=/path/to/prompt/data/ \ +PRETRAIN_DATASET=/path/to/ptx/data/ \ +PREFERENCE_DATASET=/path/to/preference/data \ +./test_data_preparation.sh +# Start benchmark +./benchmark_ppo.sh +``` diff --git a/applications/ColossalChat/benchmarks/benchmark_memory_consumption.txt b/applications/ColossalChat/benchmarks/benchmark_memory_consumption.txt new file mode 100644 index 000000000000..049285552d4f --- /dev/null +++ b/applications/ColossalChat/benchmarks/benchmark_memory_consumption.txt @@ -0,0 +1,4 @@ +Model=Opt-125m; lora_rank=0; plugin=zero2 +Max CUDA memory usage: 26123.16 MB +Model=Opt-125m; lora_rank=0; plugin=zero2 +Max CUDA memory usage: 26123.91 MB diff --git a/applications/ColossalChat/benchmarks/benchmark_performance_summarization.txt b/applications/ColossalChat/benchmarks/benchmark_performance_summarization.txt new file mode 100644 index 000000000000..b2a1ff1d77f2 --- /dev/null +++ b/applications/ColossalChat/benchmarks/benchmark_performance_summarization.txt @@ -0,0 +1,16 @@ +facebook/opt-125m; 0; zero2 +Performance summary: +Generate 768 samples, throughput: 188.48 samples/s, TFLOPS per GPU: 361.23 +Train 768 samples, throughput: 448.38 samples/s, TFLOPS per GPU: 82.84 +Overall throughput: 118.42 samples/s +Overall time per sample: 0.01 s +Make experience time per sample: 0.01 s, 62.83% +Learn time per sample: 0.00 s, 26.41% +facebook/opt-125m; 0; zero2 +Performance summary: +Generate 768 samples, throughput: 26.32 samples/s, TFLOPS per GPU: 50.45 +Train 768 samples, throughput: 71.15 samples/s, TFLOPS per GPU: 13.14 +Overall throughput: 18.86 samples/s +Overall time per sample: 0.05 s +Make experience time per sample: 0.04 s, 71.66% +Learn time per sample: 0.01 s, 26.51% diff --git a/applications/ColossalChat/benchmarks/benchmark_ppo.py b/applications/ColossalChat/benchmarks/benchmark_ppo.py new file mode 100644 index 000000000000..e1b7a313f981 --- /dev/null +++ b/applications/ColossalChat/benchmarks/benchmark_ppo.py @@ -0,0 +1,523 @@ +""" +For becnhmarking ppo. Mudified from examples/training_scripts/train_ppo.py +""" + +import argparse +import json +import os +import resource +from contextlib import nullcontext + +import torch +import torch.distributed as dist +from coati.dataset import ( + DataCollatorForPromptDataset, + DataCollatorForSupervisedDataset, + StatefulDistributedSampler, + load_tokenized_dataset, + setup_conversation_template, + setup_distributed_dataloader, +) +from coati.models import Critic, RewardModel, convert_to_lora_module, disable_dropout +from coati.trainer import PPOTrainer +from coati.trainer.callbacks import PerformanceEvaluator +from coati.trainer.utils import is_rank_0 +from coati.utils import load_checkpoint, replace_with_flash_attention +from transformers import AutoTokenizer, OPTForCausalLM +from transformers.models.opt.configuration_opt import OPTConfig + +import colossalai +from colossalai.booster import Booster +from colossalai.booster.plugin import GeminiPlugin, HybridParallelPlugin, LowLevelZeroPlugin +from colossalai.cluster import DistCoordinator +from colossalai.lazy import LazyInitContext +from colossalai.nn.lr_scheduler import CosineAnnealingWarmupLR +from colossalai.nn.optimizer import HybridAdam +from colossalai.utils import get_current_device + + +def get_model_numel(model: torch.nn.Module, plugin: str, tp: int) -> int: + numel = sum(p.numel() for p in model.parameters()) + if plugin == "3d" and tp > 1: + numel *= dist.get_world_size() + return numel + + +def get_gpt_config(model_name: str) -> OPTConfig: + model_map = { + "125m": OPTConfig.from_pretrained("facebook/opt-125m"), + "350m": OPTConfig(hidden_size=1024, ffn_dim=4096, num_hidden_layers=24, num_attention_heads=16), + "700m": OPTConfig(hidden_size=1280, ffn_dim=5120, num_hidden_layers=36, num_attention_heads=20), + "1.3b": OPTConfig.from_pretrained("facebook/opt-1.3b"), + "2.7b": OPTConfig.from_pretrained("facebook/opt-2.7b"), + "3.5b": OPTConfig(hidden_size=3072, ffn_dim=12288, num_hidden_layers=32, num_attention_heads=32), + "5.5b": OPTConfig(hidden_size=3840, ffn_dim=15360, num_hidden_layers=32, num_attention_heads=32), + "6.7b": OPTConfig.from_pretrained("facebook/opt-6.7b"), + "10b": OPTConfig(hidden_size=5120, ffn_dim=20480, num_hidden_layers=32, num_attention_heads=32), + "13b": OPTConfig.from_pretrained("facebook/opt-13b"), + } + try: + return model_map[model_name] + except KeyError: + raise ValueError(f'Unknown model "{model_name}"') + + +def benchmark_train(args): + # ============================== + # Initialize Distributed Training + # ============================== + colossalai.launch_from_torch({}) + coordinator = DistCoordinator() + + # ====================================================== + # Initialize Model, Objective, Optimizer and LR Scheduler + # ====================================================== + init_ctx = LazyInitContext(default_device=get_current_device()) if "gemini" in args.plugin else nullcontext() + + booster_policy = None + with init_ctx: + actor = OPTForCausalLM(config=get_gpt_config(args.pretrain)) + # Disable dropout + disable_dropout(actor) + ref_model = OPTForCausalLM(config=get_gpt_config(args.pretrain)) + reward_model = RewardModel(config=get_gpt_config("350m")) + critic = Critic(config=get_gpt_config("350m")) + disable_dropout(critic) + + actor_numel = get_model_numel(actor, args.plugin, args.tp) + critic_numel = get_model_numel(critic, args.plugin, args.tp) + initial_model_numel = get_model_numel(ref_model, args.plugin, args.tp) + reward_model_numel = get_model_numel(reward_model, args.plugin, args.tp) + + performance_evaluator = PerformanceEvaluator( + actor_numel, + critic_numel, + initial_model_numel, + reward_model_numel, + enable_grad_checkpoint=False, + ignore_episodes=2, + train_config={"model": "facebook/opt-" + args.pretrain, "lora_rank": args.lora_rank, "plugin": args.plugin}, + save_path="./benchmark_performance_summarization.txt", + ) + + if args.tp > 1: + if reward_model.model.config.architectures[0] != critic.model.config.architectures[0]: + raise ValueError("Reward model and critic model must have the same architecture") + if reward_model.model.config.architectures[0] == "BloomForCausalLM": + from colossalai.shardformer.policies.bloom import BloomPolicy + + booster_policy = BloomPolicy() + elif reward_model.model.config.architectures[0] == "LlamaForCausalLM": + from colossalai.shardformer.policies.llama import LlamaPolicy + + booster_policy = LlamaPolicy() + elif reward_model.model.config.architectures[0] == "GPT2LMHeadModel": + from colossalai.shardformer.policies.gpt2 import GPT2Policy + + booster_policy = GPT2Policy() + elif reward_model.model.config.architectures[0] == "ChatGLMModel": + from colossalai.shardformer.policies.chatglm2 import ChatGLMPolicy + + booster_policy = ChatGLMPolicy() + elif reward_model.model.config.architectures[0] == "OPTForCausalLM": + from colossalai.shardformer.policies.opt import OPTPolicy + + booster_policy = OPTPolicy() + else: + raise ValueError("Unknown model architecture for policy") + + if args.lora_rank > 0: + actor = convert_to_lora_module(actor, args.lora_rank, lora_train_bias=args.lora_train_bias) + critic = convert_to_lora_module(critic, args.lora_rank, lora_train_bias=args.lora_train_bias) + + if args.grad_checkpoint and args.lora_rank == 0: + actor.gradient_checkpointing_enable() + critic.model.gradient_checkpointing_enable() + coordinator.print_on_master(msg="Gradient checkpointing enabled successfully") + elif args.lora_rank > 0: + coordinator.print_on_master(msg="Gradient checkpointing will be disabled when LoRA is enabled") + + if args.use_flash_attn: + replace_with_flash_attention(model=actor) + replace_with_flash_attention(model=critic) + coordinator.print_on_master(msg="Flash-attention enabled successfully") + + # configure tokenizer + tokenizer_dir = args.tokenizer_dir if args.tokenizer_dir is not None else args.pretrain + tokenizer = AutoTokenizer.from_pretrained(tokenizer_dir) + if os.path.exists(args.conversation_template_config): + conversation_template_config = json.load(open(args.conversation_template_config, "r", encoding="utf8")) + conversation_template = setup_conversation_template( + tokenizer, chat_template_config=conversation_template_config, save_path=args.conversation_template_config + ) + stop_token_ids = ( + conversation_template.assistant_line_end if len(conversation_template.assistant_line_end) > 0 else None + ) + else: + raise ValueError("Conversation template config is not provided or incorrect") + if hasattr(tokenizer, "pad_token") and hasattr(tokenizer, "eos_token") and tokenizer.eos_token is not None: + try: + # Some tokenizers doesn't allow to set pad_token mannually e.g., Qwen + tokenizer.pad_token = tokenizer.eos_token + except AttributeError as e: + logger.warning(f"Unable to set pad token to eos token, {str(e)}") + if not hasattr(tokenizer, "pad_token") or tokenizer.pad_token is None: + logger.warning( + "The tokenizer does not have a pad token which is required. May lead to unintended behavior in training, Please consider manually set them." + ) + tokenizer.add_bos_token = False + tokenizer.add_eos_token = False + tokenizer.padding_side = "left" # left padding for generation (online learning) + + # configure generation config + actor.generation_config.update( + pad_token_id=tokenizer.eos_token_id, bos_token_id=tokenizer.bos_token_id, eos_token_id=tokenizer.eos_token_id + ) + + # configure optimizer + coordinator.print_on_master(f"setting up optimizer for actor: lr={args.lr}, weight_decay={args.weight_decay}") + actor_optim = HybridAdam( + model_params=actor.parameters(), + lr=args.lr, + betas=(0.9, 0.95), + weight_decay=args.weight_decay, + adamw_mode=True, + ) + + coordinator.print_on_master(f"setting up optimizer for critic: lr={args.lr}, weight_decay={args.weight_decay}") + critic_optim = HybridAdam( + model_params=critic.parameters(), + lr=args.critic_lr, + betas=(0.9, 0.95), + weight_decay=args.weight_decay, + adamw_mode=True, + ) + + # configure dataset + coordinator.print_on_master(f"Load dataset: {args.prompt_dataset}") + mode_map = {"train": "train", "valid": "validation", "test": "test"} + train_prompt_dataset = load_tokenized_dataset(dataset_paths=args.prompt_dataset, mode="train", mode_map=mode_map) + coordinator.print_on_master(f"prompt dataset size: {len(train_prompt_dataset)}") + data_collator = DataCollatorForPromptDataset(tokenizer=tokenizer, max_length=args.max_length - args.max_seq_len) + train_prompt_dataloader = setup_distributed_dataloader( + dataset=train_prompt_dataset, + batch_size=args.experience_batch_size, + shuffle=True, + drop_last=True, + collate_fn=data_collator, + use_tp=args.tp > 1, + ) + + if len(args.pretrain_dataset) > 0: + train_pretrain_dataset = load_tokenized_dataset( + dataset_paths=args.pretrain_dataset, mode="train", mode_map=mode_map + ) + data_collator = DataCollatorForSupervisedDataset(tokenizer=tokenizer, max_length=args.max_length) + train_pretrain_dataloader = setup_distributed_dataloader( + dataset=train_pretrain_dataset, + batch_size=args.ptx_batch_size, + shuffle=True, + drop_last=True, + collate_fn=data_collator, + use_tp=args.tp > 1, + ) + else: + train_pretrain_dataloader = None + + if args.warmup_steps is None: + args.warmup_steps = int(0.025 * args.num_episodes) + coordinator.print_on_master(f"Warmup steps is set to {args.warmup_steps}") + + actor_lr_scheduler = CosineAnnealingWarmupLR( + optimizer=actor_optim, + total_steps=args.num_episodes, + warmup_steps=args.warmup_steps, + eta_min=0.1 * args.lr, + ) + + critic_lr_scheduler = CosineAnnealingWarmupLR( + optimizer=critic_optim, + total_steps=args.num_episodes, + warmup_steps=args.warmup_steps, + eta_min=0.1 * args.lr, + ) + + # ============================== + # Initialize Booster + # ============================== + if args.plugin == "gemini": + plugin = GeminiPlugin( + precision=args.mixed_precision, + initial_scale=2**16, + max_norm=args.grad_clip, + ) + elif args.plugin == "gemini_auto": + plugin = GeminiPlugin( + precision=args.mixed_precision, + placement_policy="auto", + initial_scale=2**16, + max_norm=args.grad_clip, + ) + elif args.plugin == "zero2": + plugin = LowLevelZeroPlugin( + stage=2, + precision=args.mixed_precision, + initial_scale=2**16, + max_norm=args.grad_clip, + ) + elif args.plugin == "zero2_cpu": + plugin = LowLevelZeroPlugin( + stage=2, + precision=args.mixed_precision, + initial_scale=2**16, + cpu_offload=True, + max_norm=args.grad_clip, + ) + elif args.plugin == "3d": + plugin = HybridParallelPlugin( + tp_size=args.tp, + pp_size=1, + zero_stage=0, + precision=args.mixed_precision, + ) + custom_plugin = HybridParallelPlugin( + tp_size=args.tp, + pp_size=1, + zero_stage=0, + precision=args.mixed_precision, + custom_policy=booster_policy, + ) + else: + raise ValueError(f"Unknown plugin {args.plugin}") + + if args.plugin != "3d": + custom_plugin = plugin + + actor_booster = Booster(plugin=plugin) + ref_booster = Booster(plugin=plugin) + rm_booster = Booster(plugin=custom_plugin) + critic_booster = Booster(plugin=custom_plugin) + + default_dtype = torch.float16 if args.mixed_precision == "fp16" else torch.bfloat16 + torch.set_default_dtype(default_dtype) + actor, actor_optim, _, train_prompt_dataloader, actor_lr_scheduler = actor_booster.boost( + model=actor, + optimizer=actor_optim, + lr_scheduler=actor_lr_scheduler, + dataloader=train_prompt_dataloader, + ) + + critic, critic_optim, _, _, critic_lr_scheduler = critic_booster.boost( + model=critic, + optimizer=critic_optim, + lr_scheduler=critic_lr_scheduler, + dataloader=train_prompt_dataloader, + ) + reward_model, _, _, _, _ = rm_booster.boost(model=reward_model, dataloader=train_prompt_dataloader) + ref_model, _, _, _, _ = ref_booster.boost(model=ref_model, dataloader=train_prompt_dataloader) + + torch.set_default_dtype(torch.float) + + coordinator.print_on_master(f"Booster init max CUDA memory: {torch.cuda.max_memory_allocated() / 1024 ** 2:.2f} MB") + coordinator.print_on_master( + f"Booster init max CPU memory: {resource.getrusage(resource.RUSAGE_SELF).ru_maxrss / 1024:.2f} MB" + ) + + sampler_start_idx = 0 + start_step = 0 + + if args.rm_checkpoint_path is not None: + if "modeling" in args.rm_checkpoint_path: + rm_booster.load_model(reward_model, args.rm_checkpoint_path) + else: + _, _, _ = load_checkpoint( + load_dir=args.rm_checkpoint_path, + booster=rm_booster, + model=reward_model, + optimizer=None, + lr_scheduler=None, + ) + coordinator.print_on_master(f"Loaded reward model checkpoint {args.rm_checkpoint_path}") + + if args.checkpoint_path is not None: + if "modeling" in args.checkpoint_path: + actor_booster.load_model(actor, args.checkpoint_path) + ref_booster.load_model(ref_model, args.checkpoint_path) + coordinator.print_on_master(f"Loaded actor and reference model {args.checkpoint_path}") + else: + _, start_step, sampler_start_idx = load_checkpoint( + load_dir=args.checkpoint_path, + booster=actor_booster, + model=actor, + optimizer=actor_optim, + lr_scheduler=actor_lr_scheduler, + ) + _, _, _ = load_checkpoint( + load_dir=args.checkpoint_path, + booster=ref_booster, + model=ref_model, + optimizer=critic_optim, + lr_scheduler=critic_lr_scheduler, + ) + assert isinstance(train_prompt_dataloader.sampler, StatefulDistributedSampler) + train_prompt_dataloader.sampler.set_start_index(start_index=sampler_start_idx) + + coordinator.print_on_master( + f"Loaded actor and reference model checkpoint {args.checkpoint_path} at spisode {start_step}" + ) + coordinator.print_on_master(f"Loaded sample at index {sampler_start_idx}") + + coordinator.print_on_master( + f"Checkpoint loaded max CUDA memory: {torch.cuda.max_memory_allocated() / 1024 ** 2:.2f} MB" + ) + coordinator.print_on_master( + f"Checkpoint loaded CUDA memory: {torch.cuda.memory_allocated() / 1024 ** 2:.2f} MB" + ) + coordinator.print_on_master( + f"Checkpoint loaded max CPU memory: {resource.getrusage(resource.RUSAGE_SELF).ru_maxrss / 1024:.2f} MB" + ) + + if args.critic_checkpoint_path is not None: + if "modeling" in args.critic_checkpoint_path: + critic_booster.load_model(critic, args.critic_checkpoint_path) + else: + _, _, _ = load_checkpoint( + load_dir=args.critic_checkpoint_path, + booster=critic_booster, + model=critic, + optimizer=critic_optim, + lr_scheduler=critic_lr_scheduler, + ) + coordinator.print_on_master(f"Loaded critic checkpoint {args.critic_checkpoint_path}") + coordinator.print_on_master( + f"Checkpoint loaded max CUDA memory: {torch.cuda.max_memory_allocated() / 1024 ** 2:.2f} MB" + ) + coordinator.print_on_master( + f"Checkpoint loaded CUDA memory: {torch.cuda.memory_allocated() / 1024 ** 2:.2f} MB" + ) + coordinator.print_on_master( + f"Checkpoint loaded max CPU memory: {resource.getrusage(resource.RUSAGE_SELF).ru_maxrss / 1024:.2f} MB" + ) + + # configure trainer + trainer = PPOTrainer( + actor_booster, + critic_booster, + actor, + critic, + reward_model, + ref_model, + actor_optim, + critic_optim, + actor_lr_scheduler, + critic_lr_scheduler, + tokenizer=tokenizer, + stop_token_ids=stop_token_ids, + kl_coef=args.kl_coef, + ptx_coef=args.ptx_coef, + train_batch_size=args.train_batch_size, + buffer_limit=args.num_collect_steps * args.experience_batch_size, + max_length=args.max_length, + max_new_tokens=args.max_seq_len, + use_cache=True, + do_sample=True, + temperature=0.7, + accumulation_steps=args.accumulation_steps, + save_dir=args.save_path, + save_interval=args.save_interval, + top_k=50, + use_tp=args.tp > 1, + offload_inference_models="gemini" not in args.plugin, + callbacks=[performance_evaluator], + coordinator=coordinator, + ) + + trainer.fit( + num_episodes=args.num_episodes, + num_collect_steps=args.num_collect_steps, + num_update_steps=args.num_update_steps, + prompt_dataloader=train_prompt_dataloader, + pretrain_dataloader=train_pretrain_dataloader, + log_dir=args.log_dir, + use_wandb=args.use_wandb, + ) + + if args.lora_rank > 0 and args.merge_lora_weights: + from coati.models.lora import LORA_MANAGER + + # NOTE: set model to eval to merge LoRA weights + LORA_MANAGER.merge_weights = True + actor.eval() + critic.eval() + # save model checkpoint after fitting on only rank0 + coordinator.print_on_master("Start saving final actor model checkpoint") + actor_booster.save_model(actor, os.path.join(trainer.actor_save_dir, "modeling"), shard=True) + coordinator.print_on_master( + f"Saved final actor model checkpoint at episodes {args.num_episodes} at folder {args.save_path}" + ) + coordinator.print_on_master("Start saving final critic model checkpoint") + critic_booster.save_model(critic, os.path.join(trainer.critic_save_dir, "modeling"), shard=True) + coordinator.print_on_master( + f"Saved final critic model checkpoint at episodes {args.num_episodes} at folder {args.save_path}" + ) + memory_consumption = torch.cuda.max_memory_allocated() / 1024**2 + if is_rank_0(): + with open("./benchmark_memory_consumption.txt", "a+") as f: + f.write( + f"Model=Opt-{args.pretrain}; lora_rank={args.lora_rank}; plugin={args.plugin}\nMax CUDA memory usage: {memory_consumption:.2f} MB\n" + ) + coordinator.print_on_master(f"Max CUDA memory usage: {memory_consumption:.2f} MB") + + +if __name__ == "__main__": + parser = argparse.ArgumentParser() + parser.add_argument("--prompt_dataset", nargs="+", default=[]) + parser.add_argument("--pretrain_dataset", nargs="+", default=[]) + parser.add_argument( + "--plugin", + type=str, + default="gemini", + choices=["gemini", "gemini_auto", "zero2", "zero2_cpu", "3d"], + help="Choose which plugin to use", + ) + parser.add_argument( + "--conversation_template_config", + type=str, + default=None, + help="Path \ + to save conversation template config files.", + ) + parser.add_argument("--grad_clip", type=float, default=1.0, help="Gradient clipping value") + parser.add_argument("--weight_decay", type=float, default=0.1, help="Weight decay") + parser.add_argument("--warmup_steps", type=int, default=None, help="Warmup steps") + parser.add_argument("--tokenizer_dir", type=str, default=None) + parser.add_argument("--tp", type=int, default=1) + parser.add_argument("--pretrain", type=str, default=None) + parser.add_argument("--checkpoint_path", type=str, default=None) + parser.add_argument("--critic_checkpoint_path", type=str, default=None) + parser.add_argument("--rm_checkpoint_path", type=str, help="Reward model checkpoint path") + parser.add_argument("--save_path", type=str, default="actor_checkpoint_prompts") + parser.add_argument("--num_episodes", type=int, default=1) + parser.add_argument("--num_collect_steps", type=int, default=2) + parser.add_argument("--num_update_steps", type=int, default=5) + parser.add_argument("--save_interval", type=int, default=1000) + parser.add_argument("--train_batch_size", type=int, default=16) + parser.add_argument("--experience_batch_size", type=int, default=16) + parser.add_argument("--ptx_batch_size", type=int, default=1) + parser.add_argument("--lora_train_bias", type=str, default="none") + parser.add_argument("--mixed_precision", type=str, default="fp16", choices=["fp16", "bf16"], help="Mixed precision") + parser.add_argument("--accumulation_steps", type=int, default=8) + parser.add_argument("--lora_rank", type=int, default=0, help="low-rank adaptation matrices rank") + parser.add_argument("--merge_lora_weights", type=bool, default=True) + parser.add_argument("--lr", type=float, default=9e-6) + parser.add_argument("--critic_lr", type=float, default=9e-6) + parser.add_argument("--kl_coef", type=float, default=0.1) + parser.add_argument("--ptx_coef", type=float, default=0.0) + parser.add_argument("--max_length", type=int, default=512) + parser.add_argument("--max_seq_len", type=int, default=256) + parser.add_argument("--log_dir", default="logs", type=str) + parser.add_argument("--use_wandb", default=False, action="store_true") + parser.add_argument("--grad_checkpoint", default=False, action="store_true") + parser.add_argument("--use_flash_attn", default=False, action="store_true") + args = parser.parse_args() + benchmark_train(args) diff --git a/applications/ColossalChat/benchmarks/benchmark_ppo.sh b/applications/ColossalChat/benchmarks/benchmark_ppo.sh new file mode 100755 index 000000000000..e88757659685 --- /dev/null +++ b/applications/ColossalChat/benchmarks/benchmark_ppo.sh @@ -0,0 +1,119 @@ +#!/usr/bin/env bash + +set_n_least_used_CUDA_VISIBLE_DEVICES() { + local n=${1:-"9999"} + echo "GPU Memory Usage:" + local FIRST_N_GPU_IDS=$(nvidia-smi --query-gpu=memory.used --format=csv | + tail -n +2 | + nl -v 0 | + tee /dev/tty | + sort -g -k 2 | + awk '{print $1}' | + head -n $n) + export CUDA_VISIBLE_DEVICES=$(echo $FIRST_N_GPU_IDS | sed 's/ /,/g') + echo "Now CUDA_VISIBLE_DEVICES is set to:" + echo "CUDA_VISIBLE_DEVICES=$CUDA_VISIBLE_DEVICES" +} + +set_n_least_used_CUDA_VISIBLE_DEVICES 8 + +set -xu + +NUM_RETRY=3 +BASE_DIR=$(dirname $(dirname $(realpath $BASH_SOURCE))) +EXAMPLES_DIR=$BASE_DIR/examples +TEMP_DIR=$BASE_DIR/temp +MODEL_SAVE_PATH=$TEMP_DIR/rlhf_models +MODELS_DIR=$TEMP_DIR/models_config +# To benchmark different models, change the following line +# MODELS=('125m' '350m' '700m' '1.3b' '2.7b' '3.5b' '5.5b' '6.7b' '10b' '13b') +MODELS=('125m') +# To benchmark different strategies, change the following line +# PLUGINS=('zero2', 'zero2_cpu', '3d') +PLUGINS=('zero2') +LORA_RANK=('0') + +export OMP_NUM_THREADS=8 + +rm ./benchmark_memory_consumption.txt +rm ./benchmark_performance_summarization.txt + +# install requirements +pip install -r $EXAMPLES_DIR/requirements.txt + +random_choice() { + local arr=("$@") + local len=${#arr[@]} + local idx=$((RANDOM % len)) + echo ${arr[$idx]} +} + +echo "[Test]: testing ppo ..." + +SKIPPED_TESTS=( +) + +GRAD_CKPTS=('' '--grad_checkpoint') +GRAD_CKPTS=('') +for lora_rank in ${LORA_RANK[@]}; do + for model in ${MODELS[@]}; do + plugins=($(shuf -e "${PLUGINS[@]}")) + for plugin in ${plugins[@]}; do + if [[ " ${SKIPPED_TESTS[*]} " =~ " $model-$plugin-$lora_rank " ]]; then + echo "[Test]: Skipped $model-$plugin-$lora_rank" + continue + elif [[ " ${SKIPPED_TESTS[*]} " =~ " $model-$plugin " ]]; then + echo "[Test]: Skipped $model-$plugin" + continue + fi + pretrain=$model + tokenizer_dir="facebook/opt-125m" + grad_ckpt=$(random_choice "${GRAD_CKPTS[@]}") + tp='1' + if [[ $plugin == "3d" ]]; then + tp='4' + fi + for i in $(seq $NUM_RETRY); do + echo "[Test]: $model-$plugin-$lora_rank, attempt $i" + declare -a prompt_dataset=() + for split in $(seq -f "%05g" 0 9); do + prompt_dataset+=("$TEMP_DIR/benchmark/arrow/part-$split") + done + colossalai run --nproc_per_node 8 --master_port 28547 $BASE_DIR/benchmarks/benchmark_ppo.py \ + --pretrain $pretrain \ + --tokenizer_dir $tokenizer_dir \ + --prompt_dataset ${prompt_dataset[@]} \ + --ptx_coef 0 \ + --save_path $MODEL_SAVE_PATH \ + --conversation_template_config ./Opt.json \ + --lora_rank $lora_rank \ + --plugin $plugin \ + --num_episodes 5 \ + --num_collect_steps 1 \ + --num_update_steps 1 \ + --max_seq_len 128 \ + --max_length 512 \ + --experience_batch_size 32 \ + --train_batch_size 32 \ + --accumulation_steps 1 \ + --lr 9e-6 \ + --mixed_precision "bf16" \ + --grad_clip 1.0 \ + --use_flash_attn \ + --tp $tp \ + --lr 2e-5 \ + $grad_ckpt + passed=$? + if [ $passed -eq 0 ]; then + rm -rf $MODEL_SAVE_PATH/* + rm -rf $MODELS_DIR/* + break + fi + done + if [ $passed -ne 0 ]; then + echo "[Test]: Failed $model-$plugin-$lora_rank" + exit 1 + fi + done + done +done diff --git a/applications/ColossalChat/benchmarks/data_preparation.sh b/applications/ColossalChat/benchmarks/data_preparation.sh new file mode 100755 index 000000000000..ca2986be43d5 --- /dev/null +++ b/applications/ColossalChat/benchmarks/data_preparation.sh @@ -0,0 +1,16 @@ +SAVE_DIR="" + + +BASE_DIR=$(dirname $(dirname $(realpath $BASH_SOURCE))) +EXAMPLES_DIR=$BASE_DIR/examples +SAVE_DIR=$BASE_DIR/temp/benchmark + +rm -rf $SAVE_DIR + +python $EXAMPLES_DIR/data_preparation_scripts/prepare_prompt_dataset.py --data_input_dirs "/home/yeanbang/data/dataset/sft_data/alpaca/data_preprocessed/train" \ + --conversation_template_config ./Opt.json \ + --tokenizer_dir "facebook/opt-125m" \ + --data_cache_dir $SAVE_DIR/cache \ + --data_jsonl_output_dir $SAVE_DIR/jsonl \ + --data_arrow_output_dir $SAVE_DIR/arrow \ + --num_samples_per_datafile 30 diff --git a/applications/Chat/benchmarks/ray/1mmt_dummy.py b/applications/ColossalChat/benchmarks/ray/1mmt_dummy.py old mode 100644 new mode 100755 similarity index 100% rename from applications/Chat/benchmarks/ray/1mmt_dummy.py rename to applications/ColossalChat/benchmarks/ray/1mmt_dummy.py diff --git a/applications/Chat/benchmarks/ray/mmmt_dummy.py b/applications/ColossalChat/benchmarks/ray/mmmt_dummy.py old mode 100644 new mode 100755 similarity index 100% rename from applications/Chat/benchmarks/ray/mmmt_dummy.py rename to applications/ColossalChat/benchmarks/ray/mmmt_dummy.py diff --git a/applications/Chat/coati/__init__.py b/applications/ColossalChat/coati/__init__.py old mode 100644 new mode 100755 similarity index 100% rename from applications/Chat/coati/__init__.py rename to applications/ColossalChat/coati/__init__.py diff --git a/applications/ColossalChat/coati/dataset/__init__.py b/applications/ColossalChat/coati/dataset/__init__.py new file mode 100755 index 000000000000..e216c37e1c62 --- /dev/null +++ b/applications/ColossalChat/coati/dataset/__init__.py @@ -0,0 +1,26 @@ +from .conversation import Conversation, setup_conversation_template +from .loader import ( + DataCollatorForPreferenceDataset, + DataCollatorForPromptDataset, + DataCollatorForSupervisedDataset, + StatefulDistributedSampler, + load_tokenized_dataset, + setup_distributed_dataloader, +) +from .tokenization_utils import supervised_tokenize_sft, tokenize_prompt_dataset, tokenize_rlhf + +__all__ = [ + "tokenize_prompt_dataset", + "DataCollatorForPromptDataset", + "is_rank_0", + "DataCollatorForPreferenceDataset", + "DataCollatorForSupervisedDataset", + "StatefulDistributedSampler", + "load_tokenized_dataset", + "setup_distributed_dataloader", + "supervised_tokenize_pretrain", + "supervised_tokenize_sft", + "tokenize_rlhf", + "setup_conversation_template", + "Conversation", +] diff --git a/applications/ColossalChat/coati/dataset/conversation.py b/applications/ColossalChat/coati/dataset/conversation.py new file mode 100755 index 000000000000..15a33be93966 --- /dev/null +++ b/applications/ColossalChat/coati/dataset/conversation.py @@ -0,0 +1,143 @@ +import dataclasses +import json +import os +from typing import Any, Dict, List + +import torch.distributed as dist +from transformers import AutoTokenizer, PreTrainedTokenizer + +from colossalai.logging import get_dist_logger + +logger = get_dist_logger() + + +@dataclasses.dataclass +class Conversation: + tokenizer: PreTrainedTokenizer + system_message: str + chat_template: str + stop_ids: List[int] + + @classmethod + def from_config(cls, tokenizer: PreTrainedTokenizer, config: Dict): + """ + Setup the conversation template from config + """ + tokenizer.chat_template = config["chat_template"] + conv = cls(tokenizer, config["system_message"], config["chat_template"], config["stop_ids"]) + conv.clear() + return conv + + def clear(self): + self.messages = [] + + @classmethod + def get_conversation_template_keys(cls): + return ["system_message", "chat_template"] + + def __str__(self): + return json.dumps( + {k: self.__dict__[k] for k in self.__dict__ if k not in ["tokenizer", "messages"]}, + ensure_ascii=False, + indent=4, + ) + + def get_prompt(self, length: int = None, add_generation_prompt=False) -> Any: + """ + Retrieves the prompt for the conversation. + + Args: + length (int, optional): The number of messages to include in the prompt. Defaults to None. + get_seps_info (bool, optional): Whether to include separator information in the output. Defaults to False. + add_generation_prompt (bool, optional): Whether to add the assistant line start token in generation (for generation only). Defaults to False. + + Returns: + str or tuple: The prompt string if get_seps_info is False, otherwise a tuple containing the prompt string and separator information. + """ + + if length is None: + length = len(self.messages) + + assert length <= len(self.messages) + if self.system_message is not None: + messages = [{"role": "system", "content": self.system_message}] + self.messages[:length] + else: + messages = self.messages[:length] + prompt = self.tokenizer.apply_chat_template( + messages, tokenize=False, add_generation_prompt=add_generation_prompt + ) + return prompt + + def save_prompt(self): + return self.get_prompt() + + def append_message(self, role: str, message: str): + """ + Append a message to the conversation. + + Args: + role (str): The role of the message sender. Must be either 'user' or 'assistant'. + message (str): The content of the message. + + Raises: + AssertionError: If the role is not 'user' or 'assistant'. + """ + assert role in ["user", "assistant"] + self.messages.append({"role": role, "content": message}) + + def copy(self): + return Conversation(tokenizer=self.tokenizer, chat_template=self.chat_template) + + +def setup_conversation_template( + tokenizer: PreTrainedTokenizer, chat_template_config: Dict = None, save_path: str = None +) -> Conversation: + """ + Setup the conversation template, if chat_template is given, will replace the default chat_template of the tokenizer + with it. Otherwise, the default chat_template will be used. If the tokenizer doesn't have a default chat_template, + raise error to remind the user to set it manually. + + Args: + tokenizer: The tokenizer to use + chat_template_config: + { + "system_message": str The system message to use + "chat_template": str The chat_template to use, if can be a chat_template, a huggingface model path or a local model. + if a huggeface model path or a local model, the chat_template will be loaded from the model's tokenizer's default chat template. + "stop_ids": List[int], the token ids used to terminate generation. You need to provide this for ppo training and generation. + } + """ + if any([s not in chat_template_config.keys() for s in Conversation.get_conversation_template_keys()]): + # Try to automatically set up conversation template, if fail, it throws an error that you need to do it manually + if "system_message" not in chat_template_config: + logger.warning("No system message is provided, will not use system message.") + if "chat_template" not in chat_template_config: + logger.warning("No chat_template is provided, will try to load it from the tokenizer.") + if tokenizer.chat_template != None: + chat_template_config["chat_template"] = tokenizer.chat_template + else: + raise ValueError( + f"Load a tokenizer from {chat_template_config['chat_template']}, which doesn't have a default chat template, please set it manually." + ) + else: + try: + tokenizer = AutoTokenizer.from_pretrained(chat_template_config["chat_template"]) + if tokenizer.chat_template != None: + chat_template_config["chat_template"] = tokenizer.chat_template + else: + raise ValueError( + f"Load a tokenizer from {chat_template_config['chat_template']}, which doesn't have a default chat template, please set it manually." + ) + logger.warning( + f"chat_template is provided as a local model path or huggingface model path, loaded chat_template from \"{chat_template_config['chat_template']}\"." + ) + except OSError: + pass + except ValueError as e: + raise ValueError(e) + if not dist.is_initialized() or dist.get_rank() == 0: + os.makedirs(os.path.dirname(save_path), exist_ok=True) + with open(save_path, "w", encoding="utf8") as f: + logger.info(f"Successfully generated a conversation tempalte config, save to {save_path}.") + json.dump(chat_template_config, f, indent=4, ensure_ascii=False) + return Conversation.from_config(tokenizer, chat_template_config) diff --git a/applications/ColossalChat/coati/dataset/loader.py b/applications/ColossalChat/coati/dataset/loader.py new file mode 100755 index 000000000000..93cc1dab8d21 --- /dev/null +++ b/applications/ColossalChat/coati/dataset/loader.py @@ -0,0 +1,383 @@ +#!/usr/bin/env python3 +# -*- coding: utf-8 -*- +""" +Dataloader for sft, dpo, ppo +""" + +import math +import os +import random +from dataclasses import dataclass +from typing import Callable, Dict, Iterator, List, Optional, Sequence, Union + +import numpy as np +import torch +import torch.distributed as dist +import torch.nn.functional as F +from coati.dataset.utils import chuncate_sequence, pad_to_max_len +from datasets import Dataset as HFDataset +from datasets import dataset_dict, load_from_disk +from torch.distributed import ProcessGroup +from torch.distributed.distributed_c10d import _get_default_group +from torch.utils.data import ConcatDataset, DataLoader, Dataset, DistributedSampler +from transformers.tokenization_utils import PreTrainedTokenizer + +DatasetType = Union[Dataset, ConcatDataset, dataset_dict.Dataset] +PathType = Union[str, os.PathLike] + + +def load_tokenized_dataset( + dataset_paths: Union[PathType, List[PathType]], mode: str = "train", **kwargs +) -> Optional[DatasetType]: + """ + Load pre-tokenized dataset. + Each instance of dataset is a dictionary with + `{'input_ids': List[int], 'labels': List[int], sequence: str}` format. + """ + mode_map = kwargs.get("mode_map", {"train": "train", "dev": "validation", "test": "test"}) + assert mode in tuple(mode_map), f"Unsupported mode {mode}, it must be in {tuple(mode_map)}" + + if isinstance(dataset_paths, (str, os.PathLike)): + dataset_paths = [dataset_paths] + + datasets = [] # `List[datasets.dataset_dict.Dataset]` + for ds_path in dataset_paths: + ds_path = os.path.abspath(ds_path) + assert os.path.exists(ds_path), f"Not existed file path {ds_path}" + ds_dict = load_from_disk(dataset_path=ds_path, keep_in_memory=False) + if isinstance(ds_dict, HFDataset): + datasets.append(ds_dict) + else: + if mode_map[mode] in ds_dict: + datasets.append(ds_dict[mode_map[mode]]) + if len(datasets) == 0: + return None + if len(datasets) == 1: + return datasets.pop() + return ConcatDataset(datasets=datasets) + + +@dataclass +class DataCollatorForSupervisedDataset(object): + """ + Collate instances for supervised dataset. + Each instance is a tokenized dictionary with fields + `input_ids`(List[int]), `labels`(List[int]) and `sequence`(str). + """ + + tokenizer: PreTrainedTokenizer + max_length: int = 4096 + ignore_index: int = -100 + + def __call__(self, instances: Sequence[Dict[str, List[int]]]) -> Dict[str, torch.Tensor]: + """ + + Args: + instances (`Sequence[Dict[str, List[int]]]`): + Mini-batch samples, each sample is stored in an individual dictionary. + + Returns: + (`Dict[str, torch.Tensor]`): Contains the following `torch.Tensor`: + `input_ids`: `torch.Tensor` of shape (bsz, max_len); + `attention_mask`: `torch.BoolTensor` of shape (bsz, max_len); + `labels`: `torch.Tensor` of shape (bsz, max_len), which contains `IGNORE_INDEX`. + """ + assert isinstance(self.tokenizer.pad_token_id, int) and self.tokenizer.pad_token_id >= 0, ( + f"`{self.tokenizer.__class__.__name__}.pad_token_id` must be a valid non-negative integer index value, " + f"but now `{self.tokenizer.pad_token_id}`" + ) + + # `List[torch.Tensor]` + batch_input_ids = [ + torch.LongTensor(instance["input_ids"][: self.max_length]) + if len(instance["input_ids"]) > self.max_length + else torch.LongTensor(instance["input_ids"]) + for instance in instances + ] + batch_labels = [ + torch.LongTensor(instance["labels"][: self.max_length]) + if len(instance["labels"]) > self.max_length + else torch.LongTensor(instance["labels"]) + for instance in instances + ] + if self.tokenizer.padding_side == "right": + input_ids = torch.nn.utils.rnn.pad_sequence( + sequences=batch_input_ids, + batch_first=True, + padding_value=self.tokenizer.pad_token_id, + ) # (bsz, max_len) + labels = torch.nn.utils.rnn.pad_sequence( + sequences=batch_labels, + batch_first=True, + padding_value=self.ignore_index, + ) # (bsz, max_len) + # pad to max + to_pad = self.max_length - input_ids.size(1) + input_ids = F.pad(input_ids, (0, to_pad), value=self.tokenizer.pad_token_id) + labels = F.pad(labels, (0, to_pad), value=self.ignore_index) + elif self.tokenizer.padding_side == "left": + reversed_input_ids = [seq.flip(dims=(0,)) for seq in batch_input_ids] + reversed_input_ids = torch.nn.utils.rnn.pad_sequence( + sequences=reversed_input_ids, + batch_first=True, + padding_value=self.tokenizer.pad_token_id, + ) # (bsz, max_len) + input_ids = torch.flip(reversed_input_ids, dims=(1,)) # (bsz, max_len) + reversed_labels = [seq.flip(dims=(0,)) for seq in batch_labels] + reversed_labels = torch.nn.utils.rnn.pad_sequence( + sequences=reversed_labels, + batch_first=True, + padding_value=self.ignore_index, + ) # (bsz, max_len) + labels = torch.flip(reversed_labels, dims=(1,)) # (bsz, max_len) + else: + raise RuntimeError( + f"`{self.tokenizer.__class__.__name__}.padding_side` can only be `left` or `right`, " + f"but now `{self.tokenizer.padding_side}`" + ) + + attention_mask = input_ids.ne(self.tokenizer.pad_token_id) # `torch.BoolTensor`, (bsz, max_len) + + return dict(input_ids=input_ids, attention_mask=attention_mask, labels=labels) + + +@dataclass +class DataCollatorForPromptDataset(DataCollatorForSupervisedDataset): + def __call__(self, instances: Sequence[Dict[str, List[int]]]) -> Dict[str, torch.Tensor]: + """ + + Args: + instances (`Sequence[Dict[str, List[int]]]`): + Mini-batch samples, each sample is stored in an individual dictionary. + + Returns: + (`Dict[str, torch.Tensor]`): Contains the following `torch.Tensor`: + `input_ids`: `torch.Tensor` of shape (bsz, max_len); + `attention_mask`: `torch.BoolTensor` of shape (bsz, max_len); + """ + instances = [{"input_ids": ins["input_ids"], "labels": ins["input_ids"]} for ins in instances] + ret = super().__call__(instances=instances) + input_ids = F.pad( + ret["input_ids"], (self.max_length - ret["input_ids"].size(1), 0), value=self.tokenizer.pad_token_id + ) + attention_mask = F.pad(ret["attention_mask"], (self.max_length - ret["attention_mask"].size(1), 0), value=False) + return {"input_ids": input_ids, "attention_mask": attention_mask} + + +@dataclass +class DataCollatorForPreferenceDataset(object): + """ + Collate instances for supervised dataset. + Each instance is a tokenized dictionary with fields + `input_ids`(List[int]), `labels`(List[int]) and `sequence`(str). + """ + + tokenizer: PreTrainedTokenizer + max_length: int = 4096 + + def __call__(self, instances: Sequence[Dict[str, List[int]]]) -> Dict[str, torch.Tensor]: + """ + + Args: + instances (`Sequence[Dict[str, List[int]]]`): + Mini-batch samples, each sample is stored in an individual dictionary. + + Returns: + (`Dict[str, torch.Tensor]`): Contains the following `torch.Tensor`: + `input_ids`: `torch.Tensor` of shape (bsz, max_len); + `attention_mask`: `torch.BoolTensor` of shape (bsz, max_len); + `labels`: `torch.Tensor` of shape (bsz, max_len), which contains `IGNORE_INDEX`. + """ + assert isinstance(self.tokenizer.pad_token_id, int) and self.tokenizer.pad_token_id >= 0, ( + f"`{self.tokenizer.__class__.__name__}.pad_token_id` must be a valid non-negative integer index value, " + f"but now `{self.tokenizer.pad_token_id}`" + ) + + ( + chosen_input_ids, + chosen_loss_mask, # [batch_size * seq_len] + reject_input_ids, + reject_loss_mask, + ) = ( + chuncate_sequence([ins["chosen_input_ids"] for ins in instances], self.max_length, torch.int64), + chuncate_sequence([ins["chosen_loss_mask"] for ins in instances], self.max_length, torch.bool), + chuncate_sequence([ins["rejected_input_ids"] for ins in instances], self.max_length, torch.int64), + chuncate_sequence([ins["rejected_loss_mask"] for ins in instances], self.max_length, torch.bool), + ) + + padding_side = self.tokenizer.padding_side + chosen_attention_mask = [torch.ones_like(seq).bool() for seq in chosen_input_ids] + reject_attention_mask = [torch.ones_like(seq).bool() for seq in reject_input_ids] + + ( + chosen_input_ids, + chosen_attention_mask, + chosen_loss_mask, + reject_input_ids, + reject_attention_mask, + reject_loss_mask, + ) = ( + pad_to_max_len(chosen_input_ids, self.max_length, self.tokenizer.pad_token_id, padding_side=padding_side), + pad_to_max_len(chosen_attention_mask, self.max_length, False, padding_side=padding_side), + pad_to_max_len(chosen_loss_mask, self.max_length, False, padding_side=padding_side), + pad_to_max_len(reject_input_ids, self.max_length, self.tokenizer.pad_token_id, padding_side=padding_side), + pad_to_max_len(reject_attention_mask, self.max_length, False, padding_side=padding_side), + pad_to_max_len(reject_loss_mask, self.max_length, False, padding_side=padding_side), + ) + + return dict( + chosen_input_ids=chosen_input_ids, + chosen_attention_mask=chosen_attention_mask, + chosen_loss_mask=chosen_loss_mask, + reject_input_ids=reject_input_ids, + reject_attention_mask=reject_attention_mask, + reject_loss_mask=reject_loss_mask, + ) + + +class StatefulDistributedSampler(DistributedSampler): + """ + Stateful distributed sampler for multi-stage training. + """ + + def __init__( + self, + dataset: DatasetType, + num_replicas: Optional[int] = None, + rank: Optional[int] = None, + shuffle: bool = True, + seed: int = 0, + drop_last: bool = False, + use_tp: Optional[bool] = False, + ) -> None: + if not use_tp: + super().__init__( + dataset=dataset, + num_replicas=num_replicas, + rank=rank, + shuffle=shuffle, + seed=seed, + drop_last=drop_last, + ) + else: + # adapted from https://github.com/pytorch/pytorch/blob/4979f9c0d72490970e2019bb1d2284f83d93f76b/torch/utils/data/distributed.py#L62 + # TODO: support tp_group>1. will fix it later + num_replicas = 1 + if rank is None: + rank = dist.get_rank() + if rank < 0: + raise ValueError(f"Invalid rank {rank}, rank should be in the interval [0, 0]") + self.dataset = dataset + self.num_replicas = num_replicas + self.rank = rank + self.epoch = 0 + self.drop_last = drop_last + # If the dataset length is evenly divisible by # of replicas, then there + # is no need to drop any data, since the dataset will be split equally. + if self.drop_last and len(self.dataset) % self.num_replicas != 0: # type: ignore[arg-type] + # Split to nearest available length that is evenly divisible. + # This is to ensure each rank receives the same amount of data when + # using this Sampler. + self.num_samples = math.ceil( + (len(self.dataset) - self.num_replicas) / self.num_replicas # type: ignore[arg-type] + ) + else: + self.num_samples = math.ceil(len(self.dataset) / self.num_replicas) # type: ignore[arg-type] + self.total_size = self.num_samples * self.num_replicas + self.shuffle = shuffle + self.seed = seed + self.start_index = 0 + self.use_tp = use_tp + + def __iter__(self) -> Iterator: + if self.use_tp: + # TODO Add support for tp_group not equal to 1 + pass + # adpated from https://github.com/pytorch/pytorch/blob/4979f9c0d72490970e2019bb1d2284f83d93f76b/torch/utils/data/distributed.py#L96 + if self.shuffle: + # deterministically shuffle based on epoch and seed + g = torch.Generator() + g.manual_seed(self.seed + self.epoch) + indices = torch.randperm(len(self.dataset), generator=g).tolist() # type: ignore[arg-type] + else: + indices = list(range(len(self.dataset))) # type: ignore[arg-type] + + if not self.drop_last: + # add extra samples to make it evenly divisible + padding_size = self.total_size - len(indices) + if padding_size <= len(indices): + indices += indices[:padding_size] + else: + indices += (indices * math.ceil(padding_size / len(indices)))[:padding_size] + else: + # remove tail of data to make it evenly divisible. + indices = indices[: self.total_size] + assert len(indices) == self.total_size + + # subsample + indices = indices[ + : self.total_size : self.num_replicas + ] # num_replicas=tp_group=1, we only support tp_group==1 for now + assert len(indices) == self.num_samples + + return iter(indices) + + else: + iterator = super().__iter__() + indices = list(iterator) + indices = indices[self.start_index :] + return iter(indices) + + def __len__(self) -> int: + return self.num_samples - self.start_index + + def set_start_index(self, start_index: int) -> None: + self.start_index = start_index + + +def setup_distributed_dataloader( + dataset: DatasetType, + batch_size: int = 1, + shuffle: bool = False, + seed: int = 1024, + drop_last: bool = False, + pin_memory: bool = False, + num_workers: int = 0, + collate_fn: Callable[[Sequence[Dict[str, Union[str, List[int]]]]], Dict[str, torch.Tensor]] = None, + process_group: Optional[ProcessGroup] = None, + use_tp: Optional[bool] = False, + **kwargs, +) -> DataLoader: + """ + Setup dataloader for distributed training. + """ + _kwargs = kwargs.copy() + process_group = process_group or _get_default_group() + sampler = StatefulDistributedSampler( + dataset=dataset, + num_replicas=process_group.size() if not use_tp else 1, + rank=process_group.rank(), + shuffle=shuffle, + seed=seed, + drop_last=drop_last, + use_tp=use_tp, + ) + + # Deterministic dataloader + def seed_worker(worker_id: int) -> None: + worker_seed = seed + np.random.seed(worker_seed) + torch.manual_seed(worker_seed) + random.seed(worker_seed) + + return DataLoader( + dataset=dataset, + batch_size=batch_size, + sampler=sampler, + num_workers=num_workers, + collate_fn=collate_fn, + pin_memory=pin_memory, + drop_last=drop_last, + worker_init_fn=seed_worker, + **_kwargs, + ) diff --git a/applications/ColossalChat/coati/dataset/tokenization_utils.py b/applications/ColossalChat/coati/dataset/tokenization_utils.py new file mode 100755 index 000000000000..7606bc2a97ba --- /dev/null +++ b/applications/ColossalChat/coati/dataset/tokenization_utils.py @@ -0,0 +1,383 @@ +#!/usr/bin/env python3 +# -*- coding: utf-8 -*- +""" +tokenization utils for constructing dataset for ppo, dpo, sft, rm +""" + +import warnings +from copy import deepcopy +from typing import Any, Dict, List, Union + +from coati.dataset.conversation import Conversation +from coati.dataset.utils import split_templated_prompt_into_chunks, tokenize_and_concatenate +from datasets import dataset_dict +from torch.utils.data import ConcatDataset, Dataset +from transformers import PreTrainedTokenizer + +from colossalai.logging import get_dist_logger + +logger = get_dist_logger() + +IGNORE_INDEX = -100 + +DSType = Union[Dataset, ConcatDataset, dataset_dict.Dataset] + + +def supervised_tokenize_sft( + data_point: Dict[str, str], + tokenizer: PreTrainedTokenizer, + conversation_template: Conversation = None, + ignore_index: int = None, + max_length: int = 4096, +) -> Dict[str, Union[int, str, List[int]]]: + """ + A tokenization function to tokenize an original pretraining data point as following + and calculate corresponding labels for sft training: + "Something here can be system message[user_line_start]User line[User line end][Assistant line start]Assistant line[Assistant line end]...[Assistant line end]Something here" + ^ + end_of_system_line_position + + Args: + data_point: the data point of the following format + {"messages": [{"from": "human", "content": "xxx"}, {"from": "assistant", "content": "xxx"}]} + tokenizer: the tokenizer whose + conversation_template: the conversation template to apply + ignore_index: the ignore index when calculate loss during training + max_length: the maximum context length + """ + + if ignore_index is None: + ignore_index = IGNORE_INDEX + + messages = data_point["messages"] + template = deepcopy(conversation_template) + template.messages = [] + + for mess in messages: + from_str = mess["from"] + if from_str.lower() == "human": + from_str = "user" + elif from_str.lower() == "assistant": + from_str = "assistant" + else: + raise ValueError(f"Unsupported role {from_str.lower()}") + + template.append_message(from_str, mess["content"]) + + if len(template.messages) % 2 != 0: + template.messages = template.messages[0:-1] + + # `target_turn_index` is the number of turns which exceeds `max_length - 1` for the first time. + turns = [i for i in range(1, len(messages) // 2 + 1)] + + lo, hi = 0, len(turns) + while lo < hi: + mid = (lo + hi) // 2 + if max_length - 1 < len( + tokenizer([template.get_prompt(2 * turns[mid] - 1)], add_special_tokens=False)["input_ids"][0] + ): + hi = mid + else: + lo = mid + 1 + target_turn_index = lo + + # The tokenized length for first turn already exceeds `max_length - 1`. + if target_turn_index - 1 < 0: + warnings.warn("The tokenized length for first turn already exceeds `max_length - 1`.") + return dict( + input_ids=None, + labels=None, + inputs_decode=None, + labels_decode=None, + seq_length=None, + seq_category=None, + ) + + target_turn = turns[target_turn_index - 1] + prompt = template.get_prompt(2 * target_turn) + chunks, require_loss = split_templated_prompt_into_chunks(template.messages[: 2 * target_turn], prompt) + tokenized, starts, ends = tokenize_and_concatenate(tokenizer, chunks, require_loss) + + labels = [ignore_index] * len(tokenized) + label_decode = [] + for start, end in zip(starts, ends): + if end == len(tokenized): + tokenized = tokenized + [tokenizer.eos_token_id] + labels = labels + [ignore_index] + labels[start : end + 1] = tokenized[start : end + 1] + label_decode.append(tokenizer.decode(tokenized[start : end + 1], skip_special_tokens=False)) + + if tokenizer.bos_token_id is not None: + if tokenized[0] != tokenizer.bos_token_id: + tokenized = [tokenizer.bos_token_id] + tokenized + labels = [ignore_index] + labels + + if tokenizer.eos_token_id is not None: + # Force to add eos token at the end of the tokenized sequence + if tokenized[-1] != tokenizer.eos_token_id: + tokenized = tokenized + [tokenizer.eos_token_id] + labels = labels + [tokenizer.eos_token_id] + else: + labels[-1] = tokenizer.eos_token_id + + # For some model without bos/eos may raise the following errors + try: + inputs_decode = tokenizer.decode(tokenized) + except TypeError as e: + raise TypeError(str(e) + f"\nUnable to decode input_ids: {tokenized}") + + # Check if all labels are ignored, this may happen when the tokenized length is too long + if labels.count(ignore_index) == len(labels): + return dict( + input_ids=None, + labels=None, + inputs_decode=None, + labels_decode=None, + seq_length=None, + seq_category=None, + ) + + return dict( + input_ids=tokenized, + labels=labels, + inputs_decode=inputs_decode, + labels_decode=label_decode, + seq_length=len(tokenized), + seq_category=data_point["category"] if "category" in data_point else "None", + ) + + +def tokenize_prompt_dataset( + data_point: Dict[str, str], + tokenizer: PreTrainedTokenizer, + conversation_template: Conversation = None, + ignore_index: int = None, + max_length: int = 4096, +) -> Dict[str, Union[int, str, List[int]]]: + """ + A tokenization function to tokenize an original pretraining data point as following for ppo training: + "Something here can be system message[user_line_start]User line[User line end][Assistant line start]Assistant line[Assistant line end]...[Assistant line start]" + Args: + data_point: the data point of the following format + {"messages": [{"from": "human", "content": "xxx"}, {"from": "assistant", "content": "xxx"}]} + tokenizer: the tokenizer whose + conversation_template: the conversation template to apply + ignore_index: the ignore index when calculate loss during training + max_length: the maximum context length + """ + if ignore_index is None: + ignore_index = IGNORE_INDEX + + messages = data_point["messages"] + template = deepcopy(conversation_template) + template.messages = [] + + for mess in messages: + from_str = mess["from"] + if from_str.lower() == "human": + from_str = "user" + elif from_str.lower() == "assistant": + from_str = "assistant" + else: + raise ValueError(f"Unsupported role {from_str.lower()}") + + template.append_message(from_str, mess["content"]) + + # `target_turn_index` is the number of turns which exceeds `max_length - 1` for the first time. + target_turn = len(template.messages) + if target_turn % 2 != 1: + # exclude the answer if provided. keep only the prompt + target_turn = target_turn - 1 + + # Prepare data + prompt = template.get_prompt(target_turn, add_generation_prompt=True) + tokenized = tokenizer([prompt], add_special_tokens=False)["input_ids"][0] + if tokenizer.bos_token_id is not None: + if tokenized[0] != tokenizer.bos_token_id: + tokenized = [tokenizer.bos_token_id] + tokenized + + # Skip overlength data + if max_length - 1 < len(tokenized): + return dict( + input_ids=None, + inputs_decode=None, + seq_length=None, + seq_category=None, + ) + + # `inputs_decode` can be used to check whether the tokenization method is true. + return dict( + input_ids=tokenized, + inputs_decode=tokenizer.decode(tokenized), + seq_length=len(tokenized), + seq_category=data_point["category"] if "category" in data_point else "None", + ) + + +def apply_rlhf_data_format( + template: Conversation, tokenizer: Any, context_len: int, mask_out_target_assistant_line_end=False +): + target_turn = int(len(template.messages) / 2) + prompt = template.get_prompt(target_turn * 2) + chunks, require_loss = split_templated_prompt_into_chunks(template.messages[: 2 * target_turn], prompt) + tokenized, starts, ends = tokenize_and_concatenate(tokenizer, chunks, require_loss) + loss_mask = [0] * len(tokenized) + mask_token = tokenizer.eos_token_id or tokenizer.pad_token_id + if mask_token is None: + mask_token = 1 # If the tokenizer doesn't have eos_token or pad_token: Qwen + + label_decode = [] + for start, end in zip(starts[-1:], ends[-1:]): + # only the last round (chosen/rejected) counts + if end == len(tokenized): + tokenized = tokenized + [tokenizer.eos_token_id] + loss_mask = loss_mask + [1] + loss_mask[start : end + 1] = [1] * len(loss_mask[start : end + 1]) + label_decode.append(tokenizer.decode(tokenized[start : end + 1], skip_special_tokens=False)) + if tokenizer.bos_token_id is not None: + if tokenized[0] != tokenizer.bos_token_id: + tokenized = [tokenizer.bos_token_id] + tokenized + loss_mask = [0] + loss_mask + + if tokenizer.eos_token_id is not None: + # Force to add eos token at the end of the tokenized sequence + if tokenized[-1] != tokenizer.eos_token_id: + tokenized = tokenized + [tokenizer.eos_token_id] + loss_mask = loss_mask + [1] + else: + loss_mask[-1] = 1 + + return {"input_ids": tokenized, "loss_mask": loss_mask, "label_decode": label_decode} + + +def tokenize_rlhf( + data_point: Dict[str, str], + tokenizer: PreTrainedTokenizer, + conversation_template: Conversation = None, + ignore_index: int = None, + max_length: int = 4096, +) -> Dict[str, Union[int, str, List[int]]]: + """ + A tokenization function to tokenize an original pretraining data point as following: + {"context": [{"from": "human", "content": "xxx"}, {"from": "assistant", "content": "xxx"}], + "chosen": {"from": "assistant", "content": "xxx"}, "rejected": {"from": "assistant", "content": "xxx"}} + """ + if ignore_index is None: + ignore_index = IGNORE_INDEX + + context = data_point["context"] + template = deepcopy(conversation_template) + template.clear() + + for mess in context: + from_str = mess["from"] + if from_str.lower() == "human": + from_str = "user" + elif from_str.lower() == "assistant": + from_str = "assistant" + else: + raise ValueError(f"Unsupported role {from_str.lower()}") + + if len(template.messages) > 0 and from_str == template.messages[-1]["role"]: + # Concate adjacent message from the same role + template.messages[-1]["content"] = str(template.messages[-1]["content"] + " " + mess["content"]) + else: + template.append_message(from_str, mess["content"]) + + if len(template.messages) % 2 != 1: + warnings.warn( + "Please make sure leading context starts and ends with a line from human\nLeading context: " + + str(template.messages) + ) + return dict( + chosen_input_ids=None, + chosen_loss_mask=None, + chosen_label_decode=None, + rejected_input_ids=None, + rejected_loss_mask=None, + rejected_label_decode=None, + ) + round_of_context = int((len(template.messages) - 1) / 2) + + assert context[-1]["from"].lower() == "human", "The last message in context should be from human." + chosen = deepcopy(template) + rejected = deepcopy(template) + + for round in range(len(data_point["chosen"])): + from_str = data_point["chosen"][round]["from"] + if from_str.lower() == "human": + from_str = "user" + elif from_str.lower() == "assistant": + from_str = "assistant" + else: + raise ValueError(f"Unsupported role {from_str.lower()}") + chosen.append_message(from_str, data_point["chosen"][round]["content"]) + + for round in range(len(data_point["rejected"])): + from_str = data_point["rejected"][round]["from"] + if from_str.lower() == "human": + from_str = "user" + elif from_str.lower() == "assistant": + from_str = "assistant" + else: + raise ValueError(f"Unsupported role {from_str.lower()}") + rejected.append_message(from_str, data_point["rejected"][round]["content"]) + + ( + chosen_input_ids, + chosen_loss_mask, + chosen_label_decode, + rejected_input_ids, + rejected_loss_mask, + rejected_label_decode, + ) = (None, None, None, None, None, None) + if ( + len(tokenizer([chosen.get_prompt(len(chosen.messages))], add_special_tokens=False)["input_ids"][0]) + <= max_length - 1 + and len(tokenizer([rejected.get_prompt(len(rejected.messages))], add_special_tokens=False)["input_ids"][0]) + <= max_length - 1 + ): + chosen_data_packed = apply_rlhf_data_format(chosen, tokenizer, round_of_context) + (chosen_input_ids, chosen_loss_mask, chosen_label_decode) = ( + chosen_data_packed["input_ids"], + chosen_data_packed["loss_mask"], + chosen_data_packed["label_decode"], + ) + + rejected_data_packed = apply_rlhf_data_format( + rejected, tokenizer, round_of_context, mask_out_target_assistant_line_end=True + ) + (rejected_input_ids, rejected_loss_mask, rejected_label_decode) = ( + rejected_data_packed["input_ids"], + rejected_data_packed["loss_mask"], + rejected_data_packed["label_decode"], + ) + + # Check if loss mask is all 0s (no loss), this may happen when the tokenized length is too long + if chosen_loss_mask.count(0) == len(chosen_loss_mask) or rejected_loss_mask.count(0) == len(rejected_loss_mask): + return dict( + chosen_input_ids=None, + chosen_loss_mask=None, + chosen_label_decode=None, + rejected_input_ids=None, + rejected_loss_mask=None, + rejected_label_decode=None, + ) + + return { + "chosen_input_ids": chosen_input_ids, + "chosen_loss_mask": chosen_loss_mask, + "chosen_label_decode": chosen_label_decode, + "rejected_input_ids": rejected_input_ids, + "rejected_loss_mask": rejected_loss_mask, + "rejected_label_decode": rejected_label_decode, + } + else: + return dict( + chosen_input_ids=None, + chosen_loss_mask=None, + chosen_label_decode=None, + rejected_input_ids=None, + rejected_loss_mask=None, + rejected_label_decode=None, + ) diff --git a/applications/ColossalChat/coati/dataset/utils.py b/applications/ColossalChat/coati/dataset/utils.py new file mode 100755 index 000000000000..ada2afef0154 --- /dev/null +++ b/applications/ColossalChat/coati/dataset/utils.py @@ -0,0 +1,138 @@ +import io +import json +from typing import Any, Dict, List + +import torch +import torch.distributed as dist +import torch.nn.functional as F +from transformers import PreTrainedTokenizer + + +def is_rank_0() -> bool: + return not dist.is_initialized() or dist.get_rank() == 0 + + +def _make_r_io_base(f, mode: str): + if not isinstance(f, io.IOBase): + f = open(f, mode=mode) + return f + + +def jload(f, mode="r"): + """Load a .json file into a dictionary.""" + f = _make_r_io_base(f, mode) + jdict = json.load(f) + f.close() + return jdict + + +def read_string_by_schema(data: Dict[str, Any], schema: str) -> str: + """ + Read a feild of the dataset be schema + Args: + data: Dict[str, Any] + schema: cascaded feild names seperated by '.'. e.g. person.name.first will access data['person']['name']['first'] + """ + keys = schema.split(".") + result = data + for key in keys: + result = result.get(key, None) + if result is None: + return "" + assert isinstance(result, str), f"dataset element is not a string: {result}" + return result + + +def pad_to_max_len( + sequence: List[torch.Tensor], max_length: int, padding_value: int, batch_first: bool = True, padding_side="left" +): + """ + Args: + sequence: a batch of tensor of shape [batch_size, seq_len] if batch_first==True + """ + if padding_side == "left": + reversed_sequence = [seq.flip(dims=(0,)) for seq in sequence] + padded = torch.nn.utils.rnn.pad_sequence( + sequences=reversed_sequence, batch_first=batch_first, padding_value=padding_value + ) + to_pad = max_length - padded.size(1) + padded = F.pad(padded, (0, to_pad), value=padding_value) + return torch.flip(padded, dims=(1,)) + elif padding_side == "right": + padded = torch.nn.utils.rnn.pad_sequence( + sequences=sequence, batch_first=batch_first, padding_value=padding_value + ) + to_pad = max_length - padded.size(1) + return F.pad(padded, (0, to_pad), value=padding_value) + else: + raise RuntimeError(f"`padding_side` can only be `left` or `right`, " f"but now `{padding_side}`") + + +def chuncate_sequence(sequence: List[torch.Tensor], max_length: int, dtype: Any): + """ + Args: + sequence: a batch of tensor of shape [batch_size, seq_len] if batch_first==True + """ + return [ + torch.Tensor(seq[:max_length]).to(dtype) if len(seq) > max_length else torch.Tensor(seq).to(dtype) + for seq in sequence + ] + + +def find_first_occurrence_subsequence(seq: torch.Tensor, subseq: torch.Tensor, start_index: int = 0) -> int: + if subseq is None: + return 0 + for i in range(start_index, len(seq) - len(subseq) + 1): + if torch.all(seq[i : i + len(subseq)] == subseq): + return i + return -1 + + +def tokenize_and_concatenate(tokenizer: PreTrainedTokenizer, text: List[str], require_loss: List[bool]): + """ + Tokenizes a list of texts using the provided tokenizer and concatenates the tokenized outputs. + + Args: + tokenizer (PreTrainedTokenizer): The tokenizer to use for tokenization. + text (List[str]): The list of texts to tokenize. + require_loss (List[bool]): A list of boolean values indicating whether each text requires loss calculation. + + Returns: + Tuple[List[int], List[int], List[int]]: A tuple containing the concatenated tokenized input ids, + the start positions of loss spans, and the end positions of loss spans. + """ + input_ids = [] + loss_starts = [] + loss_ends = [] + for s, r in zip(text, require_loss): + tokenized = tokenizer(s, add_special_tokens=False)["input_ids"] + if r: + loss_starts.append(len(input_ids)) + loss_ends.append(len(input_ids) + len(tokenized)) + input_ids.extend(tokenized) + return input_ids, loss_starts, loss_ends + + +def split_templated_prompt_into_chunks(messages: List[Dict[str, str]], prompt: str): + # Seperate templated prompt into chunks by human/assistant's lines, prepare data for tokenize_and_concatenate + start_idx = 0 + chunks = [] + require_loss = [] + for line in messages: + first_occur = prompt.find(line["content"], start_idx) + if prompt[first_occur - 1] != " ": + chunks.append(prompt[start_idx:first_occur]) + chunks.append(prompt[first_occur : first_occur + len(line["content"])]) + else: + chunks.append(prompt[start_idx : first_occur - 1]) + chunks.append(prompt[first_occur - 1 : first_occur + len(line["content"])]) + start_idx = first_occur + len(line["content"]) + if line["role"].lower() == "assistant": + require_loss.append(False) + require_loss.append(True) + else: + require_loss.append(False) + require_loss.append(False) + chunks.append(prompt[start_idx:]) + require_loss.append(False) + return chunks, require_loss diff --git a/applications/Chat/coati/experience_buffer/__init__.py b/applications/ColossalChat/coati/experience_buffer/__init__.py old mode 100644 new mode 100755 similarity index 100% rename from applications/Chat/coati/experience_buffer/__init__.py rename to applications/ColossalChat/coati/experience_buffer/__init__.py diff --git a/applications/Chat/coati/experience_buffer/base.py b/applications/ColossalChat/coati/experience_buffer/base.py old mode 100644 new mode 100755 similarity index 100% rename from applications/Chat/coati/experience_buffer/base.py rename to applications/ColossalChat/coati/experience_buffer/base.py diff --git a/applications/Chat/coati/experience_buffer/naive.py b/applications/ColossalChat/coati/experience_buffer/naive.py old mode 100644 new mode 100755 similarity index 86% rename from applications/Chat/coati/experience_buffer/naive.py rename to applications/ColossalChat/coati/experience_buffer/naive.py index d47b67dbe713..b912df26818d --- a/applications/Chat/coati/experience_buffer/naive.py +++ b/applications/ColossalChat/coati/experience_buffer/naive.py @@ -1,13 +1,16 @@ import random -import warnings from typing import List import torch from coati.experience_maker.base import Experience +from colossalai.logging import get_dist_logger + from .base import ExperienceBuffer from .utils import BufferItem, make_experience_batch, split_experience_batch +logger = get_dist_logger() + class NaiveExperienceBuffer(ExperienceBuffer): """Naive experience buffer class. It stores experience. @@ -35,7 +38,7 @@ def append(self, experience: Experience) -> None: if self.limit > 0: samples_to_remove = len(self.items) - self.limit if samples_to_remove > 0: - warnings.warn(f"Experience buffer is full. Removing {samples_to_remove} samples.") + logger.warning(f"Experience buffer is full. Removing {samples_to_remove} samples.") self.items = self.items[samples_to_remove:] def clear(self) -> None: @@ -43,6 +46,12 @@ def clear(self) -> None: @torch.no_grad() def sample(self) -> Experience: + """ + Randomly samples experiences from the buffer. + + Returns: + A batch of sampled experiences. + """ items = random.sample(self.items, self.sample_batch_size) experience = make_experience_batch(items) if self.cpu_offload: diff --git a/applications/Chat/coati/experience_buffer/utils.py b/applications/ColossalChat/coati/experience_buffer/utils.py old mode 100644 new mode 100755 similarity index 94% rename from applications/Chat/coati/experience_buffer/utils.py rename to applications/ColossalChat/coati/experience_buffer/utils.py index baedbebd184f..c4807d179d90 --- a/applications/Chat/coati/experience_buffer/utils.py +++ b/applications/ColossalChat/coati/experience_buffer/utils.py @@ -26,6 +26,7 @@ class BufferItem: action_log_probs: torch.Tensor values: torch.Tensor reward: torch.Tensor + kl: torch.Tensor advantages: torch.Tensor attention_mask: Optional[torch.LongTensor] action_mask: Optional[torch.BoolTensor] @@ -34,7 +35,7 @@ class BufferItem: def split_experience_batch(experience: Experience) -> List[BufferItem]: batch_size = experience.sequences.size(0) batch_kwargs = [{} for _ in range(batch_size)] - keys = ("sequences", "action_log_probs", "values", "reward", "advantages", "attention_mask", "action_mask") + keys = ("sequences", "action_log_probs", "values", "reward", "kl", "advantages", "attention_mask", "action_mask") for key in keys: value = getattr(experience, key) if isinstance(value, torch.Tensor): @@ -63,7 +64,7 @@ def _zero_pad_sequences(sequences: List[torch.Tensor], side: str = "left") -> to def make_experience_batch(items: List[BufferItem]) -> Experience: kwargs = {} to_pad_keys = set(("action_log_probs", "action_mask")) - keys = ("sequences", "action_log_probs", "values", "reward", "advantages", "attention_mask", "action_mask") + keys = ("sequences", "action_log_probs", "values", "reward", "kl", "advantages", "attention_mask", "action_mask") for key in keys: vals = [getattr(item, key) for item in items] if key in to_pad_keys: diff --git a/applications/Chat/coati/experience_maker/__init__.py b/applications/ColossalChat/coati/experience_maker/__init__.py old mode 100644 new mode 100755 similarity index 100% rename from applications/Chat/coati/experience_maker/__init__.py rename to applications/ColossalChat/coati/experience_maker/__init__.py diff --git a/applications/Chat/coati/experience_maker/base.py b/applications/ColossalChat/coati/experience_maker/base.py old mode 100644 new mode 100755 similarity index 74% rename from applications/Chat/coati/experience_maker/base.py rename to applications/ColossalChat/coati/experience_maker/base.py index 0731f6e0f97f..55054c3a0611 --- a/applications/Chat/coati/experience_maker/base.py +++ b/applications/ColossalChat/coati/experience_maker/base.py @@ -3,7 +3,8 @@ from typing import Optional import torch -from coati.models.base import Actor, Critic, RewardModel +from coati.models import Critic, RewardModel +from transformers import PreTrainedModel @dataclass @@ -28,6 +29,7 @@ class Experience: action_log_probs: torch.Tensor values: torch.Tensor reward: torch.Tensor + kl: torch.Tensor advantages: torch.Tensor attention_mask: Optional[torch.LongTensor] action_mask: Optional[torch.BoolTensor] @@ -39,6 +41,7 @@ def to_device(self, device: torch.device) -> None: self.values = self.values.to(device) self.reward = self.reward.to(device) self.advantages = self.advantages.to(device) + self.kl = self.kl.to(device) if self.attention_mask is not None: self.attention_mask = self.attention_mask.to(device) if self.action_mask is not None: @@ -50,6 +53,7 @@ def pin_memory(self): self.values = self.values.pin_memory() self.reward = self.reward.pin_memory() self.advantages = self.advantages.pin_memory() + self.kl = self.kl.pin_memory() if self.attention_mask is not None: self.attention_mask = self.attention_mask.pin_memory() if self.action_mask is not None: @@ -58,7 +62,13 @@ def pin_memory(self): class ExperienceMaker(ABC): - def __init__(self, actor: Actor, critic: Critic, reward_model: RewardModel, initial_model: Actor) -> None: + """ + Base class for experience makers. + """ + + def __init__( + self, actor: PreTrainedModel, critic: Critic, reward_model: RewardModel, initial_model: PreTrainedModel + ) -> None: super().__init__() self.actor = actor self.critic = critic @@ -67,4 +77,14 @@ def __init__(self, actor: Actor, critic: Critic, reward_model: RewardModel, init @abstractmethod def make_experience(self, input_ids: torch.Tensor, attention_mask: torch.Tensor, **generate_kwargs) -> Experience: - pass + """ + Abstract method to generate an experience. + + Args: + input_ids (torch.Tensor): The input tensor. + attention_mask (torch.Tensor): The attention mask tensor. + **generate_kwargs: Additional keyword arguments for generating the experience. + + Returns: + Experience: The generated experience. + """ diff --git a/applications/ColossalChat/coati/experience_maker/naive.py b/applications/ColossalChat/coati/experience_maker/naive.py new file mode 100755 index 000000000000..945bb95577c7 --- /dev/null +++ b/applications/ColossalChat/coati/experience_maker/naive.py @@ -0,0 +1,180 @@ +""" +experience maker. +""" + +import torch +import torch.nn.functional as F +from coati.dataset.utils import find_first_occurrence_subsequence +from coati.models import Critic, RewardModel +from coati.models.generation import generate +from coati.models.utils import calc_action_log_probs, compute_reward +from transformers import PreTrainedModel, PreTrainedTokenizer + +from colossalai.logging import get_dist_logger + +from .base import Experience, ExperienceMaker + +logger = get_dist_logger() + +import torch.distributed as dist + + +def is_rank_0() -> bool: + return not dist.is_initialized() or dist.get_rank() == 0 + + +class NaiveExperienceMaker(ExperienceMaker): + """ + Naive experience maker. + """ + + def __init__( + self, + actor: PreTrainedModel, + critic: Critic, + reward_model: RewardModel, + initial_model: PreTrainedModel, + tokenizer: PreTrainedTokenizer, + kl_coef: float = 0.01, + gamma: float = 1.0, + lam: float = 0.95, + ) -> None: + super().__init__(actor, critic, reward_model, initial_model) + self.tokenizer = tokenizer + self.kl_coef = kl_coef + self.gamma = gamma + self.lam = lam + + @torch.no_grad() + def calculate_advantage(self, value: torch.Tensor, reward: torch.Tensor, num_actions: int) -> torch.Tensor: + """ + Calculates the advantage values for each action based on the value and reward tensors. + + Args: + value (torch.Tensor): Tensor containing the predicted values from critic. + reward (torch.Tensor): reward of the shape [B, len]. + num_actions (int): Number of actions. + + Returns: + torch.Tensor: Tensor containing the calculated advantages for each action. + """ + lastgaelam = 0 + advantages_reversed = [] + for t in reversed(range(num_actions)): + nextvalues = value[:, t + 1] if t < num_actions - 1 else 0.0 + delta = reward[:, t] + self.gamma * nextvalues - value[:, t] + lastgaelam = delta + self.gamma * self.lam * lastgaelam + advantages_reversed.append(lastgaelam) + advantages = torch.stack(advantages_reversed[::-1], dim=1) + return advantages + + @torch.no_grad() + def make_experience(self, input_ids: torch.Tensor, attention_mask: torch.Tensor, **generate_kwargs) -> Experience: + """ + Generates an experience using the given input_ids and attention_mask. + + Args: + input_ids (torch.Tensor): The input tensor containing the tokenized input sequence. + attention_mask (torch.Tensor): The attention mask tensor indicating which tokens to attend to. + **generate_kwargs: Additional keyword arguments for the generation process. + + Returns: + Experience: The generated experience object. + + """ + self.actor.eval() + self.critic.eval() + self.initial_model.eval() + self.reward_model.eval() + pad_token_id = self.tokenizer.pad_token_id + + stop_token_ids = generate_kwargs.get("stop_token_ids", None) + torch.manual_seed(41) # for tp, gurantee the same input for reward model + + sequences = generate(self.actor, input_ids, self.tokenizer, **generate_kwargs) + + # Pad to max length + sequences = F.pad(sequences, (0, generate_kwargs["max_length"] - sequences.size(1)), value=pad_token_id) + sequence_length = sequences.size(1) + + # Calculate auxiliary tensors + attention_mask = None + if pad_token_id is not None: + attention_mask = sequences.not_equal(pad_token_id).to(dtype=torch.long, device=sequences.device) + + input_len = input_ids.size(1) + if stop_token_ids is None: + # End the sequence with eos token + eos_token_id = self.tokenizer.eos_token_id + if eos_token_id is None: + action_mask = torch.ones_like(sequences, dtype=torch.bool) + else: + # Left padding may be applied, only mask action + action_mask = (sequences[:, input_len:] == eos_token_id).cumsum(dim=-1) == 0 + action_mask = F.pad(action_mask, (1 + input_len, -1), value=True) # include eos token and input + else: + # stop_token_ids are given, generation ends with stop_token_ids + action_mask = torch.ones_like(sequences, dtype=torch.bool) + for i in range(sequences.size(0)): + stop_index = find_first_occurrence_subsequence( + sequences[i][input_len:], torch.tensor(stop_token_ids).to(sequences.device) + ) + if stop_index == -1: + # Sequence does not contain stop_token_ids, this should never happen BTW + logger.warning( + "Generated sequence does not contain stop_token_ids. Please check your chat template config" + ) + else: + # Keep stop tokens + stop_index = input_len + stop_index + action_mask[i, stop_index + len(stop_token_ids) :] = False + + generation_end_index = (action_mask == True).sum(dim=-1) - 1 + action_mask[:, :input_len] = False + action_mask = action_mask[:, 1:] + action_mask = action_mask[:, -(sequences.size(1) - input_len) :] + num_actions = action_mask.size(1) + + actor_output = self.actor(input_ids=sequences, attention_mask=attention_mask)["logits"] + action_log_probs = calc_action_log_probs(actor_output, sequences, num_actions) + + base_model_output = self.initial_model(input_ids=sequences, attention_mask=attention_mask)["logits"] + + base_action_log_probs = calc_action_log_probs(base_model_output, sequences, num_actions) + + # Convert to right padding for the reward model and the critic model + input_ids_rm = torch.zeros_like(sequences, device=sequences.device) + attention_mask_rm = torch.zeros_like(sequences, device=sequences.device) + for i in range(sequences.size(0)): + sequence = sequences[i] + bos_index = (sequence != pad_token_id).nonzero().reshape([-1])[0] + eos_index = generation_end_index[i] + sequence_to_pad = sequence[bos_index:eos_index] + sequence_padded = F.pad( + sequence_to_pad, (0, sequence_length - sequence_to_pad.size(0)), value=self.tokenizer.pad_token_id + ) + input_ids_rm[i] = sequence_padded + if sequence_length - sequence_to_pad.size(0) > 0: + attention_mask_rm[i, : sequence_to_pad.size(0) + 1] = 1 + else: + attention_mask_rm[i, :] = 1 + attention_mask_rm = attention_mask_rm.to(dtype=torch.bool) + + r = self.reward_model( + input_ids=input_ids_rm.to(dtype=torch.long, device=sequences.device), + attention_mask=attention_mask_rm.to(device=sequences.device), + ) + + value = self.critic( + input_ids=input_ids_rm.to(dtype=torch.long, device=sequences.device), + attention_mask=attention_mask_rm.to(device=sequences.device), + ) + reward, kl = compute_reward(r, self.kl_coef, action_log_probs, base_action_log_probs, action_mask=action_mask) + value = value[:, -num_actions:] * action_mask + advantages = self.calculate_advantage(value, reward, num_actions) + + advantages = advantages.detach() + value = value.detach() + r = r.detach() + + return Experience(sequences, action_log_probs, value, r, kl, advantages, attention_mask, action_mask) diff --git a/applications/ColossalChat/coati/models/__init__.py b/applications/ColossalChat/coati/models/__init__.py new file mode 100755 index 000000000000..14073207f150 --- /dev/null +++ b/applications/ColossalChat/coati/models/__init__.py @@ -0,0 +1,24 @@ +from .base import BaseModel +from .critic import Critic +from .generation import generate, generate_streaming, prepare_inputs_fn, update_model_kwargs_fn +from .lora import convert_to_lora_module +from .loss import DpoLoss, LogExpLoss, LogSigLoss, PolicyLoss, ValueLoss +from .reward_model import RewardModel +from .utils import disable_dropout + +__all__ = [ + "BaseModel", + "Critic", + "RewardModel", + "PolicyLoss", + "ValueLoss", + "LogSigLoss", + "LogExpLoss", + "convert_to_lora_module", + "DpoLoss", + "generate", + "generate_streaming", + "disable_dropout", + "update_model_kwargs_fn", + "prepare_inputs_fn", +] diff --git a/applications/ColossalChat/coati/models/base.py b/applications/ColossalChat/coati/models/base.py new file mode 100755 index 000000000000..fcea9414b430 --- /dev/null +++ b/applications/ColossalChat/coati/models/base.py @@ -0,0 +1,58 @@ +""" +Base class for critic and reward model +""" + +from typing import Optional + +import torch +import torch.nn as nn +from transformers import AutoModel, PretrainedConfig + + +class BaseModel(nn.Module): + """ + Actor model base class. + + Args: + pretrained (str): path to pretrained model. + config (PretrainedConfig): PretrainedConfig used to initiate the base model. + **kwargs: all other kwargs as in AutoModel.from_pretrained + """ + + def __init__(self, pretrained: str = None, config: Optional[PretrainedConfig] = None, **kwargs) -> None: + super().__init__() + if pretrained is not None: + if config is not None: + # initialize with config and load weights from pretrained + self.model = AutoModel.from_pretrained(pretrained, config=config, **kwargs) + else: + # initialize with pretrained + self.model = AutoModel.from_pretrained(pretrained, **kwargs) + elif config is not None: + # initialize with config + self.model = AutoModel.from_config(config, **kwargs) + else: + raise ValueError("Either pretrained or config must be provided.") + + self.config = self.model.config + # create dummy input to get the size of the last hidden state + if "use_flash_attention_2" in kwargs: + self.model = self.model.cuda() + dummy_input = torch.zeros((1, 1), dtype=torch.long).to(self.model.device) + out = self.model(dummy_input) + self.last_hidden_state_size = out.last_hidden_state.shape[-1] + self.model = self.model.cpu() + # print("self.last_hidden_state_size: ",self.last_hidden_state_size) + + def resize_token_embeddings(self, *args, **kwargs): + """ + Resize the token embeddings of the model. + + Args: + *args: Variable length argument list. + **kwargs: Arbitrary keyword arguments. + + Returns: + The resized token embeddings. + """ + return self.model.resize_token_embeddings(*args, **kwargs) diff --git a/applications/ColossalChat/coati/models/critic.py b/applications/ColossalChat/coati/models/critic.py new file mode 100755 index 000000000000..80340d9bd43d --- /dev/null +++ b/applications/ColossalChat/coati/models/critic.py @@ -0,0 +1,34 @@ +""" +Critic model +""" + +from typing import Optional + +import torch +import torch.nn as nn +from coati.models import BaseModel +from transformers import PretrainedConfig + + +class Critic(BaseModel): + """ + Critic model class. + + Args: + pretrained (str): path to pretrained model. + config (PretrainedConfig): PretrainedConfig used to initiate the base model. + """ + + def __init__(self, pretrained: str = None, config: Optional[PretrainedConfig] = None, **kwargs) -> None: + super().__init__(pretrained=pretrained, config=config, **kwargs) + # et last hidden state size with dummy input + self.value_head = nn.Linear(self.last_hidden_state_size, 1) + + def forward(self, input_ids: torch.LongTensor, attention_mask: Optional[torch.Tensor] = None) -> torch.Tensor: + outputs = self.model(input_ids, attention_mask=attention_mask) + last_hidden_states = outputs["last_hidden_state"] + sequence_hidden_states = last_hidden_states[torch.arange(last_hidden_states.size(0)), :].type( + self.value_head.weight.dtype + ) + values = self.value_head(sequence_hidden_states).squeeze(-1) # ensure shape is (B, sequence length) + return values diff --git a/applications/ColossalChat/coati/models/generation.py b/applications/ColossalChat/coati/models/generation.py new file mode 100755 index 000000000000..b671ef124063 --- /dev/null +++ b/applications/ColossalChat/coati/models/generation.py @@ -0,0 +1,428 @@ +from typing import Any, Callable, List, Optional + +import torch +import torch.distributed as dist +from transformers import PreTrainedTokenizer + +try: + from transformers.generation_logits_process import ( + LogitsProcessorList, + TemperatureLogitsWarper, + TopKLogitsWarper, + TopPLogitsWarper, + ) +except ImportError: + from transformers.generation import LogitsProcessorList, TemperatureLogitsWarper, TopKLogitsWarper, TopPLogitsWarper + + +def _prepare_logits_processor( + top_k: Optional[int] = None, top_p: Optional[float] = None, temperature: Optional[float] = None +) -> LogitsProcessorList: + """ + Prepare the logits processor list based on the given parameters. + + Args: + top_k (Optional[int]): The number of highest probability logits to keep for each token. + top_p (Optional[float]): The cumulative probability threshold for selecting tokens. + temperature (Optional[float]): The temperature value to apply to the logits. + + Returns: + LogitsProcessorList: The list of logits processors. + + """ + processor_list = LogitsProcessorList() + if temperature is not None and temperature != 1.0: + processor_list.append(TemperatureLogitsWarper(temperature)) + if top_k is not None and top_k != 0: + processor_list.append(TopKLogitsWarper(top_k)) + if top_p is not None and top_p < 1.0: + processor_list.append(TopPLogitsWarper(top_p)) + return processor_list + + +def _is_sequence_finished(unfinished_sequences: torch.Tensor) -> bool: + """ + Check if the sequence generation is finished. + + Args: + unfinished_sequences (torch.Tensor): Tensor indicating the unfinished sequences. + + Returns: + bool: True if all sequences are finished, False otherwise. + """ + if dist.is_initialized() and dist.get_world_size() > 1: + # consider DP + unfinished_sequences = unfinished_sequences.clone() + dist.all_reduce(unfinished_sequences) + return unfinished_sequences.max() == 0 + + +def update_model_kwargs_fn(outputs: dict, new_mask, **model_kwargs) -> dict: + """ + Update the model keyword arguments based on the outputs and new mask. + + Args: + outputs (dict): The outputs from the model. + new_mask: The new attention mask. + **model_kwargs: Additional model keyword arguments. + + Returns: + dict: The updated model keyword arguments. + """ + + if "past_key_values" in outputs: + model_kwargs["past_key_values"] = outputs["past_key_values"] + else: + model_kwargs["past_key_values"] = None + + # update token_type_ids with last value + if "token_type_ids" in model_kwargs: + token_type_ids = model_kwargs["token_type_ids"] + model_kwargs["token_type_ids"] = torch.cat([token_type_ids, token_type_ids[:, -1].unsqueeze(-1)], dim=-1) + + # update attention mask + if "attention_mask" in model_kwargs: + attention_mask = model_kwargs["attention_mask"] + model_kwargs["attention_mask"] = torch.cat([attention_mask, new_mask], dim=-1) + + return model_kwargs + + +def prepare_inputs_fn(input_ids: torch.Tensor, pad_token_id: int, **model_kwargs) -> dict: + model_kwargs["input_ids"] = input_ids + return model_kwargs + + +def _sample( + model: Any, + input_ids: torch.Tensor, + max_length: int, + early_stopping: bool = True, + eos_token_id: Optional[int] = None, + pad_token_id: Optional[int] = None, + stop_token_ids: Optional[List[int]] = None, + top_k: Optional[int] = None, + top_p: Optional[float] = None, + temperature: Optional[float] = None, + max_new_tokens: int = None, + prepare_inputs_fn: Optional[Callable[[torch.Tensor, Any], dict]] = None, + update_model_kwargs_fn: Optional[Callable[[dict, Any], dict]] = None, + stream_interval: int = 2, + **model_kwargs, +) -> torch.Tensor: + """ + Generates new tokens using the given model and input_ids. + + Args: + model (Any): The model used for token generation. + input_ids (torch.Tensor): The input tensor containing the initial tokens. + max_length (int): The maximum length of the generated tokens. + early_stopping (bool, optional): Whether to stop generating tokens early if all sequences are finished. Defaults to True. + eos_token_id (int, optional): The ID of the end-of-sequence token. Defaults to None. + pad_token_id (int, optional): The ID of the padding token. Defaults to None. + stop_token_ids (List[int], optional): A list of token IDs that, if encountered, will stop the generation process. Defaults to None. + top_k (int, optional): The number of top-k tokens to consider during sampling. Defaults to None. + top_p (float, optional): The cumulative probability threshold for top-p sampling. Defaults to None. + temperature (float, optional): The temperature value for token sampling. Defaults to None. + max_new_tokens (int, optional): The maximum number of new tokens to generate. Defaults to None. + prepare_inputs_fn (Callable[[torch.Tensor, Any], dict], optional): A function to prepare the model inputs. Defaults to None. + update_model_kwargs_fn (Callable[[dict, Any], dict], optional): A function to update the model kwargs. Defaults to None. + stream_interval (int, optional): The interval for streaming generation. Defaults to 2. + **model_kwargs: Additional keyword arguments for the model. + + Returns: + torch.Tensor: The tensor containing the generated tokens. + """ + context_length = input_ids.size(1) + if max_new_tokens is None: + max_new_tokens = max_length - context_length + if context_length + max_new_tokens > max_length or max_new_tokens == 0: + return input_ids + + logits_processor = _prepare_logits_processor(top_k, top_p, temperature) + unfinished_sequences = input_ids.new(input_ids.shape[0]).fill_(1) + past = None + for i in range(context_length, context_length + max_new_tokens): + # Calculate attention mask + if "attention_mask" not in model_kwargs: + model_kwargs["attention_mask"] = input_ids.ne(pad_token_id) + model_inputs = ( + prepare_inputs_fn(input_ids, past=past, **model_kwargs) + if prepare_inputs_fn is not None + else {"input_ids": input_ids, "attention_mask": input_ids.ne(pad_token_id)} + ) + outputs = model(**model_inputs) + + if "past_key_values" in outputs: + past = outputs.past_key_values + elif "mems" in outputs: + past = outputs.mems + + # NOTE: this is correct only in left padding mode + next_token_logits = outputs["logits"][:, -1, :] + next_token_logits = logits_processor(input_ids, next_token_logits) + + # Sample + probs = torch.softmax(next_token_logits, dim=-1, dtype=torch.float) + next_tokens = torch.multinomial(probs, num_samples=1).squeeze(1) + + # Finished sentences should have their next token be a padding token + if eos_token_id is not None: + assert pad_token_id is not None, "If `eos_token_id` is defined, make sure that `pad_token_id` is defined." + next_tokens = next_tokens * unfinished_sequences + pad_token_id * (1 - unfinished_sequences) + + # Update generated ids, model inputs for next step + input_ids = torch.cat([input_ids, next_tokens[:, None]], dim=-1) + + if update_model_kwargs_fn is not None: + model_kwargs = update_model_kwargs_fn(outputs, model_kwargs) + + # If eos_token was found in one sentence, set sentence to finished + if eos_token_id is not None: + unfinished_sequences = unfinished_sequences.mul((next_tokens != eos_token_id).long()) + + if stop_token_ids is not None: + # If the last len(stop_token_ids) tokens of input_ids are equal to stop_token_ids, set sentence to finished. + tokens_to_check = input_ids[:, -len(stop_token_ids) :] + unfinished_sequences = unfinished_sequences.mul( + torch.any(tokens_to_check != torch.LongTensor(stop_token_ids).to(input_ids.device), dim=1).long() + ) + + # Stop when each sentence is finished if early_stopping=True + if (early_stopping and _is_sequence_finished(unfinished_sequences)) or i == context_length + max_new_tokens - 1: + if i == context_length + max_new_tokens - 1: + # Force to end with stop token ids + input_ids[input_ids[:, -1] != pad_token_id, -len(stop_token_ids) :] = ( + torch.LongTensor(stop_token_ids).to(input_ids.device).long() + ) + return input_ids + + +@torch.inference_mode() +def generate( + model: Any, + input_ids: torch.Tensor, + tokenizer: PreTrainedTokenizer, + max_length: int, + num_beams: int = 1, + do_sample: bool = True, + early_stopping: bool = True, + top_k: Optional[int] = None, + top_p: Optional[float] = None, + temperature: Optional[float] = None, + prepare_inputs_fn: Optional[Callable[[torch.Tensor, Any], dict]] = None, + update_model_kwargs_fn: Optional[Callable[[dict, Any], dict]] = None, + **model_kwargs, +) -> torch.Tensor: + """Generate token sequence. The returned sequence is input_ids + generated_tokens. + + Args: + model (nn.Module): model + input_ids (torch.Tensor): input sequence + max_length (int): max length of the returned sequence + num_beams (int, optional): number of beams. Defaults to 1. + do_sample (bool, optional): whether to do sample. Defaults to True. + early_stopping (bool, optional): if True, the sequence length may be smaller than max_length due to finding eos. Defaults to False. + top_k (Optional[int], optional): the number of highest probability vocabulary tokens to keep for top-k-filtering. Defaults to None. + top_p (Optional[float], optional): If set to float < 1, only the smallest set of most probable tokens with probabilities that add up to top_p or higher are kept for generation. Defaults to None. + temperature (Optional[float], optional): The value used to module the next token probabilities. Defaults to None. + prepare_inputs_fn (Optional[Callable[[torch.Tensor, Any], dict]], optional): Function to preprocess model inputs. Arguments of this function should be input_ids and model_kwargs. Defaults to None. + update_model_kwargs_fn (Optional[Callable[[dict, Any], dict]], optional): Function to update model_kwargs based on outputs. Arguments of this function should be outputs and model_kwargs. Defaults to None. + """ + assert tokenizer.padding_side == "left", "Current generation only supports left padding." + is_greedy_gen_mode = (num_beams == 1) and do_sample is False + is_sample_gen_mode = (num_beams == 1) and do_sample is True + is_beam_gen_mode = (num_beams > 1) and do_sample is False + if is_greedy_gen_mode: + raise NotImplementedError + elif is_sample_gen_mode: + # Run sample + res = _sample( + model, + input_ids, + max_length, + early_stopping=early_stopping, + eos_token_id=tokenizer.eos_token_id, + pad_token_id=tokenizer.pad_token_id, + top_k=top_k, + top_p=top_p, + temperature=temperature, + prepare_inputs_fn=prepare_inputs_fn, + update_model_kwargs_fn=update_model_kwargs_fn, + **model_kwargs, + ) + return res + elif is_beam_gen_mode: + raise NotImplementedError + else: + raise ValueError("Unsupported generation mode") + + +def _sample_streaming( + model: Any, + input_ids: torch.Tensor, + max_length: int, + early_stopping: bool = False, + eos_token_id: Optional[int] = None, + pad_token_id: Optional[int] = None, + stop_token_ids: Optional[List[int]] = None, + top_k: Optional[int] = None, + top_p: Optional[float] = None, + temperature: Optional[float] = None, + max_new_tokens: int = None, + prepare_inputs_fn: Optional[Callable[[torch.Tensor, Any], dict]] = None, + update_model_kwargs_fn: Optional[Callable[[dict, Any], dict]] = None, + stream_interval: int = 2, + **model_kwargs, +) -> torch.Tensor: + """ + Generates new tokens using a streaming approach. + + Args: + model (Any): The model used for token generation. + input_ids (torch.Tensor): The input tensor containing the initial tokens. + max_length (int): The maximum length of the generated sequence. + early_stopping (bool, optional): Whether to stop generating tokens for a sequence if it is finished. Defaults to False. + eos_token_id (int, optional): The ID of the end-of-sequence token. Defaults to None. + pad_token_id (int, optional): The ID of the padding token. Defaults to None. + stop_token_ids (List[int], optional): A list of token IDs that, if encountered, will mark the sequence as finished. Defaults to None. + top_k (int, optional): The number of top-k tokens to consider during sampling. Defaults to None. + top_p (float, optional): The cumulative probability threshold for top-p sampling. Defaults to None. + temperature (float, optional): The temperature value for sampling. Defaults to None. + max_new_tokens (int, optional): The maximum number of new tokens to generate. Defaults to None. + prepare_inputs_fn (Callable[[torch.Tensor, Any], dict], optional): A function to prepare the model inputs. Defaults to None. + update_model_kwargs_fn (Callable[[dict, Any], dict], optional): A function to update the model keyword arguments. Defaults to None. + stream_interval (int, optional): The interval at which to yield the generated tokens. Defaults to 2. + **model_kwargs: Additional keyword arguments to be passed to the model. + + Yields: + torch.Tensor: The generated tokens at each step. + + Returns: + torch.Tensor: The final generated tokens. + """ + + context_length = input_ids.size(1) + if max_new_tokens is None: + max_new_tokens = max_length - context_length + if context_length + max_new_tokens > max_length or max_new_tokens == 0: + return input_ids + + logits_processor = _prepare_logits_processor(top_k, top_p, temperature) + unfinished_sequences = input_ids.new(input_ids.shape[0]).fill_(1) + + past = None + for i in range(context_length, context_length + max_new_tokens): + # calculate attention mask + if "attention_mask" not in model_kwargs: + model_kwargs["attention_mask"] = input_ids.ne(pad_token_id) + model_inputs = ( + prepare_inputs_fn(input_ids, past=past, **model_kwargs) + if prepare_inputs_fn is not None + else {"input_ids": input_ids, "attention_mask": input_ids.ne(pad_token_id)} + ) + outputs = model(**model_inputs) + if "past_key_values" in outputs: + past = outputs.past_key_values + elif "mems" in outputs: + past = outputs.mems + + # NOTE: this is correct only in left padding mode + next_token_logits = outputs["logits"][:, -1, :] + next_token_logits = logits_processor(input_ids, next_token_logits) + # sample + probs = torch.softmax(next_token_logits, dim=-1, dtype=torch.float) + next_tokens = torch.multinomial(probs, num_samples=1).squeeze(1) + + # finished sentences should have their next token be a padding token + if eos_token_id is not None: + assert pad_token_id is not None, "If `eos_token_id` is defined, make sure that `pad_token_id` is defined." + next_tokens = next_tokens * unfinished_sequences + pad_token_id * (1 - unfinished_sequences) + + # update generated ids, model inputs for next step + input_ids = torch.cat([input_ids, next_tokens[:, None]], dim=-1) + + if update_model_kwargs_fn is not None: + model_kwargs = update_model_kwargs_fn(outputs, model_kwargs) + + # if eos_token was found in one sentence, set sentence to finished + if eos_token_id is not None: + unfinished_sequences = unfinished_sequences.mul((next_tokens != eos_token_id).long()) + + if stop_token_ids is not None: + # If the last len(stop_token_ids) tokens of input_ids are equal to stop_token_ids, set sentence to finished. + tokens_to_check = input_ids[:, -len(stop_token_ids) :] + unfinished_sequences = unfinished_sequences.mul( + torch.any(tokens_to_check != torch.LongTensor(stop_token_ids).to(input_ids.device), dim=1).long() + ) + + # Stop when each sentence is finished if early_stopping=True + if ( + (early_stopping and _is_sequence_finished(unfinished_sequences)) + or (i - context_length) % stream_interval == 0 + or i == context_length + max_new_tokens - 1 + ): + yield input_ids + if early_stopping and _is_sequence_finished(unfinished_sequences): + break + + +@torch.inference_mode() +def generate_streaming( + model: Any, + input_ids: torch.Tensor, + tokenizer: PreTrainedTokenizer, + max_length: int, + num_beams: int = 1, + do_sample: bool = True, + early_stopping: bool = False, + top_k: Optional[int] = None, + top_p: Optional[float] = None, + temperature: Optional[float] = None, + prepare_inputs_fn: Optional[Callable[[torch.Tensor, Any], dict]] = None, + update_model_kwargs_fn: Optional[Callable[[dict, Any], dict]] = None, + **model_kwargs, +): + """Generate token sequence. The returned sequence is input_ids + generated_tokens. + + Args: + model (nn.Module): model + input_ids (torch.Tensor): input sequence + max_length (int): max length of the returned sequence + num_beams (int, optional): number of beams. Defaults to 1. + do_sample (bool, optional): whether to do sample. Defaults to True. + early_stopping (bool, optional): if True, the sequence length may be smaller than max_length due to finding eos. Defaults to False. + top_k (Optional[int], optional): the number of highest probability vocabulary tokens to keep for top-k-filtering. Defaults to None. + top_p (Optional[float], optional): If set to float < 1, only the smallest set of most probable tokens with probabilities that add up to top_p or higher are kept for generation. Defaults to None. + temperature (Optional[float], optional): The value used to module the next token probabilities. Defaults to None. + prepare_inputs_fn (Optional[Callable[[torch.Tensor, Any], dict]], optional): Function to preprocess model inputs. Arguments of this function should be input_ids and model_kwargs. Defaults to None. + update_model_kwargs_fn (Optional[Callable[[dict, Any], dict]], optional): Function to update model_kwargs based on outputs. Arguments of this function should be outputs and model_kwargs. Defaults to None. + """ + assert tokenizer.padding_side == "left", "Current generation only supports left padding." + is_greedy_gen_mode = (num_beams == 1) and do_sample is False + is_sample_gen_mode = (num_beams == 1) and do_sample is True + is_beam_gen_mode = (num_beams > 1) and do_sample is False + if is_greedy_gen_mode: + # run greedy search + raise NotImplementedError + elif is_sample_gen_mode: + # run sample + for res in _sample_streaming( + model, + input_ids, + max_length, + early_stopping=early_stopping, + eos_token_id=tokenizer.eos_token_id, + pad_token_id=tokenizer.pad_token_id, + top_k=top_k, + top_p=top_p, + temperature=temperature, + prepare_inputs_fn=prepare_inputs_fn, + update_model_kwargs_fn=update_model_kwargs_fn, + **model_kwargs, + ): + yield res + elif is_beam_gen_mode: + raise NotImplementedError + else: + raise ValueError("Unsupported generation mode") diff --git a/applications/Chat/coati/models/lora.py b/applications/ColossalChat/coati/models/lora.py old mode 100644 new mode 100755 similarity index 85% rename from applications/Chat/coati/models/lora.py rename to applications/ColossalChat/coati/models/lora.py index e9bd7b2ed8f0..9553b00ff2a8 --- a/applications/Chat/coati/models/lora.py +++ b/applications/ColossalChat/coati/models/lora.py @@ -1,3 +1,7 @@ +""" +LORA utils +""" + import dataclasses import math import warnings @@ -8,6 +12,10 @@ import torch.nn as nn import torch.nn.functional as F +from colossalai.logging import get_dist_logger + +logger = get_dist_logger() + @dataclasses.dataclass class LoRAManager: @@ -58,6 +66,10 @@ def reset_parameters(self): nn.init.zeros_(self.lora_B) def train(self, mode: bool = True): + """ + This function runs when model.train() is invoked. It is used to prepare the linear layer for training + """ + def T(w): return w.T if self.fan_in_fan_out else w @@ -101,6 +113,16 @@ def T(w): def _lora_linear_wrapper(linear: nn.Linear, lora_rank: int) -> LoraLinear: + """ + Wraps a linear layer with LoRA functionality. + + Args: + linear (nn.Linear): The linear layer to be wrapped. + lora_rank (int): The rank of the LoRA decomposition. + + Returns: + LoraLinear: The wrapped linear layer with LoRA functionality. + """ assert ( lora_rank <= linear.in_features ), f"LoRA rank ({lora_rank}) must be less than or equal to in features ({linear.in_features})" @@ -109,6 +131,16 @@ def _lora_linear_wrapper(linear: nn.Linear, lora_rank: int) -> LoraLinear: def _convert_to_lora_recursively(module: nn.Module, lora_rank: int) -> None: + """ + Recursively converts the given module and its children to LoRA (Low-Rank Approximation) form. + + Args: + module (nn.Module): The module to convert to LoRA form. + lora_rank (int): The rank of the LoRA approximation. + + Returns: + None + """ for name, child in module.named_children(): if isinstance(child, nn.Linear): setattr(module, name, _lora_linear_wrapper(child, lora_rank)) @@ -131,23 +163,3 @@ def convert_to_lora_module(module: nn.Module, lora_rank: int, lora_train_bias: s _convert_to_lora_recursively(module, lora_rank) lora.mark_only_lora_as_trainable(module, lora_train_bias) return module - - -class LoRAModule(nn.Module): - """A LoRA module base class. All derived classes should call `convert_to_lora()` at the bottom of `__init__()`. - This class will convert all torch.nn.Linear layer to LoraLinear layer. - - Args: - lora_rank (int, optional): LoRA rank. 0 means LoRA is not applied. Defaults to 0. - lora_train_bias (str, optional): Whether LoRA train biases. - 'none' means it doesn't train biases. 'all' means it trains all biases. 'lora_only' means it only trains biases of LoRA layers. - Defaults to 'none'. - """ - - def __init__(self, lora_rank: int = 0, lora_train_bias: str = "none") -> None: - super().__init__() - self.lora_rank = lora_rank - self.lora_train_bias = lora_train_bias - - def convert_to_lora(self) -> None: - convert_to_lora_module(self, self.lora_rank, self.lora_train_bias) diff --git a/applications/ColossalChat/coati/models/loss.py b/applications/ColossalChat/coati/models/loss.py new file mode 100755 index 000000000000..aaef447a4383 --- /dev/null +++ b/applications/ColossalChat/coati/models/loss.py @@ -0,0 +1,169 @@ +""" +loss functions +""" +from typing import Optional, Tuple + +import torch +import torch.nn as nn + +from .utils import masked_mean + + +class GPTLMLoss(nn.Module): + """ + GPT Language Model Loss + """ + + def __init__(self): + super().__init__() + # NOTE: default ignore_index is -100, which is equal to IGNORE_INDEX in sft_dataset.py + self.loss = nn.CrossEntropyLoss() + + def forward(self, logits: torch.Tensor, labels: torch.Tensor) -> torch.Tensor: + shift_logits = logits[..., :-1, :].contiguous() + shift_labels = labels[..., 1:].contiguous() + # Flatten the tokens + return self.loss(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1)) + + +class PolicyLoss(nn.Module): + """ + Policy Loss for PPO + """ + + def __init__(self, clip_eps: float = 0.2, skip_threshold: float = 20.0) -> None: + super().__init__() + self.clip_eps = clip_eps + self.skip_threshold = skip_threshold + + def forward( + self, + log_probs: torch.Tensor, + old_log_probs: torch.Tensor, + advantages: torch.Tensor, + action_mask: Optional[torch.Tensor] = None, + ) -> torch.Tensor: + skip = False + ratio_ = ((log_probs - old_log_probs) * action_mask).exp() + + # note that if dropout is disabled (recommanded), ratio will always be 1. + if ratio_.mean() > self.skip_threshold: + skip = True + + ratio = ratio_.clamp(0.0, 10.0) + surr1 = ratio * advantages + surr2 = ratio.clamp(1 - self.clip_eps, 1 + self.clip_eps) * advantages + loss = -torch.min(surr1, surr2) + loss = masked_mean(loss, action_mask) + loss = loss.mean() + return loss, skip, ratio_.max() + + +class ValueLoss(nn.Module): + """ + Value Loss for PPO + """ + + def __init__(self, clip_eps: float = 0.2) -> None: + super().__init__() + self.clip_eps = clip_eps + + def forward( + self, + values: torch.Tensor, + old_values: torch.Tensor, + advantage: torch.Tensor, + action_mask: Optional[torch.Tensor] = None, + ) -> torch.Tensor: + returns = advantage + old_values + values_clipped = old_values + (values - old_values).clamp(-self.clip_eps, self.clip_eps) + surr1 = (values_clipped - returns) ** 2 + surr2 = (values - returns) ** 2 + loss = torch.max(surr1, surr2) / torch.sum(action_mask) + loss = torch.sum(loss * action_mask) + return 0.5 * loss + + +class DpoLoss(nn.Module): + """ + Dpo loss + Details: https://arxiv.org/pdf/2305.18290.pdf + """ + + def __init__(self, beta: float = 0.1): + super().__init__() + self.beta = beta + + def forward( + self, + logprob_actor_chosen: torch.Tensor, + logprob_actor_reject: torch.Tensor, + logprob_ref_chosen: torch.Tensor, + logprob_ref_reject: torch.Tensor, + chosen_mask: torch.Tensor, + reject_mask: torch.Tensor, + ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]: + """Compute the DPO loss for a batch of policy and reference model log probabilities. + + # adapted from https://github.com/huggingface/trl/blob/main/trl/trainer/dpo_trainer.py#L328 + + Args: + logprob_actor_chosen: Log probabilities of the policy model for the chosen responses. Shape: (batch_size,) + logprob_actor_reject: Log probabilities of the policy model for the rejected responses. Shape: (batch_size,) + logprob_ref_chosen: Log probabilities of the reference model for the chosen responses. Shape: (batch_size,) + logprob_ref_reject: Log probabilities of the reference model for the rejected responses. Shape: (batch_size,) + + Returns: + A tuple of three tensors: (losses, chosen_rewards, rejected_rewards). + The losses tensor contains the DPO loss for each example in the batch. + The chosen_rewards and rejected_rewards tensors contain the rewards for the chosen and rejected responses, respectively. + """ + logprob_actor_chosen = logprob_actor_chosen * chosen_mask + logprob_actor_reject = logprob_actor_reject * reject_mask + if logprob_ref_chosen is not None and logprob_ref_reject is not None: + logprob_ref_chosen = logprob_ref_chosen * chosen_mask + logprob_ref_reject = logprob_ref_reject * reject_mask + if len(logprob_ref_chosen.shape) == 2: + ref_logratios = logprob_ref_chosen.sum(-1) - logprob_ref_reject.sum(-1) + else: + ref_logratios = logprob_ref_chosen.squeeze() - logprob_ref_reject.squeeze() + else: + # If no reference model is provided + ref_logratios = 0.0 + + pi_logratios = logprob_actor_chosen.sum(-1) - logprob_actor_reject.sum(-1) + logits = pi_logratios - ref_logratios + losses = -torch.nn.functional.logsigmoid(self.beta * logits) + + # Calculate rewards for logging + if logprob_ref_chosen is not None: + chosen_rewards = self.beta * (logprob_actor_chosen.sum(-1) - logprob_ref_chosen.sum(-1)).detach() + else: + chosen_rewards = self.beta * logprob_actor_chosen.sum(-1).detach() + if logprob_ref_reject is not None: + rejected_rewards = self.beta * (logprob_actor_reject.sum(-1) - logprob_ref_reject.sum(-1)).detach() + else: + rejected_rewards = self.beta * logprob_actor_reject.sum(-1).detach() + + return losses, chosen_rewards, rejected_rewards + + +class LogSigLoss(nn.Module): + """ + Pairwise Loss for Reward Model + Details: https://arxiv.org/abs/2203.02155 + """ + + def forward(self, chosen_reward: torch.Tensor, reject_reward: torch.Tensor) -> torch.Tensor: + return -torch.nn.functional.logsigmoid(chosen_reward - reject_reward).mean() + + +class LogExpLoss(nn.Module): + """ + Pairwise Loss for Reward Model + Details: https://arxiv.org/abs/2204.05862 + """ + + def forward(self, chosen_reward: torch.Tensor, reject_reward: torch.Tensor) -> torch.Tensor: + loss = torch.log(1 + torch.exp(reject_reward - chosen_reward)).mean() + return loss diff --git a/applications/ColossalChat/coati/models/reward_model.py b/applications/ColossalChat/coati/models/reward_model.py new file mode 100755 index 000000000000..18c5eca41a71 --- /dev/null +++ b/applications/ColossalChat/coati/models/reward_model.py @@ -0,0 +1,38 @@ +""" +reward model +""" +from typing import Optional + +import torch +import torch.nn as nn +from coati.models import BaseModel +from transformers import PretrainedConfig + + +class RewardModel(BaseModel): + """ + Reward model class. + + Args: + pretrained str: huggingface or local model path + config: PretrainedConfig object + **kwargs: all other kwargs as in AutoModel.from_pretrained + """ + + def __init__(self, pretrained: str = None, config: Optional[PretrainedConfig] = None, **kwargs) -> None: + super().__init__(pretrained=pretrained, config=config, **kwargs) + self.value_head = nn.Linear(self.last_hidden_state_size, 1) + self.value_head.weight.data.normal_(mean=0.0, std=1 / (self.last_hidden_state_size + 1)) + + def forward(self, input_ids: torch.LongTensor, attention_mask: Optional[torch.Tensor] = None) -> torch.Tensor: + outputs = self.model(input_ids, attention_mask=attention_mask) + + last_hidden_states = outputs["last_hidden_state"] + sequence_lengths = torch.max(attention_mask * torch.arange(input_ids.size(1), device=input_ids.device), dim=1)[ + 0 + ] + sequence_hidden_states = last_hidden_states[torch.arange(last_hidden_states.size(0)), sequence_lengths].type( + self.value_head.weight.dtype + ) + values = self.value_head(sequence_hidden_states).squeeze(-1) # Ensure shape is (B,) + return values diff --git a/applications/ColossalChat/coati/models/utils.py b/applications/ColossalChat/coati/models/utils.py new file mode 100755 index 000000000000..ce672534c28e --- /dev/null +++ b/applications/ColossalChat/coati/models/utils.py @@ -0,0 +1,137 @@ +import json +import os +from typing import Any, Dict, Optional, Union + +import torch +import torch.nn.functional as F + + +def get_model_numel(model: torch.nn.Module) -> int: + return sum(p.numel() for p in model.parameters()) + + +def compute_reward( + r: Union[torch.Tensor, float], + kl_coef: float, + log_probs: torch.Tensor, + log_probs_base: torch.Tensor, + action_mask: Optional[torch.Tensor] = None, + reward_eps=5, +) -> torch.Tensor: + """ + Args: + log_probs: [batch_size, response_length] + log_probs_base: [batch_size, response_length] + action_mask: [batch_size, response_length] + r: float + Returns: + reward: [batch_size, response_length] + """ + log_ratio = log_probs - log_probs_base # address numerical instability issue + kl = -kl_coef * log_ratio * action_mask + reward = kl + r_clip = torch.clamp(r, -reward_eps, reward_eps) + for i in range(action_mask.size(0)): + assert action_mask[i].sum() > 0 + reward[i, : action_mask[i].sum()] += r_clip[i] + reward[i, action_mask[i].sum() :] *= 0 + return reward, ((log_ratio * (log_ratio < 10)).exp() - 1 - log_ratio) * action_mask + + +def _log_probs_from_logits(logits: torch.Tensor, labels: torch.Tensor) -> torch.Tensor: + """ + Compute the log probabilities from logits for the given labels. + + Args: + logits (torch.Tensor): The input logits. + labels (torch.Tensor): The target labels. + + Returns: + torch.Tensor: The log probabilities corresponding to the labels. + """ + log_probs = F.log_softmax(logits, dim=-1) + log_probs_labels = log_probs.gather(dim=-1, index=labels.unsqueeze(-1)) + return log_probs_labels.squeeze(-1) + + +def calc_action_log_probs(logits: torch.Tensor, sequences: torch.LongTensor, num_actions: int) -> torch.Tensor: + """Calculate action log probs. + + Args: + output (torch.Tensor): Output tensor of Actor.forward.logits. + sequences (torch.LongTensor): Input sequences. + num_actions (int): Number of actions. + + Returns: + torch.Tensor: Action log probs. + """ + log_probs = _log_probs_from_logits(logits[:, :-1, :], sequences[:, 1:]) + return log_probs[:, -num_actions:] + + +def masked_mean(tensor: torch.Tensor, mask: torch.Tensor, dim: int = 1) -> torch.Tensor: + """ + Compute the masked mean of a tensor along a specified dimension. + + Args: + tensor (torch.Tensor): The input tensor. + mask (torch.Tensor): The mask tensor with the same shape as the input tensor. + dim (int, optional): The dimension along which to compute the mean. Default is 1. + + Returns: + torch.Tensor: The masked mean tensor. + + """ + tensor = tensor * mask + tensor = tensor.sum(dim=dim) + mask_sum = mask.sum(dim=dim) + mean = tensor / (mask_sum + 1e-8) + return mean + + +def calc_masked_log_probs(logits: torch.Tensor, sequences: torch.LongTensor, mask: torch.Tensor) -> torch.Tensor: + """ + Calculate the masked log probabilities for a given sequence of logits. + + Args: + logits (torch.Tensor): The input logits tensor of shape (batch_size, sequence_length, vocab_size). + sequences (torch.LongTensor): The input sequence tensor of shape (batch_size, sequence_length). + mask (torch.Tensor): The mask tensor of shape (batch_size, sequence_length). + + Returns: + torch.Tensor: The masked log probabilities tensor of shape (batch_size, sequence_length - 1). + """ + # logits are probabilities of the next token, so we shift them to the left by one + log_probs = _log_probs_from_logits(logits[:, :-1, :], sequences[:, 1:]) + return log_probs * mask + + +def load_json(file_path: Union[str, os.PathLike]) -> Dict[str, Any]: + """ + Load file in JSON format + """ + with open(file=file_path, mode="r", encoding="utf-8") as fp: + return json.load(fp) + + +def save_json(data: Dict[str, Any], file_path: Union[str, os.PathLike]) -> None: + """ + Save as JSON format + """ + with open(file=file_path, mode="w", encoding="utf-8") as fp: + json.dump(data, fp=fp, ensure_ascii=False, indent=4) + + +def disable_dropout(model: torch.nn.Module): + """ + Disables dropout in a PyTorch model. This is used in PPO Training + + Args: + model (torch.nn.Module): The PyTorch model. + + Returns: + None + """ + for module in model.modules(): + if isinstance(module, torch.nn.Dropout): + module.p = 0.0 diff --git a/applications/Chat/coati/quant/__init__.py b/applications/ColossalChat/coati/quant/__init__.py old mode 100644 new mode 100755 similarity index 100% rename from applications/Chat/coati/quant/__init__.py rename to applications/ColossalChat/coati/quant/__init__.py diff --git a/applications/Chat/coati/quant/llama_gptq/__init__.py b/applications/ColossalChat/coati/quant/llama_gptq/__init__.py old mode 100644 new mode 100755 similarity index 100% rename from applications/Chat/coati/quant/llama_gptq/__init__.py rename to applications/ColossalChat/coati/quant/llama_gptq/__init__.py diff --git a/applications/Chat/coati/quant/llama_gptq/loader.py b/applications/ColossalChat/coati/quant/llama_gptq/loader.py old mode 100644 new mode 100755 similarity index 100% rename from applications/Chat/coati/quant/llama_gptq/loader.py rename to applications/ColossalChat/coati/quant/llama_gptq/loader.py diff --git a/applications/Chat/coati/quant/llama_gptq/model_utils.py b/applications/ColossalChat/coati/quant/llama_gptq/model_utils.py old mode 100644 new mode 100755 similarity index 100% rename from applications/Chat/coati/quant/llama_gptq/model_utils.py rename to applications/ColossalChat/coati/quant/llama_gptq/model_utils.py diff --git a/applications/Chat/coati/quant/llama_gptq/quant.py b/applications/ColossalChat/coati/quant/llama_gptq/quant.py old mode 100644 new mode 100755 similarity index 100% rename from applications/Chat/coati/quant/llama_gptq/quant.py rename to applications/ColossalChat/coati/quant/llama_gptq/quant.py diff --git a/applications/Chat/coati/quant/utils.py b/applications/ColossalChat/coati/quant/utils.py old mode 100644 new mode 100755 similarity index 100% rename from applications/Chat/coati/quant/utils.py rename to applications/ColossalChat/coati/quant/utils.py diff --git a/applications/Chat/coati/ray/README.md b/applications/ColossalChat/coati/ray/README.md old mode 100644 new mode 100755 similarity index 100% rename from applications/Chat/coati/ray/README.md rename to applications/ColossalChat/coati/ray/README.md diff --git a/applications/Chat/coati/ray/__init__.py b/applications/ColossalChat/coati/ray/__init__.py old mode 100644 new mode 100755 similarity index 100% rename from applications/Chat/coati/ray/__init__.py rename to applications/ColossalChat/coati/ray/__init__.py diff --git a/applications/Chat/coati/ray/callbacks/__init__.py b/applications/ColossalChat/coati/ray/callbacks/__init__.py old mode 100644 new mode 100755 similarity index 100% rename from applications/Chat/coati/ray/callbacks/__init__.py rename to applications/ColossalChat/coati/ray/callbacks/__init__.py diff --git a/applications/Chat/coati/ray/callbacks/base.py b/applications/ColossalChat/coati/ray/callbacks/base.py old mode 100644 new mode 100755 similarity index 100% rename from applications/Chat/coati/ray/callbacks/base.py rename to applications/ColossalChat/coati/ray/callbacks/base.py diff --git a/applications/Chat/coati/ray/callbacks/performance_evaluator.py b/applications/ColossalChat/coati/ray/callbacks/performance_evaluator.py old mode 100644 new mode 100755 similarity index 100% rename from applications/Chat/coati/ray/callbacks/performance_evaluator.py rename to applications/ColossalChat/coati/ray/callbacks/performance_evaluator.py diff --git a/applications/Chat/coati/ray/detached_replay_buffer.py b/applications/ColossalChat/coati/ray/detached_replay_buffer.py old mode 100644 new mode 100755 similarity index 100% rename from applications/Chat/coati/ray/detached_replay_buffer.py rename to applications/ColossalChat/coati/ray/detached_replay_buffer.py diff --git a/applications/Chat/coati/ray/detached_trainer_base.py b/applications/ColossalChat/coati/ray/detached_trainer_base.py old mode 100644 new mode 100755 similarity index 100% rename from applications/Chat/coati/ray/detached_trainer_base.py rename to applications/ColossalChat/coati/ray/detached_trainer_base.py diff --git a/applications/Chat/coati/ray/detached_trainer_ppo.py b/applications/ColossalChat/coati/ray/detached_trainer_ppo.py old mode 100644 new mode 100755 similarity index 100% rename from applications/Chat/coati/ray/detached_trainer_ppo.py rename to applications/ColossalChat/coati/ray/detached_trainer_ppo.py diff --git a/applications/Chat/coati/ray/experience_maker_holder.py b/applications/ColossalChat/coati/ray/experience_maker_holder.py old mode 100644 new mode 100755 similarity index 100% rename from applications/Chat/coati/ray/experience_maker_holder.py rename to applications/ColossalChat/coati/ray/experience_maker_holder.py diff --git a/applications/Chat/coati/ray/lora_constructor.py b/applications/ColossalChat/coati/ray/lora_constructor.py old mode 100644 new mode 100755 similarity index 100% rename from applications/Chat/coati/ray/lora_constructor.py rename to applications/ColossalChat/coati/ray/lora_constructor.py diff --git a/applications/Chat/coati/ray/utils.py b/applications/ColossalChat/coati/ray/utils.py old mode 100644 new mode 100755 similarity index 100% rename from applications/Chat/coati/ray/utils.py rename to applications/ColossalChat/coati/ray/utils.py diff --git a/applications/ColossalChat/coati/trainer/__init__.py b/applications/ColossalChat/coati/trainer/__init__.py new file mode 100755 index 000000000000..2eff8ca7676a --- /dev/null +++ b/applications/ColossalChat/coati/trainer/__init__.py @@ -0,0 +1,7 @@ +from .base import OLTrainer, SLTrainer +from .dpo import DPOTrainer +from .ppo import PPOTrainer +from .rm import RewardModelTrainer +from .sft import SFTTrainer + +__all__ = ["SLTrainer", "OLTrainer", "RewardModelTrainer", "SFTTrainer", "PPOTrainer", "DPOTrainer"] diff --git a/applications/Chat/coati/trainer/base.py b/applications/ColossalChat/coati/trainer/base.py old mode 100644 new mode 100755 similarity index 80% rename from applications/Chat/coati/trainer/base.py rename to applications/ColossalChat/coati/trainer/base.py index 0a41d450d41e..63c903a51940 --- a/applications/Chat/coati/trainer/base.py +++ b/applications/ColossalChat/coati/trainer/base.py @@ -1,6 +1,14 @@ +""" +Base trainers for online and offline training + SLTrainer: supervised learning trainer + pretrain, sft, dpo, reward model training + OLTrainer: online learning trainer + rlhf-ppo +""" + from abc import ABC, abstractmethod from contextlib import contextmanager -from typing import List +from typing import Callable, List import torch.nn as nn import tqdm @@ -8,8 +16,8 @@ from coati.experience_maker import Experience from torch.optim import Optimizer -from .callbacks import Callback -from .strategies import Strategy +from colossalai.booster import Booster + from .utils import is_rank_0 @@ -26,16 +34,18 @@ class SLTrainer(ABC): def __init__( self, - strategy: Strategy, + booster: Booster, max_epochs: int, model: nn.Module, optimizer: Optimizer, + start_epoch: int = 0, ) -> None: super().__init__() - self.strategy = strategy + self.booster = booster self.max_epochs = max_epochs self.model = model self.optimizer = optimizer + self.start_epoch = start_epoch @abstractmethod def _train(self, epoch): @@ -45,19 +55,20 @@ def _train(self, epoch): def _eval(self, epoch): raise NotImplementedError() + @abstractmethod def _before_fit(self): raise NotImplementedError() def fit(self, *args, **kwargs): self._before_fit(*args, **kwargs) - for epoch in tqdm.trange(self.max_epochs, desc="Epochs", disable=not is_rank_0()): + for epoch in tqdm.trange(self.start_epoch, self.max_epochs, desc="Epochs", disable=not is_rank_0()): self._train(epoch) self._eval(epoch) -class OnPolicyTrainer(ABC): +class OLTrainer(ABC): """ - Base class for on-policy rl trainers, e.g. PPO. + Base class for online learning trainers, e.g. PPO. Args: strategy (Strategy):the strategy to use for training @@ -69,14 +80,16 @@ class OnPolicyTrainer(ABC): def __init__( self, - strategy: Strategy, + actor_booster: Booster, + critic_booster: Booster, data_buffer: NaiveExperienceBuffer, sample_buffer: bool, dataloader_pin_memory: bool, - callbacks: List[Callback] = [], + callbacks: List[Callable] = [], ) -> None: super().__init__() - self.strategy = strategy + self.actor_booster = actor_booster + self.critic_booster = critic_booster self.data_buffer = data_buffer self.sample_buffer = sample_buffer self.dataloader_pin_memory = dataloader_pin_memory @@ -141,6 +154,20 @@ def _learn(self, update_step: int): """ raise NotImplementedError() + @abstractmethod + def _setup_update_phrase_dataload(self): + """ + Implement this method to setup dataloader for update phase. + """ + raise NotImplementedError() + + @abstractmethod + def _save_checkpoint(self, episode: int = 0): + """ + Implement this method to save checkpoint. + """ + raise NotImplementedError() + def _collect_phase(self, collect_step: int): self._on_make_experience_start() experience = self._make_experience(collect_step) @@ -178,11 +205,10 @@ def fit( for collect_step in tqdm.trange(num_collect_steps, desc="Collect steps", disable=not is_rank_0()): self._collect_phase(collect_step) if not self.sample_buffer: - # HACK(cwher): according to the design of boost API, dataloader should also be boosted, - # but it is impractical to adapt this pattern in RL training. Thus, I left dataloader unboosted. - # I only call strategy.setup_dataloader() to setup dataloader. - self.dataloader = self.strategy.setup_dataloader(self.data_buffer, self.dataloader_pin_memory) + self._setup_update_phrase_dataload() for update_step in tqdm.trange(num_update_steps, desc="Update steps", disable=not is_rank_0()): self._update_phase(update_step) # NOTE: this is for on-policy algorithms self.data_buffer.clear() + if self.save_interval > 0 and (episode + 1) % (self.save_interval) == 0: + self._save_checkpoint(episode + 1) diff --git a/applications/ColossalChat/coati/trainer/callbacks/__init__.py b/applications/ColossalChat/coati/trainer/callbacks/__init__.py new file mode 100644 index 000000000000..a765485072c1 --- /dev/null +++ b/applications/ColossalChat/coati/trainer/callbacks/__init__.py @@ -0,0 +1,4 @@ +from .base import Callback +from .performance_evaluator import PerformanceEvaluator + +__all__ = ["Callback", "PerformanceEvaluator"] diff --git a/applications/Chat/coati/trainer/callbacks/base.py b/applications/ColossalChat/coati/trainer/callbacks/base.py similarity index 100% rename from applications/Chat/coati/trainer/callbacks/base.py rename to applications/ColossalChat/coati/trainer/callbacks/base.py diff --git a/applications/Chat/coati/trainer/callbacks/performance_evaluator.py b/applications/ColossalChat/coati/trainer/callbacks/performance_evaluator.py similarity index 92% rename from applications/Chat/coati/trainer/callbacks/performance_evaluator.py rename to applications/ColossalChat/coati/trainer/callbacks/performance_evaluator.py index b286c766c263..86384e5e39fb 100644 --- a/applications/Chat/coati/trainer/callbacks/performance_evaluator.py +++ b/applications/ColossalChat/coati/trainer/callbacks/performance_evaluator.py @@ -14,9 +14,11 @@ def get_world_size() -> int: return 1 -def print_rank_0(*args, **kwargs) -> None: +def save_eval_result_rank_0(s: str, save_path: str, **kwargs) -> None: if not dist.is_initialized() or dist.get_rank() == 0: - print(*args, **kwargs) + with open(save_path, "a+") as f: + train_config = "; ".join([str(kwargs[key]) for key in kwargs]) + f.write(train_config + "\n" + s + "\n") def divide(x: float, y: float) -> float: @@ -74,6 +76,8 @@ def __init__( reward_model_num_params: int, enable_grad_checkpoint: bool = False, ignore_episodes: int = 0, + train_config: Optional[dict] = None, + save_path: Optional[str] = None, ) -> None: super().__init__() self.world_size = get_world_size() @@ -92,6 +96,8 @@ def __init__( self.make_experience_flop: int = 0 self.learn_num_samples: int = 0 self.learn_flop: int = 0 + self.train_config = train_config + self.save_path = save_path def on_episode_start(self, episode: int) -> None: self.disable = self.ignore_episodes > 0 and episode < self.ignore_episodes @@ -172,12 +178,14 @@ def on_fit_end(self) -> None: make_experience_time_per_sample = divide(avg_make_experience_duration, num_effective_samples) learn_time_per_sample = divide(avg_learn_duration, num_effective_samples) - print_rank_0( + save_eval_result_rank_0( f"Performance summary:\n" + f"Generate {self.make_experience_num_samples * self.world_size} samples, throughput: {avg_make_experience_throughput:.2f} samples/s, TFLOPS per GPU: {avg_make_experience_tflops:.2f}\n" + f"Train {self.learn_num_samples * self.world_size} samples, throughput: {avg_learn_throughput:.2f} samples/s, TFLOPS per GPU: {avg_learn_tflops:.2f}\n" + f"Overall throughput: {avg_overall_throughput:.2f} samples/s\n" + f"Overall time per sample: {overall_time_per_sample:.2f} s\n" + f"Make experience time per sample: {make_experience_time_per_sample:.2f} s, {make_experience_time_per_sample/overall_time_per_sample*100:.2f}%\n" - + f"Learn time per sample: {learn_time_per_sample:.2f} s, {learn_time_per_sample/overall_time_per_sample*100:.2f}%" + + f"Learn time per sample: {learn_time_per_sample:.2f} s, {learn_time_per_sample/overall_time_per_sample*100:.2f}%", + self.save_path, + **self.train_config, ) diff --git a/applications/ColossalChat/coati/trainer/dpo.py b/applications/ColossalChat/coati/trainer/dpo.py new file mode 100755 index 000000000000..cbe7d7ca811a --- /dev/null +++ b/applications/ColossalChat/coati/trainer/dpo.py @@ -0,0 +1,336 @@ +""" +Dpo trainer +""" + +from typing import Any, Optional + +import torch +from coati.models.loss import DpoLoss +from coati.models.utils import calc_masked_log_probs +from coati.trainer.utils import all_reduce_mean +from coati.utils import AccumulativeMeanMeter, save_checkpoint +from torch.optim import Optimizer +from torch.optim.lr_scheduler import _LRScheduler +from torch.utils.data import DataLoader +from tqdm import trange +from transformers import PreTrainedTokenizerBase + +from colossalai.booster import Booster +from colossalai.cluster import DistCoordinator +from colossalai.utils import get_current_device + +from .base import SLTrainer +from .utils import is_rank_0, to_device + + +class DPOTrainer(SLTrainer): + """ + Trainer for PPO algorithm. + + Args: + actor (Actor): the actor model in ppo algorithm + ref_model (Critic): the reference model in ppo algorithm + booster (Strategy): the strategy to use for training + actor_optim (Optimizer): the optimizer to use for actor model + actor_lr_scheduler (_LRScheduler): the lr scheduler to use for actor model + tokenizer (PreTrainedTokenizerBase): the tokenizer to use for encoding + max_epochs (int, defaults to 1): the max number of epochs to train + beta (float, defaults to 0.1): the beta parameter in dpo loss + accumulation_steps (int): the number of steps to accumulate gradients + start_epoch (int, defaults to 0): the start epoch, non-zero if resumed from a checkpoint + save_interval (int): the interval to save model checkpoints, default to 0, which means no checkpoint will be saved during trainning + save_dir (str): the directory to save checkpoints + coordinator (DistCoordinator): the coordinator to use for distributed logging + """ + + def __init__( + self, + actor: Any, + ref_model: Any, + booster: Booster, + actor_optim: Optimizer, + actor_lr_scheduler: _LRScheduler, + tokenizer: PreTrainedTokenizerBase, + max_epochs: int = 1, + beta: float = 0.1, + accumulation_steps: int = 1, + start_epoch: int = 0, + save_interval: int = 0, + save_dir: str = None, + coordinator: DistCoordinator = None, + ) -> None: + super().__init__(booster, max_epochs=max_epochs, model=actor, optimizer=actor_optim, start_epoch=start_epoch) + self.ref_model = ref_model + self.actor_scheduler = actor_lr_scheduler + self.tokenizer = tokenizer + self.actor_loss_fn = DpoLoss(beta) + self.save_interval = save_interval + self.coordinator = coordinator + self.save_dir = save_dir + self.num_train_step = 0 + self.accumulation_steps = accumulation_steps + self.device = get_current_device() + self.accumulative_meter = AccumulativeMeanMeter() + + def _before_fit( + self, + train_preference_dataloader: DataLoader = None, + eval_preference_dataloader: DataLoader = None, + log_dir: Optional[str] = None, + use_wandb: bool = False, + ): + """ + Args: + prompt_dataloader (DataLoader): the dataloader to use for prompt data + pretrain_dataloader (DataLoader): the dataloader to use for pretrain data + """ + self.train_dataloader = train_preference_dataloader + self.eval_dataloader = eval_preference_dataloader + self.writer = None + if use_wandb and is_rank_0(): + assert log_dir is not None, "log_dir must be provided when use_wandb is True" + import wandb + + self.wandb_run = wandb.init(project="Coati-dpo", sync_tensorboard=True) + if log_dir is not None and is_rank_0(): + import os + import time + + from torch.utils.tensorboard import SummaryWriter + + log_dir = os.path.join(log_dir, "dpo") + log_dir = os.path.join(log_dir, time.strftime("%Y-%m-%d_%H:%M:%S", time.localtime())) + self.writer = SummaryWriter(log_dir=log_dir) + + def _train(self, epoch: int): + """ + Args: + epoch int: the number of current epoch + """ + self.model.train() + self.accumulative_meter.reset() + step_bar = trange( + len(self.train_dataloader) // self.accumulation_steps, + desc=f"Epoch {epoch + 1}/{self.max_epochs}", + disable=not is_rank_0(), + ) + for i, batch in enumerate(self.train_dataloader): + batch = to_device(batch, self.device) + ( + chosen_input_ids, + chosen_attention_mask, + chosen_loss_mask, + reject_input_ids, + reject_attention_mask, + reject_loss_mask, + ) = ( + batch["chosen_input_ids"], + batch["chosen_attention_mask"], + batch["chosen_loss_mask"], + batch["reject_input_ids"], + batch["reject_attention_mask"], + batch["reject_loss_mask"], + ) + reject_loss_mask[:, -1] = False + batch_size = chosen_input_ids.size()[0] + + actor_all_logits = self.model( + input_ids=torch.cat([chosen_input_ids, reject_input_ids]), + attention_mask=torch.cat([chosen_attention_mask, reject_attention_mask]), + )["logits"].to(torch.float32) + actor_chosen_logits = actor_all_logits[:batch_size] + actor_reject_logits = actor_all_logits[batch_size:] + logprob_actor_chosen = calc_masked_log_probs(actor_chosen_logits, chosen_input_ids, chosen_loss_mask[:, 1:]) + + logprob_actor_reject = calc_masked_log_probs(actor_reject_logits, reject_input_ids, reject_loss_mask[:, 1:]) + + if self.ref_model is not None: + self.ref_model.eval() + with torch.no_grad(): + ref_all_logits = self.ref_model( + input_ids=torch.cat([chosen_input_ids, reject_input_ids]), + attention_mask=torch.cat([chosen_attention_mask, reject_attention_mask]), + )["logits"].to(torch.float32) + ref_chosen_logits = ref_all_logits[:batch_size] + ref_reject_logits = ref_all_logits[batch_size:] + logprob_ref_chosen = calc_masked_log_probs( + ref_chosen_logits, chosen_input_ids, chosen_loss_mask[:, 1:] + ) + logprob_ref_reject = calc_masked_log_probs( + ref_reject_logits, reject_input_ids, reject_loss_mask[:, 1:] + ) + else: + logprob_ref_chosen = None + logprob_ref_reject = None + + losses, chosen_rewards, rejected_rewards = self.actor_loss_fn( + logprob_actor_chosen, + logprob_actor_reject, + logprob_ref_chosen if logprob_ref_chosen is not None else None, + logprob_ref_reject if logprob_ref_reject is not None else None, + chosen_loss_mask[:, 1:], + reject_loss_mask[:, 1:], + ) + reward_accuracies = (chosen_rewards > rejected_rewards).float().mean() + + # DPO Loss + loss = losses.mean() + + self.booster.backward(loss=loss, optimizer=self.optimizer) + if self.num_train_step % self.accumulation_steps == self.accumulation_steps - 1: + self.optimizer.step() + self.optimizer.zero_grad() + self.actor_scheduler.step() + + # sync + loss_mean = all_reduce_mean(tensor=loss) + chosen_rewards_mean = all_reduce_mean(tensor=chosen_rewards) + rejected_rewards_mean = all_reduce_mean(tensor=rejected_rewards) + reward_accuracies_mean = all_reduce_mean(tensor=reward_accuracies) + self.accumulative_meter.add("chosen_rewards", chosen_rewards_mean.to(torch.float16).mean().item()) + self.accumulative_meter.add("rejected_rewards", rejected_rewards_mean.to(torch.float16).mean().item()) + self.accumulative_meter.add("loss", loss_mean.to(torch.float16).item()) + self.accumulative_meter.add("accuracy", reward_accuracies_mean.to(torch.float16).item()) + + if i % self.accumulation_steps == self.accumulation_steps - 1: + self.num_train_step += 1 + step_bar.update() + # logging + if self.writer and is_rank_0(): + self.writer.add_scalar("train/loss", self.accumulative_meter.get("loss"), self.num_train_step) + self.writer.add_scalar("train/lr", self.optimizer.param_groups[0]["lr"], self.num_train_step) + self.writer.add_scalar( + "train/chosen_rewards", self.accumulative_meter.get("chosen_rewards"), self.num_train_step + ) + self.writer.add_scalar( + "train/rejected_rewards", + self.accumulative_meter.get("rejected_rewards"), + self.num_train_step, + ) + self.writer.add_scalar( + "train/margin", + self.accumulative_meter.get("chosen_rewards") - self.accumulative_meter.get("rejected_rewards"), + self.num_train_step, + ) + self.writer.add_scalar( + "train/accuracy", + self.accumulative_meter.get("accuracy"), + self.num_train_step, + ) + self.accumulative_meter.reset() + + if (self.num_train_step + 1) % self.save_interval == 0: + # save checkpoint + self.coordinator.print_on_master("\nStart saving model checkpoint with running states") + save_checkpoint( + save_dir=self.save_dir, + booster=self.booster, + model=self.model, + optimizer=self.optimizer, + lr_scheduler=self.actor_scheduler, + epoch=epoch, + step=i + 1, + batch_size=batch_size, + coordinator=self.coordinator, + ) + self.coordinator.print_on_master( + f"Saved checkpoint at epoch {epoch} step {self.save_interval} at folder {self.save_dir}" + ) + + step_bar.close() + + def _eval(self, epoch: int): + """ + Args: + epoch int: the number of current epoch + """ + if self.eval_dataloader is None: + self.coordinator.print_on_master("No eval dataloader is provided, skip evaluation") + return + self.model.eval() + self.ref_model.eval() + self.coordinator.print_on_master("\nStart evaluation...") + + step_bar = trange( + len(self.eval_dataloader), + desc=f"Epoch {epoch + 1}/{self.max_epochs}", + disable=not is_rank_0(), + ) + + self.accumulative_meter.reset() + + with torch.no_grad(): + for i, batch in enumerate(self.eval_dataloader): + batch = to_device(batch, self.device) + ( + chosen_input_ids, + chosen_attention_mask, + chosen_loss_mask, + reject_input_ids, + reject_attention_mask, + reject_loss_mask, + ) = ( + batch["chosen_input_ids"], + batch["chosen_attention_mask"], + batch["chosen_loss_mask"], + batch["reject_input_ids"], + batch["reject_attention_mask"], + batch["reject_loss_mask"], + ) + + batch_size = chosen_input_ids.size()[0] + + actor_all_logits = self.model( + torch.cat([chosen_input_ids, reject_input_ids]), + torch.cat([chosen_attention_mask, reject_attention_mask]), + )["logits"].to(torch.float32) + actor_chosen_logits = actor_all_logits[:batch_size] + actor_reject_logits = actor_all_logits[batch_size:] + + logprob_actor_chosen = calc_masked_log_probs( + actor_chosen_logits, chosen_input_ids, chosen_loss_mask[:, 1:] + ) + + logprob_actor_reject = calc_masked_log_probs( + actor_reject_logits, reject_input_ids, reject_loss_mask[:, 1:] + ) + + self.ref_model.eval() + + ref_all_logits = self.ref_model( + torch.cat([chosen_input_ids, reject_input_ids]), + torch.cat([chosen_attention_mask, reject_attention_mask]), + )["logits"].to(torch.float32) + ref_chosen_logits = ref_all_logits[:batch_size] + ref_reject_logits = ref_all_logits[batch_size:] + logprob_ref_chosen = calc_masked_log_probs(ref_chosen_logits, chosen_input_ids, chosen_loss_mask[:, 1:]) + logprob_ref_reject = calc_masked_log_probs(ref_reject_logits, reject_input_ids, reject_loss_mask[:, 1:]) + + losses, chosen_rewards, rejected_rewards = self.actor_loss_fn( + logprob_actor_chosen, + logprob_actor_reject, + logprob_ref_chosen if logprob_ref_chosen is not None else None, + logprob_ref_reject if logprob_ref_reject is not None else None, + chosen_loss_mask[:, 1:], + reject_loss_mask[:, 1:], + ) + reward_accuracies = (chosen_rewards > rejected_rewards).float() + loss = losses.mean() + loss_mean = all_reduce_mean(tensor=loss) + chosen_rewards_mean = all_reduce_mean(tensor=chosen_rewards) + rejected_rewards_mean = all_reduce_mean(tensor=rejected_rewards) + reward_accuracies_mean = all_reduce_mean(tensor=reward_accuracies) + self.accumulative_meter.add("chosen_rewards", chosen_rewards_mean.to(torch.float16).mean().item()) + self.accumulative_meter.add("rejected_rewards", rejected_rewards_mean.to(torch.float16).mean().item()) + self.accumulative_meter.add("loss", loss_mean.to(torch.float16).item()) + self.accumulative_meter.add("accuracy", reward_accuracies_mean.to(torch.float16).item()) + self.accumulative_meter.add( + "margin", (chosen_rewards_mean - rejected_rewards_mean).to(torch.float16).mean().item() + ) + step_bar.update() + + msg = "Evaluation Result:\n" + for tag in ["loss", "chosen_rewards", "rejected_rewards", "accuracy", "margin"]: + msg = msg + f"{tag}: {self.accumulative_meter.get(tag)}\n" + self.coordinator.print_on_master(msg) + step_bar.close() diff --git a/applications/ColossalChat/coati/trainer/ppo.py b/applications/ColossalChat/coati/trainer/ppo.py new file mode 100755 index 000000000000..287767669516 --- /dev/null +++ b/applications/ColossalChat/coati/trainer/ppo.py @@ -0,0 +1,403 @@ +""" +PPO trainer +""" + +import os +from typing import Dict, List, Optional + +import torch +import wandb +from coati.experience_buffer import NaiveExperienceBuffer +from coati.experience_maker import Experience, NaiveExperienceMaker +from coati.models import Critic, RewardModel +from coati.models.loss import GPTLMLoss, PolicyLoss, ValueLoss +from coati.models.utils import calc_action_log_probs +from coati.trainer.callbacks import Callback +from coati.trainer.utils import all_reduce_mean +from coati.utils import AccumulativeMeanMeter, save_checkpoint +from torch.optim import Optimizer +from torch.optim.lr_scheduler import _LRScheduler +from torch.utils.data import DataLoader, DistributedSampler +from tqdm import tqdm +from transformers import PreTrainedModel, PreTrainedTokenizerBase + +from colossalai.booster import Booster +from colossalai.booster.plugin import GeminiPlugin +from colossalai.cluster import DistCoordinator +from colossalai.utils import get_current_device + +from .base import OLTrainer +from .utils import CycledDataLoader, is_rank_0, to_device + + +def _set_default_generate_kwargs(actor: PreTrainedModel) -> Dict: + """ + Set default keyword arguments for generation based on the actor model. + + Args: + actor (PreTrainedModel): The actor model. + + Returns: + Dict: A dictionary containing the default keyword arguments for generation. + """ + unwrapped_model = actor.unwrap() + new_kwargs = {} + # use huggingface models method directly + if hasattr(unwrapped_model, "prepare_inputs_for_generation"): + new_kwargs["prepare_inputs_fn"] = unwrapped_model.prepare_inputs_for_generation + + if hasattr(unwrapped_model, "_update_model_kwargs_for_generation"): + new_kwargs["update_model_kwargs_fn"] = unwrapped_model._update_model_kwargs_for_generation + return new_kwargs + + +class PPOTrainer(OLTrainer): + """ + Trainer for PPO algorithm. + + Args: + strategy (Booster): the strategy to use for training + actor (Actor): the actor model in ppo algorithm + critic (Critic): the critic model in ppo algorithm + reward_model (RewardModel): the reward model in rlhf algorithm to make reward of sentences + initial_model (Actor): the initial model in rlhf algorithm to generate reference logics to limit the update of actor + actor_optim (Optimizer): the optimizer to use for actor model + critic_optim (Optimizer): the optimizer to use for critic model + kl_coef (float, defaults to 0.1): the coefficient of kl divergence loss + train_batch_size (int, defaults to 8): the batch size to use for training + buffer_limit (int, defaults to 0): the max_size limitation of buffer + buffer_cpu_offload (bool, defaults to True): whether to offload buffer to cpu + eps_clip (float, defaults to 0.2): the clip coefficient of policy loss + vf_coef (float, defaults to 1.0): the coefficient of value loss + ptx_coef (float, defaults to 0.9): the coefficient of ptx loss + value_clip (float, defaults to 0.4): the clip coefficient of value loss + sample_buffer (bool, defaults to False): whether to sample from buffer + dataloader_pin_memory (bool, defaults to True): whether to pin memory for data loader + offload_inference_models (bool, defaults to True): whether to offload inference models to cpu during training process + callbacks (List[Callback], defaults to []): the callbacks to call during training process + generate_kwargs (dict, optional): the kwargs to use while model generating + """ + + def __init__( + self, + actor_booster: Booster, + critic_booster: Booster, + actor: PreTrainedModel, + critic: Critic, + reward_model: RewardModel, + initial_model: PreTrainedModel, + actor_optim: Optimizer, + critic_optim: Optimizer, + actor_lr_scheduler: _LRScheduler, + critic_lr_scheduler: _LRScheduler, + tokenizer: PreTrainedTokenizerBase, + kl_coef: float = 0.1, + ptx_coef: float = 0.9, + train_batch_size: int = 8, + buffer_limit: int = 0, + buffer_cpu_offload: bool = True, + eps_clip: float = 0.2, + vf_coef: float = 1.0, + value_clip: float = 0.2, + sample_buffer: bool = False, + dataloader_pin_memory: bool = True, + offload_inference_models: bool = True, + accumulation_steps: int = 1, + save_interval: int = 0, + save_dir: str = None, + use_tp: bool = False, + coordinator: DistCoordinator = None, + callbacks: List[Callback] = [], + **generate_kwargs, + ) -> None: + if isinstance(actor_booster, GeminiPlugin): + assert not offload_inference_models, "GeminiPlugin is not compatible with manual model.to('cpu')" + + data_buffer = NaiveExperienceBuffer(train_batch_size, buffer_limit, buffer_cpu_offload) + super().__init__( + actor_booster, critic_booster, data_buffer, sample_buffer, dataloader_pin_memory, callbacks=callbacks + ) + self.generate_kwargs = _set_default_generate_kwargs(actor) + self.generate_kwargs.update(generate_kwargs) + + self.actor = actor + self.critic = critic + self.actor_booster = actor_booster + self.critic_booster = critic_booster + self.actor_scheduler = actor_lr_scheduler + self.critic_scheduler = critic_lr_scheduler + self.tokenizer = tokenizer + self.experience_maker = NaiveExperienceMaker( + self.actor, self.critic, reward_model, initial_model, self.tokenizer, kl_coef + ) + self.train_batch_size = train_batch_size + + self.actor_loss_fn = PolicyLoss(eps_clip) + self.critic_loss_fn = ValueLoss(value_clip) + self.vf_coef = vf_coef + self.ptx_loss_fn = GPTLMLoss() + self.ptx_coef = ptx_coef + self.actor_optim = actor_optim + self.critic_optim = critic_optim + self.save_interval = save_interval + self.coordinator = coordinator + self.actor_save_dir = os.path.join(save_dir, "actor") + self.critic_save_dir = os.path.join(save_dir, "critic") + self.num_train_step = 0 + self.accumulation_steps = accumulation_steps + self.use_tp = use_tp + self.accumulative_meter = AccumulativeMeanMeter() + self.offload_inference_models = offload_inference_models + self.device = get_current_device() + + def _before_fit( + self, + prompt_dataloader: DataLoader, + pretrain_dataloader: Optional[DataLoader] = None, + log_dir: Optional[str] = None, + use_wandb: bool = False, + ): + """ + Args: + prompt_dataloader (DataLoader): the dataloader to use for prompt data + pretrain_dataloader (DataLoader): the dataloader to use for pretrain data + """ + self.prompt_dataloader = CycledDataLoader(prompt_dataloader) + self.pretrain_dataloader = CycledDataLoader(pretrain_dataloader) if pretrain_dataloader is not None else None + + self.writer = None + if use_wandb and is_rank_0(): + assert log_dir is not None, "log_dir must be provided when use_wandb is True" + import wandb + + self.wandb_run = wandb.init(project="Coati-ppo", sync_tensorboard=True) + if log_dir is not None and is_rank_0(): + import os + import time + + from torch.utils.tensorboard import SummaryWriter + + log_dir = os.path.join(log_dir, "ppo") + log_dir = os.path.join(log_dir, time.strftime("%Y-%m-%d_%H:%M:%S", time.localtime())) + self.writer = SummaryWriter(log_dir=log_dir) + + def _setup_update_phrase_dataload(self): + """ + why not use distributed_dataloader? + if tp is used, input on each rank is the same and we use the same dataloader to feed same experience to all ranks + if tp is not used, input on each rank is different and we expect different experiences to be fed to each rank + """ + self.dataloader = DataLoader( + self.data_buffer, + batch_size=self.train_batch_size, + shuffle=True, + drop_last=True, + pin_memory=self.dataloader_pin_memory, + collate_fn=self.data_buffer.collate_fn, + ) + + def _make_experience(self, collect_step: int) -> Experience: + """ + Make experience + """ + prompts = self.prompt_dataloader.next() + if self.offload_inference_models: + # TODO(ver217): this may be controlled by strategy if they are prepared by strategy + self.experience_maker.initial_model.to(self.device) + self.experience_maker.reward_model.to(self.device) + return self.experience_maker.make_experience( + input_ids=prompts["input_ids"].to(get_current_device()), + attention_mask=prompts["attention_mask"].to(get_current_device()), + **self.generate_kwargs, + ) + + def _training_step(self, experience: Experience): + """ + Args: + experience: + sequences: [batch_size, prompt_length + response_length] --- ............ + """ + self.num_train_step += 1 + self.actor.train() + self.critic.train() + num_actions = experience.action_log_probs.size(1) + # policy loss + + actor_logits = self.actor(input_ids=experience.sequences, attention_mask=experience.attention_mask)[ + "logits" + ] # [batch size, prompt_length + response_length] + action_log_probs = calc_action_log_probs(actor_logits, experience.sequences, num_actions) + + actor_loss, to_skip, max_ratio = self.actor_loss_fn( + action_log_probs, experience.action_log_probs, experience.advantages, action_mask=experience.action_mask + ) + actor_loss = (1 - self.ptx_coef) * actor_loss + if not to_skip: + self.actor_booster.backward(loss=actor_loss, optimizer=self.actor_optim) + + # ptx loss + if self.ptx_coef != 0: + batch = self.pretrain_dataloader.next() + batch = to_device(batch, self.device) + outputs = self.actor(batch["input_ids"], attention_mask=batch["attention_mask"], labels=batch["labels"]) + ptx_loss = outputs.loss + ptx_loss = self.ptx_coef * ptx_loss + self.actor_booster.backward(loss=ptx_loss, optimizer=self.actor_optim) + + # value loss + values = self.critic( + input_ids=experience.sequences, attention_mask=experience.attention_mask + ) # [batch size, prompt_length + response_length] + critic_loss = self.critic_loss_fn( + values[:, -num_actions:], experience.values, experience.advantages, action_mask=experience.action_mask + ) + critic_loss = critic_loss * self.vf_coef + self.critic_booster.backward(loss=critic_loss, optimizer=self.critic_optim) + + # sync + actor_loss_mean = all_reduce_mean(tensor=actor_loss) + critic_loss_mean = all_reduce_mean(tensor=critic_loss) + max_ratio_mean = all_reduce_mean(tensor=max_ratio) + reward_mean = all_reduce_mean(tensor=experience.reward.mean()) + value_mean = all_reduce_mean(tensor=experience.values.mean()) + advantages_mean = all_reduce_mean(tensor=experience.advantages.mean()) + kl_mean = all_reduce_mean(tensor=experience.kl.mean()) + if self.ptx_coef != 0: + ptx_loss_mean = all_reduce_mean(tensor=ptx_loss) + + self.accumulative_meter.add("actor_loss", actor_loss_mean.to(torch.float16).mean().item()) + self.accumulative_meter.add("critic_loss", critic_loss_mean.to(torch.float16).mean().item()) + self.accumulative_meter.add("max_ratio", max_ratio_mean.to(torch.float16).item()) + self.accumulative_meter.add("reward", reward_mean.to(torch.float16).mean().item()) + self.accumulative_meter.add("value", value_mean.to(torch.float16).mean().item()) + self.accumulative_meter.add("advantages", advantages_mean.to(torch.float16).item()) + self.accumulative_meter.add("skip_ratio", 1.0 if to_skip else 0.0) + self.accumulative_meter.add("kl", kl_mean.to(torch.float16).item()) + if self.ptx_coef != 0: + self.accumulative_meter.add("ptx_loss", ptx_loss_mean.to(torch.float16).mean().item()) + + if self.num_train_step % self.accumulation_steps == self.accumulation_steps - 1: + self.actor_optim.step() + self.critic_optim.step() + self.actor_optim.zero_grad() + self.critic_optim.zero_grad() + self.actor_scheduler.step() + self.critic_scheduler.step() + + # preparing logging model output and corresponding rewards. + if self.num_train_step % 10 == 1: + response_text = self.experience_maker.tokenizer.batch_decode( + experience.sequences, skip_special_tokens=True + ) + for i in range(len(response_text)): + response_text[i] = response_text[i] + f"\n\nReward: {experience.reward[i]}" + + if self.writer and is_rank_0() and "wandb_run" in self.__dict__: + # log output to wandb + my_table = wandb.Table( + columns=[f"sample response {i}" for i in range(len(response_text))], data=[response_text] + ) + try: + self.wandb_run.log({"sample_response": my_table}) + except OSError as e: + self.coordinator.print_on_master(e) + elif self.writer and is_rank_0(): + for line in response_text: + self.coordinator.print_on_master(line) + + if self.writer and is_rank_0(): + self.writer.add_scalar("train/max_ratio", self.accumulative_meter.get("max_ratio"), self.num_train_step) + self.writer.add_scalar( + "train/skip_ratio", self.accumulative_meter.get("skip_ratio"), self.num_train_step + ) + self.writer.add_scalar( + "train/actor_loss", self.accumulative_meter.get("actor_loss"), self.num_train_step + ) + self.writer.add_scalar("train/lr_actor", self.actor_optim.param_groups[0]["lr"], self.num_train_step) + self.writer.add_scalar("train/lr_critic", self.critic_optim.param_groups[0]["lr"], self.num_train_step) + self.writer.add_scalar( + "train/critic_loss", self.accumulative_meter.get("critic_loss"), self.num_train_step + ) + if self.ptx_coef != 0: + self.writer.add_scalar( + "train/ptx_loss", self.accumulative_meter.get("ptx_loss"), self.num_train_step + ) + self.writer.add_scalar("reward", self.accumulative_meter.get("reward"), self.num_train_step) + self.writer.add_scalar("approx_kl", self.accumulative_meter.get("kl"), self.num_train_step) + self.writer.add_scalar("value", self.accumulative_meter.get("value"), self.num_train_step) + self.writer.add_scalar("advantages", self.accumulative_meter.get("advantages"), self.num_train_step) + self.accumulative_meter.reset() + + def _learn(self, update_step: int): + """ + Perform the learning step of the PPO algorithm. + + Args: + update_step (int): The current update step. + + Returns: + None + """ + if self.offload_inference_models: + self.experience_maker.initial_model.to("cpu") + self.experience_maker.reward_model.to("cpu") + + # buffer may be empty at first, we should rebuild at each training + if self.sample_buffer: + experience = self.data_buffer.sample() + self._on_learn_batch_start() + experience.to_device(self.device) + self._training_step(experience) + self._on_learn_batch_end(experience) + else: + if isinstance(self.dataloader.sampler, DistributedSampler): + self.dataloader.sampler.set_epoch(update_step) + pbar = tqdm(self.dataloader, desc=f"Train epoch [{update_step + 1}]", disable=not is_rank_0()) + for experience in pbar: + self._on_learn_batch_start() + experience.to_device(self.device) + self._training_step(experience) + self._on_learn_batch_end(experience) + + def _save_checkpoint(self, episode: int = 0): + """ + Save the actor and critic checkpoints with running states. + + Args: + episode (int): The current episode number. + + Returns: + None + """ + + self.coordinator.print_on_master("\nStart saving actor checkpoint with running states") + save_checkpoint( + save_dir=self.actor_save_dir, + booster=self.actor_booster, + model=self.actor, + optimizer=self.actor_optim, + lr_scheduler=self.actor_scheduler, + epoch=0, + step=episode + 1, + batch_size=self.train_batch_size, + coordinator=self.coordinator, + ) + self.coordinator.print_on_master( + f"Saved actor checkpoint at episode {(episode + 1)} at folder {self.actor_save_dir}" + ) + + self.coordinator.print_on_master("\nStart saving critic checkpoint with running states") + save_checkpoint( + save_dir=self.critic_save_dir, + booster=self.critic_booster, + model=self.critic, + optimizer=self.critic_optim, + lr_scheduler=self.critic_scheduler, + epoch=0, + step=episode + 1, + batch_size=self.train_batch_size, + coordinator=self.coordinator, + ) + self.coordinator.print_on_master( + f"Saved critic checkpoint at episode {(episode + 1)} at folder {self.critic_save_dir}" + ) diff --git a/applications/ColossalChat/coati/trainer/rm.py b/applications/ColossalChat/coati/trainer/rm.py new file mode 100755 index 000000000000..0fb714a62bce --- /dev/null +++ b/applications/ColossalChat/coati/trainer/rm.py @@ -0,0 +1,242 @@ +""" +Reward model trianer +""" + +import os +from typing import Any, Callable, Optional + +import torch +import tqdm +from coati.models import LogSigLoss +from coati.trainer.utils import all_reduce_mean +from coati.utils import AccumulativeMeanMeter, save_checkpoint +from torch.optim import Optimizer +from torch.optim.lr_scheduler import _LRScheduler +from torch.utils.data import DataLoader +from transformers import PreTrainedTokenizerBase + +from colossalai.booster import Booster +from colossalai.cluster import DistCoordinator +from colossalai.utils import get_current_device + +from .base import SLTrainer +from .utils import is_rank_0, to_device + + +class RewardModelTrainer(SLTrainer): + """ + Trainer for PPO algorithm. + + Args: + actor (Actor): the actor model in ppo algorithm + ref_model (Critic): the reference model in ppo algorithm + booster (Strategy): the strategy to use for training + actor_optim (Optimizer): the optimizer to use for actor model + actor_lr_scheduler (_LRScheduler): the lr scheduler to use for actor model + tokenizer (PreTrainedTokenizerBase): the tokenizer to use for encoding + max_epochs (int, defaults to 1): the max number of epochs to train + beta (float, defaults to 0.1): the beta parameter in dpo loss + accumulation_steps (int): the number of steps to accumulate gradients + start_epoch (int, defaults to 0): the start epoch, non-zero if resumed from a checkpoint + save_interval (int): the interval to save model checkpoints, default to 0, which means no checkpoint will be saved during trainning + save_dir (str): the directory to save checkpoints + coordinator (DistCoordinator): the coordinator to use for distributed logging + """ + + def __init__( + self, + model: Any, + booster: Booster, + optimizer: Optimizer, + lr_scheduler: _LRScheduler, + tokenizer: PreTrainedTokenizerBase, + loss_fn: Optional[Callable] = None, + max_epochs: int = 1, + beta: float = 0.1, + accumulation_steps: int = 1, + start_epoch: int = 0, + save_interval: int = 0, + save_dir: str = None, + coordinator: DistCoordinator = None, + ) -> None: + super().__init__(booster, max_epochs=max_epochs, model=model, optimizer=optimizer, start_epoch=start_epoch) + self.actor_scheduler = lr_scheduler + self.tokenizer = tokenizer + self.loss_fn = loss_fn if loss_fn is not None else LogSigLoss(beta=beta) + self.save_interval = save_interval + self.coordinator = coordinator + self.save_dir = save_dir + self.num_train_step = 0 + self.accumulation_steps = accumulation_steps + self.device = get_current_device() + self.accumulative_meter = AccumulativeMeanMeter() + + def _before_fit( + self, + train_preference_dataloader: DataLoader = None, + eval_preference_dataloader: DataLoader = None, + log_dir: Optional[str] = None, + use_wandb: bool = False, + ): + """ + Args: + prompt_dataloader (DataLoader): the dataloader to use for prompt data + pretrain_dataloader (DataLoader): the dataloader to use for pretrain data + """ + self.train_dataloader = train_preference_dataloader + self.eval_dataloader = eval_preference_dataloader + self.writer = None + if use_wandb and is_rank_0(): + assert log_dir is not None, "log_dir must be provided when use_wandb is True" + import wandb + + self.wandb_run = wandb.init(project="Coati-rm", sync_tensorboard=True) + if log_dir is not None and is_rank_0(): + import os + import time + + from torch.utils.tensorboard import SummaryWriter + + log_dir = os.path.join(log_dir, "rm") + log_dir = os.path.join(log_dir, time.strftime("%Y-%m-%d_%H:%M:%S", time.localtime())) + self.writer = SummaryWriter(log_dir=log_dir) + + def _train(self, epoch): + self.model.train() + step_bar = tqdm.trange( + len(self.train_dataloader) // self.accumulation_steps, + desc=f"Epoch {epoch + 1}/{self.max_epochs}", + disable=not is_rank_0(), + ) + for i, batch in enumerate(self.train_dataloader): + batch = to_device(batch, self.device) + + ( + chosen_input_ids, + chosen_attention_mask, + reject_input_ids, + reject_attention_mask, + ) = ( + batch["chosen_input_ids"], + batch["chosen_attention_mask"], + batch["reject_input_ids"], + batch["reject_attention_mask"], + ) + batch_size = chosen_input_ids.size()[0] + + # Concatenate for better parrallelism + reward = self.model( + torch.cat([chosen_input_ids, reject_input_ids], dim=0), + attention_mask=torch.cat([chosen_attention_mask, reject_attention_mask], dim=0), + ) + chosen_reward = reward[:batch_size] + reject_reward = reward[batch_size:] + loss = self.loss_fn(chosen_reward, reject_reward).mean() + + self.booster.backward(loss=loss, optimizer=self.optimizer) + + accuracy = (chosen_reward > reject_reward).float() + + # Sync + loss_mean = all_reduce_mean(tensor=loss) + chosen_rewards_mean = all_reduce_mean(tensor=chosen_reward) + rejected_rewards_mean = all_reduce_mean(tensor=reject_reward) + accuracy_mean = all_reduce_mean(tensor=accuracy) + self.accumulative_meter.add("chosen_rewards", chosen_rewards_mean.to(torch.float16).mean().item()) + self.accumulative_meter.add("rejected_rewards", rejected_rewards_mean.to(torch.float16).mean().item()) + self.accumulative_meter.add("loss", loss_mean.to(torch.float16).item()) + self.accumulative_meter.add("accuracy", accuracy_mean.mean().to(torch.float16).item()) + + if (i + 1) % self.accumulation_steps == 0: + self.optimizer.step() + self.optimizer.zero_grad() + self.actor_scheduler.step() + step_bar.update() + self.num_train_step += 1 + + # Logging + if self.writer and is_rank_0(): + self.writer.add_scalar("train/loss", self.accumulative_meter.get("loss"), self.num_train_step) + self.writer.add_scalar("train/lr", self.optimizer.param_groups[0]["lr"], self.num_train_step) + self.writer.add_scalar( + "train/dist", + self.accumulative_meter.get("chosen_rewards") - self.accumulative_meter.get("rejected_rewards"), + self.num_train_step, + ) + self.writer.add_scalar( + "train/reward_chosen", self.accumulative_meter.get("chosen_rewards"), self.num_train_step + ) + self.writer.add_scalar( + "train/reward_reject", self.accumulative_meter.get("rejected_rewards"), self.num_train_step + ) + self.writer.add_scalar("train/acc", self.accumulative_meter.get("accuracy"), self.num_train_step) + + self.accumulative_meter.reset() + + # Save checkpoint + if self.save_interval > 0 and (self.num_train_step + 1) % self.save_interval == 0: + self.coordinator.print_on_master("\nStart saving model checkpoint with running states") + save_checkpoint( + save_dir=self.save_dir, + booster=self.booster, + model=self.model, + optimizer=self.optimizer, + lr_scheduler=self.actor_scheduler, + epoch=epoch, + step=i + 1, + batch_size=batch_size, + coordinator=self.coordinator, + ) + self.coordinator.print_on_master( + f"Saved checkpoint at epoch {epoch} step {(i + 1)/self.accumulation_steps} at folder {self.save_dir}" + ) + step_bar.close() + + def _eval(self, epoch): + if self.eval_dataloader is None: + self.coordinator.print_on_master("No eval dataloader is provided, skip evaluation") + return + self.model.eval() + step_bar = tqdm.trange( + len(self.eval_dataloader), desc=f"Epoch {epoch + 1}/{self.max_epochs}", disable=not is_rank_0() + ) + with torch.no_grad(): + for i, batch in enumerate(self.eval_dataloader): + batch = to_device(batch, self.device) + ( + chosen_input_ids, + chosen_attention_mask, + reject_input_ids, + reject_attention_mask, + ) = ( + batch["chosen_input_ids"], + batch["chosen_attention_mask"], + batch["reject_input_ids"], + batch["reject_attention_mask"], + ) + + chosen_reward = self.model(chosen_input_ids, attention_mask=chosen_attention_mask) + reject_reward = self.model(reject_input_ids, attention_mask=reject_attention_mask) + loss = self.loss_fn(chosen_reward, reject_reward).mean() + + # Sync + loss_mean = all_reduce_mean(tensor=loss) + chosen_rewards_mean = all_reduce_mean(tensor=chosen_reward) + rejected_rewards_mean = all_reduce_mean(tensor=reject_reward) + self.accumulative_meter.add("chosen_rewards", chosen_rewards_mean.to(torch.float16).mean().item()) + self.accumulative_meter.add("rejected_rewards", rejected_rewards_mean.to(torch.float16).mean().item()) + self.accumulative_meter.add("loss", loss_mean.to(torch.float16).item()) + + step_bar.update() + + msg = "Evaluation Result:\n" + for tag in ["loss", "chosen_rewards", "rejected_rewards"]: + msg = msg + f"{tag}: {self.accumulative_meter.get(tag)}\n" + msg = ( + msg + + f"distance: {self.accumulative_meter.get('chosen_rewards')-self.accumulative_meter.get('rejected_rewards')}\n" + ) + self.coordinator.print_on_master(msg) + with open(os.path.join(self.save_dir, f"eval_result_epoch{epoch}.txt"), "w") as f: + f.write(msg) + step_bar.close() diff --git a/applications/ColossalChat/coati/trainer/sft.py b/applications/ColossalChat/coati/trainer/sft.py new file mode 100755 index 000000000000..c95f5b65a822 --- /dev/null +++ b/applications/ColossalChat/coati/trainer/sft.py @@ -0,0 +1,170 @@ +""" +SFT trainer +""" + +import os +from typing import Optional + +import torch +from coati.trainer.utils import all_reduce_mean +from coati.utils import AccumulativeMeanMeter, save_checkpoint +from torch.optim import Optimizer +from torch.optim.lr_scheduler import _LRScheduler +from torch.utils.data import DataLoader +from tqdm import trange + +from colossalai.booster import Booster +from colossalai.cluster import DistCoordinator + +from .base import SLTrainer +from .utils import is_rank_0, to_device + + +class SFTTrainer(SLTrainer): + """ + Trainer to use while training reward model. + + Args: + model (torch.nn.Module): the model to train + strategy (Strategy): the strategy to use for training + optim(Optimizer): the optimizer to use for training + lr_scheduler(_LRScheduler): the lr scheduler to use for training + max_epochs (int, defaults to 2): the number of epochs to train + accumulation_steps (int, defaults to 8): the number of steps to accumulate gradients + """ + + def __init__( + self, + model, + booster: Booster, + optim: Optimizer, + lr_scheduler: _LRScheduler, + max_epochs: int = 2, + accumulation_steps: int = 8, + start_epoch=0, + save_interval: int = None, + save_dir: str = None, + coordinator: Optional[DistCoordinator] = None, + ) -> None: + super().__init__(booster, max_epochs, model, optim, start_epoch=start_epoch) + + self.accumulation_steps = accumulation_steps + self.scheduler = lr_scheduler + self.save_interval = save_interval + self.save_dir = save_dir + self.coordinator = coordinator + self.num_train_step = 0 + self.num_eval_step = 0 + self.accumulative_meter = AccumulativeMeanMeter() + + def _before_fit( + self, + train_dataloader: DataLoader, + eval_dataloader: Optional[DataLoader] = None, + log_dir: Optional[str] = None, + use_wandb: bool = False, + ): + """ + Args: + train_dataloader: the dataloader to use for training + eval_dataloader: the dataloader to use for evaluation + log_dir: the directory to save logs + use_wandb: whether to use wandb for logging + """ + self.train_dataloader = train_dataloader + self.eval_dataloader = eval_dataloader + + self.writer = None + if use_wandb and is_rank_0(): + assert log_dir is not None, "log_dir must be provided when use_wandb is True" + import wandb + + wandb.init(project="Coati-sft", sync_tensorboard=True) + if log_dir is not None and is_rank_0(): + import os + import time + + from torch.utils.tensorboard import SummaryWriter + + log_dir = os.path.join(log_dir, "sft") + log_dir = os.path.join(log_dir, time.strftime("%Y-%m-%d_%H:%M:%S", time.localtime())) + self.writer = SummaryWriter(log_dir=log_dir) + + def _train(self, epoch: int): + self.model.train() + step_bar = trange( + len(self.train_dataloader) // self.accumulation_steps, + desc=f"Epoch {epoch + 1}/{self.max_epochs}", + disable=not is_rank_0(), + ) + for i, batch in enumerate(self.train_dataloader): + batch = to_device(batch, torch.cuda.current_device()) + batch_size = batch["input_ids"].size(0) + outputs = self.model(batch["input_ids"], attention_mask=batch["attention_mask"], labels=batch["labels"]) + loss = outputs.loss + self.booster.backward(loss=loss, optimizer=self.optimizer) + + loss_mean = all_reduce_mean(tensor=loss) + self.accumulative_meter.add("loss", loss_mean.to(torch.float16).item()) + + # Gradient accumulation + if (i + 1) % self.accumulation_steps == 0: + self.optimizer.step() + self.optimizer.zero_grad() + self.scheduler.step() + + if self.writer: + self.writer.add_scalar("train/loss", self.accumulative_meter.get("loss"), self.num_train_step) + self.writer.add_scalar("train/lr", self.scheduler.get_last_lr()[0], self.num_train_step) + self.num_train_step += 1 + self.accumulative_meter.reset() + step_bar.update() + + # Save checkpoint + if ( + self.save_dir is not None + and self.save_interval is not None + and (self.num_train_step + 1) % self.save_interval == 0 + ): + save_checkpoint( + save_dir=self.save_dir, + booster=self.booster, + model=self.model, + optimizer=self.optimizer, + lr_scheduler=self.scheduler, + epoch=epoch, + step=self.num_train_step + 1, + batch_size=batch_size, + coordinator=self.coordinator, + ) + self.coordinator.print_on_master( + f"Saved checkpoint at epoch {epoch} step {self.num_train_step} at folder {self.save_dir}" + ) + step_bar.close() + + def _eval(self, epoch: int): + if self.eval_dataloader is None: + self.coordinator.print_on_master("No eval dataloader is provided, skip evaluation") + return + self.accumulative_meter.reset() + self.model.eval() + with torch.no_grad(): + step_bar = trange( + len(self.eval_dataloader), + desc=f"Epoch {epoch + 1}/{self.max_epochs}", + disable=not is_rank_0(), + ) + for batch in self.eval_dataloader: + batch = to_device(batch, torch.cuda.current_device()) + outputs = self.model(batch["input_ids"], attention_mask=batch["attention_mask"], labels=batch["labels"]) + loss_mean = all_reduce_mean(tensor=outputs.loss) + self.accumulative_meter.add("loss", loss_mean.item(), count_update=batch["input_ids"].size(0)) + step_bar.update() + loss_mean = self.accumulative_meter.get("loss") + msg = "Evaluation Result:\n" + for tag in ["loss"]: + msg = msg + f"{tag}: {self.accumulative_meter.get(tag)}\n" + self.coordinator.print_on_master(msg) + with open(os.path.join(self.save_dir, f"eval_result_epoch{epoch}.txt"), "w") as f: + f.write(msg) + step_bar.close() diff --git a/applications/ColossalChat/coati/trainer/utils.py b/applications/ColossalChat/coati/trainer/utils.py new file mode 100755 index 000000000000..5ce1e9ef009c --- /dev/null +++ b/applications/ColossalChat/coati/trainer/utils.py @@ -0,0 +1,113 @@ +""" +Training utilities for Coati. +""" +from typing import Any + +import torch +import torch.distributed as dist +from torch.utils._pytree import tree_map +from torch.utils.data import DataLoader + + +class CycledDataLoader: + """ + A data loader that cycles through the data when it reaches the end. + + Args: + dataloader (DataLoader): The original data loader. + + Attributes: + dataloader (DataLoader): The original data loader. + count (int): The number of times the data loader has been cycled. + dataloader_iter (iterable): The iterator for the data loader. + + Methods: + next(): Returns the next batch of data from the data loader, cycling through the data if necessary. + """ + + def __init__( + self, + dataloader: DataLoader, + ) -> None: + self.dataloader = dataloader + + self.count = 0 + self.dataloader_iter = None + + def next(self): + """ + Returns the next batch of data from the data loader, cycling through the data if necessary. + + Returns: + Any: The next batch of data from the data loader. + """ + # defer initialization + if self.dataloader_iter is None: + self.dataloader_iter = iter(self.dataloader) + + self.count += 1 + try: + return next(self.dataloader_iter) + except StopIteration: + self.count = 0 + self.dataloader_iter = iter(self.dataloader) + return next(self.dataloader_iter) + + +def is_rank_0() -> bool: + """ + Check if the current process is the rank 0 process in a distributed training setup. + + Returns: + bool: True if the current process is the rank 0 process, False otherwise. + """ + return not dist.is_initialized() or dist.get_rank() == 0 + + +def to_device(x: Any, device: torch.device) -> Any: + """ + Move the input tensor or nested structure of tensors to the specified device. + + Args: + x (Any): The input tensor or nested structure of tensors. + device (torch.device): The target device to move the tensors to. + + Returns: + Any: The tensor or nested structure of tensors moved to the target device. + """ + + def _to(t: Any): + if isinstance(t, torch.Tensor): + return t.to(device) + return t + + return tree_map(_to, x) + + +def all_reduce_mean(tensor: torch.Tensor) -> torch.Tensor: + """ + Perform all-reduce operation on the given tensor and compute the mean across all processes. + + Args: + tensor (torch.Tensor): The input tensor to be reduced. + + Returns: + torch.Tensor: The reduced tensor with mean computed across all processes. + """ + dist.all_reduce(tensor=tensor, op=dist.ReduceOp.SUM) + tensor.div_(dist.get_world_size()) + return tensor + + +def all_reduce_sum(tensor: torch.Tensor) -> torch.Tensor: + """ + Performs an all-reduce operation to sum the values of the given tensor across all processes. + + Args: + tensor (torch.Tensor): The input tensor to be reduced. + + Returns: + torch.Tensor: The reduced tensor with the sum of values across all processes. + """ + dist.all_reduce(tensor=tensor, op=dist.ReduceOp.SUM) + return tensor diff --git a/applications/ColossalChat/coati/utils/__init__.py b/applications/ColossalChat/coati/utils/__init__.py new file mode 100755 index 000000000000..5d651b202cc8 --- /dev/null +++ b/applications/ColossalChat/coati/utils/__init__.py @@ -0,0 +1,4 @@ +from .accumulative_meter import AccumulativeMeanMeter +from .ckpt_io import load_checkpoint, save_checkpoint + +__all__ = ["load_checkpoint", "save_checkpoint", "AccumulativeMeanMeter"] diff --git a/applications/ColossalChat/coati/utils/accumulative_meter.py b/applications/ColossalChat/coati/utils/accumulative_meter.py new file mode 100755 index 000000000000..9ed662d2d25a --- /dev/null +++ b/applications/ColossalChat/coati/utils/accumulative_meter.py @@ -0,0 +1,69 @@ +""" +A class that can be used to calculate the mean of a variable +""" + + +class AccumulativeMeanVariable: + """ + A class that calculates the accumulative mean of a variable. + """ + + def __init__(self): + self._sum = 0 + self._count = 0 + + def add(self, value, count_update=1): + """ + Adds a value to the sum and updates the count. + + Args: + value (float): The value to be added. + count_update (int, optional): The amount to update the count by. Defaults to 1. + """ + self._sum += value + self._count += count_update + + def get(self): + """ + Calculates and returns the accumulative mean. + + Returns: + float: The accumulative mean. + """ + return self._sum / self._count if self._count > 0 else 0 + + def reset(self): + """ + Resets the sum and count to zero. + """ + self._sum = 0 + self._count = 0 + + +class AccumulativeMeanMeter: + """ + A class for calculating and storing the accumulative mean of variables. + + Attributes: + variable_dict (dict): A dictionary to store the accumulative mean variables. + + Methods: + add(name, value, count_update=1): Adds a value to the specified variable. + get(name): Retrieves the accumulative mean value of the specified variable. + reset(): Resets all the accumulative mean variables to their initial state. + """ + + def __init__(self): + self.variable_dict = {} + + def add(self, name, value, count_update=1): + if name not in self.variable_dict: + self.variable_dict[name] = AccumulativeMeanVariable() + self.variable_dict[name].add(value, count_update=count_update) + + def get(self, name): + return self.variable_dict[name].get() + + def reset(self): + for name in self.variable_dict: + self.variable_dict[name].reset() diff --git a/applications/ColossalChat/coati/utils/ckpt_io.py b/applications/ColossalChat/coati/utils/ckpt_io.py new file mode 100755 index 000000000000..5b804f0acc14 --- /dev/null +++ b/applications/ColossalChat/coati/utils/ckpt_io.py @@ -0,0 +1,93 @@ +#!/usr/bin/env python3 +# -*- coding: utf-8 -*- + +""" +Helper functions for IO save load checkpoints +""" + +import json +import os +from typing import Any, Dict, Tuple, Union + +import torch +from torch.optim.lr_scheduler import _LRScheduler +from torch.optim.optimizer import Optimizer + +from colossalai.booster import Booster +from colossalai.cluster import DistCoordinator + + +def load_json(file_path: Union[str, os.PathLike]) -> Dict[str, Any]: + """ + Load file in JSON format + """ + with open(file=file_path, mode="r", encoding="utf-8") as fp: + return json.load(fp) + + +def save_json(data: Dict[str, Any], file_path: Union[str, os.PathLike]) -> None: + """ + Save as JSON format + """ + with open(file=file_path, mode="w", encoding="utf-8") as fp: + json.dump(data, fp=fp, ensure_ascii=False, indent=4) + + +def save_checkpoint( + save_dir: Union[str, os.PathLike], + booster: Booster, + model: torch.nn.Module, + optimizer: Optimizer, + lr_scheduler: _LRScheduler, + epoch: int, + step: int, + batch_size: int, + coordinator: DistCoordinator, +) -> None: + """ + Save model checkpoint, optimizer, LR scheduler and intermedidate running states. + """ + + save_dir = os.path.join(save_dir, f"epoch-{epoch}_step-{step}") + os.makedirs(os.path.join(save_dir, "modeling"), exist_ok=True) + + booster.save_model(model, os.path.join(save_dir, "modeling"), shard=True) + + """ + Temporary disable the following as save_optimizer causes all processes to hang in a multi-gpu environment, + working on fixing this bug + """ + + booster.save_optimizer(optimizer, os.path.join(save_dir, "optimizer"), shard=True) + booster.save_lr_scheduler(lr_scheduler, os.path.join(save_dir, "lr_scheduler")) + running_states = { + "epoch": epoch, + "step": step, + "sample_start_index": step * batch_size, + } + if coordinator.is_master(): + save_json(running_states, os.path.join(save_dir, "running_states.json")) + + +def load_checkpoint( + load_dir: Union[str, os.PathLike], + booster: Booster, + model: torch.nn.Module, + optimizer: Optimizer, + lr_scheduler: _LRScheduler, +) -> Tuple[int, int, int]: + """ + Load model checkpoint, optimizer, LR scheduler and intermedidate running states. + """ + + # Update booster params states. + booster.load_model(model=model, checkpoint=os.path.join(load_dir, "modeling")) + booster.load_optimizer(optimizer=optimizer, checkpoint=os.path.join(load_dir, "optimizer")) + booster.load_lr_scheduler(lr_scheduler=lr_scheduler, checkpoint=os.path.join(load_dir, "lr_scheduler")) + + running_states = load_json(file_path=os.path.join(load_dir, "running_states.json")) + return ( + running_states["epoch"], + running_states["step"], + running_states["sample_start_index"], + ) diff --git a/applications/ColossalChat/config/conversation_template/Qwen.json b/applications/ColossalChat/config/conversation_template/Qwen.json new file mode 100644 index 000000000000..09f706ffed90 --- /dev/null +++ b/applications/ColossalChat/config/conversation_template/Qwen.json @@ -0,0 +1,7 @@ +{ + "chat_template": "{% for message in messages %}{{'<|im_start|>' + message['role'] + '\n' + message['content'] + '<|im_end|>' + '\n'}}{% endfor %}{% if add_generation_prompt %}{{ '<|im_start|>assistant\n' }}{% endif %}", + "system_message": "A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions.\n\n", + "stop_ids": [ + null + ] +} diff --git a/applications/ColossalChat/config/conversation_template/Vicuna.json b/applications/ColossalChat/config/conversation_template/Vicuna.json new file mode 100644 index 000000000000..2b00b6529720 --- /dev/null +++ b/applications/ColossalChat/config/conversation_template/Vicuna.json @@ -0,0 +1,7 @@ +{ + "chat_template": "{% if messages[0]['role'] == 'system' %}{% set loop_messages = messages[1:] %}{% set system_message = messages[0]['content'] %}{% elif true == true and not '<>' in messages[0]['content'] %}{% set loop_messages = messages %}{% set system_message = 'You are a helpful, respectful and honest assistant. Always answer as helpfully as possible, while being safe. Your answers should not include any harmful, unethical, racist, sexist, toxic, dangerous, or illegal content. Please ensure that your responses are socially unbiased and positive in nature.\\n\\nIf a question does not make any sense, or is not factually coherent, explain why instead of answering something not correct. If you don\\'t know the answer to a question, please don\\'t share false information.' %}{% else %}{% set loop_messages = messages %}{% set system_message = false %}{% endif %}{% for message in loop_messages %}{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}{% endif %}{% if loop.index0 == 0 and system_message != false %}{% set content = '<>\\n' + system_message + '\\n<>\\n\\n' + message['content'] %}{% else %}{% set content = message['content'] %}{% endif %}{% if message['role'] == 'user' %}{{ bos_token + '[INST] ' + content.strip() + ' [/INST]' }}{% elif message['role'] == 'system' %}{{ '<>\\n' + content.strip() + '\\n<>\\n\\n' }}{% elif message['role'] == 'assistant' %}{{ ' ' + content.strip() + ' ' + eos_token }}{% endif %}{% endfor %}", + "system_message": "A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions.\n\n", + "stop_ids": [ + 2 + ] +} diff --git a/applications/ColossalChat/config/conversation_template/Yi.json b/applications/ColossalChat/config/conversation_template/Yi.json new file mode 100644 index 000000000000..9716413b53ad --- /dev/null +++ b/applications/ColossalChat/config/conversation_template/Yi.json @@ -0,0 +1,7 @@ +{ + "chat_template": "{% if not add_generation_prompt is defined %}{% set add_generation_prompt = false %}{% endif %}{% for message in messages %}{{'<|im_start|>' + message['role'] + '\n' + message['content'] + '<|im_end|>' + '\n'}}{% endfor %}{% if add_generation_prompt %}{{ '<|im_start|>assistant\n' }}{% endif %}", + "system_message": "A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions.\n\n", + "stop_ids": [ + 2 + ] +} diff --git a/applications/ColossalChat/config/conversation_template/chatGLM2.json b/applications/ColossalChat/config/conversation_template/chatGLM2.json new file mode 100644 index 000000000000..a2638dbe7439 --- /dev/null +++ b/applications/ColossalChat/config/conversation_template/chatGLM2.json @@ -0,0 +1,7 @@ +{ + "chat_template": "{% for message in messages %}{{'<|im_start|>' + message['role'] + '\n' + message['content'] + '<|im_end|>' + '\n'}}{% endfor %}{% if add_generation_prompt %}{{ '<|im_start|>assistant\n' }}{% endif %}", + "system_message": "A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions.\n\n", + "stop_ids": [ + 2 + ] +} diff --git a/applications/ColossalChat/config/conversation_template/colossal-llama2.json b/applications/ColossalChat/config/conversation_template/colossal-llama2.json new file mode 100644 index 000000000000..cc7f1e5d76fc --- /dev/null +++ b/applications/ColossalChat/config/conversation_template/colossal-llama2.json @@ -0,0 +1,7 @@ +{ + "chat_template": "{% if not add_generation_prompt is defined %}{% set add_generation_prompt = false %}{% endif %}{% for message in messages %}{% if message['role'] == 'user' %}{{'Human: ' + bos_token + message['content'].strip() + eos_token }}{% elif message['role'] == 'system' %}{{ message['content'].strip() + '\\n\\n' }}{% elif message['role'] == 'assistant' %}{{ 'Assistant: ' + bos_token + message['content'].strip() + eos_token }}{% endif %}{% endfor %}{% if add_generation_prompt %}{{ 'Assistant: ' + bos_token }}{% endif %}", + "system_message": "A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions.\n\n", + "stop_ids": [ + 2 + ] +} diff --git a/applications/ColossalChat/config/conversation_template/llama2.json b/applications/ColossalChat/config/conversation_template/llama2.json new file mode 100644 index 000000000000..80558f976e3b --- /dev/null +++ b/applications/ColossalChat/config/conversation_template/llama2.json @@ -0,0 +1,7 @@ +{ + "chat_template": "{% if messages[0]['role'] == 'system' %}{% set loop_messages = messages[1:] %}{% set system_message = messages[0]['content'] %}{% else %}{% set loop_messages = messages %}{% set system_message = false %}{% endif %}{% for message in loop_messages %}{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}{% endif %}{% if loop.index0 == 0 and system_message != false %}{% set content = '<>\\n' + system_message + '\\n<>\\n\\n' + message['content'] %}{% else %}{% set content = message['content'] %}{% endif %}{% if message['role'] == 'user' %}{{ bos_token + '[INST] ' + content.strip() + ' [/INST]' }}{% elif message['role'] == 'assistant' %}{{ ' ' + content.strip() + ' ' + eos_token }}{% endif %}{% endfor %}", + "system_message": "A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions.\n\n", + "stop_ids": [ + 2 + ] +} diff --git a/applications/ColossalChat/config/conversation_template/mistral.json b/applications/ColossalChat/config/conversation_template/mistral.json new file mode 100644 index 000000000000..b48c3a3f27af --- /dev/null +++ b/applications/ColossalChat/config/conversation_template/mistral.json @@ -0,0 +1,7 @@ +{ + "chat_template": "{{ bos_token }}{% for message in messages %}{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}{% endif %}{% if message['role'] == 'user' %}{{ '[INST] ' + message['content'] + ' [/INST]' }}{% elif message['role'] == 'assistant' %}{{ message['content'] + eos_token}}{% else %}{{ raise_exception('Only user and assistant roles are supported!') }}{% endif %}{% endfor %}", + "system_message": null, + "stop_ids": [ + 2 + ] +} diff --git a/applications/ColossalChat/config/conversation_template/zephyr.json b/applications/ColossalChat/config/conversation_template/zephyr.json new file mode 100644 index 000000000000..2ab14111108b --- /dev/null +++ b/applications/ColossalChat/config/conversation_template/zephyr.json @@ -0,0 +1,7 @@ +{ + "chat_template": "{% for message in messages %}\n{% if message['role'] == 'user' %}\n{{ '<|user|>\n' + message['content'] + eos_token }}\n{% elif message['role'] == 'system' %}\n{{ '<|system|>\n' + message['content'] + eos_token }}\n{% elif message['role'] == 'assistant' %}\n{{ '<|assistant|>\n' + message['content'] + eos_token }}\n{% endif %}\n{% if loop.last and add_generation_prompt %}\n{{ '<|assistant|>' }}\n{% endif %}\n{% endfor %}", + "system_message": "A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions.\n\n", + "stop_ids": [ + 2 + ] +} diff --git a/applications/ColossalChat/examples/README.md b/applications/ColossalChat/examples/README.md new file mode 100755 index 000000000000..cfed3f1f3a75 --- /dev/null +++ b/applications/ColossalChat/examples/README.md @@ -0,0 +1,565 @@ +# Examples + +## Table of Contents + +- [Examples](#examples) + - [Table of Contents](#table-of-contents) + - [Install Requirements](#install-requirements) + - [Get Start with ColossalRun](#get-start-with-colossalrun) + - [Training Configuration](#training-configuration) + - [RLHF Stage 1: Supervised Instruction Tuning](#rlhf-training-stage1---supervised-instructs-tuning) + - [Step 1: Data Collection](#step-1-data-collection) + - [Step 2: Preprocessing](#step-2-preprocessing) + - [Step 3: Training](#step-3-training) + - [RLHF Stage 2: Training Reward Model](#rlhf-training-stage2---training-reward-model) + - [Step 1: Data Collection](#step-1-data-collection-1) + - [Step 2: Preprocessing](#step-2-preprocessing-1) + - [Step 3: Training](#step-3-training-1) + - [Features and Tricks in RM Training](#features-and-tricks-in-rm-training) + - [RLHF Stage 3: Proximal Policy Optimization](#rlhf-training-stage3---proximal-policy-optimization) + - [Step 1: Data Collection](#step-1-data-collection-2) + - [Step 2: Preprocessing](#step-2-preprocessing-2) + - [Step 3: Training](#step-3-training-3) + - [PPO Training Results](#sample-training-results-using-default-script) + - [Reward](#reward) + - [KL Divergence](#approximate-kl-divergence) + - [Note on PPO Training](#note-on-ppo-training) + - [Alternative Option For RLHF: Direct Preference Optimization](#alternative-option-for-rlhf-direct-preference-optimization) + - [DPO Stage 1: Supervised Instruction Tuning](#dpo-training-stage1---supervised-instructs-tuning) + - [DPO Stage 2: DPO Training](#dpo-training-stage2---dpo-training) + - [Hardware Requirements](#hardware-requirements) + - [Inference example](#inference-example) + - [Attention](#attention) + +--- + +## Install requirements + +```shell +pip install -r requirements.txt +``` + + +## Get Start with ColossalRun + +You can use colossalai run to launch multi-nodes training: +``` +colossalai run --nproc_per_node YOUR_GPU_PER_NODE --hostfile YOUR_HOST_FILE \ +train.py --OTHER_CONFIGURATIONS +``` +Here is a sample hostfile: + +``` +hostname1 +hostname2 +hostname3 +hostname4 +``` + +Make sure master node can access all nodes (including itself) by ssh without password. Here are some other arguments. + +- nnodes: number of nodes used in the training +- nproc-per-node: specifies the number of processes to be launched per node +- rdzv-endpoint: address of the host node + +### Training Configuration + +This section gives a simple introduction on different training strategies that you can use and how to use them with our boosters and plugins to reduce training time and VRAM consumption. For more detail regarding training strategies, please refer to [here](https://colossalai.org/docs/concepts/paradigms_of_parallelism). For details regarding boosters and plugins, please refer to [here](https://colossalai.org/docs/basics/booster_plugins). + + +
      Gemini + +This plugin implements Zero-3 with chunk-based and heterogeneous memory management. It can train large models without much loss in speed. It also does not support local gradient accumulation. More details can be found in [Gemini Doc](https://colossalai.org/docs/features/zero_with_chunk). + +Below shows how to use the gemini in SFT training. +``` +colossalai run --nproc_per_node 4 --master_port 28534 --hostfile ./hostfile train_sft.py \ + --pretrain $PRETRAINED_MODEL_PATH \ + --tokenizer_dir $PRETRAINED_TOKENIZER_PATH \ + --dataset ${dataset[@]} \ + --save_interval 5000 \ + --save_path $SAVE_DIR \ + --config_file $CONFIG_FILE \ + --plugin gemini \ + --batch_size 4 \ + --max_epochs 1 \ + --accumulation_steps 1 \ # the gradient accumulation has to be disabled + --lr 2e-5 \ + --max_len 2048 \ + --use_wandb +``` + +
      + +
      Gemini-Auto + +This option use gemini and will automatically offload tensors with low priority to cpu. It also does not support local gradient accumulation. More details can be found in [Gemini Doc](https://colossalai.org/docs/features/zero_with_chunk). + +Below shows how to use the gemin-auto in SFT training. +``` +colossalai run --nproc_per_node 4 --master_port 28534 --hostfile ./hostfile train_sft.py \ + --pretrain $PRETRAINED_MODEL_PATH \ + --tokenizer_dir $PRETRAINED_TOKENIZER_PATH \ + --dataset ${dataset[@]} \ + --save_interval 5000 \ + --save_path $SAVE_DIR \ + --config_file $CONFIG_FILE \ + --plugin gemini_auto \ + --batch_size 4 \ + --max_epochs 1 \ + --accumulation_steps 1 \ # the gradient accumulation has to be disabled + --lr 2e-5 \ + --max_len 2048 \ + --use_wandb +``` + +
      + + + +
      Zero2 + +This option will distribute the optimizer parameters and the gradient to multiple GPUs and won't offload weights to cpu. It uses reduce and gather to synchronize gradients and weights. It does not support local gradient accumulation. Though you can accumulate gradient if you insist, it cannot reduce communication cost. That is to say, it's not a good idea to use Zero-2 with pipeline parallelism. + +Below shows how to use the zero2 in SFT training. +``` +colossalai run --nproc_per_node 4 --master_port 28534 --hostfile ./hostfile train_sft.py \ + --pretrain $PRETRAINED_MODEL_PATH \ + --tokenizer_dir $PRETRAINED_TOKENIZER_PATH \ + --dataset ${dataset[@]} \ + --save_interval 5000 \ + --save_path $SAVE_DIR \ + --config_file $CONFIG_FILE \ + --plugin zero2 \ + --batch_size 4 \ + --max_epochs 1 \ + --accumulation_steps 4 \ + --lr 2e-5 \ + --max_len 2048 \ + --use_wandb +``` + +
      + + +
      Zero2CPU + +This option will distribute the optimizer parameters and the gradient to multiple GPUs as well as offload parameters to cpu. It does not support local gradient accumulation. Though you can accumulate gradient if you insist, it cannot reduce communication cost. + +Below shows how to use the zero2-cpu in SFT training. +``` +colossalai run --nproc_per_node 4 --master_port 28534 --hostfile ./hostfile train_sft.py \ + --pretrain $PRETRAINED_MODEL_PATH \ + --tokenizer_dir $PRETRAINED_TOKENIZER_PATH \ + --dataset ${dataset[@]} \ + --save_interval 5000 \ + --save_path $SAVE_DIR \ + --config_file $CONFIG_FILE \ + --plugin zero2_cpu \ + --batch_size 4 \ + --max_epochs 1 \ + --accumulation_steps 4 \ + --lr 2e-5 \ + --max_len 2048 \ + --use_wandb +``` + +
      + +
      Tensor Parallelism + +This option support Tensor Parallelism (TP). Note that if you want to use TP, zero and pipeline parallelism will be disabled. TP split large model weights/optimizer parameters/gradients into multiple small ones and distributes them to multiple GPUs, hence it is recommended to use TP when your model is large (e.g. 20B and above) or your training algorithm consumes a lot of memory (e.g. PPO). + +Below shows how to use the TP in PPO training. +``` +colossalai run --nproc_per_node 4 --hostfile hostfile --master_port 30039 train_ppo.py \ + --pretrain $PRETRAINED_MODEL_PATH \ + --rm_pretrain $PRETRAINED_MODEL_PATH \ + --tokenizer_dir $PRETRAINED_TOKENIZER_PATH \ + --rm_checkpoint_path $REWARD_MODEL_PATH \ + --prompt_dataset ${prompt_dataset[@]} \ + --pretrain_dataset ${ptx_dataset[@]} \ + --ptx_batch_size 1 \ + --ptx_coef 0.0 \ + --plugin "zero2" \ + --save_interval 200 \ + --save_path $SAVE_DIR \ + --num_episodes 2000 \ + --num_collect_steps 4 \ + --num_update_steps 1 \ + --experience_batch_size 8 \ + --train_batch_size 4 \ + --accumulation_steps 8 \ + --tp 4 \ # TP size, nproc_per_node must be divisible by it + --lr 9e-6 \ + --mixed_precision "bf16" \ + --grad_clip 1.0 \ + --weight_decay 0.01 \ + --warmup_steps 100 \ + --grad_checkpoint \ + --use_wandb +``` + +
      + + +
      Gradient Checkpointing + +This option saves VRAM consumption by selectively recomputing some of the intermediate value on-the-fly during the backward pass, rather than storing them in memory. + +To enable gradient checkpointing, add --grad_checkpoint to your training script. +``` +colossalai run --nproc_per_node 4 --master_port 28534 --hostfile ./hostfile train_sft.py \ + --pretrain $PRETRAINED_MODEL_PATH \ + --tokenizer_dir $PRETRAINED_TOKENIZER_PATH \ + --dataset ${dataset[@]} \ + --save_interval 5000 \ + --save_path $SAVE_DIR \ + --config_file $CONFIG_FILE \ + --plugin zero2_cpu \ + --batch_size 4 \ + --max_epochs 1 \ + --accumulation_steps 4 \ + --lr 2e-5 \ + --max_len 2048 \ + --grad_checkpoint \ # This enables gradient checkpointing + --use_wandb +``` + +
      + +
      Flash Attention + +Details about flash attention can be found in the paper: [FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness](https://arxiv.org/abs/2205.14135). + +To enable flash attention, add --use_flash_attn to your training script. +``` +colossalai run --nproc_per_node 4 --master_port 28534 --hostfile ./hostfile train_sft.py \ + --pretrain $PRETRAINED_MODEL_PATH \ + --tokenizer_dir $PRETRAINED_TOKENIZER_PATH \ + --dataset ${dataset[@]} \ + --save_interval 5000 \ + --save_path $SAVE_DIR \ + --config_file $CONFIG_FILE \ + --plugin zero2_cpu \ + --batch_size 4 \ + --max_epochs 1 \ + --accumulation_steps 4 \ + --lr 2e-5 \ + --max_len 2048 \ + --use_flash_attn \ # This enables flash attention + --use_wandb +``` + +
      + +
      Low Rank Adaption + +Details about Low Rank Adaption (LoRA) can be found in the paper: [LoRA: Low-Rank Adaptation of Large Language Models](https://arxiv.org/abs/2106.09685). It dramatically reduce the VRAM consumption at the cost of sacrifice model capability. It is suitable for training LLM with constrained resources. + +To enable LoRA, set --lora_rank to a positive value (usually between 20 and 64). +``` +colossalai run --nproc_per_node 4 --master_port 28534 --hostfile ./hostfile train_sft.py \ + --pretrain $PRETRAINED_MODEL_PATH \ + --tokenizer_dir $PRETRAINED_TOKENIZER_PATH \ + --dataset ${dataset[@]} \ + --save_interval 5000 \ + --save_path $SAVE_DIR \ + --config_file $CONFIG_FILE \ + --plugin zero2_cpu \ + --batch_size 4 \ + --max_epochs 1 \ + --accumulation_steps 4 \ + --lr 2e-5 \ + --max_len 2048 \ + --lora_rank 32 \ # This enables LoRA + --use_wandb +``` + +
      + +
      Other Training Arguments + +- grad_clip: gradient larger than this value will be clipped. +- weight_decay: weight decay hyper-parameter. +- warmup_steps: number of warmup steps used in setting up the learning rate scheduler. +- pretrain: pretrain model path, weights will be loaded from this pretrained model unless checkpoint_path is provided. +- tokenizer_dir: specify where to load the tokenizer, if not provided, tokenizer will be loaded from pretrain model path. +- dataset: a list of strings, each is a path to a folder contains buffered dataset files in arrow format. +- checkpoint_path: if provided, will load weights from the checkpoint_path. +- config_file: path to store the training config file. +- save_dir: path to store the model checkpoints. +- max_length: input will be padded/truncate to max_length before feeding to the model. +- max_epochs: number of epoch to train. +- batch_size: training batch size. +- mixed_precision: precision to use in training. Support 'fp16' and 'bf16'. Note that some device may not support the 'bf16' option, please refer to [Nvidia](https://developer.nvidia.com/) to check compatibility. +- save_interval: save the model weights as well as optimizer/scheduler states every save_interval steps/episodes. +- merge_lora_weights: whether to merge lora weights before saving the model +- lr: the learning rate used in training. +- accumulation_steps: accumulate gradient every accumulation_steps. +- log_dir: path to store the log. +- use_wandb: if this flag is up, you can view logs on wandb. + +
      + +### RLHF Training Stage1 - Supervised Instructs Tuning + +Stage1 is supervised instructs fine-tuning (SFT). This step is a crucial part of the RLHF training process, as it involves training a machine learning model using human-provided instructions to learn the initial behavior for the task at hand. Here's a detailed guide on how to SFT your LLM with ColossalChat: + +#### Step 1: Data Collection +The first step in Stage 1 is to collect a dataset of human demonstrations of the following format. + +```json +[ + {"messages": + [ + { + "from": "human", + "content": "what are some pranks with a pen i can do?" + }, + { + "from": "assistant", + "content": "Are you looking for practical joke ideas?" + }, + ... + ] + }, + ... +] +``` + +#### Step 2: Preprocessing +Once you have collected your SFT dataset, you will need to preprocess it. This involves four steps: data cleaning, data deduplication, formatting and tokenization. In this section, we will focus on formatting and tokenization. + +In this code we provide a flexible way for users to set the conversation template for formatting chat data using Huggingface's newest feature--- chat template. Please follow the following steps to define your chat template and preprocess your data. + +- Step 1: (Optional). Define your conversation template. You need to provide a conversation template config file similar to the config files under the ./config/conversation_template directory. This config should include the following fields. + ```json + { + "chat_template": (Optional), A string of chat_template used for formatting chat data. If not set (None), will use the default chat template of the provided tokenizer. If a path to a huggingface model or local model is provided, will use the chat_template of that model. To use a custom chat template, you need to manually set this field. For more details on how to write a chat template in Jinja format, please read https://huggingface.co/docs/transformers/main/chat_templating, + "system_message": A string of system message to be added at the beginning of the prompt. If no is provided (None), no system message will be added, + "stop_ids": (Optional), A list of string indicating the end of assistant's response during the rollout stage of PPO training. It's recommended to set this manually for PPO training. If not set, will set to tokenizer.eos_token_ids automatically, + } + ``` + On your first run of the data preparation script, you only need to define the "chat_template" (if you want to use custom chat template) and the "system message" (if you want to use a custom system message), + +- Step 2: Run the data preparation script--- [prepare_sft_dataset.sh](./examples/data_preparation_scripts/prepare_sft_dataset.sh). Note that whether or not you have skipped the first step, you need to provide the path to the conversation template config file (via the conversation_template_config arg). If you skipped the first step, an auto-generated conversation template will be stored at the designated file path. + +- Step 3: (Optional) Check the correctness of the processed data. We provided an easy way for you to do a manual checking on the processed data by checking the "$SAVE_DIR/jsonl/part-XXXX.jsonl" files. + +Finishing the above steps, you have converted the raw conversation to the designated chat format and tokenized the formatted conversation, calculate input_ids, labels, attention_masks and buffer those into binary dataset files under "$SAVE_DIR/arrow/part-XXXX" folders. + +For example, our Colossal-LLaMA-2 format looks like, +``` + A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions. + +Human: what are some pranks with a pen i can do? Assistant: Are you looking for practical joke ideas? +... +``` + +#### Step 3: Training +Choose a suitable model architecture for your task. Note that your model should be compatible with the tokenizer that you used to tokenize the SFT dataset. You can run [train_sft.sh](./examples/training_scripts/train_sft.sh) to start a supervised instructs fine-tuning. Please refer to the [training configuration](#training-configuration) section for details regarding supported training options. + +### RLHF Training Stage2 - Training Reward Model + +Stage2 trains a reward model, which obtains corresponding scores by manually ranking different outputs for the same prompt and supervises the training of the reward model. + +#### Step 1: Data Collection +Below shows the preference dataset format used in training the reward model. + +```json +[ + {"context": [ + { + "from": "human", + "content": "Introduce butterflies species in Oregon." + } + ] + "chosen": [ + { + "from": "assistant", + "content": "About 150 species of butterflies live in Oregon, with about 100 species are moths..." + }, + ... + ], + "rejected": [ + { + "from": "assistant", + "content": "Are you interested in just the common butterflies? There are a few common ones which will be easy to find..." + }, + ... + ] + }, + ... +] +``` + +#### Step 2: Preprocessing +Similar to the second step in the previous stage, we format the reward data into the same structured format as used in step 2 of the SFT stage. You can run [prepare_preference_dataset.sh](./examples/data_preparation_scripts/prepare_preference_dataset.sh) to prepare the preference data for reward model training. + +#### Step 3: Training +You can run [train_rm.sh](./examples/training_scripts/train_rm.sh) to start the reward model training. Please refer to the [training configuration](#training-configuration) section for details regarding supported training options. + +#### Features and Tricks in RM Training + +- We recommend using the [Anthropic/hh-rlhf](https://huggingface.co/datasets/Anthropic/hh-rlhf)and[rm-static](https://huggingface.co/datasets/Dahoas/rm-static) datasets for training the reward model. +- We support 2 kinds of loss function named `log_sig`(used by OpenAI) and `log_exp`(used by Anthropic). +- We log the training accuracy `train/acc`, `reward_chosen` and `reward_rejected` to monitor progress during training. +- We use cosine-reducing lr-scheduler for RM training. +- We set value_head as 1 liner layer and initialize the weight of value_head using N(0，1/(d_model + 1)) distribution. + +#### Note on Reward Model Training + +Before you move on the next stage, please check the following list to ensure that your reward model is stable and robust. You can check the reward chart and the accuracy chart on wandb. +- The mean reward for chosen data is much higher than those for rejected data +- The accuracy is larger than 0.5 by a significant margin (usually should be greater than 0.6) +- Optional：check the reward is positive for chosen data vice versa + +Your training reward curves should look similar to the following charts. +

      +image +

      + +### RLHF Training Stage3 - Proximal Policy Optimization + +In stage3 we will use reinforcement learning algorithm--- Proximal Policy Optimization (PPO), which is the most complex part of the training process: + +

      + +

      + +#### Step 1: Data Collection +PPO uses two kind of training data--- the prompt data and the pretrain data (optional). The first dataset is mandatory, data samples within the prompt dataset ends with a line from "human" and thus the "assistant" needs to generate a response to answer to the "human". Note that you can still use conversation that ends with a line from the "assistant", in that case, the last line will be dropped. Here is an example of the prompt dataset format. + +```json +[ + {"messages": + [ + { + "from": "human", + "content": "what are some pranks with a pen i can do?" + } + ... + ] + }, +] +``` + +The second dataset--- pretrained dataset is optional, provide it if you want to use the ptx loss introduced in the [InstructGPT paper](https://arxiv.org/abs/2203.02155). It follows the following format. + +```json + [ + { + "source": "", # system instruction + "Target": "Provide a list of the top 10 most popular mobile games in Asia\nThe top 10 most popular mobile games in Asia are:\n1) PUBG Mobile\n2) Pokemon Go\n3) Candy Crush Saga\n4) Free Fire\n5) Clash of Clans\n6) Mario Kart Tour\n7) Arena of Valor\n8) Fantasy Westward Journey\n9) Subway Surfers\n10) ARK Survival Evolved", + }, + ... + ] + ``` +#### Step 2: Preprocessing +To prepare the prompt dataset for PPO training, simply run [prepare_prompt_dataset.sh](./examples/data_preparation_scripts/prepare_prompt_dataset.sh) + +You can use the SFT dataset you prepared in the SFT stage or prepare a new one from different source for the ptx dataset. The ptx data is used to calculate ptx loss, which stablize the training according to the [InstructGPT paper](https://arxiv.org/pdf/2203.02155.pdf). + +#### Step 3: Training +You can run the [train_ppo.sh](./examples/training_scripts/train_ppo.sh) to start PPO training. Here are some unique arguments for PPO, please refer to the training configuration section for other training configuration. Please refer to the [training configuration](#training-configuration) section for details regarding supported training options. + +```bash +--pretrain $PRETRAINED_MODEL_PATH \ +--rm_pretrain $PRETRAINED_MODEL_PATH \ # reward model architectural +--tokenizer_dir $PRETRAINED_TOKENIZER_PATH \ +--rm_checkpoint_path $REWARD_MODEL_PATH \ # reward model checkpoint path +--prompt_dataset ${prompt_dataset[@]} \ # List of string, prompt dataset +--conversation_template_config $CONVERSATION_TEMPLATE_CONFIG_PATH \ # path to the conversation template config file +--pretrain_dataset ${ptx_dataset[@]} \ # List of string, the sft dataset +--ptx_batch_size 1 \ # batch size for calculate ptx loss +--ptx_coef 0.0 \ # none-zero if ptx loss is enable +--num_episodes 2000 \ # number of episodes to train +--num_collect_steps 1 \ +--num_update_steps 1 \ +--experience_batch_size 8 \ +--train_batch_size 4 \ +--accumulation_steps 2 +``` + +Each episode has two phases, the collect phase and the update phase. During the collect phase, we will collect experiences (answers generated by actor), store those in ExperienceBuffer. Then data in ExperienceBuffer is used during the update phase to update parameter of actor and critic. + +- Without tensor parallelism, +``` +experience buffer size += num_process * num_collect_steps * experience_batch_size += train_batch_size * accumulation_steps * num_process +``` + +- With tensor parallelism, +``` +num_tp_group = num_process / tp +experience buffer size += num_tp_group * num_collect_steps * experience_batch_size += train_batch_size * accumulation_steps * num_tp_group +``` + +### Sample Training Results Using Default Script +#### Reward +

      +image +

      + +### Note on PPO Training +#### Q1: My reward is negative +Answer: Check your reward model trained in stage 1. If the reward model only generate negative reward, we actually will expect a negative reward. However, even though the reward is negative, the reward should go up. + +#### Q2: My actor loss is negative +Answer: This is normal for actor loss as PPO doesn't restrict the actor loss to be positive. + +#### Q3: My reward doesn't go up (decreases) +Answer: The causes to this problem are two-fold. Check your reward model, make sure that it gives positive and strong reward for good cases and negative, strong reward for bad responses. You should also try different hyperparameter settings. + +#### Q4: Generation is garbage +Answer: Yes, this happens and is well documented by other implementations. After training for too many episodes, the actor gradually deviate from its original state, which may leads to decrease in language modeling capabilities. A way to fix this is to add supervised loss during PPO. Set ptx_coef to a none-zero value (between 0 and 1), which balances PPO loss and sft loss. + +## Alternative Option For RLHF: Direct Preference Optimization + +For those seeking an alternative to Reinforcement Learning from Human Feedback (RLHF), Direct Preference Optimization (DPO) presents a compelling option. DPO, as detailed in the paper (available at [https://arxiv.org/abs/2305.18290](https://arxiv.org/abs/2305.18290)), DPO offers an low-cost way to perform RLHF and usually request less computation resources compares to PPO. + +### DPO Training Stage1 - Supervised Instructs Tuning + +Please refer the [sft section](#dpo-training-stage1---supervised-instructs-tuning) in the PPO part. + +### DPO Training Stage2 - DPO Training +#### Step 1: Data Collection & Preparation +For DPO training, you only need the preference dataset. Please follow the instruction in the [preference dataset preparation section](#rlhf-training-stage2---training-reward-model) to prepare the preference data for DPO training. + +#### Step 2: Training +You can run the [train_dpo.sh](./examples/training_scripts/train_dpo.sh) to start DPO training. Please refer to the [training configuration](#training-configuration) section for details regarding supported training options. + +#### DPO Result +

      +image +

      + +## Hardware Requirements +For PPO, we suggest using Tensor Parallelism. The following table shows the VRAM consumption of training a 7B model on a dummy dataset with 2048 sequence length and 512 layout length with different tp_size (equal to the number of GPUs). In this experiment, we use H800 GPU with 80GB VRAM. +| PPO | tp=8 | tp=4 | +|-------|---------------|---------------| +| bs=1 | 18485.19 MB | 42934.45 MB | +| bs=4 | 25585.65 MB | 42941.93 MB | +| bs=16 | 41408.28 MB | 56778.97 MB | +| bs=30 | 64047.42 MB | failed | + +For DPO, we recommend using zero2 or zero2-cpu. We tested the VRAM consumption on a dummy dataset with 2048 sequence length. + +- 1 H800 GPU + - zero2-cpu, batch size=2, VRAM Usage=49873.90 MB + - zero2-cpu, batch size=4, VRAM Usage=60998.22 MB +- 4 H800 GPUs + - zero2, batch size=4, VRAM Usage=67544.47 MB + +## Inference example + +We support different inference options, including int8 and int4 quantization. +For details, see [`inference/`](https://github.com/hpcaitech/ColossalAI/tree/main/applications/Chat/inference). + +## Attention + +The examples are demos for the whole training process. You need to change the hyper-parameters to reach great performance. diff --git a/applications/Chat/examples/community/README.md b/applications/ColossalChat/examples/community/README.md old mode 100644 new mode 100755 similarity index 100% rename from applications/Chat/examples/community/README.md rename to applications/ColossalChat/examples/community/README.md diff --git a/applications/Chat/examples/community/peft/README.md b/applications/ColossalChat/examples/community/peft/README.md old mode 100644 new mode 100755 similarity index 100% rename from applications/Chat/examples/community/peft/README.md rename to applications/ColossalChat/examples/community/peft/README.md diff --git a/applications/Chat/examples/community/peft/easy_dataset.py b/applications/ColossalChat/examples/community/peft/easy_dataset.py old mode 100644 new mode 100755 similarity index 100% rename from applications/Chat/examples/community/peft/easy_dataset.py rename to applications/ColossalChat/examples/community/peft/easy_dataset.py diff --git a/applications/Chat/examples/community/peft/easy_models.py b/applications/ColossalChat/examples/community/peft/easy_models.py old mode 100644 new mode 100755 similarity index 100% rename from applications/Chat/examples/community/peft/easy_models.py rename to applications/ColossalChat/examples/community/peft/easy_models.py diff --git a/applications/Chat/examples/community/peft/train_peft_prompts.py b/applications/ColossalChat/examples/community/peft/train_peft_prompts.py old mode 100644 new mode 100755 similarity index 100% rename from applications/Chat/examples/community/peft/train_peft_prompts.py rename to applications/ColossalChat/examples/community/peft/train_peft_prompts.py diff --git a/applications/Chat/examples/community/peft/train_peft_sft.py b/applications/ColossalChat/examples/community/peft/train_peft_sft.py old mode 100644 new mode 100755 similarity index 100% rename from applications/Chat/examples/community/peft/train_peft_sft.py rename to applications/ColossalChat/examples/community/peft/train_peft_sft.py diff --git a/applications/Chat/examples/community/ray/README.md b/applications/ColossalChat/examples/community/ray/README.md old mode 100644 new mode 100755 similarity index 100% rename from applications/Chat/examples/community/ray/README.md rename to applications/ColossalChat/examples/community/ray/README.md diff --git a/applications/Chat/examples/community/ray/ray_job_script.py b/applications/ColossalChat/examples/community/ray/ray_job_script.py old mode 100644 new mode 100755 similarity index 100% rename from applications/Chat/examples/community/ray/ray_job_script.py rename to applications/ColossalChat/examples/community/ray/ray_job_script.py diff --git a/applications/Chat/examples/community/ray/train_prompts_on_ray.py b/applications/ColossalChat/examples/community/ray/train_prompts_on_ray.py old mode 100644 new mode 100755 similarity index 100% rename from applications/Chat/examples/community/ray/train_prompts_on_ray.py rename to applications/ColossalChat/examples/community/ray/train_prompts_on_ray.py diff --git a/applications/ColossalChat/examples/data_preparation_scripts/prepare_dataset.py b/applications/ColossalChat/examples/data_preparation_scripts/prepare_dataset.py new file mode 100644 index 000000000000..64093f88d7ca --- /dev/null +++ b/applications/ColossalChat/examples/data_preparation_scripts/prepare_dataset.py @@ -0,0 +1,268 @@ +#!/usr/bin/env python3 +# -*- coding: utf-8 -*- +""" +Prepare dataset scripts + +Usage: +- For SFT dataset preparation (SFT) +python prepare_dataset.py --type sft \ + --data_input_dirs /PATH/TO/SFT/DATASET \ + --conversation_template_config /PATH/TO/CHAT/TEMPLATE/CONFIG.json \ + --tokenizer_dir "" \ + --data_cache_dir $SAVE_DIR/cache \ + --data_jsonl_output_dir $SAVE_DIR/jsonl \ + --data_arrow_output_dir $SAVE_DIR/arrow \ + +- For prompt dataset preparation (PPO) +python prepare_dataset.py --type prompt \ + --data_input_dirs /PATH/TO/SFT/DATASET \ + --conversation_template_config /PATH/TO/CHAT/TEMPLATE/CONFIG.json \ + --tokenizer_dir "" \ + --data_cache_dir $SAVE_DIR/cache \ + --data_jsonl_output_dir $SAVE_DIR/jsonl \ + --data_arrow_output_dir $SAVE_DIR/arrow \ + +- For Preference dataset preparation (DPO and Reward model training) +python prepare_dataset.py --type preference \ + --data_input_dirs /PATH/TO/SFT/DATASET \ + --conversation_template_config /PATH/TO/CHAT/TEMPLATE/CONFIG.json \ + --tokenizer_dir "" \ + --data_cache_dir $SAVE_DIR/cache \ + --data_jsonl_output_dir $SAVE_DIR/jsonl \ + --data_arrow_output_dir $SAVE_DIR/arrow \ +""" + +import argparse +import json +import math +import os +import random +import time +from multiprocessing import cpu_count + +from coati.dataset import setup_conversation_template, supervised_tokenize_sft, tokenize_prompt_dataset, tokenize_rlhf +from datasets import dataset_dict, load_dataset +from transformers import AutoTokenizer + +from colossalai.logging import get_dist_logger + +logger = get_dist_logger() + + +def main(): + parser = argparse.ArgumentParser() + parser.add_argument( + "--type", + type=str, + required=True, + default=None, + choices=["sft", "prompt", "preference"], + help="Type of dataset, chose from 'sft', 'prompt', 'preference'.", + ) + parser.add_argument( + "--data_input_dirs", + type=str, + required=True, + default=None, + help="Comma(i.e., ',') separated list of all data directories containing `.jsonl` data files.", + ) + parser.add_argument( + "--tokenizer_dir", type=str, required=True, default=None, help="A directory containing the tokenizer" + ) + parser.add_argument( + "--conversation_template_config", + type=str, + default="conversation_template_config", + help="Path \ + to save conversation template config files.", + ) + parser.add_argument("--data_cache_dir", type=str, default="cache", help="Data cache directory") + parser.add_argument( + "--data_jsonl_output_dir", + type=str, + default="jsonl_output", + help="Output directory of spliced dataset with jsonl format", + ) + parser.add_argument( + "--data_arrow_output_dir", + type=str, + default="arrow_output", + help="Output directory of spliced dataset with arrow format", + ) + parser.add_argument("--max_length", type=int, default=4096, help="Max length of each spliced tokenized sequence") + parser.add_argument("--num_spliced_dataset_bins", type=int, default=10, help="Number of spliced dataset bins") + parser.add_argument( + "--num_samples_per_datafile", + type=int, + default=-1, + help="Number of samples to be generated from each data file. -1 denote all samples.", + ) + args = parser.parse_args() + + if args.num_spliced_dataset_bins >= 100000: + raise ValueError("Too many spliced divisions, must be smaller than 100000") + + assert not os.path.exists(args.data_cache_dir), f"Find existed data cache dir {args.data_cache_dir}" + assert not os.path.exists( + args.data_jsonl_output_dir + ), f"Find existed jsonl data output dir {args.data_jsonl_output_dir}" + assert not os.path.exists( + args.data_arrow_output_dir + ), f"Find existed arrow data output dir {args.data_arrow_output_dir}" + os.makedirs(args.data_jsonl_output_dir) + os.makedirs(args.data_arrow_output_dir) + + # Prepare to all input datasets + input_data_paths = [] + input_data_dirs = args.data_input_dirs.split(",") + for ds_dir in input_data_dirs: + ds_dir = os.path.abspath(ds_dir) + assert os.path.exists(ds_dir), f"Not find data dir {ds_dir}" + ds_files = [name for name in os.listdir(ds_dir) if name.endswith(".jsonl")] + ds_paths = [os.path.join(ds_dir, name) for name in ds_files] + input_data_paths.extend(ds_paths) + + # Prepare to data splitting. + train_splits = [] + split_interval = math.ceil(100 / args.num_spliced_dataset_bins) + for i in range(0, 100, split_interval): + start = i + end = i + split_interval + if end > 100: + end = 100 + train_splits.append(f"train[{start}%:{end}%]") + + # Prepare the tokenizer. + tokenizer = AutoTokenizer.from_pretrained(args.tokenizer_dir, use_fast=False, trust_remote_code=True) + if os.path.exists(args.conversation_template_config): + chat_template_config = json.load(open(args.conversation_template_config, "r", encoding="utf8")) + else: + chat_template_config = { + "system_message": "A chat between a curious human and an artificial intelligence assistant. " + "The assistant gives helpful, detailed, and polite answers to the human's questions.\n\n" + } # Use default system message + if args.type == "preference": + if "stop_ids" not in chat_template_config: + # Ask the user to define stop_ids for PPO training + dummy_messages = [ + {"role": "user", "content": "Hello, how are you?"}, + {"role": "assistant", "content": "I'm doing great. How can I help you today?"}, + {"role": "user", "content": "Who made you?"}, + {"role": "assistant", "content": "I am a chatbot trained by Colossal-AI."}, + ] + dummy_prompt = tokenizer.apply_chat_template(dummy_messages, tokenize=False) + tokenized = tokenizer(dummy_prompt, add_special_tokens=False)["input_ids"] + tokens = tokenizer.convert_ids_to_tokens(tokenized, skip_special_tokens=False) + corresponding_str = [tokenizer.convert_tokens_to_string([token]) for token in tokens] + token_id_mapping = [{"token": s, "id": tokenized[i]} for i, s in enumerate(corresponding_str)] + stop_ids = input( + "For PPO, we recommend to provide stop_ids for the properly stop the generation during roll out stage. " + "stop_ids are the ids of repetitive pattern that indicate the end of the assistant's response. " + "Here is an example of formatted prompt and token-id mapping, you can set stop_ids by entering a list " + "of integers, separate by space, press `Enter` to end. Or you can press `Enter` without input if you are " + "not using PPO or you prefer to not set the stop_ids, in that case, stop_ids will be set to tokenizer.eos_token_id. " + f"\nPrompt:\n{dummy_prompt}\nToken-id Mapping:\n{token_id_mapping}\nstop_ids:" + ) + if stop_ids == "": + chat_template_config["stop_ids"] = [tokenizer.eos_token_id] + else: + try: + chat_template_config["stop_ids"] = [int(s) for s in stop_ids.split()] + except ValueError: + raise ValueError("Invalid input, please provide a list of integers.") + else: + # Set stop_ids to eos_token_id for other dataset types if not exist + if "stop_ids" not in chat_template_config: + chat_template_config["stop_ids"] = [tokenizer.eos_token_id] + + conversation_template = setup_conversation_template( + tokenizer, chat_template_config=chat_template_config, save_path=args.conversation_template_config + ) + if hasattr(tokenizer, "pad_token") and hasattr(tokenizer, "eos_token") and tokenizer.eos_token is not None: + try: + # Some tokenizers doesn't allow to set pad_token mannually e.g., Qwen + tokenizer.pad_token = tokenizer.eos_token + except AttributeError as e: + logger.warning(f"Unable to set pad token to eos token, {str(e)}") + if not hasattr(tokenizer, "pad_token") or tokenizer.pad_token is None: + logger.warning( + "The tokenizer does not have a pad token which is required. May lead to unintended behavior in training, Please consider manually set them." + ) + + list_dataset = load_dataset( + path="json", + data_files=input_data_paths, + cache_dir=os.path.join(args.data_cache_dir, "raw"), + keep_in_memory=False, + split=train_splits, + num_proc=cpu_count(), + ) + + if args.type == "sft": + preparation_function = supervised_tokenize_sft + elif args.type == "prompt": + preparation_function = tokenize_prompt_dataset + elif args.type == "preference": + preparation_function = tokenize_rlhf + else: + raise ValueError("Unknow dataset type. Please choose one from ['sft', 'prompt', 'preference']") + + for index, dataset in enumerate(list_dataset): + assert isinstance(dataset, dataset_dict.Dataset) + if len(dataset) == 0: + # Hack: Skip empty dataset. If dataset contains less than num_of_rank samples, some rank may have empty dataset and leads to error + continue + if args.num_samples_per_datafile > 0: + # limit the number of samples in each dataset + dataset = dataset.select( + random.sample(range(len(dataset)), min(args.num_samples_per_datafile, len(dataset))) + ) + logger.info(f"Start to process part-{index}/{len(list_dataset)} of all original datasets.") + dataset = dataset.map( + function=preparation_function, + fn_kwargs={ + "tokenizer": tokenizer, + "conversation_template": conversation_template, + "max_length": args.max_length, + }, + keep_in_memory=False, + num_proc=min(len(dataset), cpu_count()), + ) + + dataset = dataset.filter( + lambda data: data["chosen_input_ids" if args.type == "preference" else "input_ids"] is not None + ) + + # Save each jsonl spliced dataset. + output_index = "0" * (5 - len(str(index))) + str(index) + output_name = f"part-{output_index}" + output_jsonl_path = os.path.join(args.data_jsonl_output_dir, output_name + ".jsonl") + st = time.time() + with open(file=output_jsonl_path, mode="w", encoding="utf-8") as fp_writer: + count = 0 + for data_point in dataset: + if count % 500 == 0: + logger.info(f"processing {count} spliced data points for {fp_writer.name}") + count += 1 + fp_writer.write(json.dumps(data_point, ensure_ascii=False) + "\n") + logger.info( + f"Current file {fp_writer.name}; " + f"Data size: {len(dataset)}; " + f"Time cost: {round((time.time() - st) / 60, 6)} minutes." + ) + # Save each arrow spliced dataset + output_arrow_path = os.path.join(args.data_arrow_output_dir, output_name) + logger.info(f"Start to save {output_arrow_path}") + dataset = load_dataset( + path="json", + data_files=[output_jsonl_path], + cache_dir=os.path.join(args.data_cache_dir, "tokenized"), + keep_in_memory=False, + num_proc=cpu_count(), + split="train", + ) + dataset.save_to_disk(dataset_path=output_arrow_path, num_proc=min(len(dataset), cpu_count())) + + +if __name__ == "__main__": + main() diff --git a/applications/ColossalChat/examples/data_preparation_scripts/prepare_preference_dataset.sh b/applications/ColossalChat/examples/data_preparation_scripts/prepare_preference_dataset.sh new file mode 100755 index 000000000000..999d7778be52 --- /dev/null +++ b/applications/ColossalChat/examples/data_preparation_scripts/prepare_preference_dataset.sh @@ -0,0 +1,13 @@ +SAVE_DIR="" + +rm -rf $SAVE_DIR/cache +rm -rf $SAVE_DIR/jsonl +rm -rf $SAVE_DIR/arrow + +python prepare_dataset.py --type preference \ + --data_input_dirs "PATH/TO/PREFERENCE/DATA" \ + --conversation_template_config /PATH/TO/CHAT/TEMPLATE/CONFIG.json \ + --tokenizer_dir "" \ + --data_cache_dir $SAVE_DIR/cache \ + --data_jsonl_output_dir $SAVE_DIR/jsonl \ + --data_arrow_output_dir $SAVE_DIR/arrow diff --git a/applications/ColossalChat/examples/data_preparation_scripts/prepare_prompt_dataset.sh b/applications/ColossalChat/examples/data_preparation_scripts/prepare_prompt_dataset.sh new file mode 100755 index 000000000000..8d3d6c2c2d80 --- /dev/null +++ b/applications/ColossalChat/examples/data_preparation_scripts/prepare_prompt_dataset.sh @@ -0,0 +1,13 @@ +SAVE_DIR="" + +rm -rf $SAVE_DIR/cache +rm -rf $SAVE_DIR/jsonl +rm -rf $SAVE_DIR/arrow + +python prepare_dataset.py --type prompt \ + --data_input_dirs /PATH/TO/PROMPT/DATASET \ + --conversation_template_config /PATH/TO/CHAT/TEMPLATE/CONFIG.json \ + --tokenizer_dir "" \ + --data_cache_dir $SAVE_DIR/cache \ + --data_jsonl_output_dir $SAVE_DIR/jsonl \ + --data_arrow_output_dir $SAVE_DIR/arrow diff --git a/applications/ColossalChat/examples/data_preparation_scripts/prepare_sft_dataset.sh b/applications/ColossalChat/examples/data_preparation_scripts/prepare_sft_dataset.sh new file mode 100755 index 000000000000..cf937db2a84b --- /dev/null +++ b/applications/ColossalChat/examples/data_preparation_scripts/prepare_sft_dataset.sh @@ -0,0 +1,13 @@ +SAVE_DIR="" + +rm -rf $SAVE_DIR/cache +rm -rf $SAVE_DIR/jsonl +rm -rf $SAVE_DIR/arrow + +python prepare_dataset.py --type sft \ + --data_input_dirs /PATH/TO/SFT/DATASET \ + --conversation_template_config /PATH/TO/CHAT/TEMPLATE/CONFIG.json \ + --tokenizer_dir "" \ + --data_cache_dir $SAVE_DIR/cache \ + --data_jsonl_output_dir $SAVE_DIR/jsonl \ + --data_arrow_output_dir $SAVE_DIR/arrow \ diff --git a/applications/ColossalChat/examples/inference/chatio.py b/applications/ColossalChat/examples/inference/chatio.py new file mode 100755 index 000000000000..26784f3a3411 --- /dev/null +++ b/applications/ColossalChat/examples/inference/chatio.py @@ -0,0 +1,168 @@ +""" +command line IO utils for chatbot +""" + +import abc +import re + +from prompt_toolkit import PromptSession +from prompt_toolkit.auto_suggest import AutoSuggestFromHistory +from prompt_toolkit.completion import WordCompleter +from prompt_toolkit.history import InMemoryHistory +from rich.console import Console +from rich.live import Live +from rich.markdown import Markdown + + +class ChatIO(abc.ABC): + @abc.abstractmethod + def prompt_for_input(self, role: str) -> str: + """Prompt for input from a role.""" + + @abc.abstractmethod + def prompt_for_output(self, role: str): + """Prompt for output from a role.""" + + @abc.abstractmethod + def stream_output(self, output_stream): + """Stream output.""" + + +class SimpleChatIO(ChatIO): + def prompt_for_input(self, role) -> str: + return input(f"{role}: ") + + def prompt_for_output(self, role: str): + print(f"{role}: ", end="", flush=True) + + def stream_output(self, output_stream): + pre = 0 + for outputs in output_stream: + outputs = outputs.strip() + outputs = outputs.split(" ") + now = len(outputs) - 1 + if now > pre: + print(" ".join(outputs[pre:now]), end=" ", flush=True) + pre = now + print(" ".join(outputs[pre:]), flush=True) + return " ".join(outputs) + + +class RichChatIO(ChatIO): + def __init__(self): + self._prompt_session = PromptSession(history=InMemoryHistory()) + self._completer = WordCompleter(words=["!exit", "!reset"], pattern=re.compile("$")) + self._console = Console() + + def prompt_for_input(self, role) -> str: + self._console.print(f"[bold]{role}:") + prompt_input = self._prompt_session.prompt( + completer=self._completer, + multiline=False, + auto_suggest=AutoSuggestFromHistory(), + key_bindings=None, + ) + self._console.print() + return prompt_input + + def prompt_for_output(self, role: str) -> str: + self._console.print(f"[bold]{role}:") + + def stream_output(self, output_stream): + """Stream output from a role.""" + # Create a Live context for updating the console output + with Live(console=self._console, refresh_per_second=60) as live: + # Read lines from the stream + for outputs in output_stream: + accumulated_text = outputs + if not accumulated_text: + continue + # Render the accumulated text as Markdown + # NOTE: this is a workaround for the rendering "unstandard markdown" + # in rich. The chatbots output treat "\n" as a new line for + # better compatibility with real-world text. However, rendering + # in markdown would break the format. It is because standard markdown + # treat a single "\n" in normal text as a space. + # Our workaround is adding two spaces at the end of each line. + # This is not a perfect solution, as it would + # introduce trailing spaces (only) in code block, but it works well + # especially for console output, because in general the console does not + # care about trailing spaces. + lines = [] + for line in accumulated_text.splitlines(): + lines.append(line) + if line.startswith("```"): + # Code block marker - do not add trailing spaces, as it would + # break the syntax highlighting + lines.append("\n") + else: + lines.append(" \n") + markdown = Markdown("".join(lines)) + # Update the Live console output + live.update(markdown) + self._console.print() + return outputs + + +class DummyChatIO(ChatIO): + """ + Dummy ChatIO class for testing + """ + + def __init__(self): + self.roles = [] + self._console = Console() + + def prompt_for_input(self, role) -> str: + self.roles.append(role) + if len(self.roles) == 1: + ret = "Hello" + elif len(self.roles) == 2: + ret = "What's the value of 1+1?" + else: + ret = "exit" + self._console.print(f"[bold]{role}:{ret}") + return ret + + def prompt_for_output(self, role: str) -> str: + self._console.print(f"[bold]{role}:") + + def stream_output(self, output_stream): + """Stream output from a role.""" + # Create a Live context for updating the console output + with Live(console=self._console, refresh_per_second=60) as live: + # Read lines from the stream + for outputs in output_stream: + accumulated_text = outputs + if not accumulated_text: + continue + # Render the accumulated text as Markdown + # NOTE: this is a workaround for the rendering "unstandard markdown" + # in rich. The chatbots output treat "\n" as a new line for + # better compatibility with real-world text. However, rendering + # in markdown would break the format. It is because standard markdown + # treat a single "\n" in normal text as a space. + # Our workaround is adding two spaces at the end of each line. + # This is not a perfect solution, as it would + # introduce trailing spaces (only) in code block, but it works well + # especially for console output, because in general the console does not + # care about trailing spaces. + lines = [] + for line in accumulated_text.splitlines(): + lines.append(line) + if line.startswith("```"): + # Code block marker - do not add trailing spaces, as it would + # break the syntax highlighting + lines.append("\n") + else: + lines.append(" \n") + markdown = Markdown("".join(lines)) + # Update the Live console output + live.update(markdown) + self._console.print() + return outputs + + +simple_io = SimpleChatIO() +rich_io = RichChatIO() +dummy_io = DummyChatIO() diff --git a/applications/ColossalChat/examples/inference/inference.py b/applications/ColossalChat/examples/inference/inference.py new file mode 100755 index 000000000000..103bd8d95016 --- /dev/null +++ b/applications/ColossalChat/examples/inference/inference.py @@ -0,0 +1,195 @@ +import argparse +import json +import os +from typing import Dict + +import torch +from chatio import dummy_io, rich_io, simple_io +from coati.dataset.conversation import setup_conversation_template +from coati.models import generate_streaming +from transformers import AutoModelForCausalLM, AutoTokenizer, PreTrainedModel + +from colossalai.logging import get_dist_logger + +logger = get_dist_logger() + + +def get_gpu_memory(max_gpus=None): + """ + Get the available memory for each GPU. + + Args: + max_gpus (int, optional): The maximum number of GPUs to consider. Defaults to None. + + Returns: + list: A list of available memory for each GPU. + """ + gpu_memory = [] + num_gpus = torch.cuda.device_count() if max_gpus is None else min(max_gpus, torch.cuda.device_count()) + + for gpu_id in range(num_gpus): + # Code to get GPU memory goes here + with torch.cuda.device(gpu_id): + device = torch.cuda.current_device() + gpu_properties = torch.cuda.get_device_properties(device) + total_memory = gpu_properties.total_memory / (1024**3) + allocated_memory = torch.cuda.memory_allocated() / (1024**3) + available_memory = total_memory - allocated_memory + gpu_memory.append(available_memory) + return gpu_memory + + +def load_model_and_tokenizer(model_path, tokenizer_path, device="cuda", **kwargs): + """ + Load the model and tokenizer from the specified paths and move the model to the specified device. + + Args: + model_path (str): The path to the pre-trained model. + tokenizer_path (str): The path to the pre-trained tokenizer. + device (str, optional): The device to move the model to. Defaults to "cuda". + **kwargs: Additional keyword arguments to be passed to the `AutoModelForCausalLM.from_pretrained` function. + + Returns: + tuple: A tuple containing the loaded model and tokenizer. + """ + + model = AutoModelForCausalLM.from_pretrained(model_path, **kwargs) + tokenizer = AutoTokenizer.from_pretrained(tokenizer_path) + tokenizer.pad_token = tokenizer.eos_token + model.to(device) + + return model, tokenizer + + +def _set_default_generate_kwargs(model: PreTrainedModel) -> Dict: + """ + Set default keyword arguments for generation based on the given model. + + Args: + model (PreTrainedModel): The model used for generation. + + Returns: + Dict: A dictionary containing the default keyword arguments for generation. + """ + unwrapped_model = model + new_kwargs = {} + # Use huggingface models method directly + if hasattr(unwrapped_model, "prepare_inputs_for_generation"): + new_kwargs["prepare_inputs_fn"] = unwrapped_model.prepare_inputs_for_generation + + if hasattr(unwrapped_model, "_update_model_kwargs_for_generation"): + new_kwargs["update_model_kwargs_fn"] = unwrapped_model._update_model_kwargs_for_generation + return new_kwargs + + +def generation_wrapper(*args, **kwargs): + input_ids = args[1] + tokenizer = args[2] + for output in generate_streaming(*args, **kwargs): + yield tokenizer.batch_decode(output[:, input_ids.size(1) :], skip_special_tokens=True)[0] + + +def main(args): + conversation_template_config = json.load(open(args.conversation_template_config, "r", encoding="utf8")) + + max_new_tokens = args.max_new_tokens + model_max_length = args.model_max_length + model, tokenizer = load_model_and_tokenizer( + args.model_path, args.tokenizer_path or args.model_path, local_files_only=True + ) + + assert max_new_tokens <= model_max_length + if hasattr(tokenizer, "pad_token") and hasattr(tokenizer, "eos_token") and tokenizer.eos_token is not None: + try: + # Some tokenizers doesn't allow to set pad_token mannually e.g., Qwen + tokenizer.pad_token = tokenizer.eos_token + except AttributeError as e: + logger.warning(f"Unable to set pad token to eos token, {str(e)}") + tokenizer.padding_side = "left" + + model_kwargs = { + "max_new_tokens": max_new_tokens, + # 'early_stopping': True, + # 'top_k': -1, + # 'top_p': 1.0, + # 'temperature': 1.0, + # 'temperature':0.1, + } + round = 1 + + conv = setup_conversation_template(tokenizer, conversation_template_config, args.conversation_template_config) + + while True: + if args.io == "simple": + chat_io = simple_io + elif args.io == "rich": + chat_io = rich_io + elif args.io == "dummy": + chat_io = dummy_io + else: + raise ValueError(f"Unknown io type: {args.io}") + # raw_text = print(">>> Human:", end=" ") + inp = chat_io.prompt_for_input("user") + + if not inp: + print("prompt should not be empty!") + continue + + if inp.strip() == "clear": + conv.clear() + os.system("clear") + continue + + if inp.strip() == "exit": + print("End of chat.") + break + + query_text = inp.strip() + + conv.append_message("user", query_text) + + chat_io.prompt_for_output("assistant") + + prompt = conv.get_prompt(add_generation_prompt=True) + print(prompt + "") + input_ids = tokenizer(prompt, return_tensors="pt", add_special_tokens=False)["input_ids"].to( + torch.cuda.current_device() + ) + default_generate_kwargs = _set_default_generate_kwargs(model) + model_kwargs.update(default_generate_kwargs) + output_stream = generation_wrapper( + model, + input_ids, + tokenizer, + max_length=model_max_length, + temperature=0.7, + early_stopping=True, + stop_token_ids=conversation_template_config["stop_ids"], + **model_kwargs, + ) + + # print(f">>> Assistant:", end=" ") + outputs = chat_io.stream_output(output_stream) + + conv.append_message("assistant", outputs.strip()) + + with open("round.txt", mode="a", encoding="utf-8") as f: + f.write("\n\n" + "=" * 10 + "\n") + f.write(f"round {round}:\n{conv.save_prompt()}\n\n") + f.write("=" * 10 + "\n") + + # print(f">>> Assistant:", end=" ") + + round += 1 + + +if __name__ == "__main__": + parser = argparse.ArgumentParser() + parser.add_argument("--model_path", type=str, default=None) + parser.add_argument("--tokenizer_path", type=str, default=None) + parser.add_argument("--conversation_template_config", type=str, default=None) + parser.add_argument("--model_max_length", type=int, default=2048) + parser.add_argument("--max_new_tokens", type=int, default=512) + parser.add_argument("--io", type=str, default="rich", choices=["simple", "rich", "dummy"]) + args = parser.parse_args() + main(args) diff --git a/applications/Chat/inference/README.md b/applications/ColossalChat/examples/inference/web_chatbot/README.md old mode 100644 new mode 100755 similarity index 100% rename from applications/Chat/inference/README.md rename to applications/ColossalChat/examples/inference/web_chatbot/README.md diff --git a/applications/Chat/inference/locustfile.py b/applications/ColossalChat/examples/inference/web_chatbot/locustfile.py old mode 100644 new mode 100755 similarity index 100% rename from applications/Chat/inference/locustfile.py rename to applications/ColossalChat/examples/inference/web_chatbot/locustfile.py diff --git a/applications/Chat/inference/requirements.txt b/applications/ColossalChat/examples/inference/web_chatbot/requirements.txt old mode 100644 new mode 100755 similarity index 100% rename from applications/Chat/inference/requirements.txt rename to applications/ColossalChat/examples/inference/web_chatbot/requirements.txt diff --git a/applications/Chat/inference/server.py b/applications/ColossalChat/examples/inference/web_chatbot/server.py old mode 100644 new mode 100755 similarity index 79% rename from applications/Chat/inference/server.py rename to applications/ColossalChat/examples/inference/web_chatbot/server.py index 7c6a61b9e7f2..aec342802b02 --- a/applications/Chat/inference/server.py +++ b/applications/ColossalChat/examples/inference/web_chatbot/server.py @@ -5,6 +5,7 @@ import torch import uvicorn +from coati.models import generate_streaming from coati.quant import llama_load_quant, low_resource_init from fastapi import FastAPI, Request from fastapi.middleware.cors import CORSMiddleware @@ -13,10 +14,9 @@ from slowapi.errors import RateLimitExceeded from slowapi.util import get_remote_address from sse_starlette.sse import EventSourceResponse -from transformers import AutoTokenizer, LlamaConfig, LlamaForCausalLM -from utils import ChatPromptProcessor, Dialogue, LockedIterator, load_json, sample_streamingly, update_model_kwargs_fn +from transformers import AutoConfig, AutoModelForCausalLM, AutoTokenizer +from utils import ChatPromptProcessor, Dialogue, LockedIterator, load_json, update_model_kwargs_fn -CONTEXT = "Below is an instruction that describes a task. Write a response that appropriately completes the request. Do not generate new instructions." MAX_LEN = 512 running_lock = Lock() @@ -54,20 +54,22 @@ class GenerationTaskReq(BaseModel): ) -def generate_streamingly(prompt, max_new_tokens, top_k, top_p, temperature): - inputs = {k: v.cuda() for k, v in tokenizer(prompt, return_tensors="pt").items()} +def generate_streamingly(prompt, max_length, max_new_tokens, top_k, top_p, temperature): + input_ids = tokenizer(prompt, return_tensors="pt")["input_ids"] # TODO(ver217): streaming generation does not support repetition_penalty now model_kwargs = { - "max_generate_tokens": max_new_tokens, + "max_new_tokens": max_new_tokens, "early_stopping": True, "top_k": top_k, "top_p": top_p, "temperature": temperature, - "prepare_inputs_fn": model.prepare_inputs_for_generation, + "prepare_inputs_fn": None, "update_model_kwargs_fn": update_model_kwargs_fn, } is_first_word = True - generator = LockedIterator(sample_streamingly(model, **inputs, **model_kwargs), running_lock) + generator = LockedIterator( + generate_streaming(model, input_ids, tokenizer, max_length, **model_kwargs), running_lock + ) for output in generator: output = output.cpu() tokens = tokenizer.convert_ids_to_tokens(output, skip_special_tokens=True) @@ -101,9 +103,10 @@ async def event_generator(request: Request, generator: Generator): @app.post("/generate/stream") @limiter.limit("1/second") def generate(data: GenerationTaskReq, request: Request): - prompt = prompt_processor.preprocess_prompt(data.history, data.max_new_tokens) + prompt = prompt_processor.preprocess_prompt(data.history) event_source = event_generator( - request, generate_streamingly(prompt, data.max_new_tokens, data.top_k, data.top_p, data.temperature) + request, + generate_streamingly(prompt, data.max_length, data.max_new_tokens, data.top_k, data.top_p, data.temperature), ) return EventSourceResponse(event_source) @@ -133,6 +136,11 @@ def generate_no_stream(data: GenerationTaskReq, request: Request): "pretrained", help="Path to pretrained model. Can be a local path or a model name from the HuggingFace model hub.", ) + parser.add_argument( + "--tokenizer_path", + help="Path to pretrained tokenizer. Can be a local path or a model name from the HuggingFace model hub.", + default=None, + ) parser.add_argument( "--quant", choices=["8bit", "4bit"], @@ -162,26 +170,29 @@ def generate_no_stream(data: GenerationTaskReq, request: Request): if args.quant == "4bit": assert args.gptq_checkpoint is not None, "Please specify a GPTQ checkpoint." - tokenizer = AutoTokenizer.from_pretrained(args.pretrained) + if args.tokenizer_path is None: + args.tokenizer_path = args.pretrained + tokenizer = AutoTokenizer.from_pretrained(args.tokenizer_path, local_files_only=True) if args.profanity_file is not None: censored_words = load_json(args.profanity_file) else: censored_words = [] - prompt_processor = ChatPromptProcessor(tokenizer, CONTEXT, MAX_LEN, censored_words=censored_words) + prompt_processor = ChatPromptProcessor(censored_words=censored_words) if args.quant == "4bit": with low_resource_init(): - config = LlamaConfig.from_pretrained(args.pretrained) - model = LlamaForCausalLM(config) + config = AutoConfig.from_pretrained(args.pretrained) + model = AutoModelForCausalLM(config) model = llama_load_quant(model, args.gptq_checkpoint, 4, args.gptq_group_size) model.cuda() else: - model = LlamaForCausalLM.from_pretrained( + model = AutoModelForCausalLM.from_pretrained( args.pretrained, load_in_8bit=(args.quant == "8bit"), torch_dtype=torch.float16, device_map="auto", + local_files_only=True, ) if args.quant != "8bit": model.half() # seems to fix bugs for some users. @@ -190,3 +201,8 @@ def generate_no_stream(data: GenerationTaskReq, request: Request): config = uvicorn.Config(app, host=args.http_host, port=args.http_port) server = uvicorn.Server(config=config) server.run() + + +""" +python server.py /home/lcyab/data/models/experiments5/checkpoint/experiment5-2023-10-20-21-53-51/modeling/ --tokenizer_path /mnt/vepfs/lcxyc/leaderboard_models/Colossal-LLaMA-2-7b-base/ +""" diff --git a/applications/ColossalChat/examples/inference/web_chatbot/utils.py b/applications/ColossalChat/examples/inference/web_chatbot/utils.py new file mode 100755 index 000000000000..82a1a7255164 --- /dev/null +++ b/applications/ColossalChat/examples/inference/web_chatbot/utils.py @@ -0,0 +1,78 @@ +import copy +import json +from threading import Lock +from typing import List + +import jieba +import torch +from coati.dataset.conversation import default_conversation +from pydantic import BaseModel, Field + + +def update_model_kwargs_fn(outputs: dict, **model_kwargs) -> dict: + if "past_key_values" in outputs: + model_kwargs["past"] = outputs["past_key_values"] + else: + model_kwargs["past"] = None + + # update token_type_ids with last value + if "token_type_ids" in model_kwargs: + token_type_ids = model_kwargs["token_type_ids"] + model_kwargs["token_type_ids"] = torch.cat([token_type_ids, token_type_ids[:, -1].unsqueeze(-1)], dim=-1) + + # update attention mask + if "attention_mask" in model_kwargs: + attention_mask = model_kwargs["attention_mask"] + model_kwargs["attention_mask"] = torch.cat( + [attention_mask, attention_mask.new_ones((attention_mask.shape[0], 1))], dim=-1 + ) + + return model_kwargs + + +class Dialogue(BaseModel): + instruction: str = Field(min_length=1, example="Count up from 1 to 500.") + response: str = Field(example="") + + +class ChatPromptProcessor: + SAFE_RESPONSE = "The input/response contains inappropriate content, please rephrase your prompt." + + def __init__(self, censored_words: List[str] = []): + self.censored_words = set([word.lower() for word in censored_words]) + self.conv = copy.deepcopy(default_conversation) + + def preprocess_prompt(self, history: List[Dialogue]) -> str: + self.conv.clear() + for round in history: + self.conv.append_message(self.conv.roles[0], round.instruction) + if len(round.instruction) > 0: + self.conv.append_message(self.conv.roles[1], round.response) + return self.conv.get_prompt() + + def postprocess_output(self, output: str) -> str: + return output.strip() + + def has_censored_words(self, text: str) -> bool: + if len(self.censored_words) == 0: + return False + intersection = set(jieba.cut(text.lower())) & self.censored_words + return len(intersection) > 0 + + +class LockedIterator: + def __init__(self, it, lock: Lock) -> None: + self.lock = lock + self.it = iter(it) + + def __iter__(self): + return self + + def __next__(self): + with self.lock: + return next(self.it) + + +def load_json(path: str): + with open(path) as f: + return json.load(f) diff --git a/applications/Chat/examples/ray/1mmt_prompt.py b/applications/ColossalChat/examples/ray/1mmt_prompt.py old mode 100644 new mode 100755 similarity index 100% rename from applications/Chat/examples/ray/1mmt_prompt.py rename to applications/ColossalChat/examples/ray/1mmt_prompt.py diff --git a/applications/Chat/examples/ray/mmmt_prompt.py b/applications/ColossalChat/examples/ray/mmmt_prompt.py old mode 100644 new mode 100755 similarity index 100% rename from applications/Chat/examples/ray/mmmt_prompt.py rename to applications/ColossalChat/examples/ray/mmmt_prompt.py diff --git a/applications/Chat/examples/ray/requirements.txt b/applications/ColossalChat/examples/ray/requirements.txt old mode 100644 new mode 100755 similarity index 100% rename from applications/Chat/examples/ray/requirements.txt rename to applications/ColossalChat/examples/ray/requirements.txt diff --git a/applications/Chat/examples/ray/test_ci.sh b/applications/ColossalChat/examples/ray/test_ci.sh similarity index 100% rename from applications/Chat/examples/ray/test_ci.sh rename to applications/ColossalChat/examples/ray/test_ci.sh diff --git a/applications/Chat/examples/requirements.txt b/applications/ColossalChat/examples/requirements.txt similarity index 51% rename from applications/Chat/examples/requirements.txt rename to applications/ColossalChat/examples/requirements.txt index 5474dfa16b3e..838590f4b103 100644 --- a/applications/Chat/examples/requirements.txt +++ b/applications/ColossalChat/examples/requirements.txt @@ -1,3 +1,4 @@ pandas>=1.4.1 sentencepiece -colossalai==0.3.3 +colossalai +prompt_toolkit diff --git a/applications/ColossalChat/examples/training_scripts/hostfile b/applications/ColossalChat/examples/training_scripts/hostfile new file mode 100755 index 000000000000..d4118dda9783 --- /dev/null +++ b/applications/ColossalChat/examples/training_scripts/hostfile @@ -0,0 +1 @@ +10.20.1.82 diff --git a/applications/ColossalChat/examples/training_scripts/train_dpo.py b/applications/ColossalChat/examples/training_scripts/train_dpo.py new file mode 100755 index 000000000000..b9287eb1a407 --- /dev/null +++ b/applications/ColossalChat/examples/training_scripts/train_dpo.py @@ -0,0 +1,326 @@ +import argparse +import json +import os +import resource +from contextlib import nullcontext + +import torch +from coati.dataset import ( + DataCollatorForPreferenceDataset, + StatefulDistributedSampler, + load_tokenized_dataset, + setup_distributed_dataloader, +) +from coati.models import convert_to_lora_module, disable_dropout +from coati.trainer import DPOTrainer +from coati.utils import load_checkpoint +from transformers import AutoModelForCausalLM, AutoTokenizer + +import colossalai +from colossalai.booster import Booster +from colossalai.booster.plugin import GeminiPlugin, HybridParallelPlugin, LowLevelZeroPlugin +from colossalai.cluster import DistCoordinator +from colossalai.logging import get_dist_logger +from colossalai.nn.lr_scheduler import CosineAnnealingWarmupLR +from colossalai.nn.optimizer import HybridAdam + +logger = get_dist_logger() + + +def train(args): + # check lora compatibility + if "gemini" in args.plugin and args.lora_rank > 0: + raise ValueError("LoRA is not supported in GeminiPlugin. Please use other plugin") + if args.plugin == "gemini_auto" and args.accumulation_steps > 1: + raise ValueError("Gradient accumulation is not supported in GeminiPlugin. Please use other plugin") + + # ============================== + # Initialize Distributed Training + # ============================== + colossalai.launch_from_torch({}) + coordinator = DistCoordinator() + + # ============================== + # Initialize Booster + # ============================== + if args.plugin == "ddp": + """ + Default torch ddp plugin without any acceleration, for + debugging purpose acceleration, for debugging purpose + """ + plugin = TorchDDPPlugin(find_unused_parameters=True) + elif args.plugin == "gemini": + plugin = GeminiPlugin( + precision=args.mixed_precision, + placement_policy="static", + initial_scale=2**16, + max_norm=args.grad_clip, + enable_gradient_accumulation=True, + ) + elif args.plugin == "gemini_auto": + plugin = GeminiPlugin( + precision=args.mixed_precision, + placement_policy="auto", + initial_scale=2**16, + max_norm=args.grad_clip, + ) + elif args.plugin == "zero2": + plugin = LowLevelZeroPlugin( + stage=2, + precision=args.mixed_precision, + initial_scale=2**16, + max_norm=args.grad_clip, + ) + elif args.plugin == "zero2_cpu": + plugin = LowLevelZeroPlugin( + stage=2, + precision=args.mixed_precision, + initial_scale=2**16, + cpu_offload=True, + max_norm=args.grad_clip, + ) + elif args.plugin == "3d": + plugin = HybridParallelPlugin( + tp_size=args.tp, + pp_size=1, + zero_stage=0, + parallel_output=False, + precision=args.mixed_precision, + ) + else: + raise ValueError(f"Unknown plugin {args.plugin}") + + booster = Booster(plugin=plugin) + ref_booster = Booster(plugin=plugin) + + # ====================================================== + # Initialize Model, Objective, Optimizer and LR Scheduler + # ====================================================== + # Temp Fix: Disable lazy init due to version conflict + # init_ctx = ( + # LazyInitContext(default_device=get_current_device()) if isinstance(plugin, (GeminiPlugin,)) else nullcontext() + # ) + + init_ctx = nullcontext() + with init_ctx: + if args.use_flash_attn: + model = AutoModelForCausalLM.from_pretrained( + args.pretrain, + torch_dtype=torch.bfloat16 if args.mixed_precision == "bf16" else torch.float16, + use_flash_attention_2=True, + ) + coordinator.print_on_master(msg="Flash-attention enabled successfully") + else: + model = AutoModelForCausalLM.from_pretrained(args.pretrain) + disable_dropout(model) + if args.enable_reference_model: + if args.use_flash_attn: + ref_model = AutoModelForCausalLM.from_pretrained( + args.pretrain, + torch_dtype=torch.bfloat16 if args.mixed_precision == "bf16" else torch.float16, + use_flash_attention_2=True, + ) + else: + ref_model = AutoModelForCausalLM.from_pretrained(args.pretrain) + disable_dropout(ref_model) + else: + ref_model = None + + if args.lora_rank > 0: + model = convert_to_lora_module(model, args.lora_rank, lora_train_bias=args.lora_train_bias) + + if args.grad_checkpoint and args.lora_rank == 0: + model.gradient_checkpointing_enable() + coordinator.print_on_master(msg="Gradient checkpointing enabled successfully") + elif args.lora_rank > 0: + coordinator.print_on_master(msg="Gradient checkpointing will be disabled when LoRA is enabled") + + # configure tokenizer + tokenizer_dir = args.tokenizer_dir if args.tokenizer_dir is not None else args.pretrain + tokenizer = AutoTokenizer.from_pretrained(tokenizer_dir, use_fast=False, trust_remote_code=True) + if hasattr(tokenizer, "pad_token") and hasattr(tokenizer, "eos_token") and tokenizer.eos_token is not None: + try: + # Some tokenizers doesn't allow to set pad_token mannually e.g., Qwen + tokenizer.pad_token = tokenizer.eos_token + except AttributeError as e: + logger.warning(f"Unable to set pad token to eos token, {str(e)}") + if not hasattr(tokenizer, "pad_token") or tokenizer.pad_token is None: + logger.warning( + "The tokenizer does not have a pad token which is required. May lead to unintended behavior in training, Please consider manually set them." + ) + + tokenizer.add_bos_token = False + tokenizer.add_eos_token = False + + # configure optimizer + optim = HybridAdam( + model_params=model.parameters(), + lr=args.lr, + betas=(0.9, 0.95), + weight_decay=args.weight_decay, + adamw_mode=True, + ) + + # configure dataset + coordinator.print_on_master(f"Load dataset: {args.dataset}") + mode_map = {"train": "train", "valid": "validation", "test": "test"} + train_dataset = load_tokenized_dataset(dataset_paths=args.dataset, mode="train", mode_map=mode_map) + data_collator = DataCollatorForPreferenceDataset(tokenizer=tokenizer, max_length=args.max_length) + train_dataloader = setup_distributed_dataloader( + dataset=train_dataset, + batch_size=args.batch_size, + shuffle=True, + drop_last=True, + collate_fn=data_collator, + use_tp=args.tp > 1, + ) + + num_update_steps_per_epoch = len(train_dataloader) // args.accumulation_steps + if args.warmup_steps is None: + args.warmup_steps = int(args.max_epochs * 0.025 * (len(train_dataloader) // args.accumulation_steps)) + coordinator.print_on_master(f"Warmup steps is set to {args.warmup_steps}") + + lr_scheduler = CosineAnnealingWarmupLR( + optimizer=optim, + total_steps=args.max_epochs * num_update_steps_per_epoch, + warmup_steps=args.warmup_steps, + eta_min=0.1 * args.lr, + ) + + default_dtype = torch.float16 if args.mixed_precision == "fp16" else torch.bfloat16 + torch.set_default_dtype(default_dtype) + model, optim, _, train_dataloader, lr_scheduler = booster.boost( + model=model, + optimizer=optim, + lr_scheduler=lr_scheduler, + dataloader=train_dataloader, + ) + if ref_model is not None: + ref_model, _, _, _, _ = ref_booster.boost(model=ref_model, dataloader=train_dataloader) + torch.set_default_dtype(torch.float) + + coordinator.print_on_master(f"Booster init max CUDA memory: {torch.cuda.max_memory_allocated() / 1024 ** 2:.2f} MB") + coordinator.print_on_master( + f"Booster init max CPU memory: {resource.getrusage(resource.RUSAGE_SELF).ru_maxrss / 1024:.2f} MB" + ) + + start_epoch = 0 + sampler_start_idx = 0 + start_step = 0 + if args.checkpoint_path is not None: + if "modeling" in args.checkpoint_path: + coordinator.print_on_master(f"Continued pretrain from checkpoint {args.checkpoint_path}") + booster.load_model(model, args.checkpoint_path) + else: + coordinator.print_on_master(f"Load model checkpoint from {args.checkpoint_path}") + start_epoch, start_step, sampler_start_idx = load_checkpoint( + load_dir=args.checkpoint_path, + booster=booster, + model=model, + optimizer=optim, + lr_scheduler=lr_scheduler, + ) + assert isinstance(train_dataloader.sampler, StatefulDistributedSampler) + train_dataloader.sampler.set_start_index(start_index=sampler_start_idx) + + coordinator.print_on_master( + f"Loaded checkpoint {args.checkpoint_path} at epoch {start_epoch} step {start_step}" + ) + coordinator.print_on_master(f"Loaded sample at index {sampler_start_idx}") + + coordinator.print_on_master( + f"Checkpoint loaded max CUDA memory: {torch.cuda.max_memory_allocated() / 1024 ** 2:.2f} MB" + ) + coordinator.print_on_master( + f"Checkpoint loaded CUDA memory: {torch.cuda.memory_allocated() / 1024 ** 2:.2f} MB" + ) + coordinator.print_on_master( + f"Checkpoint loaded max CPU memory: {resource.getrusage(resource.RUSAGE_SELF).ru_maxrss / 1024:.2f} MB" + ) + + trainer = DPOTrainer( + actor=model, + ref_model=ref_model, + booster=booster, + actor_optim=optim, + actor_lr_scheduler=lr_scheduler, + tokenizer=tokenizer, + max_epochs=args.max_epochs, + accumulation_steps=args.accumulation_steps, + start_epoch=start_epoch, + save_interval=args.save_interval, + save_dir=args.save_dir, + coordinator=coordinator, + ) + + trainer.fit( + train_preference_dataloader=train_dataloader, + eval_preference_dataloader=None, + log_dir=args.log_dir, + use_wandb=args.use_wandb, + ) + + if args.lora_rank > 0 and args.merge_lora_weights: + from coati.models.lora import LORA_MANAGER + + # NOTE: set model to eval to merge LoRA weights + LORA_MANAGER.merge_weights = True + model.eval() + # save model checkpoint after fitting on only rank0 + coordinator.print_on_master("Start saving final model checkpoint") + booster.save_model(model, os.path.join(args.save_dir, "modeling"), shard=True) + coordinator.print_on_master(f"Saved final model checkpoint at epoch {args.max_epochs} at folder {args.save_dir}") + + coordinator.print_on_master(f"Max CUDA memory usage: {torch.cuda.max_memory_allocated()/1024**2:.2f} MB") + + +if __name__ == "__main__": + # ============================== + # Parse Arguments + # ============================== + parser = argparse.ArgumentParser() + parser.add_argument( + "--plugin", + type=str, + default="gemini", + choices=["gemini", "gemini_auto", "zero2", "zero2_cpu", "3d"], + help="Choose which plugin to use", + ) + parser.add_argument("--grad_clip", type=float, default=1.0, help="Gradient clipping value") + parser.add_argument("--weight_decay", type=float, default=0.1, help="Weight decay") + parser.add_argument("--warmup_steps", type=int, default=None, help="Warmup steps") + parser.add_argument("--tp", type=int, default=1) + parser.add_argument("--pretrain", type=str, default=None) + parser.add_argument("--model_type", type=str, default=None) + parser.add_argument("--tokenizer_dir", type=str, default=None) + parser.add_argument("--dataset", nargs="+", default=[]) + parser.add_argument( + "--checkpoint_path", type=str, default=None, help="Checkpoint path if need to resume training form a checkpoint" + ) + parser.add_argument("--config_file", type=str, default="config_file", help="Config file") + parser.add_argument("--save_dir", type=str, default="output") + parser.add_argument("--max_length", type=int, default=2048, help="Model max length") + parser.add_argument("--max_epochs", type=int, default=3) + parser.add_argument("--batch_size", type=int, default=4) + parser.add_argument("--enable_reference_model", type=bool, default=True) + parser.add_argument("--mixed_precision", type=str, default="fp16", choices=["fp16", "bf16"], help="Mixed precision") + parser.add_argument("--lora_rank", type=int, default=0, help="low-rank adaptation matrices rank") + parser.add_argument( + "--lora_train_bias", + type=str, + default="none", + help="'none' means it doesn't train biases. 'all' means it trains all biases. 'lora_only' means it only trains biases of LoRA layers", + ) + parser.add_argument("--save_interval", type=int, default=1000, help="number of step between two checkpoints") + parser.add_argument("--merge_lora_weights", type=bool, default=True) + parser.add_argument("--lr", type=float, default=5e-6) + parser.add_argument("--accumulation_steps", type=int, default=8) + parser.add_argument("--log_dir", default="logs", type=str) + parser.add_argument("--use_wandb", default=False, action="store_true") + parser.add_argument("--grad_checkpoint", default=False, action="store_true") + parser.add_argument("--use_flash_attn", default=False, action="store_true") + args = parser.parse_args() + os.makedirs(os.path.dirname(args.config_file), exist_ok=True) + with open(args.config_file, "w") as f: + json.dump(args.__dict__, f, indent=4) + train(args) diff --git a/applications/ColossalChat/examples/training_scripts/train_dpo.sh b/applications/ColossalChat/examples/training_scripts/train_dpo.sh new file mode 100755 index 000000000000..80fc30c3d955 --- /dev/null +++ b/applications/ColossalChat/examples/training_scripts/train_dpo.sh @@ -0,0 +1,62 @@ +#!/bin/bash +set_n_least_used_CUDA_VISIBLE_DEVICES() { + local n=${1:-"9999"} + echo "GPU Memory Usage:" + local FIRST_N_GPU_IDS=$(nvidia-smi --query-gpu=memory.used --format=csv | + tail -n +2 | + nl -v 0 | + tee /dev/tty | + sort -g -k 2 | + awk '{print $1}' | + head -n $n) + export CUDA_VISIBLE_DEVICES=$(echo $FIRST_N_GPU_IDS | sed 's/ /,/g') + echo "Now CUDA_VISIBLE_DEVICES is set to:" + echo "CUDA_VISIBLE_DEVICES=$CUDA_VISIBLE_DEVICES" +} +set_n_least_used_CUDA_VISIBLE_DEVICES 8 +# export CUDA_VISIBLE_DEVICES=6 + +PROJECT_NAME="dpo" +PARENT_SAVE_DIR="" # Path to a folder to save checkpoints +PARENT_TENSORBOARD_DIR="" # Path to a folder to save logs +PARENT_CONFIG_FILE="" # Path to a folder to save training config logs +PRETRAINED_MODEL_PATH="" # huggingface or local model path +PRETRAINED_TOKENIZER_PATH="" # huggingface or local tokenizer path + +declare -a dataset=( + YOUR/DATA/DIR/arrow/part-00000 + YOUR/DATA/DIR/arrow/part-00001 + YOUR/DATA/DIR/arrow/part-00002 + YOUR/DATA/DIR/arrow/part-00003 + YOUR/DATA/DIR/arrow/part-00004 + YOUR/DATA/DIR/arrow/part-00005 + YOUR/DATA/DIR/arrow/part-00006 + YOUR/DATA/DIR/arrow/part-00007 + YOUR/DATA/DIR/arrow/part-00008 + YOUR/DATA/DIR/arrow/part-00009 +) + +TIMESTAMP=$(date +%Y-%m-%d-%H-%M-%S) +FULL_PROJECT_NAME="${PROJECT_NAME}-${TIMESTAMP}" +SAVE_DIR="${PARENT_SAVE_DIR}${FULL_PROJECT_NAME}" +CONFIG_FILE="${PARENT_CONFIG_FILE}-${FULL_PROJECT_NAME}.json" + +colossalai run --nproc_per_node 8 --hostfile hostfile --master_port 31312 train_dpo.py \ + --pretrain $PRETRAINED_MODEL_PATH \ + --checkpoint_path $PRETRAINED_MODEL_PATH \ + --tokenizer_dir $PRETRAINED_TOKENIZER_PATH \ + --dataset ${dataset[@]} \ + --plugin "zero2" \ + --save_interval 1000 \ + --save_dir $SAVE_DIR \ + --config_file $CONFIG_FILE \ + --max_epochs 1 \ + --accumulation_steps 4 \ + --batch_size 2 \ + --lr 1e-6 \ + --mixed_precision "bf16" \ + --grad_clip 1.0 \ + --weight_decay 0.01 \ + --warmup_steps 100 \ + --grad_checkpoint \ + --use_wandb diff --git a/applications/ColossalChat/examples/training_scripts/train_ppo.py b/applications/ColossalChat/examples/training_scripts/train_ppo.py new file mode 100755 index 000000000000..7c91fa347847 --- /dev/null +++ b/applications/ColossalChat/examples/training_scripts/train_ppo.py @@ -0,0 +1,506 @@ +import argparse +import json +import os +import resource +from contextlib import nullcontext + +import torch +import torch.distributed as dist +from coati.dataset import ( + DataCollatorForPromptDataset, + DataCollatorForSupervisedDataset, + StatefulDistributedSampler, + load_tokenized_dataset, + setup_conversation_template, + setup_distributed_dataloader, +) +from coati.models import Critic, RewardModel, convert_to_lora_module, disable_dropout +from coati.trainer import PPOTrainer +from coati.utils import load_checkpoint +from transformers import AutoModelForCausalLM, AutoTokenizer + +import colossalai +from colossalai.booster import Booster +from colossalai.booster.plugin import GeminiPlugin, HybridParallelPlugin, LowLevelZeroPlugin +from colossalai.cluster import DistCoordinator +from colossalai.logging import get_dist_logger +from colossalai.nn.lr_scheduler import CosineAnnealingWarmupLR +from colossalai.nn.optimizer import HybridAdam + +logger = get_dist_logger() + + +def train(args): + # check lora compatibility + if "gemini" in args.plugin and args.lora_rank > 0: + raise ValueError("LoRA is not supported in GeminiPlugin. Please use other plugin") + if args.plugin == "gemini_auto" and args.accumulation_steps > 1: + raise ValueError("Gradient accumulation is not supported in GeminiPlugin. Please use other plugin") + # ============================== + # Initialize Distributed Training + # ============================== + colossalai.launch_from_torch({}) + coordinator = DistCoordinator() + + # ====================================================== + # Initialize Model, Objective, Optimizer and LR Scheduler + # ====================================================== + # Temp Fix: Disable lazy init due to version conflict + # init_ctx = ( + # LazyInitContext(default_device=get_current_device()) if isinstance(plugin, (GeminiPlugin,)) else nullcontext() + # ) + + init_ctx = nullcontext() + booster_policy = None + with init_ctx: + if args.use_flash_attn: + actor = AutoModelForCausalLM.from_pretrained( + args.pretrain, + torch_dtype=torch.bfloat16 if args.mixed_precision == "bf16" else torch.float16, + use_flash_attention_2=True, + local_files_only=True, + ) + ref_model = AutoModelForCausalLM.from_pretrained( + args.pretrain, + torch_dtype=torch.bfloat16 if args.mixed_precision == "bf16" else torch.float16, + use_flash_attention_2=True, + local_files_only=True, + ) + reward_model = RewardModel( + args.rm_pretrain, + torch_dtype=torch.bfloat16 if args.mixed_precision == "bf16" else torch.float16, + use_flash_attention_2=True, + ) + critic = Critic( + args.rm_pretrain, + torch_dtype=torch.bfloat16 if args.mixed_precision == "bf16" else torch.float16, + use_flash_attention_2=True, + ) + coordinator.print_on_master(msg="Flash-attention enabled successfully") + else: + actor = AutoModelForCausalLM.from_pretrained(args.pretrain, local_files_only=True) + ref_model = AutoModelForCausalLM.from_pretrained(args.pretrain, local_files_only=True) + reward_model = RewardModel(args.rm_pretrain) + critic = Critic(args.rm_pretrain) + # Disable dropout + disable_dropout(actor) + disable_dropout(critic) + + if args.tp > 1: + if reward_model.model.config.architectures[0] != critic.model.config.architectures[0]: + raise ValueError("Reward model and critic model must have the same architecture") + if reward_model.model.config.architectures[0] == "BloomForCausalLM": + from colossalai.shardformer.policies.bloom import BloomPolicy + + booster_policy = BloomPolicy() + elif reward_model.model.config.architectures[0] == "LlamaForCausalLM": + from colossalai.shardformer.policies.llama import LlamaPolicy + + booster_policy = LlamaPolicy() + elif reward_model.model.config.architectures[0] == "GPT2LMHeadModel": + from colossalai.shardformer.policies.gpt2 import GPT2Policy + + booster_policy = GPT2Policy() + elif reward_model.model.config.architectures[0] == "ChatGLMModel": + from colossalai.shardformer.policies.chatglm2 import ChatGLMPolicy + + booster_policy = ChatGLMPolicy() + elif reward_model.model.config.architectures[0] == "OPTForCausalLM": + from colossalai.shardformer.policies.opt import OPTPolicy + + booster_policy = OPTPolicy() + else: + raise ValueError("Unknown model architecture for policy") + + if args.lora_rank > 0: + actor = convert_to_lora_module(actor, args.lora_rank, lora_train_bias=args.lora_train_bias) + critic = convert_to_lora_module(critic, args.lora_rank, lora_train_bias=args.lora_train_bias) + + if args.grad_checkpoint and args.lora_rank == 0: + actor.gradient_checkpointing_enable() + critic.model.gradient_checkpointing_enable() + coordinator.print_on_master(msg="Gradient checkpointing enabled successfully") + elif args.lora_rank > 0: + coordinator.print_on_master(msg="Gradient checkpointing will be disabled when LoRA is enabled") + + # configure tokenizer + tokenizer_dir = args.tokenizer_dir if args.tokenizer_dir is not None else args.pretrain + tokenizer = AutoTokenizer.from_pretrained(tokenizer_dir, use_fast=False, trust_remote_code=True) + if os.path.exists(args.conversation_template_config): + with open(args.conversation_template_config, "r", encoding="utf8") as f: + conversation_template_config = json.load(f) + dist.barrier() + conversation_template = setup_conversation_template( + tokenizer, chat_template_config=conversation_template_config, save_path=args.conversation_template_config + ) + stop_ids = conversation_template.stop_ids if len(conversation_template.stop_ids) > 0 else None + else: + raise ValueError("Conversation template config is not provided or incorrect") + if hasattr(tokenizer, "pad_token") and hasattr(tokenizer, "eos_token") and tokenizer.eos_token is not None: + try: + # Some tokenizers doesn't allow to set pad_token mannually e.g., Qwen + tokenizer.pad_token = tokenizer.eos_token + except AttributeError as e: + logger.warning(f"Unable to set pad token to eos token, {str(e)}") + if not hasattr(tokenizer, "pad_token") or tokenizer.pad_token is None: + logger.warning( + "The tokenizer does not have a pad token which is required. May lead to unintended behavior in training, Please consider manually set them." + ) + + tokenizer.add_bos_token = False + tokenizer.add_eos_token = False + tokenizer.padding_side = "left" # left padding for generation (online learning) + + # configure generation config + actor.generation_config.update( + pad_token_id=tokenizer.eos_token_id, bos_token_id=tokenizer.bos_token_id, eos_token_id=tokenizer.eos_token_id + ) + + # configure optimizer + coordinator.print_on_master(f"setting up optimizer for actor: lr={args.lr}, weight_decay={args.weight_decay}") + actor_optim = HybridAdam( + model_params=actor.parameters(), + lr=args.lr, + betas=(0.9, 0.95), + weight_decay=args.weight_decay, + adamw_mode=True, + ) + + coordinator.print_on_master(f"setting up optimizer for critic: lr={args.lr}, weight_decay={args.weight_decay}") + critic_optim = HybridAdam( + model_params=critic.parameters(), + lr=args.critic_lr, + betas=(0.9, 0.95), + weight_decay=args.weight_decay, + adamw_mode=True, + ) + + # configure dataset + coordinator.print_on_master(f"Load dataset: {args.prompt_dataset}") + mode_map = {"train": "train", "valid": "validation", "test": "test"} + train_prompt_dataset = load_tokenized_dataset(dataset_paths=args.prompt_dataset, mode="train", mode_map=mode_map) + data_collator = DataCollatorForPromptDataset(tokenizer=tokenizer, max_length=args.max_length - args.max_seq_len) + train_prompt_dataloader = setup_distributed_dataloader( + dataset=train_prompt_dataset, + batch_size=args.experience_batch_size, + shuffle=True, + drop_last=True, + collate_fn=data_collator, + use_tp=args.tp > 1, + ) + + if len(args.ptx_dataset) > 0: + train_ptx_dataset = load_tokenized_dataset(dataset_paths=args.ptx_dataset, mode="train", mode_map=mode_map) + data_collator = DataCollatorForSupervisedDataset(tokenizer=tokenizer, max_length=args.max_length) + train_pretrain_dataloader = setup_distributed_dataloader( + dataset=train_ptx_dataset, + batch_size=args.ptx_batch_size, + shuffle=True, + drop_last=True, + collate_fn=data_collator, + use_tp=args.tp > 1, + ) + else: + train_pretrain_dataloader = None + + if args.warmup_steps is None: + args.warmup_steps = int(0.025 * args.num_episodes) + coordinator.print_on_master(f"Warmup steps is set to {args.warmup_steps}") + + actor_lr_scheduler = CosineAnnealingWarmupLR( + optimizer=actor_optim, + total_steps=args.num_episodes, + warmup_steps=args.warmup_steps, + eta_min=0.1 * args.lr, + ) + + critic_lr_scheduler = CosineAnnealingWarmupLR( + optimizer=critic_optim, + total_steps=args.num_episodes, + warmup_steps=args.warmup_steps, + eta_min=0.1 * args.lr, + ) + + # ============================== + # Initialize Booster + # ============================== + if args.plugin == "ddp": + """ + Default torch ddp plugin without any acceleration, for + debugging purpose acceleration, for debugging purpose + """ + plugin = TorchDDPPlugin(find_unused_parameters=True) + elif args.plugin == "gemini": + plugin = GeminiPlugin( + precision=args.mixed_precision, + placement_policy="static", + initial_scale=2**16, + max_norm=args.grad_clip, + enable_gradient_accumulation=True, + ) + elif args.plugin == "gemini_auto": + plugin = GeminiPlugin( + precision=args.mixed_precision, + placement_policy="auto", + initial_scale=2**16, + max_norm=args.grad_clip, + ) + elif args.plugin == "zero2": + plugin = LowLevelZeroPlugin( + stage=2, + precision=args.mixed_precision, + initial_scale=2**16, + max_norm=args.grad_clip, + ) + elif args.plugin == "zero2_cpu": + plugin = LowLevelZeroPlugin( + stage=2, + precision=args.mixed_precision, + initial_scale=2**16, + cpu_offload=True, + max_norm=args.grad_clip, + ) + elif args.plugin == "3d": + plugin = HybridParallelPlugin( + tp_size=args.tp, + pp_size=1, + zero_stage=0, + parallel_output=False, + precision=args.mixed_precision, + ) + custom_plugin = HybridParallelPlugin( + tp_size=args.tp, + pp_size=1, + zero_stage=0, + parallel_output=False, + precision=args.mixed_precision, + custom_policy=booster_policy, + ) + else: + raise ValueError(f"Unknown plugin {args.plugin}") + + if args.plugin != "3d": + custom_plugin = plugin + + actor_booster = Booster(plugin=plugin) + ref_booster = Booster(plugin=plugin) + rm_booster = Booster(plugin=custom_plugin) + critic_booster = Booster(plugin=custom_plugin) + + default_dtype = torch.float16 if args.mixed_precision == "fp16" else torch.bfloat16 + torch.set_default_dtype(default_dtype) + actor, actor_optim, _, train_prompt_dataloader, actor_lr_scheduler = actor_booster.boost( + model=actor, + optimizer=actor_optim, + lr_scheduler=actor_lr_scheduler, + dataloader=train_prompt_dataloader, + ) + + critic, critic_optim, _, _, critic_lr_scheduler = critic_booster.boost( + model=critic, + optimizer=critic_optim, + lr_scheduler=critic_lr_scheduler, + dataloader=train_prompt_dataloader, + ) + reward_model, _, _, _, _ = rm_booster.boost(model=reward_model, dataloader=train_prompt_dataloader) + ref_model, _, _, _, _ = ref_booster.boost(model=ref_model, dataloader=train_prompt_dataloader) + + torch.set_default_dtype(torch.float) + + coordinator.print_on_master(f"Booster init max CUDA memory: {torch.cuda.max_memory_allocated() / 1024 ** 2:.2f} MB") + coordinator.print_on_master( + f"Booster init max CPU memory: {resource.getrusage(resource.RUSAGE_SELF).ru_maxrss / 1024:.2f} MB" + ) + + sampler_start_idx = 0 + start_step = 0 + + if args.rm_checkpoint_path is not None: + if "modeling" in args.rm_checkpoint_path: + rm_booster.load_model(reward_model, args.rm_checkpoint_path) + else: + _, _, _ = load_checkpoint( + load_dir=args.rm_checkpoint_path, + booster=rm_booster, + model=reward_model, + optimizer=None, + lr_scheduler=None, + ) + coordinator.print_on_master(f"Loaded reward model checkpoint {args.rm_checkpoint_path}") + + if args.checkpoint_path is not None: + if "modeling" in args.checkpoint_path: + actor_booster.load_model(actor, args.checkpoint_path) + ref_booster.load_model(ref_model, args.checkpoint_path) + coordinator.print_on_master(f"Loaded actor and reference model {args.checkpoint_path}") + else: + _, start_step, sampler_start_idx = load_checkpoint( + load_dir=args.checkpoint_path, + booster=actor_booster, + model=actor, + optimizer=actor_optim, + lr_scheduler=actor_lr_scheduler, + ) + _, _, _ = load_checkpoint( + load_dir=args.checkpoint_path, + booster=ref_booster, + model=ref_model, + optimizer=critic_optim, + lr_scheduler=critic_lr_scheduler, + ) + assert isinstance(train_prompt_dataloader.sampler, StatefulDistributedSampler) + train_prompt_dataloader.sampler.set_start_index(start_index=sampler_start_idx) + + coordinator.print_on_master( + f"Loaded actor and reference model checkpoint {args.checkpoint_path} at spisode {start_step}" + ) + coordinator.print_on_master(f"Loaded sample at index {sampler_start_idx}") + + coordinator.print_on_master( + f"Checkpoint loaded max CUDA memory: {torch.cuda.max_memory_allocated() / 1024 ** 2:.2f} MB" + ) + coordinator.print_on_master( + f"Checkpoint loaded CUDA memory: {torch.cuda.memory_allocated() / 1024 ** 2:.2f} MB" + ) + coordinator.print_on_master( + f"Checkpoint loaded max CPU memory: {resource.getrusage(resource.RUSAGE_SELF).ru_maxrss / 1024:.2f} MB" + ) + + if args.critic_checkpoint_path is not None: + if "modeling" in args.critic_checkpoint_path: + critic_booster.load_model(critic, args.critic_checkpoint_path) + else: + _, _, _ = load_checkpoint( + load_dir=args.critic_checkpoint_path, + booster=critic_booster, + model=critic, + optimizer=critic_optim, + lr_scheduler=critic_lr_scheduler, + ) + coordinator.print_on_master(f"Loaded critic checkpoint {args.critic_checkpoint_path}") + coordinator.print_on_master( + f"Checkpoint loaded max CUDA memory: {torch.cuda.max_memory_allocated() / 1024 ** 2:.2f} MB" + ) + coordinator.print_on_master( + f"Checkpoint loaded CUDA memory: {torch.cuda.memory_allocated() / 1024 ** 2:.2f} MB" + ) + coordinator.print_on_master( + f"Checkpoint loaded max CPU memory: {resource.getrusage(resource.RUSAGE_SELF).ru_maxrss / 1024:.2f} MB" + ) + + # configure trainer + trainer = PPOTrainer( + actor_booster, + critic_booster, + actor, + critic, + reward_model, + ref_model, + actor_optim, + critic_optim, + actor_lr_scheduler, + critic_lr_scheduler, + tokenizer=tokenizer, + stop_token_ids=stop_ids, + kl_coef=args.kl_coef, + ptx_coef=args.ptx_coef, + train_batch_size=args.train_batch_size, + buffer_limit=args.num_collect_steps * args.experience_batch_size, + max_length=args.max_length, + max_new_tokens=args.max_seq_len, + use_cache=True, + do_sample=True, + temperature=0.7, + accumulation_steps=args.accumulation_steps, + save_dir=args.save_path, + save_interval=args.save_interval, + top_k=50, + use_tp=args.tp > 1, + offload_inference_models="gemini" not in args.plugin, + coordinator=coordinator, + ) + + trainer.fit( + num_episodes=args.num_episodes, + num_collect_steps=args.num_collect_steps, + num_update_steps=args.num_update_steps, + prompt_dataloader=train_prompt_dataloader, + pretrain_dataloader=train_pretrain_dataloader, + log_dir=args.log_dir, + use_wandb=args.use_wandb, + ) + + if args.lora_rank > 0 and args.merge_lora_weights: + from coati.models.lora import LORA_MANAGER + + # NOTE: set model to eval to merge LoRA weights + LORA_MANAGER.merge_weights = True + actor.eval() + critic.eval() + # save model checkpoint after fitting on only rank0 + coordinator.print_on_master("Start saving final actor model checkpoint") + actor_booster.save_model(actor, os.path.join(trainer.actor_save_dir, "modeling"), shard=True) + coordinator.print_on_master( + f"Saved final actor model checkpoint at episodes {args.num_episodes} at folder {args.save_path}" + ) + coordinator.print_on_master("Start saving final critic model checkpoint") + critic_booster.save_model(critic, os.path.join(trainer.critic_save_dir, "modeling"), shard=True) + coordinator.print_on_master( + f"Saved final critic model checkpoint at episodes {args.num_episodes} at folder {args.save_path}" + ) + coordinator.print_on_master(f"Max CUDA memory usage: {torch.cuda.max_memory_allocated()/1024**2:.2f} MB") + + +if __name__ == "__main__": + parser = argparse.ArgumentParser() + parser.add_argument("--prompt_dataset", nargs="+", default=[]) + parser.add_argument("--ptx_dataset", nargs="+", default=[]) + parser.add_argument( + "--plugin", + type=str, + default="gemini", + choices=["gemini", "gemini_auto", "zero2", "zero2_cpu", "3d"], + help="Choose which plugin to use", + ) + parser.add_argument( + "--conversation_template_config", + type=str, + default=None, + help="Path \ + to save conversation template config files.", + ) + parser.add_argument("--grad_clip", type=float, default=1.0, help="Gradient clipping value") + parser.add_argument("--weight_decay", type=float, default=0.1, help="Weight decay") + parser.add_argument("--warmup_steps", type=int, default=None, help="Warmup steps") + parser.add_argument("--tokenizer_dir", type=str, default=None) + parser.add_argument("--tp", type=int, default=1) + parser.add_argument("--pretrain", type=str, default=None) + parser.add_argument("--rm_pretrain", type=str, default=None) + parser.add_argument("--checkpoint_path", type=str, default=None) + parser.add_argument("--critic_checkpoint_path", type=str, default=None) + parser.add_argument("--rm_checkpoint_path", type=str, help="Reward model checkpoint path") + parser.add_argument("--save_path", type=str, default="actor_checkpoint_prompts") + parser.add_argument("--num_episodes", type=int, default=1) + parser.add_argument("--num_collect_steps", type=int, default=2) + parser.add_argument("--num_update_steps", type=int, default=5) + parser.add_argument("--save_interval", type=int, default=1000) + parser.add_argument("--train_batch_size", type=int, default=16) + parser.add_argument("--experience_batch_size", type=int, default=16) + parser.add_argument("--ptx_batch_size", type=int, default=4) + parser.add_argument("--lora_train_bias", type=str, default="none") + parser.add_argument("--mixed_precision", type=str, default="fp16", choices=["fp16", "bf16"], help="Mixed precision") + parser.add_argument("--accumulation_steps", type=int, default=8) + parser.add_argument("--lora_rank", type=int, default=0, help="low-rank adaptation matrices rank") + parser.add_argument("--merge_lora_weights", type=bool, default=True) + parser.add_argument("--lr", type=float, default=9e-6) + parser.add_argument("--critic_lr", type=float, default=9e-6) + parser.add_argument("--kl_coef", type=float, default=0.1) + parser.add_argument("--ptx_coef", type=float, default=0.0) + parser.add_argument("--max_length", type=int, default=2048) + parser.add_argument("--max_seq_len", type=int, default=256) + parser.add_argument("--log_dir", default="logs", type=str) + parser.add_argument("--use_wandb", default=False, action="store_true") + parser.add_argument("--grad_checkpoint", default=False, action="store_true") + parser.add_argument("--use_flash_attn", default=False, action="store_true") + args = parser.parse_args() + train(args) diff --git a/applications/ColossalChat/examples/training_scripts/train_ppo.sh b/applications/ColossalChat/examples/training_scripts/train_ppo.sh new file mode 100755 index 000000000000..91633978e6ff --- /dev/null +++ b/applications/ColossalChat/examples/training_scripts/train_ppo.sh @@ -0,0 +1,82 @@ +#!/bin/bash +set_n_least_used_CUDA_VISIBLE_DEVICES() { + local n=${1:-"9999"} + echo "GPU Memory Usage:" + local FIRST_N_GPU_IDS=$(nvidia-smi --query-gpu=memory.used --format=csv | + tail -n +2 | + nl -v 0 | + tee /dev/tty | + sort -g -k 2 | + awk '{print $1}' | + head -n $n) + export CUDA_VISIBLE_DEVICES=$(echo $FIRST_N_GPU_IDS | sed 's/ /,/g') + echo "Now CUDA_VISIBLE_DEVICES is set to:" + echo "CUDA_VISIBLE_DEVICES=$CUDA_VISIBLE_DEVICES" +} +set_n_least_used_CUDA_VISIBLE_DEVICES 8 + +PROJECT_NAME="ppo" + +PARENT_SAVE_DIR="" # Path to a folder to save checkpoints +PARENT_TENSORBOARD_DIR="" # Path to a folder to save logs +PARENT_CONFIG_FILE="" # Path to a folder to save training config logs +PRETRAINED_MODEL_PATH="" # local pretrained model path (from RLHF step 1: SFT) +PRETRAINED_TOKENIZER_PATH="" # huggingface or local tokenizer path +REWARD_MODEL_PATH="" # local reward model path (from RLHF step 2: Train Reward Model) +CONVERSATION_TEMPLATE_CONFIG_PATH="" # path to the conversation config file + +declare -a prompt_dataset=( + YOUR/PROMPT/DATA/DIR/arrow/part-00000 + YOUR/PROMPT/DATA/DIR/arrow/part-00001 + YOUR/PROMPT/DATA/DIR/arrow/part-00002 + YOUR/PROMPT/DATA/DIR/arrow/part-00003 + YOUR/PROMPT/DATA/DIR/arrow/part-00004 + YOUR/PROMPT/DATA/DIR/arrow/part-00005 + YOUR/PROMPT/DATA/DIR/arrow/part-00006 + YOUR/PROMPT/DATA/DIR/arrow/part-00007 + YOUR/PROMPT/DATA/DIR/arrow/part-00008 + YOUR/PROMPT/DATA/DIR/arrow/part-00009 +) + +declare -a ptx_dataset=( + YOUR/SFT/DATA/DIR/arrow/part-00000 + YOUR/SFT/DATA/DIR/arrow/part-00001 + YOUR/SFT/DATA/DIR/arrow/part-00002 + YOUR/SFT/DATA/DIR/arrow/part-00003 + YOUR/SFT/DATA/DIR/arrow/part-00004 + YOUR/SFT/DATA/DIR/arrow/part-00005 + YOUR/SFT/DATA/DIR/arrow/part-00006 + YOUR/SFT/DATA/DIR/arrow/part-00007 + YOUR/SFT/DATA/DIR/arrow/part-00008 + YOUR/SFT/DATA/DIR/arrow/part-00009 +) + +TIMESTAMP=$(date +%Y-%m-%d-%H-%M-%S) +FULL_PROJECT_NAME="${PROJECT_NAME}-${TIMESTAMP}" +SAVE_DIR="${PARENT_SAVE_DIR}${FULL_PROJECT_NAME}" +CONFIG_FILE="${PARENT_CONFIG_FILE}-${FULL_PROJECT_NAME}.json" + +colossalai run --nproc_per_node 8 --hostfile hostfile --master_port 31312 train_ppo.py \ + --pretrain $PRETRAINED_MODEL_PATH \ + --rm_pretrain $PRETRAINED_MODEL_PATH \ + --tokenizer_dir $PRETRAINED_TOKENIZER_PATH \ + --rm_checkpoint_path $REWARD_MODEL_PATH \ + --prompt_dataset ${prompt_dataset[@]} \ + --conversation_template_config $CONVERSATION_TEMPLATE_CONFIG_PATH \ + --ptx_coef 0.0 \ + --plugin "zero2" \ + --save_interval 500 \ + --save_path $SAVE_DIR \ + --num_episodes 2000 \ + --num_collect_steps 2 \ + --num_update_steps 1 \ + --experience_batch_size 4 \ + --train_batch_size 4 \ + --accumulation_steps 2 \ + --lr 9e-6 \ + --mixed_precision "bf16" \ + --grad_clip 0.1\ + --weight_decay 0.01 \ + --warmup_steps 40 \ + --grad_checkpoint \ + --use_wandb diff --git a/applications/ColossalChat/examples/training_scripts/train_rm.py b/applications/ColossalChat/examples/training_scripts/train_rm.py new file mode 100755 index 000000000000..a0c710f2bb7f --- /dev/null +++ b/applications/ColossalChat/examples/training_scripts/train_rm.py @@ -0,0 +1,342 @@ +import argparse +import json +import math +import os +import resource +from contextlib import nullcontext + +import torch +from coati.dataset import ( + DataCollatorForPreferenceDataset, + StatefulDistributedSampler, + load_tokenized_dataset, + setup_distributed_dataloader, +) +from coati.models import LogExpLoss, LogSigLoss, RewardModel, convert_to_lora_module +from coati.trainer import RewardModelTrainer +from coati.utils import load_checkpoint +from transformers import AutoTokenizer + +import colossalai +from colossalai.booster import Booster +from colossalai.booster.plugin import GeminiPlugin, HybridParallelPlugin, LowLevelZeroPlugin, TorchDDPPlugin +from colossalai.cluster import DistCoordinator +from colossalai.nn.lr_scheduler import CosineAnnealingWarmupLR +from colossalai.nn.optimizer import HybridAdam + + +def train(args): + # check lora compatibility + if "gemini" in args.plugin and args.lora_rank > 0: + raise ValueError("LoRA is not supported in GeminiPlugin. Please use other plugin") + if args.plugin == "gemini_auto" and args.accumulation_steps > 1: + raise ValueError("Gradient accumulation is not supported in GeminiPlugin. Please use other plugin") + # ============================== + # Initialize Distributed Training + # ============================== + colossalai.launch_from_torch({}) + coordinator = DistCoordinator() + + # ====================================================== + # Initialize Model, Objective, Optimizer and LR Scheduler + # ====================================================== + # Temp Fix: Disable lazy init due to version conflict + # init_ctx = ( + # LazyInitContext(default_device=get_current_device()) if isinstance(plugin, (GeminiPlugin,)) else nullcontext() + # ) + + init_ctx = nullcontext() + booster_policy = None + with init_ctx: + if args.use_flash_attn: + model = RewardModel( + args.pretrain, + torch_dtype=torch.bfloat16 if args.mixed_precision == "bf16" else torch.float16, + use_flash_attention_2=True, + ) + coordinator.print_on_master(msg="Flash-attention enabled successfully") + else: + model = RewardModel(args.pretrain) + + if args.tp > 1: + if model.model.config.architectures[0] == "BloomForCausalLM": + from colossalai.shardformer.policies.bloom import BloomPolicy + + booster_policy = BloomPolicy() + elif model.model.config.architectures[0] == "LlamaForCausalLM": + from colossalai.shardformer.policies.llama import LlamaPolicy + + booster_policy = LlamaPolicy() + elif model.model.config.architectures[0] == "GPT2LMHeadModel": + from colossalai.shardformer.policies.gpt2 import GPT2Policy + + booster_policy = GPT2Policy() + elif model.model.config.architectures[0] == "ChatGLMModel": + from colossalai.shardformer.policies.chatglm2 import ChatGLMPolicy + + booster_policy = ChatGLMPolicy() + elif model.model.config.architectures[0] == "OPTForCausalLM": + from colossalai.shardformer.policies.opt import OPTPolicy + + booster_policy = OPTPolicy() + else: + raise ValueError("Unknown model architecture for policy") + + if args.lora_rank > 0: + model = convert_to_lora_module(model, args.lora_rank, lora_train_bias=args.lora_train_bias) + + # ============================== + # Initialize Booster + # ============================== + if args.plugin == "ddp": + """ + Default torch ddp plugin without any acceleration, for + debugging purpose acceleration, for debugging purpose + """ + plugin = TorchDDPPlugin(find_unused_parameters=True) + elif args.plugin == "gemini": + plugin = GeminiPlugin( + precision=args.mixed_precision, + placement_policy="static", + initial_scale=2**16, + max_norm=args.grad_clip, + enable_gradient_accumulation=True, + ) + elif args.plugin == "gemini_auto": + plugin = GeminiPlugin( + precision=args.mixed_precision, + placement_policy="auto", + initial_scale=2**16, + max_norm=args.grad_clip, + ) + elif args.plugin == "zero2": + plugin = LowLevelZeroPlugin( + stage=2, + precision=args.mixed_precision, + initial_scale=2**16, + max_norm=args.grad_clip, + ) + elif args.plugin == "zero2_cpu": + plugin = LowLevelZeroPlugin( + stage=2, + precision=args.mixed_precision, + initial_scale=2**16, + cpu_offload=True, + max_norm=args.grad_clip, + ) + elif args.plugin == "3d": + plugin = HybridParallelPlugin( + tp_size=args.tp, + pp_size=1, + zero_stage=0, + parallel_output=False, + precision=args.mixed_precision, + custom_policy=booster_policy, + ) + else: + raise ValueError(f"Unknown plugin {args.plugin}") + + booster = Booster(plugin=plugin) + + if args.grad_checkpoint and args.lora_rank == 0: + model.model.gradient_checkpointing_enable() # TODO: support gradient checkpoint for the last linear layer + coordinator.print_on_master(msg="Gradient checkpointing enabled successfully") + elif args.lora_rank > 0: + coordinator.print_on_master(msg="Gradient checkpointing will be disabled when LoRA is enabled") + + # configure tokenizer + tokenizer_dir = args.tokenizer_dir if args.tokenizer_dir is not None else args.pretrain + tokenizer = AutoTokenizer.from_pretrained(tokenizer_dir, use_fast=False, trust_remote_code=True) + if hasattr(tokenizer, "pad_token") and hasattr(tokenizer, "eos_token") and tokenizer.eos_token is not None: + try: + # Some tokenizers doesn't allow to set pad_token mannually e.g., Qwen + tokenizer.pad_token = tokenizer.eos_token + except AttributeError as e: + logger.warning(f"Unable to set pad token to eos token, {str(e)}") + if not hasattr(tokenizer, "pad_token") or tokenizer.pad_token is None: + logger.warning( + "The tokenizer does not have a pad token which is required. May lead to unintended behavior in training, Please consider manually set them." + ) + tokenizer.padding_side = "right" + tokenizer.add_bos_token = False + tokenizer.add_eos_token = False + + # configure loss function + if args.loss_fn == "log_sig": + loss_fn = LogSigLoss() + elif args.loss_fn == "log_exp": + loss_fn = LogExpLoss() + else: + raise ValueError(f'Unsupported loss function "{args.loss_fn}"') + + # configure optimizer + optim = HybridAdam( + model_params=model.parameters(), + lr=args.lr, + betas=(0.9, 0.95), + weight_decay=args.weight_decay, + adamw_mode=True, + ) + + # configure dataset + coordinator.print_on_master(f"Load dataset: {args.dataset}") + mode_map = {"train": "train", "valid": "validation", "test": "test"} + train_dataset = load_tokenized_dataset(dataset_paths=args.dataset, mode="train", mode_map=mode_map) + data_collator = DataCollatorForPreferenceDataset(tokenizer=tokenizer, max_length=args.max_length) + train_dataloader = setup_distributed_dataloader( + dataset=train_dataset, + batch_size=args.batch_size, + shuffle=True, + drop_last=True, + collate_fn=data_collator, + use_tp=args.tp > 1, + ) + + num_update_steps_per_epoch = len(train_dataloader) // args.accumulation_steps + math.ceil(args.max_epochs * num_update_steps_per_epoch) + + if args.warmup_steps is None: + args.warmup_steps = int(args.max_epochs * 0.025 * (len(train_dataloader) // args.accumulation_steps)) + coordinator.print_on_master(f"Warmup steps is set to {args.warmup_steps}") + + lr_scheduler = CosineAnnealingWarmupLR( + optimizer=optim, + total_steps=args.max_epochs * num_update_steps_per_epoch, + warmup_steps=args.warmup_steps, + eta_min=0.1 * args.lr, + ) + + default_dtype = torch.float16 if args.mixed_precision == "fp16" else torch.bfloat16 + torch.set_default_dtype(default_dtype) + model, optim, _, train_dataloader, lr_scheduler = booster.boost( + model=model, + optimizer=optim, + lr_scheduler=lr_scheduler, + dataloader=train_dataloader, + ) + torch.set_default_dtype(torch.float) + + coordinator.print_on_master(f"Booster init max CUDA memory: {torch.cuda.max_memory_allocated() / 1024 ** 2:.2f} MB") + coordinator.print_on_master( + f"Booster init max CPU memory: {resource.getrusage(resource.RUSAGE_SELF).ru_maxrss / 1024:.2f} MB" + ) + + start_epoch = 0 + sampler_start_idx = 0 + start_step = 0 + if args.checkpoint_path is not None: + if "modeling" in args.checkpoint_path: + coordinator.print_on_master(f"Continued pretrain from checkpoint {args.checkpoint_path}") + booster.load_model(model, args.checkpoint_path) + else: + coordinator.print_on_master(f"Load model checkpoint from {args.checkpoint_path}") + start_epoch, start_step, sampler_start_idx = load_checkpoint( + load_dir=args.checkpoint_path, + booster=booster, + model=model, + optimizer=optim, + lr_scheduler=lr_scheduler, + ) + assert isinstance(train_dataloader.sampler, StatefulDistributedSampler) + train_dataloader.sampler.set_start_index(start_index=sampler_start_idx) + + coordinator.print_on_master( + f"Loaded checkpoint {args.checkpoint_path} at epoch {start_epoch} step {start_step}" + ) + coordinator.print_on_master(f"Loaded sample at index {sampler_start_idx}") + + coordinator.print_on_master( + f"Checkpoint loaded max CUDA memory: {torch.cuda.max_memory_allocated() / 1024 ** 2:.2f} MB" + ) + coordinator.print_on_master( + f"Checkpoint loaded CUDA memory: {torch.cuda.memory_allocated() / 1024 ** 2:.2f} MB" + ) + coordinator.print_on_master( + f"Checkpoint loaded max CPU memory: {resource.getrusage(resource.RUSAGE_SELF).ru_maxrss / 1024:.2f} MB" + ) + + trainer = RewardModelTrainer( + model, + booster, + optim, + lr_scheduler, + tokenizer, + loss_fn=loss_fn, + max_epochs=args.max_epochs, + accumulation_steps=args.accumulation_steps, + start_epoch=start_epoch, + save_interval=args.save_interval, + save_dir=args.save_dir, + coordinator=coordinator, + ) + + trainer.fit( + train_preference_dataloader=train_dataloader, + eval_preference_dataloader=None, + log_dir=args.log_dir, + use_wandb=args.use_wandb, + ) + + if args.lora_rank > 0 and args.merge_lora_weights: + from coati.models.lora import LORA_MANAGER + + # NOTE: set model to eval to merge LoRA weights + LORA_MANAGER.merge_weights = True + model.eval() + # save model checkpoint after fitting on only rank0 + coordinator.print_on_master("Start saving final model checkpoint") + booster.save_model(model, os.path.join(args.save_dir, "modeling"), shard=True) + coordinator.print_on_master(f"Saved final model checkpoint at epoch {args.max_epochs} at folder {args.save_dir}") + + coordinator.print_on_master(f"Max CUDA memory usage: {torch.cuda.max_memory_allocated()/1024**2:.2f} MB") + + +if __name__ == "__main__": + # ============================== + # Parse Arguments + # ============================== + parser = argparse.ArgumentParser() + parser.add_argument( + "--plugin", + type=str, + default="gemini", + choices=["gemini", "gemini_auto", "zero2", "zero2_cpu", "3d", "ddp"], + help="Choose which plugin to use", + ) + parser.add_argument("--grad_clip", type=float, default=1.0, help="Gradient clipping value") + parser.add_argument("--weight_decay", type=float, default=0.1, help="Weight decay") + parser.add_argument("--warmup_steps", type=int, default=None, help="Warmup steps") + parser.add_argument("--tp", type=int, default=1) + parser.add_argument("--pretrain", type=str, default=None) + parser.add_argument("--tokenizer_dir", type=str, default=None) + parser.add_argument("--dataset", nargs="+", default=[]) + parser.add_argument( + "--checkpoint_path", type=str, default=None, help="Checkpoint path if need to resume training form a checkpoint" + ) + parser.add_argument("--config_file", type=str, default="config_file", help="Config file") + parser.add_argument("--save_dir", type=str, default="output") + parser.add_argument("--max_length", type=int, default=2048, help="Model max length") + parser.add_argument("--max_epochs", type=int, default=3) + parser.add_argument("--batch_size", type=int, default=4) + parser.add_argument("--mixed_precision", type=str, default="fp16", choices=["fp16", "bf16"], help="Mixed precision") + parser.add_argument("--loss_fn", type=str, default="log_sig", choices=["log_sig", "log_exp"], help="Loss function") + parser.add_argument("--lora_rank", type=int, default=0, help="low-rank adaptation matrices rank") + parser.add_argument( + "--lora_train_bias", + type=str, + default="none", + help="'none' means it doesn't train biases. 'all' means it trains all biases. 'lora_only' means it only trains biases of LoRA layers", + ) + parser.add_argument("--save_interval", type=int, default=1000, help="number of step between two checkpoints") + parser.add_argument("--merge_lora_weights", type=bool, default=True) + parser.add_argument("--lr", type=float, default=5e-6) + parser.add_argument("--accumulation_steps", type=int, default=8) + parser.add_argument("--log_dir", default="logs", type=str) + parser.add_argument("--use_wandb", default=False, action="store_true") + parser.add_argument("--grad_checkpoint", default=False, action="store_true") + parser.add_argument("--use_flash_attn", default=False, action="store_true") + args = parser.parse_args() + os.makedirs(os.path.dirname(args.config_file), exist_ok=True) + with open(args.config_file, "w") as f: + json.dump(args.__dict__, f, indent=4) + train(args) diff --git a/applications/ColossalChat/examples/training_scripts/train_rm.sh b/applications/ColossalChat/examples/training_scripts/train_rm.sh new file mode 100755 index 000000000000..e06d9092fe4c --- /dev/null +++ b/applications/ColossalChat/examples/training_scripts/train_rm.sh @@ -0,0 +1,61 @@ +#!/bin/bash +set_n_least_used_CUDA_VISIBLE_DEVICES() { + local n=${1:-"9999"} + echo "GPU Memory Usage:" + local FIRST_N_GPU_IDS=$(nvidia-smi --query-gpu=memory.used --format=csv | + tail -n +2 | + nl -v 0 | + tee /dev/tty | + sort -g -k 2 | + awk '{print $1}' | + head -n $n) + export CUDA_VISIBLE_DEVICES=$(echo $FIRST_N_GPU_IDS | sed 's/ /,/g') + echo "Now CUDA_VISIBLE_DEVICES is set to:" + echo "CUDA_VISIBLE_DEVICES=$CUDA_VISIBLE_DEVICES" +} +set_n_least_used_CUDA_VISIBLE_DEVICES 8 + +PROJECT_NAME="rm" +PARENT_SAVE_DIR="" # Path to a folder to save checkpoints +PARENT_TENSORBOARD_DIR="" # Path to a folder to save logs +PARENT_CONFIG_FILE="" # Path to a folder to save training config logs +PRETRAINED_MODEL_PATH="" # huggingface or local model path +PRETRAINED_TOKENIZER_PATH="" # huggingface or local tokenizer path + +declare -a dataset=( + YOUR/PREFERENCE/DATA/DIR/arrow/part-00000 + YOUR/PREFERENCE/DATA/DIR/arrow/part-00001 + YOUR/PREFERENCE/DATA/DIR/arrow/part-00002 + YOUR/PREFERENCE/DATA/DIR/arrow/part-00003 + YOUR/PREFERENCE/DATA/DIR/arrow/part-00004 + YOUR/PREFERENCE/DATA/DIR/arrow/part-00005 + YOUR/PREFERENCE/DATA/DIR/arrow/part-00006 + YOUR/PREFERENCE/DATA/DIR/arrow/part-00007 + YOUR/PREFERENCE/DATA/DIR/arrow/part-00008 + YOUR/PREFERENCE/DATA/DIR/arrow/part-00009 +) + +TIMESTAMP=$(date +%Y-%m-%d-%H-%M-%S) +FULL_PROJECT_NAME="${PROJECT_NAME}-${TIMESTAMP}" +SAVE_DIR="${PARENT_SAVE_DIR}${FULL_PROJECT_NAME}" +CONFIG_FILE="${PARENT_CONFIG_FILE}-${FULL_PROJECT_NAME}.json" + +colossalai run --nproc_per_node 8 --hostfile hostfile --master_port 31312 train_rm.py \ + --pretrain $PRETRAINED_MODEL_PATH \ + --checkpoint_path /home/yeanbang/data/experiments/rm/hhh_aligh/ckptllama2-rm-2024-01-17-14-43-24/epoch-1_step-1317/modeling \ + --tokenizer_dir $PRETRAINED_TOKENIZER_PATH \ + --dataset ${dataset[@]} \ + --plugin "zero2" \ + --save_interval 1000 \ + --save_dir $SAVE_DIR \ + --config_file $CONFIG_FILE \ + --max_epochs 3 \ + --accumulation_steps 1 \ + --batch_size 8 \ + --lr 5e-6 \ + --mixed_precision "bf16" \ + --grad_clip 1.0 \ + --weight_decay 0.01 \ + --warmup_steps 40 \ + --grad_checkpoint \ + --use_wandb diff --git a/applications/ColossalChat/examples/training_scripts/train_sft.py b/applications/ColossalChat/examples/training_scripts/train_sft.py new file mode 100755 index 000000000000..fcd1a429cc5f --- /dev/null +++ b/applications/ColossalChat/examples/training_scripts/train_sft.py @@ -0,0 +1,311 @@ +import argparse +import json +import math +import os +import resource +from contextlib import nullcontext + +import torch +from coati.dataset import DataCollatorForSupervisedDataset, load_tokenized_dataset, setup_distributed_dataloader +from coati.models import convert_to_lora_module +from coati.trainer import SFTTrainer +from coati.utils import load_checkpoint +from transformers import AutoModelForCausalLM, AutoTokenizer + +import colossalai +from colossalai.booster import Booster +from colossalai.booster.plugin import GeminiPlugin, HybridParallelPlugin, LowLevelZeroPlugin, TorchDDPPlugin +from colossalai.cluster import DistCoordinator +from colossalai.nn.lr_scheduler import CosineAnnealingWarmupLR +from colossalai.nn.optimizer import HybridAdam + + +def train(args): + # check lora compatibility + if "gemini" in args.plugin and args.lora_rank > 0: + raise ValueError("LoRA is not supported in GeminiPlugin. Please use other plugin") + if args.plugin == "gemini_auto" and args.accumulation_steps > 1: + raise ValueError("Gradient accumulation is not supported in GeminiPlugin. Please use other plugin") + # ============================== + # Initialize Distributed Training + # ============================== + colossalai.launch_from_torch({}) + coordinator = DistCoordinator() + + # ============================== + # Initialize Booster + # ============================== + if args.plugin == "ddp": + """ + Default torch ddp plugin without any acceleration, for + debugging purpose acceleration, for debugging purpose + """ + plugin = TorchDDPPlugin(find_unused_parameters=True) + elif args.plugin == "gemini": + plugin = GeminiPlugin( + precision=args.mixed_precision, + placement_policy="static", + initial_scale=2**16, + max_norm=args.grad_clip, + enable_gradient_accumulation=True, + ) + elif args.plugin == "gemini_auto": + plugin = GeminiPlugin( + precision=args.mixed_precision, + placement_policy="auto", + initial_scale=2**16, + max_norm=args.grad_clip, + ) + elif args.plugin == "zero2": + plugin = LowLevelZeroPlugin( + stage=2, + precision=args.mixed_precision, + initial_scale=2**16, + max_norm=args.grad_clip, + ) + elif args.plugin == "zero2_cpu": + plugin = LowLevelZeroPlugin( + stage=2, + precision=args.mixed_precision, + initial_scale=2**16, + cpu_offload=True, + max_norm=args.grad_clip, + ) + elif args.plugin == "3d": + plugin = HybridParallelPlugin( + tp_size=args.tp, + pp_size=1, + zero_stage=0, + parallel_output=False, + max_norm=args.grad_clip, + precision=args.mixed_precision, + ) + else: + raise ValueError(f"Unknown plugin {args.plugin}") + + booster = Booster(plugin=plugin) + + # ====================================================== + # Initialize Model, Objective, Optimizer and LR Scheduler + # ====================================================== + # Temp Fix: Disable lazy init due to version conflict + # init_ctx = ( + # LazyInitContext(default_device=get_current_device()) if isinstance(plugin, (GeminiPlugin,)) else nullcontext() + # ) + + init_ctx = nullcontext() + with init_ctx: + if args.use_flash_attn: + model = AutoModelForCausalLM.from_pretrained( + args.pretrain, + torch_dtype=torch.bfloat16 if args.mixed_precision == "bf16" else torch.float16, + use_flash_attention_2=True, + ) + coordinator.print_on_master(msg="Flash-attention enabled successfully") + else: + model = AutoModelForCausalLM.from_pretrained(args.pretrain) + if args.lora_rank > 0: + model = convert_to_lora_module(model, args.lora_rank, lora_train_bias=args.lora_train_bias) + + if args.grad_checkpoint and args.lora_rank == 0: + # lora layers are not supported by gradient checkpointing + model.gradient_checkpointing_enable() + coordinator.print_on_master(msg="Gradient checkpointing enabled successfully") + elif args.lora_rank > 0: + coordinator.print_on_master(msg="Gradient checkpointing will be disabled when LoRA is enabled") + + # configure tokenizer + tokenizer = AutoTokenizer.from_pretrained( + args.tokenizer_dir or args.pretrain, use_fast=False, trust_remote_code=True + ) + if hasattr(tokenizer, "pad_token") and hasattr(tokenizer, "eos_token") and tokenizer.eos_token is not None: + try: + # Some tokenizers doesn't allow to set pad_token mannually e.g., Qwen + tokenizer.pad_token = tokenizer.eos_token + except AttributeError as e: + logger.warning(f"Unable to set pad token to eos token, {str(e)}") + if not hasattr(tokenizer, "pad_token") or tokenizer.pad_token is None: + logger.warning( + "The tokenizer does not have a pad token which is required. May lead to unintended behavior in training, Please consider manually set them." + ) + + tokenizer.add_bos_token = False + tokenizer.add_eos_token = False + + coordinator.print_on_master(f"Configuration file will be saved at: {args.config_file}") + coordinator.print_on_master(f"Model checkpoint will be saved at: {args.save_path}") + + # configure optimizer + optim = HybridAdam( + model_params=model.parameters(), + lr=args.lr, + betas=(0.9, 0.95), + weight_decay=args.weight_decay, + adamw_mode=True, + ) + + # configure dataset + coordinator.print_on_master( + f"Max CUDA memory before data loader: {torch.cuda.max_memory_allocated() / 1024 ** 2:.2f} MB" + ) + dataset = load_tokenized_dataset(dataset_paths=args.dataset, mode="train") + data_collator = DataCollatorForSupervisedDataset(tokenizer=tokenizer, max_length=args.max_len) + train_dataloader = setup_distributed_dataloader( + dataset=dataset, + batch_size=args.batch_size, + shuffle=True, + drop_last=True, + collate_fn=data_collator, + use_tp=args.tp > 1, + ) + coordinator.print_on_master( + f"Max CUDA memory after data loader: {torch.cuda.max_memory_allocated() / 1024 ** 2:.2f} MB" + ) + + num_update_steps_per_epoch = len(train_dataloader) // args.accumulation_steps + math.ceil(args.max_epochs * num_update_steps_per_epoch) + + if args.warmup_steps is None: + args.warmup_steps = int(args.max_epochs * 0.025 * (len(train_dataloader) // args.accumulation_steps)) + coordinator.print_on_master(f"Warmup steps is set to {args.warmup_steps}") + + lr_scheduler = CosineAnnealingWarmupLR( + optimizer=optim, + total_steps=args.max_epochs * num_update_steps_per_epoch, + warmup_steps=args.warmup_steps, + eta_min=0.1 * args.lr, + ) + + # Flash attention will be disabled because it does NOT support fp32. + default_dtype = torch.float16 if args.mixed_precision == "fp16" else torch.bfloat16 + torch.set_default_dtype(default_dtype) + model, optim, _, train_dataloader, lr_scheduler = booster.boost( + model=model, + optimizer=optim, + lr_scheduler=lr_scheduler, + dataloader=train_dataloader, + ) + # model = model.to(get_current_device()) + torch.set_default_dtype(torch.float) + + coordinator.print_on_master(f"Booster init max CUDA memory: {torch.cuda.max_memory_allocated() / 1024 ** 2:.2f} MB") + coordinator.print_on_master( + f"Booster init max CPU memory: {resource.getrusage(resource.RUSAGE_SELF).ru_maxrss / 1024:.2f} MB" + ) + + start_epoch = 0 + sampler_start_idx = 0 + start_step = 0 + if args.checkpoint_path is not None: + if "modeling" in args.checkpoint_path: + coordinator.print_on_master(f"Continued pretrain from checkpoint {args.checkpoint_path}") + booster.load_model(model, args.checkpoint_path) + else: + coordinator.print_on_master(f"Load model checkpoint from {args.checkpoint_path}") + start_epoch, start_step, sampler_start_idx = load_checkpoint( + load_dir=args.checkpoint_path, + booster=booster, + model=model, + optimizer=optim, + lr_scheduler=lr_scheduler, + ) + train_dataloader.sampler.set_start_index(start_index=sampler_start_idx) + + coordinator.print_on_master( + f"Loaded checkpoint {args.checkpoint_path} at epoch {start_epoch} step {start_step}" + ) + coordinator.print_on_master(f"Loaded sample at index {sampler_start_idx}") + + coordinator.print_on_master( + f"Checkpoint loaded max CUDA memory: {torch.cuda.max_memory_allocated() / 1024 ** 2:.2f} MB" + ) + coordinator.print_on_master( + f"Checkpoint loaded CUDA memory: {torch.cuda.memory_allocated() / 1024 ** 2:.2f} MB" + ) + coordinator.print_on_master( + f"Checkpoint loaded max CPU memory: {resource.getrusage(resource.RUSAGE_SELF).ru_maxrss / 1024:.2f} MB" + ) + + trainer = SFTTrainer( + model=model, + booster=booster, + optim=optim, + lr_scheduler=lr_scheduler, + max_epochs=args.max_epochs, + accumulation_steps=args.accumulation_steps, + start_epoch=start_epoch, + save_interval=args.save_interval, + save_dir=args.save_path, + coordinator=coordinator, + ) + + trainer.fit( + train_dataloader=train_dataloader, + eval_dataloader=None, + log_dir=args.log_dir, + use_wandb=args.use_wandb, + ) + + if args.lora_rank > 0 and args.merge_lora_weights: + from coati.models.lora import LORA_MANAGER + + # NOTE: set model to eval to merge LoRA weights + LORA_MANAGER.merge_weights = True + model.eval() + # save model checkpoint after fitting on only rank0 + coordinator.print_on_master("Start saving final model checkpoint") + + booster.save_model(model, os.path.join(args.save_path, "modeling"), shard=True) + coordinator.print_on_master(f"Saved final model checkpoint at epoch {args.max_epochs} at folder {args.save_path}") + + coordinator.print_on_master(f"Max CUDA memory usage: {torch.cuda.max_memory_allocated()/1024**2:.2f} MB") + + +if __name__ == "__main__": + # ============================== + # Parse Arguments + # ============================== + parser = argparse.ArgumentParser() + parser.add_argument( + "--plugin", + type=str, + default="gemini", + choices=["gemini", "gemini_auto", "zero2", "zero2_cpu", "3d", "ddp"], + help="Choose which plugin to use", + ) + parser.add_argument("--grad_clip", type=float, default=1.0, help="Gradient clipping value") + parser.add_argument("--weight_decay", type=float, default=0.1, help="Weight decay") + parser.add_argument("--warmup_steps", type=int, default=None, help="Warmup steps") + parser.add_argument("--tp", type=int, default=1) + parser.add_argument("--pretrain", type=str, default=None) + parser.add_argument("--tokenizer_dir", type=str, default=None) + parser.add_argument("--dataset", nargs="+", default=[]) + parser.add_argument( + "--checkpoint_path", type=str, default=None, help="Checkpoint path if need to resume training form a checkpoint" + ) + parser.add_argument("--save_path", type=str, default="output") + parser.add_argument("--max_epochs", type=int, default=3) + parser.add_argument("--batch_size", type=int, default=4) + parser.add_argument("--max_len", type=int, default=512) + parser.add_argument("--mixed_precision", type=str, default="fp16", choices=["fp16", "bf16"], help="Mixed precision") + parser.add_argument("--lora_rank", type=int, default=0, help="low-rank adaptation matrices rank") + parser.add_argument( + "--lora_train_bias", + type=str, + default="none", + help="'none' means it doesn't train biases. 'all' means it trains all biases. 'lora_only' means it only trains biases of LoRA layers", + ) + parser.add_argument("--save_interval", type=int, default=1000, help="number of step between two checkpoints") + parser.add_argument("--merge_lora_weights", type=bool, default=True) + parser.add_argument("--lr", type=float, default=5e-6) + parser.add_argument("--config_file", type=str, default="config_file", help="Config file") + parser.add_argument("--accumulation_steps", type=int, default=8) + parser.add_argument("--log_dir", default="logs", type=str) + parser.add_argument("--use_wandb", default=False, action="store_true") + parser.add_argument("--grad_checkpoint", default=False, action="store_true") + parser.add_argument("--use_flash_attn", default=False, action="store_true") + args = parser.parse_args() + os.makedirs(os.path.dirname(args.config_file), exist_ok=True) + with open(args.config_file, "w") as f: + json.dump(args.__dict__, f, indent=4) + train(args) diff --git a/applications/ColossalChat/examples/training_scripts/train_sft.sh b/applications/ColossalChat/examples/training_scripts/train_sft.sh new file mode 100755 index 000000000000..d5c394377616 --- /dev/null +++ b/applications/ColossalChat/examples/training_scripts/train_sft.sh @@ -0,0 +1,59 @@ +set_n_least_used_CUDA_VISIBLE_DEVICES() { + local n=${1:-"9999"} + echo "GPU Memory Usage:" + local FIRST_N_GPU_IDS=$(nvidia-smi --query-gpu=memory.used --format=csv | + tail -n +2 | + nl -v 0 | + tee /dev/tty | + sort -g -k 2 | + awk '{print $1}' | + head -n $n) + export CUDA_VISIBLE_DEVICES=$(echo $FIRST_N_GPU_IDS | sed 's/ /,/g') + echo "Now CUDA_VISIBLE_DEVICES is set to:" + echo "CUDA_VISIBLE_DEVICES=$CUDA_VISIBLE_DEVICES" +} + + +# export CUDA_VISIBLE_DEVICES=4,5,6 +set_n_least_used_CUDA_VISIBLE_DEVICES 4 +PROJECT_NAME="sft" +PARENT_SAVE_DIR="" # Path to a folder to save checkpoints +PARENT_TENSORBOARD_DIR="" # Path to a folder to save logs +PARENT_CONFIG_FILE="" # Path to a folder to save training config logs +PRETRAINED_MODEL_PATH="" # huggingface or local model path +PRETRAINED_TOKENIZER_PATH="" # huggingface or local tokenizer path +declare -a dataset=( + YOUR/SFT/DATA/DIR/arrow/part-00000 + YOUR/SFT/DATA/DIR/arrow/part-00001 + YOUR/SFT/DATA/DIR/arrow/part-00002 + YOUR/SFT/DATA/DIR/arrow/part-00003 + YOUR/SFT/DATA/DIR/arrow/part-00004 + YOUR/SFT/DATA/DIR/arrow/part-00005 + YOUR/SFT/DATA/DIR/arrow/part-00006 + YOUR/SFT/DATA/DIR/arrow/part-00007 + YOUR/SFT/DATA/DIR/arrow/part-00008 + YOUR/SFT/DATA/DIR/arrow/part-00009 +) + +TIMESTAMP=$(date +%Y-%m-%d-%H-%M-%S) +FULL_PROJECT_NAME="${PROJECT_NAME}-${TIMESTAMP}" +SAVE_DIR="${PARENT_SAVE_DIR}${FULL_PROJECT_NAME}" +CONFIG_FILE="${PARENT_CONFIG_FILE}-${FULL_PROJECT_NAME}.json" + +# the real batch size for gradient descent is number_of_node_in_hostfile * nproc_per_node * train_batch_size +colossalai run --nproc_per_node 4 --master_port 31312 --hostfile ./hostfile train_sft.py \ + --pretrain $PRETRAINED_MODEL_PATH \ + --tokenizer_dir $PRETRAINED_TOKENIZER_PATH \ + --save_interval 4000 \ + --dataset ${dataset[@]} \ + --save_path $SAVE_DIR \ + --config_file $CONFIG_FILE \ + --lora_rank 0 \ + --plugin zero2 \ + --batch_size 8 \ + --max_epochs 1 \ + --accumulation_steps 1 \ + --lr 2e-5 \ + --max_len 2048 \ + --grad_checkpoint \ + --use_wandb diff --git a/applications/Chat/pytest.ini b/applications/ColossalChat/pytest.ini old mode 100644 new mode 100755 similarity index 100% rename from applications/Chat/pytest.ini rename to applications/ColossalChat/pytest.ini diff --git a/applications/ColossalChat/requirements.txt b/applications/ColossalChat/requirements.txt new file mode 100755 index 000000000000..de5f6160e827 --- /dev/null +++ b/applications/ColossalChat/requirements.txt @@ -0,0 +1,24 @@ +transformers==4.34.1 +huggingface_hub==0.17.3 +tqdm +datasets +loralib +colossalai>=0.3.6 +torch>=1.12.1 +langchain +tokenizers +fastapi +sse_starlette +wandb +sentencepiece +gpustat +packaging +autoflake==2.2.1 +black==23.9.1 +tensorboard +six==1.16.0 +datasets +ninja==1.11.1 +sentencepiece==0.1.99 +flash-attn +tiktoken diff --git a/applications/Chat/setup.py b/applications/ColossalChat/setup.py old mode 100644 new mode 100755 similarity index 97% rename from applications/Chat/setup.py rename to applications/ColossalChat/setup.py index eb44b6203ef8..37503920ade6 --- a/applications/Chat/setup.py +++ b/applications/ColossalChat/setup.py @@ -32,7 +32,7 @@ def fetch_version(): license="Apache Software License 2.0", url="https://github.com/hpcaitech/Coati", install_requires=fetch_requirements("requirements.txt"), - python_requires=">=3.6", + python_requires=">=3.7", classifiers=[ "Programming Language :: Python :: 3", "License :: OSI Approved :: Apache Software License", diff --git a/applications/Chat/tests/__init__.py b/applications/ColossalChat/tests/__init__.py old mode 100644 new mode 100755 similarity index 100% rename from applications/Chat/tests/__init__.py rename to applications/ColossalChat/tests/__init__.py diff --git a/applications/ColossalChat/tests/generate_dummy_datasets_for_testing.py b/applications/ColossalChat/tests/generate_dummy_datasets_for_testing.py new file mode 100644 index 000000000000..9f85b4beb65d --- /dev/null +++ b/applications/ColossalChat/tests/generate_dummy_datasets_for_testing.py @@ -0,0 +1,72 @@ +import argparse +import json +import os + +sft_seed = { + "messages": [ + {"from": "human", "content": "Give three tips for staying healthy."}, + { + "from": "assistant", + "content": "1.Eat a balanced diet and make sure to include plenty of fruits and vegetables. \n2. Exercise regularly to keep your body active and strong. \n3. Get enough sleep and maintain a consistent sleep schedule.", + }, + ] +} +prompt_seed = { + "messages": [ + {"from": "human", "content": "Describe the impacts of climate change on communities living in coastal areas."}, + { + "from": "assistant", + "content": "Climate change has caused an increase in sea levels, which has caused coastal erosion and flooding of low-lying areas. This has led to displacement of people from their homes, as well as increased risk of epidemics of waterborne illnesses. Coastal cities have also seen an increase in extreme weather events such as hurricanes and tropical storms, which can cause extensive damage to infrastructure, homes, and businesses. As a result of climate change, some coastal areas are becoming uninhabitable, forcing communities to seek alternative living arrangements.", + }, + ] +} +preference_seed = { + "context": [ + {"from": "human", "content": "What kind of noises did dinosaurs make?"}, + { + "from": "assistant", + "content": "Humans and dinosaurs didn't live at the same time, so it's really hard to say. The best place to find out what noises dinosaurs made would be", + }, + {"from": "human", "content": "yes they did"}, + { + "from": "assistant", + "content": "to guess, and that would probably require lots of reading and a certain amount of imagination, so we're not really prepared to do that.", + }, + {"from": "human", "content": "you cant read"}, + ], + "chosen": [{"from": "assistant", "content": "You can read?"}], + "rejected": [{"from": "assistant", "content": "there's a lot of stuff humans don't know"}], +} + +if __name__ == "__main__": + parser = argparse.ArgumentParser() + parser.add_argument( + "--data_dir", + type=str, + required=True, + default=None, + help="The output dir", + ) + parser.add_argument( + "--data_type", + type=str, + required=True, + default=None, + help="The type of data", + ) + args = parser.parse_args() + if args.data_type == "sft": + seed = sft_seed + elif args.data_type == "prompt": + seed = prompt_seed + elif args.data_type == "preference": + seed = preference_seed + else: + raise ValueError(f"Unknown data type {args.data_type}") + + line = json.dumps(seed, ensure_ascii=False) + "\n" + for idx in [1, 2, 3]: + with open(os.path.join(args.data_dir, f"{idx}.jsonl"), "w", encoding="utf8") as f: + for i in range(1000): + f.write(line) + f.write(line) diff --git a/applications/ColossalChat/tests/llama.json b/applications/ColossalChat/tests/llama.json new file mode 100644 index 000000000000..482ff9e6528c --- /dev/null +++ b/applications/ColossalChat/tests/llama.json @@ -0,0 +1,8 @@ +{ + "chat_template": "{% if not add_generation_prompt is defined %}{% set add_generation_prompt = false %}{% endif %}{% for message in messages %}{% if message['role'] == 'user' %}{{'Human: ' + bos_token + message['content'].strip() + eos_token }}{% elif message['role'] == 'system' %}{{ message['content'].strip() + '\\n\\n' }}{% elif message['role'] == 'assistant' %}{{ 'Assistant: ' + bos_token + message['content'].strip() + eos_token }}{% endif %}{% endfor %}{% if add_generation_prompt %}{{ 'Assistant: ' + bos_token }}{% endif %}", + "system_message": "A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions.\n\n", + "stop_ids": [ + 29871, + 2 + ] +} diff --git a/applications/ColossalChat/tests/opt.json b/applications/ColossalChat/tests/opt.json new file mode 100644 index 000000000000..6d47666bb056 --- /dev/null +++ b/applications/ColossalChat/tests/opt.json @@ -0,0 +1,17 @@ +{ + "chat_template": "{% for message in messages %}{% if message['role'] == 'user' %}{{'Human: ' + bos_token + message['content'].strip() + eos_token }}{% elif message['role'] == 'system' %}{{ message['content'].strip() + '\\n\\n' }}{% elif message['role'] == 'assistant' %}{{ 'Assistant: ' + bos_token + message['content'].strip() + eos_token }}{% endif %}{% endfor %}", + "system_message": "A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions.\n\n", + "human_line_start": [ + 2 + ], + "human_line_end": [ + 2 + ], + "assistant_line_start": [ + 2 + ], + "assistant_line_end": [ + 2 + ], + "end_of_system_line_position": 0 +} diff --git a/applications/ColossalChat/tests/test_data/dpo/test_dpo_data.jsonl b/applications/ColossalChat/tests/test_data/dpo/test_dpo_data.jsonl new file mode 100644 index 000000000000..2e11a91c643f --- /dev/null +++ b/applications/ColossalChat/tests/test_data/dpo/test_dpo_data.jsonl @@ -0,0 +1 @@ +{"context": [{"from": "human", "content": "If I were to give you some coordinates, could you tell me where the location is?"}], "chosen": [{"from": "assistant", "content": "Certainly! I can tell you if the given coordinates are within any geographic location, such as a country or city. However, I cannot give accurate directions to any specific location without more information. Are you looking for a more general geographic location, or do you want more precise directions?"}], "rejected": [{"from": "assistant", "content": "Yes, I can determine the location and give you some details about the area. Please provide the latitude and longitude coordinates, and I will return a description of the location based on these coordinates. Please note that precise geographical coordinates are often imprecise, so please provide the best possible coordinates you have available. Here are some examples:\n-The coordinates 41.543554,-71.09471 locate the Boston Commons in Boston, Massachusetts\n-The coordinates 37.42242, -122.083733 locate the Palace of Fine Arts in San Francisco, California \n-The coordinates 25.774145, -80.190262 locate Canal Street in Miami Beach, Florida\n-The coordinates 35. journalisten nl 48.056944, -106.277778 locate the Wyoming State Penitentiary in Rawlins, Wyoming\n-The coordinates 51.507351,-0.127758 locate Tower Bridge in London, England\n-The coordinates 44.885618, -79.422594 locate the Stratosphere Casino, Hotel and Tower in Las Vegas, Nevada"}]} diff --git a/applications/ColossalChat/tests/test_data/sft/test_sft_data.jsonl b/applications/ColossalChat/tests/test_data/sft/test_sft_data.jsonl new file mode 100644 index 000000000000..21c4d9dc76ec --- /dev/null +++ b/applications/ColossalChat/tests/test_data/sft/test_sft_data.jsonl @@ -0,0 +1 @@ +{"messages": [{"from": "human", "content": "what are some pranks with a pen I can do?"}, {"from": "assistant", "content": "Are you looking for practical joke ideas?"}, {"from": "human", "content": "Yes, exactly!"}, {"from": "assistant", "content": "You could try attaching a small buzzer to a pen. When someone tries to use it, they'll be surprised by the sound!"}, {"from": "human", "content": "That's a funny one! Any other ideas?"}, {"from": "assistant", "content": "Another idea is the classic 'ink explosion' prank. Replace the ink of a pen with disappearing ink, and watch the confusion when the ink vanishes from the paper."}]} diff --git a/applications/ColossalChat/tests/test_data_preparation.sh b/applications/ColossalChat/tests/test_data_preparation.sh new file mode 100755 index 000000000000..a7689cdc6688 --- /dev/null +++ b/applications/ColossalChat/tests/test_data_preparation.sh @@ -0,0 +1,260 @@ +#!/usr/bin/env bash +set_n_least_used_CUDA_VISIBLE_DEVICES() { + local n=${1:-"9999"} + echo "GPU Memory Usage:" + local FIRST_N_GPU_IDS=$(nvidia-smi --query-gpu=memory.used --format=csv | + tail -n +2 | + nl -v 0 | + tee /dev/tty | + sort -g -k 2 | + awk '{print $1}' | + head -n $n) + export CUDA_VISIBLE_DEVICES=$(echo $FIRST_N_GPU_IDS | sed 's/ /,/g') + echo "Now CUDA_VISIBLE_DEVICES is set to:" + echo "CUDA_VISIBLE_DEVICES=$CUDA_VISIBLE_DEVICES" +} + +set_n_least_used_CUDA_VISIBLE_DEVICES 4 + +set -xu + +if [ -z "$SFT_DATASET" ]; then + echo "Please set \$SFT_DATASET to the path to sft dataset." + exit 1 +fi + +if [ -z "$PROMPT_DATASET" ]; then + echo "Please set \$PROMPT_DATASET to the path to prompts." + exit 1 +fi + +if [ -z "$PREFERENCE_DATASET" ]; then + echo "Please set \$SFT_DATASET to the path to sft dataset." + exit 1 +fi + +NUM_RETRY=3 +BASE_DIR=$(dirname $(dirname $(realpath $BASH_SOURCE))) +BASE_TEMP_DIR=$BASE_DIR/temp +TEST_DIR=$BASE_DIR/tests +EXAMPLES_DIR=$BASE_DIR/examples +DATA_SAVE_PATH=$BASE_TEMP_DIR/rlhf_data +CONFIG_DIR=$BASE_DIR/config +# Skip those tests due to CI tests timeout +MODELS=('llama') + +if [ ! -d "$BASE_TEMP_DIR" ]; then + mkdir "$BASE_TEMP_DIR" + echo "Directory created successfully" +else + echo "Directory already exists" +fi + +if [ ! -d "$DATA_SAVE_PATH" ]; then + mkdir "$DATA_SAVE_PATH" + echo "Directory created successfully" +else + echo "Directory already exists" +fi + + +export OMP_NUM_THREADS=8 + +# install requirements +pip install -r $EXAMPLES_DIR/requirements.txt + +get_data_input_dirs() { + local data_type=$1 + if [[ $data_type == "sft" ]]; then + echo "$SFT_DATASET" + elif [[ $data_type == "prompt" ]]; then + echo "$PROMPT_DATASET" + elif [[ $data_type == "preference" ]]; then + echo "$PREFERENCE_DATASET" + else + echo "Unknown data type $data_type" + exit 1 + fi +} + +get_conversation_template_config() { + local model=$1 + if [[ $model == "llama" ]]; then + echo "$TEST_DIR/llama.json" + elif [[ $model == "opt" ]]; then + echo "$TEST_DIR/opt.json" + else + echo "Unknown model $model" + exit 1 + fi +} + +get_tokenizer_dirs() { + local model=$1 + if [[ $model == "llama" ]]; then + echo "hf-internal-testing/llama-tokenizer" + elif [[ $model == "opt" ]]; then + echo "facebook/opt-125m" + else + echo "Unknown model $model" + exit 1 + fi +} + +random_choice() { + local arr=("$@") + local len=${#arr[@]} + local idx=$((RANDOM % len)) + echo ${arr[$idx]} +} + +echo "Prepare dummy data for testing..." +python $TEST_DIR/generate_dummy_datasets_for_testing.py \ + --data_dir $(get_data_input_dirs sft) \ + --data_type "sft" + +python $TEST_DIR/generate_dummy_datasets_for_testing.py \ + --data_dir $(get_data_input_dirs preference) \ + --data_type "preference" + +python $TEST_DIR/generate_dummy_datasets_for_testing.py \ + --data_dir $(get_data_input_dirs prompt) \ + --data_type "prompt" + +echo "[Test]: testing prepare_preference_dataset.py ..." + +# FIXME: This is a hack to skip tests that are not working +SKIPPED_TESTS=( +) + +# test prepare_preference_dataset +for model in ${MODELS[@]}; do + data_type="preference" + if [[ " ${SKIPPED_TESTS[*]} " =~ " $model-$data_type " ]]; then + echo "[Test]: Skipped $model-$data_type" + continue + fi + cache_dir=$DATA_SAVE_PATH/tokenized_${model}_${data_type}/cache + jsonl_dir=$DATA_SAVE_PATH/tokenized_${model}_${data_type}/jsonl + arrow_dir=$DATA_SAVE_PATH/tokenized_${model}_${data_type}/arrow + rm -rf $cache_dir + rm -rf $jsonl_dir + rm -rf $arrow_dir + data_input_dirs=$(get_data_input_dirs $data_type) + tokenizer_dir=$(get_tokenizer_dirs $model) + conversation_template=$(get_conversation_template_config $model) + for i in $(seq $NUM_RETRY); do + echo "[Test]: $model-$data_type, attempt $i" + python $EXAMPLES_DIR/data_preparation_scripts/prepare_dataset.py \ + --type preference \ + --data_input_dirs $data_input_dirs \ + --conversation_template_config $conversation_template \ + --tokenizer_dir $tokenizer_dir \ + --data_cache_dir $cache_dir \ + --data_jsonl_output_dir $jsonl_dir \ + --data_arrow_output_dir $arrow_dir \ + --max_length 400 \ + --num_samples_per_datafile 100 \ + --num_spliced_dataset_bins 1 + passed=$? + if [ $passed -eq 0 ]; then + break + fi + done + if [ $passed -ne 0 ]; then + echo "[Test]: Failed $model-$data_type" + exit 1 + fi +done + +echo "[Test]: testing prepare_sft_dataset.py ..." + +# FIXME: This is a hack to skip tests that are not working +SKIPPED_TESTS=( +) + +# test prepare_sft_dataset +for model in ${MODELS[@]}; do + data_type="sft" + if [[ " ${SKIPPED_TESTS[*]} " =~ " $model-$data_type " ]]; then + echo "[Test]: Skipped $model-$data_type" + continue + fi + cache_dir=$DATA_SAVE_PATH/tokenized_${model}_${data_type}/cache + jsonl_dir=$DATA_SAVE_PATH/tokenized_${model}_${data_type}/jsonl + arrow_dir=$DATA_SAVE_PATH/tokenized_${model}_${data_type}/arrow + data_input_dirs=$(get_data_input_dirs $data_type) + tokenizer_dir=$(get_tokenizer_dirs $model) + conversation_template=$(get_conversation_template_config $model) + for i in $(seq $NUM_RETRY); do + rm -rf $cache_dir + rm -rf $jsonl_dir + rm -rf $arrow_dir + echo "[Test]: $model-$data_type, attempt $i" + python $EXAMPLES_DIR/data_preparation_scripts/prepare_dataset.py \ + --type sft \ + --data_input_dirs $data_input_dirs \ + --conversation_template_config $conversation_template \ + --tokenizer_dir $tokenizer_dir \ + --data_cache_dir $cache_dir \ + --data_jsonl_output_dir $jsonl_dir \ + --data_arrow_output_dir $arrow_dir \ + --max_length 400 \ + --num_samples_per_datafile 100 \ + --num_spliced_dataset_bins 1 + passed=$? + if [ $passed -eq 0 ]; then + break + fi + done + if [ $passed -ne 0 ]; then + echo "[Test]: Failed $model-$data_type" + exit 1 + fi +done + +echo "[Test]: testing prepare_prompt_dataset.py ..." + +# FIXME: This is a hack to skip tests that are not working +SKIPPED_TESTS=( +) + +# test prepare_prompt_dataset +for model in ${MODELS[@]}; do + data_type="prompt" + if [[ " ${SKIPPED_TESTS[*]} " =~ " $model-$data_type " ]]; then + echo "[Test]: Skipped $model-$data_type" + continue + fi + cache_dir=$DATA_SAVE_PATH/tokenized_${model}_${data_type}/cache + jsonl_dir=$DATA_SAVE_PATH/tokenized_${model}_${data_type}/jsonl + arrow_dir=$DATA_SAVE_PATH/tokenized_${model}_${data_type}/arrow + data_input_dirs=$(get_data_input_dirs $data_type) + tokenizer_dir=$(get_tokenizer_dirs $model) + conversation_template=$(get_conversation_template_config $model) + for i in $(seq $NUM_RETRY); do + rm -rf $cache_dir + rm -rf $jsonl_dir + rm -rf $arrow_dir + echo "[Test]: $model-$data_type, attempt $i" + python $EXAMPLES_DIR/data_preparation_scripts/prepare_dataset.py \ + --type prompt \ + --data_input_dirs $data_input_dirs \ + --conversation_template_config $conversation_template \ + --tokenizer_dir $tokenizer_dir \ + --data_cache_dir $cache_dir \ + --data_jsonl_output_dir $jsonl_dir \ + --data_arrow_output_dir $arrow_dir \ + --max_length 400 \ + --num_samples_per_datafile 100 \ + --num_spliced_dataset_bins 1 + passed=$? + if [ $passed -eq 0 ]; then + break + fi + done + if [ $passed -ne 0 ]; then + echo "[Test]: Failed $model-$data_type" + exit 1 + fi +done diff --git a/applications/ColossalChat/tests/test_lora.py b/applications/ColossalChat/tests/test_lora.py new file mode 100755 index 000000000000..4ea9e1a15c59 --- /dev/null +++ b/applications/ColossalChat/tests/test_lora.py @@ -0,0 +1,69 @@ +import torch +import torch.nn as nn +import torch.optim as optim +from coati.models import convert_to_lora_module +from torch.utils.data import DataLoader, TensorDataset + + +class SimpleNN(nn.Module): + def __init__(self, input_size, hidden_size, num_classes): + super(SimpleNN, self).__init__() + self.fc1 = nn.Linear(input_size, hidden_size) + self.relu = nn.ReLU() + self.fc2 = nn.Linear(hidden_size, num_classes) + + def forward(self, x): + out = self.fc1(x) + out = self.relu(out) + out = self.fc2(out) + return out + + +def test_overfit(): + input_size = 1000 + hidden_size = 200 + num_classes = 5 + batch_size = 64 + learning_rate = 0.01 + num_epochs = 200 + + # Synthesized dataset + X = torch.randn(batch_size, input_size) + Y = torch.randint(0, num_classes, (batch_size,)) + + # Convert to DataLoader + dataset = TensorDataset(X, Y) + loader = DataLoader(dataset, batch_size=batch_size, shuffle=True) + + # Build and convert model + model = SimpleNN(input_size, hidden_size, num_classes) + weight_to_compare = model.fc1.weight.detach().clone() + model = convert_to_lora_module(model, lora_rank=30) + + # Loss and optimizer + criterion = nn.CrossEntropyLoss() + optimizer = optim.Adam(model.parameters(), lr=learning_rate) + + # Train the model + for _ in range(num_epochs): + for i, (inputs, labels) in enumerate(loader): + # Forward pass + outputs = model(inputs) + loss = criterion(outputs, labels) + print(loss) + # Backward and optimize + optimizer.zero_grad() + loss.backward() + optimizer.step() + + # Check if model has overfitted + outputs = model(X) + _, predicted = torch.max(outputs.data, 1) + total = labels.size(0) + correct = (predicted == Y).sum().item() + assert (correct / total > 0.95, "The model has not overfitted to the synthesized dataset") + assert (weight_to_compare - model.fc1.weight).sum() < 0.01 + + +if __name__ == "__main__": + test_overfit() diff --git a/applications/ColossalChat/tests/test_templating.sh b/applications/ColossalChat/tests/test_templating.sh new file mode 100755 index 000000000000..7fefede47539 --- /dev/null +++ b/applications/ColossalChat/tests/test_templating.sh @@ -0,0 +1,97 @@ + +BASE_DIR=$(dirname $(dirname $(realpath $BASH_SOURCE))) +BASE_TEMP_DIR=$BASE_DIR/temp +EXAMPLES_DIR=$BASE_DIR/examples +TEST_DATA_DIR=$BASE_DIR/tests/test_data +DATA_SAVE_PATH=$BASE_TEMP_DIR/tests +CONFIG_DIR=$BASE_DIR/config + +MODELS=("colossal-llama2" "llama2" "zephyr" "mistral" "chatGLM2" "Qwen" "Vicuna" "Yi") + +get_pretrain() { + local model=$1 + if [[ $model == "colossal-llama2" ]]; then + echo "hpcai-tech/Colossal-LLaMA-2-7b-base" + elif [[ $model == "llama2" ]]; then + echo "hf-internal-testing/llama-tokenizer" + elif [[ $model == "zephyr" ]]; then + echo "HuggingFaceH4/zephyr-7b-beta" + elif [[ $model == "mistral" ]]; then + echo "mistralai/Mistral-7B-Instruct-v0.2" + elif [[ $model == "chatGLM2" ]]; then + echo "THUDM/chatglm2-6b" + elif [[ $model == "Qwen" ]]; then + echo "Qwen/Qwen-7B-Chat" + elif [[ $model == "Vicuna" ]]; then + echo "lmsys/vicuna-7b-v1.5" + elif [[ $model == "Yi" ]]; then + echo "01-ai/Yi-6B-Chat" + else + echo "Unknown model $model" + exit 1 + fi +} + +get_conversation_template_config() { + local model=$1 + echo "$CONFIG_DIR/conversation_template/$model.json" +} + +# Test SFT data Preparation +for model in ${MODELS[@]}; do + echo "Testing SFT data templating for $model" + SAVE_DIR=$DATA_SAVE_PATH/sft/$model + rm -rf $SAVE_DIR/cache + rm -rf $SAVE_DIR/jsonl + rm -rf $SAVE_DIR/arrow + pretrain=$(get_pretrain $model) + conversation_template_config=$(get_conversation_template_config $model) + python $EXAMPLES_DIR/data_preparation_scripts/prepare_dataset.py --type sft --data_input_dirs $TEST_DATA_DIR/sft \ + --tokenizer_dir $pretrain \ + --conversation_template_config $conversation_template_config \ + --data_cache_dir $SAVE_DIR/cache \ + --data_jsonl_output_dir $SAVE_DIR/jsonl \ + --data_arrow_output_dir $SAVE_DIR/arrow + passed=$? + if [ $passed -ne 0 ]; then + echo "[Test]: Failed in the SFT data templating for $model" + exit 1 + fi + python $BASE_DIR/tests/verify_chat_data.py --data_source $TEST_DATA_DIR/sft/test_sft_data.jsonl \ + --to_verify_file $SAVE_DIR/jsonl/part-00005.jsonl --data_type sft + passed=$? + if [ $passed -ne 0 ]; then + echo "[Test]: Failed in the SFT data templating test for $model" + exit 1 + fi +done + + +# Test DPO/PPO data Preparation +for model in ${MODELS[@]}; do + echo "Testing DPO/PPO data templating for $model" + SAVE_DIR=$DATA_SAVE_PATH/dpo/$model + rm -rf $SAVE_DIR/cache + rm -rf $SAVE_DIR/jsonl + rm -rf $SAVE_DIR/arrow + pretrain=$(get_pretrain $model) + conversation_template_config=$(get_conversation_template_config $model) + python $EXAMPLES_DIR/data_preparation_scripts/prepare_dataset.py --type preference --data_input_dirs $TEST_DATA_DIR/dpo \ + --tokenizer_dir $pretrain \ + --conversation_template_config $conversation_template_config \ + --data_cache_dir $SAVE_DIR/cache \ + --data_jsonl_output_dir $SAVE_DIR/jsonl \ + --data_arrow_output_dir $SAVE_DIR/arrow + passed=$? + if [ $passed -ne 0 ]; then + echo "[Test]: Failed in the DPO data templating for $model" + exit 1 + fi + python $BASE_DIR/tests/verify_chat_data.py --data_source $TEST_DATA_DIR/dpo/test_dpo_data.jsonl \ + --to_verify_file $SAVE_DIR/jsonl/part-00005.jsonl --data_type dpo + passed=$? + if [ $passed -ne 0 ]; then + echo "[Test]: Failed in the DPO data templating test for $model" + exit 1 + fi +done diff --git a/applications/ColossalChat/tests/test_train.sh b/applications/ColossalChat/tests/test_train.sh new file mode 100755 index 000000000000..5ba4904711ea --- /dev/null +++ b/applications/ColossalChat/tests/test_train.sh @@ -0,0 +1,397 @@ +#!/usr/bin/env bash + +set_n_least_used_CUDA_VISIBLE_DEVICES() { + local n=${1:-"9999"} + echo "GPU Memory Usage:" + local FIRST_N_GPU_IDS=$(nvidia-smi --query-gpu=memory.used --format=csv | + tail -n +2 | + nl -v 0 | + tee /dev/tty | + sort -g -k 2 | + awk '{print $1}' | + head -n $n) + export CUDA_VISIBLE_DEVICES=$(echo $FIRST_N_GPU_IDS | sed 's/ /,/g') + echo "Now CUDA_VISIBLE_DEVICES is set to:" + echo "CUDA_VISIBLE_DEVICES=$CUDA_VISIBLE_DEVICES" +} + +set_n_least_used_CUDA_VISIBLE_DEVICES 4 + +set -xu + + +NUM_RETRY=3 +BASE_DIR=$(dirname $(dirname $(realpath $BASH_SOURCE))) +EXAMPLES_DIR=$BASE_DIR/examples +CONFIG_DIR=$BASE_DIR/config +TEMP_DIR=$BASE_DIR/temp +TEST_DIR=$BASE_DIR/tests +MODEL_SAVE_PATH=$TEMP_DIR/rlhf_models +MODELS_DIR=$TEMP_DIR/models_config +# Skip those tests due to CI tests timeout +MODELS=('llama') +PLUGINS=('gemini' 'gemini_auto' 'zero2' 'zero2_cpu' '3d') +LORA_RANK=('0') # skip to reduce CI execution time, can pass all locally + +export OMP_NUM_THREADS=8 + +get_pretrain() { + local model=$1 + if [[ $model == "llama" ]]; then + echo "nickypro/tinyllama-110M" + elif [[ $model == "opt" ]]; then + echo "facebook/opt-125m" + else + echo "Unknown model $model" + exit 1 + fi +} + +get_tokenizer_dirs() { + local model=$1 + if [[ $model == "llama" ]]; then + echo "hf-internal-testing/llama-tokenizer" + elif [[ $model == "opt" ]]; then + echo "facebook/opt-125m" + else + echo "Unknown model $model" + exit 1 + fi +} + + +get_conversation_template_config() { + local model=$1 + if [[ $model == "llama" ]]; then + echo "$TEST_DIR/llama.json" + elif [[ $model == "opt" ]]; then + echo "$TEST_DIR/opt.json" + else + echo "Unknown model $model" + exit 1 + fi +} + +random_choice() { + local arr=("$@") + local len=${#arr[@]} + local idx=$((RANDOM % len)) + echo ${arr[$idx]} +} + + +echo "[Test]: testing sft ..." + +SKIPPED_TESTS=( + llama-3d-20 # 3d plugin doesn't support lora + llama-gemini_auto-20 # gemini_auto plugin doesn't support lora + llama-gemini-20 # gemini doesn't support lora +) + +GRAD_CKPTS=('--grad_checkpoint') +for lora_rank in ${LORA_RANK[@]}; do + for model in ${MODELS[@]}; do + for plugin in ${PLUGINS[@]}; do + if [[ " ${SKIPPED_TESTS[*]} " =~ " $model-$plugin-$lora_rank " ]]; then + echo "[Test]: Skipped $model-$plugin-$lora_rank" + continue + elif [[ " ${SKIPPED_TESTS[*]} " =~ " $model-$plugin " ]]; then + echo "[Test]: Skipped $model-$plugin" + continue + fi + pretrain=$(get_pretrain $model) + tokenizer_dir=$(get_tokenizer_dirs $model) + grad_ckpt=$(random_choice "${GRAD_CKPTS[@]}") + tp='1' + bs='2' + if [[ $plugin == "3d" ]]; then + tp='4' + bs='8' + fi + grad_accu='2' + # Check if the plugin is either "gemini_auto" or "gemini" and set grad_accu to '1' + if [[ $plugin == "gemini_auto" ]]; then + grad_accu='1' + fi + + for i in $(seq $NUM_RETRY); do + echo "[Test]: $model-$plugin-$lora_rank, attempt $i" + declare -a dataset=() + for split in $(seq -f "%05g" 0 0); do + dataset+=("$TEMP_DIR/rlhf_data/tokenized_${model}_sft/arrow/part-$split") + done + colossalai run --nproc_per_node 4 --master_port 31332 $EXAMPLES_DIR/training_scripts/train_sft.py \ + --pretrain $pretrain \ + --tokenizer_dir $tokenizer_dir \ + --dataset ${dataset[@]} \ + --save_path $MODEL_SAVE_PATH \ + --config_file $MODELS_DIR/config.jsonl \ + --lora_rank $lora_rank \ + --plugin $plugin \ + --batch_size $bs \ + --max_epochs 1 \ + --accumulation_steps $grad_accu \ + --tp $tp \ + --lr 2e-5 \ + $grad_ckpt \ + --max_len 400 \ + --use_flash_attn + passed=$? + if [ $passed -eq 0 ]; then + rm -rf $MODEL_SAVE_PATH/* + rm -rf $MODELS_DIR/* + break + fi + done + if [ $passed -ne 0 ]; then + echo "[Test]: Failed $model-$plugin-$lora_rank" + exit 1 + fi + done + done +done + +echo "[Test]: testing reward model ..." + +SKIPPED_TESTS=( + llama-3d-20 # 3d plugin doesn't support lora + llama-gemini_auto-20 # gemini_auto plugin doesn't support lora + llama-gemini-20 # gemini doesn't support lora +) + +GRAD_CKPTS=('--grad_checkpoint') +for lora_rank in ${LORA_RANK[@]}; do + for model in ${MODELS[@]}; do + for plugin in ${PLUGINS[@]}; do + if [[ " ${SKIPPED_TESTS[*]} " =~ " $model-$plugin-$lora_rank " ]]; then + echo "[Test]: Skipped $model-$plugin-$lora_rank" + continue + elif [[ " ${SKIPPED_TESTS[*]} " =~ " $model-$plugin " ]]; then + echo "[Test]: Skipped $model-$plugin" + continue + fi + pretrain=$(get_pretrain $model) + tokenizer_dir=$(get_tokenizer_dirs $model) + grad_ckpt=$(random_choice "${GRAD_CKPTS[@]}") + tp='1' + bs='2' + if [[ $plugin == "3d" ]]; then + tp='4' + bs='8' + fi + grad_accu='2' + # gemini_auto and gemini doesn't support gradient accumulation + if [[ $plugin == "gemini_auto" ]]; then + grad_accu='1' + fi + for i in $(seq $NUM_RETRY); do + echo "[Test]: $model-$plugin-$lora_rank, attempt $i" + declare -a dataset=() + for split in $(seq -f "%05g" 0 0); do + dataset+=("$TEMP_DIR/rlhf_data/tokenized_${model}_preference/arrow/part-$split") + done + colossalai run --nproc_per_node 4 --master_port 31332 $EXAMPLES_DIR/training_scripts/train_rm.py \ + --pretrain $pretrain \ + --tokenizer_dir $tokenizer_dir \ + --dataset ${dataset[@]} \ + --save_dir $MODEL_SAVE_PATH \ + --config_file $MODELS_DIR/config.jsonl \ + --lora_rank $lora_rank \ + --plugin $plugin \ + --batch_size $bs \ + --max_epochs 1 \ + --accumulation_steps $grad_accu \ + --tp $tp \ + --lr 2e-5 \ + $grad_ckpt \ + --max_len 400 \ + --use_flash_attn + passed=$? + if [ $passed -eq 0 ]; then + rm -rf $MODEL_SAVE_PATH/* + rm -rf $MODELS_DIR/* + break + fi + done + if [ $passed -ne 0 ]; then + echo "[Test]: Failed $model-$plugin-$lora_rank" + exit 1 + fi + done + done +done + + +echo "[Test]: testing ppo ..." + + +SKIPPED_TESTS=( + llama-3d-20 # 3d plugin doesn't support lora + llama-gemini-20 # gemini doesn't support lora +) + +GRAD_CKPTS=('--grad_checkpoint') +for lora_rank in ${LORA_RANK[@]}; do + for model in ${MODELS[@]}; do + for plugin in ${PLUGINS[@]}; do + if [[ $plugin == "gemini_auto" ]]; then + echo "[Test]: Skipped $model-$plugin" + continue # gemini_auto plugin doesn't support generation + fi + if [[ " ${SKIPPED_TESTS[*]} " =~ " $model-$plugin-$lora_rank " ]]; then + echo "[Test]: Skipped $model-$plugin-$lora_rank" + continue + elif [[ " ${SKIPPED_TESTS[*]} " =~ " $model-$plugin " ]]; then + echo "[Test]: Skipped $model-$plugin" + continue + fi + pretrain=$(get_pretrain $model) + tokenizer_dir=$(get_tokenizer_dirs $model) + grad_ckpt=$(random_choice "${GRAD_CKPTS[@]}") + tp='1' + bs='4' + ebs='8' + conversation_template=$(get_conversation_template_config $model) + if [[ $plugin == "3d" ]]; then + tp='4' + bs='16' + ebs='32' + fi + grad_accu='2' + # gemini_auto and gemini doesn't support gradient accumulation + if [[ $plugin == "gemini_auto" ]]; then + grad_accu='1' + fi + # gemini_auto and gemini doesn't support generation + if [[ $plugin == "gemini_auto" ]]; then + # gemini-auto doesn't support generation + echo "[Test]: Skipped $model-$plugin" + continue + fi + for i in $(seq $NUM_RETRY); do + echo "[Test]: $model-$plugin-$lora_rank, attempt $i" + declare -a prompt_dataset=() + for split in $(seq -f "%05g" 0 0); do + prompt_dataset+=("$TEMP_DIR/rlhf_data/tokenized_${model}_prompt/arrow/part-$split") + done + declare -a ptx_dataset=() + for split in $(seq -f "%05g" 0 0); do + ptx_dataset+=("$TEMP_DIR/rlhf_data/tokenized_${model}_sft/arrow/part-$split") + done + colossalai run --nproc_per_node 4 --master_port 31332 $EXAMPLES_DIR/training_scripts/train_ppo.py \ + --pretrain $pretrain \ + --rm_pretrain $pretrain \ + --tokenizer_dir $tokenizer_dir \ + --conversation_template_config $conversation_template \ + --prompt_dataset ${prompt_dataset[@]} \ + --ptx_dataset ${ptx_dataset[@]} \ + --ptx_batch_size 1 \ + --ptx_coef 0.2 \ + --save_path $MODEL_SAVE_PATH \ + --lora_rank $lora_rank \ + --plugin $plugin \ + --num_episodes 5 \ + --num_collect_steps 1 \ + --num_update_steps 1 \ + --experience_batch_size $ebs \ + --train_batch_size $bs \ + --accumulation_steps $grad_accu \ + --lr 9e-6 \ + --mixed_precision "bf16" \ + --grad_clip 1.0 \ + --tp $tp \ + --lr 2e-5 \ + $grad_ckpt \ + --max_len 400 \ + --max_seq_len 10 \ + --use_flash_attn + passed=$? + if [ $passed -eq 0 ]; then + rm -rf $MODEL_SAVE_PATH/* + rm -rf $MODELS_DIR/* + break + fi + done + if [ $passed -ne 0 ]; then + echo "[Test]: Failed $model-$plugin-$lora_rank" + exit 1 + fi + done + done +done + + +echo "[Test]: testing DPO ..." + +SKIPPED_TESTS=( + llama-3d-20 # 3d plugin doesn't support lora + llama-gemini_auto-20 # gemini_auto plugin doesn't support lora + llama-gemini-20 # gemini doesn't support lora +) +GRAD_CKPTS=('--grad_checkpoint') +for lora_rank in ${LORA_RANK[@]}; do + for model in ${MODELS[@]}; do + for plugin in ${PLUGINS[@]}; do + if [[ " ${SKIPPED_TESTS[*]} " =~ " $model-$plugin-$lora_rank " ]]; then + echo "[Test]: Skipped $model-$plugin-$lora_rank" + continue + elif [[ " ${SKIPPED_TESTS[*]} " =~ " $model-$plugin " ]]; then + echo "[Test]: Skipped $model-$plugin" + continue + fi + pretrain=$(get_pretrain $model) + tokenizer_dir=$(get_tokenizer_dirs $model) + grad_ckpt=$(random_choice "${GRAD_CKPTS[@]}") + tp='1' + bs='2' + if [[ $plugin == "3d" ]]; then + tp='4' + bs='8' + fi + grad_accu='2' + # gemini_auto and gemini doesn't support gradient accumulation + if [[ $plugin == "gemini_auto" ]]; then + grad_accu='1' + fi + # gemini_auto doesn't support generation + # (need to calculate ref_model logits through forwarding in inference mode) + if [[ $plugin == "gemini_auto" ]]; then + echo "[Test]: Skipped $model-$plugin" + continue + fi + for i in $(seq $NUM_RETRY); do + echo "[Test]: $model-$plugin-$lora_rank, attempt $i" + declare -a dataset=() + for split in $(seq -f "%05g" 0 0); do + dataset+=("$TEMP_DIR/rlhf_data/tokenized_${model}_preference/arrow/part-$split") + done + colossalai run --nproc_per_node 4 --master_port 31332 $EXAMPLES_DIR/training_scripts/train_dpo.py \ + --pretrain $pretrain \ + --tokenizer_dir $tokenizer_dir \ + --dataset ${dataset[@]} \ + --save_dir $MODEL_SAVE_PATH \ + --config_file $MODELS_DIR/config.jsonl \ + --lora_rank $lora_rank \ + --plugin $plugin \ + --batch_size $bs \ + --max_epochs 1 \ + --accumulation_steps $grad_accu \ + --tp $tp \ + --lr 2e-5 \ + $grad_ckpt \ + --max_len 400 \ + --use_flash_attn + passed=$? + if [ $passed -eq 0 ]; then + rm -rf $MODEL_SAVE_PATH/* + rm -rf $MODELS_DIR/* + break + fi + done + if [ $passed -ne 0 ]; then + echo "[Test]: Failed $model-$plugin-$lora_rank" + exit 1 + fi + done + done +done diff --git a/applications/ColossalChat/tests/verify_chat_data.py b/applications/ColossalChat/tests/verify_chat_data.py new file mode 100644 index 000000000000..98ae0c1b2d28 --- /dev/null +++ b/applications/ColossalChat/tests/verify_chat_data.py @@ -0,0 +1,64 @@ +import argparse +import json + +if __name__ == "__main__": + parser = argparse.ArgumentParser() + parser.add_argument( + "--data_source", + type=str, + required=True, + default=None, + help="The raw data file", + ) + parser.add_argument( + "--to_verify_file", + type=str, + required=True, + default=None, + help="The file that contains the data to be verified", + ) + parser.add_argument( + "--data_type", + type=str, + required=True, + default=None, + help="The data type", + ) + args = parser.parse_args() + + # Read data + data = [] + with open(args.data_source, "r", encoding="utf8") as f: + for line in f.readlines(): + data.append(json.loads(line)) + to_verify_data = [] + with open(args.to_verify_file, "r", encoding="utf8") as f: + for line in f.readlines(): + to_verify_data.append(json.loads(line)) + + if args.data_type == "sft": + target_lable = [msg["content"].strip() for msg in data[0]["messages"] if msg["from"] == "assistant"] + target_negative_label = [msg["content"].strip() for msg in data[0]["messages"] if msg["from"] == "human"] + + # Read to verify file + + to_verify_lable = to_verify_data[0]["labels_decode"] + for label in target_lable: + assert any([label in s for s in to_verify_lable]), f"Label {label} not in target label {to_verify_lable}" + for label in target_negative_label: + assert all( + [label not in s for s in to_verify_lable] + ), f"Negative label {label} in target label {to_verify_lable}" + elif args.data_type == "dpo": + chosen_lable = data[0]["chosen"][0]["content"].strip() + rejected_lable = data[0]["rejected"][0]["content"].strip() + + # Read to verify file + to_verify_lable_chosen = to_verify_data[0]["chosen_label_decode"] + to_verify_lable_rejected = to_verify_data[0]["rejected_label_decode"] + assert any( + [chosen_lable in s for s in to_verify_lable_chosen] + ), f"Chosen label {chosen_lable} not in target chosen label {to_verify_lable_chosen}" + assert any( + [rejected_lable in s for s in to_verify_lable_rejected] + ), f"Rejected label {rejected_lable} not in target rejected label {to_verify_lable_chosen}" diff --git a/applications/Chat/version.txt b/applications/ColossalChat/version.txt old mode 100644 new mode 100755 similarity index 100% rename from applications/Chat/version.txt rename to applications/ColossalChat/version.txt diff --git a/applications/ColossalEval/colossal_eval/evaluate/gpt_evaluate.py b/applications/ColossalEval/colossal_eval/evaluate/gpt_evaluate.py index a0b1ed1143f0..19907daaff7f 100644 --- a/applications/ColossalEval/colossal_eval/evaluate/gpt_evaluate.py +++ b/applications/ColossalEval/colossal_eval/evaluate/gpt_evaluate.py @@ -670,7 +670,7 @@ def calculate_scores_form_logprobs(logprobs: Dict[str, Any]) -> float: def calculate_scores_form_response(response: str, evaluation: Dict[str, Any]) -> int: """ Calculate the score from the response returned by gpt-3.5-turbo or gpt-4. - Different from text-davinci-003, this fuction directly calculates the score according to the plain response returned by gpt-3.5-turbo or gpt-4. + Different from text-davinci-003, this function directly calculates the score according to the plain response returned by gpt-3.5-turbo or gpt-4. Although text-davinci-003 can return log probabilities, it costs ten times as much as gpt-3.5-turbo. Args: diff --git a/applications/ColossalMoE/colossal_moe/models/mixtral_policy.py b/applications/ColossalMoE/colossal_moe/models/mixtral_policy.py index 218b05b27fad..c01e02c49a60 100644 --- a/applications/ColossalMoE/colossal_moe/models/mixtral_policy.py +++ b/applications/ColossalMoE/colossal_moe/models/mixtral_policy.py @@ -109,8 +109,8 @@ def set_pipeline_forward(self, model_cls: nn.Module, new_forward: Callable, poli else: module = self.model.model - layers_per_stage = self.distribute_layers(len(module.layers), stage_manager.num_stages) - stage_index = Policy.get_stage_index(layers_per_stage, stage_manager.stage) + layers_per_stage = stage_manager.distribute_layers(len(module.layers)) + stage_index = stage_manager.get_stage_index(layers_per_stage) method_replacement = {"forward": partial(new_forward, stage_manager=stage_manager, stage_index=stage_index)} self.append_or_create_method_replacement( description=method_replacement, policy=policy, target_key=model_cls @@ -129,10 +129,10 @@ def get_held_layers(self) -> List[Module]: stage_manager = self.pipeline_stage_manager held_layers = [] - layers_per_stage = self.distribute_layers(len(module.layers), stage_manager.num_stages) + layers_per_stage = stage_manager.distribute_layers(len(module.layers)) if stage_manager.is_first_stage(): held_layers.append(module.embed_tokens) - start_idx, end_idx = self.get_stage_index(layers_per_stage, stage_manager.stage) + start_idx, end_idx = stage_manager.get_stage_index(layers_per_stage) held_layers.extend(module.layers[start_idx:end_idx]) if stage_manager.is_last_stage(): held_layers.append(module.norm) diff --git a/applications/ColossalMoE/train.py b/applications/ColossalMoE/train.py index c567038ec252..850236726a27 100644 --- a/applications/ColossalMoE/train.py +++ b/applications/ColossalMoE/train.py @@ -128,13 +128,13 @@ def parse_args(): parser.add_argument( "--comm_overlap", action="store_true", - help="Use communication overlap for MoE. Recommended to enable for muiti-node training.", + help="Use communication overlap for MoE. Recommended to enable for multi-node training.", ) # hierarchical all-to-all parser.add_argument( "--hierarchical_alltoall", action="store_true", - help="Use hierarchical all-to-all for MoE. Recommended to enable for muiti-node training.", + help="Use hierarchical all-to-all for MoE. Recommended to enable for multi-node training.", ) args = parser.parse_args() @@ -238,7 +238,6 @@ def main(): lambda x, y: x.loss, optimizer, return_loss=True, - return_outputs=True, ) # Backward and optimize if is_pp_last_stage: @@ -268,7 +267,7 @@ def main(): # ): # coordinator.print_on_master(f"Apply load balance") # apply_load_balance(model, optimizer) - # save ckeckpoint + # save checkpoint if (step + 1) % args.save_interval == 0: coordinator.print_on_master(f"Saving model checkpoint to {args.output_path}") save_checkpoint( diff --git a/applications/ColossalQA/colossalqa/data_loader/document_loader.py b/applications/ColossalQA/colossalqa/data_loader/document_loader.py index cbcd6ad1d2d3..219c7ca41675 100644 --- a/applications/ColossalQA/colossalqa/data_loader/document_loader.py +++ b/applications/ColossalQA/colossalqa/data_loader/document_loader.py @@ -52,7 +52,7 @@ def __init__(self, files: List, **kwargs) -> None: def load_data(self, path: str) -> None: """ Load data. Please refer to https://python.langchain.com/docs/modules/data_connection/document_loaders/ - for sepcific format requirements. + for specific format requirements. Args: path: path to a file To load files with glob path, here are some examples. diff --git a/applications/ColossalQA/colossalqa/local/colossalcloud_llm.py b/applications/ColossalQA/colossalqa/local/colossalcloud_llm.py index 14e33820d9c9..3629778698fb 100644 --- a/applications/ColossalQA/colossalqa/local/colossalcloud_llm.py +++ b/applications/ColossalQA/colossalqa/local/colossalcloud_llm.py @@ -101,7 +101,7 @@ def _call(self, prompt: str, stop=None, **kwargs: Any) -> str: return resp_text def text_completion(self, prompt, gen_config, auth_config): - # Complusory Parameters + # Required Parameters endpoint = auth_config.pop("endpoint") max_new_tokens = gen_config.pop("max_new_tokens") # Optional Parameters diff --git a/applications/ColossalQA/colossalqa/local/llm.py b/applications/ColossalQA/colossalqa/local/llm.py index bab702d14b13..30a456c3d9c7 100644 --- a/applications/ColossalQA/colossalqa/local/llm.py +++ b/applications/ColossalQA/colossalqa/local/llm.py @@ -33,7 +33,7 @@ class ColossalAPI: def __init__(self, model_type: str, model_path: str, ckpt_path: str = None) -> None: """ - Configurate model + Configure model """ if model_type + model_path + (ckpt_path or "") in ColossalAPI.__instances: return @@ -47,7 +47,7 @@ def __init__(self, model_type: str, model_path: str, ckpt_path: str = None) -> N self.model.load_state_dict(state_dict) self.model.to(torch.cuda.current_device()) - # Configurate tokenizer + # Configure tokenizer self.tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True) self.model.eval() @@ -87,7 +87,7 @@ def generate(self, input: str, **kwargs) -> str: class VllmAPI: def __init__(self, host: str = "localhost", port: int = 8077) -> None: - # Configurate api for model served through web + # Configure api for model served through web self.host = host self.port = port self.url = f"http://{self.host}:{self.port}/generate" diff --git a/applications/ColossalQA/colossalqa/retrieval_conversation_universal.py b/applications/ColossalQA/colossalqa/retrieval_conversation_universal.py index b23058d6dbe3..6e77bb2aee17 100644 --- a/applications/ColossalQA/colossalqa/retrieval_conversation_universal.py +++ b/applications/ColossalQA/colossalqa/retrieval_conversation_universal.py @@ -36,7 +36,7 @@ def __init__( text_splitter_chunk_overlap=10, ) -> None: """ - Warpper for multilingual retrieval qa class (Chinese + English) + Wrapper for multilingual retrieval qa class (Chinese + English) Args: embedding_model_path: local or huggingface embedding model embedding_model_device: diff --git a/applications/ColossalQA/colossalqa/retriever.py b/applications/ColossalQA/colossalqa/retriever.py index 22a75050f03b..6a0c69859ac7 100644 --- a/applications/ColossalQA/colossalqa/retriever.py +++ b/applications/ColossalQA/colossalqa/retriever.py @@ -59,7 +59,7 @@ def add_documents( Add documents to retriever Args: docs: the documents to add - cleanup: choose from "incremental" (update embeddings, skip existing embeddings) and "full" (destory and rebuild retriever) + cleanup: choose from "incremental" (update embeddings, skip existing embeddings) and "full" (destroy and rebuild retriever) mode: choose from "by source" (documents are grouped by source) and "merge" (documents are merged into one vector store) """ if cleanup == "full": diff --git a/applications/ColossalQA/colossalqa/utils.py b/applications/ColossalQA/colossalqa/utils.py index cd8c3e5acec8..49d99014b372 100644 --- a/applications/ColossalQA/colossalqa/utils.py +++ b/applications/ColossalQA/colossalqa/utils.py @@ -49,7 +49,7 @@ def destroy_sql_database(sql_engine: Union[Engine, str]) -> None: def detect_lang_naive(s): """ - Naive function for language detection, should be replaced by an independant layer + Naive function for language detection, should be replaced by an independent layer """ remove_nota = "[’·°–!\"#$%&'()*+,-./:;<=>?@，。?★、…【】（）《》？“”‘’！[\\]^_`{|}~]+" s = re.sub(remove_nota, "", s) diff --git a/colossalai/_analyzer/fx/tracer/tracer.py b/colossalai/_analyzer/fx/tracer/tracer.py index 17dce767269d..36e8780af33a 100644 --- a/colossalai/_analyzer/fx/tracer/tracer.py +++ b/colossalai/_analyzer/fx/tracer/tracer.py @@ -237,7 +237,7 @@ def _tracer_override(self): # override the tracer to support custom modules and checkpointing if self.trace_act_ckpt: orig_ckpt_func_apply = torch.utils.checkpoint.CheckpointFunction.apply - orig_ckpt_func_without_reentrant = torch.utils.checkpoint._checkpoint_without_reentrant + orig_ckpt_func_without_reentrant = torch.utils.checkpoint._checkpoint_without_reentrant_generator def checkpoint(run_function, preserve_rng_state=False, *args): self.ckpt_regions.append(self.ckpt_idx) diff --git a/colossalai/booster/plugin/hybrid_parallel_plugin.py b/colossalai/booster/plugin/hybrid_parallel_plugin.py index c37a6b4df72d..29cec7cfd146 100644 --- a/colossalai/booster/plugin/hybrid_parallel_plugin.py +++ b/colossalai/booster/plugin/hybrid_parallel_plugin.py @@ -26,7 +26,7 @@ from colossalai.interface import AMPModelMixin, ModelWrapper, OptimizerWrapper from colossalai.pipeline.schedule import InterleavedSchedule, OneForwardOneBackwardSchedule from colossalai.pipeline.stage_manager import PipelineStageManager -from colossalai.shardformer import ShardConfig, ShardFormer +from colossalai.shardformer import GradientCheckpointConfig, ShardConfig, ShardFormer from colossalai.shardformer.layer.utils import SeqParallelUtils from colossalai.shardformer.policies.base_policy import Policy from colossalai.tensor.d_tensor.api import is_distributed_tensor @@ -34,7 +34,8 @@ from .pp_plugin_base import PipelinePluginBase -DP_AXIS, PP_AXIS, TP_AXIS = 0, 1, 2 +DP_AXIS, PP_AXIS, TP_AXIS, SP_AXIS = 0, 1, 2, 3 +SUPPORT_SP_MODE = ["split_gather", "ring", "all_to_all"] PRECISION_TORCH_TYPE = {"fp16": torch.float16, "fp32": torch.float32, "bf16": torch.bfloat16} @@ -53,6 +54,7 @@ def __init__( shard_config: ShardConfig, dp_group: ProcessGroup, tp_group: ProcessGroup, + sp_group: ProcessGroup, use_ddp: bool, ddp_config: dict, custom_policy: Policy, @@ -61,6 +63,7 @@ def __init__( self.shard_config = shard_config self.dp_group = dp_group self.tp_group = tp_group + self.sp_group = sp_group self.use_dpp = use_ddp self.require_grad_sync = True @@ -168,13 +171,24 @@ def sync_sp_grads(self, grads: Optional[List[torch.Tensor]] = None): Returns: None """ - if self.tp_group.size() > 1 and self.shard_config.enable_sequence_parallelism: + + if self.shard_config.enable_sequence_parallelism: + if self.shard_config.sequence_parallelism_mode == "all_to_all": + return + + if self.shard_config.sequence_parallelism_mode in ["split_gather", "ring"]: + # If sequence parallelism is enabled and mode is split_gather or ring, gradients are synchronized + # across the tensor parallelism group. + group = self.tp_group + else: + raise ValueError(f"Unknown sequence parallelism mode: {self.shard_config.sequence_parallelism_mode}") + if grads is not None: # Synchronize provided gradient tensors across the tensor parallelism group. - SeqParallelUtils.allreduce_partial_data_grad(tp_group=self.tp_group, grads=grads) + SeqParallelUtils.allreduce_partial_data_grad(process_group=group, grads=grads) else: # Synchronize gradients from the model across the tensor parallelism group. - SeqParallelUtils.allreduce_partial_data_grad(tp_group=self.tp_group, model=self.module) + SeqParallelUtils.allreduce_partial_data_grad(process_group=group, model=self.module) def forward(self, *args, **kwargs): if self.convert_fn is not None: @@ -727,10 +741,9 @@ def _get_grads_to_sync(all_working_grads) -> Union[List[Tensor], None]: # Get all working gradients and gradients to be synchronized. all_working_grads = _get_all_working_grads() grads_to_sync = _get_grads_to_sync(all_working_grads) - if self.require_grad_sync and grads_to_sync is not None: # Synchronize sequence parallelism gradients if required. - SeqParallelUtils.allreduce_partial_data_grad(tp_group=self.tp_pg, grads=grads_to_sync) + SeqParallelUtils.allreduce_partial_data_grad(process_group=self.tp_pg, grads=grads_to_sync) else: return @@ -891,6 +904,7 @@ class HybridParallelPlugin(PipelinePluginBase): Args: tp_size (int): The size of tensor parallelism. Tensor parallelism will not be used when tp_size is set to 1. pp_size (int): The number of pipeline stages in pipeline parallelism. Pipeline parallelism will not be used when pp_size is set to 1. + sp_size (int): The size of sequence parallelism. precision (str, optional): Specifies the precision of parameters during training. Auto-mixied precision will be used when this argument is set to 'fp16' or 'bf16', otherwise model is trained with 'fp32'. Defaults to 'fp16'. @@ -903,6 +917,7 @@ class HybridParallelPlugin(PipelinePluginBase): enable_flash_attention (bool, optional): Whether to switch on flash attention in Shardformer. Defaults to False. enable_jit_fused (bool, optional): Whether to switch on JIT in Shardformer. Default to False. enable_sequence_parallelism (bool): Whether to turn on sequence parallelism in Shardformer. Defaults to False. + sequence_parallelism_mode (str): The Sequence parallelism mode. Can only be choosed from ["split_gather", "ring", "all_to_all"]. Defaults to "split_gather". enable_sequence_overlap (bool): Whether to turn on sequence overlap in Shardformer. Defaults to False. parallel_output (bool): Whether to keep the output parallel when enabling tensor parallelism. Default to True. num_microbatches (int, optional): Number of microbatches when using pipeline parallelism. Defaults to None. @@ -930,6 +945,7 @@ class HybridParallelPlugin(PipelinePluginBase): custom_policy (Policy, optional): Custom policy for Shardformer. Defaults to None. pp_style (str, optional): The style for pipeline parallelism. Defaults to '1f1b'. num_model_chunks (int, optional): The number of model chunks for interleaved pipeline parallelism. Defaults to 1. + gradient_checkpoint_config (GradientCheckpointConfig, optional): Configuration for gradient checkpointing. Defaults to None. enable_metadata_cache (bool, optional): Whether to enable metadata cache for pipeline parallelism. Defaults to True. """ @@ -937,6 +953,7 @@ def __init__( self, tp_size: int, pp_size: int, + sp_size: int = None, precision: str = "fp16", zero_stage: int = 0, enable_all_optimization: bool = False, @@ -944,6 +961,7 @@ def __init__( enable_flash_attention: bool = False, enable_jit_fused: bool = False, enable_sequence_parallelism: bool = False, + sequence_parallelism_mode: str = None, enable_sequence_overlap: bool = False, parallel_output: bool = True, num_microbatches: Optional[int] = None, @@ -969,19 +987,47 @@ def __init__( custom_policy: Policy = None, pp_style: str = "1f1b", num_model_chunks: int = 1, + gradient_checkpoint_config: Optional[GradientCheckpointConfig] = None, enable_metadata_cache: bool = True, ) -> None: super().__init__() assert ( dist.get_world_size() % (tp_size * pp_size) == 0 - ), f"world size {dist.get_world_size()} is not divisible by tp_size {tp_size} * pp_size {pp_size}" + ), f"World size {dist.get_world_size()} is not divisible by tp_size {tp_size} * pp_size {pp_size}" if enable_sequence_parallelism: - assert tp_size > 1, "Sequence parallelism must be enabled when using tensor parallelism" + self.sequence_parallelism_mode = sequence_parallelism_mode if sequence_parallelism_mode is not None else "1" + assert ( + self.sequence_parallelism_mode in SUPPORT_SP_MODE + ), f"Sequence parallelism mode {self.sequence_parallelism_mode} is not in the supported list {SUPPORT_SP_MODE}" + if self.sequence_parallelism_mode in ["split_gather", "ring"]: + assert ( + tp_size > 1 + ), f"Sequence parallelism mode {self.sequence_parallelism_mode} must be enabled when using tensor parallelism" + if sp_size != 1: + warnings.warn( + f"The sp_size will be the same as tp_size in sequence parallelism mode {self.sequence_parallelism_mode}, will ignore the given sequence parallelism size." + ) + self.sp_size = 1 + self.dp_size = dist.get_world_size() // (tp_size * pp_size) + elif self.sequence_parallelism_mode in ["all_to_all"]: + assert ( + tp_size == 1 + ), f"Sequence parallelism mode {self.sequence_parallelism_mode} cannot be used with tensor parallelism" + assert ( + pp_size == 1 + ), f"Sequence parallelism mode {self.sequence_parallelism_mode} cannot be used with pipeline parallelism" + self.sp_size = dist.get_world_size() if sp_size is None else sp_size + self.dp_size = dist.get_world_size() // (self.sp_size * pp_size) + else: + self.dp_size = dist.get_world_size() // (tp_size * pp_size) + assert ( + sp_size == 1 or sp_size is None + ), f"sp_size can only be set to a >1 number when enable_sequence_parallelism is True" + self.sp_size = 1 self.tp_size = tp_size self.pp_size = pp_size - self.dp_size = dist.get_world_size() // (tp_size * pp_size) self.precision = precision self.zero_stage = zero_stage self.cpu_offload = cpu_offload @@ -990,7 +1036,7 @@ def __init__( self.enable_flash_attention = enable_flash_attention self.enable_jit_fused = enable_jit_fused self.enable_sequence_parallelism = enable_sequence_parallelism - self.pg_mesh = ProcessGroupMesh(self.dp_size, self.pp_size, self.tp_size) + self.pg_mesh = ProcessGroupMesh(self.dp_size, self.pp_size, self.tp_size, self.sp_size) self.stage_manager = None self.schedule = None self.custom_policy = custom_policy @@ -1031,9 +1077,14 @@ def __init__( self.tp_group = self.pg_mesh.get_group_along_axis(TP_AXIS) self.dp_group = self.pg_mesh.get_group_along_axis(DP_AXIS) self.pp_group = self.pg_mesh.get_group_along_axis(PP_AXIS) + if self.enable_sequence_parallelism and self.sequence_parallelism_mode in ["split_gather", "ring"]: + self.sp_group = self.pg_mesh.get_group_along_axis(TP_AXIS) + else: + self.sp_group = self.pg_mesh.get_group_along_axis(SP_AXIS) self.shard_config = ShardConfig( tensor_parallel_process_group=self.tp_group, + sequence_parallel_process_group=self.sp_group, pipeline_stage_manager=self.stage_manager, enable_tensor_parallelism=self.tp_size > 1, enable_all_optimization=self.enable_all_optimization, @@ -1041,8 +1092,10 @@ def __init__( enable_flash_attention=self.enable_flash_attention, enable_jit_fused=self.enable_jit_fused, enable_sequence_parallelism=enable_sequence_parallelism, + sequence_parallelism_mode=sequence_parallelism_mode, enable_sequence_overlap=enable_sequence_overlap, parallel_output=parallel_output, + gradient_checkpoint_config=gradient_checkpoint_config, ) self.amp_config = dict( initial_scale=initial_scale, @@ -1110,13 +1163,23 @@ def configure( ) -> Tuple[Module, OptimizerWrapper, Callable, DataLoader, LRScheduler]: param_info = get_param_info(optimizer) if not isinstance(model, ModelWrapper): - use_ddp = self.dp_size > 1 and self.pp_size == 1 and self.zero_stage == 0 + use_ddp = (self.dp_size > 1 and self.pp_size == 1 and self.zero_stage == 0) or ( + self.dp_size == 1 + and self.pp_size == 1 + and self.enable_sequence_parallelism + and self.sequence_parallelism_mode == "all_to_all" + ) + if self.enable_sequence_parallelism and self.sequence_parallelism_mode == "all_to_all": + dp_group = self.pg_mesh.create_group_along_axis([DP_AXIS, SP_AXIS]) + else: + dp_group = self.dp_group model = HybridParallelModule( model, precision=self.precision, shard_config=self.shard_config, - dp_group=self.dp_group, + dp_group=dp_group, tp_group=self.tp_group, + sp_group=self.sp_group, use_ddp=use_ddp, ddp_config=self.ddp_config, custom_policy=self.custom_policy, @@ -1146,7 +1209,8 @@ def configure( tp_process_group=self.tp_group, ) else: - if self.dp_size == 1: + zero_dp_size = dist.get_world_size(dp_group) + if zero_dp_size == 1: warnings.warn( "Use Zero Optimizer when data parallel size is 1 may introduce unnecessary overhead. " "If you are not intended to use cpu_offload, please consider set zero_stage=0." @@ -1158,7 +1222,7 @@ def configure( model, use_pipeline=self.enable_pipeline_parallelism, param_info=param_info, - dp_process_group=self.dp_group, + dp_process_group=dp_group, tp_process_group=self.tp_group, pp_process_group=self.pp_group, verbose=True, @@ -1183,6 +1247,9 @@ def execute_pipeline( ) -> dict: assert self.enable_pipeline_parallelism, "pipeline parallelism is not enabled" + if return_outputs: + warnings.warn("return_outputs may lead to significant extra memory consumption.") + # Create a context for gradient synchronization based on the optimizer type. # If it's a HybridParallelZeroOptimizer, use optimizer.no_sync(); otherwise, use model.no_sync(). # This is to avoid redundant gradient reduction in pipeline parallelism (multiple microbatch values should be reduced once), diff --git a/colossalai/booster/plugin/moe_hybrid_parallel_plugin.py b/colossalai/booster/plugin/moe_hybrid_parallel_plugin.py index ae372dd034e0..83888e5069a7 100644 --- a/colossalai/booster/plugin/moe_hybrid_parallel_plugin.py +++ b/colossalai/booster/plugin/moe_hybrid_parallel_plugin.py @@ -254,6 +254,9 @@ def __init__( self.tp_group = self.pg_mesh.get_group_along_axis(TP_AXIS) self.dp_group = self.pg_mesh.get_group_along_axis(DP_AXIS) self.pp_group = self.pg_mesh.get_group_along_axis(PP_AXIS) + # TODO: Currently moe only support partially sequence parallel + self.sp_group = self.pg_mesh.get_group_along_axis(TP_AXIS) + self.shard_config = ShardConfig( tensor_parallel_process_group=self.tp_group, pipeline_stage_manager=self.stage_manager, @@ -365,6 +368,7 @@ def configure( shard_config=self.shard_config, dp_group=self.dp_group, tp_group=self.tp_group, + sp_group=self.sp_group, use_ddp=use_ddp, ddp_config=self.ddp_config, custom_policy=self.custom_policy, diff --git a/colossalai/cluster/process_group_mesh.py b/colossalai/cluster/process_group_mesh.py index 1f32541a7b21..43095af50b8f 100644 --- a/colossalai/cluster/process_group_mesh.py +++ b/colossalai/cluster/process_group_mesh.py @@ -161,7 +161,7 @@ def get_ranks_in_group(self, group: ProcessGroup) -> List[int]: @staticmethod def get_coords_along_axis( - base_coord: Tuple[int, ...], axis: int, indices_at_axis: List[int] + base_coord: Tuple[int, ...], axis: Union[int, List[int]], indices_at_axis: Union[List[int], List[List[int]]] ) -> List[Tuple[int, ...]]: """Get coordinates along the given axis. @@ -173,13 +173,35 @@ def get_coords_along_axis( Returns: List[Tuple[int, ...]]: Coordinates along the axis. """ - coords_in_group = [] - for idx in indices_at_axis: - coords_in_group.append(base_coord[:axis] + (idx,) + base_coord[axis + 1 :]) + if isinstance(axis, int): + axis = [ + axis, + ] + assert isinstance(indices_at_axis[0], int) + indices_at_axis = [ + indices_at_axis, + ] + + def add_index(base_coord, axis, indices_at_axis): + coords_in_group = [] + for idx in indices_at_axis: + coords_in_group.append(base_coord[:axis] + (idx,) + base_coord[axis + 1 :]) + return coords_in_group + + coords_in_group = [base_coord] + for ax, indices_at_ax in zip(axis, indices_at_axis): + new_coords_in_group = [] + for coords in coords_in_group: + new_coords_in_group += add_index(coords, ax, indices_at_ax) + coords_in_group = new_coords_in_group + return coords_in_group def create_group_along_axis( - self, axis: int, indices_at_axis: Optional[List[int]] = None, backend: Optional[str] = None + self, + axis: Union[int, List[int]], + indices_at_axis: Optional[Union[List[int], List[List[int]]]] = None, + backend: Optional[str] = None, ) -> ProcessGroup: """Create all process groups along the given axis, and return the one which the current process belongs to. @@ -191,10 +213,21 @@ def create_group_along_axis( Returns: ProcessGroup: The process group along the given axis which the current process belongs to. """ - indices_at_axis = indices_at_axis or list(range(self._shape[axis])) + if isinstance(axis, int): + axis = [ + axis, + ] + if indices_at_axis is not None: + assert isinstance(indices_at_axis[0], int) + indices_at_axis = [ + indices_at_axis, + ] + + indices_at_axis = indices_at_axis or [list(range(self._shape[ax])) for ax in axis] reduced_shape = list(self._shape) # the choices on the axis are reduced to 1, since it's determined by `indices_at_axis` - reduced_shape[axis] = 1 + for ax in axis: + reduced_shape[ax] = 1 target_group = None # use Cartesian product to generate all combinations of coordinates for base_coord in itertools.product(*[range(s) for s in reduced_shape]): diff --git a/colossalai/inference/engine/policies/bloom.py b/colossalai/inference/engine/policies/bloom.py index f35b50189e82..5bc47c3c1a49 100644 --- a/colossalai/inference/engine/policies/bloom.py +++ b/colossalai/inference/engine/policies/bloom.py @@ -114,12 +114,12 @@ def get_held_layers(self) -> List[Module]: stage_manager = self.pipeline_stage_manager held_layers = [] - layers_per_stage = self.distribute_layers(len(module.h), stage_manager.num_stages) + layers_per_stage = stage_manager.distribute_layers(len(module.h)) if stage_manager.is_first_stage(): held_layers.append(module.word_embeddings) held_layers.append(module.word_embeddings_layernorm) held_layers.append(self.model.lm_head) - start_idx, end_idx = self.get_stage_index(layers_per_stage, stage_manager.stage) + start_idx, end_idx = stage_manager.get_stage_index(layers_per_stage) held_layers.extend(module.h[start_idx:end_idx]) if stage_manager.is_last_stage(): held_layers.append(module.ln_f) diff --git a/colossalai/inference/engine/policies/chatglm2.py b/colossalai/inference/engine/policies/chatglm2.py index 3e1d94f4785c..c7c6f3b927e1 100644 --- a/colossalai/inference/engine/policies/chatglm2.py +++ b/colossalai/inference/engine/policies/chatglm2.py @@ -69,11 +69,11 @@ def get_held_layers(self) -> List[nn.Module]: stage_manager = self.pipeline_stage_manager held_layers = [] - layers_per_stage = self.distribute_layers(module.num_layers, stage_manager.num_stages) + layers_per_stage = stage_manager.distribute_layers(module.num_layers) if stage_manager.is_first_stage(): held_layers.append(module.embedding) held_layers.append(module.output_layer) - start_idx, end_idx = self.get_stage_index(layers_per_stage, stage_manager.stage) + start_idx, end_idx = stage_manager.get_stage_index(layers_per_stage) held_layers.extend(module.encoder.layers[start_idx:end_idx]) if stage_manager.is_last_stage(): if module.encoder.post_layer_norm: diff --git a/colossalai/inference/engine/policies/llama.py b/colossalai/inference/engine/policies/llama.py index 11517d7e8a13..a57a4e50cdb9 100644 --- a/colossalai/inference/engine/policies/llama.py +++ b/colossalai/inference/engine/policies/llama.py @@ -194,11 +194,11 @@ def get_held_layers(self) -> List[Module]: stage_manager = self.pipeline_stage_manager held_layers = [] - layers_per_stage = self.distribute_layers(len(module.layers), stage_manager.num_stages) + layers_per_stage = stage_manager.distribute_layers(len(module.layers)) if stage_manager.is_first_stage(): held_layers.append(module.embed_tokens) held_layers.append(self.model.lm_head) - start_idx, end_idx = self.get_stage_index(layers_per_stage, stage_manager.stage) + start_idx, end_idx = stage_manager.get_stage_index(layers_per_stage) held_layers.extend(module.layers[start_idx:end_idx]) if stage_manager.is_last_stage(): held_layers.append(module.norm) diff --git a/colossalai/kernel/kernel_loader.py b/colossalai/kernel/kernel_loader.py index 148c3e3fc08a..353e29b3d122 100644 --- a/colossalai/kernel/kernel_loader.py +++ b/colossalai/kernel/kernel_loader.py @@ -6,7 +6,7 @@ CpuAdamX86Extension, FlashAttentionDaoCudaExtension, FlashAttentionNpuExtension, - FlashAttentionXformersCudaExtension, + FlashAttentionSdpaCudaExtension, FusedOptimizerCudaExtension, LayerNormCudaExtension, MoeCudaExtension, @@ -65,9 +65,9 @@ def load(self, ext_name: str = None): else: usable_exts = [] for ext in exts: - if ext.is_hardware_available(): + if ext.is_available(): # make sure the machine is compatible during kernel loading - ext.assert_hardware_compatible() + ext.assert_compatible() usable_exts.append(ext) assert len(usable_exts) != 0, f"No usable kernel found for {self.__class__.__name__} on the current machine." @@ -106,4 +106,20 @@ class ScaledUpperTriangleMaskedSoftmaxLoader(KernelLoader): class FlashAttentionLoader(KernelLoader): - REGISTRY = [FlashAttentionNpuExtension, FlashAttentionDaoCudaExtension, FlashAttentionXformersCudaExtension] + REGISTRY = [ + FlashAttentionNpuExtension, + FlashAttentionDaoCudaExtension, + FlashAttentionSdpaCudaExtension, + ] + + +class FlashAttentionWithPaddingMaskLoader(KernelLoader): + REGISTRY = [FlashAttentionNpuExtension, FlashAttentionDaoCudaExtension] + + +class FlashAttentionWithCustomMaskLoader(KernelLoader): + REGISTRY = [FlashAttentionNpuExtension, FlashAttentionSdpaCudaExtension] + + +class FlashAttentionForFloatAndCustomMaskLoader(KernelLoader): + REGISTRY = [FlashAttentionSdpaCudaExtension] diff --git a/colossalai/legacy/tensor/tensor_spec.py b/colossalai/legacy/tensor/tensor_spec.py index 5bdd384e5e15..44d8d04b9540 100644 --- a/colossalai/legacy/tensor/tensor_spec.py +++ b/colossalai/legacy/tensor/tensor_spec.py @@ -1,4 +1,4 @@ -from dataclasses import dataclass +from dataclasses import dataclass, field from typing import Optional from colossalai.legacy.tensor.distspec import DistPlacementPattern, _DistSpec @@ -17,5 +17,5 @@ class ColoTensorSpec: """ pg: ProcessGroup - dist_attr: Optional[_DistSpec] = _DistSpec(DistPlacementPattern.REPLICATE) + dist_attr: Optional[_DistSpec] = field(default_factory=lambda: _DistSpec(DistPlacementPattern.REPLICATE)) compute_attr: Optional[ComputeSpec] = None diff --git a/colossalai/nn/layer/colo_attention.py b/colossalai/nn/layer/colo_attention.py deleted file mode 100644 index 0b7011e8e2d8..000000000000 --- a/colossalai/nn/layer/colo_attention.py +++ /dev/null @@ -1,209 +0,0 @@ -import enum -import math -import warnings -from dataclasses import dataclass -from typing import Iterable, Optional, Tuple - -import torch -import torch.nn.functional as F -from einops import rearrange - -from colossalai.accelerator import get_accelerator -from colossalai.kernel.kernel_loader import FlashAttentionLoader - - -@dataclass -class SeqLenInfo: - seqlens: Iterable[int] = None - indices: torch.Tensor = None - max_seqlen: int = None - cu_seqlens: torch.Tensor = None - - @staticmethod - def materialize( - attn_mask: torch.Tensor = None, size: Tuple[int] = None, device=get_accelerator().get_current_device() - ): - if attn_mask is not None: - indices = torch.nonzero(attn_mask.flatten(), as_tuple=False).flatten().to(device) - seqlens = attn_mask.sum(dim=-1, dtype=torch.int32).flatten() - else: - batch_size, tgt_len = size[0], size[1] - indices = torch.arange(batch_size * tgt_len, dtype=torch.long, device=device) - seqlens = torch.LongTensor([tgt_len] * batch_size, device=device) - max_seqlen = max(seqlens) - cu_seqlens = F.pad(torch.cumsum(seqlens, dim=0, dtype=torch.int32), (1, 0)).to(device) - return SeqLenInfo(seqlens.tolist(), indices, max_seqlen, cu_seqlens) - - -class AttnMaskType(enum.Enum): - padding = 1 - causal = 2 - paddedcausal = 3 - - -class Unpad(torch.autograd.Function): - """ - Adapted from - https://github.com/HazyResearch/flash-attention/blob/main/flash_attn/bert_padding.py - """ - - @staticmethod - def forward(ctx, tensor: torch.Tensor, indices: torch.Tensor): - ctx.save_for_backward(indices) - # [b, s, ...] - assert tensor.ndim >= 3 - ctx.bsz = tensor.shape[0] - out = rearrange(tensor, "b s ... -> (b s) ...") - ctx.shape = out.shape - # [ntokens, ...] - return out[indices] - - @staticmethod - def backward(ctx, grad_output): - (indices,) = ctx.saved_tensors - # [ntokens, ...] - grad = torch.zeros(ctx.shape, dtype=grad_output.dtype, device=grad_output.device) - grad[indices] = grad_output - grad = rearrange(grad, "(b s) ... -> b s ...", b=ctx.bsz) - # [b, s, ...] - return grad, None - - -class Repad(torch.autograd.Function): - """ - Adapted from - https://github.com/HazyResearch/flash-attention/blob/main/flash_attn/bert_padding.py - """ - - @staticmethod - def forward(ctx, tensor: torch.Tensor, indices: torch.Tensor, batch_size: int, seq_len: int): - ctx.save_for_backward(indices) - # [ntokens, ...] - tensor = tensor - out = torch.zeros((batch_size * seq_len, *tensor.shape[1:]), dtype=tensor.dtype, device=tensor.device) - # [b*s, ...] - out[indices] = tensor - return out - - @staticmethod - def backward(ctx, grad_output): - (indices,) = ctx.saved_tensors - # [b*s, ...] - grad = grad_output[indices] - # [ntokens, ...] - return grad, None, None, None - - -class ColoAttention(torch.nn.Module): - def __init__(self, embed_dim: int, num_heads: int, dropout: float = 0.0, scale=None): - super().__init__() - assert ( - embed_dim % num_heads == 0 - ), f"the embed dim ({embed_dim}) is not divisible by the number of attention heads ({num_heads})." - if scale is not None: - self.scale = scale - else: - self.scale = 1 / math.sqrt(embed_dim // num_heads) - self.dropout = dropout - - self.attn = FlashAttentionLoader().load() - - @staticmethod - def unpad(tensor: torch.Tensor, indices: torch.Tensor) -> torch.Tensor: - return Unpad.apply(tensor, indices) - - @staticmethod - def repad(tensor: torch.Tensor, indices: torch.Tensor, batch_size: int, seq_len: int) -> torch.Tensor: - return Repad.apply(tensor, indices, batch_size, seq_len) - - def forward( - self, - query: torch.Tensor, - key: torch.Tensor, - value: torch.Tensor, - attn_mask: Optional[torch.Tensor] = None, - origin_attn_mask: Optional[torch.Tensor] = None, - attn_mask_type: Optional[AttnMaskType] = None, - bias: Optional[torch.Tensor] = None, - ): - """ - ColoAttention - - Args: - q: (batch, q_seqlen, nheads, headdim) - k: (batch, kv_seqlen, nheads, headdim) - v: (batch, kv_seqlen, nheads, headdim) - origin_attn_mask: (nheads, q_seqlen, kv_seqlen) - bias: will not be used - Return: - attn_out: (batch, q_seqlen, nheads, headdim). - """ - # if flash attention is not applicable, switch to memory effcient attention - if self.attn.__name__ == "flash_attention" and ( - query.dtype not in [torch.float16, torch.bfloat16] or bias != None - ): - warnings.warn( - f"flash-attn expects fp16 or bf16 but got {query.dtype}, switching to xformers' implementation." - ) - self.attn = FlashAttentionLoader().load(ext_name="flash_attention_xformers_cuda") - - padded = attn_mask_type is not None and attn_mask_type.value % 2 == 1 - causal = attn_mask_type is not None and attn_mask_type.value > 1 - - batch_size, tgt_len, src_len = query.shape[0], query.shape[1], key.shape[1] - # unpad - seq_len_info_q = None - seq_len_info_kv = None - if padded: - # bert style, unpad process - assert ( - attn_mask is not None - ), f"attention mask {attn_mask} is not valid for attention mask type {attn_mask_type}." - assert attn_mask.dim() == 2, ( - "attention mask is supposed to have shape (batch_size, seq_len), " - + f"but got {attn_mask.dim()} dimensions." - ) - - # bert style - if tgt_len == src_len: - seq_len_info_q = SeqLenInfo.materialize(attn_mask=attn_mask, device=query.device) - if batch_size > 1: - query, key, value = self.unpad( - torch.stack([query, key, value], dim=2), seq_len_info_q.indices - ).unbind(dim=1) - else: - query, key, value = torch.stack([query, key, value], dim=2).squeeze(0).unbind(dim=1) - seq_len_info_kv = seq_len_info_q - else: - seq_len_info_q = SeqLenInfo.materialize(size=(batch_size, tgt_len), device=query.device) - seq_len_info_kv = SeqLenInfo.materialize(attn_mask=attn_mask, device=query.device) - if batch_size > 1: - query = rearrange(query, "b s ... -> c (b s) ...", c=1) - key, value = self.unpad(torch.stack([query, key, value], dim=2), seq_len_info_kv.indices).unbind( - dim=1 - ) - else: - query, key, value = torch.stack([query, key, value], dim=2).squeeze(0).unbind(dim=1) - - out = self.attn( - query, - key, - value, - seq_len_info_q=seq_len_info_q, - seq_len_info_kv=seq_len_info_kv, - origin_attn_mask=origin_attn_mask, - dropout_p=self.dropout, - scale=self.scale, - causal=causal, - padded=padded, - ) - - # repad - if padded: - if batch_size > 1: - out = self.repad(out, seq_len_info_q.indices, batch_size, tgt_len) - out = rearrange(out, "(b s) h d -> b s h d", b=batch_size) - - if len(out.shape) == 4: - out = rearrange(out, "b s h d -> b s (h d)") - return out diff --git a/colossalai/nn/layer/scaled_softmax.py b/colossalai/nn/layer/scaled_softmax.py index a8d72ddd90c9..2e802db2dbca 100644 --- a/colossalai/nn/layer/scaled_softmax.py +++ b/colossalai/nn/layer/scaled_softmax.py @@ -8,6 +8,14 @@ from colossalai.kernel.kernel_loader import ScaledMaskedSoftmaxLoader, ScaledUpperTriangleMaskedSoftmaxLoader +# NOTE: These kernels are compiled on specific GPU arch and not widely applicable. +# try: +# from colossalai._C import scaled_masked_softmax as scaled_masked_softmax, scaled_upper_triangle_masked_softmax_cuda as scaled_upper_triang_masked_softmax +# except ImportError: + +scaled_masked_softmax = None +scaled_upper_triang_masked_softmax = None + class AttnMaskType(enum.Enum): padding = 1 diff --git a/colossalai/pipeline/schedule/one_f_one_b.py b/colossalai/pipeline/schedule/one_f_one_b.py index bf2f01b10e9b..58008b98f24e 100644 --- a/colossalai/pipeline/schedule/one_f_one_b.py +++ b/colossalai/pipeline/schedule/one_f_one_b.py @@ -7,7 +7,7 @@ from torch.utils._pytree import tree_map from colossalai.accelerator import get_accelerator -from colossalai.interface import ModelWrapper, OptimizerWrapper +from colossalai.interface import OptimizerWrapper from colossalai.pipeline.p2p import PipelineP2PCommunication, create_send_metadata from colossalai.pipeline.stage_manager import PipelineStageManager from colossalai.utils import get_current_device @@ -327,9 +327,7 @@ def run_forward_only( self.send_forward(output_obj) if outputs is not None: - if isinstance(model, ModelWrapper): - model = model.unwrap() - outputs = merge_batch(outputs, getattr(model, "batch_size_dim", 0)) + outputs = merge_batch(outputs) return {"loss": accum_loss, "outputs": outputs} def run_forward_backward( @@ -412,9 +410,7 @@ def run_forward_backward( assert all(len(v) == 0 for v in input_objs) and all(len(v) == 0 for v in output_objs) if outputs is not None: - if isinstance(model, ModelWrapper): - model = model.unwrap() - outputs = merge_batch(outputs, getattr(model, "batch_size_dim", 0)) + outputs = merge_batch(outputs) return {"loss": accum_loss, "outputs": outputs} def forward_backward_step( diff --git a/colossalai/pipeline/stage_manager.py b/colossalai/pipeline/stage_manager.py index c8f9042084da..b0556669b2bc 100644 --- a/colossalai/pipeline/stage_manager.py +++ b/colossalai/pipeline/stage_manager.py @@ -1,6 +1,7 @@ import contextlib -from typing import Dict, List, Optional, Tuple +from typing import Dict, List, Optional, Tuple, Union +import numpy as np import torch.distributed as dist from torch.distributed import ProcessGroup @@ -29,6 +30,8 @@ def __init__( ) -> None: assert enable_interleave or num_model_chunks == 1, "num_model_chunks must be 1 when enable_interleave is False" + self.num_layers_per_stage = None + self.pg_mesh = pg_mesh self.pipeline_axis = pipeline_axis self.prev_rank: Optional[Tuple[int, ...]] = None @@ -69,6 +72,88 @@ def __init__( # for shardformer, hold model chunk id self.model_chunk_id: Optional[int] = None + @property + def control_distribute_layers(self) -> bool: + return self.num_layers_per_stage is not None + + def set_distribution_config(self, num_model_layers: int, num_layers_per_stage: List[int]) -> None: + """Set the distribution configuration. + This allows user to customize the number of layers for each stage. + + Args: + num_model_layers (int): Number of layers in the model. + num_layers_per_stage (List[int]): Number of layers for each stage. + """ + assert all([0 < num_layers < num_model_layers for num_layers in num_layers_per_stage]) + assert sum(num_layers_per_stage) == num_model_layers + assert len(num_layers_per_stage) == self.num_stages * (self.num_model_chunks if self.is_interleave else 1) + self.num_model_layers = num_model_layers + self.num_layers_per_stage = num_layers_per_stage + + def distribute_layers( + self, num_layers: int, num_stages: Optional[int] = None, num_model_chunks: Optional[int] = None + ) -> List[int]: + """Divide layers into stages""" + num_stages = self.num_stages if num_stages is None else num_stages + num_model_chunks = ( + (self.num_model_chunks if self.is_interleave else 1) if num_model_chunks is None else num_model_chunks + ) + + if self.control_distribute_layers: + assert num_layers == self.num_model_layers + return self.num_layers_per_stage + + else: + quotient = num_layers // (num_stages * num_model_chunks) + remainder = num_layers % (num_stages * num_model_chunks) + + # calculate the num_layers per stage + layers_per_stage = [quotient] * num_stages * num_model_chunks + + # deal with the rest layers + if remainder > 0: + start_position = (num_stages * num_model_chunks) // 2 - remainder // 2 + for i in range(start_position, start_position + remainder): + layers_per_stage[i] += 1 + return layers_per_stage + + def get_stage_index( + self, + layers_per_stage: List[int], + stage: Optional[int] = None, + num_model_chunks: Optional[int] = None, + num_stages: Optional[int] = None, + ) -> Union[Tuple[int, int], List[Tuple[int, int]]]: + """ + Get the start index and end index of layers for each stage. + + Args: + layers_per_stage (List[int]): number of layers for each stage + stage (int): the stage index + num_stages (int): number of stages + num_model_chunks (int): number of model chunks + + Returns: + - Tuple[int, int]: the start index and end index of this stage + - List[Tuple[int, int]]: the start index and end index of this stage for each model chunk + + """ + stage = self.stage if stage is None else stage + num_model_chunks = ( + (self.num_model_chunks if self.is_interleave else 1) if num_model_chunks is None else num_model_chunks + ) + num_stages = self.num_stages if num_stages is None else num_stages + + num_layers_per_stage_accumulated = np.insert(np.cumsum(layers_per_stage), 0, 0) + + stage_indices = [] + for model_chunk in range(num_model_chunks): + start_idx = num_layers_per_stage_accumulated[stage + model_chunk * num_stages] + end_idx = num_layers_per_stage_accumulated[stage + model_chunk * num_stages + 1] + stage_indices.append([start_idx, end_idx]) + + return stage_indices[0] if num_model_chunks == 1 else stage_indices + def is_first_stage(self, ignore_chunk: bool = False) -> bool: """Is the current stage the first stage. diff --git a/colossalai/shardformer/__init__.py b/colossalai/shardformer/__init__.py index 77c2af8d18f7..234e7131728f 100644 --- a/colossalai/shardformer/__init__.py +++ b/colossalai/shardformer/__init__.py @@ -1 +1 @@ -from .shard import ShardConfig, ShardFormer +from .shard import GradientCheckpointConfig, ModelSharder, PipelineGradientCheckpointConfig, ShardConfig, ShardFormer diff --git a/colossalai/shardformer/layer/__init__.py b/colossalai/shardformer/layer/__init__.py index 56e8b08c4e4a..7b8aa53800f0 100644 --- a/colossalai/shardformer/layer/__init__.py +++ b/colossalai/shardformer/layer/__init__.py @@ -1,3 +1,5 @@ +from ._operation import all_to_all_comm +from .attn import AttnMaskType, ColoAttention from .dropout import DropoutForParallelInput, DropoutForReplicatedInput from .embedding import Embedding1D, VocabParallelEmbedding1D from .linear import Linear1D_Col, Linear1D_Row @@ -23,4 +25,7 @@ "FusedRMSNorm", "FusedLinear1D_Col", "ParallelModule", + "AttnMaskType", + "ColoAttention", + "all_to_all_comm", ] diff --git a/colossalai/shardformer/layer/_operation.py b/colossalai/shardformer/layer/_operation.py index 241770901ed7..82d37bb4cf94 100644 --- a/colossalai/shardformer/layer/_operation.py +++ b/colossalai/shardformer/layer/_operation.py @@ -167,6 +167,97 @@ def backward(ctx, grad_output): return grad_input, grad_weight, grad_bias, None, None, None +def _ring_as_gather(func, input_to_gather=None, input_local=None, process_group=None, gather_dim=1, keep_item=False): + # currently only support one single tensor as output + group_size = dist.get_world_size(process_group) + cur_rank = dist.get_rank(process_group) + + # output_tensors = [torch.empty((input_shape[0], input_shape[1], weight_shape[0])) for _ in range(group_size)] + + # initialization of ring communication + recv_rank = cur_rank + 1 if cur_rank + 1 < group_size else 0 + send_rank = cur_rank - 1 if cur_rank > 0 else group_size - 1 + rank_map = list(dist.get_process_group_ranks(process_group)) + recv_rank = rank_map[recv_rank] + send_rank = rank_map[send_rank] + recv_tensors = {} + send_tensors = {} + for k, v in input_to_gather.items(): + recv_tensors[k] = torch.empty_like(v) + send_tensors[k] = v.clone() + + def communicate_step(): + comm_ops = [] + for k in recv_tensors: + comm_ops.append(dist.P2POp(dist.irecv, recv_tensors[k], recv_rank, group=process_group)) + comm_ops.append(dist.P2POp(dist.isend, send_tensors[k], send_rank, group=process_group)) + return dist.batch_isend_irecv(comm_ops) + + def switch_step(): + for k in recv_tensors: + send_tensors[k], recv_tensors[k] = recv_tensors[k], send_tensors[k] + + output_tensors = [] + + handles = communicate_step() + # first round: special case, retrive from local tensor + output_tensors.append(func(**input_to_gather, **input_local)) + for i in range(group_size - 2): + for handle in handles: + handle.wait() + + switch_step() + + handles = communicate_step() + + # actual computation + output_tensors.append(func(**send_tensors, **input_local)) + + # final round: special case, no need to send/recv again + for handle in handles: + handle.wait() + output_tensors.append(func(**recv_tensors, **input_local)) + + return torch.cat(output_tensors[group_size - cur_rank :] + output_tensors[: group_size - cur_rank], dim=gather_dim) + + +class _GatherForwardReduceScatterBackward(torch.autograd.Function): + """Gather input from sequence parallel in forward and reduce-scatter gradient in backward + + Args: + input_ (`torch.Tensor`): The input tensor from sequence parallel region. + process_group (`torch.distributed.ProcessGroup`): The process group used for collective communication. + overlap (`bool`): Whther to overlap the all_gather op and gradient calculate in backward. + + """ + + @staticmethod + def forward(ctx, input_, process_group, dim): + ctx.process_group = process_group + ctx.dim = dim + + return _gather(input_, dim, process_group) + + @staticmethod + def backward(ctx, grad_output): + dim = ctx.dim + process_group = ctx.process_group + + # do reduce-scatter + new_shape = list(grad_output.shape) + assert ( + new_shape[dim] % dist.get_world_size(process_group) == 0 + ), f"The dimension to split ({new_shape[dim]}) is not a multiple of tensor parallel size ({dist.get_world_size(process_group)}). " + new_shape[dim] = new_shape[dim] // dist.get_world_size(process_group) + grad_list = [ + item.contiguous() for item in torch.chunk(grad_output, dist.get_world_size(process_group), dim=dim) + ] + output = torch.empty(new_shape, dtype=grad_output.dtype, device=grad_output.device) + dist.reduce_scatter(output, grad_list, group=process_group) + + return output, None, None + + class _LinearWithGatherForwardReduceScatterBackward(torch.autograd.Function): """Gather input from sequence parallel in forward and reduce-scatter gradient in backward @@ -178,7 +269,7 @@ class _LinearWithGatherForwardReduceScatterBackward(torch.autograd.Function): """ @staticmethod - def forward(ctx, input_, weight, bias, process_group, async_grad_reduce_scatter, dim, overlap=True): + def forward(ctx, input_, weight, bias, process_group, async_grad_reduce_scatter, dim, overlap=True, ring=False): ctx.save_for_backward(input_, weight, bias) ctx.use_bias = bias is not None ctx.process_group = process_group @@ -186,12 +277,25 @@ def forward(ctx, input_, weight, bias, process_group, async_grad_reduce_scatter, ctx.dim = dim ctx.overlap = overlap - input_parallel = _gather(input_, dim, process_group) + if ring is True: + input_to_gather = {"input": input_} + input_local = {"weight": weight} - if bias is not None: - output = F.linear(input_parallel, weight, bias) + output = _ring_as_gather( + F.linear, + input_to_gather=input_to_gather, + input_local=input_local, + process_group=process_group, + ) + + if bias is not None: + output += bias else: - output = F.linear(input_parallel, weight) + input_parallel = _gather(input_, dim, process_group) + if bias is not None: + output = F.linear(input_parallel, weight, bias) + else: + output = F.linear(input_parallel, weight) return output @@ -294,11 +398,146 @@ def backward(ctx, grad_output): # wait until reduce-scatter finished reducescatter_handle.wait() - return output, grad_weight, grad_bias, None, None, None, None + return output, grad_weight, grad_bias, None, None, None, None, None + + +def _ring_as_reducescatter( + func, input_to_reducescatter=None, input_local=None, process_group=None, reducescatter_dim=1 +): + # currently only support one single tensor as output + group_size = dist.get_world_size(process_group) + cur_rank = dist.get_rank(process_group) + + # initialization of ring communication + recv_rank = cur_rank - 1 if cur_rank > 0 else group_size - 1 + send_rank = cur_rank + 1 if cur_rank + 1 < group_size else 0 + rank_map = list(dist.get_process_group_ranks(process_group)) + recv_rank = rank_map[recv_rank] + send_rank = rank_map[send_rank] + input_tensors = [] + for _ in range(group_size): + input_tensors.append({}) + for k, v in input_to_reducescatter.items(): + input_shape = v.shape + assert input_shape[reducescatter_dim] % group_size == 0 + _input_tensors = list(torch.split(v, input_shape[reducescatter_dim] // group_size, dim=reducescatter_dim)) + for i in range(group_size): + input_tensors[i][k] = _input_tensors[i] + input_tensors = input_tensors[cur_rank:] + input_tensors[:cur_rank] + input_tensors.reverse() + + output_tensor = func(**input_tensors[0], **input_local) + recv_tensor = torch.empty_like(output_tensor) + send_tensor = output_tensor.clone() + + def communicate_step(): + recv_op = dist.P2POp(dist.irecv, recv_tensor, recv_rank, group=process_group) + send_op = dist.P2POp(dist.isend, send_tensor, send_rank, group=process_group) + return dist.batch_isend_irecv([recv_op, send_op]) + + handles = communicate_step() + # first round: special case, retrive from local tensor + for i in range(group_size - 2): + # actual computation + output_tensor = func(**input_tensors[i + 1], **input_local) + + for handle in handles: + handle.wait() + output_tensor += recv_tensor + + tmp_tensor = send_tensor + send_tensor = output_tensor + output_tensor = tmp_tensor + + handles = communicate_step() + + # final round: special case, no need to send/recv again + output_tensor = func(**input_tensors[-1], **input_local) + for handle in handles: + handle.wait() + output_tensor += recv_tensor + return output_tensor class _LinearWithReduceScatterForwardGatherBackward(torch.autograd.Function): - """Gather input from sequence parallel in forward and reduce-scatter gradient in backward + """Reduce-scatter input from sequence parallel in forward and gather gradient in backward with ring + + Args: + input_ (`torch.Tensor`): The input tensor from sequence parallel region. + process_group (`torch.distributed.ProcessGroup`): The process group used for collective communication. + overlap (`bool`): Whther to overlap the all_gather op and gradient calculate in backward. + + """ + + @staticmethod + def forward(ctx, input_, weight, bias, process_group, dim, ring): + ctx.save_for_backward(input_, weight, bias) + ctx.use_bias = bias is not None + ctx.process_group = process_group + ctx.dim = dim + + if ring is True: + input_to_reducescatter = {"input": input_} + input_local = {"weight": weight} + + if bias is not None: + input_to_reducescatter["bias"] = bias + + output = _ring_as_reducescatter( + F.linear, + input_to_reducescatter=input_to_reducescatter, + input_local=input_local, + process_group=process_group, + ) + else: + if bias is not None: + partial_output = F.linear(input_, weight, bias) + else: + partial_output = F.linear(input_, weight) + + output_shape = list(partial_output.shape) + assert ( + output_shape[dim] % dist.get_world_size(process_group) == 0 + ), f"The dimension to split ({output_shape[dim]}) is not a multiple of tensor parallel size ({dist.get_world_size(process_group)}). " + output_shape[dim] = output_shape[dim] // dist.get_world_size(process_group) + + output_list = [ + item.contiguous() for item in torch.chunk(partial_output, dist.get_world_size(process_group), dim=dim) + ] + output = torch.empty(output_shape, dtype=partial_output.dtype, device=partial_output.device).contiguous() + dist.reduce_scatter(output, output_list, group=process_group) + + return output + + @staticmethod + def backward(ctx, grad_output): + input_, weight, bias = ctx.saved_tensors + use_bias = ctx.use_bias + dim = ctx.dim + process_group = ctx.process_group + + # In order to be hooked into Gemini's '__torch_function__', adding a view operation to weight and bias. Used in FusedLayerNorm + if use_bias: + bias = bias.view(bias.shape) + + grad_output = _gather(grad_output, dim, process_group) + + # TODO Need to fully optimize + total_input = input_ + grad_input = grad_output.matmul(weight) + grad_output = grad_output.contiguous() + # Convert the tensor shapes to 2D for execution compatibility + if len(grad_output.shape) > 2: + grad_output = grad_output.view(-1, grad_output.shape[-1]) + total_input = total_input.view(-1, total_input.shape[-1]) + grad_weight = grad_output.t().matmul(total_input) + grad_bias = grad_output.sum(dim=0) if use_bias else None + + return grad_input, grad_weight, grad_bias, None, None, None + + +class _ReduceScatterForwardGatherBackward(torch.autograd.Function): + """Reduce-scatter input from sequence parallel in forward and gather gradient in backward Args: input_ (`torch.Tensor`): The input tensor from sequence parallel region. @@ -343,7 +582,7 @@ class _MatmulWithGatherForwardReduceScatterBackward(torch.autograd.Function): """ @staticmethod - def forward(ctx, input_, weight, bias, process_group, async_grad_reduce_scatter, dim, overlap): + def forward(ctx, input_, weight, bias, process_group, async_grad_reduce_scatter, dim, overlap, ring): ctx.save_for_backward(input_, weight, bias) ctx.use_bias = bias is not None ctx.process_group = process_group @@ -351,9 +590,24 @@ def forward(ctx, input_, weight, bias, process_group, async_grad_reduce_scatter, ctx.dim = dim ctx.overlap = overlap - input_parallel = _gather(input_, dim, process_group) + if ring is True: + input_to_gather = {} + input_local = {} + input_to_gather["input"] = input_ + input_local["other"] = weight - output = torch.matmul(input_parallel, weight) + output = _ring_as_gather( + torch.matmul, + input_to_gather=input_to_gather, + input_local=input_local, + process_group=process_group, + gather_dim=dim, + ) + + else: + input_parallel = _gather(input_, dim, process_group) + + output = torch.matmul(input_parallel, weight) if bias is not None: output = output + bias @@ -433,7 +687,7 @@ def backward(ctx, grad_output): # wait until reduce-scatter finished reducescatter_handle.wait() - return output, grad_weight, grad_bias, None, None, None, None + return output, grad_weight, grad_bias, None, None, None, None, None class _SplitForwardGatherBackward(torch.autograd.Function): @@ -448,14 +702,17 @@ class _SplitForwardGatherBackward(torch.autograd.Function): """ @staticmethod - def forward(ctx, input_, dim, process_group): + def forward(ctx, input_, dim, process_group, grad_scale=None): ctx.process_group = process_group ctx.dim = dim + ctx.grad_scale = grad_scale return _split(input_, dim, process_group) @staticmethod def backward(ctx, grad_output): - return _gather(grad_output, ctx.dim, ctx.process_group), None, None + if ctx.grad_scale is not None: + grad_output = grad_output * ctx.grad_scale + return _gather(grad_output, ctx.dim, ctx.process_group), None, None, None class _ReduceForward(torch.autograd.Function): @@ -505,14 +762,50 @@ class _GatherForwardSplitBackward(torch.autograd.Function): """ @staticmethod - def forward(ctx, input_, dim, process_group): + def forward(ctx, input_, dim, process_group, grad_scale=None): ctx.process_group = process_group ctx.dim = dim + ctx.grad_scale = grad_scale return _gather(input_, dim, process_group) @staticmethod def backward(ctx, grad_output): - return _split(grad_output, ctx.dim, ctx.process_group), None, None + if ctx.grad_scale is not None: + grad_output = grad_output * ctx.grad_scale + return _split(grad_output, ctx.dim, ctx.process_group), None, None, None + + +class _AllToAll(torch.autograd.Function): + """All-to-all communication. + + Args: + input_: input matrix + process_group: communication group + scatter_dim: scatter dimension + gather_dim: gather dimension + """ + + @staticmethod + def forward(ctx, input_, process_group, scatter_dim, gather_dim): + ctx.process_group = process_group + ctx.scatter_dim = scatter_dim + ctx.gather_dim = gather_dim + world_size = dist.get_world_size(process_group) + bsz, _, _ = input_.shape + + # using all_to_all_single when batch size is 1 + if bsz == 1: + return _all_to_all_single(input_, world_size, process_group, scatter_dim, gather_dim) + else: + return _all_to_all(input_, world_size, process_group, scatter_dim, gather_dim) + + @staticmethod + def backward(ctx, *grad_output): + process_group = ctx.process_group + scatter_dim = ctx.gather_dim + gather_dim = ctx.scatter_dim + return_grad = _AllToAll.apply(*grad_output, process_group, scatter_dim, gather_dim) + return (return_grad, None, None, None) class HookParameter(torch.autograd.Function): @@ -608,6 +901,40 @@ def _reduce_scatter(input_, dim=1, process_group=None): return output +def _all_to_all(input_, world_size, group, scatter_dim, gather_dim): + input_list = [t.contiguous() for t in torch.tensor_split(input_, world_size, scatter_dim)] + output_list = [torch.empty_like(input_list[0]) for _ in range(world_size)] + dist.all_to_all(output_list, input_list, group=group) + return torch.cat(output_list, dim=gather_dim).contiguous() + + +def _all_to_all_single(input_, seq_world_size, group, scatter_dim, gather_dim): + inp_shape = list(input_.shape) + inp_shape[scatter_dim] = inp_shape[scatter_dim] // seq_world_size + if scatter_dim < 2: + input_t = input_.reshape([seq_world_size, inp_shape[scatter_dim]] + inp_shape[scatter_dim + 1 :]).contiguous() + else: + input_t = ( + input_.reshape([-1, seq_world_size, inp_shape[scatter_dim]] + inp_shape[scatter_dim + 1 :]) + .transpose(0, 1) + .contiguous() + ) + + output = torch.empty_like(input_t) + dist.all_to_all_single(output, input_t, group=group) + + if scatter_dim < 2: + output = output.transpose(0, 1).contiguous() + + return output.reshape( + inp_shape[:gather_dim] + + [ + inp_shape[gather_dim] * seq_world_size, + ] + + inp_shape[gather_dim + 1 :] + ).contiguous() + + def matmul_with_async_comm(input_, weight, bias, process_group, async_grad_allreduce): return MatmulWithAsyncCommunication.apply(input_, weight, bias, process_group, async_grad_allreduce) @@ -617,31 +944,39 @@ def linear_with_async_comm(input_, weight, bias, process_group, async_grad_allre def linear_gather_forward_reducescatter_backward( - input_, weight, bias, process_group, async_grad_reduce_scatter, dim, overlap + input_, weight, bias, process_group, async_grad_reduce_scatter, dim, overlap, ring=False ): return _LinearWithGatherForwardReduceScatterBackward.apply( - input_, weight, bias, process_group, async_grad_reduce_scatter, dim, overlap + input_, weight, bias, process_group, async_grad_reduce_scatter, dim, overlap, ring ) -def linear_reducescatter_forward_gather_backward(input_, process_group, dim): - return _LinearWithReduceScatterForwardGatherBackward.apply(input_, process_group, dim) +def gather_forward_reducescatter_backward(input_, process_group, dim): + return _GatherForwardReduceScatterBackward.apply(input_, process_group, dim) + + +def reducescatter_forward_gather_backward(input_, process_group, dim): + return _ReduceScatterForwardGatherBackward.apply(input_, process_group, dim) + + +def linear_reducescatter_forward_gather_backward(input_, weight, bias=None, process_group=None, dim=1, ring=False): + return _LinearWithReduceScatterForwardGatherBackward.apply(input_, weight, bias, process_group, dim, ring) def matmul_gather_forward_reducescatter_backward( - input_, weight, bias, process_group, async_grad_reduce_scatter, dim, overlap + input_, weight, bias, process_group, async_grad_reduce_scatter, dim, overlap, ring=False ): return _MatmulWithGatherForwardReduceScatterBackward.apply( - input_, weight, bias, process_group, async_grad_reduce_scatter, dim, overlap + input_, weight, bias, process_group, async_grad_reduce_scatter, dim, overlap, ring ) -def gather_forward_split_backward(input_, dim, process_group): - return _GatherForwardSplitBackward.apply(input_, dim, process_group) +def gather_forward_split_backward(input_, dim, process_group, grad_scale=None): + return _GatherForwardSplitBackward.apply(input_, dim, process_group, grad_scale) -def split_forward_gather_backward(input_, dim, process_group): - return _SplitForwardGatherBackward.apply(input_, dim, process_group) +def split_forward_gather_backward(input_, dim, process_group, grad_scale=None): + return _SplitForwardGatherBackward.apply(input_, dim, process_group, grad_scale) def reduce_forward(input_, process_group): @@ -650,3 +985,7 @@ def reduce_forward(input_, process_group): def reduce_backward(input_, process_group): return _ReduceBackward.apply(input_, process_group) + + +def all_to_all_comm(input_, process_group=None, scatter_dim=2, gather_dim=1): + return _AllToAll.apply(input_, process_group, scatter_dim, gather_dim) diff --git a/colossalai/shardformer/layer/attn.py b/colossalai/shardformer/layer/attn.py new file mode 100644 index 000000000000..f3f6e59d3d6a --- /dev/null +++ b/colossalai/shardformer/layer/attn.py @@ -0,0 +1,269 @@ +from enum import Enum +from typing import Callable, Dict, Optional, Tuple + +import torch +import torch.nn.functional as F + +from colossalai.kernel.kernel_loader import ( + FlashAttentionForFloatAndCustomMaskLoader, + FlashAttentionLoader, + FlashAttentionWithCustomMaskLoader, + FlashAttentionWithPaddingMaskLoader, + KernelLoader, +) + +__all__ = [ + "AttnMaskType", + "ColoAttention", +] + + +class AttnMaskType(Enum): + CUSTOM = 0 + PADDED = 1 + CAUSAL = 2 + PADDED_CAUSAL = 3 + + +def invert_mask(mask: torch.Tensor) -> torch.Tensor: + """Invert the mask tensor. + + Args: + mask (torch.Tensor): Mask tensor. Shape should be [B, 1, Sq, Skv] + + Returns: + torch.Tensor: Inverted mask tensor. + """ + inverted_mask = 1.0 - mask + return inverted_mask.masked_fill(inverted_mask.bool(), torch.finfo(mask.dtype).min) + + +# adapted from https://github.com/Dao-AILab/flash-attention/blob/main/flash_attn/bert_padding.py +def get_pad_info(padding_mask: torch.Tensor) -> Tuple[int, torch.Tensor, torch.Tensor]: + """Get padding information from padding mask. + + Args: + padding_mask (torch.Tensor): Padding mask tensor. Shape should be [B, S] + + Returns: + Tuple[int, torch.Tensor, torch.Tensor]: Tuple of (max_seq_len, cu_seqlens, indices) + """ + seqlens_in_batch = padding_mask.sum(dim=-1, dtype=torch.int32) + indices = torch.nonzero(padding_mask.flatten(), as_tuple=False).flatten() + max_seqlen_in_batch = seqlens_in_batch.max().item() + cu_seqlens = F.pad(torch.cumsum(seqlens_in_batch, dim=0, dtype=torch.torch.int32), (1, 0)) + return max_seqlen_in_batch, cu_seqlens, indices + + +class ColoAttention: + _kernel_dispatch_map: Optional[Dict[torch.dtype, Dict[Optional[AttnMaskType], Callable]]] = None + + @staticmethod + def _init_kernels_dispatch(): + if ColoAttention._kernel_dispatch_map is None: + # fp16/bf16 + half_dispatch_map = { + None: FlashAttentionLoader(), + AttnMaskType.CUSTOM: FlashAttentionWithCustomMaskLoader(), + AttnMaskType.PADDED: FlashAttentionWithPaddingMaskLoader(), + AttnMaskType.CAUSAL: FlashAttentionLoader(), + AttnMaskType.PADDED_CAUSAL: FlashAttentionWithPaddingMaskLoader(), + } + # fp32 + float_dispatch_map = { + None: FlashAttentionForFloatAndCustomMaskLoader(), + AttnMaskType.CUSTOM: FlashAttentionForFloatAndCustomMaskLoader(), + AttnMaskType.CAUSAL: FlashAttentionForFloatAndCustomMaskLoader(), + } + ColoAttention._kernel_dispatch_map = { + torch.float16: half_dispatch_map, + torch.bfloat16: half_dispatch_map, + torch.float32: float_dispatch_map, + } + + @staticmethod + def _dispatch_kernel(dtype: torch.dtype, mask_type: Optional[AttnMaskType]) -> Callable: + ColoAttention._init_kernels_dispatch() + if ( + dtype not in ColoAttention._kernel_dispatch_map + or mask_type not in ColoAttention._kernel_dispatch_map[dtype] + ): + raise ValueError( + "FlashAttention kernel is not available for dtype {} and mask_type {}".format(dtype, mask_type) + ) + # lazy load + if isinstance(ColoAttention._kernel_dispatch_map[dtype][mask_type], KernelLoader): + ColoAttention._kernel_dispatch_map[dtype][mask_type] = ColoAttention._kernel_dispatch_map[dtype][ + mask_type + ].load() + return ColoAttention._kernel_dispatch_map[dtype][mask_type] + + @staticmethod + def prepare_attn_kwargs( + shape_4d: Tuple[int], + dtype: torch.dtype, + device: torch.device, + q_padding_mask: Optional[torch.Tensor] = None, + kv_padding_mask: Optional[torch.Tensor] = None, + is_causal: bool = False, + ) -> Dict[str, torch.Tensor]: + """Return a dictionary of keyword arguments for attention function. It supports 4 mask type. + 1. custom mask: no padding mask and is_causal=False, return {}, users should handle attention mask by themselves. + 2. padded mask: recv padding mask and is_causal=False, return {attention_mask, attention_mask_type, cu_seqlens_q, cu_seqlens_kv, max_seqlen_q, max_seqlen_kv, q_indices, kv_indices}. + 3. causal mask: no padding mask and is_causal=True, return {attention_mask, attention_mask_type}. + 4. padded causal mask: recv padding mask and is_causal=True, return {attention_mask, attention_mask_type, cu_seqlens_q, cu_seqlens_kv, max_seqlen_q, max_seqlen_kv, q_indices, kv_indices}. + + Args: + shape_4d (Tuple[int]): Should be (B, 1, Sq, Skv) + dtype (torch.dtype): Dtype of attention mask, generally should be ``hidden_states.dtype`` + device (torch.device): Device of attention mask, generally should be ``hidden_states.device`` + q_padding_mask (Optional[torch.Tensor], optional): Padding mask of query. It should be a long tensor or int tensor. + The shape should be [B, Sq]. ``1`` means valid token, and ``0`` means padding token. Defaults to None. + kv_padding_mask (Optional[torch.Tensor], optional): Padding mask of key and value. It should be a long tensor or int tensor. + The shape should be [B, Skv]. ``1`` means valid token, and ``0`` means padding token. + If it's None and ``q_padding_mask`` is not None, it will be set to ``q_padding_mask``. Defaults to None. + is_causal (bool, optional): Whether to use causal attention mask. Defaults to False. + + Returns: + Dict[str, torch.Tensor]: Dictionary of keyword arguments for attention function. + """ + if q_padding_mask is None and not is_causal: + return {} + assert len(shape_4d) == 4 and shape_4d[1] == 1 + b, _, s_q, s_kv = shape_4d + outputs = {} + if (q_padding_mask is None or q_padding_mask.bool().all()) and ( + kv_padding_mask is None or kv_padding_mask.bool().all() + ): + # no padding + assert is_causal + outputs["attention_mask_type"] = AttnMaskType.CAUSAL + attention_mask = torch.ones(s_q, s_kv, dtype=dtype, device=device).tril(diagonal=0).expand(b, s_q, s_kv) + else: + if kv_padding_mask is None: + # self attention + kv_padding_mask = q_padding_mask + assert q_padding_mask.shape == (b, s_q) and kv_padding_mask.shape == ( + b, + s_kv, + ), f"q_padding_mask shape {q_padding_mask.shape} and kv_padding_mask shape {kv_padding_mask.shape} should be the same. ({shape_4d})" + attention_mask = torch.einsum("bi,bj->bij", q_padding_mask, kv_padding_mask).to(dtype=dtype, device=device) + max_seqlen_q, cu_seqlens_q, q_indices = get_pad_info(q_padding_mask) + max_seqlen_kv, cu_seqlens_kv, kv_indices = get_pad_info(kv_padding_mask) + outputs.update( + { + "cu_seqlens_q": cu_seqlens_q, + "cu_seqlens_kv": cu_seqlens_kv, + "max_seqlen_q": max_seqlen_q, + "max_seqlen_kv": max_seqlen_kv, + "q_indices": q_indices, + "kv_indices": kv_indices, + } + ) + if is_causal: + outputs["attention_mask_type"] = AttnMaskType.PADDED_CAUSAL + attention_mask = attention_mask * attention_mask.new_ones(s_q, s_kv).tril(diagonal=0) + else: + outputs["attention_mask_type"] = AttnMaskType.PADDED + attention_mask = invert_mask(attention_mask).unsqueeze(1) + outputs["attention_mask"] = attention_mask + return outputs + + @staticmethod + def attention( + q: torch.Tensor, + k: torch.Tensor, + v: torch.Tensor, + attention_mask: Optional[torch.Tensor] = None, + attention_mask_type: AttnMaskType = AttnMaskType.CUSTOM, + cu_seqlens_q: Optional[torch.Tensor] = None, + cu_seqlens_kv: Optional[torch.Tensor] = None, + max_seqlen_q: Optional[int] = None, + max_seqlen_kv: Optional[int] = None, + q_indices: Optional[torch.Tensor] = None, + kv_indices: Optional[torch.Tensor] = None, + dropout_p: float = 0.0, + scale: Optional[float] = None, + ) -> torch.Tensor: + """Flash Attention function. It supports 4 mask type. + 1. custom mask: recv attention_mask + 2. padded mask: recv attention_mask, attention_mask_type, cu_seqlens_q, cu_seqlens_kv, max_seqlen_q, max_seqlen_kv, indices + 3. causal mask: recv attention_mask, attention_mask_type + 4. padded causal mask: recv attention_mask, attention_mask_type, cu_seqlens_q, cu_seqlens_kv, max_seqlen_q, max_seqlen_kv, indices + + Args: + q (torch.Tensor): Query tensor. Shape should be [B, N, Sq, D] + k (torch.Tensor): Key tensor. Shape should be [B, N, Skv, D] + v (torch.Tensor): Value tensor. Shape should be [B, N, Skv, D] + attention_mask (Optional[torch.Tensor], optional): Attention mask tensor. Shape should be [B, 1, Sq, Skv]. Defaults to None. + attention_mask_type (AttnMaskType, optional): Attention mask type. Defaults to AttnMaskType.CUSTOM. + cu_seqlens_q (Optional[torch.Tensor], optional): The cumulative sequence lengths + of the sequences in the batch, used to index into q. + Shape should be [B+1]. Defaults to None. + cu_seqlens_kv (Optional[torch.Tensor], optional): The cumulative sequence lengths + of the sequences in the batch, used to index into kv. + Shape should be [B+1]. Defaults to None. + max_seqlen_q (Optional[int], optional): Maximum query sequence length in the batch. Defaults to None. + max_seqlen_kv (Optional[int], optional): Maximum key/value sequence length in the batch. Defaults to None. + indices (Optional[torch.Tensor], optional): The indices of non-masked tokens from the flattened input sequence. + Shape should be [NUM_TOKENS]. Defaults to None. + dropout_p (float, optional): Dropout probability. Defaults to 0.0. + scale (Optional[float], optional): Scaling factor applied prior to softmax. Defaults to None. + + Returns: + torch.Tensor: Output tensor. Shape should be [B, N, Sq, D] + """ + # known issue: sdpa does not support attention mask which contains whole row of masked tokens, which leads to nan + # this case is usaul when padding mask is used and self attention is performed + # thus, we don't use sdpa when padding mask is used + # sanity check + if attention_mask is not None: + assert torch.is_floating_point(attention_mask), "attention_mask should be a floating point tensor." + if attention_mask_type in (AttnMaskType.CUSTOM, AttnMaskType.CAUSAL): + assert ( + cu_seqlens_q is None + and cu_seqlens_kv is None + and max_seqlen_q is None + and max_seqlen_kv is None + and q_indices is None + and kv_indices is None + ) + if attention_mask_type == AttnMaskType.CUSTOM: + assert not torch.all(attention_mask != 0, dim=-1).any() + elif attention_mask_type in ( + AttnMaskType.PADDED, + AttnMaskType.PADDED_CAUSAL, + ): + assert ( + cu_seqlens_q is not None + and cu_seqlens_kv is not None + and max_seqlen_q is not None + and max_seqlen_kv is not None + and q_indices is not None + and kv_indices is not None + ) + else: + # if attention_mask is None, attention_mask_type should be the default value + assert attention_mask_type == AttnMaskType.CUSTOM + # kernel dispatch + mask_type = attention_mask_type if attention_mask is not None else None + attn_func = ColoAttention._dispatch_kernel(q.dtype, mask_type) + is_causal = attention_mask is not None and attention_mask_type in ( + AttnMaskType.CAUSAL, + AttnMaskType.PADDED_CAUSAL, + ) + return attn_func( + q, + k, + v, + dropout_p=dropout_p, + scale=scale, + attention_mask=attention_mask, + is_causal=is_causal, + cu_seqlens_q=cu_seqlens_q, + cu_seqlens_kv=cu_seqlens_kv, + max_seqlen_q=max_seqlen_q, + max_seqlen_kv=max_seqlen_kv, + q_indices=q_indices, + kv_indices=kv_indices, + ) diff --git a/colossalai/shardformer/layer/linear.py b/colossalai/shardformer/layer/linear.py index eeb0ef39975f..7c8619ad8f5c 100644 --- a/colossalai/shardformer/layer/linear.py +++ b/colossalai/shardformer/layer/linear.py @@ -23,11 +23,13 @@ ) from ._operation import ( + gather_forward_reducescatter_backward, gather_forward_split_backward, linear_gather_forward_reducescatter_backward, linear_reducescatter_forward_gather_backward, linear_with_async_comm, reduce_forward, + reducescatter_forward_gather_backward, split_forward_gather_backward, ) from .parallel_module import ParallelModule @@ -74,7 +76,7 @@ def __init__( device: torch.device = None, process_group: ProcessGroup = None, gather_output: bool = False, - seq_parallel: bool = False, + seq_parallel_mode: str = None, seq_parallel_dim: int = 1, overlap: torch.cuda.Stream = None, skip_bias_add: bool = False, @@ -89,7 +91,7 @@ def __init__( self.in_features = in_features self.out_features = out_features self.gather_output = gather_output - self.seq_parallel = seq_parallel + self.seq_parallel_mode = seq_parallel_mode self.seq_parallel_dim = seq_parallel_dim self.overlap = overlap self.skip_bias_add = skip_bias_add @@ -196,12 +198,18 @@ def forward(self, input_: Tensor) -> Tuple[Tensor, Tensor]: # Matrix multiply. bias = self.bias if not self.skip_bias_add else None - if self.seq_parallel: + + if self.seq_parallel_mode is None: + output_parallel = linear_with_async_comm(input_parallel, self.weight, bias, self.process_group, True) + elif self.seq_parallel_mode == "split_gather": + input_parallel = gather_forward_reducescatter_backward( + input_parallel, self.process_group, self.seq_parallel_dim + ) + output_parallel = linear_with_async_comm(input_parallel, self.weight, bias, self.process_group, False) + elif self.seq_parallel_mode == "ring": output_parallel = linear_gather_forward_reducescatter_backward( - input_parallel, self.weight, bias, self.process_group, True, self.seq_parallel_dim, self.overlap + input_parallel, self.weight, bias, self.process_group, True, self.seq_parallel_dim, self.overlap, True ) - else: - output_parallel = linear_with_async_comm(input_parallel, self.weight, bias, self.process_group, True) if self.gather_output: # All-gather across the partitions. @@ -225,7 +233,8 @@ class Linear1D_Row(ParallelModule): dtype (`torch.dtype`): The dtype of parameters, defaults to None. parallel_input (bool): If set to ``True``, it's assumed that the input is split, defaults to False. process_group (`torch.distributed.ProcessGroup`): The process group to be used for weight sharding and communication, defaults to None. - seq_parallel (`bool`): If set to ``True``, it will use sequence parallel, defaults to False. + seq_parallel_mode (`str`): The type of sp mode, it will use sequence parallel when `seq_parallel_mode` is not None. Defaults to None. + seq_parallel_dim (`int`): Which dim will sequence parallelism split and gather the sequence. skip_bias_add (bool): If set to ``True``, it will skip bias add for linear layer, which is preserved for kernel fusion, defaults to False weight_initializer (:class:`typing.Callable`, optional): @@ -245,7 +254,7 @@ def __init__( dtype: torch.dtype = None, device: torch.device = None, process_group: ProcessGroup = None, - seq_parallel: bool = False, + seq_parallel_mode: str = None, seq_parallel_dim: int = 1, parallel_input: bool = True, skip_bias_add: bool = False, @@ -265,7 +274,7 @@ def __init__( self.parallel_input = parallel_input self.skip_bias_add = skip_bias_add self.process_group = process_group - self.seq_parallel = seq_parallel + self.seq_parallel_mode = seq_parallel_mode self.seq_parallel_dim = seq_parallel_dim self.num_partitions = dist.get_world_size(self.process_group) @@ -403,18 +412,26 @@ def forward(self, input_: Tensor) -> Tensor: output_parallel_list[i], group=self.process_group, async_op=True ) handle_list.append(handle) - # output_parallel_list[i] = reduce_input(output_parallel_list[i], ParallelMode.PARALLEL_1D) for handle in handle_list: handle.wait() output = torch.cat(output_parallel_list, dim=-1) else: - output_parallel = linear_with_async_comm(input_, self.weight, None, None, False) - if self.seq_parallel: - output = linear_reducescatter_forward_gather_backward( + if self.seq_parallel_mode is None: + output_parallel = linear_with_async_comm(input_, self.weight, None, self.process_group, False) + output = reduce_forward(output_parallel, self.process_group) + elif self.seq_parallel_mode == "split_gather": + output_parallel = linear_with_async_comm(input_, self.weight, None, self.process_group, False) + output = reducescatter_forward_gather_backward( output_parallel, self.process_group, self.seq_parallel_dim ) - else: - output = reduce_forward(output_parallel, self.process_group) + elif self.seq_parallel_mode == "ring": + output = linear_reducescatter_forward_gather_backward( + input_, + self.weight, + process_group=self.process_group, + dim=self.seq_parallel_dim, + ring=True, + ) if not self.skip_bias_add: if self.bias is not None: diff --git a/colossalai/shardformer/layer/qkv_fused_linear.py b/colossalai/shardformer/layer/qkv_fused_linear.py index 12476d050600..dc3634238f74 100644 --- a/colossalai/shardformer/layer/qkv_fused_linear.py +++ b/colossalai/shardformer/layer/qkv_fused_linear.py @@ -25,12 +25,12 @@ from ._operation import ( gather_forward_split_backward, - linear_reducescatter_forward_gather_backward, linear_with_async_comm, matmul_gather_forward_reducescatter_backward, matmul_with_async_comm, reduce_backward, reduce_forward, + reducescatter_forward_gather_backward, split_forward_gather_backward, ) from .parallel_module import ParallelModule @@ -150,7 +150,7 @@ class GPT2FusedLinearConv1D_Col(ParallelModule): device (`torch.device`): The device of parameters, defaults to None. n_fused (int): The number items fused, defaults to 3 (QKV). process_group (`torch.distributed.ProcessGroup`): The process group to be used for weight sharding and communication, defaults to None. - seq_parallel (`bool`): If set to ``True``, it will use sequence parallel, defaults to False. + seq_parallel_mode (str): If set to ``None``, it will not use sequence parallel, otherwise will use corresponding mode of sequence parallel, defaults to None. gather_output (bool, optional): If true, call all-gather on output and make Y available to all GPUs, otherwise, every GPU will have its output which is :math:`Y_i = XA_i`, defaults to False @@ -175,7 +175,7 @@ def __init__( process_group: ProcessGroup = None, async_communication: bool = False, gather_output: bool = False, - seq_parallel: bool = False, + seq_parallel_mode: str = None, overlap: bool = False, skip_bias_add: bool = False, n_fused: int = 3, @@ -190,7 +190,7 @@ def __init__( self.in_features = in_features self.out_features = out_features self.gather_output = gather_output - self.seq_parallel = seq_parallel + self.seq_parallel_mode = seq_parallel_mode self.overlap = overlap self.skip_bias_add = skip_bias_add self.device = device @@ -312,17 +312,22 @@ def forward(self, input_: Tensor) -> Tuple[Tensor, Tensor]: # Matrix multiply. bias = self.bias if not self.skip_bias_add else None - if self.seq_parallel: - input_parallel = input_ - output_parallel = matmul_gather_forward_reducescatter_backward( - input_parallel, self.weight, bias, self.process_group, True, 1, self.overlap - ) - else: + if self.seq_parallel_mode is None: # Set up backprop all-reduce. input_parallel = reduce_backward(input_, self.process_group) output_parallel = matmul_with_async_comm( input_parallel, self.weight, bias, self.process_group, self.async_communication ) + elif self.seq_parallel_mode == "split_gather": + input_parallel = input_ + output_parallel = matmul_gather_forward_reducescatter_backward( + input_parallel, self.weight, bias, self.process_group, True, 1, self.overlap + ) + elif self.seq_parallel_mode == "ring": + input_parallel = input_ + output_parallel = matmul_gather_forward_reducescatter_backward( + input_parallel, self.weight, bias, self.process_group, True, 1, self.overlap, True + ) if self.gather_output: # All-gather across the partitions. @@ -347,7 +352,7 @@ class GPT2FusedLinearConv1D_Row(ParallelModule): dtype (`torch.dtype`): The dtype of parameters, defaults to None. parallel_input (bool): If set to ``True``, it's assumed that the input is split, defaults to False. skip_bias_add (bool): If set to ``True``, it will skip bias add for linear layer, - seq_parallel (`bool`): If set to ``True``, it will use sequence parallel, defaults to False. + seq_parallel_mode (str): If set to ``None``, it will not use sequence parallel, otherwise will use corresponding mode of sequence parallel, defaults to None. which is preserved for kernel fusion, defaults to False weight_initializer (:class:`typing.Callable`, optional): The initializer of weight, defaults to kaiming uniform initializer. @@ -366,7 +371,7 @@ def __init__( dtype: torch.dtype = None, device: torch.device = None, process_group: ProcessGroup = None, - seq_parallel: bool = False, + seq_parallel_mode: str = None, parallel_input: bool = True, skip_bias_add: bool = False, weight: Optional[Parameter] = None, @@ -385,7 +390,7 @@ def __init__( self.parallel_input = parallel_input self.skip_bias_add = skip_bias_add self.process_group = process_group - self.seq_parallel = seq_parallel + self.seq_parallel_mode = seq_parallel_mode self.num_partitions = dist.get_world_size(self.process_group) if skip_bias_add and not bias: @@ -528,11 +533,15 @@ def forward(self, input_: Tensor) -> Tensor: handle.wait() output = torch.cat(output_parallel_list, dim=-1) else: - output_parallel = torch.matmul(input_, self.weight) - if self.seq_parallel: - output = linear_reducescatter_forward_gather_backward(output_parallel, self.process_group, 1) - else: + if self.seq_parallel_mode is None: + output_parallel = torch.matmul(input_, self.weight) output = reduce_forward(output_parallel, self.process_group) + elif self.seq_parallel_mode == "split_gather": + output_parallel = torch.matmul(input_, self.weight) + output = reducescatter_forward_gather_backward(output_parallel, self.process_group, 1) + elif self.seq_parallel_mode == "ring": + output_parallel = torch.matmul(input_, self.weight) + output = reducescatter_forward_gather_backward(output_parallel, self.process_group, 1) if not self.skip_bias_add: if self.bias is not None: @@ -702,7 +711,6 @@ def from_native_module( # process_group=process_group, # is_transposed=False) # linear_1d.bias.data.copy_(sharded_bias.data) - print(linear_1d.weight.shape) return linear_1d def reset_parameters(self, weight_initializer, bias_initializer) -> None: diff --git a/colossalai/shardformer/layer/utils.py b/colossalai/shardformer/layer/utils.py index 0d2cc1b3370d..9c6ced4454dc 100644 --- a/colossalai/shardformer/layer/utils.py +++ b/colossalai/shardformer/layer/utils.py @@ -35,17 +35,21 @@ def is_sp_partial_derived_param(param): return getattr(param, "partial_derived", False) @staticmethod - def allreduce_partial_data_grad(tp_group: ProcessGroup, model: nn.Module = None, grads: List[torch.Tensor] = None): + def allreduce_partial_data_grad( + process_group: ProcessGroup, + model: nn.Module = None, + grads: List[torch.Tensor] = None, + ): """ Allreduce partial derived gradients across the specified process group. This function performs gradient synchronization for parameters that are marked as partially derived in sequence parallelism. Args: - tp_group (ProcessGroup): The process group for gradient synchronization. + process_group (ProcessGroup): The process group for gradient synchronization. model (nn.Module): The model from which gradients will be synchronized. grads (List[torch.Tensor]): The list of gradients to be synchronized. - + only_sp_partial (bool): Whether handle all the parameters or only parameters marked as partial derived. Raises: AssertionError: If both `model` and `grads` are provided or neither is provided. """ @@ -53,22 +57,26 @@ def allreduce_partial_data_grad(tp_group: ProcessGroup, model: nn.Module = None, assert (model is not None) ^ (grads is not None), "Exactly one of model and grads must be not None." # Get the size of the process group, which determines whether synchronization is needed. - tp_size = get_world_size(tp_group) if tp_group is not None else 1 + group_size = get_world_size(process_group) if process_group is not None else 1 - if tp_size == 1: + if group_size == 1: # If the process group size is 1, no synchronization is required. return if model is not None: # If `model` is provided, extract partial derived gradients from the model's parameters. grads = [] + for p in model.parameters(): - if p.grad is not None and SeqParallelUtils.is_sp_partial_derived_param(p): - grads.append(p.grad.data) + if p.grad is not None: + if SeqParallelUtils.is_sp_partial_derived_param(p): + grads.append(p.grad.data) # Flatten and reduce the gradients using the specified process group. + if len(grads) == 0: + return coalesced = _flatten_dense_tensors(grads) - dist.all_reduce(coalesced, op=dist.ReduceOp.SUM, group=tp_group) + dist.all_reduce(coalesced, op=dist.ReduceOp.SUM, group=process_group) # Unflatten the synchronized gradients and update the model's gradients. for buf, synced in zip(grads, _unflatten_dense_tensors(coalesced, grads)): @@ -76,7 +84,7 @@ def allreduce_partial_data_grad(tp_group: ProcessGroup, model: nn.Module = None, else: # If `grads` are provided explicitly, synchronize those gradients directly. coalesced = _flatten_dense_tensors(grads) - dist.all_reduce(coalesced, op=dist.ReduceOp.SUM, group=tp_group) + dist.all_reduce(coalesced, op=dist.ReduceOp.SUM, group=process_group) for buf, synced in zip(grads, _unflatten_dense_tensors(coalesced, grads)): buf.copy_(synced) diff --git a/colossalai/shardformer/modeling/bert.py b/colossalai/shardformer/modeling/bert.py index 7411e1d0ec46..0838fcee682e 100644 --- a/colossalai/shardformer/modeling/bert.py +++ b/colossalai/shardformer/modeling/bert.py @@ -186,13 +186,14 @@ def bert_model_forward( # split the input tensor along sequence dimension # [batch_size, seq_len, hidden_size] -> [batch_size, seq_len/TP_size, hidden_size] if shard_config is not None and shard_config.enable_sequence_parallelism: - hidden_states = split_forward_gather_backward( - hidden_states, dim=1, process_group=shard_config.tensor_parallel_process_group - ) - if encoder_hidden_states is not None: - encoder_hidden_states = split_forward_gather_backward( - encoder_hidden_states, dim=1, process_group=shard_config.tensor_parallel_process_group + if shard_config.sequence_parallelism_mode == "split_gather": + hidden_states = split_forward_gather_backward( + hidden_states, dim=1, process_group=shard_config.tensor_parallel_process_group ) + if encoder_hidden_states is not None: + encoder_hidden_states = split_forward_gather_backward( + encoder_hidden_states, dim=1, process_group=shard_config.tensor_parallel_process_group + ) for idx, encoder_layer in enumerate(self.encoder.layer[start_idx:end_idx], start=start_idx): if stage_manager.is_first_stage() and idx == 0: @@ -240,9 +241,10 @@ def custom_forward(*inputs): # When sequence parallelism done, gather the output tensor in forward and split it in backward if shard_config is not None and shard_config.enable_sequence_parallelism: - hidden_states = gather_forward_split_backward( - hidden_states, dim=1, process_group=shard_config.tensor_parallel_process_group - ) + if shard_config.sequence_parallelism_mode == "split_gather": + hidden_states = gather_forward_split_backward( + hidden_states, dim=1, process_group=shard_config.tensor_parallel_process_group + ) if output_hidden_states: all_hidden_states = all_hidden_states + (hidden_states,) diff --git a/colossalai/shardformer/modeling/blip2.py b/colossalai/shardformer/modeling/blip2.py index d5c10541a28f..bd84c87c667d 100644 --- a/colossalai/shardformer/modeling/blip2.py +++ b/colossalai/shardformer/modeling/blip2.py @@ -3,6 +3,8 @@ import torch import torch.nn as nn +from colossalai.shardformer.layer import ColoAttention + def forward_fn(): def forward( @@ -62,8 +64,6 @@ def forward( def get_blip2_flash_attention_forward(): from transformers.models.blip_2.modeling_blip_2 import Blip2Attention - from colossalai.nn.layer.colo_attention import ColoAttention - def forward( self: Blip2Attention, hidden_states: torch.Tensor, @@ -71,16 +71,25 @@ def forward( output_attentions: Optional[bool] = False, ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]: """Input shape: Batch x Time x Channel""" - + assert head_mask is None, "head_mask is not supported in FlashAttention" bsz, tgt_len, embed_dim = hidden_states.size() mixed_qkv = self.qkv(hidden_states) - mixed_qkv = mixed_qkv.reshape(bsz, tgt_len, 3, self.num_heads, -1).permute(2, 0, 1, 3, 4) - query_states, key_states, value_states = mixed_qkv[0], mixed_qkv[1], mixed_qkv[2] + mixed_qkv = mixed_qkv.reshape(bsz, tgt_len, 3, self.num_heads, self.head_dim).permute(2, 0, 3, 1, 4) + query_states, key_states, value_states = ( + mixed_qkv[0], + mixed_qkv[1], + mixed_qkv[2], + ) - attention = ColoAttention( - embed_dim=self.embed_dim, num_heads=self.num_heads, dropout=self.dropout.p, scale=self.scale + dropout_p = self.dropout.p if self.training else 0.0 + context_layer = ColoAttention.attention( + query_states, + key_states, + value_states, + dropout_p=dropout_p, + scale=self.scale, ) - context_layer = attention(query_states, key_states, value_states) + context_layer = context_layer.permute(0, 2, 1, 3).reshape(bsz, tgt_len, self.embed_dim) output = self.projection(context_layer) outputs = (output, None) @@ -93,7 +102,11 @@ def forward( def get_jit_fused_blip2_QFormer_self_output_forward(): from transformers.models.blip_2.modeling_blip_2 import Blip2QFormerSelfOutput - def forward(self: Blip2QFormerSelfOutput, hidden_states: torch.Tensor, input_tensor: torch.Tensor) -> torch.Tensor: + def forward( + self: Blip2QFormerSelfOutput, + hidden_states: torch.Tensor, + input_tensor: torch.Tensor, + ) -> torch.Tensor: hidden_states = self.dense(hidden_states) hidden_states = self.dropout_add(hidden_states, input_tensor, self.dropout.p, self.dropout.training) hidden_states = self.LayerNorm(hidden_states) @@ -105,7 +118,11 @@ def forward(self: Blip2QFormerSelfOutput, hidden_states: torch.Tensor, input_ten def get_jit_fused_blip2_QFormer_output_forward(): from transformers.models.blip_2.modeling_blip_2 import Blip2QFormerOutput - def forward(self: Blip2QFormerOutput, hidden_states: torch.Tensor, input_tensor: torch.Tensor) -> torch.Tensor: + def forward( + self: Blip2QFormerOutput, + hidden_states: torch.Tensor, + input_tensor: torch.Tensor, + ) -> torch.Tensor: hidden_states = self.dense(hidden_states) hidden_states = self.dropout_add(hidden_states, input_tensor, self.dropout.p, self.dropout.training) hidden_states = self.LayerNorm(hidden_states) diff --git a/colossalai/shardformer/modeling/bloom.py b/colossalai/shardformer/modeling/bloom.py index d94c30d29e71..fe70376e144d 100644 --- a/colossalai/shardformer/modeling/bloom.py +++ b/colossalai/shardformer/modeling/bloom.py @@ -213,10 +213,11 @@ def bloom_model_forward( # split the input tensor along sequence dimension # [batch_size, seq_len, hidden_size] -> [batch_size, seq_len/TP_size, hidden_size] - if shard_config.enable_sequence_parallelism: - hidden_states = split_forward_gather_backward( - hidden_states, dim=1, process_group=shard_config.tensor_parallel_process_group - ) + if shard_config and shard_config.enable_sequence_parallelism: + if shard_config.sequence_parallelism_mode == "split_gather": + hidden_states = split_forward_gather_backward( + hidden_states, dim=1, process_group=shard_config.tensor_parallel_process_group + ) start_idx, end_idx = stage_index[0], stage_index[1] for i, (block, layer_past) in enumerate( @@ -261,10 +262,11 @@ def custom_forward(*inputs): all_self_attentions = all_self_attentions + (outputs[2 if use_cache else 1],) # When sequence parallelism done, gather the output tensor in forward and split it in backward - if shard_config.enable_sequence_parallelism: - hidden_states = gather_forward_split_backward( - hidden_states, dim=1, process_group=shard_config.tensor_parallel_process_group - ) + if shard_config and shard_config.enable_sequence_parallelism: + if shard_config.sequence_parallelism_mode == "split_gather": + hidden_states = gather_forward_split_backward( + hidden_states, dim=1, process_group=shard_config.tensor_parallel_process_group + ) if stage_manager.is_last_stage(): # Add last hidden state diff --git a/colossalai/shardformer/modeling/chatglm2.py b/colossalai/shardformer/modeling/chatglm2.py index d13bd34926a5..9207b34d0d1c 100644 --- a/colossalai/shardformer/modeling/chatglm2.py +++ b/colossalai/shardformer/modeling/chatglm2.py @@ -1,4 +1,5 @@ """ PyTorch ChatGLM model. """ + from typing import List, Optional, Tuple import torch @@ -9,63 +10,49 @@ from colossalai.pipeline.stage_manager import PipelineStageManager from colossalai.shardformer import ShardConfig +from colossalai.shardformer.layer import AttnMaskType, ColoAttention from colossalai.shardformer.layer._operation import gather_forward_split_backward, split_forward_gather_backward from colossalai.shardformer.modeling.chatglm2_6b.modeling_chatglm import ChatGLMForConditionalGeneration, ChatGLMModel def get_flash_core_attention_forward(): - from colossalai.nn.layer.colo_attention import AttnMaskType, ColoAttention - from .chatglm2_6b.modeling_chatglm import CoreAttention def forward(self: CoreAttention, query_layer, key_layer, value_layer, attention_mask): - pytorch_major_version = int(torch.__version__.split(".")[0]) - if pytorch_major_version >= 2: - query_layer, key_layer, value_layer = [k.permute(1, 2, 0, 3) for k in [query_layer, key_layer, value_layer]] - if attention_mask is None and query_layer.shape[2] == key_layer.shape[2]: - context_layer = torch.nn.functional.scaled_dot_product_attention( - query_layer, key_layer, value_layer, is_causal=True - ) - else: - if attention_mask is not None: - attention_mask = ~attention_mask - context_layer = torch.nn.functional.scaled_dot_product_attention( - query_layer, key_layer, value_layer, attention_mask - ) - context_layer = context_layer.permute(2, 0, 1, 3) - new_context_layer_shape = context_layer.size()[:-2] + (self.hidden_size_per_partition,) - context_layer = context_layer.reshape(*new_context_layer_shape) - else: - # Raw attention scores - query_layer = query_layer.permute(1, 0, 2, 3).contiguous() - key_layer = key_layer.permute(1, 0, 2, 3).contiguous() - value_layer = value_layer.permute(1, 0, 2, 3).contiguous() - - scale = 1.0 / self.norm_factor - if self.coeff is not None: - scale = scale * self.coeff - - flash_attention_mask = None - attn_mask_type = None - if attention_mask is None: - attn_mask_type = AttnMaskType.causal - else: - flash_attention_mask = ~(attention_mask[:, :, -1].squeeze(1).to(torch.bool)).contiguous() - if not torch.all(flash_attention_mask): - attn_mask_type = AttnMaskType.paddedcausal - - attention = ColoAttention( - embed_dim=self.hidden_size_per_partition, - num_heads=self.num_attention_heads_per_partition, - dropout=self.attention_dropout.p, - scale=scale, + query_layer, key_layer, value_layer = [k.permute(1, 2, 0, 3) for k in [query_layer, key_layer, value_layer]] + if attention_mask is None and query_layer.shape[2] == key_layer.shape[2]: + attention_mask_type = AttnMaskType.CAUSAL + attn_bias = torch.zeros( + query_layer.shape[0], + 1, + query_layer.shape[2], + key_layer.shape[2], + dtype=query_layer.dtype, + device=query_layer.device, ) - context_layer = attention( - query_layer, key_layer, value_layer, attn_mask=flash_attention_mask, attn_mask_type=attn_mask_type + temp_mask = ( + torch.ones(query_layer.shape[2], key_layer.shape[2], dtype=torch.bool, device=query_layer.device) + .tril(diagonal=0) + .expand(query_layer.shape[0], 1, -1, -1) ) - - context_layer = context_layer.permute(1, 0, -1).contiguous() - + attn_bias.masked_fill_(temp_mask.logical_not(), torch.finfo(query_layer.dtype).min) + else: + attention_mask_type = AttnMaskType.CUSTOM + if attention_mask is not None: + attn_bias = torch.zeros_like(attention_mask, dtype=query_layer.dtype) + attn_bias.masked_fill_(attention_mask, torch.finfo(query_layer.dtype).min) + dropout_p = self.attention_dropout.p if self.training else 0.0 + context_layer = ColoAttention.attention( + query_layer, + key_layer, + value_layer, + attention_mask=attn_bias, + attention_mask_type=attention_mask_type, + dropout_p=dropout_p, + ) + context_layer = context_layer.permute(2, 0, 1, 3) + new_context_layer_shape = context_layer.size()[:-2] + (self.hidden_size_per_partition,) + context_layer = context_layer.reshape(*new_context_layer_shape) return context_layer return forward @@ -169,11 +156,17 @@ def chatglm_model_forward( if self.pre_seq_len is not None: if past_key_values is None: past_key_values = self.get_prompt( - batch_size=batch_size, device=input_ids.device, dtype=inputs_embeds.dtype + batch_size=batch_size, + device=input_ids.device, + dtype=inputs_embeds.dtype, ) if attention_mask is not None: attention_mask = torch.cat( - [attention_mask.new_ones((batch_size, self.pre_seq_len)), attention_mask], dim=-1 + [ + attention_mask.new_ones((batch_size, self.pre_seq_len)), + attention_mask, + ], + dim=-1, ) if full_attention_mask is None: if (attention_mask is not None and not attention_mask.all()) or (past_key_values and seq_length != 1): @@ -198,17 +191,23 @@ def chatglm_model_forward( all_hidden_states = () if output_hidden_states else None start_idx, end_idx = stage_index[0], stage_index[1] - if shard_config.enable_sequence_parallelism: - hidden_states = split_forward_gather_backward( - hidden_states, dim=0, process_group=shard_config.tensor_parallel_process_group - ) + if shard_config and shard_config.enable_sequence_parallelism: + if shard_config.sequence_parallelism_mode == "split_gather": + hidden_states = split_forward_gather_backward( + hidden_states, dim=0, process_group=shard_config.tensor_parallel_process_group + ) for idx in range(start_idx, end_idx): layer = self.encoder._get_layer(idx) if output_hidden_states: all_hidden_states = all_hidden_states + (hidden_states,) if self.encoder.gradient_checkpointing and self.encoder.training: layer_ret = torch.utils.checkpoint.checkpoint( - layer, hidden_states, attention_mask, rotary_pos_emb, past_key_values[idx], use_cache + layer, + hidden_states, + attention_mask, + rotary_pos_emb, + past_key_values[idx], + use_cache, ) else: layer_ret = layer( @@ -222,10 +221,11 @@ def chatglm_model_forward( if use_cache: presents = presents + (kv_cache,) - if shard_config.enable_sequence_parallelism: - hidden_states = gather_forward_split_backward( - hidden_states, dim=0, process_group=shard_config.tensor_parallel_process_group - ) + if shard_config and shard_config.enable_sequence_parallelism: + if shard_config.sequence_parallelism_mode == "split_gather": + hidden_states = gather_forward_split_backward( + hidden_states, dim=0, process_group=shard_config.tensor_parallel_process_group + ) if output_hidden_states: all_hidden_states = all_hidden_states + (hidden_states,) if stage_manager.is_last_stage(): @@ -234,7 +234,14 @@ def chatglm_model_forward( hidden_states = self.encoder.final_layernorm(hidden_states) if not return_dict: return tuple( - v for v in [hidden_states, presents, all_hidden_states, all_self_attentions] if v is not None + v + for v in [ + hidden_states, + presents, + all_hidden_states, + all_self_attentions, + ] + if v is not None ) return BaseModelOutputWithPast( last_hidden_state=hidden_states, @@ -368,7 +375,9 @@ def forward( # Run encoder. # [seq_len, batch_size, hidden_size] -> [seq_len/TP_size, batch_size, hidden_size] inputs_embeds = split_forward_gather_backward( - inputs_embeds, dim=0, process_group=shard_config.tensor_parallel_process_group + inputs_embeds, + dim=0, + process_group=shard_config.tensor_parallel_process_group, ) hidden_states, presents, all_hidden_states, all_self_attentions = self.encoder( inputs_embeds, @@ -380,7 +389,9 @@ def forward( ) hidden_states = gather_forward_split_backward( - hidden_states, dim=0, process_group=shard_config.tensor_parallel_process_group + hidden_states, + dim=0, + process_group=shard_config.tensor_parallel_process_group, ) if not return_dict: diff --git a/colossalai/shardformer/modeling/gpt2.py b/colossalai/shardformer/modeling/gpt2.py index 1e22d9094eae..1306c8aa6299 100644 --- a/colossalai/shardformer/modeling/gpt2.py +++ b/colossalai/shardformer/modeling/gpt2.py @@ -21,12 +21,82 @@ from transformers.utils import logging from colossalai.pipeline.stage_manager import PipelineStageManager +from colossalai.shardformer.layer import ColoAttention from colossalai.shardformer.layer._operation import gather_forward_split_backward, split_forward_gather_backward from colossalai.shardformer.shard import ShardConfig from ..layer import cross_entropy_1d from ..layer._operation import gather_forward_split_backward +logger = logging.get_logger(__name__) + + +def _get_attention_mask( + self: GPT2Model, + shard_config: ShardConfig, + hidden_states: torch.Tensor, + past_key_values: Optional[Tuple[Tuple[torch.Tensor]]], + attention_mask: Optional[torch.FloatTensor], + encoder_hidden_states: Optional[torch.Tensor], + encoder_attention_mask: Optional[torch.FloatTensor], +) -> Tuple[Optional[Union[torch.Tensor, dict]], Optional[Union[torch.Tensor, dict]]]: + batch_size, seq_len = hidden_states.shape[:2] + # If a 2D or 3D attention mask is provided for the cross-attention + # we need to make broadcastable to [batch_size, num_heads, seq_length, seq_length] + if self.config.add_cross_attention and encoder_hidden_states is not None: + encoder_batch_size, encoder_sequence_length, _ = encoder_hidden_states.size() + if shard_config.enable_flash_attention: + encoder_attention_mask = ColoAttention.prepare_attn_kwargs( + (encoder_batch_size, 1, seq_len, encoder_sequence_length), + dtype=hidden_states.dtype, + dtype2=encoder_hidden_states.dtype, + q_padding_mask=attention_mask, + kv_padding_mask=encoder_attention_mask, + ) + else: + encoder_hidden_shape = (encoder_batch_size, encoder_sequence_length) + if encoder_attention_mask is None: + encoder_attention_mask = torch.ones(encoder_hidden_shape, device=encoder_hidden_states.device) + encoder_attention_mask = self.invert_attention_mask(encoder_attention_mask) + else: + if shard_config.enable_flash_attention: + encoder_attention_mask = {"attention_mask": None} + else: + encoder_attention_mask = None + # GPT2Attention mask. + past_key_values_length = 0 + if past_key_values is not None and past_key_values[0] is not None: + past_key_values_length = past_key_values[0][0].shape[2] + if shard_config.enable_flash_attention: + if attention_mask is not None: + attention_mask = attention_mask.view(batch_size, -1) + attention_mask = ColoAttention.prepare_attn_kwargs( + (batch_size, 1, seq_len, seq_len + past_key_values_length), + hidden_states.dtype, + hidden_states.device, + attention_mask, + is_causal=True, + ) + elif attention_mask is not None: + if batch_size <= 0: + raise ValueError("batch_size has to be defined and > 0") + attention_mask = attention_mask.view(batch_size, -1) + # We create a 3D attention mask from a 2D tensor mask. + # Sizes are [batch_size, 1, 1, to_seq_length] + # So we can broadcast to [batch_size, num_heads, from_seq_length, to_seq_length] + # this attention mask is more simple than the triangular masking of causal attention + # used in OpenAI GPT, we just need to prepare the broadcast dimension here. + attention_mask = attention_mask[:, None, None, :] + + # Since attention_mask is 1.0 for positions we want to attend and 0.0 for + # masked positions, this operation will create a tensor which is 0.0 for + # positions we want to attend and the dtype's smallest value for masked positions. + # Since we are adding it to the raw scores before the softmax, this is + # effectively the same as removing these entirely. + attention_mask = attention_mask.to(dtype=self.dtype) # fp16 compatibility + attention_mask = (1.0 - attention_mask) * torch.finfo(self.dtype).min + return attention_mask, encoder_attention_mask + class GPT2PipelineForwards: """ @@ -83,10 +153,10 @@ def gpt2_model_forward( elif input_ids is not None: input_shape = input_ids.size() input_ids = input_ids.view(-1, input_shape[-1]) - batch_size = input_ids.shape[0] + input_ids.shape[0] elif inputs_embeds is not None: input_shape = inputs_embeds.size()[:-1] - batch_size = inputs_embeds.shape[0] + inputs_embeds.shape[0] else: raise ValueError("You have to specify either input_ids or inputs_embeds") @@ -99,38 +169,7 @@ def gpt2_model_forward( input_shape = hidden_states.size()[:-1] device = hidden_states.device hidden_states = hidden_states.view((-1,) + hidden_states.shape[-2:]) - batch_size = hidden_states.shape[0] - - # GPT2Attention mask. - if attention_mask is not None: - if batch_size <= 0: - raise ValueError("batch_size has to be defined and > 0") - attention_mask = attention_mask.view(batch_size, -1) - # We create a 3D attention mask from a 2D tensor mask. - # Sizes are [batch_size, 1, 1, to_seq_length] - # So we can broadcast to [batch_size, num_heads, from_seq_length, to_seq_length] - # this attention mask is more simple than the triangular masking of causal attention - # used in OpenAI GPT, we just need to prepare the broadcast dimension here. - attention_mask = attention_mask[:, None, None, :] - - # Since attention_mask is 1.0 for positions we want to attend and 0.0 for - # masked positions, this operation will create a tensor which is 0.0 for - # positions we want to attend and the dtype's smallest value for masked positions. - # Since we are adding it to the raw scores before the softmax, this is - # effectively the same as removing these entirely. - attention_mask = attention_mask.to(dtype=self.dtype) # fp16 compatibility - attention_mask = (1.0 - attention_mask) * torch.finfo(self.dtype).min - - # If a 2D or 3D attention mask is provided for the cross-attention - # we need to make broadcastable to [batch_size, num_heads, seq_length, seq_length] - if self.config.add_cross_attention and encoder_hidden_states is not None: - encoder_batch_size, encoder_sequence_length, _ = encoder_hidden_states.size() - encoder_hidden_shape = (encoder_batch_size, encoder_sequence_length) - if encoder_attention_mask is None: - encoder_attention_mask = torch.ones(encoder_hidden_shape, device=device) - encoder_attention_mask = self.invert_attention_mask(encoder_attention_mask) - else: - encoder_attention_mask = None + hidden_states.shape[0] # Prepare head mask if needed # 1.0 in head_mask indicate we keep the head @@ -156,6 +195,16 @@ def gpt2_model_forward( output_shape = input_shape + (hidden_states.size(-1),) + attention_mask, encoder_attention_mask = _get_attention_mask( + self, + shard_config, + hidden_states, + past_key_values, + attention_mask, + encoder_hidden_states, + encoder_attention_mask, + ) + if self.gradient_checkpointing and self.training: if use_cache: logger.warning_once( @@ -169,10 +218,13 @@ def gpt2_model_forward( # split the input tensor along sequence dimension # [batch_size, seq_len, hidden_size] -> [batch_size, seq_len/TP_size, hidden_size] - if shard_config.enable_sequence_parallelism: - hidden_states = split_forward_gather_backward( - hidden_states, dim=1, process_group=shard_config.tensor_parallel_process_group - ) + if shard_config and shard_config.enable_sequence_parallelism: + if shard_config.sequence_parallelism_mode == "split_gather": + hidden_states = split_forward_gather_backward( + hidden_states, + dim=1, + process_group=shard_config.tensor_parallel_process_group, + ) # Going through held blocks. start_idx, end_idx = stage_index[0], stage_index[1] @@ -180,7 +232,7 @@ def gpt2_model_forward( block = self.h[i] torch.cuda.set_device(hidden_states.device) # Ensure that attention_mask is always on the same device as hidden_states - if attention_mask is not None: + if torch.is_tensor(attention_mask): attention_mask = attention_mask.to(hidden_states.device) if isinstance(head_mask, torch.Tensor): head_mask = head_mask.to(hidden_states.device) @@ -227,10 +279,13 @@ def custom_forward(*inputs): all_cross_attentions = all_cross_attentions + (outputs[3 if use_cache else 2],) # When sequence parallelism done, gather the output tensor in forward and split it in backward - if shard_config.enable_sequence_parallelism: - hidden_states = gather_forward_split_backward( - hidden_states, dim=1, process_group=shard_config.tensor_parallel_process_group - ) + if shard_config and shard_config.enable_sequence_parallelism: + if shard_config.sequence_parallelism_mode == "split_gather": + hidden_states = gather_forward_split_backward( + hidden_states, + dim=1, + process_group=shard_config.tensor_parallel_process_group, + ) if stage_manager.is_last_stage(): hidden_states = self.ln_f(hidden_states) @@ -245,7 +300,13 @@ def custom_forward(*inputs): if not return_dict: return tuple( v - for v in [hidden_states, presents, all_hidden_states, all_self_attentions, all_cross_attentions] + for v in [ + hidden_states, + presents, + all_hidden_states, + all_self_attentions, + all_cross_attentions, + ] if v is not None ) @@ -331,9 +392,11 @@ def gpt2_lmhead_model_forward( loss_fct = CrossEntropyLoss() shift_logits = shift_logits.view(-1, shift_logits.size(-1)) shift_labels = shift_labels.view(-1) - if shard_config.enable_tensor_parallelism: + if shard_config.enable_tensor_parallelism and shard_config.parallel_output: loss = cross_entropy_1d( - shift_logits, shift_labels, process_group=shard_config.tensor_parallel_process_group + shift_logits, + shift_labels, + process_group=shard_config.tensor_parallel_process_group, ) else: loss = loss_fct(shift_logits, shift_labels) @@ -733,27 +796,18 @@ def gpt2_for_sequence_classification_forward( def get_gpt2_flash_attention_forward(): from transformers.models.gpt2.modeling_gpt2 import GPT2Attention - from colossalai.nn.layer.colo_attention import AttnMaskType, ColoAttention - - def split_heads(tensor, num_heads, attn_head_size): - """ - Splits hidden_size dim into attn_head_size and num_heads - """ - new_shape = tensor.size()[:-1] + (num_heads, attn_head_size) - tensor = tensor.view(new_shape) - return tensor - def forward( self: GPT2Attention, hidden_states: Optional[Tuple[torch.FloatTensor]], layer_past: Optional[Tuple[torch.Tensor]] = None, - attention_mask: Optional[torch.FloatTensor] = None, + attention_mask: Optional[dict] = None, head_mask: Optional[torch.FloatTensor] = None, encoder_hidden_states: Optional[torch.Tensor] = None, - encoder_attention_mask: Optional[torch.FloatTensor] = None, + encoder_attention_mask: Optional[dict] = None, use_cache: Optional[bool] = False, output_attentions: Optional[bool] = False, ) -> Tuple[Union[torch.Tensor, Tuple[torch.Tensor]], ...]: + assert head_mask is None, "FlashAttention does not support head_mask" if encoder_hidden_states is not None: if not hasattr(self, "q_attn"): raise ValueError( @@ -766,10 +820,9 @@ def forward( attention_mask = encoder_attention_mask else: query, key, value = self.c_attn(hidden_states).split(self.split_size, dim=2) - - query = split_heads(query, self.num_heads, self.head_dim) - key = split_heads(key, self.num_heads, self.head_dim) - value = split_heads(value, self.num_heads, self.head_dim) + query = self._split_heads(query, self.num_heads, self.head_dim) + key = self._split_heads(key, self.num_heads, self.head_dim) + value = self._split_heads(value, self.num_heads, self.head_dim) if layer_past is not None: past_key, past_value = layer_past @@ -781,29 +834,14 @@ def forward( else: present = None - if not self.is_cross_attention: - attn_mask_type = AttnMaskType.causal - flash_attention_mask = None - if attention_mask != None: - flash_attention_mask = ~(attention_mask[:, :, -1].squeeze(1).to(torch.bool)).contiguous() - if not torch.all(flash_attention_mask): - if attn_mask_type == AttnMaskType.causal: - attn_mask_type == AttnMaskType.paddedcausal - else: - attn_mask_type = AttnMaskType.padding - - scale = value.size(-1) ** -0.5 + scale = 1.0 + if self.scale_attn_weights: + scale /= value.size(-1) ** 0.5 if self.scale_attn_by_inverse_layer_idx: - scale = scale * (1 / float(self.layer_idx + 1)) - - # use coloattention - if not hasattr(self, "attention"): - self.attention = ColoAttention( - embed_dim=self.embed_dim, num_heads=self.num_heads, dropout=self.attn_dropout.p, scale=scale - ) - - attn_output = self.attention(query, key, value, attn_mask=flash_attention_mask, attn_mask_type=attn_mask_type) - + scale /= float(self.layer_idx + 1) + dropout_p = self.attn_dropout.p if self.training else 0.0 + attn_output = ColoAttention.attention(query, key, value, **attention_mask, dropout_p=dropout_p, scale=scale) + attn_output = self._merge_heads(attn_output, self.num_heads, self.head_dim) attn_output = self.c_proj(attn_output) attn_output = self.resid_dropout(attn_output) outputs = (attn_output, present, None) @@ -813,9 +851,9 @@ def forward( return forward -def gpt2_sequence_parallel_forward_fn(shard_config: ShardConfig): +def get_gpt_model_forward_for_flash_attn(shard_config: ShardConfig): def forward( - self, + self: GPT2Model, input_ids: Optional[torch.LongTensor] = None, past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None, attention_mask: Optional[torch.FloatTensor] = None, @@ -840,12 +878,13 @@ def forward( if input_ids is not None and inputs_embeds is not None: raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time") elif input_ids is not None: + self.warn_if_padding_and_no_attention_mask(input_ids, attention_mask) input_shape = input_ids.size() input_ids = input_ids.view(-1, input_shape[-1]) - batch_size = input_ids.shape[0] + input_ids.shape[0] elif inputs_embeds is not None: input_shape = inputs_embeds.size()[:-1] - batch_size = inputs_embeds.shape[0] + inputs_embeds.shape[0] else: raise ValueError("You have to specify either input_ids or inputs_embeds") @@ -862,39 +901,201 @@ def forward( else: past_length = past_key_values[0][0].size(-2) if position_ids is None: - position_ids = torch.arange(past_length, input_shape[-1] + past_length, dtype=torch.long, device=device) + position_ids = torch.arange( + past_length, + input_shape[-1] + past_length, + dtype=torch.long, + device=device, + ) position_ids = position_ids.unsqueeze(0).view(-1, input_shape[-1]) - # GPT2Attention mask. - if attention_mask is not None: - if batch_size <= 0: - raise ValueError("batch_size has to be defined and > 0") - attention_mask = attention_mask.view(batch_size, -1) - # We create a 3D attention mask from a 2D tensor mask. - # Sizes are [batch_size, 1, 1, to_seq_length] - # So we can broadcast to [batch_size, num_heads, from_seq_length, to_seq_length] - # this attention mask is more simple than the triangular masking of causal attention - # used in OpenAI GPT, we just need to prepare the broadcast dimension here. - attention_mask = attention_mask[:, None, None, :] - - # Since attention_mask is 1.0 for positions we want to attend and 0.0 for - # masked positions, this operation will create a tensor which is 0.0 for - # positions we want to attend and the dtype's smallest value for masked positions. - # Since we are adding it to the raw scores before the softmax, this is - # effectively the same as removing these entirely. - attention_mask = attention_mask.to(dtype=self.dtype) # fp16 compatibility - attention_mask = (1.0 - attention_mask) * torch.finfo(self.dtype).min - - # If a 2D or 3D attention mask is provided for the cross-attention - # we need to make broadcastable to [batch_size, num_heads, seq_length, seq_length] - if self.config.add_cross_attention and encoder_hidden_states is not None: - encoder_batch_size, encoder_sequence_length, _ = encoder_hidden_states.size() - encoder_hidden_shape = (encoder_batch_size, encoder_sequence_length) - if encoder_attention_mask is None: - encoder_attention_mask = torch.ones(encoder_hidden_shape, device=device) - encoder_attention_mask = self.invert_attention_mask(encoder_attention_mask) + # Prepare head mask if needed + # 1.0 in head_mask indicate we keep the head + # attention_probs has shape bsz x n_heads x N x N + # head_mask has shape n_layer x batch x n_heads x N x N + head_mask = self.get_head_mask(head_mask, self.config.n_layer) + + if inputs_embeds is None: + inputs_embeds = self.wte(input_ids) + position_embeds = self.wpe(position_ids) + hidden_states = inputs_embeds + position_embeds + + if token_type_ids is not None: + token_type_embeds = self.wte(token_type_ids) + hidden_states = hidden_states + token_type_embeds + + hidden_states = self.drop(hidden_states) + + output_shape = (-1,) + input_shape[1:] + (hidden_states.size(-1),) + + attention_mask, encoder_attention_mask = _get_attention_mask( + self, + shard_config, + hidden_states, + past_key_values, + attention_mask, + encoder_hidden_states, + encoder_attention_mask, + ) + + if self.gradient_checkpointing and self.training: + if use_cache: + logger.warning_once( + "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..." + ) + use_cache = False + + presents = () if use_cache else None + all_self_attentions = () if output_attentions else None + all_cross_attentions = () if output_attentions and self.config.add_cross_attention else None + all_hidden_states = () if output_hidden_states else None + for i, (block, layer_past) in enumerate(zip(self.h, past_key_values)): + # Model parallel + if self.model_parallel: + torch.cuda.set_device(hidden_states.device) + # Ensure layer_past is on same device as hidden_states (might not be correct) + if layer_past is not None: + layer_past = tuple(past_state.to(hidden_states.device) for past_state in layer_past) + # Ensure that attention_mask is always on the same device as hidden_states + if torch.is_tensor(attention_mask): + attention_mask = attention_mask.to(hidden_states.device) + if isinstance(head_mask, torch.Tensor): + head_mask = head_mask.to(hidden_states.device) + if output_hidden_states: + all_hidden_states = all_hidden_states + (hidden_states,) + + if self.gradient_checkpointing and self.training: + + def create_custom_forward(module): + def custom_forward(*inputs): + # None for past_key_value + return module(*inputs, use_cache, output_attentions) + + return custom_forward + + outputs = torch.utils.checkpoint.checkpoint( + create_custom_forward(block), + hidden_states, + None, + attention_mask, + head_mask[i], + encoder_hidden_states, + encoder_attention_mask, + ) + else: + outputs = block( + hidden_states, + layer_past=layer_past, + attention_mask=attention_mask, + head_mask=head_mask[i], + encoder_hidden_states=encoder_hidden_states, + encoder_attention_mask=encoder_attention_mask, + use_cache=use_cache, + output_attentions=output_attentions, + ) + + hidden_states = outputs[0] + if use_cache is True: + presents = presents + (outputs[1],) + + if output_attentions: + all_self_attentions = all_self_attentions + (outputs[2 if use_cache else 1],) + if self.config.add_cross_attention: + all_cross_attentions = all_cross_attentions + (outputs[3 if use_cache else 2],) + + # Model Parallel: If it's the last layer for that device, put things on the next device + if self.model_parallel: + for k, v in self.device_map.items(): + if i == v[-1] and "cuda:" + str(k) != self.last_device: + hidden_states = hidden_states.to("cuda:" + str(k + 1)) + + hidden_states = self.ln_f(hidden_states) + + hidden_states = hidden_states.view(output_shape) + # Add last hidden state + if output_hidden_states: + all_hidden_states = all_hidden_states + (hidden_states,) + + if not return_dict: + return tuple( + v + for v in [ + hidden_states, + presents, + all_hidden_states, + all_self_attentions, + all_cross_attentions, + ] + if v is not None + ) + + return BaseModelOutputWithPastAndCrossAttentions( + last_hidden_state=hidden_states, + past_key_values=presents, + hidden_states=all_hidden_states, + attentions=all_self_attentions, + cross_attentions=all_cross_attentions, + ) + + return forward + + +def gpt2_sequence_parallel_forward_fn(shard_config: ShardConfig): + def forward( + self, + input_ids: Optional[torch.LongTensor] = None, + past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None, + attention_mask: Optional[torch.FloatTensor] = None, + token_type_ids: Optional[torch.LongTensor] = None, + position_ids: Optional[torch.LongTensor] = None, + head_mask: Optional[torch.FloatTensor] = None, + inputs_embeds: Optional[torch.FloatTensor] = None, + encoder_hidden_states: Optional[torch.Tensor] = None, + encoder_attention_mask: Optional[torch.FloatTensor] = None, + use_cache: Optional[bool] = None, + output_attentions: Optional[bool] = None, + output_hidden_states: Optional[bool] = None, + return_dict: Optional[bool] = None, + ) -> Union[Tuple, BaseModelOutputWithPastAndCrossAttentions]: + output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions + output_hidden_states = ( + output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states + ) + use_cache = use_cache if use_cache is not None else self.config.use_cache + return_dict = return_dict if return_dict is not None else self.config.use_return_dict + + if input_ids is not None and inputs_embeds is not None: + raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time") + elif input_ids is not None: + input_shape = input_ids.size() + input_ids = input_ids.view(-1, input_shape[-1]) + input_ids.shape[0] + elif inputs_embeds is not None: + input_shape = inputs_embeds.size()[:-1] + inputs_embeds.shape[0] else: - encoder_attention_mask = None + raise ValueError("You have to specify either input_ids or inputs_embeds") + + device = input_ids.device if input_ids is not None else inputs_embeds.device + + if token_type_ids is not None: + token_type_ids = token_type_ids.view(-1, input_shape[-1]) + if position_ids is not None: + position_ids = position_ids.view(-1, input_shape[-1]) + + if past_key_values is None: + past_length = 0 + past_key_values = tuple([None] * len(self.h)) + else: + past_length = past_key_values[0][0].size(-2) + if position_ids is None: + position_ids = torch.arange( + past_length, + input_shape[-1] + past_length, + dtype=torch.long, + device=device, + ) + position_ids = position_ids.unsqueeze(0).view(-1, input_shape[-1]) # Prepare head mask if needed # 1.0 in head_mask indicate we keep the head @@ -914,6 +1115,15 @@ def forward( hidden_states = self.drop(hidden_states) output_shape = input_shape + (hidden_states.size(-1),) + attention_mask, encoder_attention_mask = _get_attention_mask( + self, + shard_config, + hidden_states, + past_key_values, + attention_mask, + encoder_hidden_states, + encoder_attention_mask, + ) if self.gradient_checkpointing and self.training: if use_cache: @@ -931,7 +1141,9 @@ def forward( # split the input tensor along sequence dimension # [batch_size, seq_len, hidden_size] -> [batch_size, seq_len/TP_size, hidden_size] hidden_states = split_forward_gather_backward( - hidden_states, dim=1, process_group=shard_config.tensor_parallel_process_group + hidden_states, + dim=1, + process_group=shard_config.sequence_parallel_process_group, ) for i, (block, layer_past) in enumerate(zip(self.h, past_key_values)): @@ -942,7 +1154,7 @@ def forward( if layer_past is not None: layer_past = tuple(past_state.to(hidden_states.device) for past_state in layer_past) # Ensure that attention_mask is always on the same device as hidden_states - if attention_mask is not None: + if torch.is_tensor(attention_mask): attention_mask = attention_mask.to(hidden_states.device) if isinstance(head_mask, torch.Tensor): head_mask = head_mask.to(hidden_states.device) @@ -996,7 +1208,9 @@ def custom_forward(*inputs): # When sequence parallelism done, gather the output tensor in forward and split it in backward hidden_states = gather_forward_split_backward( - hidden_states, dim=1, process_group=shard_config.tensor_parallel_process_group + hidden_states, + dim=1, + process_group=shard_config.sequence_parallel_process_group, ) hidden_states = self.ln_f(hidden_states) @@ -1008,7 +1222,13 @@ def custom_forward(*inputs): if not return_dict: return tuple( v - for v in [hidden_states, presents, all_hidden_states, all_self_attentions, all_cross_attentions] + for v in [ + hidden_states, + presents, + all_hidden_states, + all_self_attentions, + all_cross_attentions, + ] if v is not None ) @@ -1078,15 +1298,11 @@ def forward( shift_logits = lm_logits[..., :-1, :].contiguous() shift_labels = labels[..., 1:].contiguous() # Flatten the tokens - loss_fct = CrossEntropyLoss() shift_logits = shift_logits.view(-1, shift_logits.size(-1)) shift_labels = shift_labels.view(-1) - if shard_config.enable_tensor_parallelism: - loss = cross_entropy_1d( - shift_logits, shift_labels, process_group=shard_config.tensor_parallel_process_group - ) - else: - loss = loss_fct(shift_logits, shift_labels) + loss = cross_entropy_1d( + shift_logits, shift_labels, process_group=shard_config.tensor_parallel_process_group + ) if not shard_config.parallel_output: lm_logits = gather_forward_split_backward(lm_logits, -1, shard_config.tensor_parallel_process_group) diff --git a/colossalai/shardformer/modeling/gptj.py b/colossalai/shardformer/modeling/gptj.py index 1990d7df3279..5c254d1e76bd 100644 --- a/colossalai/shardformer/modeling/gptj.py +++ b/colossalai/shardformer/modeling/gptj.py @@ -19,9 +19,54 @@ from transformers.utils import is_torch_fx_proxy, logging from colossalai.pipeline.stage_manager import PipelineStageManager +from colossalai.shardformer.layer import ColoAttention from colossalai.shardformer.layer._operation import gather_forward_split_backward, split_forward_gather_backward from colossalai.shardformer.shard import ShardConfig +logger = logging.get_logger(__name__) + + +def _get_attention_mask( + self: GPTJModel, + shard_config: ShardConfig, + hidden_states: torch.Tensor, + past_key_values: Optional[Tuple[Tuple[torch.Tensor]]], + attention_mask: Optional[torch.FloatTensor], +) -> Optional[Union[torch.Tensor, dict]]: + batch_size, seq_len = hidden_states.shape[:2] + past_key_values_length = 0 + if past_key_values is not None and past_key_values[0] is not None: + past_key_values_length = past_key_values[0][0].shape[2] + if shard_config.enable_flash_attention: + if attention_mask is not None: + attention_mask = attention_mask.view(batch_size, -1) + attention_mask = ColoAttention.prepare_attn_kwargs( + (batch_size, 1, seq_len, seq_len + past_key_values_length), + hidden_states.dtype, + hidden_states.device, + attention_mask, + is_causal=True, + ) + elif attention_mask is not None: + if batch_size <= 0: + raise ValueError("batch_size has to be defined and > 0") + attention_mask = attention_mask.view(batch_size, -1) + # We create a 3D attention mask from a 2D tensor mask. + # Sizes are [batch_size, 1, 1, to_seq_length] + # So we can broadcast to [batch_size, num_heads, from_seq_length, to_seq_length] + # this attention mask is more simple than the triangular masking of causal attention + # used in OpenAI GPT, we just need to prepare the broadcast dimension here. + attention_mask = attention_mask[:, None, None, :] + + # Since attention_mask is 1.0 for positions we want to attend and 0.0 for + # masked positions, this operation will create a tensor which is 0.0 for + # positions we want to attend and the dtype's smallest value for masked positions. + # Since we are adding it to the raw scores before the softmax, this is + # effectively the same as removing these entirely. + attention_mask = attention_mask.to(dtype=self.dtype) # fp16 compatibility + attention_mask = (1.0 - attention_mask) * torch.finfo(self.dtype).min + return attention_mask + class GPTJPipelineForwards: """ @@ -96,26 +141,6 @@ def gptj_model_forward( batch_size, seq_length = input_shape[0], input_shape[1] device = hidden_states.device - # Attention mask. - if attention_mask is not None: - if batch_size <= 0: - raise ValueError("batch_size has to be defined and > 0") - attention_mask = attention_mask.view(batch_size, -1) - # We create a 3D attention mask from a 2D tensor mask. - # Sizes are [batch_size, 1, 1, to_seq_length] - # So we can broadcast to [batch_size, num_heads, from_seq_length, to_seq_length] - # this attention mask is more simple than the triangular masking of causal attention - # used in OpenAI GPT, we just need to prepare the broadcast dimension here. - attention_mask = attention_mask[:, None, None, :] - - # Since attention_mask is 1.0 for positions we want to attend and 0.0 for - # masked positions, this operation will create a tensor which is 0.0 for - # positions we want to attend and the dtype's smallest value for masked positions. - # Since we are adding it to the raw scores before the softmax, this is - # effectively the same as removing these entirely. - attention_mask = attention_mask.to(dtype=self.dtype) # fp16 compatibility - attention_mask = (1.0 - attention_mask) * torch.finfo(self.dtype).min - # Prepare head mask if needed # 1.0 in head_mask indicate we keep the head # attention_probs has shape bsz x num_attention_heads x N x N @@ -139,6 +164,8 @@ def gptj_model_forward( output_shape = input_shape + (hidden_states.size(-1),) + attention_mask = _get_attention_mask(self, shard_config, hidden_states, past_key_values, attention_mask) + if self.gradient_checkpointing and self.training: if use_cache: logger.warning_once( @@ -154,7 +181,9 @@ def gptj_model_forward( # [batch_size, seq_len, hidden_size] -> [batch_size, seq_len/TP_size, hidden_size] if shard_config.enable_sequence_parallelism: hidden_states = split_forward_gather_backward( - hidden_states, dim=1, process_group=shard_config.tensor_parallel_process_group + hidden_states, + dim=1, + process_group=shard_config.tensor_parallel_process_group, ) # Going through held blocks. @@ -209,7 +238,9 @@ def custom_forward(*inputs): # When sequence parallelism done, gather the output tensor in forward and split it in backward if shard_config.enable_sequence_parallelism: hidden_states = gather_forward_split_backward( - hidden_states, dim=1, process_group=shard_config.tensor_parallel_process_group + hidden_states, + dim=1, + process_group=shard_config.tensor_parallel_process_group, ) if stage_manager.is_last_stage(): @@ -223,7 +254,14 @@ def custom_forward(*inputs): if stage_manager.is_last_stage(): if not return_dict: return tuple( - v for v in [hidden_states, presents, all_hidden_states, all_self_attentions] if v is not None + v + for v in [ + hidden_states, + presents, + all_hidden_states, + all_self_attentions, + ] + if v is not None ) return BaseModelOutputWithPast( @@ -530,24 +568,11 @@ def gptj_for_question_answering_forward( def get_gptj_flash_attention_forward(): from transformers.models.gptj.modeling_gptj import GPTJAttention - from colossalai.nn.layer.colo_attention import AttnMaskType, ColoAttention - - def split_heads(tensor, num_attention_heads, attn_head_size, rotary): - """ - Splits hidden dim into attn_head_size and num_attention_heads - """ - new_shape = tensor.size()[:-1] + (num_attention_heads, attn_head_size) - tensor = tensor.view(new_shape) - if rotary or len(tensor.shape) in [4, 5]: - return tensor - else: - raise ValueError(f"Input tensor rank should be one of [4, 5], but is: {len(tensor.shape)}") - def forward( self: GPTJAttention, hidden_states: torch.FloatTensor, layer_past: Optional[Tuple[torch.Tensor]] = None, - attention_mask: Optional[torch.FloatTensor] = None, + attention_mask: Optional[dict] = None, position_ids: Optional[torch.LongTensor] = None, head_mask: Optional[torch.FloatTensor] = None, use_cache: Optional[bool] = False, @@ -556,13 +581,14 @@ def forward( Tuple[torch.Tensor, Tuple[torch.Tensor]], Optional[Tuple[torch.Tensor, Tuple[torch.Tensor], Tuple[torch.Tensor, ...]]], ]: + assert head_mask is None, "head_mask is not supported for FlashAttention" query = self.q_proj(hidden_states) key = self.k_proj(hidden_states) value = self.v_proj(hidden_states) - query = split_heads(query, self.num_attention_heads, self.head_dim, True) - key = split_heads(key, self.num_attention_heads, self.head_dim, True) - value = split_heads(value, self.num_attention_heads, self.head_dim, False) + query = self._split_heads(query, self.num_attention_heads, self.head_dim, True) + key = self._split_heads(key, self.num_attention_heads, self.head_dim, True) + value = self._split_heads(value, self.num_attention_heads, self.head_dim, False) if is_torch_fx_proxy(position_ids) or torch.jit.is_tracing(): # The logic to conditionally copy to GPU could not be traced, so we do this @@ -591,46 +617,202 @@ def forward( key = apply_rotary_pos_emb(key, sin, cos) query = apply_rotary_pos_emb(query, sin, cos) - # key = key.permute(0, 2, 1, 3) - # query = query.permute(0, 2, 1, 3) - key = key.to(dtype=value.dtype) # fp16 compatibility - query = query.to(dtype=value.dtype) + key = key.permute(0, 2, 1, 3) + query = query.permute(0, 2, 1, 3) if layer_past is not None: past_key = layer_past[0] past_value = layer_past[1] - key = torch.cat((past_key, key), dim=1) - value = torch.cat((past_value, value), dim=1) + key = torch.cat((past_key, key), dim=-2) + value = torch.cat((past_value, value), dim=-2) if use_cache is True: present = (key, value) else: present = None - # use AttnMaskType and ColoAttention - attn_mask_type = AttnMaskType.causal - flash_attention_mask = None - if attention_mask != None: - if attn_mask_type == AttnMaskType.causal: - attn_mask_type == AttnMaskType.paddedcausal - else: - attn_mask_type = AttnMaskType.padding - flash_attention_mask = ~(attention_mask[:, :, -1].squeeze(1).to(torch.bool)).contiguous() + dropout_p = self.attn_dropout.p if self.training else 0.0 + attn_output = ColoAttention.attention(query, key, value, **attention_mask, dropout_p=dropout_p) + attn_output = self._merge_heads(attn_output, self.num_attention_heads, self.head_dim) + attn_output = self.out_proj(attn_output) + attn_output = self.resid_dropout(attn_output) + outputs = (attn_output, present, None) - # use coloattention - scale = value.size(-1) ** -0.5 + return outputs # a, present, (attentions) + + return forward - attention = ColoAttention( - embed_dim=self.embed_dim, num_heads=self.num_attention_heads, dropout=self.attn_dropout.p, scale=scale + +def gptj_model_forward_for_flash_attention(shard_config: ShardConfig): + def forward( + self, + input_ids: Optional[torch.LongTensor] = None, + past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None, + attention_mask: Optional[torch.FloatTensor] = None, + token_type_ids: Optional[torch.LongTensor] = None, + position_ids: Optional[torch.LongTensor] = None, + head_mask: Optional[torch.FloatTensor] = None, + inputs_embeds: Optional[torch.FloatTensor] = None, + use_cache: Optional[bool] = None, + output_attentions: Optional[bool] = None, + output_hidden_states: Optional[bool] = None, + return_dict: Optional[bool] = None, + ) -> Union[Tuple, BaseModelOutputWithPast]: + output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions + output_hidden_states = ( + output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states ) + use_cache = use_cache if use_cache is not None else self.config.use_cache + return_dict = return_dict if return_dict is not None else self.config.use_return_dict + + if input_ids is not None and inputs_embeds is not None: + raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time") + elif input_ids is not None: + self.warn_if_padding_and_no_attention_mask(input_ids, attention_mask) + input_shape = input_ids.size() + input_ids = input_ids.view(-1, input_shape[-1]) + input_ids.shape[0] + elif inputs_embeds is not None: + input_shape = inputs_embeds.size()[:-1] + inputs_embeds.shape[0] + else: + raise ValueError("You have to specify either input_ids or inputs_embeds") - attn_output = attention(query, key, value, attn_mask=flash_attention_mask, attn_mask_type=attn_mask_type) + device = input_ids.device if input_ids is not None else inputs_embeds.device - attn_output = self.out_proj(attn_output) - attn_output = self.resid_dropout(attn_output) - outputs = (attn_output, present, None) + if token_type_ids is not None: + token_type_ids = token_type_ids.view(-1, input_shape[-1]) - return outputs # a, present, (attentions) + if position_ids is not None: + position_ids = position_ids.view(-1, input_shape[-1]).long() + + if past_key_values is None: + past_length = 0 + past_key_values = tuple([None] * len(self.h)) + else: + past_length = past_key_values[0][0].size(-2) + + if position_ids is None: + position_ids = torch.arange( + past_length, + input_shape[-1] + past_length, + dtype=torch.long, + device=device, + ) + position_ids = position_ids.unsqueeze(0).view(-1, input_shape[-1]) + + # Prepare head mask if needed + # 1.0 in head_mask indicate we keep the head + # attention_probs has shape bsz x num_attention_heads x N x N + # head_mask has shape n_layer x batch x num_attention_heads x N x N + head_mask = self.get_head_mask(head_mask, self.config.n_layer) + + if inputs_embeds is None: + inputs_embeds = self.wte(input_ids) + + hidden_states = inputs_embeds + + if token_type_ids is not None: + token_type_embeds = self.wte(token_type_ids) + hidden_states = hidden_states + token_type_embeds + + hidden_states = self.drop(hidden_states) + + attention_mask = _get_attention_mask(self, shard_config, hidden_states, past_key_values, attention_mask) + + output_shape = (-1,) + input_shape[1:] + (hidden_states.size(-1),) + + if self.gradient_checkpointing and self.training: + if use_cache: + logger.warning_once( + "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..." + ) + use_cache = False + + presents = () if use_cache else None + all_self_attentions = () if output_attentions else None + all_hidden_states = () if output_hidden_states else None + for i, (block, layer_past) in enumerate(zip(self.h, past_key_values)): + # Model parallel + if self.model_parallel: + torch.cuda.set_device(hidden_states.device) + # Ensure layer_past is on same device as hidden_states (might not be correct) + if layer_past is not None: + layer_past = tuple(past_state.to(hidden_states.device) for past_state in layer_past) + # Ensure that attention_mask is always on the same device as hidden_states + if attention_mask is not None: + attention_mask = attention_mask.to(hidden_states.device) + if isinstance(head_mask, torch.Tensor): + head_mask = head_mask.to(hidden_states.device) + if output_hidden_states: + all_hidden_states = all_hidden_states + (hidden_states,) + + if self.gradient_checkpointing and self.training: + + def create_custom_forward(module): + def custom_forward(*inputs): + # None for past_key_value + return module(*inputs, use_cache, output_attentions) + + return custom_forward + + outputs = torch.utils.checkpoint.checkpoint( + create_custom_forward(block), + hidden_states, + None, + attention_mask, + position_ids, + head_mask[i], + ) + else: + outputs = block( + hidden_states=hidden_states, + layer_past=layer_past, + attention_mask=attention_mask, + position_ids=position_ids, + head_mask=head_mask[i], + use_cache=use_cache, + output_attentions=output_attentions, + ) + + hidden_states = outputs[0] + if use_cache is True: + presents = presents + (outputs[1],) + + if output_attentions: + all_self_attentions = all_self_attentions + (outputs[2 if use_cache else 1],) + + # Model Parallel: If it's the last layer for that device, put things on the next device + if self.model_parallel: + for k, v in self.device_map.items(): + if i == v[-1] and "cuda:" + str(k) != self.last_device: + hidden_states = hidden_states.to("cuda:" + str(k + 1)) + + hidden_states = self.ln_f(hidden_states) + + hidden_states = hidden_states.view(output_shape) + # Add last hidden state + if output_hidden_states: + all_hidden_states = all_hidden_states + (hidden_states,) + + if not return_dict: + return tuple( + v + for v in [ + hidden_states, + presents, + all_hidden_states, + all_self_attentions, + ] + if v is not None + ) + + return BaseModelOutputWithPast( + last_hidden_state=hidden_states, + past_key_values=presents, + hidden_states=all_hidden_states, + attentions=all_self_attentions, + ) return forward @@ -662,10 +844,10 @@ def forward( elif input_ids is not None: input_shape = input_ids.size() input_ids = input_ids.view(-1, input_shape[-1]) - batch_size = input_ids.shape[0] + input_ids.shape[0] elif inputs_embeds is not None: input_shape = inputs_embeds.size()[:-1] - batch_size = inputs_embeds.shape[0] + inputs_embeds.shape[0] else: raise ValueError("You have to specify either input_ids or inputs_embeds") @@ -684,29 +866,14 @@ def forward( past_length = past_key_values[0][0].size(-2) if position_ids is None: - position_ids = torch.arange(past_length, input_shape[-1] + past_length, dtype=torch.long, device=device) + position_ids = torch.arange( + past_length, + input_shape[-1] + past_length, + dtype=torch.long, + device=device, + ) position_ids = position_ids.unsqueeze(0).view(-1, input_shape[-1]) - # Attention mask. - if attention_mask is not None: - if batch_size <= 0: - raise ValueError("batch_size has to be defined and > 0") - attention_mask = attention_mask.view(batch_size, -1) - # We create a 3D attention mask from a 2D tensor mask. - # Sizes are [batch_size, 1, 1, to_seq_length] - # So we can broadcast to [batch_size, num_heads, from_seq_length, to_seq_length] - # this attention mask is more simple than the triangular masking of causal attention - # used in OpenAI GPT, we just need to prepare the broadcast dimension here. - attention_mask = attention_mask[:, None, None, :] - - # Since attention_mask is 1.0 for positions we want to attend and 0.0 for - # masked positions, this operation will create a tensor which is 0.0 for - # positions we want to attend and the dtype's smallest value for masked positions. - # Since we are adding it to the raw scores before the softmax, this is - # effectively the same as removing these entirely. - attention_mask = attention_mask.to(dtype=self.dtype) # fp16 compatibility - attention_mask = (1.0 - attention_mask) * torch.finfo(self.dtype).min - # Prepare head mask if needed # 1.0 in head_mask indicate we keep the head # attention_probs has shape bsz x num_attention_heads x N x N @@ -725,6 +892,7 @@ def forward( hidden_states = self.drop(hidden_states) output_shape = input_shape + (hidden_states.size(-1),) + attention_mask = _get_attention_mask(self, shard_config, hidden_states, past_key_values, attention_mask) if self.gradient_checkpointing and self.training: if use_cache: @@ -740,7 +908,9 @@ def forward( # split the input tensor along sequence dimension # [batch_size, seq_len, hidden_size] -> [batch_size, seq_len/TP_size, hidden_size] hidden_states = split_forward_gather_backward( - hidden_states, dim=1, process_group=shard_config.tensor_parallel_process_group + hidden_states, + dim=1, + process_group=shard_config.tensor_parallel_process_group, ) for i, (block, layer_past) in enumerate(zip(self.h, past_key_values)): @@ -801,7 +971,9 @@ def custom_forward(*inputs): # When sequence parallelism done, gather the output tensor in forward and split it in backward hidden_states = gather_forward_split_backward( - hidden_states, dim=1, process_group=shard_config.tensor_parallel_process_group + hidden_states, + dim=1, + process_group=shard_config.tensor_parallel_process_group, ) hidden_states = self.ln_f(hidden_states) @@ -812,7 +984,16 @@ def custom_forward(*inputs): all_hidden_states = all_hidden_states + (hidden_states,) if not return_dict: - return tuple(v for v in [hidden_states, presents, all_hidden_states, all_self_attentions] if v is not None) + return tuple( + v + for v in [ + hidden_states, + presents, + all_hidden_states, + all_self_attentions, + ] + if v is not None + ) return BaseModelOutputWithPast( last_hidden_state=hidden_states, diff --git a/colossalai/shardformer/modeling/llama.py b/colossalai/shardformer/modeling/llama.py index eb8e9f748527..0f1b4ad0a5d5 100644 --- a/colossalai/shardformer/modeling/llama.py +++ b/colossalai/shardformer/modeling/llama.py @@ -1,22 +1,35 @@ +import math import warnings from typing import List, Optional, Tuple, Union import torch import torch.nn.functional as F +import torch.utils.checkpoint +from torch import nn from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss from transformers.modeling_outputs import ( BaseModelOutputWithPast, CausalLMOutputWithPast, SequenceClassifierOutputWithPast, ) -from transformers.models.llama.modeling_llama import LlamaForCausalLM, LlamaForSequenceClassification, LlamaModel +from transformers.models.llama.modeling_llama import ( + LlamaForCausalLM, + LlamaForSequenceClassification, + LlamaModel, + apply_rotary_pos_emb, + repeat_kv, +) from transformers.utils import logging from colossalai.pipeline.stage_manager import PipelineStageManager +from colossalai.shardformer.layer._operation import ( + all_to_all_comm, + gather_forward_split_backward, + split_forward_gather_backward, +) from colossalai.shardformer.shard import ShardConfig -from ..layer import cross_entropy_1d -from ..layer._operation import gather_forward_split_backward +from ..layer import ColoAttention, cross_entropy_1d try: from transformers.models.llama.modeling_llama import _prepare_4d_causal_attention_mask @@ -106,18 +119,25 @@ def llama_model_forward( # embed positions, for the first stage, hidden_states is the input embeddings, # for the other stages, hidden_states is the output of the previous stage - if attention_mask is None: - attention_mask = torch.ones( - (batch_size, seq_length_with_past), dtype=torch.bool, device=hidden_states.device - ) - if LATEST_VERSION: - attention_mask = _prepare_4d_causal_attention_mask( - attention_mask, (batch_size, seq_length), hidden_states, past_key_values_length + if shard_config.enable_flash_attention: + # in this case, attention_mask is a dict rather than a tensor + mask_shape = (batch_size, 1, seq_length_with_past, seq_length_with_past) + attention_mask = ColoAttention.prepare_attn_kwargs( + mask_shape, hidden_states.dtype, hidden_states.device, q_padding_mask=attention_mask, is_causal=True ) else: - attention_mask = self._prepare_decoder_attention_mask( - attention_mask, (batch_size, seq_length), hidden_states, past_key_values_length - ) + if attention_mask is None: + attention_mask = torch.ones( + (batch_size, seq_length_with_past), dtype=torch.bool, device=hidden_states.device + ) + if LATEST_VERSION: + attention_mask = _prepare_4d_causal_attention_mask( + attention_mask, (batch_size, seq_length), hidden_states, past_key_values_length + ) + else: + attention_mask = self._prepare_decoder_attention_mask( + attention_mask, (batch_size, seq_length), hidden_states, past_key_values_length + ) if self.gradient_checkpointing and self.training: if use_cache: @@ -132,13 +152,25 @@ def llama_model_forward( next_decoder_cache = () if use_cache else None start_idx, end_idx = stage_index[0], stage_index[1] + num_ckpt_layers = 0 + if self.gradient_checkpointing and self.training: + num_ckpt_layers = end_idx - start_idx + # TODO: We can replace `gradient_checkpointing_enable` fn and initialize a gradient_checkpointing (List[bool]) for each layer + if shard_config.gradient_checkpoint_config is not None: + num_ckpt_layers = shard_config.gradient_checkpoint_config.get_num_ckpt_layers( + stage=stage_manager.stage, + num_layers=end_idx - start_idx, + model_chunk_id=stage_manager.model_chunk_id if stage_manager.is_interleave else 0, + ) + assert num_ckpt_layers <= end_idx - start_idx + for idx, decoder_layer in enumerate(self.layers[start_idx:end_idx], start=start_idx): if output_hidden_states: all_hidden_states += (hidden_states,) past_key_value = past_key_values[idx] if past_key_values is not None else None - if self.gradient_checkpointing and self.training: + if idx - start_idx < num_ckpt_layers: def create_custom_forward(module): def custom_forward(*inputs): @@ -263,6 +295,7 @@ def llama_for_causal_lm_forward( stage_manager=stage_manager, hidden_states=hidden_states, stage_index=stage_index, + shard_config=shard_config, ) past_key_values = None @@ -279,7 +312,7 @@ def llama_for_causal_lm_forward( shift_labels = shift_labels.view(-1) # Enable model parallelism shift_labels = shift_labels.to(shift_logits.device) - if shard_config.enable_tensor_parallelism: + if shard_config.enable_tensor_parallelism and shard_config.parallel_output: new_vocab_size = logits.shape[-1] shift_logits = shift_logits.view(-1, new_vocab_size) loss = cross_entropy_1d( @@ -289,9 +322,6 @@ def llama_for_causal_lm_forward( shift_logits = shift_logits.view(-1, self.config.vocab_size) loss = loss_fct(shift_logits, shift_labels) - if not shard_config.parallel_output: - logits = gather_forward_split_backward(logits, -1, shard_config.tensor_parallel_process_group) - if not return_dict: output = (logits,) + outputs[1:] return (loss,) + output if loss is not None else output @@ -356,6 +386,7 @@ def llama_for_sequence_classification_forward( stage_manager=stage_manager, hidden_states=hidden_states, stage_index=stage_index, + shard_config=shard_config, ) if input_ids is not None: @@ -421,11 +452,9 @@ def llama_for_sequence_classification_forward( return {"hidden_states": hidden_states} -def get_llama_flash_attention_forward(shard_config: ShardConfig): +def get_llama_flash_attention_forward(shard_config, sp_mode, sp_group, sp_size): from transformers.models.llama.modeling_llama import LlamaAttention, apply_rotary_pos_emb - from colossalai.nn.layer.colo_attention import AttnMaskType, ColoAttention - llama_version = 2 try: from transformers.models.llama.modeling_llama import repeat_kv @@ -436,7 +465,7 @@ def get_llama_flash_attention_forward(shard_config: ShardConfig): def forward( self: LlamaAttention, hidden_states: torch.Tensor, - attention_mask: Optional[torch.Tensor] = None, + attention_mask: Optional[dict] = None, position_ids: Optional[torch.LongTensor] = None, past_key_value: Optional[Tuple[torch.Tensor]] = None, output_attentions: bool = False, @@ -444,18 +473,30 @@ def forward( **kwargs, ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]: bsz, q_len, _ = hidden_states.size() + + if sp_mode in ["split_gather", "ring"]: + q_len *= sp_size assert q_len % 4 == 0, "Flash Attention Error: The sequence length should be a multiple of 4." - query_states = self.q_proj(hidden_states).view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2) - key_states = self.k_proj(hidden_states).view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2) - value_states = self.v_proj(hidden_states).view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2) + query_states = self.q_proj(hidden_states) + key_states = self.k_proj(hidden_states) + value_states = self.v_proj(hidden_states) + + # sp: all-to-all comminucation when introducing sequence parallel + if sp_mode == "all_to_all": + query_states = all_to_all_comm(query_states, sp_group) + key_states = all_to_all_comm(key_states, sp_group) + value_states = all_to_all_comm(value_states, sp_group) + bsz, q_len, _ = query_states.size() + + query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2) + key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2) + value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2) kv_seq_len = key_states.shape[-2] if past_key_value is not None: kv_seq_len += past_key_value[0].shape[-2] - cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len) - query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids) if past_key_value is not None: @@ -470,32 +511,14 @@ def forward( key_states = repeat_kv(key_states, self.num_key_value_groups) value_states = repeat_kv(value_states, self.num_key_value_groups) - me_input_shape = (bsz, q_len, self.num_heads, self.head_dim) - query_states = query_states.transpose(1, 2).contiguous().view(*me_input_shape) - key_states = key_states.transpose(1, 2).contiguous().view(*me_input_shape) - value_states = value_states.transpose(1, 2).contiguous().view(*me_input_shape) - - flash_attention_mask = None - attn_mask_type = AttnMaskType.causal - if not getattr(shard_config, "causal_lm", False) and attention_mask != None: - if attention_mask.size() != (bsz, 1, q_len, kv_seq_len): - raise ValueError( - f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}" - ) - flash_attention_mask = ~(attention_mask[:, :, -1].squeeze(1).to(torch.bool)).contiguous() - attn_mask_type = AttnMaskType.paddedcausal - - if not hasattr(self, "attention"): - self.attention = ColoAttention(embed_dim=self.hidden_size, num_heads=self.num_heads) - attn_output = self.attention( - query_states, - key_states, - value_states, - attn_mask=flash_attention_mask, - attn_mask_type=attn_mask_type, - origin_attn_mask=attention_mask, - ) + assert isinstance(attention_mask, dict), "Flash Attention Error: attention_mask should be a dict." + attn_output = ColoAttention.attention(query_states, key_states, value_states, **attention_mask) + attn_output = attn_output.transpose(1, 2).contiguous() + attn_output = attn_output.reshape(bsz, q_len, self.hidden_size) + # sp: all-to-all comminucation when introducing sequence parallel + if sp_mode == "all_to_all": + attn_output = all_to_all_comm(attn_output, sp_group, scatter_dim=1, gather_dim=2) attn_output = self.o_proj(attn_output) return attn_output, None, past_key_value @@ -503,6 +526,137 @@ def forward( return forward +def get_llama_model_forward_for_flash_attn(shard_config: ShardConfig): + logger = logging.get_logger(__name__) + assert shard_config.enable_flash_attention, "Flash Attention is not enabled." + + def forward( + self: LlamaModel, + input_ids: torch.LongTensor = None, + attention_mask: Optional[torch.Tensor] = None, + position_ids: Optional[torch.LongTensor] = None, + past_key_values: Optional[List[torch.FloatTensor]] = None, + inputs_embeds: Optional[torch.FloatTensor] = None, + use_cache: Optional[bool] = None, + output_attentions: Optional[bool] = None, + output_hidden_states: Optional[bool] = None, + return_dict: Optional[bool] = None, + ) -> Union[Tuple, BaseModelOutputWithPast]: + output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions + output_hidden_states = ( + output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states + ) + use_cache = use_cache if use_cache is not None else self.config.use_cache + + return_dict = return_dict if return_dict is not None else self.config.use_return_dict + + # retrieve input_ids and inputs_embeds + if input_ids is not None and inputs_embeds is not None: + raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time") + elif input_ids is not None: + batch_size, seq_length = input_ids.shape + elif inputs_embeds is not None: + batch_size, seq_length, _ = inputs_embeds.shape + else: + raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds") + + seq_length_with_past = seq_length + past_key_values_length = 0 + + if past_key_values is not None: + past_key_values_length = past_key_values[0][0].shape[2] + seq_length_with_past = seq_length_with_past + past_key_values_length + + if position_ids is None: + device = input_ids.device if input_ids is not None else inputs_embeds.device + position_ids = torch.arange( + past_key_values_length, seq_length + past_key_values_length, dtype=torch.long, device=device + ) + position_ids = position_ids.unsqueeze(0).view(-1, seq_length) + else: + position_ids = position_ids.view(-1, seq_length).long() + + if inputs_embeds is None: + inputs_embeds = self.embed_tokens(input_ids) + # embed positions + hidden_states = inputs_embeds + + # in this case, attention_mask is a dict rather than a tensor + mask_shape = (batch_size, 1, seq_length_with_past, seq_length_with_past) + attention_mask = ColoAttention.prepare_attn_kwargs( + mask_shape, hidden_states.dtype, hidden_states.device, q_padding_mask=attention_mask, is_causal=True + ) + + if self.gradient_checkpointing and self.training: + if use_cache: + logger.warning_once( + "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..." + ) + use_cache = False + + # decoder layers + all_hidden_states = () if output_hidden_states else None + all_self_attns = () if output_attentions else None + next_decoder_cache = () if use_cache else None + + for idx, decoder_layer in enumerate(self.layers): + if output_hidden_states: + all_hidden_states += (hidden_states,) + + past_key_value = past_key_values[idx] if past_key_values is not None else None + + if self.gradient_checkpointing and self.training: + + def create_custom_forward(module): + def custom_forward(*inputs): + # None for past_key_value + return module(*inputs, past_key_value, output_attentions) + + return custom_forward + + layer_outputs = torch.utils.checkpoint.checkpoint( + create_custom_forward(decoder_layer), + hidden_states, + attention_mask, + position_ids, + ) + else: + layer_outputs = decoder_layer( + hidden_states, + attention_mask=attention_mask, + position_ids=position_ids, + past_key_value=past_key_value, + output_attentions=output_attentions, + use_cache=use_cache, + ) + + hidden_states = layer_outputs[0] + + if use_cache: + next_decoder_cache += (layer_outputs[2 if output_attentions else 1],) + + if output_attentions: + all_self_attns += (layer_outputs[1],) + + hidden_states = self.norm(hidden_states) + + # add hidden states from the last decoder layer + if output_hidden_states: + all_hidden_states += (hidden_states,) + + next_cache = next_decoder_cache if use_cache else None + if not return_dict: + return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None) + return BaseModelOutputWithPast( + last_hidden_state=hidden_states, + past_key_values=next_cache, + hidden_states=all_hidden_states, + attentions=all_self_attns, + ) + + return forward + + def get_lm_forward_with_dist_cross_entropy(shard_config: ShardConfig): from transformers import LlamaForCausalLM @@ -578,23 +732,15 @@ def forward( # Shift so that tokens < n predict n shift_logits = logits[..., :-1, :].contiguous() shift_labels = labels[..., 1:].contiguous() - # Flatten the tokens - loss_fct = CrossEntropyLoss() shift_labels = shift_labels.view(-1) # Enable model parallelism shift_labels = shift_labels.to(shift_logits.device) - if shard_config.enable_tensor_parallelism: - new_vocab_size = logits.shape[-1] - shift_logits = shift_logits.view(-1, new_vocab_size) - loss = cross_entropy_1d( - shift_logits, shift_labels, process_group=shard_config.tensor_parallel_process_group - ) - else: - shift_logits = shift_logits.view(-1, self.config.vocab_size) - loss = loss_fct(shift_logits, shift_labels) - if not shard_config.parallel_output: - logits = gather_forward_split_backward(logits, -1, shard_config.tensor_parallel_process_group) + new_vocab_size = logits.shape[-1] + shift_logits = shift_logits.view(-1, new_vocab_size) + loss = cross_entropy_1d( + shift_logits, shift_labels, process_group=shard_config.tensor_parallel_process_group + ) if not return_dict: output = (logits,) + outputs[1:] @@ -609,3 +755,261 @@ def forward( ) return forward + + +def get_llama_seq_parallel_attention_forward(sp_mode, sp_size, sp_group): + def forward( + self, + hidden_states: torch.Tensor, + attention_mask: Optional[torch.Tensor] = None, + position_ids: Optional[torch.LongTensor] = None, + past_key_value: Optional[Tuple[torch.Tensor]] = None, + output_attentions: bool = False, + use_cache: bool = False, + ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]: + bsz, q_len, _ = hidden_states.size() + # sp: modify sp_len when sequence parallel mode is ring + if sp_mode in ["split_gather", "ring"]: + q_len *= sp_size + if self.config.pretraining_tp > 1: + key_value_slicing = (self.num_key_value_heads * self.head_dim) // self.config.pretraining_tp + query_slices = self.q_proj.weight.split( + (self.num_heads * self.head_dim) // self.config.pretraining_tp, dim=0 + ) + key_slices = self.k_proj.weight.split(key_value_slicing, dim=0) + value_slices = self.v_proj.weight.split(key_value_slicing, dim=0) + + query_states = [F.linear(hidden_states, query_slices[i]) for i in range(self.config.pretraining_tp)] + query_states = torch.cat(query_states, dim=-1) + + key_states = [F.linear(hidden_states, key_slices[i]) for i in range(self.config.pretraining_tp)] + key_states = torch.cat(key_states, dim=-1) + + value_states = [F.linear(hidden_states, value_slices[i]) for i in range(self.config.pretraining_tp)] + value_states = torch.cat(value_states, dim=-1) + + else: + query_states = self.q_proj(hidden_states) + key_states = self.k_proj(hidden_states) + value_states = self.v_proj(hidden_states) + + # sp: all-to-all comminucation when introducing sequence parallel + if sp_mode == "all_to_all": + query_states = all_to_all_comm(query_states, sp_group) + key_states = all_to_all_comm(key_states, sp_group) + value_states = all_to_all_comm(value_states, sp_group) + bsz, q_len, _ = query_states.size() + + query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2) + key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2) + value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2) + + kv_seq_len = key_states.shape[-2] + if past_key_value is not None: + kv_seq_len += past_key_value[0].shape[-2] + cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len) + query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids) + + if past_key_value is not None: + # reuse k, v, self_attention + key_states = torch.cat([past_key_value[0], key_states], dim=2) + value_states = torch.cat([past_key_value[1], value_states], dim=2) + + past_key_value = (key_states, value_states) if use_cache else None + + # repeat k/v heads if n_kv_heads < n_heads + key_states = repeat_kv(key_states, self.num_key_value_groups) + value_states = repeat_kv(value_states, self.num_key_value_groups) + + attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim) + + if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len): + raise ValueError( + f"Attention weights should be of size {(bsz, self.num_heads, q_len, kv_seq_len)}, but is" + f" {attn_weights.size()}" + ) + + if attention_mask is not None: + if attention_mask.size() != (bsz, 1, q_len, kv_seq_len): + raise ValueError( + f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}" + ) + attn_weights = attn_weights + attention_mask + + # upcast attention to fp32 + attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype) + attn_output = torch.matmul(attn_weights, value_states) + + if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim): + raise ValueError( + f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is" + f" {attn_output.size()}" + ) + + attn_output = attn_output.transpose(1, 2).contiguous() + # sp: all-to-all comminucation when introducing sequence parallel + if sp_mode == "all_to_all": + attn_output = attn_output.reshape(bsz, q_len, self.num_heads * self.head_dim) + attn_output = all_to_all_comm(attn_output, sp_group, scatter_dim=1, gather_dim=2) + else: + attn_output = attn_output.reshape(bsz, q_len, self.hidden_size) + + if self.config.pretraining_tp > 1: + attn_output = attn_output.split(self.hidden_size // self.config.pretraining_tp, dim=2) + o_proj_slices = self.o_proj.weight.split(self.hidden_size // self.config.pretraining_tp, dim=1) + attn_output = sum([F.linear(attn_output[i], o_proj_slices[i]) for i in range(self.config.pretraining_tp)]) + else: + attn_output = self.o_proj(attn_output) + + if not output_attentions: + attn_weights = None + return attn_output, attn_weights, past_key_value + + return forward + + +def get_llama_seq_parallel_model_forward(sp_mode, sp_size, sp_group): + logger = logging.get_logger(__name__) + + def forward( + self, + input_ids: torch.LongTensor = None, + attention_mask: Optional[torch.Tensor] = None, + position_ids: Optional[torch.LongTensor] = None, + past_key_values: Optional[List[torch.FloatTensor]] = None, + inputs_embeds: Optional[torch.FloatTensor] = None, + use_cache: Optional[bool] = None, + output_attentions: Optional[bool] = None, + output_hidden_states: Optional[bool] = None, + return_dict: Optional[bool] = None, + ) -> Union[Tuple, BaseModelOutputWithPast]: + output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions + output_hidden_states = ( + output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states + ) + use_cache = use_cache if use_cache is not None else self.config.use_cache + + return_dict = return_dict if return_dict is not None else self.config.use_return_dict + + # retrieve input_ids and inputs_embeds + if input_ids is not None and inputs_embeds is not None: + raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time") + elif input_ids is not None: + batch_size, seq_length = input_ids.shape + elif inputs_embeds is not None: + batch_size, seq_length, _ = inputs_embeds.shape + else: + raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds") + + seq_length_with_past = seq_length + past_key_values_length = 0 + + if past_key_values is not None: + past_key_values_length = past_key_values[0][0].shape[2] + # modify past_key_values_length when using sequence parallel + past_key_values_length *= sp_size + seq_length_with_past = seq_length_with_past + past_key_values_length + + if position_ids is None: + device = input_ids.device if input_ids is not None else inputs_embeds.device + position_ids = torch.arange( + past_key_values_length, seq_length + past_key_values_length, dtype=torch.long, device=device + ) + position_ids = position_ids.unsqueeze(0).view(-1, seq_length) + else: + position_ids = position_ids.view(-1, seq_length).long() + + if inputs_embeds is None: + inputs_embeds = self.embed_tokens(input_ids) + + if sp_mode in ["ring", "split_gather"]: + inputs_embeds = split_forward_gather_backward(inputs_embeds, 1, sp_group) + elif sp_mode == "all_to_all": + inputs_embeds = split_forward_gather_backward(inputs_embeds, 1, sp_group, 1 / sp_size) + + if attention_mask is None: + attention_mask = torch.ones( + (batch_size, seq_length_with_past), dtype=torch.bool, device=inputs_embeds.device + ) + + attention_mask = self._prepare_decoder_attention_mask( + attention_mask, attention_mask.shape, inputs_embeds, past_key_values_length + ) + + hidden_states = inputs_embeds + + if (self.gradient_checkpointing or sp_mode in ["ring", "all_to_all"]) and self.training: + if use_cache: + logger.warning_once( + "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..." + ) + use_cache = False + + # decoder layers + all_hidden_states = () if output_hidden_states else None + all_self_attns = () if output_attentions else None + next_decoder_cache = () if use_cache else None + + for idx, decoder_layer in enumerate(self.layers): + if output_hidden_states: + all_hidden_states += (hidden_states,) + + past_key_value = past_key_values[idx] if past_key_values is not None else None + + if (self.gradient_checkpointing or sp_mode in ["ring", "all_to_all"]) and self.training: + + def create_custom_forward(module): + def custom_forward(*inputs): + # None for past_key_value + return module(*inputs, past_key_value, output_attentions) + + return custom_forward + + layer_outputs = torch.utils.checkpoint.checkpoint( + create_custom_forward(decoder_layer), + hidden_states, + attention_mask, + position_ids, + ) + + else: + layer_outputs = decoder_layer( + hidden_states, + attention_mask=attention_mask, + position_ids=position_ids, + past_key_value=past_key_value, + output_attentions=output_attentions, + use_cache=use_cache, + ) + + hidden_states = layer_outputs[0] + + if use_cache: + next_decoder_cache += (layer_outputs[2 if output_attentions else 1],) + + if output_attentions: + all_self_attns += (layer_outputs[1],) + + hidden_states = self.norm(hidden_states) + + if sp_mode == "ring" or sp_mode == "split_gather": + hidden_states = gather_forward_split_backward(hidden_states, 1, sp_group) + elif sp_mode == "all_to_all": + hidden_states = gather_forward_split_backward(hidden_states, 1, sp_group, grad_scale=sp_size) + + # add hidden states from the last decoder layer + if output_hidden_states: + all_hidden_states += (hidden_states,) + + next_cache = next_decoder_cache if use_cache else None + if not return_dict: + return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None) + + return BaseModelOutputWithPast( + last_hidden_state=hidden_states, + past_key_values=next_cache, + hidden_states=all_hidden_states, + attentions=all_self_attns, + ) + + return forward diff --git a/colossalai/shardformer/modeling/opt.py b/colossalai/shardformer/modeling/opt.py index d0e267eacd25..a265264303ad 100644 --- a/colossalai/shardformer/modeling/opt.py +++ b/colossalai/shardformer/modeling/opt.py @@ -18,6 +18,37 @@ from transformers.utils import logging from colossalai.pipeline.stage_manager import PipelineStageManager +from colossalai.shardformer.layer import ColoAttention +from colossalai.shardformer.shard import ShardConfig + +logger = logging.get_logger(__name__) + + +def _get_attention_mask( + self: OPTModel, + shard_config: ShardConfig, + hidden_states: torch.Tensor, + past_key_values_length: int, + attention_mask: Optional[torch.FloatTensor], +): + batch_size, seq_length = hidden_states.shape[:2] + mask_seq_length = past_key_values_length + seq_length + if shard_config.enable_flash_attention: + attention_mask = ColoAttention.prepare_attn_kwargs( + (batch_size, 1, seq_length, mask_seq_length), + hidden_states.dtype, + hidden_states.device, + attention_mask, + is_causal=True, + ) + else: + attention_mask = self.decoder._prepare_decoder_attention_mask( + attention_mask, + (batch_size, seq_length), + hidden_states, + past_key_values_length, + ) + return attention_mask class OPTPipelineForwards: @@ -26,46 +57,6 @@ class OPTPipelineForwards: under pipeline setting. """ - @staticmethod - def _prepare_decoder_attention_mask(attention_mask, input_shape, _dtype, device, past_key_values_length): - # create causal mask - # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len] - from transformers.models.opt.modeling_opt import _make_causal_mask - - combined_attention_mask = None - if input_shape[-1] > 1: - combined_attention_mask = _make_causal_mask( - input_shape, - _dtype, - device, - past_key_values_length=past_key_values_length, - ) - - if attention_mask is not None: - # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len] - expanded_attn_mask = OPTPipelineForwards._expand_mask(attention_mask, _dtype, tgt_len=input_shape[-1]).to( - device - ) - combined_attention_mask = ( - expanded_attn_mask if combined_attention_mask is None else expanded_attn_mask + combined_attention_mask - ) - - return combined_attention_mask - - @staticmethod - def _expand_mask(mask: torch.Tensor, dtype: torch.dtype, tgt_len: Optional[int] = None): - """ - Expands attention_mask from `[bsz, seq_len]` to `[bsz, 1, tgt_seq_len, src_seq_len]`. - """ - bsz, src_len = mask.size() - tgt_len = tgt_len if tgt_len is not None else src_len - - expanded_mask = mask[:, None, None, :].expand(bsz, 1, tgt_len, src_len).to(dtype) - - inverted_mask = 1.0 - expanded_mask - - return inverted_mask.masked_fill(inverted_mask.to(torch.bool), torch.finfo(dtype).min) - @staticmethod def opt_model_forward( self: OPTModel, @@ -81,6 +72,7 @@ def opt_model_forward( stage_manager: Optional[PipelineStageManager] = None, hidden_states: Optional[torch.FloatTensor] = None, stage_index: Optional[List[int]] = None, + shard_config: Optional[ShardConfig] = None, ) -> Union[Tuple, BaseModelOutputWithPast]: """ This forward method is modified based on transformers.models.opt.modeling_opt.OPTModel.forward @@ -119,7 +111,7 @@ def opt_model_forward( if decoder.project_in is not None: inputs_embeds = decoder.project_in(inputs_embeds) device = input_ids.device if input_ids is not None else inputs_embeds.device - _dtype = inputs_embeds.dtype + inputs_embeds.dtype else: if hidden_states is None: @@ -127,7 +119,7 @@ def opt_model_forward( input_shape = hidden_states.size()[:-1] batch_size, seq_length = input_shape[0], input_shape[1] device = hidden_states.device - _dtype = hidden_states.dtype + hidden_states.dtype past_key_values_length = past_key_values[0][0].shape[2] if past_key_values is not None else 0 # required mask seq length can be calculated via length of past @@ -141,13 +133,24 @@ def opt_model_forward( f"{mask_seq_length} (sum of the lengths of current and past inputs)" ) - causal_attention_mask = OPTPipelineForwards._prepare_decoder_attention_mask( - attention_mask, input_shape, _dtype, device, past_key_values_length - ) - if stage_manager.is_first_stage(): + causal_attention_mask = _get_attention_mask( + self, + shard_config, + inputs_embeds, + past_key_values_length, + attention_mask, + ) pos_embeds = decoder.embed_positions(attention_mask, past_key_values_length) hidden_states = inputs_embeds + pos_embeds + else: + causal_attention_mask = _get_attention_mask( + self, + shard_config, + hidden_states, + past_key_values_length, + attention_mask, + ) if decoder.gradient_checkpointing and decoder.training: if use_cache: @@ -249,7 +252,16 @@ def custom_forward(*inputs): if stage_manager.is_last_stage(): if not return_dict: - return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None) + return tuple( + v + for v in [ + hidden_states, + next_cache, + all_hidden_states, + all_self_attns, + ] + if v is not None + ) return BaseModelOutputWithPast( last_hidden_state=hidden_states, @@ -276,6 +288,7 @@ def opt_for_causal_lm_forward( stage_manager: Optional[PipelineStageManager] = None, hidden_states: Optional[torch.FloatTensor] = None, stage_index: Optional[List[int]] = None, + shard_config: Optional[ShardConfig] = None, ) -> Union[Tuple, CausalLMOutputWithPast]: r""" This function is modified on the basis of transformers.models.opt.modeling_opt.OPTForCausalLM.forward. @@ -303,6 +316,7 @@ def opt_for_causal_lm_forward( stage_manager=stage_manager, hidden_states=hidden_states, stage_index=stage_index, + shard_config=shard_config, ) if stage_manager.is_last_stage(): logits = self.lm_head(outputs[0]).contiguous() @@ -347,6 +361,7 @@ def opt_for_sequence_classification_forward( stage_manager: Optional[PipelineStageManager] = None, hidden_states: Optional[torch.FloatTensor] = None, stage_index: Optional[List[int]] = None, + shard_config: Optional[ShardConfig] = None, ) -> Union[Tuple, SequenceClassifierOutputWithPast]: r""" This function is modified on the basis of transformers.models.opt.modeling_opt.OPTForSequenceClassification.forward. @@ -371,6 +386,7 @@ def opt_for_sequence_classification_forward( stage_manager=stage_manager, hidden_states=hidden_states, stage_index=stage_index, + shard_config=shard_config, ) if stage_manager.is_last_stage(): @@ -448,6 +464,7 @@ def opt_for_question_answering_forward( stage_manager: Optional[PipelineStageManager] = None, hidden_states: Optional[torch.FloatTensor] = None, stage_index: Optional[List[int]] = None, + shard_config: Optional[ShardConfig] = None, ) -> Union[Tuple, QuestionAnsweringModelOutput]: r""" This function is modified on the basis of transformers.models.opt.modeling_opt.OPTForQuestionAnswering.forward. @@ -469,6 +486,7 @@ def opt_for_question_answering_forward( stage_manager=stage_manager, hidden_states=hidden_states, stage_index=stage_index, + shard_config=shard_config, ) if stage_manager.is_last_stage(): hidden_states = transformer_outputs[0] @@ -511,49 +529,47 @@ def opt_for_question_answering_forward( return {"hidden_states": hidden_states} -def get_opt_flash_attention_forward(): +def get_opt_flash_attention_forward(shard_config: ShardConfig): from transformers.models.opt.modeling_opt import OPTAttention - from colossalai.nn.layer.colo_attention import AttnMaskType, ColoAttention - def forward( self: OPTAttention, hidden_states: torch.Tensor, key_value_states: Optional[torch.Tensor] = None, past_key_value: Optional[Tuple[torch.Tensor]] = None, - attention_mask: Optional[torch.Tensor] = None, + attention_mask: Optional[dict] = None, layer_head_mask: Optional[torch.Tensor] = None, output_attentions: bool = False, ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]: """Input shape: Batch x Time x Channel""" - + assert layer_head_mask is None, "layer_head_mask is not supported for FlashAttention" # if key_value_states are provided this layer is used as a cross-attention layer # for the decoder is_cross_attention = key_value_states is not None + bsz, tgt_len, _ = hidden_states.size() - attention_input_shape = (bsz, -1, self.num_heads, self.head_dim) # get query proj - query_states = self.q_proj(hidden_states).view(*attention_input_shape) + query_states = self.q_proj(hidden_states) # get key, value proj if is_cross_attention and past_key_value is not None: - # reuse k, v, cross_attentions - key_states = past_key_value[0].transpose(1, 2).contiguous().view(*attention_input_shape) - value_states = past_key_value[1].transpose(1, 2).contiguous().view(*attention_input_shape) + # reuse k,v, cross_attentions + key_states = past_key_value[0] + value_states = past_key_value[1] elif is_cross_attention: # cross_attentions - key_states = self.k_proj(key_value_states).view(*attention_input_shape) - value_states = self.v_proj(key_value_states).view(*attention_input_shape) + key_states = self._shape(self.k_proj(key_value_states), -1, bsz) + value_states = self._shape(self.v_proj(key_value_states), -1, bsz) elif past_key_value is not None: # reuse k, v, self_attention - key_states = self.k_proj(hidden_states).view(*attention_input_shape) - value_states = self.v_proj(hidden_states).view(*attention_input_shape) - key_states = torch.cat([past_key_value[0], key_states], dim=1) - value_states = torch.cat([past_key_value[1], value_states], dim=1) + key_states = self._shape(self.k_proj(hidden_states), -1, bsz) + value_states = self._shape(self.v_proj(hidden_states), -1, bsz) + key_states = torch.cat([past_key_value[0], key_states], dim=2) + value_states = torch.cat([past_key_value[1], value_states], dim=2) else: # self_attention - key_states = self.k_proj(hidden_states).view(*attention_input_shape) - value_states = self.v_proj(hidden_states).view(*attention_input_shape) + key_states = self._shape(self.k_proj(hidden_states), -1, bsz) + value_states = self._shape(self.v_proj(hidden_states), -1, bsz) if self.is_decoder: # if cross_attention save Tuple(torch.Tensor, torch.Tensor) of all cross attention key/value_states. @@ -565,38 +581,181 @@ def forward( # if encoder bi-directional self-attention `past_key_value` is always `None` past_key_value = (key_states, value_states) - src_len = key_states.size(1) - if layer_head_mask != None: - if layer_head_mask.size() != (self.num_heads,): - raise ValueError( - f"Head mask for a single layer should be of size {(self.num_heads,)}, but is" - f" {layer_head_mask.size()}" - ) - - flash_attention_mask = None - attn_mask_type = AttnMaskType.causal - if attention_mask != None: - if attention_mask.size() != (bsz, 1, tgt_len, src_len): - raise ValueError( - f"Attention mask should be of size {(bsz, 1, tgt_len, src_len)}, but is {attention_mask.size()}" - ) - flash_attention_mask = ~(attention_mask[:, :, -1].squeeze(1).to(torch.bool)).contiguous() - if not torch.all(flash_attention_mask): - attn_mask_type = AttnMaskType.paddedcausal + query_states = self._shape(query_states, tgt_len, bsz) - attention = ColoAttention( - embed_dim=self.embed_dim, num_heads=self.num_heads, dropout=self.dropout, scale=self.scaling - ) - attn_output = attention( - query_states, key_states, value_states, attn_mask=flash_attention_mask, attn_mask_type=attn_mask_type + dropout_p = self.dropout if self.training else 0.0 + attn_output = ColoAttention.attention( + query_states, + key_states, + value_states, + **attention_mask, + dropout_p=dropout_p, + scale=self.scaling, ) + attn_output = attn_output.transpose(1, 2) + + # Use the `embed_dim` from the config (stored in the class) rather than `hidden_state` because `attn_output` can be + # partitioned aross GPUs when using tensor-parallelism. + attn_output = attn_output.reshape(bsz, tgt_len, self.embed_dim) + attn_output = self.out_proj(attn_output) + return attn_output, None, past_key_value return forward +def get_opt_decoder_forward_for_flash_attention(shard_config: ShardConfig): + from transformers.models.opt.modeling_opt import OPTDecoder + + def forward( + self: OPTDecoder, + input_ids: torch.LongTensor = None, + attention_mask: Optional[torch.Tensor] = None, + head_mask: Optional[torch.Tensor] = None, + past_key_values: Optional[List[torch.FloatTensor]] = None, + inputs_embeds: Optional[torch.FloatTensor] = None, + use_cache: Optional[bool] = None, + output_attentions: Optional[bool] = None, + output_hidden_states: Optional[bool] = None, + return_dict: Optional[bool] = None, + ) -> Union[Tuple, BaseModelOutputWithPast]: + output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions + output_hidden_states = ( + output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states + ) + use_cache = use_cache if use_cache is not None else self.config.use_cache + + return_dict = return_dict if return_dict is not None else self.config.use_return_dict + + # retrieve input_ids and inputs_embeds + if input_ids is not None and inputs_embeds is not None: + raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time") + elif input_ids is not None: + input_shape = input_ids.size() + input_ids = input_ids.view(-1, input_shape[-1]) + elif inputs_embeds is not None: + input_shape = inputs_embeds.size()[:-1] + else: + raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds") + + if inputs_embeds is None: + inputs_embeds = self.embed_tokens(input_ids) + + batch_size, seq_length = input_shape + past_key_values_length = past_key_values[0][0].shape[2] if past_key_values is not None else 0 + # required mask seq length can be calculated via length of past + mask_seq_length = past_key_values_length + seq_length + + # embed positions + if attention_mask is None: + attention_mask = torch.ones(batch_size, mask_seq_length, device=inputs_embeds.device) + elif attention_mask.shape[1] != mask_seq_length: + raise ValueError( + f"The provided attention mask has length {attention_mask.shape[1]}, but its length should be " + f"{mask_seq_length} (sum of the lengths of current and past inputs)" + ) + causal_attention_mask = _get_attention_mask( + self, shard_config, inputs_embeds, past_key_values_length, attention_mask + ) + pos_embeds = self.embed_positions(attention_mask, past_key_values_length) + + if self.project_in is not None: + inputs_embeds = self.project_in(inputs_embeds) + + hidden_states = inputs_embeds + pos_embeds + + if self.gradient_checkpointing and self.training: + if use_cache: + logger.warning_once( + "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..." + ) + use_cache = False + + # decoder layers + all_hidden_states = () if output_hidden_states else None + all_self_attns = () if output_attentions else None + next_decoder_cache = () if use_cache else None + + # check if head_mask has a correct number of layers specified if desired + for attn_mask, mask_name in zip([head_mask], ["head_mask"]): + if attn_mask is not None: + if attn_mask.size()[0] != (len(self.layers)): + raise ValueError( + f"The `{mask_name}` should be specified for {len(self.layers)} layers, but it is for" + f" {head_mask.size()[0]}." + ) + + for idx, decoder_layer in enumerate(self.layers): + # add LayerDrop (see https://arxiv.org/abs/1909.11556 for description) + if output_hidden_states: + all_hidden_states += (hidden_states,) + + if self.training: + dropout_probability = torch.rand([]) + if dropout_probability < self.layerdrop: + continue + + past_key_value = past_key_values[idx] if past_key_values is not None else None + + if self.gradient_checkpointing and self.training: + + def create_custom_forward(module): + def custom_forward(*inputs): + # None for past_key_value + return module(*inputs, output_attentions, None) + + return custom_forward + + layer_outputs = torch.utils.checkpoint.checkpoint( + create_custom_forward(decoder_layer), + hidden_states, + causal_attention_mask, + head_mask[idx] if head_mask is not None else None, + None, + ) + else: + layer_outputs = decoder_layer( + hidden_states, + attention_mask=causal_attention_mask, + layer_head_mask=(head_mask[idx] if head_mask is not None else None), + past_key_value=past_key_value, + output_attentions=output_attentions, + use_cache=use_cache, + ) + + hidden_states = layer_outputs[0] + + if use_cache: + next_decoder_cache += (layer_outputs[2 if output_attentions else 1],) + + if output_attentions: + all_self_attns += (layer_outputs[1],) + + if self.final_layer_norm is not None: + hidden_states = self.final_layer_norm(hidden_states) + + if self.project_out is not None: + hidden_states = self.project_out(hidden_states) + + # add hidden states from the last decoder layer + if output_hidden_states: + all_hidden_states += (hidden_states,) + + next_cache = next_decoder_cache if use_cache else None + if not return_dict: + return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None) + return BaseModelOutputWithPast( + last_hidden_state=hidden_states, + past_key_values=next_cache, + hidden_states=all_hidden_states, + attentions=all_self_attns, + ) + + return forward + + def get_jit_fused_opt_decoder_layer_forward(): from transformers.models.opt.modeling_opt import OPTDecoderLayer diff --git a/colossalai/shardformer/modeling/vit.py b/colossalai/shardformer/modeling/vit.py index ab141a74aef8..e9c256a13571 100644 --- a/colossalai/shardformer/modeling/vit.py +++ b/colossalai/shardformer/modeling/vit.py @@ -1,4 +1,3 @@ -import math from typing import List, Optional, Tuple, Union import torch @@ -6,6 +5,7 @@ from transformers.utils import logging from colossalai.pipeline.stage_manager import PipelineStageManager +from colossalai.shardformer.layer import ColoAttention def _encoder_forward( @@ -98,7 +98,9 @@ def pp_forward( pixel_values = pixel_values.to(expected_dtype) embedding_output = self.embeddings( - pixel_values, bool_masked_pos=bool_masked_pos, interpolate_pos_encoding=interpolate_pos_encoding + pixel_values, + bool_masked_pos=bool_masked_pos, + interpolate_pos_encoding=interpolate_pos_encoding, ) hidden_states = embedding_output else: @@ -336,34 +338,27 @@ def pp_forward( def get_vit_flash_self_attention_forward(): from transformers.models.vit.modeling_vit import ViTSelfAttention - from colossalai.nn.layer.colo_attention import ColoAttention - - def transpose_for_scores(x: torch.Tensor, num_attention_heads, attention_head_size) -> torch.Tensor: - new_x_shape = x.size()[:-1] + (num_attention_heads, attention_head_size) - x = x.view(new_x_shape) - return x - def forward( self: ViTSelfAttention, hidden_states: torch.Tensor, head_mask: Optional[torch.Tensor] = None, output_attentions: bool = False, ) -> Union[Tuple[torch.Tensor, torch.Tensor], Tuple[torch.Tensor]]: + assert head_mask is None, "head_mask is not supported for FlashAttention" mixed_query_layer = self.query(hidden_states) - key_layer = transpose_for_scores(self.key(hidden_states), self.num_attention_heads, self.attention_head_size) - value_layer = transpose_for_scores( - self.value(hidden_states), self.num_attention_heads, self.attention_head_size - ) - query_layer = transpose_for_scores(mixed_query_layer, self.num_attention_heads, self.attention_head_size) + key_layer = self.transpose_for_scores(self.key(hidden_states)) + value_layer = self.transpose_for_scores(self.value(hidden_states)) + query_layer = self.transpose_for_scores(mixed_query_layer) - scale = 1.0 / math.sqrt(self.attention_head_size) - attention = ColoAttention( - embed_dim=self.all_head_size, num_heads=self.num_attention_heads, dropout=self.dropout.p, scale=scale - ) - context_layer = attention(query_layer, key_layer, value_layer) + dropout_p = self.dropout.p if self.training else 0.0 + context_layer = ColoAttention.attention(query_layer, key_layer, value_layer, dropout_p=dropout_p) + + context_layer = context_layer.permute(0, 2, 1, 3).contiguous() + new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,) + context_layer = context_layer.view(new_context_layer_shape) - outputs = (context_layer,) + outputs = (context_layer, None) if output_attentions else (context_layer,) return outputs diff --git a/colossalai/shardformer/modeling/whisper.py b/colossalai/shardformer/modeling/whisper.py index cb8b45ae7d01..7ccc79276cf7 100644 --- a/colossalai/shardformer/modeling/whisper.py +++ b/colossalai/shardformer/modeling/whisper.py @@ -13,41 +13,74 @@ SequenceClassifierOutput, ) from transformers.models.whisper.modeling_whisper import ( + WhisperDecoder, WhisperEncoder, WhisperForAudioClassification, WhisperForConditionalGeneration, WhisperModel, + shift_tokens_right, ) from transformers.utils import logging from colossalai.pipeline.stage_manager import PipelineStageManager +from colossalai.shardformer.layer import ColoAttention +from colossalai.shardformer.shard import ShardConfig + +logger = logging.get_logger(__name__) + + +def _get_attention_mask( + self: WhisperDecoder, + shard_config: ShardConfig, + hidden_states: torch.Tensor, + past_key_values_length: int, + attention_mask: Optional[torch.FloatTensor], +): + batch_size, seq_length = hidden_states.shape[:2] + mask_seq_length = past_key_values_length + seq_length + if shard_config.enable_flash_attention: + attention_mask = ColoAttention.prepare_attn_kwargs( + (batch_size, 1, seq_length, mask_seq_length), + hidden_states.dtype, + hidden_states.device, + attention_mask, + is_causal=True, + ) + else: + attention_mask = self._prepare_decoder_attention_mask( + attention_mask, + (batch_size, seq_length), + hidden_states, + past_key_values_length, + ) + return attention_mask def get_whisper_flash_attention_forward(): from transformers.models.whisper.modeling_whisper import WhisperAttention - from colossalai.nn.layer.colo_attention import AttnMaskType, ColoAttention - - def shape(tensor: torch.Tensor, seq_len: int, bsz: int, num_heads: int, head_dim: int): - return tensor.view(bsz, seq_len, num_heads, head_dim).contiguous() - def forward( self: WhisperAttention, hidden_states: torch.Tensor, key_value_states: Optional[torch.Tensor] = None, past_key_value: Optional[Tuple[torch.Tensor]] = None, - attention_mask: Optional[torch.Tensor] = None, + attention_mask: Optional[dict] = None, layer_head_mask: Optional[torch.Tensor] = None, output_attentions: bool = False, ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]: """Input shape: Batch x Time x Channel""" - + assert layer_head_mask is None, "layer_head_mask is not supported for FlashAttention" + # for encoder, attention_mask is None + if attention_mask is None: + attention_mask = {} # if key_value_states are provided this layer is used as a cross-attention layer # for the decoder is_cross_attention = key_value_states is not None bsz, tgt_len, _ = hidden_states.size() + # get query proj + query_states = self.q_proj(hidden_states) # get key, value proj # `past_key_value[0].shape[2] == key_value_states.shape[1]` # is checking that the `sequence_length` of the `past_key_value` is the same as @@ -55,25 +88,25 @@ def forward( if ( is_cross_attention and past_key_value is not None - and past_key_value[0].shape[1] == key_value_states.shape[1] + and past_key_value[0].shape[2] == key_value_states.shape[1] ): # reuse k,v, cross_attentions key_states = past_key_value[0] value_states = past_key_value[1] elif is_cross_attention: # cross_attentions - key_states = shape(self.k_proj(key_value_states), -1, bsz, self.num_heads, self.head_dim) - value_states = shape(self.v_proj(key_value_states), -1, bsz, self.num_heads, self.head_dim) + key_states = self._shape(self.k_proj(key_value_states), -1, bsz) + value_states = self._shape(self.v_proj(key_value_states), -1, bsz) elif past_key_value is not None: # reuse k, v, self_attention - key_states = shape(self.k_proj(hidden_states), -1, bsz, self.num_heads, self.head_dim) - value_states = shape(self.v_proj(hidden_states), -1, bsz, self.num_heads, self.head_dim) - key_states = torch.cat([past_key_value[0], key_states], dim=1) - value_states = torch.cat([past_key_value[1], value_states], dim=1) + key_states = self._shape(self.k_proj(hidden_states), -1, bsz) + value_states = self._shape(self.v_proj(hidden_states), -1, bsz) + key_states = torch.cat([past_key_value[0], key_states], dim=2) + value_states = torch.cat([past_key_value[1], value_states], dim=2) else: # self_attention - key_states = shape(self.k_proj(hidden_states), -1, bsz, self.num_heads, self.head_dim) - value_states = shape(self.v_proj(hidden_states), -1, bsz, self.num_heads, self.head_dim) + key_states = self._shape(self.k_proj(hidden_states), -1, bsz) + value_states = self._shape(self.v_proj(hidden_states), -1, bsz) if self.is_decoder: # if cross_attention save Tuple(torch.Tensor, torch.Tensor) of all cross attention key/value_states. @@ -85,42 +118,178 @@ def forward( # if encoder bi-directional self-attention `past_key_value` is always `None` past_key_value = (key_states, value_states) - # get query proj - query_states = shape(self.q_proj(hidden_states), tgt_len, bsz, self.num_heads, self.head_dim) + query_states = self._shape(query_states, tgt_len, bsz) - src_len = key_states.size(1) - if layer_head_mask is not None: - if layer_head_mask.size() != (self.num_heads,): - raise ValueError( - f"Head mask for a single layer should be of size {(self.num_heads,)}, but is" - f" {layer_head_mask.size()}" - ) + dropout_p = self.dropout if self.training else 0.0 + attn_output = ColoAttention.attention( + query_states, + key_states, + value_states, + **attention_mask, + dropout_p=dropout_p, + scale=self.scaling, + ) + attn_output = attn_output.transpose(1, 2) - attn_type = None - flash_attention_mask = None + # Use the `embed_dim` from the config (stored in the class) rather than `hidden_state` because `attn_output` can be + # partitioned across GPUs when using tensor-parallelism. + attn_output = attn_output.reshape(bsz, tgt_len, self.embed_dim) - if self.is_decoder: - if attention_mask is not None: - if attention_mask.size() != (bsz, 1, tgt_len, src_len): - raise ValueError( - f"Attention mask should be of size {(bsz, 1, tgt_len, src_len)}, but is {attention_mask.size()}" - ) - flash_attention_mask = ~(attention_mask[:, :, -1].squeeze(1).to(torch.bool).contiguous()) - if not torch.all(flash_attention_mask): - attn_type = AttnMaskType.paddedcausal - else: - attn_type = AttnMaskType.causal + attn_output = self.out_proj(attn_output) - attention = ColoAttention( - embed_dim=self.embed_dim, num_heads=self.num_heads, dropout=self.dropout, scale=self.scaling - ) - attn_output = attention( - query_states, key_states, value_states, attn_mask=flash_attention_mask, attn_mask_type=attn_type + return attn_output, None, past_key_value + + return forward + + +def get_whisper_decoder_forward_for_flash_attention(shard_config: ShardConfig): + def forward( + self: WhisperDecoder, + input_ids=None, + attention_mask=None, + encoder_hidden_states=None, + head_mask=None, + cross_attn_head_mask=None, + past_key_values=None, + inputs_embeds=None, + use_cache=None, + output_attentions=None, + output_hidden_states=None, + return_dict=None, + ): + output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions + output_hidden_states = ( + output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states ) + use_cache = use_cache if use_cache is not None else self.config.use_cache + return_dict = return_dict if return_dict is not None else self.config.use_return_dict - attn_output = self.out_proj(attn_output) + # retrieve input_ids and inputs_embeds + if input_ids is not None and inputs_embeds is not None: + raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time") + elif input_ids is not None: + input_shape = input_ids.size() + input_ids = input_ids.view(-1, input_shape[-1]) + elif inputs_embeds is not None: + input_shape = inputs_embeds.size()[:-1] + else: + raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds") - return attn_output, None, past_key_value + # past_key_values_length + past_key_values_length = past_key_values[0][0].shape[2] if past_key_values is not None else 0 + + if inputs_embeds is None: + inputs_embeds = self.embed_tokens(input_ids) + + attention_mask = _get_attention_mask(self, shard_config, inputs_embeds, past_key_values_length, attention_mask) + + # embed positions + if input_ids is not None: + positions = self.embed_positions(input_ids, past_key_values_length=past_key_values_length) + else: + positions = self.embed_positions(inputs_embeds, past_key_values_length=past_key_values_length) + + hidden_states = inputs_embeds + positions + hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training) + + if self.gradient_checkpointing and self.training: + if use_cache: + logger.warning_once( + "`use_cache = True` is incompatible with gradient checkpointing. Setting `use_cache = False`..." + ) + use_cache = False + # decoder layers + all_hidden_states = () if output_hidden_states else None + all_self_attns = () if output_attentions else None + all_cross_attentions = () if (output_attentions and encoder_hidden_states is not None) else None + next_decoder_cache = () if use_cache else None + + # check if head_mask/cross_attn_head_mask has a correct number of layers specified if desired + for attn_mask, mask_name in zip([head_mask, cross_attn_head_mask], ["head_mask", "cross_attn_head_mask"]): + if attn_mask is not None: + assert attn_mask.size()[0] == (len(self.layers)), ( + f"The `{mask_name}` should be specified for {len(self.layers)} layers, but it is for" + f" {head_mask.size()[0]}." + ) + for idx, decoder_layer in enumerate(self.layers): + # add LayerDrop (see https://arxiv.org/abs/1909.11556 for description) + if output_hidden_states: + all_hidden_states += (hidden_states,) + if self.training: + dropout_probability = torch.rand([]) + if dropout_probability < self.layerdrop: + continue + + past_key_value = past_key_values[idx] if past_key_values is not None else None + + if self.gradient_checkpointing and self.training: + + def create_custom_forward(module): + def custom_forward(*inputs): + # None for past_key_value + return module(*inputs, output_attentions, use_cache) + + return custom_forward + + layer_outputs = torch.utils.checkpoint.checkpoint( + create_custom_forward(decoder_layer), + hidden_states, + attention_mask, + encoder_hidden_states, + None, # encoder attention mask + head_mask[idx] if head_mask is not None else None, + (cross_attn_head_mask[idx] if cross_attn_head_mask is not None else None), + None, # past_key_value + ) + else: + layer_outputs = decoder_layer( + hidden_states, + attention_mask=attention_mask, + encoder_hidden_states=encoder_hidden_states, + layer_head_mask=(head_mask[idx] if head_mask is not None else None), + cross_attn_layer_head_mask=( + cross_attn_head_mask[idx] if cross_attn_head_mask is not None else None + ), + past_key_value=past_key_value, + output_attentions=output_attentions, + use_cache=use_cache, + ) + hidden_states = layer_outputs[0] + + if use_cache: + next_decoder_cache += (layer_outputs[3 if output_attentions else 1],) + + if output_attentions: + all_self_attns += (layer_outputs[1],) + + if encoder_hidden_states is not None: + all_cross_attentions += (layer_outputs[2],) + + hidden_states = self.layer_norm(hidden_states) + # add hidden states from the last decoder layer + if output_hidden_states: + all_hidden_states += (hidden_states,) + + next_cache = next_decoder_cache if use_cache else None + if not return_dict: + return tuple( + v + for v in [ + hidden_states, + next_cache, + all_hidden_states, + all_self_attns, + all_cross_attentions, + ] + if v is not None + ) + return BaseModelOutputWithPastAndCrossAttentions( + last_hidden_state=hidden_states, + past_key_values=next_cache, + hidden_states=all_hidden_states, + attentions=all_self_attns, + cross_attentions=all_cross_attentions, + ) return forward @@ -292,6 +461,7 @@ def whisper_encoder_forward( all_attentions=None, stage_index: Optional[List[int]] = None, decoder_starting_stage: Optional[int] = None, + shard_config: Optional[ShardConfig] = None, ): r""" Args: @@ -403,7 +573,9 @@ def custom_forward(*inputs): if not return_dict: return tuple(v for v in [hidden_states, encoder_states, all_attentions] if v is not None) return BaseModelOutput( - last_hidden_state=hidden_states, hidden_states=encoder_states, attentions=all_attentions + last_hidden_state=hidden_states, + hidden_states=encoder_states, + attentions=all_attentions, ) else: @@ -411,7 +583,7 @@ def custom_forward(*inputs): @staticmethod def whisper_decoder_forward( - self, + self: WhisperDecoder, input_ids=None, attention_mask=None, encoder_hidden_states=None, @@ -427,6 +599,7 @@ def whisper_decoder_forward( hidden_states: Optional[torch.FloatTensor] = None, stage_index: Optional[List[int]] = None, decoder_starting_stage: Optional[int] = None, + shard_config: Optional[ShardConfig] = None, ): r""" Args: @@ -535,8 +708,12 @@ def whisper_decoder_forward( else: positions = self.embed_positions(inputs_embeds, past_key_values_length=past_key_values_length) - attention_mask = self._prepare_decoder_attention_mask( - attention_mask, input_shape, inputs_embeds, past_key_values_length + attention_mask = _get_attention_mask( + self, + shard_config, + inputs_embeds, + past_key_values_length, + attention_mask, ) hidden_states = inputs_embeds + positions @@ -556,8 +733,12 @@ def whisper_decoder_forward( ) input_shape = hidden_states.size()[:-1] - attention_mask = self._prepare_decoder_attention_mask( - attention_mask, input_shape, hidden_states, past_key_values_length + attention_mask = _get_attention_mask( + self, + shard_config, + hidden_states, + past_key_values_length, + attention_mask, ) start_idx, end_idx = stage_index[0], stage_index[1] @@ -590,7 +771,7 @@ def custom_forward(*inputs): encoder_hidden_states, None, # encoder attention mask head_mask[idx] if head_mask is not None else None, - cross_attn_head_mask[idx] if cross_attn_head_mask is not None else None, + (cross_attn_head_mask[idx] if cross_attn_head_mask is not None else None), None, # past_key_value ) else: @@ -626,7 +807,13 @@ def custom_forward(*inputs): if not return_dict: return tuple( v - for v in [hidden_states, next_cache, all_hidden_states, all_self_attns, all_cross_attentions] + for v in [ + hidden_states, + next_cache, + all_hidden_states, + all_self_attns, + all_cross_attentions, + ] if v is not None ) return BaseModelOutputWithPastAndCrossAttentions( @@ -666,6 +853,7 @@ def whisper_model_forward( encoder_hidden_states: Optional[torch.FloatTensor] = None, stage_index: Optional[List[int]] = None, decoder_starting_stage: Optional[int] = None, + shard_config: Optional[ShardConfig] = None, ): r""" Returns: @@ -735,7 +923,7 @@ def whisper_model_forward( elif return_dict and not isinstance(encoder_outputs, BaseModelOutput): encoder_outputs = BaseModelOutput( last_hidden_state=encoder_outputs[0], - hidden_states=encoder_outputs[1] if len(encoder_outputs) > 1 else None, + hidden_states=(encoder_outputs[1] if len(encoder_outputs) > 1 else None), attentions=encoder_outputs[2] if len(encoder_outputs) > 2 else None, ) @@ -767,6 +955,7 @@ def whisper_model_forward( hidden_states=hidden_states, stage_index=stage_index, decoder_starting_stage=decoder_starting_stage, + shard_config=shard_config, ) # Directly return outputs of overloaded Whisper forward if not at last stage. @@ -810,6 +999,7 @@ def whisper_for_conditional_generation_forward( encoder_hidden_states: Optional[torch.FloatTensor] = None, stage_index: Optional[List[int]] = None, decoder_starting_stage: Optional[int] = None, + shard_config: Optional[ShardConfig] = None, ) -> Union[Tuple[torch.Tensor], Seq2SeqLMOutput]: r""" labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*): @@ -870,6 +1060,7 @@ def whisper_for_conditional_generation_forward( encoder_hidden_states=encoder_hidden_states, stage_index=stage_index, decoder_starting_stage=decoder_starting_stage, + shard_config=shard_config, ) if not in_decoder: return outputs @@ -920,6 +1111,7 @@ def whisper_for_audio_classification_forward( all_attentions=None, stage_index: Optional[List[int]] = None, decoder_starting_stage: Optional[int] = None, + shard_config: Optional[ShardConfig] = None, ): r""" This function is modified on the basis of transformers.models.whisper.modeling_whisper.WhisperForAudioClassification.forward. diff --git a/colossalai/shardformer/policies/base_policy.py b/colossalai/shardformer/policies/base_policy.py index 1d2b7a570681..d67ab0a3c6bb 100644 --- a/colossalai/shardformer/policies/base_policy.py +++ b/colossalai/shardformer/policies/base_policy.py @@ -2,9 +2,8 @@ from abc import ABC, abstractmethod from dataclasses import dataclass -from typing import Any, Callable, Dict, List, Optional, Tuple, Union +from typing import Any, Callable, Dict, List, Optional, Union -import numpy as np import torch.nn as nn from torch import Tensor from torch.nn import Module @@ -196,50 +195,3 @@ def get_shared_params(self) -> List[Dict[int, Tensor]]: List[Dict[int, Tensor]]: List of parameters that should be shared across stages. E.g. [{0: module.model.embed_tokens.weight, 3: module.lm_head.weight}] """ return [] - - @staticmethod - def distribute_layers(num_layers: int, num_stages: int) -> List[int]: - """Divide layers into stages""" - quotient = num_layers // num_stages - remainder = num_layers % num_stages - - # calculate the num_layers per stage - layers_per_stage = [quotient] * num_stages - - # deal with the rest layers - if remainder > 0: - start_position = num_stages // 2 - remainder // 2 - for i in range(start_position, start_position + remainder): - layers_per_stage[i] += 1 - return layers_per_stage - - @staticmethod - def get_stage_index( - layers_per_stage: List[int], - stage: int, - num_model_chunks: int = 1, - num_stages: int = 0, - ) -> Union[Tuple[int, int], List[Tuple[int, int]]]: - """ - Get the start index and end index of layers for each stage. - - Args: - layers_per_stage (List[int]): number of layers for each stage - stage (int): the stage index - num_stages (int): number of stages - num_model_chunks (int): number of model chunks - - Returns: - - Tuple[int, int]: the start index and end index of this stage - - List[Tuple[int, int]]: the start index and end index of this stage for each model chunk - - """ - num_layers_per_stage_accumulated = np.insert(np.cumsum(layers_per_stage), 0, 0) - - stage_indices = [] - for model_chunk in range(num_model_chunks): - start_idx = num_layers_per_stage_accumulated[stage + model_chunk * num_stages] - end_idx = num_layers_per_stage_accumulated[stage + model_chunk * num_stages + 1] - stage_indices.append([start_idx, end_idx]) - - return stage_indices[0] if num_model_chunks == 1 else stage_indices diff --git a/colossalai/shardformer/policies/bert.py b/colossalai/shardformer/policies/bert.py index 0ab63b7650c1..0a61d8cff410 100644 --- a/colossalai/shardformer/policies/bert.py +++ b/colossalai/shardformer/policies/bert.py @@ -1,3 +1,4 @@ +import warnings from functools import partial from typing import Callable, Dict, List @@ -66,8 +67,17 @@ def module_policy(self): else: norm_cls = col_nn.LayerNorm - use_sequence_parallel = self.shard_config.enable_sequence_parallelism + sp_mode = self.shard_config.sequence_parallelism_mode if self.shard_config.enable_sequence_parallelism else None + assert sp_mode != "all_to_all", "all_to_all sequence parallelism is not supported for Bert" + if sp_mode == "ring": + warnings.warn( + f"For Bert, sequence parallelism is currently not support mode {sp_mode}, will set to be split_gather" + ) + sp_mode = "split_gather" + overlap = self.shard_config.enable_sequence_overlap + sp_partial_derived = sp_mode == "split_gather" + if self.shard_config.enable_tensor_parallelism: policy[BertLayer] = ModulePolicyDescription( attribute_replacement={ @@ -84,17 +94,26 @@ def module_policy(self): SubModuleReplacementDescription( suffix="attention.self.query", target_module=col_nn.Linear1D_Col, - kwargs={"seq_parallel": use_sequence_parallel, "overlap": overlap}, + kwargs={ + "seq_parallel_mode": sp_mode, + "overlap": overlap, + }, ), SubModuleReplacementDescription( suffix="attention.self.key", target_module=col_nn.Linear1D_Col, - kwargs={"seq_parallel": use_sequence_parallel, "overlap": overlap}, + kwargs={ + "seq_parallel_mode": sp_mode, + "overlap": overlap, + }, ), SubModuleReplacementDescription( suffix="attention.self.value", target_module=col_nn.Linear1D_Col, - kwargs={"seq_parallel": use_sequence_parallel, "overlap": overlap}, + kwargs={ + "seq_parallel_mode": sp_mode, + "overlap": overlap, + }, ), SubModuleReplacementDescription( suffix="attention.self.dropout", @@ -103,7 +122,7 @@ def module_policy(self): SubModuleReplacementDescription( suffix="attention.output.dense", target_module=col_nn.Linear1D_Row, - kwargs={"seq_parallel": use_sequence_parallel}, + kwargs={"seq_parallel_mode": sp_mode}, ), SubModuleReplacementDescription( suffix="attention.output.dropout", @@ -112,12 +131,15 @@ def module_policy(self): SubModuleReplacementDescription( suffix="intermediate.dense", target_module=col_nn.Linear1D_Col, - kwargs={"seq_parallel": use_sequence_parallel, "overlap": overlap}, + kwargs={ + "seq_parallel_mode": sp_mode, + "overlap": overlap, + }, ), SubModuleReplacementDescription( suffix="output.dense", target_module=col_nn.Linear1D_Row, - kwargs={"seq_parallel": use_sequence_parallel}, + kwargs={"seq_parallel_mode": sp_mode}, ), SubModuleReplacementDescription( suffix="output.dropout", @@ -139,7 +161,7 @@ def module_policy(self): ] ) - if use_sequence_parallel: + if sp_mode == "split_gather": self.append_or_create_method_replacement( description={"forward": bert_sequence_parallel_forward_fn(self.shard_config)}, policy=policy, @@ -153,12 +175,12 @@ def module_policy(self): SubModuleReplacementDescription( suffix="attention.output.LayerNorm", target_module=norm_cls, - kwargs={"sp_partial_derived": use_sequence_parallel}, + kwargs={"sp_partial_derived": sp_partial_derived}, ), SubModuleReplacementDescription( suffix="output.LayerNorm", target_module=norm_cls, - kwargs={"sp_partial_derived": use_sequence_parallel}, + kwargs={"sp_partial_derived": sp_partial_derived}, ), ], policy=policy, @@ -214,7 +236,9 @@ def add_lm_head_policy(self, base_policy): if self.shard_config.enable_tensor_parallelism: self.append_or_create_submodule_replacement( description=SubModuleReplacementDescription( - suffix="decoder", target_module=col_nn.Linear1D_Col, kwargs={"gather_output": True} + suffix="decoder", + target_module=col_nn.Linear1D_Col, + kwargs={"gather_output": True}, ), policy=base_policy, target_key=BertLMPredictionHead, @@ -241,7 +265,9 @@ def add_lm_prediction_policy(self, base_policy): "_load_from_state_dict": col_nn.ParallelModule._load_from_state_dict, } self.append_or_create_method_replacement( - description=method_replacement, policy=base_policy, target_key=BertLMPredictionHead + description=method_replacement, + policy=base_policy, + target_key=BertLMPredictionHead, ) return base_policy @@ -263,25 +289,25 @@ def set_pipeline_forward(self, model_cls: nn.Module, new_forward: Callable, poli module = self.model.bert if stage_manager.is_interleave: - layers_per_stage = self.distribute_layers( - len(module.encoder.layer), stage_manager.num_stages * stage_manager.num_model_chunks - ) - stage_manager.stage_indices = Policy.get_stage_index( - layers_per_stage, - stage_manager.stage, - num_model_chunks=stage_manager.num_model_chunks, - num_stages=stage_manager.num_stages, - ) + layers_per_stage = stage_manager.distribute_layers(len(module.encoder.layer)) + stage_manager.stage_indices = stage_manager.get_stage_index(layers_per_stage) method_replacement = { - "forward": partial(new_forward, stage_manager=stage_manager, shard_config=self.shard_config) + "forward": partial( + new_forward, + stage_manager=stage_manager, + shard_config=self.shard_config, + ) } else: - layers_per_stage = Policy.distribute_layers(len(module.encoder.layer), stage_manager.num_stages) - stage_index = Policy.get_stage_index(layers_per_stage, stage_manager.stage) + layers_per_stage = stage_manager.distribute_layers(len(module.encoder.layer)) + stage_index = stage_manager.get_stage_index(layers_per_stage) method_replacement = { "forward": partial( - new_forward, stage_manager=stage_manager, stage_index=stage_index, shard_config=self.shard_config + new_forward, + stage_manager=stage_manager, + stage_index=stage_index, + shard_config=self.shard_config, ) } @@ -300,15 +326,8 @@ def get_held_layers(self) -> List[Module]: held_layers = [] if stage_manager.is_interleave: assert stage_manager.num_model_chunks is not None - layers_per_stage = self.distribute_layers( - len(module.encoder.layer), stage_manager.num_stages * stage_manager.num_model_chunks - ) - stage_indices = Policy.get_stage_index( - layers_per_stage, - stage_manager.stage, - num_model_chunks=stage_manager.num_model_chunks, - num_stages=stage_manager.num_stages, - ) + layers_per_stage = stage_manager.distribute_layers(len(module.encoder.layer)) + stage_indices = stage_manager.get_stage_index(layers_per_stage) if stage_manager.is_first_stage(ignore_chunk=True): held_layers.append(module.embeddings) for start_idx, end_idx in stage_indices: @@ -317,10 +336,10 @@ def get_held_layers(self) -> List[Module]: held_layers.append(module.pooler) else: - layers_per_stage = self.distribute_layers(len(module.encoder.layer), stage_manager.num_stages) + layers_per_stage = stage_manager.distribute_layers(len(module.encoder.layer)) if stage_manager.is_first_stage(): held_layers.append(module.embeddings) - start_idx, end_idx = Policy.get_stage_index(layers_per_stage, stage_manager.stage) + start_idx, end_idx = stage_manager.get_stage_index(layers_per_stage) held_layers.extend(module.encoder.layer[start_idx:end_idx]) if stage_manager.is_last_stage(): held_layers.append(module.pooler) @@ -336,7 +355,9 @@ def module_policy(self): if self.pipeline_stage_manager: self.set_pipeline_forward( - model_cls=BertModel, new_forward=BertPipelineForwards.bert_model_forward, policy=policy + model_cls=BertModel, + new_forward=BertPipelineForwards.bert_model_forward, + policy=policy, ) return policy @@ -399,7 +420,9 @@ def module_policy(self): if self.pipeline_stage_manager: self.set_pipeline_forward( - model_cls=BertLMHeadModel, new_forward=BertPipelineForwards.bert_lm_head_model_forward, policy=policy + model_cls=BertLMHeadModel, + new_forward=BertPipelineForwards.bert_lm_head_model_forward, + policy=policy, ) return policy @@ -437,7 +460,9 @@ def module_policy(self): if self.pipeline_stage_manager: self.set_pipeline_forward( - model_cls=BertForMaskedLM, new_forward=BertPipelineForwards.bert_for_masked_lm_forward, policy=policy + model_cls=BertForMaskedLM, + new_forward=BertPipelineForwards.bert_for_masked_lm_forward, + policy=policy, ) return policy diff --git a/colossalai/shardformer/policies/bloom.py b/colossalai/shardformer/policies/bloom.py index eddfafdcbcdc..2becadc3fb19 100644 --- a/colossalai/shardformer/policies/bloom.py +++ b/colossalai/shardformer/policies/bloom.py @@ -1,3 +1,4 @@ +import warnings from functools import partial from typing import Callable, Dict, List @@ -55,8 +56,18 @@ def module_policy(self): norm_cls = col_nn.FusedLayerNorm else: norm_cls = col_nn.LayerNorm - use_sequence_parallel = self.shard_config.enable_sequence_parallelism + + sp_mode = self.shard_config.sequence_parallelism_mode if self.shard_config.enable_sequence_parallelism else None + assert sp_mode != "all_to_all", "all_to_all sequence parallelism is not supported for BLOOM" + if sp_mode == "ring": + warnings.warn( + f"For BLOOM, sequence parallelism is currently not support mode {sp_mode}, will set to be split_gather" + ) + sp_mode = "split_gather" + overlap = self.shard_config.enable_sequence_overlap + sp_partial_derived = sp_mode == "split_gather" + if self.shard_config.enable_tensor_parallelism: policy[BloomBlock] = ModulePolicyDescription( attribute_replacement={ @@ -70,12 +81,12 @@ def module_policy(self): SubModuleReplacementDescription( suffix="self_attention.query_key_value", target_module=col_nn.Linear1D_Col, - kwargs={"seq_parallel": use_sequence_parallel, "overlap": overlap}, + kwargs={"seq_parallel_mode": sp_mode, "overlap": overlap}, ), SubModuleReplacementDescription( suffix="self_attention.dense", target_module=col_nn.Linear1D_Row, - kwargs={"seq_parallel": use_sequence_parallel}, + kwargs={"seq_parallel_mode": sp_mode}, ), SubModuleReplacementDescription( suffix="self_attention.attention_dropout", @@ -84,12 +95,12 @@ def module_policy(self): SubModuleReplacementDescription( suffix="mlp.dense_h_to_4h", target_module=col_nn.Linear1D_Col, - kwargs={"seq_parallel": use_sequence_parallel, "overlap": overlap}, + kwargs={"seq_parallel_mode": sp_mode, "overlap": overlap}, ), SubModuleReplacementDescription( suffix="mlp.dense_4h_to_h", target_module=col_nn.Linear1D_Row, - kwargs={"seq_parallel": use_sequence_parallel}, + kwargs={"seq_parallel_mode": sp_mode}, ), ], ) @@ -132,19 +143,19 @@ def module_policy(self): SubModuleReplacementDescription( suffix="input_layernorm", target_module=norm_cls, - kwargs={"sp_partial_derived": use_sequence_parallel}, + kwargs={"sp_partial_derived": sp_partial_derived}, ), SubModuleReplacementDescription( suffix="post_attention_layernorm", target_module=norm_cls, - kwargs={"sp_partial_derived": use_sequence_parallel}, + kwargs={"sp_partial_derived": sp_partial_derived}, ), ], policy=policy, target_key=BloomBlock, ) - if use_sequence_parallel: + if sp_mode == "split_gather": self.append_or_create_method_replacement( description={"forward": get_bloom_sequence_parallel_forward_fn(self.shard_config)}, policy=policy, @@ -203,8 +214,8 @@ def set_pipeline_forward(self, model_cls: nn.Module, new_forward: Callable, poli else: module = self.model.transformer - layers_per_stage = Policy.distribute_layers(len(module.h), stage_manager.num_stages) - stage_index = Policy.get_stage_index(layers_per_stage, stage_manager.stage) + layers_per_stage = stage_manager.distribute_layers(len(module.h)) + stage_index = stage_manager.get_stage_index(layers_per_stage) method_replacement = { "forward": partial( new_forward, stage_manager=stage_manager, stage_index=stage_index, shard_config=self.shard_config @@ -226,11 +237,11 @@ def get_held_layers(self) -> List[Module]: stage_manager = self.pipeline_stage_manager held_layers = [] - layers_per_stage = self.distribute_layers(len(module.h), stage_manager.num_stages) + layers_per_stage = stage_manager.distribute_layers(len(module.h)) if stage_manager.is_first_stage(): held_layers.append(module.word_embeddings) held_layers.append(module.word_embeddings_layernorm) - start_idx, end_idx = self.get_stage_index(layers_per_stage, stage_manager.stage) + start_idx, end_idx = stage_manager.get_stage_index(layers_per_stage) held_layers.extend(module.h[start_idx:end_idx]) if stage_manager.is_last_stage(): held_layers.append(module.ln_f) diff --git a/colossalai/shardformer/policies/chatglm2.py b/colossalai/shardformer/policies/chatglm2.py index d1ad9f91478b..dabc14bffc95 100644 --- a/colossalai/shardformer/policies/chatglm2.py +++ b/colossalai/shardformer/policies/chatglm2.py @@ -1,3 +1,4 @@ +import warnings from functools import partial from typing import Callable, Dict, List, Union @@ -55,8 +56,17 @@ def module_policy(self) -> Dict[Union[str, nn.Module], ModulePolicyDescription]: norm_cls = col_nn.RMSNorm else: norm_cls = col_nn.LayerNorm - use_sequence_parallel = self.shard_config.enable_sequence_parallelism + + sp_mode = self.shard_config.sequence_parallelism_mode if self.shard_config.enable_sequence_parallelism else None + assert sp_mode != "all_to_all", "all_to_all sequence parallelism is not supported for ChatGLM2" + if sp_mode == "ring": + warnings.warn( + f"For ChatGLM2, sequence parallelism is currently not support mode {sp_mode}, will set to be split_gather" + ) + sp_mode = "split_gather" overlap = self.shard_config.enable_sequence_overlap + sp_partial_derived = sp_mode == "split_gather" + if self.shard_config.enable_tensor_parallelism: policy[ChatGLMModel] = ModulePolicyDescription( attribute_replacement={}, @@ -91,12 +101,12 @@ def module_policy(self) -> Dict[Union[str, nn.Module], ModulePolicyDescription]: SubModuleReplacementDescription( suffix="self_attention.query_key_value", target_module=col_nn.Linear1D_Col, - kwargs={"seq_parallel": use_sequence_parallel, "seq_parallel_dim": 0, "overlap": overlap}, + kwargs={"seq_parallel_mode": sp_mode, "seq_parallel_dim": 0, "overlap": overlap}, ), SubModuleReplacementDescription( suffix="self_attention.dense", target_module=col_nn.Linear1D_Row, - kwargs={"seq_parallel": use_sequence_parallel, "seq_parallel_dim": 0}, + kwargs={"seq_parallel_mode": sp_mode, "seq_parallel_dim": 0}, ), SubModuleReplacementDescription( suffix="self_attention.core_attention.attention_dropout", @@ -110,12 +120,12 @@ def module_policy(self) -> Dict[Union[str, nn.Module], ModulePolicyDescription]: SubModuleReplacementDescription( suffix="input_layernorm", target_module=norm_cls, - kwargs={"sp_partial_derived": use_sequence_parallel}, + kwargs={"sp_partial_derived": sp_partial_derived}, ), SubModuleReplacementDescription( suffix="post_attention_layernorm", target_module=norm_cls, - kwargs={"sp_partial_derived": use_sequence_parallel}, + kwargs={"sp_partial_derived": sp_partial_derived}, ), ], policy=policy, @@ -145,7 +155,7 @@ def module_policy(self) -> Dict[Union[str, nn.Module], ModulePolicyDescription]: ) # use sequence parallel - if use_sequence_parallel: + if sp_mode == "split_gather": self.append_or_create_method_replacement( description={"forward": get_chatglm_sequence_parallel_forward_fn(self.shard_config)}, policy=policy, @@ -179,10 +189,10 @@ def get_held_layers(self) -> List[nn.Module]: stage_manager = self.pipeline_stage_manager held_layers = [] - layers_per_stage = self.distribute_layers(module.num_layers, stage_manager.num_stages) + layers_per_stage = stage_manager.distribute_layers(module.num_layers) if stage_manager.is_first_stage(): held_layers.append(module.embedding) - start_idx, end_idx = self.get_stage_index(layers_per_stage, stage_manager.stage) + start_idx, end_idx = stage_manager.get_stage_index(layers_per_stage) held_layers.extend(module.encoder.layers[start_idx:end_idx]) if stage_manager.is_last_stage(): if module.encoder.post_layer_norm: @@ -204,8 +214,8 @@ def set_pipeline_forward(self, model_cls: nn.Module, new_forward: Callable, poli else: module = self.model.transformer - layers_per_stage = Policy.distribute_layers(module.num_layers, stage_manager.num_stages) - stage_index = Policy.get_stage_index(layers_per_stage, stage_manager.stage) + layers_per_stage = stage_manager.distribute_layers(module.num_layers) + stage_index = stage_manager.get_stage_index(layers_per_stage) method_replacement = { "forward": partial( new_forward, stage_manager=stage_manager, stage_index=stage_index, shard_config=self.shard_config diff --git a/colossalai/shardformer/policies/falcon.py b/colossalai/shardformer/policies/falcon.py index 5c148880f980..fe61c406fae3 100644 --- a/colossalai/shardformer/policies/falcon.py +++ b/colossalai/shardformer/policies/falcon.py @@ -161,8 +161,8 @@ def set_pipeline_forward(self, model_cls: nn.Module, new_forward: Callable, poli else: module = self.model.transformer - layers_per_stage = Policy.distribute_layers(len(module.h), stage_manager.num_stages) - stage_index = Policy.get_stage_index(layers_per_stage, stage_manager.stage) + layers_per_stage = stage_manager.distribute_layers(len(module.h)) + stage_index = stage_manager.get_stage_index(layers_per_stage) method_replacement = { "forward": partial( new_forward, stage_manager=stage_manager, stage_index=stage_index, shard_config=self.shard_config @@ -181,10 +181,10 @@ def get_held_layers(self) -> List[Module]: module = self.model.transformer stage_manager = self.pipeline_stage_manager held_layers = [] - layers_per_stage = self.distribute_layers(len(module.h), stage_manager.num_stages) + layers_per_stage = stage_manager.distribute_layers(len(module.h)) if stage_manager.is_first_stage(): held_layers.append(module.word_embeddings) - start_idx, end_idx = self.get_stage_index(layers_per_stage, stage_manager.stage) + start_idx, end_idx = stage_manager.get_stage_index(layers_per_stage) held_layers.extend(module.h[start_idx:end_idx]) if stage_manager.is_last_stage(): held_layers.append(module.ln_f) diff --git a/colossalai/shardformer/policies/gpt2.py b/colossalai/shardformer/policies/gpt2.py index 303766993e3d..380a432dc8b8 100644 --- a/colossalai/shardformer/policies/gpt2.py +++ b/colossalai/shardformer/policies/gpt2.py @@ -1,3 +1,4 @@ +import warnings from functools import partial from typing import Callable, Dict, List @@ -8,6 +9,7 @@ from ..modeling.gpt2 import ( GPT2PipelineForwards, get_gpt2_flash_attention_forward, + get_gpt_model_forward_for_flash_attn, get_lm_forward_with_dist_cross_entropy, gpt2_sequence_parallel_forward_fn, ) @@ -49,8 +51,25 @@ def module_policy(self): norm_cls = col_nn.FusedLayerNorm else: norm_cls = col_nn.LayerNorm - use_sequence_parallel = self.shard_config.enable_sequence_parallelism + + sp_mode = self.shard_config.sequence_parallelism_mode if self.shard_config.enable_sequence_parallelism else None + assert sp_mode != "all_to_all", "all_to_all sequence parallelism is not supported for GPT2" + if sp_mode == "ring": + warnings.warn( + f"For GPT2, sequence parallelism is currently not support mode {sp_mode}, will set to be split_gather" + ) + sp_mode = "split_gather" overlap = self.shard_config.enable_sequence_overlap + sp_partial_derived = sp_mode in ["split_gather", "ring"] + use_flash_attention = self.shard_config.enable_flash_attention + # todo: currently sp cannot be used with flashattention + if sp_mode in ["split_gather", "ring", "all_to_all"]: + if use_flash_attention: + warnings.warn( + f"Sequence parallelism mode {sp_mode} cannot be used with FlashAttention, will disable FlashAttention automatically." + ) + self.shard_config.enable_flash_attention = False + use_flash_attention = False if self.shard_config.enable_tensor_parallelism: policy[GPT2Model] = ModulePolicyDescription( sub_module_replacement=[ @@ -75,24 +94,34 @@ def module_policy(self): SubModuleReplacementDescription( suffix="attn.c_attn", target_module=col_nn.GPT2FusedLinearConv1D_Col, - kwargs={"n_fused": 3, "seq_parallel": use_sequence_parallel, "overlap": overlap}, + kwargs={ + "n_fused": 3, + "seq_parallel_mode": sp_mode, + "overlap": overlap, + }, ), SubModuleReplacementDescription( suffix="attn.c_proj", target_module=col_nn.GPT2FusedLinearConv1D_Row, kwargs={ - "seq_parallel": use_sequence_parallel, + "seq_parallel_mode": sp_mode, }, ), SubModuleReplacementDescription( suffix="mlp.c_fc", target_module=col_nn.GPT2FusedLinearConv1D_Col, - kwargs={"n_fused": 1, "seq_parallel": use_sequence_parallel, "overlap": overlap}, + kwargs={ + "n_fused": 1, + "seq_parallel_mode": sp_mode, + "overlap": overlap, + }, ), SubModuleReplacementDescription( suffix="mlp.c_proj", target_module=col_nn.GPT2FusedLinearConv1D_Row, - kwargs={"seq_parallel": use_sequence_parallel}, + kwargs={ + "seq_parallel_mode": sp_mode, + }, ), SubModuleReplacementDescription( suffix="attn.attn_dropout", @@ -124,25 +153,25 @@ def module_policy(self): SubModuleReplacementDescription( suffix="ln_1", target_module=norm_cls, - kwargs={"sp_partial_derived": use_sequence_parallel}, + kwargs={"sp_partial_derived": sp_partial_derived}, ), SubModuleReplacementDescription( suffix="ln_2", target_module=norm_cls, - kwargs={"sp_partial_derived": use_sequence_parallel}, + kwargs={"sp_partial_derived": sp_partial_derived}, ), SubModuleReplacementDescription( suffix="ln_cross_attn", target_module=norm_cls, ignore_if_not_exist=True, - kwargs={"sp_partial_derived": use_sequence_parallel}, + kwargs={"sp_partial_derived": sp_partial_derived}, ), ], policy=policy, target_key=GPT2Block, ) - if self.shard_config.enable_flash_attention: + if use_flash_attention: self.append_or_create_method_replacement( description={ "forward": get_gpt2_flash_attention_forward(), @@ -150,8 +179,12 @@ def module_policy(self): policy=policy, target_key=GPT2Attention, ) + if not self.shard_config.pipeline_stage_manager: + policy[GPT2Model].method_replacement = { + "forward": get_gpt_model_forward_for_flash_attn(self.shard_config) + } - if self.shard_config.enable_sequence_parallelism: + if sp_mode is not None: policy[GPT2Model].method_replacement = {"forward": gpt2_sequence_parallel_forward_fn(self.shard_config)} return policy @@ -172,15 +205,8 @@ def get_held_layers(self) -> List[nn.Module]: held_layers = [] if stage_manager.is_interleave: assert stage_manager.num_model_chunks is not None - layers_per_stage = self.distribute_layers( - len(module.h), stage_manager.num_stages * stage_manager.num_model_chunks - ) - stage_indices = Policy.get_stage_index( - layers_per_stage, - stage_manager.stage, - num_model_chunks=stage_manager.num_model_chunks, - num_stages=stage_manager.num_stages, - ) + layers_per_stage = stage_manager.distribute_layers(len(module.h)) + stage_indices = stage_manager.get_stage_index(layers_per_stage) if stage_manager.is_first_stage(ignore_chunk=True): held_layers.append(module.wte) held_layers.append(module.wpe) @@ -190,12 +216,12 @@ def get_held_layers(self) -> List[nn.Module]: if stage_manager.is_last_stage(ignore_chunk=True): held_layers.append(module.ln_f) else: - layers_per_stage = self.distribute_layers(len(module.h), stage_manager.num_stages) + layers_per_stage = stage_manager.distribute_layers(len(module.h)) if stage_manager.is_first_stage(): held_layers.append(module.wte) held_layers.append(module.wpe) held_layers.append(module.drop) - start_idx, end_idx = self.get_stage_index(layers_per_stage, stage_manager.stage) + start_idx, end_idx = stage_manager.get_stage_index(layers_per_stage) held_layers.extend(module.h[start_idx:end_idx]) if stage_manager.is_last_stage(): held_layers.append(module.ln_f) @@ -213,24 +239,24 @@ def set_pipeline_forward(self, model_cls: nn.Module, new_forward: Callable, poli module = self.model.transformer if stage_manager.is_interleave: - layers_per_stage = self.distribute_layers( - len(module.h), stage_manager.num_stages * stage_manager.num_model_chunks - ) - stage_manager.stage_indices = Policy.get_stage_index( - layers_per_stage, - stage_manager.stage, - num_model_chunks=stage_manager.num_model_chunks, - num_stages=stage_manager.num_stages, - ) + layers_per_stage = stage_manager.distribute_layers(len(module.h)) + stage_manager.stage_indices = stage_manager.get_stage_index(layers_per_stage) method_replacement = { - "forward": partial(new_forward, stage_manager=stage_manager, shard_config=self.shard_config) + "forward": partial( + new_forward, + stage_manager=stage_manager, + shard_config=self.shard_config, + ) } else: - layers_per_stage = Policy.distribute_layers(len(module.h), stage_manager.num_stages) - stage_index = Policy.get_stage_index(layers_per_stage, stage_manager.stage) + layers_per_stage = stage_manager.distribute_layers(len(module.h)) + stage_index = stage_manager.get_stage_index(layers_per_stage) method_replacement = { "forward": partial( - new_forward, stage_manager=stage_manager, stage_index=stage_index, shard_config=self.shard_config + new_forward, + stage_manager=stage_manager, + stage_index=stage_index, + shard_config=self.shard_config, ) } self.append_or_create_method_replacement(description=method_replacement, policy=policy, target_key=model_cls) @@ -245,7 +271,9 @@ def module_policy(self): if self.pipeline_stage_manager is not None: self.set_pipeline_forward( - model_cls=GPT2Model, new_forward=GPT2PipelineForwards.gpt2_model_forward, policy=policy + model_cls=GPT2Model, + new_forward=GPT2PipelineForwards.gpt2_model_forward, + policy=policy, ) return policy @@ -269,12 +297,17 @@ def module_policy(self): GPT2LMHeadModel: ModulePolicyDescription( sub_module_replacement=[ SubModuleReplacementDescription( - suffix="lm_head", target_module=col_nn.Linear1D_Col, kwargs={"gather_output": False} + suffix="lm_head", + target_module=col_nn.Linear1D_Col, + kwargs={"gather_output": not self.shard_config.parallel_output}, ) ], - method_replacement={"forward": get_lm_forward_with_dist_cross_entropy(self.shard_config)}, ) } + if self.shard_config.parallel_output: + addon_module[GPT2LMHeadModel].method_replacement = { + "forward": get_lm_forward_with_dist_cross_entropy(self.shard_config) + } module_policy.update(addon_module) if self.pipeline_stage_manager is not None: @@ -298,7 +331,12 @@ def get_shared_params(self) -> List[Dict[int, Tensor]]: if stage_manager is not None: if stage_manager.num_stages > 1 and id(module.transformer.wte.weight) == id(module.lm_head.weight): first_stage, last_stage = 0, stage_manager.num_stages - 1 - return [{first_stage: module.transformer.wte.weight, last_stage: module.lm_head.weight}] + return [ + { + first_stage: module.transformer.wte.weight, + last_stage: module.lm_head.weight, + } + ] return [] @@ -314,7 +352,9 @@ def module_policy(self): GPT2DoubleHeadsModel: ModulePolicyDescription( sub_module_replacement=[ SubModuleReplacementDescription( - suffix="lm_head", target_module=col_nn.Linear1D_Col, kwargs={"gather_output": True} + suffix="lm_head", + target_module=col_nn.Linear1D_Col, + kwargs={"gather_output": True}, ) ] ) @@ -349,7 +389,12 @@ def get_shared_params(self) -> List[Dict[int, Tensor]]: if stage_manager is not None: if stage_manager.num_stages > 1 and id(module.transformer.wte.weight) == id(module.lm_head.weight): first_stage, last_stage = 0, stage_manager.num_stages - 1 - return [{first_stage: module.transformer.wte.weight, last_stage: module.lm_head.weight}] + return [ + { + first_stage: module.transformer.wte.weight, + last_stage: module.lm_head.weight, + } + ] return [] @@ -391,7 +436,10 @@ def module_policy(self): addon_module = { GPT2ForTokenClassification: ModulePolicyDescription( sub_module_replacement=[ - SubModuleReplacementDescription(suffix="dropout", target_module=col_nn.DropoutForParallelInput) + SubModuleReplacementDescription( + suffix="dropout", + target_module=col_nn.DropoutForParallelInput, + ) ] ) } diff --git a/colossalai/shardformer/policies/gptj.py b/colossalai/shardformer/policies/gptj.py index 9feb826c4624..eab4c214a41f 100644 --- a/colossalai/shardformer/policies/gptj.py +++ b/colossalai/shardformer/policies/gptj.py @@ -6,7 +6,11 @@ import colossalai.shardformer.layer as col_nn -from ..modeling.gptj import GPTJPipelineForwards, get_gptj_flash_attention_forward +from ..modeling.gptj import ( + GPTJPipelineForwards, + get_gptj_flash_attention_forward, + gptj_model_forward_for_flash_attention, +) from .base_policy import ModulePolicyDescription, Policy, SubModuleReplacementDescription __all__ = [ @@ -71,17 +75,26 @@ def module_policy(self): SubModuleReplacementDescription( suffix="attn.k_proj", target_module=col_nn.Linear1D_Col, - kwargs={"seq_parallel": use_sequence_parallel, "overlap": overlap}, + kwargs={ + "seq_parallel": use_sequence_parallel, + "overlap": overlap, + }, ), SubModuleReplacementDescription( suffix="attn.q_proj", target_module=col_nn.Linear1D_Col, - kwargs={"seq_parallel": use_sequence_parallel, "overlap": overlap}, + kwargs={ + "seq_parallel": use_sequence_parallel, + "overlap": overlap, + }, ), SubModuleReplacementDescription( suffix="attn.v_proj", target_module=col_nn.Linear1D_Col, - kwargs={"seq_parallel": use_sequence_parallel, "overlap": overlap}, + kwargs={ + "seq_parallel": use_sequence_parallel, + "overlap": overlap, + }, ), SubModuleReplacementDescription( suffix="attn.out_proj", @@ -143,6 +156,12 @@ def module_policy(self): policy=policy, target_key=GPTJAttention, ) + if not self.shard_config.pipeline_stage_manager: + self.append_or_create_method_replacement( + description={"forward": gptj_model_forward_for_flash_attention(self.shard_config)}, + policy=policy, + target_key=GPTJModel, + ) return policy @@ -160,11 +179,11 @@ def get_held_layers(self) -> List[nn.Module]: stage_manager = self.pipeline_stage_manager held_layers = [] - layers_per_stage = self.distribute_layers(len(module.h), stage_manager.num_stages) + layers_per_stage = stage_manager.distribute_layers(len(module.h)) if stage_manager.is_first_stage(): held_layers.append(module.wte) held_layers.append(module.drop) - start_idx, end_idx = self.get_stage_index(layers_per_stage, stage_manager.stage) + start_idx, end_idx = stage_manager.get_stage_index(layers_per_stage) held_layers.extend(module.h[start_idx:end_idx]) if stage_manager.is_last_stage(): held_layers.append(module.ln_f) @@ -181,11 +200,14 @@ def set_pipeline_forward(self, model_cls: nn.Module, new_forward: Callable, poli else: module = self.model.transformer - layers_per_stage = Policy.distribute_layers(len(module.h), stage_manager.num_stages) - stage_index = Policy.get_stage_index(layers_per_stage, stage_manager.stage) + layers_per_stage = stage_manager.distribute_layers(len(module.h)) + stage_index = stage_manager.get_stage_index(layers_per_stage) method_replacement = { "forward": partial( - new_forward, stage_manager=stage_manager, stage_index=stage_index, shard_config=self.shard_config + new_forward, + stage_manager=stage_manager, + stage_index=stage_index, + shard_config=self.shard_config, ) } self.append_or_create_method_replacement(description=method_replacement, policy=policy, target_key=model_cls) @@ -203,7 +225,9 @@ def module_policy(self): if self.pipeline_stage_manager is not None: self.set_pipeline_forward( - model_cls=GPTJModel, new_forward=GPTJPipelineForwards.gptj_model_forward, policy=policy + model_cls=GPTJModel, + new_forward=GPTJPipelineForwards.gptj_model_forward, + policy=policy, ) return policy @@ -230,7 +254,9 @@ def module_policy(self): GPTJForCausalLM: ModulePolicyDescription( sub_module_replacement=[ SubModuleReplacementDescription( - suffix="lm_head", target_module=col_nn.Linear1D_Col, kwargs={"gather_output": True} + suffix="lm_head", + target_module=col_nn.Linear1D_Col, + kwargs={"gather_output": True}, ) ] ) @@ -239,7 +265,9 @@ def module_policy(self): if self.pipeline_stage_manager is not None: self.set_pipeline_forward( - model_cls=GPTJForCausalLM, new_forward=GPTJPipelineForwards.gptj_causallm_model_forward, policy=policy + model_cls=GPTJForCausalLM, + new_forward=GPTJPipelineForwards.gptj_causallm_model_forward, + policy=policy, ) return policy @@ -256,7 +284,12 @@ def get_shared_params(self) -> List[Dict[int, Tensor]]: if stage_manager is not None: if stage_manager.num_stages > 1 and id(module.transformer.wte.weight) == id(module.lm_head.weight): first_stage, last_stage = 0, stage_manager.num_stages - 1 - return [{first_stage: module.transformer.wte.weight, last_stage: module.lm_head.weight}] + return [ + { + first_stage: module.transformer.wte.weight, + last_stage: module.lm_head.weight, + } + ] return [] diff --git a/colossalai/shardformer/policies/llama.py b/colossalai/shardformer/policies/llama.py index 42bf0825b045..bb4551b2c31c 100644 --- a/colossalai/shardformer/policies/llama.py +++ b/colossalai/shardformer/policies/llama.py @@ -11,6 +11,9 @@ from ..modeling.llama import ( LlamaPipelineForwards, get_llama_flash_attention_forward, + get_llama_model_forward_for_flash_attn, + get_llama_seq_parallel_attention_forward, + get_llama_seq_parallel_model_forward, get_lm_forward_with_dist_cross_entropy, ) from .base_policy import ModulePolicyDescription, Policy, SubModuleReplacementDescription @@ -44,9 +47,74 @@ def module_policy(self) -> Dict[Union[str, nn.Module], ModulePolicyDescription]: else: norm_cls = RMSNorm - if self.shard_config.enable_sequence_parallelism: + if self.pipeline_stage_manager is not None: self.shard_config.enable_sequence_parallelism = False - warnings.warn("Llama doesn't support sequence parallelism now, will ignore the sequence parallelism flag.") + self.shard_config.enable_sequence_overlap = False + self.shard_config.sequence_parallelism_mode = None + warnings.warn( + f"For llama, sequence parallelism is currently not compatible with pipeline parallelism, set to be False" + ) + sp_mode = self.shard_config.sequence_parallelism_mode if self.shard_config.enable_sequence_parallelism else None + sp_size = self.shard_config.sequence_parallel_size if self.shard_config.enable_sequence_parallelism else None + sp_group = ( + self.shard_config.sequence_parallel_process_group if self.shard_config.enable_sequence_parallelism else None + ) + sp_partial_derived = sp_mode in ["split_gather", "ring"] + + use_flash_attention = self.shard_config.enable_flash_attention + # Currently sp cannot to be used with flashattention + if sp_mode in ["split_gather", "ring", "all_to_all"]: + if use_flash_attention: + warnings.warn( + f"Sequence parallelism mode {sp_mode} need to be used with FlashAttention, will disable FlashAttention automatically." + ) + use_flash_attention = False + + if sp_mode in ["split_gather", "ring"]: + self.append_or_create_method_replacement( + description={ + "forward": get_llama_seq_parallel_model_forward( + sp_mode=sp_mode, sp_size=sp_size, sp_group=sp_group + ), + }, + policy=policy, + target_key=LlamaModel, + ) + self.append_or_create_method_replacement( + description={ + "forward": get_llama_seq_parallel_attention_forward(sp_mode, sp_size, sp_group), + }, + policy=policy, + target_key=LlamaAttention, + ) + elif sp_mode == "all_to_all": + decoder_attribute_replacement = { + "num_heads": self.model.config.num_attention_heads // sp_size, + } + if getattr(self.model.config, "num_key_value_heads", False): + decoder_attribute_replacement["num_key_value_heads"] = self.model.config.num_key_value_heads // sp_size + + policy[LlamaAttention] = ModulePolicyDescription( + attribute_replacement=decoder_attribute_replacement, + ) + self.append_or_create_method_replacement( + description={ + "forward": get_llama_seq_parallel_attention_forward(sp_mode, sp_size, sp_group), + }, + policy=policy, + target_key=LlamaAttention, + ) + self.append_or_create_method_replacement( + description={ + "forward": get_llama_seq_parallel_model_forward( + sp_mode=sp_mode, + sp_size=sp_size, + sp_group=sp_group, + ), + }, + policy=policy, + target_key=LlamaModel, + ) if self.shard_config.enable_tensor_parallelism: decoder_attribute_replacement = { @@ -64,30 +132,37 @@ def module_policy(self) -> Dict[Union[str, nn.Module], ModulePolicyDescription]: SubModuleReplacementDescription( suffix="self_attn.q_proj", target_module=Linear1D_Col, + kwargs=dict(seq_parallel_mode=sp_mode), ), SubModuleReplacementDescription( suffix="self_attn.k_proj", target_module=Linear1D_Col, + kwargs=dict(seq_parallel_mode=sp_mode), ), SubModuleReplacementDescription( suffix="self_attn.v_proj", target_module=Linear1D_Col, + kwargs=dict(seq_parallel_mode=sp_mode), ), SubModuleReplacementDescription( suffix="self_attn.o_proj", target_module=Linear1D_Row, + kwargs=dict(seq_parallel_mode=sp_mode), ), SubModuleReplacementDescription( suffix="mlp.gate_proj", target_module=Linear1D_Col, + kwargs=dict(seq_parallel_mode=sp_mode), ), SubModuleReplacementDescription( suffix="mlp.up_proj", target_module=Linear1D_Col, + kwargs=dict(seq_parallel_mode=sp_mode), ), SubModuleReplacementDescription( suffix="mlp.down_proj", target_module=Linear1D_Row, + kwargs=dict(seq_parallel_mode=sp_mode), ), ], ) @@ -107,10 +182,12 @@ def module_policy(self) -> Dict[Union[str, nn.Module], ModulePolicyDescription]: SubModuleReplacementDescription( suffix="input_layernorm", target_module=norm_cls, + kwargs={"sp_partial_derived": sp_partial_derived}, ), SubModuleReplacementDescription( suffix="post_attention_layernorm", target_module=norm_cls, + kwargs={"sp_partial_derived": sp_partial_derived}, ), ], policy=policy, @@ -121,20 +198,30 @@ def module_policy(self) -> Dict[Union[str, nn.Module], ModulePolicyDescription]: description=SubModuleReplacementDescription( suffix="norm", target_module=norm_cls, + kwargs={"sp_partial_derived": sp_partial_derived}, ), policy=policy, target_key=LlamaModel, ) # use flash attention - if self.shard_config.enable_flash_attention: + if use_flash_attention: self.append_or_create_method_replacement( description={ - "forward": get_llama_flash_attention_forward(self.shard_config), + "forward": get_llama_flash_attention_forward(self.shard_config, sp_mode, sp_group, sp_size), }, policy=policy, target_key=LlamaAttention, ) + if self.pipeline_stage_manager is None: + # replace llama model forward method + self.append_or_create_method_replacement( + description={ + "forward": get_llama_model_forward_for_flash_attn(self.shard_config), + }, + policy=policy, + target_key=LlamaModel, + ) return policy @@ -154,30 +241,20 @@ def set_pipeline_forward(self, model_cls: nn.Module, new_forward: Callable, poli module = self.model.model if stage_manager.is_interleave: - layers_per_stage = self.distribute_layers( - len(module.layers), stage_manager.num_stages * stage_manager.num_model_chunks - ) - stage_manager.stage_indices = Policy.get_stage_index( - layers_per_stage, - stage_manager.stage, - num_model_chunks=stage_manager.num_model_chunks, - num_stages=stage_manager.num_stages, - ) + layers_per_stage = stage_manager.distribute_layers(len(module.layers)) + stage_manager.stage_indices = stage_manager.get_stage_index(layers_per_stage) method_replacement = { "forward": partial(new_forward, stage_manager=stage_manager, shard_config=self.shard_config) } else: - layers_per_stage = Policy.distribute_layers(len(module.layers), stage_manager.num_stages) - stage_index = Policy.get_stage_index(layers_per_stage, stage_manager.stage) + layers_per_stage = stage_manager.distribute_layers(len(module.layers)) + stage_index = stage_manager.get_stage_index(layers_per_stage) method_replacement = { "forward": partial( new_forward, stage_manager=stage_manager, stage_index=stage_index, shard_config=self.shard_config ) } - self.append_or_create_method_replacement( - description=method_replacement, policy=policy, target_key=model_cls - ) self.append_or_create_method_replacement(description=method_replacement, policy=policy, target_key=model_cls) @@ -194,15 +271,8 @@ def get_held_layers(self) -> List[Module]: held_layers = [] if stage_manager.is_interleave: assert stage_manager.num_model_chunks is not None - layers_per_stage = self.distribute_layers( - len(module.layers), stage_manager.num_stages * stage_manager.num_model_chunks - ) - stage_indices = Policy.get_stage_index( - layers_per_stage, - stage_manager.stage, - num_model_chunks=stage_manager.num_model_chunks, - num_stages=stage_manager.num_stages, - ) + layers_per_stage = stage_manager.distribute_layers(len(module.layers)) + stage_indices = stage_manager.get_stage_index(layers_per_stage) if stage_manager.is_first_stage(ignore_chunk=True): held_layers.append(module.embed_tokens) for start_idx, end_idx in stage_indices: @@ -211,10 +281,10 @@ def get_held_layers(self) -> List[Module]: held_layers.append(module.norm) else: - layers_per_stage = self.distribute_layers(len(module.layers), stage_manager.num_stages) + layers_per_stage = stage_manager.distribute_layers(len(module.layers)) if stage_manager.is_first_stage(): held_layers.append(module.embed_tokens) - start_idx, end_idx = self.get_stage_index(layers_per_stage, stage_manager.stage) + start_idx, end_idx = stage_manager.get_stage_index(layers_per_stage) held_layers.extend(module.layers[start_idx:end_idx]) if stage_manager.is_last_stage(): held_layers.append(module.norm) @@ -250,18 +320,23 @@ def module_policy(self): policy = super().module_policy() - setattr(self.shard_config, "causal_lm", True) - - if self.shard_config.enable_tensor_parallelism: + if self.shard_config.enable_tensor_parallelism and not self.shard_config.enable_sequence_parallelism: # add a new item for casual lm new_item = { LlamaForCausalLM: ModulePolicyDescription( sub_module_replacement=[ - SubModuleReplacementDescription(suffix="lm_head", target_module=Linear1D_Col) + SubModuleReplacementDescription( + suffix="lm_head", + target_module=Linear1D_Col, + kwargs={"gather_output": not self.shard_config.parallel_output}, + ) ], - method_replacement={"forward": get_lm_forward_with_dist_cross_entropy(self.shard_config)}, ) } + if self.shard_config.parallel_output: + new_item[LlamaForCausalLM].method_replacement = { + "forward": get_lm_forward_with_dist_cross_entropy(self.shard_config) + } policy.update(new_item) if self.pipeline_stage_manager: diff --git a/colossalai/shardformer/policies/opt.py b/colossalai/shardformer/policies/opt.py index a542808ba794..98e584be861b 100644 --- a/colossalai/shardformer/policies/opt.py +++ b/colossalai/shardformer/policies/opt.py @@ -9,7 +9,12 @@ from .._utils import getattr_ from ..modeling.jit import get_jit_fused_dropout_add_func -from ..modeling.opt import OPTPipelineForwards, get_jit_fused_opt_decoder_layer_forward, get_opt_flash_attention_forward +from ..modeling.opt import ( + OPTPipelineForwards, + get_jit_fused_opt_decoder_layer_forward, + get_opt_decoder_forward_for_flash_attention, + get_opt_flash_attention_forward, +) from .base_policy import ModulePolicyDescription, Policy, SubModuleReplacementDescription __all__ = [ @@ -27,6 +32,7 @@ def __init__(self) -> None: import transformers from packaging.version import Version + # TODO: remove this version check when transformers>=4.36.0 assert Version(transformers.__version__) <= Version( "4.33.0" ), "The OPT model should run on a transformers version not greater than 4.33.0." @@ -111,7 +117,9 @@ def module_policy(self): # optimization configuration self.append_or_create_submodule_replacement( description=SubModuleReplacementDescription( - suffix="final_layer_norm", target_module=norm_cls, ignore_if_not_exist=True + suffix="final_layer_norm", + target_module=norm_cls, + ignore_if_not_exist=True, ), policy=policy, target_key=OPTDecoder, @@ -119,10 +127,14 @@ def module_policy(self): self.append_or_create_submodule_replacement( description=[ SubModuleReplacementDescription( - suffix="self_attn_layer_norm", target_module=norm_cls, ignore_if_not_exist=True + suffix="self_attn_layer_norm", + target_module=norm_cls, + ignore_if_not_exist=True, ), SubModuleReplacementDescription( - suffix="final_layer_norm", target_module=norm_cls, ignore_if_not_exist=True + suffix="final_layer_norm", + target_module=norm_cls, + ignore_if_not_exist=True, ), ], policy=policy, @@ -133,11 +145,19 @@ def module_policy(self): if self.shard_config.enable_flash_attention: self.append_or_create_method_replacement( description={ - "forward": get_opt_flash_attention_forward(), + "forward": get_opt_flash_attention_forward(self.shard_config), }, policy=policy, target_key=OPTAttention, ) + if not self.shard_config.pipeline_stage_manager: + self.append_or_create_method_replacement( + description={ + "forward": get_opt_decoder_forward_for_flash_attention(self.shard_config), + }, + policy=policy, + target_key=OPTDecoder, + ) # use jit fused operator if self.shard_config.enable_jit_fused: @@ -166,12 +186,12 @@ def get_held_layers(self) -> List[nn.Module]: stage_manager = self.pipeline_stage_manager held_layers = [] - layers_per_stage = self.distribute_layers(len(module.layers), stage_manager.num_stages) + layers_per_stage = stage_manager.distribute_layers(len(module.layers)) if stage_manager.is_first_stage(): held_layers.append(module.embed_tokens) held_layers.append(module.embed_positions) held_layers.append(module.project_in) - start_idx, end_idx = self.get_stage_index(layers_per_stage, stage_manager.stage) + start_idx, end_idx = stage_manager.get_stage_index(layers_per_stage) held_layers.extend(module.layers[start_idx:end_idx]) if stage_manager.is_last_stage(): held_layers.append(module.final_layer_norm) @@ -188,9 +208,16 @@ def set_pipeline_forward(self, model_cls: nn.Module, new_forward: Callable, poli else: module = self.model.model.decoder - layers_per_stage = Policy.distribute_layers(len(module.layers), stage_manager.num_stages) - stage_index = Policy.get_stage_index(layers_per_stage, stage_manager.stage) - method_replacement = {"forward": partial(new_forward, stage_manager=stage_manager, stage_index=stage_index)} + layers_per_stage = stage_manager.distribute_layers(len(module.layers)) + stage_index = stage_manager.get_stage_index(layers_per_stage) + method_replacement = { + "forward": partial( + new_forward, + stage_manager=stage_manager, + stage_index=stage_index, + shard_config=self.shard_config, + ) + } self.append_or_create_method_replacement( description=method_replacement, policy=policy, target_key=model_cls ) @@ -203,7 +230,9 @@ def module_policy(self): policy = super().module_policy() if self.pipeline_stage_manager: self.set_pipeline_forward( - model_cls=OPTModel, new_forward=OPTPipelineForwards.opt_model_forward, policy=policy + model_cls=OPTModel, + new_forward=OPTPipelineForwards.opt_model_forward, + policy=policy, ) return policy @@ -223,14 +252,18 @@ def module_policy(self): if self.shard_config.enable_tensor_parallelism: self.append_or_create_submodule_replacement( description=SubModuleReplacementDescription( - suffix="lm_head", target_module=Linear1D_Col, kwargs=dict(gather_output=True) + suffix="lm_head", + target_module=Linear1D_Col, + kwargs=dict(gather_output=True), ), policy=policy, target_key=OPTForCausalLM, ) if self.pipeline_stage_manager: self.set_pipeline_forward( - model_cls=OPTForCausalLM, new_forward=OPTPipelineForwards.opt_for_causal_lm_forward, policy=policy + model_cls=OPTForCausalLM, + new_forward=OPTPipelineForwards.opt_for_causal_lm_forward, + policy=policy, ) return policy @@ -246,7 +279,12 @@ def get_shared_params(self) -> List[Dict[int, Tensor]]: if self.pipeline_stage_manager and self.pipeline_stage_manager.num_stages > 1: num_stages = self.pipeline_stage_manager.num_stages if id(opt_model.model.decoder.embed_tokens.weight) == id(opt_model.lm_head.weight): - return [{0: opt_model.model.decoder.embed_tokens.weight, num_stages - 1: opt_model.lm_head.weight}] + return [ + { + 0: opt_model.model.decoder.embed_tokens.weight, + num_stages - 1: opt_model.lm_head.weight, + } + ] return [] def postprocess(self): diff --git a/colossalai/shardformer/policies/t5.py b/colossalai/shardformer/policies/t5.py index e183b0632f88..0c8ec15fa0a9 100644 --- a/colossalai/shardformer/policies/t5.py +++ b/colossalai/shardformer/policies/t5.py @@ -1,3 +1,5 @@ +from __future__ import annotations + import warnings from functools import partial from typing import Callable, Dict, List, Tuple @@ -241,15 +243,16 @@ def module_policy(self): def postprocess(self): return self.model - @staticmethod def distribute_t5_layers( - num_encoder_layers: int, num_decoder_layers: int, num_stages: int + self, num_encoder_layers: int, num_decoder_layers: int, num_stages: int ) -> Tuple[List[int], int]: """ Distribute t5 layers into stages when pipeline parallel is used. Return the layer distribution as a list and the starting stage of decoder. If decoder doesn't exist, returned decoder starting stage is set to num_encoder_layers. """ + stage_manager = self.pipeline_stage_manager + assert stage_manager is not None, "Pipeline stage manager is not set." # number of encoder layers must be a positive integer if num_encoder_layers <= 0: @@ -261,7 +264,7 @@ def distribute_t5_layers( # in the case of T5EncoderModel, set decoder starting stage to num_stages since it doesn't exist if num_decoder_layers == 0: - return Policy.distribute_layers(num_encoder_layers, num_stages), num_stages + return stage_manager.distribute_layers(num_encoder_layers, num_stages), num_stages # the number of stages distributed between encoder and decoder is optimized in this way: # num_encoder_stages = argmin(abs(num_encoder_layers / encoder_stages - num_decoder_layers / decoder_stages)) @@ -272,22 +275,26 @@ def objective(num_encoder_stages): num_encoder_stages = np.argmin([objective(i) for i in range(1, num_stages)]) + 1 num_decoder_stages = num_stages - num_encoder_stages - encoder_distribution = Policy.distribute_layers(num_encoder_layers, num_encoder_stages) - decoder_distribution = Policy.distribute_layers(num_decoder_layers, num_decoder_stages) + encoder_distribution = stage_manager.distribute_layers(num_encoder_layers, num_encoder_stages) + decoder_distribution = stage_manager.distribute_layers(num_decoder_layers, num_decoder_stages) return encoder_distribution + decoder_distribution, num_encoder_stages - @staticmethod def get_t5_stage_index( - layers_per_stage: List[int], stage: int, decoder_starting_stage: int - ) -> Tuple[bool, int, int]: + self, layers_per_stage: List[int], stage: int, decoder_starting_stage: int + ) -> Tuple[int, int]: """ Input the distribution of layers among stages, the current stage and the first stage of decoder. Return the starting/ending idx of layers in encoder/decoder """ + stage_manager = self.pipeline_stage_manager + assert stage_manager is not None, "Pipeline stage manager is not set." + if stage < decoder_starting_stage: - return Policy.get_stage_index(layers_per_stage[:decoder_starting_stage], stage) + return stage_manager.get_stage_index(layers_per_stage[:decoder_starting_stage], stage) else: - return Policy.get_stage_index(layers_per_stage[decoder_starting_stage:], stage - decoder_starting_stage) + return stage_manager.get_stage_index( + layers_per_stage[decoder_starting_stage:], stage - decoder_starting_stage + ) def get_held_layers(self) -> List[nn.Module]: """Get pipeline layers for current stage.""" @@ -302,12 +309,10 @@ def get_held_layers(self) -> List[nn.Module]: num_decoder_layers = len(decoder.block) if decoder else 0 held_layers = [] - layers_per_stage, decoder_starting_stage = T5BasePolicy.distribute_t5_layers( + layers_per_stage, decoder_starting_stage = self.distribute_t5_layers( num_encoder_layers, num_decoder_layers, stage_manager.num_stages ) - start_idx, end_idx = T5BasePolicy.get_t5_stage_index( - layers_per_stage, stage_manager.stage, decoder_starting_stage - ) + start_idx, end_idx = self.get_t5_stage_index(layers_per_stage, stage_manager.stage, decoder_starting_stage) if stage_manager.stage < decoder_starting_stage: # current stage is in t5's encoder @@ -343,10 +348,10 @@ def set_pipeline_forward(self, model_cls: nn.Module, new_forward: Callable, poli num_encoder_layers = len(encoder.block) num_decoder_layers = len(decoder.block) if decoder else 0 - layers_per_stage, decoder_starting_stage = T5BasePolicy.distribute_t5_layers( + layers_per_stage, decoder_starting_stage = self.distribute_t5_layers( num_encoder_layers, num_decoder_layers, stage_manager.num_stages ) - stage_index = T5BasePolicy.get_t5_stage_index(layers_per_stage, stage_manager.stage, decoder_starting_stage) + stage_index = self.get_t5_stage_index(layers_per_stage, stage_manager.stage, decoder_starting_stage) method_replacement = { "forward": partial( @@ -386,7 +391,7 @@ def get_shared_params(self) -> List[Dict[int, Tensor]]: module = self.model stage_manager = self.pipeline_stage_manager if stage_manager is not None and stage_manager.num_stages > 1: - _, decoder_starting_stage = T5BasePolicy.distribute_t5_layers( + _, decoder_starting_stage = self.distribute_t5_layers( len(module.encoder.block), len(module.decoder.block), stage_manager.num_stages ) @@ -434,7 +439,7 @@ def get_shared_params(self) -> List[Dict[int, Tensor]]: module = self.model stage_manager = self.pipeline_stage_manager if stage_manager is not None and stage_manager.num_stages > 1: - _, decoder_starting_stage = T5BasePolicy.distribute_t5_layers( + _, decoder_starting_stage = self.distribute_t5_layers( len(module.encoder.block), len(module.decoder.block), stage_manager.num_stages ) diff --git a/colossalai/shardformer/policies/vit.py b/colossalai/shardformer/policies/vit.py index 584d4e2652c0..905398c4d51e 100644 --- a/colossalai/shardformer/policies/vit.py +++ b/colossalai/shardformer/policies/vit.py @@ -134,10 +134,10 @@ def get_held_layers(self) -> List[nn.Module]: stage_manager = self.pipeline_stage_manager held_layers = [] - layers_per_stage = self.distribute_layers(len(module.encoder.layer), stage_manager.num_stages) + layers_per_stage = stage_manager.distribute_layers(len(module.encoder.layer)) if stage_manager.is_first_stage(): held_layers.append(module.embeddings) - start_idx, end_idx = self.get_stage_index(layers_per_stage, stage_manager.stage) + start_idx, end_idx = stage_manager.get_stage_index(layers_per_stage) held_layers.extend(module.encoder.layer[start_idx:end_idx]) return held_layers @@ -149,8 +149,8 @@ def set_pipeline_forward(self, model_cls: nn.Module, pipeline_forward: Callable, else: module = self.model.vit - layers_per_stage = Policy.distribute_layers(len(module.encoder.layer), stage_manager.num_stages) - stage_index = Policy.get_stage_index(layers_per_stage, stage_manager.stage) + layers_per_stage = stage_manager.distribute_layers(len(module.encoder.layer)) + stage_index = stage_manager.get_stage_index(layers_per_stage) method_replacement = {"forward": pipeline_forward(stage_manager=stage_manager, stage_index=stage_index)} self.append_or_create_method_replacement( description=method_replacement, policy=policy, target_key=model_cls diff --git a/colossalai/shardformer/policies/whisper.py b/colossalai/shardformer/policies/whisper.py index b5b5db79d9de..c63f6d1cc549 100644 --- a/colossalai/shardformer/policies/whisper.py +++ b/colossalai/shardformer/policies/whisper.py @@ -13,6 +13,7 @@ WhisperPipelineForwards, get_jit_fused_whisper_decoder_layer_forward, get_jit_fused_whisper_encoder_layer_forward, + get_whisper_decoder_forward_for_flash_attention, get_whisper_flash_attention_forward, ) from .base_policy import ModulePolicyDescription, Policy, SubModuleReplacementDescription @@ -31,6 +32,7 @@ def __init__(self) -> None: import transformers from packaging.version import Version + # TODO: remove this version check when transformers>=4.36.0 assert Version(transformers.__version__) <= Version( "4.33.0" ), "The Whisper model should run on a transformers version not greater than 4.33.0." @@ -240,6 +242,14 @@ def module_policy(self): policy=policy, target_key=WhisperAttention, ) + if not self.shard_config.pipeline_stage_manager: + self.append_or_create_method_replacement( + description={ + "forward": get_whisper_decoder_forward_for_flash_attention(self.shard_config), + }, + policy=policy, + target_key=WhisperDecoder, + ) # use jit fused operator if self.shard_config.enable_jit_fused: @@ -269,7 +279,9 @@ def add_lm_head_policy(self, base_policy): if self.shard_config.enable_tensor_parallelism: self.append_or_create_submodule_replacement( description=SubModuleReplacementDescription( - suffix="proj_out", target_module=col_nn.Linear1D_Col, kwargs={"gather_output": True} + suffix="proj_out", + target_module=col_nn.Linear1D_Col, + kwargs={"gather_output": True}, ), policy=base_policy, target_key=WhisperForConditionalGeneration, @@ -280,15 +292,16 @@ def add_lm_head_policy(self, base_policy): def postprocess(self): return self.model - @staticmethod def distribute_whisper_layers( - num_encoder_layers: int, num_decoder_layers: int, num_stages: int + self, num_encoder_layers: int, num_decoder_layers: int, num_stages: int ) -> Tuple[List[int], int]: """ Distribute whisper layers into stages when pipeline parallel is used. Return the layer distribution as a list and the starting stage of decoder. If decoder doesn't exist, returned decoder starting stage is set to num_encoder_layers. """ + stage_manager = self.pipeline_stage_manager + assert stage_manager is not None, "pipeline_stage_manager is None" # number of encoder layers must be a positive integer if num_encoder_layers <= 0: @@ -300,7 +313,7 @@ def distribute_whisper_layers( # in the case of whisperEncoderModel, set decoder starting stage to num_stages since it doesn't exist if num_decoder_layers == 0: - return Policy.distribute_layers(num_encoder_layers, num_stages), num_stages + return stage_manager.distribute_layers(num_encoder_layers, num_stages), num_stages # the number of stages distributed between encoder and decoder is optimized in this way: # num_encoder_stages = argmin(abs(num_encoder_layers / encoder_stages - num_decoder_layers / decoder_stages)) @@ -311,22 +324,27 @@ def objective(num_encoder_stages): num_encoder_stages = np.argmin([objective(i) for i in range(1, num_stages)]) + 1 num_decoder_stages = num_stages - num_encoder_stages - encoder_distribution = Policy.distribute_layers(num_encoder_layers, num_encoder_stages) - decoder_distribution = Policy.distribute_layers(num_decoder_layers, num_decoder_stages) + encoder_distribution = stage_manager.distribute_layers(num_encoder_layers, num_encoder_stages) + decoder_distribution = stage_manager.distribute_layers(num_decoder_layers, num_decoder_stages) return encoder_distribution + decoder_distribution, num_encoder_stages - @staticmethod def get_whisper_stage_index( - layers_per_stage: List[int], stage: int, decoder_starting_stage: int - ) -> Tuple[bool, int, int]: + self, layers_per_stage: List[int], stage: int, decoder_starting_stage: int + ) -> Tuple[int, int]: """ Input the distribution of layers among stages, the current stage and the first stage of decoder. Return the starting/ending idx of layers in encoder/decoder """ + stage_manager = self.pipeline_stage_manager + assert stage_manager is not None, "pipeline_stage_manager is None" + if stage < decoder_starting_stage: - return Policy.get_stage_index(layers_per_stage[:decoder_starting_stage], stage) + return stage_manager.get_stage_index(layers_per_stage[:decoder_starting_stage], stage) else: - return Policy.get_stage_index(layers_per_stage[decoder_starting_stage:], stage - decoder_starting_stage) + return stage_manager.get_stage_index( + layers_per_stage[decoder_starting_stage:], + stage - decoder_starting_stage, + ) def get_held_layers(self) -> List[nn.Module]: assert self.pipeline_stage_manager is not None, "pipeline_stage_manager is None" @@ -354,12 +372,10 @@ def get_held_layers(self) -> List[nn.Module]: num_decoder_layers = 0 held_layers = [] - layers_per_stage, decoder_starting_stage = WhisperPolicy.distribute_whisper_layers( + layers_per_stage, decoder_starting_stage = self.distribute_whisper_layers( num_encoder_layers, num_decoder_layers, stage_manager.num_stages ) - start_idx, end_idx = WhisperPolicy.get_whisper_stage_index( - layers_per_stage, stage_manager.stage, decoder_starting_stage - ) + start_idx, end_idx = self.get_whisper_stage_index(layers_per_stage, stage_manager.stage, decoder_starting_stage) if stage_manager.stage < decoder_starting_stage: # current stage is in whisper's encoder @@ -409,12 +425,10 @@ def set_pipeline_forward(self, model_cls: nn.Module, new_forward: Callable, poli else: num_decoder_layers = 0 - layers_per_stage, decoder_starting_stage = WhisperPolicy.distribute_whisper_layers( + layers_per_stage, decoder_starting_stage = self.distribute_whisper_layers( num_encoder_layers, num_decoder_layers, stage_manager.num_stages ) - stage_index = WhisperPolicy.get_whisper_stage_index( - layers_per_stage, stage_manager.stage, decoder_starting_stage - ) + stage_index = self.get_whisper_stage_index(layers_per_stage, stage_manager.stage, decoder_starting_stage) method_replacement = { "forward": partial( @@ -422,6 +436,7 @@ def set_pipeline_forward(self, model_cls: nn.Module, new_forward: Callable, poli stage_manager=stage_manager, stage_index=stage_index, decoder_starting_stage=decoder_starting_stage, + shard_config=self.shard_config, ) } self.append_or_create_method_replacement(description=method_replacement, policy=policy, target_key=model_cls) @@ -436,7 +451,9 @@ def module_policy(self): if self.pipeline_stage_manager is not None: self.set_pipeline_forward( - model_cls=WhisperModel, new_forward=WhisperPipelineForwards.whisper_model_forward, policy=policy + model_cls=WhisperModel, + new_forward=WhisperPipelineForwards.whisper_model_forward, + policy=policy, ) return policy @@ -493,7 +510,7 @@ def get_shared_params(self) -> List[Dict[int, Tensor]]: stage_manager = self.pipeline_stage_manager if stage_manager is not None and stage_manager.num_stages > 1: - _, decoder_starting_stage = WhisperPolicy.distribute_whisper_layers( + _, decoder_starting_stage = self.distribute_whisper_layers( num_encoder_layers, num_decoder_layers, stage_manager.num_stages ) shared_params = [] diff --git a/colossalai/shardformer/shard/__init__.py b/colossalai/shardformer/shard/__init__.py index acf8a95a41ca..dff2118c1c1a 100644 --- a/colossalai/shardformer/shard/__init__.py +++ b/colossalai/shardformer/shard/__init__.py @@ -1,5 +1,6 @@ +from .grad_ckpt_config import GradientCheckpointConfig, PipelineGradientCheckpointConfig from .shard_config import ShardConfig from .sharder import ModelSharder from .shardformer import ShardFormer -__all__ = ["ShardConfig", "ModelSharder", "ShardFormer"] +__all__ = ["ShardConfig", "ModelSharder", "ShardFormer", "PipelineGradientCheckpointConfig", "GradientCheckpointConfig"] diff --git a/colossalai/shardformer/shard/grad_ckpt_config.py b/colossalai/shardformer/shard/grad_ckpt_config.py new file mode 100644 index 000000000000..9c6c2b54ea39 --- /dev/null +++ b/colossalai/shardformer/shard/grad_ckpt_config.py @@ -0,0 +1,87 @@ +from dataclasses import dataclass +from typing import List, Optional + + +@dataclass +class GradientCheckpointConfig: + gradient_checkpointing_ratio: float = 0.0 + + def get_num_ckpt_layers(self, num_layers: int) -> int: + return int(self.gradient_checkpointing_ratio * num_layers) + + +@dataclass +class PipelineGradientCheckpointConfig(GradientCheckpointConfig): + r""" + The pipeline gradient config is designed to provide more flexibility for users to control gradient checkpoint in pipeline parallelism. + Combined with PipelineStageManager.set_distribution_config, user can fully control the distribution of layers and checkpointed layers in pipeline parallelism. + Refer to https://github.com/hpcaitech/ColossalAI/issues/5509 for more details. + + It provides the following features: + 1. `gradient_checkpointing_ratio`: This is used to control gradient checkpointing more precisely, e.g., set 50% of the layers to use gradient checkpointing. + 2. Customize # ckpt layers assigned to each stage. This takes precedence over `gradient_checkpointing_ratio`. + + """ + """ + Args: + gradient_checkpointing_ratio (Optional[float]): The ratio of gradient checkpointing. It can only be used in pipeline parallelism. Defaults to None. + num_stages (Optional[int]): Number of stages in the pipeline. Defaults to None. For sanity check. + num_model_chunks (Optional[int]): Number of model chunks (1F1B or Interleaved). Defaults to None. For sanity check. + num_model_layers (Optional[int]): Number of model layers. Defaults to None. For sanity check. + num_ckpt_layers_per_stage (Optional[List[int]]): Number of checkpointed layers for each stage. Defaults to None. + + Example 1: + num_stages = 8 + num_layers = 80 + num_model_chunks = 1 + num_layers_per_stage = [9, 9, 9, 10, 11, 10, 11, 11] + num_ckpt_layers_per_stage = [4, 4, 2, 2, 0, 0, 0, 0] + + Example 2: + num_stages = 4 + num_layers = 80 + num_model_chunks = 2 + num_layers_per_stage = [9, 9, 9, 10, 11, 10, 11, 11] + # device 0 holds num_layers_per_stage[0] and num_layers_per_stage[4] layers + ... + + """ + num_stages: Optional[int] = None + num_model_chunks: Optional[int] = None + num_model_layers: Optional[int] = None + num_ckpt_layers_per_stage: Optional[List[int]] = None + + def __post_init__(self): + if self._enable_gradient_checkpointing_ratio: + if not (0 <= self.gradient_checkpointing_ratio <= 1): + raise ValueError("gradient_checkpointing_ratio should be in 0% to 100%") + + if self._enable_customized_ckpt_layers_per_stage: + assert ( + self.num_stages is not None and self.num_model_chunks is not None and self.num_model_layers is not None + ) + assert len(self.num_ckpt_layers_per_stage) == self.num_stages * self.num_model_chunks + assert all( + [0 <= num_ckpt_layers < self.num_model_layers for num_ckpt_layers in self.num_ckpt_layers_per_stage] + ) + self.gradient_checkpointing_ratio = sum(self.num_ckpt_layers_per_stage) / self.num_model_layers + + @property + def _enable_gradient_checkpointing_ratio(self) -> bool: + return self.gradient_checkpointing_ratio is not None + + @property + def _enable_customized_ckpt_layers_per_stage(self) -> bool: + return self.num_ckpt_layers_per_stage is not None + + def get_num_ckpt_layers(self, stage: int, num_layers: int, model_chunk_id: int = 0) -> int: + if not self._enable_gradient_checkpointing_ratio and not self._enable_customized_ckpt_layers_per_stage: + raise RuntimeError("No checkpointed layers information is provided") + + if self._enable_customized_ckpt_layers_per_stage: + assert stage <= self.num_stages and model_chunk_id <= self.num_model_chunks + num_ckpt_layers = self.num_ckpt_layers_per_stage[stage + model_chunk_id * self.num_stages] + assert num_ckpt_layers <= num_layers + return num_ckpt_layers + else: + return int(self.gradient_checkpointing_ratio * num_layers) diff --git a/colossalai/shardformer/shard/shard_config.py b/colossalai/shardformer/shard/shard_config.py index da27341d9c29..7489873c2ed6 100644 --- a/colossalai/shardformer/shard/shard_config.py +++ b/colossalai/shardformer/shard/shard_config.py @@ -1,3 +1,4 @@ +import warnings from dataclasses import dataclass, field from typing import Any, Dict, Optional @@ -6,7 +7,10 @@ from colossalai.pipeline.stage_manager import PipelineStageManager +from .grad_ckpt_config import GradientCheckpointConfig + __all__ = ["ShardConfig"] +SUPPORT_SP_MODE = ["split_gather", "ring", "all_to_all"] @dataclass @@ -23,18 +27,22 @@ class ShardConfig: enable_jit_fused (bool, optional): Whether to switch on JIT fused operators. Defaults to False. enable_sequence_parallelism (bool): Whether to turn on sequence parallelism, which partitions non-tensor-parallel regions along the sequence dimension. Defaults to False. enable_sequence_overlap (bool): Whether to turn on sequence overlap, which overlap the computation and communication in sequence parallelism. It can only be used when enable_sequence_parallelism is True. Defaults to False. + gradient_checkpoint_config (Optional[GradientCheckpointConfig]): The gradient checkpoint config. Defaults to None. enable_all_optimization (bool): Whether to turn on all optimization tools including 'fused normalization', 'flash attention', 'JIT fused operators', 'sequence parallelism' and 'sequence overlap'. Defaults to False. """ tensor_parallel_process_group: Optional[ProcessGroup] = None + sequence_parallel_process_group: Optional[ProcessGroup] = None pipeline_stage_manager: Optional[PipelineStageManager] = None enable_tensor_parallelism: bool = True + enable_all_optimization: bool = False enable_fused_normalization: bool = False enable_flash_attention: bool = False enable_jit_fused: bool = False - enable_all_optimization: bool = False enable_sequence_parallelism: bool = False + sequence_parallelism_mode: str = None enable_sequence_overlap: bool = False parallel_output: bool = True + gradient_checkpoint_config: Optional[GradientCheckpointConfig] = None extra_kwargs: Dict[str, Any] = field(default_factory=dict) # TODO padding vocab # make_vocab_size_divisible_by: int = 128 @@ -46,21 +54,56 @@ class ShardConfig: def tensor_parallel_size(self): return self._tensor_parallel_size + @property + def sequence_parallel_size(self): + return self._sequence_parallel_size + def __post_init__(self): - if not self.enable_tensor_parallelism and self.enable_sequence_parallelism: - raise ValueError( - "enable_sequence_parallelism can only be set to True when enable_tensor_parallelism is True" + # turn on all optimization if all_optimization is set to True + if self.enable_all_optimization: + self._turn_on_all_optimization() + + if self.enable_sequence_parallelism: + self.sequence_parallelism_mode = ( + "split_gather" if self.sequence_parallelism_mode is None else self.sequence_parallelism_mode ) - if not self.enable_sequence_parallelism and self.enable_sequence_overlap: - raise ValueError("enable_sequence_overlap can only be set to True when enable_sequence_parallelism is True") + assert ( + self.sequence_parallelism_mode in SUPPORT_SP_MODE + ), f"Sequence parallelism mode {self.sequence_parallelism_mode} is not in the supported list {SUPPORT_SP_MODE}" + if self.sequence_parallelism_mode in ["split_gather", "ring"]: + assert ( + self.enable_tensor_parallelism + ), f"sequence parallelism mode {self.sequence_parallelism_mode} can only be used when enable_tensor_parallelism is True" + elif self.sequence_parallelism_mode in ["all_to_all"]: + assert ( + not self.enable_tensor_parallelism + ), f"sequence parallelism mode {self.sequence_parallelism_mode} can only be used when enable_tensor_parallelism is False" + if self.enable_sequence_overlap: + self.enable_sequence_overlap = False + warnings.warn( + f"The enable_sequence_overlap flag will be ignored in sequence parallelism mode {self.sequence_parallelism_mode}" + ) + else: + if self.sequence_parallelism_mode: + self.sequence_parallelism_mode = None + warnings.warn( + f"The sequence_parallelism_mode will be ignored when enable_sequence_parallelism is False" + ) + assert ( + not self.enable_sequence_overlap + ), f"enable_sequence_overlap can only be set to True when enable_sequence_parallelism is True" + + # get the tensor parallel size if not self.enable_tensor_parallelism: self._tensor_parallel_size = 1 else: - # get the parallel size self._tensor_parallel_size = dist.get_world_size(self.tensor_parallel_process_group) - # turn on all optimization if all_optimization is set to True - if self.enable_all_optimization: - self._turn_on_all_optimization() + + # get the sequence parallel size + if not self.enable_sequence_parallelism: + self._sequence_parallel_size = 1 + else: + self._sequence_parallel_size = dist.get_world_size(self.sequence_parallel_process_group) def _turn_on_all_optimization(self): """ @@ -70,8 +113,10 @@ def _turn_on_all_optimization(self): self.enable_fused_normalization = True self.enable_flash_attention = True self.enable_jit_fused = True - self.enable_sequence_parallelism = True - self.enable_sequence_overlap = True + # This can cause non-in-place param sharding when used without ZeRO. + # It may also slow down training when seq len is small. Plz enable manually. + # self.enable_sequence_parallelism = True + # self.enable_sequence_overlap = True def _infer(self): """ diff --git a/colossalai/tensor/d_tensor/layout_converter.py b/colossalai/tensor/d_tensor/layout_converter.py index abe4a86d8198..667a7b78e4f5 100644 --- a/colossalai/tensor/d_tensor/layout_converter.py +++ b/colossalai/tensor/d_tensor/layout_converter.py @@ -440,7 +440,10 @@ def layout_converting( total_steps = 0 transform_path = [] comm_action_sequence: List[CommSpec] = [] - spec_pairs = (str(source_spec.sharding_sequence), str(target_spec.sharding_sequence)) + + src_shape = source_layout.get_sharded_shape_per_device() + dst_shape = target_layout.get_sharded_shape_per_device() + spec_pairs = ((str(source_spec.sharding_sequence), src_shape), (str(target_spec.sharding_sequence), dst_shape)) if spec_pairs in self.cached_solution: # Solution Cache hit diff --git a/colossalai/testing/comparison.py b/colossalai/testing/comparison.py index 4f2a4878e7ce..e415b5fc3aa3 100644 --- a/colossalai/testing/comparison.py +++ b/colossalai/testing/comparison.py @@ -40,7 +40,12 @@ def assert_equal_in_group(tensor: Tensor, process_group: ProcessGroup = None): assert torch.all(a == b), f"expected tensors on rank {i} and {i + 1} to be equal but they are not, {a} vs {b}" -def check_state_dict_equal(d1: OrderedDict, d2: OrderedDict, ignore_device: bool = True, ignore_dtype: bool = False): +def check_state_dict_equal( + d1: OrderedDict, + d2: OrderedDict, + ignore_device: bool = True, + ignore_dtype: bool = False, +): assert len(list(d1.keys())) == len( list(d2.keys()) ), f"Number of keys unequal: {len(list(d1.keys()))} vs {len(list(d2.keys()))}" @@ -94,7 +99,12 @@ def check_state_dict_equal_pytree(d1: OrderedDict, d2: OrderedDict, ignore_devic def assert_hf_output_close( - out1: Any, out2: Any, ignore_keys: List[str] = None, track_name: str = "", atol=1e-5, rtol=1e-5 + out1: Any, + out2: Any, + ignore_keys: List[str] = None, + track_name: str = "", + atol=1e-5, + rtol=1e-5, ): """ Check if two outputs from huggingface are equal. @@ -113,7 +123,12 @@ def assert_hf_output_close( if ignore_keys is not None and k in ignore_keys: continue assert_hf_output_close( - out1[k], out2[k], track_name=f"{track_name}.{k}", ignore_keys=ignore_keys, atol=atol, rtol=rtol + out1[k], + out2[k], + track_name=f"{track_name}.{k}", + ignore_keys=ignore_keys, + atol=atol, + rtol=rtol, ) elif isinstance(out1, (list, tuple)) and isinstance(out2, (list, tuple)): # if two values are list @@ -121,12 +136,17 @@ def assert_hf_output_close( assert len(out1) == len(out2) for i in range(len(out1)): assert_hf_output_close( - out1[i], out2[i], track_name=f"{track_name}.{i}", ignore_keys=ignore_keys, atol=atol, rtol=rtol + out1[i], + out2[i], + track_name=f"{track_name}.{i}", + ignore_keys=ignore_keys, + atol=atol, + rtol=rtol, ) elif isinstance(out1, Tensor) and isinstance(out2, Tensor): if out1.shape != out2.shape: raise AssertionError(f"{track_name}: shape mismatch: {out1.shape} vs {out2.shape}") - assert torch.allclose( + assert_close( out1, out2, atol=atol, rtol=rtol ), f"{track_name}: tensor value mismatch\nvalue 1: {out1}\nvalue 2: {out2}, \nmean error: {torch.abs(out1 - out2).mean()}" else: diff --git a/colossalai/zero/low_level/low_level_optim.py b/colossalai/zero/low_level/low_level_optim.py index a2433d1b261c..bbbaf13b53ef 100644 --- a/colossalai/zero/low_level/low_level_optim.py +++ b/colossalai/zero/low_level/low_level_optim.py @@ -79,6 +79,7 @@ def __init__( master_weights: bool = True, # master weights ): super(LowLevelZeroOptimizer, self).__init__(optim=optimizer) + self._dtype = self.optim.param_groups[0]["params"][0].dtype self._logger = get_dist_logger() self._verbose = verbose @@ -494,7 +495,6 @@ def backward(self, loss, retain_graph=False): # clear reduced grads if self._overlap_communication: get_accelerator().synchronize() - self.zero_grad() def backward_by_grad(self, tensor, grad): diff --git a/docs/README-zh-Hans.md b/docs/README-zh-Hans.md index 93045ea6adc6..6d243a80852d 100644 --- a/docs/README-zh-Hans.md +++ b/docs/README-zh-Hans.md @@ -24,6 +24,7 @@
      ## 新闻 +* [2024/03] [314 Billion Parameter Grok-1 Inference Accelerated by 3.8x, Efficient and Easy-to-Use PyTorch+HuggingFace version is Here](https://hpc-ai.com/blog/314-billion-parameter-grok-1-inference-accelerated-by-3.8x-efficient-and-easy-to-use-pytorchhuggingface-version-is-here) * [2024/03] [Open-Sora: Revealing Complete Model Parameters, Training Details, and Everything for Sora-like Video Generation Models](https://hpc-ai.com/blog/open-sora-v1.0) * [2024/03] [Open-Sora：Sora Replication Solution with 46% Cost Reduction, Sequence Expansion to Nearly a Million](https://hpc-ai.com/blog/open-sora) * [2024/01] [Inference Performance Improved by 46%, Open Source Solution Breaks the Length Limit of LLM for Multi-Round Conversations](https://hpc-ai.com/blog/Colossal-AI-SwiftInfer) @@ -71,6 +72,7 @@
    • 推理
        +
      • Grok-1: 3140亿参数PyTorch + HuggingFace推理
      • SwiftInfer:打破LLM多轮对话的长度限制，推理加速46%
      • GPT-3
      • 1750亿参数OPT在线推理服务
        • @@ -358,6 +360,18 @@ Colossal-AI 为您提供了一系列并行组件。我们的目标是让您的 ## 推理 +### Grok-1 +

        + +

        + + - 3140亿参数Grok-1推理加速3.8倍，高效易用的PyTorch+HuggingFace版 + +[[代码]](https://github.com/hpcaitech/ColossalAI/tree/main/examples/language/grok-1) +[[博客]](https://hpc-ai.com/blog/314-billion-parameter-grok-1-inference-accelerated-by-3.8x-efficient-and-easy-to-use-pytorchhuggingface-version-is-here) +[[HuggingFace Grok-1 PyTorch 模型权重]](https://huggingface.co/hpcai-tech/grok-1) +[[ModelScope Grok-1 PyTorch 模型权重]](https://www.modelscope.cn/models/colossalai/grok-1-pytorch/summary) +

        diff --git a/docs/source/en/advanced_tutorials/train_gpt_using_hybrid_parallelism.md b/docs/source/en/advanced_tutorials/train_gpt_using_hybrid_parallelism.md index e87eafb6eec7..0133dfd86ddf 100644 --- a/docs/source/en/advanced_tutorials/train_gpt_using_hybrid_parallelism.md +++ b/docs/source/en/advanced_tutorials/train_gpt_using_hybrid_parallelism.md @@ -178,7 +178,7 @@ def train_epoch( for _ in pbar: if use_pipeline: outputs = booster.execute_pipeline( - train_dataloader_iter, model, _criterion, optimizer, return_loss=True, return_outputs=True + train_dataloader_iter, model, _criterion, optimizer, return_loss=True ) # Backward and optimize if is_pp_last_stage: diff --git a/docs/source/en/advanced_tutorials/train_vit_with_hybrid_parallelism.md b/docs/source/en/advanced_tutorials/train_vit_with_hybrid_parallelism.md index 93fed61c34da..dfc2cd596d79 100644 --- a/docs/source/en/advanced_tutorials/train_vit_with_hybrid_parallelism.md +++ b/docs/source/en/advanced_tutorials/train_vit_with_hybrid_parallelism.md @@ -231,7 +231,7 @@ def run_forward_backward( if isinstance(booster.plugin, HybridParallelPlugin) and booster.plugin.pp_size > 1: # run pipeline forward backward when enabling pp in hybrid parallel plugin output_dict = booster.execute_pipeline( - data_iter, model, criterion, optimizer, return_loss=True, return_outputs=True + data_iter, model, criterion, optimizer, return_loss=True ) loss, outputs = output_dict["loss"], output_dict["outputs"] else: diff --git a/docs/source/en/features/pipeline_parallel.md b/docs/source/en/features/pipeline_parallel.md index 31b20335e529..d6f3cdfafe16 100644 --- a/docs/source/en/features/pipeline_parallel.md +++ b/docs/source/en/features/pipeline_parallel.md @@ -198,8 +198,7 @@ def train_epoch(epoch: int, model: nn.Module, optimizer: Optimizer, _criterion: model, _criterion, optimizer, - return_loss=True, - return_outputs=True) + return_loss=True) # Backward and optimize if is_pp_last_stage: loss = outputs['loss'] diff --git a/docs/source/en/features/shardformer.md b/docs/source/en/features/shardformer.md index 1e633ebc0f7c..672945ea2336 100644 --- a/docs/source/en/features/shardformer.md +++ b/docs/source/en/features/shardformer.md @@ -271,7 +271,7 @@ However, if pipeline parallel is enabled, there are several usages different fro 3. Do forward and backward passing through calling `Booster.execute_pipeline` method: ```python outputs = booster.execute_pipeline( - train_dataloader_iter, model, _criterion, optimizer, return_loss=True, return_outputs=True + train_dataloader_iter, model, _criterion, optimizer, return_loss=True ) ``` Backward passing has been completed by this method, so there is no need to call `loss.backward()` after executing this method. diff --git a/docs/source/zh-Hans/advanced_tutorials/train_gpt_using_hybrid_parallelism.md b/docs/source/zh-Hans/advanced_tutorials/train_gpt_using_hybrid_parallelism.md index ae941b489b90..cf7d191723e1 100644 --- a/docs/source/zh-Hans/advanced_tutorials/train_gpt_using_hybrid_parallelism.md +++ b/docs/source/zh-Hans/advanced_tutorials/train_gpt_using_hybrid_parallelism.md @@ -175,7 +175,7 @@ def train_epoch( for _ in pbar: if use_pipeline: outputs = booster.execute_pipeline( - train_dataloader_iter, model, _criterion, optimizer, return_loss=True, return_outputs=True + train_dataloader_iter, model, _criterion, optimizer, return_loss=True ) # Backward and optimize if is_pp_last_stage: diff --git a/docs/source/zh-Hans/advanced_tutorials/train_vit_with_hybrid_parallelism.md b/docs/source/zh-Hans/advanced_tutorials/train_vit_with_hybrid_parallelism.md index c25b109e7106..f32f6c367fe3 100644 --- a/docs/source/zh-Hans/advanced_tutorials/train_vit_with_hybrid_parallelism.md +++ b/docs/source/zh-Hans/advanced_tutorials/train_vit_with_hybrid_parallelism.md @@ -234,7 +234,7 @@ def run_forward_backward( if isinstance(booster.plugin, HybridParallelPlugin) and booster.plugin.pp_size > 1: # run pipeline forward backward when enabling pp in hybrid parallel plugin output_dict = booster.execute_pipeline( - data_iter, model, criterion, optimizer, return_loss=True, return_outputs=True + data_iter, model, criterion, optimizer, return_loss=True ) loss, outputs = output_dict["loss"], output_dict["outputs"] else: diff --git a/docs/source/zh-Hans/features/pipeline_parallel.md b/docs/source/zh-Hans/features/pipeline_parallel.md index e688020556d8..38e1fbfc56b4 100644 --- a/docs/source/zh-Hans/features/pipeline_parallel.md +++ b/docs/source/zh-Hans/features/pipeline_parallel.md @@ -193,8 +193,7 @@ def train_epoch(epoch: int, model: nn.Module, optimizer: Optimizer, _criterion: model, _criterion, optimizer, - return_loss=True, - return_outputs=True) + return_loss=True) # Backward and optimize if is_pp_last_stage: loss = outputs['loss'] diff --git a/docs/source/zh-Hans/features/shardformer.md b/docs/source/zh-Hans/features/shardformer.md index 972c48b0c7a2..a7bcbd9f2816 100644 --- a/docs/source/zh-Hans/features/shardformer.md +++ b/docs/source/zh-Hans/features/shardformer.md @@ -264,7 +264,7 @@ elif args.plugin == "hybrid_parallel": 3. 通过调用`Booster.execute_pipeline` 方法来执行前向和后向传递: ```python outputs = booster.execute_pipeline( - train_dataloader_iter, model, _criterion, optimizer, return_loss=True, return_outputs=True + train_dataloader_iter, model, _criterion, optimizer, return_loss=True ) ``` 该方法会自动执行后向传递，所以在执行该方法后不需要再调用 `loss.backward()`方法。 diff --git a/examples/images/vit/vit_benchmark.py b/examples/images/vit/vit_benchmark.py index 0780173241aa..fdae9ee01537 100644 --- a/examples/images/vit/vit_benchmark.py +++ b/examples/images/vit/vit_benchmark.py @@ -119,9 +119,7 @@ def criterion(outputs, inputs): if hasattr(booster.plugin, "stage_manager") and booster.plugin.stage_manager is not None: # run pipeline forward backward batch = iter([batch]) - outputs = booster.execute_pipeline( - batch, model, criterion, optimizer, return_loss=True, return_outputs=True - ) + outputs = booster.execute_pipeline(batch, model, criterion, optimizer, return_loss=True) else: outputs = model(**batch) loss = criterion(outputs, None) diff --git a/examples/language/bert/finetune.py b/examples/language/bert/finetune.py index 0b1e77ffff06..bd6c393a7ddc 100644 --- a/examples/language/bert/finetune.py +++ b/examples/language/bert/finetune.py @@ -148,7 +148,7 @@ def train_epoch( for _ in pbar: if use_pipeline: outputs = booster.execute_pipeline( - train_dataloader_iter, model, _criterion, optimizer, return_loss=True, return_outputs=True + train_dataloader_iter, model, _criterion, optimizer, return_loss=True ) # Backward and optimize if is_pp_last_device: diff --git a/examples/language/gpt/hybridparallelism/finetune.py b/examples/language/gpt/hybridparallelism/finetune.py index eb56ee530a0a..888f47aaaab0 100644 --- a/examples/language/gpt/hybridparallelism/finetune.py +++ b/examples/language/gpt/hybridparallelism/finetune.py @@ -145,7 +145,7 @@ def train_epoch( for _ in pbar: if use_pipeline: outputs = booster.execute_pipeline( - train_dataloader_iter, model, _criterion, optimizer, return_loss=True, return_outputs=True + train_dataloader_iter, model, _criterion, optimizer, return_loss=True ) # Backward and optimize if is_pp_last_stage: diff --git a/examples/language/grok-1/README.md b/examples/language/grok-1/README.md index c523f941262d..4c37934a0c4d 100644 --- a/examples/language/grok-1/README.md +++ b/examples/language/grok-1/README.md @@ -1,6 +1,17 @@ # Grok-1 Inference -## Install + - 314 Billion Parameter Grok-1 Inference Accelerated by 3.8x, an easy-to-use Python + PyTorch + HuggingFace version for Inference. + +[[code]](https://github.com/hpcaitech/ColossalAI/tree/main/examples/language/grok-1) +[[blog]](https://hpc-ai.com/blog/314-billion-parameter-grok-1-inference-accelerated-by-3.8x-efficient-and-easy-to-use-pytorchhuggingface-version-is-here) +[[HuggingFace Grok-1 PyTorch model weights]](https://huggingface.co/hpcai-tech/grok-1) +[[ModelScope Grok-1 PyTorch model weights]](https://www.modelscope.cn/models/colossalai/grok-1-pytorch/summary) + +

        + +

        + +## Installation ```bash # Make sure you install colossalai from the latest source code @@ -11,33 +22,36 @@ cd examples/language/grok-1 pip install -r requirements.txt ``` -## Tokenizer preparation - -You should download the tokenizer from the official grok-1 repository. - -```bash -wget https://github.com/xai-org/grok-1/raw/main/tokenizer.model -``` - ## Inference You need 8x A100 80GB or equivalent GPUs to run the inference. -We provide two scripts for inference. `run_inference_fast.sh` uses tensor parallelism provided by ColossalAI, and it is faster. `run_inference_slow.sh` uses auto device provided by transformers, and it is slower. +We provide two scripts for inference. `run_inference_fast.sh` uses tensor parallelism provided by ColossalAI, which is faster for generation, while `run_inference_slow.sh` uses auto device provided by transformers, which is relatively slower. -Command format: +Command example: ```bash -./run_inference_fast.sh -./run_inference_slow.sh +./run_inference_fast.sh +./run_inference_slow.sh ``` -`model_name_or_path` can be a local path or a model name from Hugging Face model hub. We provided weights on model hub, named `hpcaitech/grok-1`. - -Command example: - +`MODEL_NAME_OR_PATH` can be a model name from Hugging Face model hub or a local path to PyTorch-version model checkpoints. We have provided pytorch-version checkpoint on [HuggingFace model hub](https://huggingface.co/hpcai-tech/grok-1), named `hpcai-tech/grok-1`. And you could also download the weights in advance using `git`: ```bash -./run_inference_fast.sh hpcaitech/grok-1 tokenizer.model +git lfs install +git clone https://huggingface.co/hpcai-tech/grok-1 ``` -It will take 5-10 minutes to load checkpoints. Don't worry, it's not stuck. +It will take, depending on your Internet speed, several hours to tens of hours to download checkpoints (about 600G!), and 5-10 minutes to load checkpoints when it's ready to launch the inference. Don't worry, it's not stuck. + + +## Performance + +For request of batch size set to 1 and maximum length set to 100: + +| Method | Initialization-Duration(sec) | Average-Generation-Latency(sec) | +|-------------------------|------------------------------|---------------------------------| +| ColossalAI | 431.45 | 14.92 | +| HuggingFace Auto-Device | 426.96 | 48.38 | +| JAX | 147.61 | 56.25 | + +Tested on 8x80G NVIDIA H800. diff --git a/examples/language/grok-1/inference.py b/examples/language/grok-1/inference.py index ca0ad0d4fe95..58ba3872f856 100644 --- a/examples/language/grok-1/inference.py +++ b/examples/language/grok-1/inference.py @@ -1,26 +1,30 @@ import time import torch -from sentencepiece import SentencePieceProcessor -from transformers import AutoModelForCausalLM -from utils import get_defualt_parser, inference, print_output +from transformers import AutoModelForCausalLM, AutoTokenizer +from utils import get_default_parser, inference, print_output if __name__ == "__main__": - parser = get_defualt_parser() + parser = get_default_parser() args = parser.parse_args() start = time.time() torch.set_default_dtype(torch.bfloat16) + + tokenizer = AutoTokenizer.from_pretrained(args.pretrained, trust_remote_code=True) + model = AutoModelForCausalLM.from_pretrained( args.pretrained, trust_remote_code=True, device_map="auto", torch_dtype=torch.bfloat16, ) - sp = SentencePieceProcessor(model_file=args.tokenizer) + model.eval() + init_time = time.time() - start + for text in args.text: output = inference( model, - sp, + tokenizer, text, max_new_tokens=args.max_new_tokens, do_sample=args.do_sample, @@ -28,5 +32,14 @@ top_k=args.top_k, top_p=args.top_p, ) - print_output(text, sp.decode(output)) - print(f"Overall time: {time.time() - start} seconds.") + print_output(text, tokenizer.decode(output)) + + overall_time = time.time() - start + gen_latency = overall_time - init_time + avg_gen_latency = gen_latency / len(args.text) + print( + f"Initializing time: {init_time:.2f} seconds.\n" + f"Overall time: {overall_time:.2f} seconds. \n" + f"Generation latency: {gen_latency:.2f} seconds. \n" + f"Average generation latency: {avg_gen_latency:.2f} seconds. \n" + ) diff --git a/examples/language/grok-1/inference_tp.py b/examples/language/grok-1/inference_tp.py index 99de60e1f6be..e10c4929cdbf 100644 --- a/examples/language/grok-1/inference_tp.py +++ b/examples/language/grok-1/inference_tp.py @@ -2,9 +2,8 @@ import torch from grok1_policy import Grok1ForCausalLMPolicy -from sentencepiece import SentencePieceProcessor -from transformers import AutoModelForCausalLM -from utils import get_defualt_parser, inference, print_output +from transformers import AutoModelForCausalLM, AutoTokenizer +from utils import get_default_parser, inference, print_output import colossalai from colossalai.booster import Booster @@ -14,7 +13,7 @@ from colossalai.utils import get_current_device if __name__ == "__main__": - parser = get_defualt_parser() + parser = get_default_parser() args = parser.parse_args() start = time.time() colossalai.launch_from_torch({}) @@ -28,16 +27,21 @@ ) booster = Booster(plugin=plugin) torch.set_default_dtype(torch.bfloat16) + + tokenizer = AutoTokenizer.from_pretrained(args.pretrained, trust_remote_code=True) + with LazyInitContext(default_device=get_current_device()): model = AutoModelForCausalLM.from_pretrained( args.pretrained, trust_remote_code=True, torch_dtype=torch.bfloat16 ) model, *_ = booster.boost(model) - sp = SentencePieceProcessor(model_file=args.tokenizer) + model.eval() + init_time = time.time() - start + for text in args.text: output = inference( model.unwrap(), - sp, + tokenizer, text, max_new_tokens=args.max_new_tokens, do_sample=args.do_sample, @@ -46,5 +50,14 @@ top_p=args.top_p, ) if coordinator.is_master(): - print_output(text, sp.decode(output)) - coordinator.print_on_master(f"Overall time: {time.time() - start} seconds.") + print_output(text, tokenizer.decode(output)) + + overall_time = time.time() - start + gen_latency = overall_time - init_time + avg_gen_latency = gen_latency / len(args.text) + coordinator.print_on_master( + f"Initializing time: {init_time:.2f} seconds.\n" + f"Overall time: {overall_time:.2f} seconds. \n" + f"Generation latency: {gen_latency:.2f} seconds. \n" + f"Average generation latency: {avg_gen_latency:.2f} seconds. \n" + ) diff --git a/examples/language/grok-1/requirements.txt b/examples/language/grok-1/requirements.txt index 15d5ea53a15e..2d94924ce3b4 100644 --- a/examples/language/grok-1/requirements.txt +++ b/examples/language/grok-1/requirements.txt @@ -1,4 +1,3 @@ torch>=2.1.0,<2.2.0 colossalai>=0.3.6 -sentencepiece==0.1.99 transformers==4.35.0 diff --git a/examples/language/grok-1/run_inference_fast.sh b/examples/language/grok-1/run_inference_fast.sh index 0dc398c53e33..15a755b97c27 100755 --- a/examples/language/grok-1/run_inference_fast.sh +++ b/examples/language/grok-1/run_inference_fast.sh @@ -1,11 +1,9 @@ #!/usr/bin/env bash -PRETRAINED=${1:-"hpcaitech/grok-1"} -TOKENIZER=${2:-"tokenizer.model"} +PRETRAINED=${1:-"hpcai-tech/grok-1"} torchrun --standalone --nproc_per_node 8 inference_tp.py --pretrained "$PRETRAINED" \ - --tokenizer "$TOKENIZER" \ - --max_new_tokens 64 \ + --max_new_tokens 100 \ --text "The company's annual conference, featuring keynote speakers and exclusive product launches, will be held at the Los Angeles Convention Center from October 20th to October 23rd, 2021. Extract the date mentioned in the above sentence." \ "将以下句子翻译成英语。 我喜欢看电影和读书。" \ "All books have the same weight, 10 books weigh 5kg, what is the weight of 2 books?" diff --git a/examples/language/grok-1/run_inference_slow.sh b/examples/language/grok-1/run_inference_slow.sh index c64dd93b9e62..f5dba44b0db9 100755 --- a/examples/language/grok-1/run_inference_slow.sh +++ b/examples/language/grok-1/run_inference_slow.sh @@ -1,11 +1,9 @@ #!/usr/bin/env bash -PRETRAINED=${1:-"hpcaitech/grok-1"} -TOKENIZER=${2:-"tokenizer.model"} +PRETRAINED=${1:-"hpcai-tech/grok-1"} python3 inference.py --pretrained "$PRETRAINED" \ - --tokenizer "$TOKENIZER" \ - --max_new_tokens 64 \ + --max_new_tokens 100 \ --text "The company's annual conference, featuring keynote speakers and exclusive product launches, will be held at the Los Angeles Convention Center from October 20th to October 23rd, 2021. Extract the date mentioned in the above sentence." \ "将以下句子翻译成英语。 我喜欢看电影和读书。" \ "All books have the same weight, 10 books weigh 5kg, what is the weight of 2 books?" diff --git a/examples/language/grok-1/utils.py b/examples/language/grok-1/utils.py index f113f852eff6..29c86e411db1 100644 --- a/examples/language/grok-1/utils.py +++ b/examples/language/grok-1/utils.py @@ -20,9 +20,9 @@ def print_output(text, output): @torch.no_grad() -def inference(model, sp, text, **generate_kwargs): - input_ids = sp.encode(text) - input_ids = torch.tensor([input_ids]).cuda() +def inference(model, tokenizer, text, **generate_kwargs): + input_ids = tokenizer(text, return_tensors="pt").input_ids + input_ids = input_ids.cuda() attention_mask = torch.ones_like(input_ids) inputs = { "input_ids": input_ids, @@ -33,7 +33,7 @@ def inference(model, sp, text, **generate_kwargs): return outputs[0].tolist() -def get_defualt_parser(): +def get_default_parser(): parser = argparse.ArgumentParser() parser.add_argument("--pretrained", type=str, default="hpcaitech/grok-1") parser.add_argument("--tokenizer", type=str, default="tokenizer.model") diff --git a/examples/language/llama2/finetune.py b/examples/language/llama2/finetune.py index 3dbd0cf357b4..69b4ebe42bf7 100644 --- a/examples/language/llama2/finetune.py +++ b/examples/language/llama2/finetune.py @@ -270,9 +270,7 @@ def main(): ) as pbar: for step in pbar: if use_pipeline: - outputs = booster.execute_pipeline( - dataloader_iter, model, _criterion, optimizer, return_loss=True, return_outputs=True - ) + outputs = booster.execute_pipeline(dataloader_iter, model, _criterion, optimizer, return_loss=True) loss = outputs["loss"] else: batch = next(dataloader_iter) diff --git a/examples/language/llama2/pretrain.py b/examples/language/llama2/pretrain.py index fe7d958307e9..970cd5290f9f 100644 --- a/examples/language/llama2/pretrain.py +++ b/examples/language/llama2/pretrain.py @@ -185,7 +185,7 @@ def main(): microbatch_size=1, enable_jit_fused=False, zero_stage=0, - precision="fp32", + precision=args.mixed_precision, initial_scale=1, ) else: @@ -285,9 +285,7 @@ def main(): ) as pbar: for step in pbar: if use_pipeline: - outputs = booster.execute_pipeline( - dataloader_iter, model, _criterion, optimizer, return_loss=True, return_outputs=True - ) + outputs = booster.execute_pipeline(dataloader_iter, model, _criterion, optimizer, return_loss=True) loss = outputs["loss"] else: batch = next(dataloader_iter) diff --git a/examples/language/openmoe/benchmark/benchmark_cai.py b/examples/language/openmoe/benchmark/benchmark_cai.py index 770c500d86bf..a6d5f8bf2c0e 100644 --- a/examples/language/openmoe/benchmark/benchmark_cai.py +++ b/examples/language/openmoe/benchmark/benchmark_cai.py @@ -270,7 +270,6 @@ def main(): lambda x, y: x.loss, optimizer, return_loss=True, - return_outputs=True, ) # Backward and optimize if is_pp_last_stage: diff --git a/examples/language/openmoe/model/modeling_openmoe.py b/examples/language/openmoe/model/modeling_openmoe.py index eee3b505a22a..fdd8442f506b 100644 --- a/examples/language/openmoe/model/modeling_openmoe.py +++ b/examples/language/openmoe/model/modeling_openmoe.py @@ -96,7 +96,7 @@ def set_openmoe_args( load_balance_beam_width (int, optional): Expert load balance search's beam width. Defaults to 8. load_balance_group_swap_factor (float, optional): Expert load balance group swap factor. Longer value encourages less swap. Defaults to 0.4. enable_kernel (bool, optional): Use kernel optimization. Defaults to False. - enable_comm_overlap (bool, optional): Use communication overlap for MoE. Recommended to enable for muiti-node training. Defaults to False. + enable_comm_overlap (bool, optional): Use communication overlap for MoE. Recommended to enable for multi-node training. Defaults to False. enable_hierarchical_alltoall (bool, optional): Use hierarchical alltoall for MoE. Defaults to False. """ moe_args = dict( diff --git a/examples/language/openmoe/model/openmoe_policy.py b/examples/language/openmoe/model/openmoe_policy.py index 17a51d6398d0..8ef07bdb91b5 100644 --- a/examples/language/openmoe/model/openmoe_policy.py +++ b/examples/language/openmoe/model/openmoe_policy.py @@ -96,8 +96,8 @@ def set_pipeline_forward(self, model_cls: nn.Module, new_forward: Callable, poli else: module = self.model.model - layers_per_stage = self.distribute_layers(len(module.layers), stage_manager.num_stages) - stage_index = Policy.get_stage_index(layers_per_stage, stage_manager.stage) + layers_per_stage = stage_manager.distribute_layers(len(module.layers)) + stage_index = stage_manager.get_stage_index(layers_per_stage) method_replacement = {"forward": partial(new_forward, stage_manager=stage_manager, stage_index=stage_index)} self.append_or_create_method_replacement( description=method_replacement, policy=policy, target_key=model_cls @@ -116,18 +116,17 @@ def get_held_layers(self) -> List[Module]: stage_manager = self.pipeline_stage_manager held_layers = [] - layers_per_stage = self.distribute_layers(len(module.layers), stage_manager.num_stages) + layers_per_stage = stage_manager.distribute_layers(len(module.layers)) if stage_manager.is_first_stage(): held_layers.append(module.embed_tokens) - start_idx, end_idx = self.get_stage_index(layers_per_stage, stage_manager.stage) + start_idx, end_idx = stage_manager.get_stage_index(layers_per_stage) held_layers.extend(module.layers[start_idx:end_idx]) if stage_manager.is_last_stage(): held_layers.append(module.norm) return held_layers - @staticmethod - def distribute_layers(num_layers: int, num_stages: int) -> List[int]: + def distribute_layers(self, num_layers: int, num_stages: int) -> List[int]: """Divide layers into stages""" if num_layers == 24 and num_stages == 4: return [7, 7, 7, 3] @@ -139,7 +138,7 @@ def distribute_layers(num_layers: int, num_stages: int) -> List[int]: return [8, 4] else: print(f"num_layers: {num_layers}, num_stages: {num_stages} not optimized, use origin pp policy") - return Policy.distribute_layers(num_layers, num_stages) + return super().distribute_layers(num_layers, num_stages) class OpenMoeModelPolicy(OpenMoePolicy): diff --git a/examples/language/openmoe/train.py b/examples/language/openmoe/train.py index 89c4d5420994..92f4e066a7a5 100644 --- a/examples/language/openmoe/train.py +++ b/examples/language/openmoe/train.py @@ -190,13 +190,13 @@ def parse_args(): parser.add_argument( "--comm_overlap", action="store_true", - help="Use communication overlap for MoE. Recommended to enable for muiti-node training.", + help="Use communication overlap for MoE. Recommended to enable for multi-node training.", ) # hierarchical all-to-all parser.add_argument( "--hierarchical_alltoall", action="store_true", - help="Use hierarchical all-to-all for MoE. Recommended to enable for muiti-node training.", + help="Use hierarchical all-to-all for MoE. Recommended to enable for multi-node training.", ) args = parser.parse_args() @@ -340,7 +340,6 @@ def main(): lambda x, y: x.loss, optimizer, return_loss=True, - return_outputs=True, ) # Backward and optimize if is_pp_last_stage: @@ -367,7 +366,7 @@ def main(): ): coordinator.print_on_master(f"Apply load balance") apply_load_balance(model, optimizer) - # save ckeckpoint + # save checkpoint if (step + 1) % args.save_interval == 0: coordinator.print_on_master(f"Saving model checkpoint to {args.output_path}") booster.save_model(model, args.output_path, shard=True) diff --git a/examples/language/opt/opt_train_demo.py b/examples/language/opt/opt_train_demo.py index fddbc1b408e7..05336bec42c5 100644 --- a/examples/language/opt/opt_train_demo.py +++ b/examples/language/opt/opt_train_demo.py @@ -41,9 +41,7 @@ def train_epoch(epoch, model, optimizer, _criterion, lr_scheduler, dataloader, b # Forward pass for _ in pbar: if use_pipeline: - outputs = booster.execute_pipeline( - dataloader, model, _criterion, optimizer, return_loss=True, return_outputs=True - ) + outputs = booster.execute_pipeline(dataloader, model, _criterion, optimizer, return_loss=True) # Backward and optimize if is_pp_last_stage: loss = outputs["loss"] diff --git a/examples/tutorial/auto_parallel/auto_parallel_with_resnet.py b/examples/tutorial/auto_parallel/auto_parallel_with_resnet.py index 3c5b786b561a..42b43bbc06f4 100644 --- a/examples/tutorial/auto_parallel/auto_parallel_with_resnet.py +++ b/examples/tutorial/auto_parallel/auto_parallel_with_resnet.py @@ -17,7 +17,7 @@ def synthesize_data(): def main(): - colossalai.launch_from_torch(config="./config.py") + colossalai.legacy.launch_from_torch(config="./config.py") logger = get_dist_logger() diff --git a/examples/tutorial/hybrid_parallel/train.py b/examples/tutorial/hybrid_parallel/train.py index 95f1bf8ee17c..70445dff98af 100644 --- a/examples/tutorial/hybrid_parallel/train.py +++ b/examples/tutorial/hybrid_parallel/train.py @@ -41,9 +41,9 @@ def __len__(self): def main(): # launch from torch - parser = colossalai.get_default_parser() + parser = colossalai.legacy.get_default_parser() args = parser.parse_args() - colossalai.launch_from_torch(config=args.config) + colossalai.legacy.launch_from_torch(config=args.config) # get logger logger = get_dist_logger() diff --git a/examples/tutorial/large_batch_optimizer/train.py b/examples/tutorial/large_batch_optimizer/train.py index dd114b5af86d..c78292e0c072 100644 --- a/examples/tutorial/large_batch_optimizer/train.py +++ b/examples/tutorial/large_batch_optimizer/train.py @@ -37,14 +37,14 @@ def __len__(self): def main(): # initialize distributed setting - parser = colossalai.get_default_parser() + parser = colossalai.legacy.get_default_parser() parser.add_argument( "--optimizer", choices=["lars", "lamb"], help="Choose your large-batch optimizer", required=True ) args = parser.parse_args() # launch from torch - colossalai.launch_from_torch(config=args.config) + colossalai.legacy.launch_from_torch(config=args.config) # get logger logger = get_dist_logger() @@ -73,7 +73,7 @@ def main(): ) # initialize - engine, train_dataloader, test_dataloader, _ = colossalai.initialize( + engine, train_dataloader, test_dataloader, _ = colossalai.legacy.initialize( model=model, optimizer=optimizer, criterion=criterion, diff --git a/examples/tutorial/opt/opt/run_clm.py b/examples/tutorial/opt/opt/run_clm.py old mode 100755 new mode 100644 index 3f0d048795e6..ae8a0f4a044e --- a/examples/tutorial/opt/opt/run_clm.py +++ b/examples/tutorial/opt/opt/run_clm.py @@ -72,7 +72,7 @@ def get_time_stamp(): def parse_args(): - parser = colossalai.get_default_parser() + parser = colossalai.legacy.get_default_parser() parser.add_argument("-s", "--synthetic", action="store_true") parser.add_argument( "--dataset_name", @@ -289,7 +289,7 @@ def __len__(self): def main(): args = parse_args() disable_existing_loggers() - colossalai.launch_from_torch(config=dict()) + colossalai.legacy.launch_from_torch(config=dict()) logger = get_dist_logger() is_main_process = dist.get_rank() == 0 diff --git a/examples/tutorial/sequence_parallel/model/layers/bert_layer.py b/examples/tutorial/sequence_parallel/model/layers/bert_layer.py index 1ef16ee6ad79..c7782bf04fd6 100644 --- a/examples/tutorial/sequence_parallel/model/layers/bert_layer.py +++ b/examples/tutorial/sequence_parallel/model/layers/bert_layer.py @@ -1,9 +1,9 @@ import torch import torch.nn as nn -from colossalai.kernel.cuda_native import LayerNorm from colossalai.kernel.jit import bias_dropout_add_fused_inference, bias_dropout_add_fused_train from colossalai.legacy.nn.layer.parallel_sequence import TransformerSelfAttentionRing +from colossalai.nn.layer.layernorm import MixedFusedLayerNorm as LayerNorm from .dropout import get_bias_dropout_add from .mlp import TransformerMLP diff --git a/examples/tutorial/sequence_parallel/train.py b/examples/tutorial/sequence_parallel/train.py index f25fc818981a..431af35860c4 100644 --- a/examples/tutorial/sequence_parallel/train.py +++ b/examples/tutorial/sequence_parallel/train.py @@ -48,7 +48,7 @@ def pipeline_data_process_func(stage_output, micro_batch_data): def main(): # initialize parse_args() - colossalai.launch_from_torch(config="./config.py", seed=1234, backend="nccl") + colossalai.legacy.launch_from_torch(config="./config.py", seed=1234, backend="nccl") logger = get_dist_logger() @@ -136,7 +136,7 @@ def main(): logger.info(f"LR Scheduler is built with {warmup_steps} warmup steps and {gpc.config.DECAY_ITERS} decay steps") # # init - engine, *dummy = colossalai.initialize(model, optimizer, criterion, verbose=True) + engine, *dummy = colossalai.legacy.initialize(model, optimizer, criterion, verbose=True) # build timer timer = MultiTimer() diff --git a/extensions/README.md b/extensions/README.md index 6f5feb55c2af..b9bde7742be9 100644 --- a/extensions/README.md +++ b/extensions/README.md @@ -101,13 +101,13 @@ class MyExtension(_Extension): self._support_jit = True self.priority = 10 - def is_hardware_available(self) -> bool: + def is_available(self) -> bool: """ Return if the required hardware can be found. """ ... - def assert_hardware_compatible(self) -> None: + def assert_compatible(self) -> None: """ Check if the hardware required by the kernel is compatible. """ diff --git a/extensions/__init__.py b/extensions/__init__.py index 9343cadda194..0dbadba81905 100644 --- a/extensions/__init__.py +++ b/extensions/__init__.py @@ -1,9 +1,5 @@ from .cpu_adam import CpuAdamArmExtension, CpuAdamX86Extension -from .flash_attention import ( - FlashAttentionDaoCudaExtension, - FlashAttentionNpuExtension, - FlashAttentionXformersCudaExtension, -) +from .flash_attention import FlashAttentionDaoCudaExtension, FlashAttentionNpuExtension, FlashAttentionSdpaCudaExtension from .layernorm import LayerNormCudaExtension from .moe import MoeCudaExtension from .optimizer import FusedOptimizerCudaExtension @@ -18,7 +14,7 @@ ScaledMaskedSoftmaxCudaExtension, ScaledUpperTriangleMaskedSoftmaxCudaExtension, FlashAttentionDaoCudaExtension, - FlashAttentionXformersCudaExtension, + FlashAttentionSdpaCudaExtension, FlashAttentionNpuExtension, ] @@ -31,6 +27,6 @@ "ScaledMaskedSoftmaxCudaExtension", "ScaledUpperTriangleMaskedSoftmaxCudaExtension", "FlashAttentionDaoCudaExtension", - "FlashAttentionXformersCudaExtension", + "FlashAttentionSdpaCudaExtension", "FlashAttentionNpuExtension", ] diff --git a/extensions/base_extension.py b/extensions/base_extension.py index c815a7f2ac4a..0c79c0a9e9f5 100644 --- a/extensions/base_extension.py +++ b/extensions/base_extension.py @@ -58,13 +58,13 @@ def get_jit_extension_folder_path(): return cache_directory @abstractmethod - def is_hardware_available(self) -> bool: + def is_available(self) -> bool: """ Check if the hardware required by the kernel is available. """ @abstractmethod - def assert_hardware_compatible(self) -> None: + def assert_compatible(self) -> None: """ Check if the hardware required by the kernel is compatible. """ diff --git a/extensions/cpu_adam/cpu_adam_arm.py b/extensions/cpu_adam/cpu_adam_arm.py index 35bff3b55928..61c4f3ed0697 100644 --- a/extensions/cpu_adam/cpu_adam_arm.py +++ b/extensions/cpu_adam/cpu_adam_arm.py @@ -7,11 +7,11 @@ class CpuAdamArmExtension(_CppExtension): def __init__(self): super().__init__(name="cpu_adam_arm") - def is_hardware_available(self) -> bool: + def is_available(self) -> bool: # only arm allowed return platform.machine() == "aarch64" - def assert_hardware_compatible(self) -> None: + def assert_compatible(self) -> None: arch = platform.machine() assert ( arch == "aarch64" diff --git a/extensions/cpu_adam/cpu_adam_x86.py b/extensions/cpu_adam/cpu_adam_x86.py index a38194167b01..9bbc8d85126d 100644 --- a/extensions/cpu_adam/cpu_adam_x86.py +++ b/extensions/cpu_adam/cpu_adam_x86.py @@ -8,15 +8,15 @@ class CpuAdamX86Extension(_CudaExtension): def __init__(self): super().__init__(name="cpu_adam_x86") - def is_hardware_available(self) -> bool: - return platform.machine() == "x86_64" and super().is_hardware_available() + def is_available(self) -> bool: + return platform.machine() == "x86_64" and super().is_available() - def assert_hardware_compatible(self) -> None: + def assert_compatible(self) -> None: arch = platform.machine() assert ( arch == "x86_64" ), f"[extension] The {self.name} kernel requires the CPU architecture to be x86_64 but got {arch}" - super().assert_hardware_compatible() + super().assert_compatible() # necessary 4 functions def sources_files(self): diff --git a/extensions/cuda_extension.py b/extensions/cuda_extension.py index 842cd9713a99..f1e0095b29b6 100644 --- a/extensions/cuda_extension.py +++ b/extensions/cuda_extension.py @@ -22,7 +22,7 @@ def nvcc_flags(self) -> List[str]: This function should return a list of nvcc compilation flags for extensions. """ - def is_hardware_available(self) -> bool: + def is_available(self) -> bool: # cuda extension can only be built if cuda is available try: import torch @@ -32,7 +32,7 @@ def is_hardware_available(self) -> bool: cuda_available = False return cuda_available - def assert_hardware_compatible(self) -> None: + def assert_compatible(self) -> None: from torch.utils.cpp_extension import CUDA_HOME if not CUDA_HOME: diff --git a/extensions/flash_attention/__init__.py b/extensions/flash_attention/__init__.py index 18abb6191035..ea5b442aa58d 100644 --- a/extensions/flash_attention/__init__.py +++ b/extensions/flash_attention/__init__.py @@ -1,20 +1,14 @@ from .flash_attention_dao_cuda import FlashAttentionDaoCudaExtension from .flash_attention_npu import FlashAttentionNpuExtension -from .flash_attention_xformers_cuda import FlashAttentionXformersCudaExtension +from .flash_attention_sdpa_cuda import FlashAttentionSdpaCudaExtension try: + # TODO: remove this after updating openmoe example import flash_attention # noqa HAS_FLASH_ATTN = True except: HAS_FLASH_ATTN = False -try: - import xformers # noqa - - HAS_MEM_EFF_ATTN = True -except: - HAS_MEM_EFF_ATTN = False - -__all__ = ["FlashAttentionDaoCudaExtension", "FlashAttentionXformersCudaExtension", "FlashAttentionNpuExtension"] +__all__ = ["FlashAttentionDaoCudaExtension", "FlashAttentionSdpaCudaExtension", "FlashAttentionNpuExtension"] diff --git a/extensions/flash_attention/flash_attention_dao_cuda.py b/extensions/flash_attention/flash_attention_dao_cuda.py index 1b7f8ac4736a..a2f2a52f1af4 100644 --- a/extensions/flash_attention/flash_attention_dao_cuda.py +++ b/extensions/flash_attention/flash_attention_dao_cuda.py @@ -5,17 +5,20 @@ class FlashAttentionDaoCudaExtension(_Extension): def __init__(self): super().__init__(name="flash_attention_dao_cuda", support_aot=False, support_jit=False, priority=10) - def is_hardware_available(self) -> bool: + def is_available(self) -> bool: # cuda extension can only be built if cuda is available try: import torch + from flash_attn import flash_attn_func, flash_attn_varlen_kvpacked_func # noqa + from flash_attn.bert_padding import index_first_axis, pad_input # noqa + cuda_available = torch.cuda.is_available() except: cuda_available = False return cuda_available - def assert_hardware_compatible(self) -> bool: + def assert_compatible(self) -> bool: pass def build_aot(self) -> None: @@ -29,65 +32,65 @@ def build_jit(self) -> None: ) def load(self): - try: - from flash_attn.flash_attn_interface import flash_attn_func, flash_attn_varlen_func - except ImportError: - raise ModuleNotFoundError( - ( - "We rely on the third-party flash-attn library for flash attention. Please install flash-attn via 'pip install flash-attn --no-build-isolation'" - ) - ) - from typing import Optional import torch + from einops import rearrange + from flash_attn import flash_attn_func, flash_attn_varlen_kvpacked_func + from flash_attn.bert_padding import index_first_axis, pad_input + + def _unpad_input(hidden_states: torch.Tensor, indices: torch.Tensor): + return index_first_axis(rearrange(hidden_states, "b s ... -> (b s) ..."), indices) def flash_attention( q: torch.Tensor, k: torch.Tensor, v: torch.Tensor, - seq_len_info_q: "SeqLenInfo", - seq_len_info_kv: "SeqLenInfo", - origin_attn_mask: Optional[torch.Tensor] = None, - bias: Optional[torch.Tensor] = None, dropout_p: float = 0.0, - scale: float = None, - causal: bool = False, - padded: bool = False, + scale: Optional[float] = None, + attention_mask: Optional[torch.Tensor] = None, + is_causal: bool = False, + cu_seqlens_q: Optional[torch.Tensor] = None, + cu_seqlens_kv: Optional[torch.Tensor] = None, + max_seqlen_q: Optional[int] = None, + max_seqlen_kv: Optional[int] = None, + q_indices: Optional[torch.Tensor] = None, + kv_indices: Optional[torch.Tensor] = None, ): - """ - Arguments: - q: (batch, q_seqlen, nheads, headdim) - k: (batch, kv_seqlen, nheads, headdim) - v: (batch, kv_seqlen, nheads, headdim) - batch_size: int. - seq_len: int. - dropout_p: float. Dropout probability. - sm_scale: float. The scaling of QK^T before applying softmax. - Default to 1 / sqrt(headdim). - causal: bool. Whether to apply causal attention mask (e.g., for auto-regressive modeling). - Return: - attn_out: (batch, q_seqlen, nheads, headdim). - """ - # check if the input is in allowed dtypes - if padded: - if seq_len_info_kv == None: - seq_len_info_kv = seq_len_info_q - - attn_out = flash_attn_varlen_func( + # [B, N, S, D] -> [B, S, N, D] + q = q.transpose(1, 2) + k = k.transpose(1, 2) + v = v.transpose(1, 2) + b, s_q = q.shape[:2] + if cu_seqlens_q is not None: + # padded / padded causal + # unpad input: [B, S, N, D] -> [T, N, D] + q = _unpad_input(q, q_indices) + kv = _unpad_input(torch.stack(tensors=(k, v), dim=2), kv_indices) + attn_output = flash_attn_varlen_kvpacked_func( + q, + kv, + cu_seqlens_q, + cu_seqlens_kv, + max_seqlen_q, + max_seqlen_kv, + dropout_p=dropout_p, + softmax_scale=scale, + causal=is_causal, + ) + # pad output: [T, N, D] -> [B, S, N, D] + attn_output = pad_input(attn_output, q_indices, b, s_q) + else: + # causal / no attn mask + attn_output = flash_attn_func( q, k, v, - seq_len_info_q.cu_seqlens, - seq_len_info_kv.cu_seqlens, - seq_len_info_q.max_seqlen, - seq_len_info_kv.max_seqlen, - dropout_p, - scale, - causal, + dropout_p=dropout_p, + softmax_scale=scale, + causal=is_causal, ) - else: - attn_out = flash_attn_func(q, k, v, dropout_p=dropout_p, softmax_scale=scale, causal=causal) - return attn_out + # [B, S, N, D] -> [B, N, S, D] + return attn_output.transpose(1, 2) return flash_attention diff --git a/extensions/flash_attention/flash_attention_npu.py b/extensions/flash_attention/flash_attention_npu.py index 58d0f9306e3d..0e01cefa1112 100644 --- a/extensions/flash_attention/flash_attention_npu.py +++ b/extensions/flash_attention/flash_attention_npu.py @@ -5,15 +5,15 @@ class FlashAttentionNpuExtension(_Extension): def __init__(self): super().__init__(name="flash_attention_npu", support_aot=False, support_jit=False) - def is_hardware_available(self) -> bool: + def is_available(self) -> bool: try: - import torch_npu # noqa + import torch_npu - return True + return hasattr(torch_npu, "npu_fusion_attention") except: return False - def assert_hardware_compatible(self) -> bool: + def assert_compatible(self) -> bool: pass def build_aot(self) -> None: @@ -27,47 +27,36 @@ def build_jit(self) -> None: ) def load(self): + from typing import Optional + import torch - from einops import rearrange + import torch_npu - def npu_sdpa_attention( + def flash_attention( q: torch.Tensor, k: torch.Tensor, v: torch.Tensor, - seq_len_info_q=None, - seq_len_info_kv=None, - origin_attn_mask: torch.Tensor = None, dropout_p: float = 0.0, - scale: float = 1.0, - causal=None, - padded=None, + scale: Optional[float] = None, + attention_mask: Optional[torch.Tensor] = None, + is_causal: bool = False, + cu_seqlens_q: Optional[torch.Tensor] = None, + cu_seqlens_kv: Optional[torch.Tensor] = None, + max_seqlen_q: Optional[int] = None, + max_seqlen_kv: Optional[int] = None, + q_indices: Optional[torch.Tensor] = None, + kv_indices: Optional[torch.Tensor] = None, ): - """ - The scaled dot product attention. - - Arguments: - q: (batch, q_seqlen, nheads, headdim) - k: (batch, kv_seqlen, nheads, headdim) - v: (batch, kv_seqlen, nheads, headdim) - batch_size: int. - seq_len: int. - dropout_p: float. Dropout probability. - scale: float. The scaling of QK^T before applying softmax. - Default to 1. - Return: - attn_out: (batch, q_seqlen, nheads, headdim). - """ - q, k, v = [rearrange(x, "b s h d -> b h s d").contiguous() for x in (q, k, v)] - output = torch.nn.functional.scaled_dot_product_attention( + num_heads = q.size(1) + return torch_npu.npu_fusion_attention( q, k, v, - attn_mask=origin_attn_mask, - dropout_p=dropout_p, - is_causal=origin_attn_mask is None, + num_heads, + "BNSD", + atten_mask=attention_mask.bool(), scale=scale, - ) - output = rearrange(output, "b h s d -> b s (h d)") - return output + keep_prob=1 - dropout_p, + )[0] - return npu_sdpa_attention + return flash_attention diff --git a/extensions/flash_attention/flash_attention_sdpa_cuda.py b/extensions/flash_attention/flash_attention_sdpa_cuda.py new file mode 100644 index 000000000000..d3323a6aae27 --- /dev/null +++ b/extensions/flash_attention/flash_attention_sdpa_cuda.py @@ -0,0 +1,56 @@ +from ..base_extension import _Extension + + +class FlashAttentionSdpaCudaExtension(_Extension): + def __init__(self): + super().__init__(name="flash_attention_sdpa_cuda", support_aot=False, support_jit=False) + + def is_available(self) -> bool: + # cuda extension can only be built if cuda is available + try: + import torch + + cuda_available = torch.cuda.is_available() + except: + cuda_available = False + return cuda_available + + def assert_compatible(self) -> bool: + pass + + def build_aot(self) -> None: + raise NotImplementedError("Flash attention SDPA does not require ahead-of-time compilation.") + + def build_jit(self) -> None: + raise NotImplementedError("Flash attention SDPA does not require just-in-time compilation.") + + def load(self): + from typing import Optional + + import torch + + def flash_attention( + q: torch.Tensor, + k: torch.Tensor, + v: torch.Tensor, + dropout_p: float = 0.0, + scale: Optional[float] = None, + attention_mask: Optional[torch.Tensor] = None, + is_causal: bool = False, + cu_seqlens_q: Optional[torch.Tensor] = None, + cu_seqlens_kv: Optional[torch.Tensor] = None, + max_seqlen_q: Optional[int] = None, + max_seqlen_kv: Optional[int] = None, + q_indices: Optional[torch.Tensor] = None, + kv_indices: Optional[torch.Tensor] = None, + ): + return torch.nn.functional.scaled_dot_product_attention( + q, + k, + v, + attn_mask=attention_mask, + dropout_p=dropout_p, + scale=scale, + ) + + return flash_attention diff --git a/extensions/flash_attention/flash_attention_xformers_cuda.py b/extensions/flash_attention/flash_attention_xformers_cuda.py deleted file mode 100644 index 27cd823de14b..000000000000 --- a/extensions/flash_attention/flash_attention_xformers_cuda.py +++ /dev/null @@ -1,94 +0,0 @@ -from ..base_extension import _Extension - - -class FlashAttentionXformersCudaExtension(_Extension): - def __init__(self): - super().__init__(name="flash_attention_xformers_cuda", support_aot=False, support_jit=False) - - def is_hardware_available(self) -> bool: - # cuda extension can only be built if cuda is available - try: - import torch - - cuda_available = torch.cuda.is_available() - except: - cuda_available = False - return cuda_available - - def assert_hardware_compatible(self) -> bool: - pass - - def build_aot(self) -> None: - raise NotImplementedError( - "We rely on the third-party xformers library for flash attention (https://github.com/facebookresearch/xformers). Please install xformers according to the GitHub Readme." - ) - - def build_jit(self) -> None: - raise NotImplementedError( - "We rely on the third-party xformers library for flash attention (https://github.com/facebookresearch/xformers). Please install xformers according to the GitHub Readme." - ) - - def load(self): - try: - from xformers.ops.fmha import MemoryEfficientAttentionCutlassOp, memory_efficient_attention - from xformers.ops.fmha.attn_bias import ( - BlockDiagonalCausalMask, - BlockDiagonalMask, - LowerTriangularMask, - LowerTriangularMaskWithTensorBias, - ) - except ImportError: - raise ModuleNotFoundError( - ( - "We rely on the third-party xformers library for flash attention (https://github.com/facebookresearch/xformers). Please install xformers according to the GitHub Readme." - ) - ) - from typing import Optional - - import torch - - allow_alibi = True - for op in MemoryEfficientAttentionCutlassOp: - allow_alibi = allow_alibi & (LowerTriangularMaskWithTensorBias in op.SUPPORTED_ATTN_BIAS_TYPES) - - def mem_eff_attention( - q: torch.Tensor, - k: torch.Tensor, - v: torch.Tensor, - seq_len_info_q: "SeqLenInfo", - seq_len_info_kv: "SeqLenInfo", - origin_attn_mask: Optional[torch.Tensor] = None, - bias: Optional[torch.Tensor] = None, - dropout_p: float = 0.0, - scale: float = None, - causal: bool = False, - padded: bool = False, - ): - attn_bias = None - if padded: # bert style - if not causal: - attn_bias = BlockDiagonalMask.from_seqlens(seq_len_info_q.seqlens, seq_len_info_kv.seqlens) - else: - attn_bias = BlockDiagonalCausalMask.from_seqlens(seq_len_info_q.seqlens, seq_len_info_kv.seqlens) - elif causal: # gpt style - attn_bias = LowerTriangularMask() - - if bias is not None: # alibi / relative position embedding - assert allow_alibi, "flash attention with bias is not supported in this system." - assert causal, "attention with bias is only supported for causal attention so far." - attn_bias = attn_bias.add_bias(bias) - - if padded: - q = q.unsqueeze(0) - k = k.unsqueeze(0) - v = v.unsqueeze(0) - - out = memory_efficient_attention(q, k, v, attn_bias=attn_bias, p=dropout_p, scale=scale) - - # shape: (b*s, n, d) - if padded: - out = out.squeeze(0) - - return out - - return mem_eff_attention diff --git a/setup.py b/setup.py index ef89481e6b1e..c16709ad1c1c 100644 --- a/setup.py +++ b/setup.py @@ -80,8 +80,8 @@ def get_version() -> str: for ext_cls in ALL_EXTENSIONS: ext = ext_cls() - if ext.support_aot and ext.is_hardware_available(): - ext.assert_hardware_compatible() + if ext.support_aot and ext.is_available(): + ext.assert_compatible() op_names.append(ext.name) ext_modules.append(ext.build_aot()) diff --git a/tests/kit/model_zoo/transformers/gpt.py b/tests/kit/model_zoo/transformers/gpt.py index 24f9627c269c..ab5d97420292 100644 --- a/tests/kit/model_zoo/transformers/gpt.py +++ b/tests/kit/model_zoo/transformers/gpt.py @@ -18,8 +18,23 @@ def data_gen(): # tokenized_input = tokenizer(input, return_tensors='pt') # input_ids = tokenized_input['input_ids'] # attention_mask = tokenized_input['attention_mask'] - input_ids = torch.tensor([[15496, 11, 616, 3290, 318, 13779, 318, 13779]], dtype=torch.int64) - attention_mask = torch.tensor([[1, 1, 1, 1, 1, 1, 1, 1]], dtype=torch.int64) + # input_ids = torch.tensor([[15496, 11, 616, 3290, 318, 13779, 318, 13779]], dtype=torch.int64) + # attention_mask = torch.tensor([[1, 1, 1, 1, 1, 1, 1, 1]], dtype=torch.int64) + input_ids = torch.tensor( + [ + [15496, 11, 616, 3290, 318, 13779, 318, 13779, 15496, 11, 616, 3290, 318, 13779, 318, 13779], + [15496, 11, 616, 3290, 318, 13779, 318, 13779, 15496, 11, 616, 3290, 318, 13779, 318, 13779], + ], + dtype=torch.int64, + ) + attention_mask = torch.tensor( + [ + [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], + [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], + ], + dtype=torch.int64, + ) + return dict(input_ids=input_ids, attention_mask=attention_mask) @@ -35,9 +50,9 @@ def data_gen_for_question_answering(): # question answering data gen # `labels` is the type not the token id for token classification, 0 or 1 data = data_gen() - start_positions = torch.tensor([0], dtype=torch.int64) + start_positions = torch.tensor([[0], [0]], dtype=torch.int64) data["start_positions"] = start_positions - end_positions = torch.tensor([1], dtype=torch.int64) + end_positions = torch.tensor([[1], [1]], dtype=torch.int64) data["end_positions"] = end_positions return data @@ -46,14 +61,20 @@ def data_gen_for_token_classification(): # token classification data gen # `labels` is the type not the token id for token classification, 0 or 1 data = data_gen() - data["labels"] = torch.tensor([[0, 0, 0, 0, 0, 0, 0, 1]], dtype=torch.int64) + data["labels"] = torch.tensor( + [ + [0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 1], + [0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 1], + ], + dtype=torch.int64, + ) return data def data_gen_for_sequence_classification(): # sequence classification data gen data = data_gen() - data["labels"] = torch.tensor([1], dtype=torch.int64) + data["labels"] = torch.tensor([[1], [1]], dtype=torch.int64) return data @@ -61,12 +82,18 @@ def date_gen_for_double_heads(): num_choices = 2 batch_size = 2 input_ids = torch.tensor( - [[15496, 11, 616, 3290, 318, 13779, 318, 13779], [15496, 11, 616, 3290, 318, 13779, 318, 13779]], + [ + [15496, 11, 616, 3290, 318, 13779, 318, 13779, 15496, 11, 616, 3290, 318, 13779, 318, 13779], + [15496, 11, 616, 3290, 318, 13779, 318, 13779, 15496, 11, 616, 3290, 318, 13779, 318, 13779], + ], + dtype=torch.int64, + ) + attention_mask = torch.tensor( + [[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]], dtype=torch.int64, ) - attention_mask = torch.tensor([[1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1]], dtype=torch.int64) - mc_labels = torch.zeros(input_ids.shape[0], dtype=torch.int64) + mc_labels = torch.zeros(input_ids.shape[0], dtype=torch.int64) mc_token_ids = torch.arange(0, num_choices, dtype=torch.int64) mc_token_ids = mc_token_ids.expand((batch_size, num_choices)) multiple_choice_inputs_ids = input_ids.unsqueeze(1).expand(-1, num_choices, -1).contiguous() @@ -103,6 +130,7 @@ def date_gen_for_double_heads(): hidden_dropout=0, problem_type="single_label_classification", pad_token_id=50256, + tie_word_embeddings=True, ) config_for_token_classification = copy.deepcopy(config) diff --git a/tests/kit/model_zoo/transformers/llama.py b/tests/kit/model_zoo/transformers/llama.py index 4730642705ff..58b5b0487a82 100644 --- a/tests/kit/model_zoo/transformers/llama.py +++ b/tests/kit/model_zoo/transformers/llama.py @@ -28,9 +28,19 @@ def data_gen(): # ----------------------------------- input_ids = torch.Tensor( - [[1, 15043, 29892, 590, 11203, 338, 274, 1082], [1, 15043, 29892, 590, 11203, 338, 274, 1082]] + [ + [1, 15043, 29892, 590, 11203, 338, 274, 1082, 1, 15043, 29892, 590, 11203, 338, 274, 1082], + [1, 15043, 29892, 590, 11203, 338, 274, 1082, 1, 15043, 29892, 590, 11203, 338, 274, 1082], + ] ).long() - attention_mask = torch.Tensor([[1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1]]).long() + + attention_mask = torch.Tensor( + [ + [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], + [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], + ] + ).long() + return dict(input_ids=input_ids, attention_mask=attention_mask) # label is needed for casual lm @@ -49,9 +59,9 @@ def data_gen_for_casual_lm(): loss_fn_for_seq_classification = lambda output: output["logits"].mean() config = LlamaConfig( - num_hidden_layers=4, - hidden_size=128, - intermediate_size=256, + num_hidden_layers=8, + hidden_size=32, + intermediate_size=64, num_attention_heads=4, max_position_embeddings=128, num_labels=16, diff --git a/tests/test_booster/test_plugin/test_3d_plugin.py b/tests/test_booster/test_plugin/test_3d_plugin.py index d629e769d715..52cb8c46ed41 100644 --- a/tests/test_booster/test_plugin/test_3d_plugin.py +++ b/tests/test_booster/test_plugin/test_3d_plugin.py @@ -74,7 +74,7 @@ def _criterion(outputs, inputs): loss = criterion(outputs[output_key]) return loss - booster.execute_pipeline(data_iter, model, _criterion, optimizer, return_loss=True, return_outputs=False) + booster.execute_pipeline(data_iter, model, _criterion, optimizer, return_loss=True) optimizer.step() except Exception as e: @@ -83,7 +83,7 @@ def _criterion(outputs, inputs): @parameterize("init_method", ["none", "lazy"]) def check_3d_plugin(init_method: str = "none", early_stop: bool = True): - """check gemini plugin over model zoo + """check hybrid plugin over model zoo Args: early_stop (bool, optional): Whether to stop when getting the first error. Defaults to True. @@ -271,9 +271,9 @@ def run_dist(rank, world_size, port, early_stop: bool = True): @rerun_if_address_is_in_use() -def test_gemini_plugin(early_stop: bool = True): +def test_3d_plugin(early_stop: bool = True): spawn(run_dist, 4, early_stop=early_stop) if __name__ == "__main__": - test_gemini_plugin(early_stop=False) + test_3d_plugin(early_stop=False) diff --git a/tests/test_checkpoint_io/test_gemini_checkpoint_io.py b/tests/test_checkpoint_io/test_gemini_checkpoint_io.py index ece3b40360e8..ac6f8caef816 100644 --- a/tests/test_checkpoint_io/test_gemini_checkpoint_io.py +++ b/tests/test_checkpoint_io/test_gemini_checkpoint_io.py @@ -44,7 +44,10 @@ def exam_state_dict_with_origin(placement_config, model_name, use_safetensors: b (model_fn, data_gen_fn, output_transform_fn, _, _) = next(iter(model_zoo.get_sub_registry(model_name).values())) bert_model = model_fn() - enable_all_optimization = True if tp_size > 1 else False + + enable_flash_attention = True if tp_size > 1 else False + enable_fused_normalization = True if tp_size > 1 else False + enable_jit_fused = True if tp_size > 1 else False with shared_tempdir() as tempdir: pretrained_path = os.path.join(tempdir, "pretrained") @@ -54,7 +57,9 @@ def exam_state_dict_with_origin(placement_config, model_name, use_safetensors: b plugin = GeminiPlugin( **placement_config, tp_size=tp_size, - enable_all_optimization=enable_all_optimization, + enable_flash_attention=enable_flash_attention, + enable_fused_normalization=enable_fused_normalization, + enable_jit_fused=enable_jit_fused, extra_dp_size=extra_dp_size, ) booster = Booster(plugin=plugin) @@ -80,7 +85,9 @@ def exam_state_dict_with_origin(placement_config, model_name, use_safetensors: b def exam_state_dict(placement_config, shard: bool, model_name: str, size_per_shard: int, tp_size: int, zero_size: int): (model_fn, data_gen_fn, output_transform_fn, _, _) = next(iter(model_zoo.get_sub_registry(model_name).values())) criterion = lambda x: x.mean() - enable_all_optimization = True if tp_size > 1 else False + enable_flash_attention = True if tp_size > 1 else False + enable_fused_normalization = True if tp_size > 1 else False + enable_jit_fused = True if tp_size > 1 else False extra_dp_size = dist.get_world_size() // (zero_size * tp_size) plugin = GeminiPlugin( **placement_config, @@ -88,7 +95,9 @@ def exam_state_dict(placement_config, shard: bool, model_name: str, size_per_sha initial_scale=(2**14), tp_size=tp_size, extra_dp_size=extra_dp_size, - enable_all_optimization=enable_all_optimization, + enable_flash_attention=enable_flash_attention, + enable_fused_normalization=enable_fused_normalization, + enable_jit_fused=enable_jit_fused, ) booster = Booster(plugin=plugin) diff --git a/tests/test_checkpoint_io/test_hybrid_parallel_plugin_checkpoint_io.py b/tests/test_checkpoint_io/test_hybrid_parallel_plugin_checkpoint_io.py index b5cb31715aed..d8a625b98a66 100644 --- a/tests/test_checkpoint_io/test_hybrid_parallel_plugin_checkpoint_io.py +++ b/tests/test_checkpoint_io/test_hybrid_parallel_plugin_checkpoint_io.py @@ -74,9 +74,7 @@ def _preprocess_data(data): data = data_gen_fn() model.train() if booster.plugin.stage_manager is not None: - booster.execute_pipeline( - _preprocess_data(data), model, _criterion, optimizer, return_loss=True, return_outputs=False - ) + booster.execute_pipeline(_preprocess_data(data), model, _criterion, optimizer, return_loss=True) else: output = model(**_preprocess_data(data)) loss = criterion(output) @@ -108,16 +106,13 @@ def _preprocess_data(data): data_for_shard = data_gen_fn() data_for_origin = data_gen_fn() if booster.plugin.stage_manager is not None: - booster.execute_pipeline( - _preprocess_data(data_for_shard), model, _criterion, optimizer, return_loss=True, return_outputs=False - ) + booster.execute_pipeline(_preprocess_data(data_for_shard), model, _criterion, optimizer, return_loss=True) booster.execute_pipeline( _preprocess_data(data_for_origin), new_model, _criterion, new_optimizer, return_loss=True, - return_outputs=False, ) else: old_model_loss = criterion(model(**_preprocess_data(data_for_shard))) diff --git a/tests/test_cluster/test_process_group_mesh.py b/tests/test_cluster/test_process_group_mesh.py index 08542d1f64fa..3d206622d644 100644 --- a/tests/test_cluster/test_process_group_mesh.py +++ b/tests/test_cluster/test_process_group_mesh.py @@ -84,6 +84,30 @@ def check_process_group_mesh_with_cases(): 2: [2], 3: [3], } + TPxPP_RANKS_IN_GROUP = { + 0: [0, 1, 2, 3], + 1: [0, 1, 2, 3], + 2: [0, 1, 2, 3], + 3: [0, 1, 2, 3], + } + DPxTP_RANKS_IN_GROUP = { + 0: [0, 1], + 1: [0, 1], + 2: [2, 3], + 3: [2, 3], + } + TPxPP_PARTIAL_INDICES = { + 0: [[0, 1], [0]], + 1: [[1], [0, 1]], + 2: [[0], [0, 1]], + 3: [[0, 1], [1]], + } + TPxPP_RANKS_IN_GROUP_PARTIAL = { + 0: [0, 1], + 1: [1, 3], + 2: [0, 2], + 3: [2, 3], + } pg_mesh = ProcessGroupMesh(DP_SIZE, PP_SIZE, TP_SIZE) @@ -107,6 +131,12 @@ def check_process_group_mesh_with_cases(): assert pg_mesh.get_ranks_in_group(pp_group) == PP_RANKS_IN_GROUP[rank] dp_group = pg_mesh.get_group_along_axis(DP_DIM) assert pg_mesh.get_ranks_in_group(dp_group) == DP_RANKS_IN_GROUP[rank] + dpxtp_group = pg_mesh.create_group_along_axis([DP_DIM, TP_DIM]) + assert pg_mesh.get_ranks_in_group(dpxtp_group) == DPxTP_RANKS_IN_GROUP[rank] + tpxpp_group = pg_mesh.create_group_along_axis([TP_DIM, PP_DIM]) + assert pg_mesh.get_ranks_in_group(tpxpp_group) == TPxPP_RANKS_IN_GROUP[rank] + tpxpp_group_partial = pg_mesh.create_group_along_axis([TP_DIM, PP_DIM], TPxPP_PARTIAL_INDICES[rank]) + assert pg_mesh.get_ranks_in_group(tpxpp_group_partial) == TPxPP_RANKS_IN_GROUP_PARTIAL[rank] # check prev rank if RANK_TO_COORDINATE[rank][TP_DIM] != 0: diff --git a/tests/test_moe/test_moe_checkpoint.py b/tests/test_moe/test_moe_checkpoint.py index d6dad2d7fb41..10e63592ac07 100644 --- a/tests/test_moe/test_moe_checkpoint.py +++ b/tests/test_moe/test_moe_checkpoint.py @@ -49,7 +49,6 @@ def run_fwd_bwd( lambda x, y: x.loss, optimizer, return_loss=True, - return_outputs=True, ) # Backward and optimize if is_pp_last_stage: diff --git a/tests/test_optimizer/test_nvme.py b/tests/test_optimizer/test_nvme.py index 4ff16bb9b7c9..603b7b6fa325 100644 --- a/tests/test_optimizer/test_nvme.py +++ b/tests/test_optimizer/test_nvme.py @@ -1,3 +1,4 @@ +import pytest import torch from colossalai.nn.optimizer import CPUAdam, HybridAdam @@ -17,6 +18,8 @@ def check_params_equal(model, torch_model): assert torch.allclose(p, torch_p, atol=1e-3), f"diff: {torch.abs(p - torch_p)}" +# TODO Something wrong with ci when running this test. +@pytest.mark.skip(reason="skip because of something wrong with CI") @clear_cache_before_run() @parameterize("nvme_offload_fraction", [0.0, 0.5, 1.0]) @parameterize("nvme_offload_dir", ["./offload", None]) diff --git a/tests/test_pipeline/test_pipeline_utils/test_t5_pipeline_utils.py b/tests/test_pipeline/test_pipeline_utils/test_t5_pipeline_utils.py index 3723c9c1014a..1b7b0073f62e 100644 --- a/tests/test_pipeline/test_pipeline_utils/test_t5_pipeline_utils.py +++ b/tests/test_pipeline/test_pipeline_utils/test_t5_pipeline_utils.py @@ -1,4 +1,23 @@ +import random + +from colossalai.pipeline.stage_manager import PipelineStageManager from colossalai.shardformer.policies.t5 import T5BasePolicy +from colossalai.shardformer.shard.shard_config import ShardConfig + + +class _ShardConfig(ShardConfig): + def __post_init__(self): + pass + + +class _PipelineStageManager(PipelineStageManager): + def __init__(self): + self.is_interleave = False + self.num_layers_per_stage = None + + @property + def num_stages(self): + return random.randint(5, 10) def test_t5_pipeline_distribution(): @@ -10,9 +29,15 @@ def test_t5_pipeline_distribution(): "decoder_starting_stage": [1, 1, 2, 2, 3, 1, 5, 2], } + stage_manager = _PipelineStageManager() + shard_config = _ShardConfig(pipeline_stage_manager=stage_manager) + policy = T5BasePolicy() + policy.set_shard_config(shard_config) for i in range(num_test_cases): - _, decoder_starting_stage = T5BasePolicy.distribute_t5_layers( - test_dict["num_encoder_layers"][i], test_dict["num_decoder_layers"][i], test_dict["num_stages"][i] + _, decoder_starting_stage = policy.distribute_t5_layers( + test_dict["num_encoder_layers"][i], + test_dict["num_decoder_layers"][i], + test_dict["num_stages"][i], ) assert test_dict["decoder_starting_stage"][i] == decoder_starting_stage @@ -32,14 +57,18 @@ def test_t5_pipeline_layers(): } for i in range(num_test_cases): - layers_per_stage, decoder_starting_stage = T5BasePolicy.distribute_t5_layers( - test_dict["num_encoder_layers"][i], test_dict["num_decoder_layers"][i], test_dict["num_stages"][i] + stage_manager = _PipelineStageManager() + shard_config = _ShardConfig(pipeline_stage_manager=stage_manager) + policy = T5BasePolicy() + policy.set_shard_config(shard_config) + layers_per_stage, decoder_starting_stage = policy.distribute_t5_layers( + test_dict["num_encoder_layers"][i], + test_dict["num_decoder_layers"][i], + test_dict["num_stages"][i], ) for stage in range(test_dict["num_stages"][i]): start_idx, end_idx = test_dict["layers_per_stage"][i][stage] - predicted_start, predicted_end = T5BasePolicy.get_t5_stage_index( - layers_per_stage, stage, decoder_starting_stage - ) + predicted_start, predicted_end = policy.get_t5_stage_index(layers_per_stage, stage, decoder_starting_stage) assert start_idx == predicted_start assert end_idx == predicted_end diff --git a/tests/test_pipeline/test_pipeline_utils/test_whisper_pipeline_utils.py b/tests/test_pipeline/test_pipeline_utils/test_whisper_pipeline_utils.py index f6be8f6feac2..9f8c1ad32d23 100644 --- a/tests/test_pipeline/test_pipeline_utils/test_whisper_pipeline_utils.py +++ b/tests/test_pipeline/test_pipeline_utils/test_whisper_pipeline_utils.py @@ -1,4 +1,23 @@ +import random + +from colossalai.pipeline.stage_manager import PipelineStageManager from colossalai.shardformer.policies.whisper import WhisperPolicy +from colossalai.shardformer.shard.shard_config import ShardConfig + + +class _ShardConfig(ShardConfig): + def __post_init__(self): + pass + + +class _PipelineStageManager(PipelineStageManager): + def __init__(self): + self.is_interleave = False + self.num_layers_per_stage = None + + @property + def num_stages(self): + return random.randint(5, 10) def test_whisper_pipeline_distribution(): @@ -10,9 +29,15 @@ def test_whisper_pipeline_distribution(): "decoder_starting_stage": [1, 1, 2, 2, 3, 1, 5, 2], } + stage_manager = _PipelineStageManager() + shard_config = _ShardConfig(pipeline_stage_manager=stage_manager) + policy = WhisperPolicy() + policy.set_shard_config(shard_config) for i in range(num_test_cases): - _, decoder_starting_stage = WhisperPolicy.distribute_whisper_layers( - test_dict["num_encoder_layers"][i], test_dict["num_decoder_layers"][i], test_dict["num_stages"][i] + _, decoder_starting_stage = policy.distribute_whisper_layers( + test_dict["num_encoder_layers"][i], + test_dict["num_decoder_layers"][i], + test_dict["num_stages"][i], ) assert test_dict["decoder_starting_stage"][i] == decoder_starting_stage @@ -31,14 +56,20 @@ def test_whisper_pipeline_layers(): ], } + stage_manager = _PipelineStageManager() + shard_config = _ShardConfig(pipeline_stage_manager=stage_manager) + policy = WhisperPolicy() + policy.set_shard_config(shard_config) for i in range(num_test_cases): - layers_per_stage, decoder_starting_stage = WhisperPolicy.distribute_whisper_layers( - test_dict["num_encoder_layers"][i], test_dict["num_decoder_layers"][i], test_dict["num_stages"][i] + layers_per_stage, decoder_starting_stage = policy.distribute_whisper_layers( + test_dict["num_encoder_layers"][i], + test_dict["num_decoder_layers"][i], + test_dict["num_stages"][i], ) for stage in range(test_dict["num_stages"][i]): start_idx, end_idx = test_dict["layers_per_stage"][i][stage] - predicted_start, predicted_end = WhisperPolicy.get_whisper_stage_index( + predicted_start, predicted_end = policy.get_whisper_stage_index( layers_per_stage, stage, decoder_starting_stage ) assert start_idx == predicted_start diff --git a/tests/test_pipeline/test_schedule/test_interleaved.py b/tests/test_pipeline/test_schedule/test_interleaved.py index 0e81818eb239..f8820688e610 100644 --- a/tests/test_pipeline/test_schedule/test_interleaved.py +++ b/tests/test_pipeline/test_schedule/test_interleaved.py @@ -103,9 +103,7 @@ def criterion(x, *args, **kwargs): torch_loss = criterion(torch_output) torch_loss.backward() - pp_ret = schedule.forward_backward_step( - sharded_model, iter(input_list), criterion, pp_optimizer, return_loss=True, return_outputs=True - ) + pp_ret = schedule.forward_backward_step(sharded_model, iter(input_list), criterion, pp_optimizer, return_loss=True) # check loss if stage_manager.is_last_stage(ignore_chunk=True): @@ -134,7 +132,7 @@ def criterion(x, *args, **kwargs): torch_loss = criterion(torch_output) pp_ret = schedule.forward_backward_step( - sharded_model, iter(input_list), criterion, pp_optimizer, return_loss=True, return_outputs=True + sharded_model, iter(input_list), criterion, pp_optimizer, return_loss=True ) if stage_manager.is_last_stage(ignore_chunk=True): assert torch.allclose(torch_loss, pp_ret["loss"]) diff --git a/tests/test_pipeline/test_schedule/test_oneF_oneB.py b/tests/test_pipeline/test_schedule/test_oneF_oneB.py index a08dc6d277d0..590800780ab4 100644 --- a/tests/test_pipeline/test_schedule/test_oneF_oneB.py +++ b/tests/test_pipeline/test_schedule/test_oneF_oneB.py @@ -99,9 +99,7 @@ def custom_fwd(self, x): torch_output = torch_model(input_list[0]) torch_loss = criterion(torch_output) torch_loss.backward() - pp_ret = schedule.forward_backward_step( - sharded_model, iter(input_list), criterion, pp_optimizer, return_loss=True, return_outputs=True - ) + pp_ret = schedule.forward_backward_step(sharded_model, iter(input_list), criterion, pp_optimizer, return_loss=True) # check loss if stage_manager.is_last_stage(): @@ -130,7 +128,7 @@ def custom_fwd(self, x): torch_loss = criterion(torch_output) pp_ret = schedule.forward_backward_step( - sharded_model, iter(input_list), criterion, pp_optimizer, return_loss=True, return_outputs=True + sharded_model, iter(input_list), criterion, pp_optimizer, return_loss=True ) if stage_manager.is_last_stage(): assert torch.allclose(torch_loss, pp_ret["loss"]) diff --git a/tests/test_shardformer/test_flash_attention.py b/tests/test_shardformer/test_flash_attention.py new file mode 100644 index 000000000000..f9eab132f6f6 --- /dev/null +++ b/tests/test_shardformer/test_flash_attention.py @@ -0,0 +1,147 @@ +import math +from copy import copy + +import torch +from torch.testing import assert_close + +from colossalai.kernel.kernel_loader import ( + FlashAttentionLoader, + FlashAttentionWithCustomMaskLoader, + FlashAttentionWithPaddingMaskLoader, +) +from colossalai.shardformer.layer import AttnMaskType, ColoAttention +from colossalai.shardformer.layer.attn import invert_mask +from colossalai.testing import clear_cache_before_run, parameterize +from colossalai.utils import get_current_device, set_seed + +DTYPE = [torch.float16, torch.bfloat16] +B, N, S, D = 2, 8, 256, 32 + +TOL_MAP = { + torch.float16: {"atol": 5e-4, "rtol": 2e-3}, + torch.bfloat16: {}, +} + + +def attention_ref(q: torch.Tensor, k: torch.Tensor, v: torch.Tensor, attn_mask=None, dropout_p=0.0): + head_dim = q.size(-1) + attn_weights = torch.matmul(q, k.transpose(2, 3)) / math.sqrt(head_dim) + if attn_mask is not None: + attn_weights = attn_weights + attn_mask + attn_weights = torch.softmax(attn_weights, dim=-1, dtype=torch.float).to(q.dtype) + attn_weights = torch.dropout(attn_weights, p=dropout_p, train=True) + attn_output = torch.matmul(attn_weights, v) + return attn_output + + +def gen_padded_kwargs(dtype: torch.dtype): + padding_mask = torch.ones((B, S), dtype=torch.int, device=get_current_device()) + padding_mask[0, : S // 4] = 0 + return ( + ColoAttention.prepare_attn_kwargs((B, 1, S, S), dtype, padding_mask.device, q_padding_mask=padding_mask), + padding_mask, + ) + + +def gen_padded_causal_kwargs(dtype: torch.dtype): + padding_mask = torch.ones((B, S), dtype=torch.int, device=get_current_device()) + padding_mask[0, S // 2 :] = 0 + return ( + ColoAttention.prepare_attn_kwargs( + (B, 1, S, S), dtype, padding_mask.device, q_padding_mask=padding_mask, is_causal=True + ), + padding_mask, + ) + + +def gen_causal_kwargs(dtype: torch.dtype): + return ColoAttention.prepare_attn_kwargs((B, 1, S, S), dtype, get_current_device(), is_causal=True), None + + +def gen_custom_kwargs(dtype: torch.dtype): + attn_mask = torch.ones((B, S, S), dtype=dtype, device=get_current_device()) + attn_mask[0, : S // 2, S // 2 :] = 0 + attn_mask[0, S // 2 :, : S // 2] = 0 + attn_mask[1, :, S // 4 :] = 0 + attn_mask = invert_mask(attn_mask).unsqueeze(1) + assert not torch.all(attn_mask != 0, dim=-1).any() + return {"attention_mask": attn_mask}, None + + +def post_process_kwargs_for_raw_attn(attn_kwargs: dict): + if "attention_mask_type" in attn_kwargs: + attn_kwargs = copy(attn_kwargs) + mask_type = attn_kwargs.pop("attention_mask_type") + attn_kwargs["is_causal"] = mask_type in (AttnMaskType.CAUSAL, AttnMaskType.PADDED_CAUSAL) + return attn_kwargs + + +def check_attn_func(dtype: torch.dtype, attn_func, attn_kwargs: dict, padding_mask=None): + tols = TOL_MAP[dtype] + q = torch.rand((B, N, S, D), dtype=dtype, device=get_current_device(), requires_grad=True) + k = torch.rand((B, N, S, D), dtype=dtype, device=get_current_device(), requires_grad=True) + v = torch.rand((B, N, S, D), dtype=dtype, device=get_current_device(), requires_grad=True) + q_flash = q.clone().detach().requires_grad_(True) + k_flash = k.clone().detach().requires_grad_(True) + v_flash = v.clone().detach().requires_grad_(True) + attn_mask = attn_kwargs.get("attention_mask", None) + ref_output = attention_ref(q, k, v, attn_mask) + output = attn_func(q_flash, k_flash, v_flash, **attn_kwargs) + if padding_mask is not None: + # [B, Sq] -> [B, 1, Sq, 1] + padding_mask = padding_mask[:, None, :, None].logical_not() + ref_output = ref_output.masked_fill(padding_mask, 0) + output = output.masked_fill(padding_mask, 0) + assert_close(output, ref_output, **tols) + output.mean().backward() + ref_output.mean().backward() + assert_close(q.grad, q_flash.grad, **tols) + assert_close(k.grad, k_flash.grad, **tols) + assert_close(v.grad, v_flash.grad, **tols) + + +@clear_cache_before_run() +@parameterize("dtype", DTYPE) +def test_flash_attn_func(dtype: torch.dtype): + torch.backends.cudnn.deterministic = True + set_seed(0) + # (func, name, need_postprocess) + avail_attn_funcs = [(ColoAttention.attention, "coloattn", False)] + avail_custom_mask_attn_funcs = [(ColoAttention.attention, "coloattn", False)] + avail_padding_mask_attn_funcs = [(ColoAttention.attention, "coloattn", False)] + for ext_cls in FlashAttentionLoader.REGISTRY: + ext = ext_cls() + if ext.is_available(): + ext.assert_compatible() + avail_attn_funcs.append((ext.load(), ext.name, True)) + for ext_cls in FlashAttentionWithCustomMaskLoader.REGISTRY: + ext = ext_cls() + if ext.is_available(): + ext.assert_compatible() + avail_custom_mask_attn_funcs.append((ext.load(), ext.name, True)) + for ext_cls in FlashAttentionWithPaddingMaskLoader.REGISTRY: + ext = ext_cls() + if ext.is_available(): + ext.assert_compatible() + avail_padding_mask_attn_funcs.append((ext.load(), ext.name, True)) + + test_sets = { + "none": (lambda dtype: ({}, None), avail_attn_funcs), + "padded": (gen_padded_kwargs, avail_padding_mask_attn_funcs), + "padded_causal": (gen_padded_causal_kwargs, avail_padding_mask_attn_funcs), + "causal": (gen_causal_kwargs, avail_attn_funcs), + "custom": (gen_custom_kwargs, avail_custom_mask_attn_funcs), + } + + for mask_type, (gen_kwargs_func, attn_funcs) in test_sets.items(): + attn_kwargs, padding_mask = gen_kwargs_func(dtype) + for attn_func, name, need_postprocess in attn_funcs: + print(f"{dtype}, {name}, {mask_type}") + if need_postprocess: + check_attn_func(dtype, attn_func, post_process_kwargs_for_raw_attn(attn_kwargs), padding_mask) + else: + check_attn_func(dtype, attn_func, attn_kwargs, padding_mask) + + +if __name__ == "__main__": + test_flash_attn_func() diff --git a/tests/test_shardformer/test_layer/test_gpt2_qkv_fused_linear_1d.py b/tests/test_shardformer/test_layer/test_gpt2_qkv_fused_linear_1d.py index e056860ede57..e9aa0dbedbc8 100644 --- a/tests/test_shardformer/test_layer/test_gpt2_qkv_fused_linear_1d.py +++ b/tests/test_shardformer/test_layer/test_gpt2_qkv_fused_linear_1d.py @@ -56,13 +56,18 @@ def rearrange(tensor: torch.Tensor, dim: int): return rearanged_tensor -def check_linear_conv_1d_col(lazy_init: bool, seq_parallel: bool, overlap: bool): +def check_linear_conv_1d_col(lazy_init: bool, seq_parallel_mode: str, overlap: bool): ctx = LazyInitContext() if lazy_init else nullcontext() linear = Conv1D(192, 48).cuda() with ctx: linear_copy = Conv1D(192, 48).cuda() linear_conv_col = GPT2FusedLinearConv1D_Col.from_native_module( - linear_copy, process_group=None, gather_output=True, seq_parallel=seq_parallel, n_fused=3, overlap=overlap + linear_copy, + process_group=None, + gather_output=True, + seq_parallel_mode=seq_parallel_mode, + n_fused=3, + overlap=overlap, ) assert linear.weight.shape == torch.Size([48, 192]) @@ -79,7 +84,9 @@ def check_linear_conv_1d_col(lazy_init: bool, seq_parallel: bool, overlap: bool) # check computation correctness x = torch.rand(1, 4, 48).cuda() out = linear(x) - x_for_shard = x.expand_as(x.clone()) if seq_parallel is False else torch.chunk(x.clone(), 2, dim=1)[dist.get_rank()] + x_for_shard = ( + x.expand_as(x.clone()) if seq_parallel_mode is None else torch.chunk(x.clone(), 2, dim=1)[dist.get_rank()] + ) gather_out = linear_conv_col(x_for_shard) assert_close(rearrange(out, -1), gather_out) @@ -91,14 +98,14 @@ def check_linear_conv_1d_col(lazy_init: bool, seq_parallel: bool, overlap: bool) assert_close(target_grad, linear_conv_col.weight.grad) -def check_linear_conv_1d_row(lazy_init: bool, seq_parallel: bool): +def check_linear_conv_1d_row(lazy_init: bool, seq_parallel_mode: bool): ctx = LazyInitContext() if lazy_init else nullcontext() linear = Conv1D(192, 48).cuda() with ctx: linear_copy = Conv1D(192, 48).cuda() linear_row = GPT2FusedLinearConv1D_Row.from_native_module( - linear_copy, process_group=None, parallel_input=False, seq_parallel=seq_parallel + linear_copy, process_group=None, parallel_input=False, seq_parallel_mode=seq_parallel_mode ) assert linear.weight.shape == torch.Size([48, 192]) @@ -115,7 +122,7 @@ def check_linear_conv_1d_row(lazy_init: bool, seq_parallel: bool): x = torch.rand(1, 4, 48).cuda() out = linear(x) gather_out = linear_row(x) - target_out = out if seq_parallel is False else torch.chunk(out.clone(), 2, dim=1)[dist.get_rank()] + target_out = out if seq_parallel_mode is None else torch.chunk(out.clone(), 2, dim=1)[dist.get_rank()] assert_close(target_out, gather_out) # check backward correctness @@ -128,11 +135,11 @@ def check_linear_conv_1d_row(lazy_init: bool, seq_parallel: bool): @parameterize("lazy_init", [False, True]) -@parameterize("seq_parallel", [False, True]) +@parameterize("seq_parallel_mode", ["split_gather", None]) @parameterize("overlap", [True]) -def check_gpt2_qkv_fused_linear_1d(lazy_init: bool, seq_parallel: bool, overlap: bool): - check_linear_conv_1d_col(lazy_init, seq_parallel, overlap) - check_linear_conv_1d_row(lazy_init, seq_parallel) +def check_gpt2_qkv_fused_linear_1d(lazy_init: bool, seq_parallel_mode: bool, overlap: bool): + check_linear_conv_1d_col(lazy_init, seq_parallel_mode, overlap) + check_linear_conv_1d_row(lazy_init, seq_parallel_mode) def run_dist(rank, world_size, port): diff --git a/tests/test_shardformer/test_layer/test_linear_1d.py b/tests/test_shardformer/test_layer/test_linear_1d.py index defa4afb919b..21d3190de7ae 100644 --- a/tests/test_shardformer/test_layer/test_linear_1d.py +++ b/tests/test_shardformer/test_layer/test_linear_1d.py @@ -15,13 +15,13 @@ os.environ["CUDA_DEVICE_MAX_CONNECTIONS"] = "1" -def check_linear_1d_col(lazy_init: bool, seq_parallel: bool, overlap: bool): +def check_linear_1d_col(lazy_init: bool, seq_parallel_mode: bool, overlap: bool): ctx = LazyInitContext() if lazy_init else nullcontext() linear = nn.Linear(32, 128).cuda() with ctx: linear_copy = nn.Linear(32, 128).cuda() linear_col = Linear1D_Col.from_native_module( - linear_copy, process_group=None, gather_output=True, seq_parallel=seq_parallel, overlap=overlap + linear_copy, process_group=None, gather_output=True, seq_parallel_mode=seq_parallel_mode, overlap=overlap ) # ensure that the parameters are distributed @@ -43,7 +43,9 @@ def check_linear_1d_col(lazy_init: bool, seq_parallel: bool, overlap: bool): x = torch.rand(2, 4, 32).cuda() x_for_unshard = x.expand_as(x.clone()) x_for_unshard.requires_grad_(True) - x_for_shard = x.expand_as(x.clone()) if seq_parallel is False else torch.chunk(x.clone(), 2, dim=1)[dist.get_rank()] + x_for_shard = ( + x.expand_as(x.clone()) if seq_parallel_mode is None else torch.chunk(x.clone(), 2, dim=1)[dist.get_rank()] + ) x_for_shard.requires_grad_(True) out = linear(x_for_unshard) @@ -63,20 +65,20 @@ def check_linear_1d_col(lazy_init: bool, seq_parallel: bool, overlap: bool): assert x_for_unshard.grad is not None target_unshard_gard = ( x_for_unshard.grad - if seq_parallel is False + if seq_parallel_mode is None else torch.chunk(x_for_unshard.grad.clone(), 2, dim=1)[dist.get_rank()] ) assert_close(target_unshard_gard, x_for_shard.grad) -def check_linear_1d_row(lazy_init: bool, seq_parallel: bool): +def check_linear_1d_row(lazy_init: bool, seq_parallel_mode: bool): ctx = LazyInitContext() if lazy_init else nullcontext() linear = nn.Linear(32, 128).cuda() with ctx: linear_copy = nn.Linear(32, 128).cuda() linear_row = Linear1D_Row.from_native_module( - linear_copy, process_group=None, parallel_input=False, seq_parallel=seq_parallel + linear_copy, process_group=None, parallel_input=False, seq_parallel_mode=seq_parallel_mode ) assert linear_row.weight.shape == torch.Size([128, 16]) @@ -98,7 +100,7 @@ def check_linear_1d_row(lazy_init: bool, seq_parallel: bool): # run forward out = linear(x_for_unshard) gather_out = linear_row(x_for_shard) - target_out = out if seq_parallel is False else torch.chunk(out.clone(), 2, dim=1)[dist.get_rank()] + target_out = out if seq_parallel_mode is None else torch.chunk(out.clone(), 2, dim=1)[dist.get_rank()] assert_close(target_out, gather_out) # check backward correctness @@ -115,7 +117,7 @@ def check_linear_1d_row(lazy_init: bool, seq_parallel: bool): assert_close(x_for_unshard.grad, x_for_shard.grad) -def check_linear_col_plus_row(lazy_init: bool, seq_parallel: bool, overlap: bool): +def check_linear_col_plus_row(lazy_init: bool, seq_parallel_mode: bool, overlap: bool): ctx = LazyInitContext() if lazy_init else nullcontext() linear_1 = nn.Linear(32, 128).cuda() @@ -125,10 +127,10 @@ def check_linear_col_plus_row(lazy_init: bool, seq_parallel: bool, overlap: bool linear_1_copy = nn.Linear(32, 128).cuda() linear_2_copy = nn.Linear(128, 32).cuda() linear_col = Linear1D_Col.from_native_module( - linear_1_copy, process_group=None, gather_output=False, seq_parallel=seq_parallel, overlap=overlap + linear_1_copy, process_group=None, gather_output=False, seq_parallel_mode=seq_parallel_mode, overlap=overlap ) linear_row = Linear1D_Row.from_native_module( - linear_2_copy, process_group=None, parallel_input=True, seq_parallel=seq_parallel + linear_2_copy, process_group=None, parallel_input=True, seq_parallel_mode=seq_parallel_mode ) linear_1.load_state_dict(linear_col.state_dict()) @@ -141,13 +143,17 @@ def check_linear_col_plus_row(lazy_init: bool, seq_parallel: bool, overlap: bool x = torch.rand(2, 4, 32).cuda() x_for_unshard = x.expand_as(x.clone()) x_for_unshard.requires_grad_(True) - x_for_shard = x.expand_as(x.clone()) if seq_parallel is False else torch.chunk(x.clone(), 2, dim=1)[dist.get_rank()] + x_for_shard = ( + x.expand_as(x.clone()) if seq_parallel_mode is None else torch.chunk(x.clone(), 2, dim=1)[dist.get_rank()] + ) x_for_shard.requires_grad_(True) # run forward unshard_out = linear_2(linear_1(x_for_unshard)) shard_out = linear_row(linear_col(x_for_shard)) - target_out = unshard_out if seq_parallel is False else torch.chunk(unshard_out.clone(), 2, dim=1)[dist.get_rank()] + target_out = ( + unshard_out if seq_parallel_mode is None else torch.chunk(unshard_out.clone(), 2, dim=1)[dist.get_rank()] + ) assert_close(target_out, shard_out) # check backward correctness @@ -163,19 +169,19 @@ def check_linear_col_plus_row(lazy_init: bool, seq_parallel: bool, overlap: bool assert x_for_unshard.grad is not None target_unshard_gard = ( x_for_unshard.grad - if seq_parallel is False + if seq_parallel_mode is None else torch.chunk(x_for_unshard.grad.clone(), 2, dim=1)[dist.get_rank()] ) assert_close(target_unshard_gard, x_for_shard.grad) @parameterize("lazy_init", [False, True]) -@parameterize("seq_parallel", [False, True]) +@parameterize("seq_parallel_mode", [None, "split_gather"]) @parameterize("overlap", [True]) -def run_dist_linear_test(lazy_init, seq_parallel, overlap): - check_linear_1d_col(lazy_init, seq_parallel, overlap) - check_linear_1d_row(lazy_init, seq_parallel) - check_linear_col_plus_row(lazy_init, seq_parallel, overlap) +def run_dist_linear_test(lazy_init, seq_parallel_mode, overlap): + check_linear_1d_col(lazy_init, seq_parallel_mode, overlap) + check_linear_1d_row(lazy_init, seq_parallel_mode) + check_linear_col_plus_row(lazy_init, seq_parallel_mode, overlap) def check_dist_linear(rank, world_size, port): diff --git a/tests/test_shardformer/test_layer/test_sequence_parallel.py b/tests/test_shardformer/test_layer/test_sequence_parallel.py new file mode 100644 index 000000000000..13b1a13e7f94 --- /dev/null +++ b/tests/test_shardformer/test_layer/test_sequence_parallel.py @@ -0,0 +1,178 @@ +import copy + +import torch +import torch.distributed as dist +import torch.nn as nn +import torch.nn.functional as F +from torch import Tensor +from torch.testing import assert_close + +import colossalai +from colossalai.shardformer.layer import all_to_all_comm +from colossalai.testing import parameterize, rerun_if_address_is_in_use, spawn + + +class SequenceParallelAttention(torch.nn.Module): + """Initialization. + + Arguments: + local_attention (Module): local attention with q,k,v + sequence_process_group (ProcessGroup): sequence parallel process group + scatter_idx (int): scatter_idx for all2all comm + gather_idx (int): gather_idx for all2all comm + """ + + def __init__( + self, + heads_num: torch.Tensor, + hidden_dim: torch.Tensor, + enable_sequence_parallellism: bool = False, + sequence_process_group: dist.ProcessGroup = None, + scatter_idx: int = 2, + gather_idx: int = 1, + ) -> None: + super(SequenceParallelAttention, self).__init__() + self.spg = sequence_process_group + self.scatter_idx = scatter_idx + self.gather_idx = gather_idx + self.heads_num = heads_num + self.hidden_dim = hidden_dim + assert hidden_dim % heads_num == 0 + self.head_dim = hidden_dim // heads_num + self.enable_sequence_parallellism = enable_sequence_parallellism + + self.q = nn.Linear(hidden_dim, hidden_dim) + self.k = nn.Linear(hidden_dim, hidden_dim) + self.v = nn.Linear(hidden_dim, hidden_dim) + self.out = nn.Linear(hidden_dim, hidden_dim) + + def attn(self, q, k, v): + batch_size, seq_len = q.shape[0], q.shape[1] + + scale = self.head_dim**0.5 + qk = torch.matmul(q, k.transpose(-2, -1)) / scale + weights = F.softmax(qk, dim=-1) + + attention_score = torch.matmul(weights, v) + + return attention_score + + def forward(self, x) -> Tensor: + bsz, q_len, _ = x.size() + + seq_len = q_len * dist.get_world_size(self.spg) if self.enable_sequence_parallellism else q_len + num_heads = ( + self.heads_num // dist.get_world_size(self.spg) if self.enable_sequence_parallellism else self.heads_num + ) + + # in shape : e.g., [s/p:h:] + query_states = self.q(x) + key_states = self.k(x) + value_states = self.v(x) + + if self.enable_sequence_parallellism: + query_states = all_to_all_comm(query_states, self.spg, self.scatter_idx, self.gather_idx) + key_states = all_to_all_comm(key_states, self.spg, self.scatter_idx, self.gather_idx) + value_states = all_to_all_comm(value_states, self.spg, self.scatter_idx, self.gather_idx) + + query_states = query_states.view(bsz, seq_len, num_heads, self.head_dim).transpose(1, 2) + key_states = key_states.view(bsz, seq_len, num_heads, self.head_dim).transpose(1, 2) + value_states = value_states.view(bsz, seq_len, num_heads, self.head_dim).transpose(1, 2) + # out shape : e.g., [s:h/p:] + attn_score = self.attn(query_states, key_states, value_states) + attn_score = attn_score.transpose(1, 2).contiguous() + attn_score = attn_score.reshape(bsz, seq_len, num_heads * self.head_dim) + if self.enable_sequence_parallellism: + attn_score = all_to_all_comm(attn_score, self.spg, self.gather_idx, self.scatter_idx) + + # output e.g., [s/p::h] + output = self.out(attn_score) + + return output + + +def seq_parallel_attn(seq_len, hidden_dim, head_num, batch_size): + seq_len = seq_len + hidden_dim = hidden_dim + head_num = head_num + batch_size = batch_size + world_size = dist.get_world_size() + + x = torch.randn(batch_size, seq_len, hidden_dim).cuda() + x_unshard = x.clone() + x_unshard.requires_grad_(True) + x_input = torch.chunk(x.clone(), world_size, dim=1)[dist.get_rank()] + x_input.requires_grad_(True) + + # Multi-head Attention + mha = SequenceParallelAttention(head_num, hidden_dim).cuda() + # Multi-head Attention forward + mha_out = mha(x_unshard) + + # Sequence parallel Attention + sp_attn = SequenceParallelAttention(head_num, hidden_dim, True).cuda() + sp_attn.load_state_dict(copy.deepcopy(mha.state_dict())) + # Sequence parallel Attention forward + dist_attn_out = sp_attn(x_input) + + # gather the output of sequence parallel attention + out_list = [torch.empty_like(dist_attn_out) for _ in range(world_size)] + dist.all_gather(out_list, dist_attn_out) + seq_out = torch.cat(out_list, dim=1) + + # forward result check + assert_close(seq_out, mha_out) + + # Multi-head Attention backward + mha_out.sum().backward() + q_grad = mha.q.weight.grad + k_grad = mha.k.weight.grad + v_grad = mha.v.weight.grad + o_grad = mha.out.weight.grad + x_grad = x_unshard.grad + + # Sequence parallel Attention backward + dist_attn_out.sum().backward() + q_grad_seq = sp_attn.q.weight.grad + k_grad_seq = sp_attn.k.weight.grad + v_grad_seq = sp_attn.v.weight.grad + o_grad_seq = sp_attn.out.weight.grad + x_grad_seq = x_input.grad + # all_reduce the grad of sequence parallel attention weight + dist.all_reduce(q_grad_seq) + dist.all_reduce(k_grad_seq) + dist.all_reduce(v_grad_seq) + dist.all_reduce(o_grad_seq) + # gather the grad of sequence parallel attention input + x_grad_seq_list = [torch.empty_like(x_grad_seq) for _ in range(world_size)] + dist.all_gather(x_grad_seq_list, x_grad_seq) + x_grad_seq_gather = torch.cat(x_grad_seq_list, dim=1) + + # backward result check + assert_close(q_grad_seq, q_grad) + assert_close(k_grad_seq, k_grad) + assert_close(v_grad_seq, v_grad, atol=1e-4, rtol=1e-4) + assert_close(o_grad_seq, o_grad) + assert_close(x_grad_seq_gather, x_grad) + + +@parameterize("seq_len", [128]) +@parameterize("hidden_dim", [64]) +@parameterize("head_num", [4]) +@parameterize("batch_size", [1]) +def run_seq_parallel_attn(seq_len, hidden_dim, head_num, batch_size): + seq_parallel_attn(seq_len, hidden_dim, head_num, batch_size) + + +def check_all2all_attn(rank, world_size, port): + colossalai.launch(config={}, rank=rank, world_size=world_size, host="localhost", port=port, backend="nccl") + run_seq_parallel_attn() + + +@rerun_if_address_is_in_use() +def test_all_to_all_attention(): + spawn(check_all2all_attn, nprocs=4) + + +if __name__ == "__main__": + test_all_to_all_attention() diff --git a/tests/test_shardformer/test_model/_utils.py b/tests/test_shardformer/test_model/_utils.py index 62d4d1bf3c7c..d5fc2c30f294 100644 --- a/tests/test_shardformer/test_model/_utils.py +++ b/tests/test_shardformer/test_model/_utils.py @@ -1,5 +1,4 @@ import copy -import math from contextlib import nullcontext from typing import Any, Callable, Dict, List, Optional @@ -31,6 +30,7 @@ def build_model( enable_jit_fused=False, enable_sequence_parallelism=False, use_lazy_init: bool = False, + dtype=torch.float32, ): # create new model ctx = LazyInitContext() if use_lazy_init else nullcontext() @@ -51,7 +51,7 @@ def build_model( model_copy = copy.deepcopy(org_model) shard_former = ShardFormer(shard_config=shard_config) sharded_model, shared_params = shard_former.optimize(model_copy) - return org_model.cuda(), sharded_model.cuda() + return org_model.cuda().to(dtype), sharded_model.cuda().to(dtype) def build_pipeline_model( @@ -122,7 +122,6 @@ def build_model_from_hybrid_plugin(model_fn: Callable, loss_fn: Callable, test_c sharded_model = copy.deepcopy(org_model) if use_lazy_init: ctx.materialize(org_model) - org_model = org_model.cuda() org_optimizer = Adam(org_model.parameters(), lr=1e-3) sharded_optimizer = Adam(sharded_model.parameters(), lr=1e-3) @@ -132,7 +131,14 @@ def build_model_from_hybrid_plugin(model_fn: Callable, loss_fn: Callable, test_c booster = Booster(plugin=plugin) sharded_model, sharded_optimizer, criterion, _, _ = booster.boost(sharded_model, sharded_optimizer, criterion) - return org_model, org_optimizer, sharded_model, sharded_optimizer, criterion, booster + return ( + org_model, + org_optimizer, + sharded_model, + sharded_optimizer, + criterion, + booster, + ) def run_forward_backward_with_hybrid_plugin( @@ -154,39 +160,47 @@ def _criterion(outputs, inputs): data = data_gen_fn() - if booster.plugin.shard_config.enable_sequence_parallelism and booster.plugin.tp_size != 0: - seq_len = data["input_ids"].shape[-1] - lcm = booster.plugin.tp_size * seq_len // math.gcd(booster.plugin.tp_size, seq_len) - times = lcm // seq_len - input_shape = data["input_ids"].shape - for k, v in data.items(): - if v.shape == input_shape: - data[k] = v.repeat((1,) * (v.dim() - 1) + (times,)) + shard_test_data = {} + for k, v in data.items(): + shard_test_data[k] = data[k].clone() + unshard_test_data = {} + for k, v in data.items(): + unshard_test_data[k] = data[k].clone() sharded_model.train() if booster.plugin.stage_manager is not None: - for k, v in data.items(): + for k, v in shard_test_data.items(): if torch.is_tensor(v) or "Tensor" in v.__class__.__name__: new_shape = [1] * v.dim() new_shape[0] = 4 - data[k] = v.to("cuda").repeat(*new_shape) + shard_test_data[k] = v.to("cuda").repeat(*new_shape) - data_iter = iter([data]) + data_iter = iter([shard_test_data]) sharded_output = booster.execute_pipeline( - data_iter, sharded_model, _criterion, sharded_optimizer, return_loss=True, return_outputs=True + data_iter, + sharded_model, + _criterion, + sharded_optimizer, + return_loss=True, + return_outputs=True, ) sharded_loss = sharded_output["loss"] - else: - data = {k: v.cuda() for k, v in data.items()} - sharded_output = sharded_model(**data) + else: + shard_test_data = {k: v.cuda() for k, v in shard_test_data.items()} + sharded_output = sharded_model(**shard_test_data) sharded_loss = criterion(sharded_output) sharded_optimizer.backward(sharded_loss) org_model.train() - data = {k: v.cuda() for k, v in data.items()} - org_output = org_model(**data) - + if booster.plugin.stage_manager is not None: + for k, v in unshard_test_data.items(): + if torch.is_tensor(v) or "Tensor" in v.__class__.__name__: + new_shape = [1] * v.dim() + new_shape[0] = 4 + unshard_test_data[k] = v.to("cuda").repeat(*new_shape) + unshard_test_data = {k: v.cuda() for k, v in unshard_test_data.items()} + org_output = org_model(**unshard_test_data) org_loss = criterion(org_output) org_loss.backward() @@ -199,7 +213,6 @@ def check_output_hidden_state( stage_manager: Optional[PipelineStageManager] = None, atol: float = 1e-5, rtol: float = 1e-3, - dim: int = 0, ): org_hidden_state = org_output.last_hidden_state @@ -313,7 +326,9 @@ def check_grad( def unwrap_model( - module: Module, base_model_class_name: Optional[str] = None, base_model_attribute_name: Optional[str] = None + module: Module, + base_model_class_name: Optional[str] = None, + base_model_attribute_name: Optional[str] = None, ): if isinstance(module, HybridParallelModule): module = module.unwrap() diff --git a/tests/test_shardformer/test_model/test_shard_bert.py b/tests/test_shardformer/test_model/test_shard_bert.py index 768bd95bdb42..919557797fcd 100644 --- a/tests/test_shardformer/test_model/test_shard_bert.py +++ b/tests/test_shardformer/test_model/test_shard_bert.py @@ -100,6 +100,28 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn, @parameterize( "test_config", [ + { + "tp_size": 4, + "pp_size": 1, + "num_microbatches": 1, + "enable_sequence_parallelism": True, + "sequence_parallelism_mode": "ring", + "enable_flash_attention": False, + "use_lazy_init": True, + "precision": "fp32", + "initial_scale": 1, + }, + { + "tp_size": 4, + "pp_size": 1, + "num_microbatches": 1, + "enable_sequence_parallelism": True, + "sequence_parallelism_mode": "split_gather", + "enable_flash_attention": False, + "use_lazy_init": True, + "precision": "fp16", + "initial_scale": 1, + }, { "tp_size": 2, "pp_size": 1, @@ -154,7 +176,6 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn, ) def run_bert_test(test_config): sub_model_zoo = model_zoo.get_sub_registry("transformers_bert") - for name, (model_fn, data_gen_fn, output_transform_fn, loss_fn, _) in sub_model_zoo.items(): check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn, test_config) diff --git a/tests/test_shardformer/test_model/test_shard_blip2.py b/tests/test_shardformer/test_model/test_shard_blip2.py index 02c15460ecb3..2c56b0435a6d 100644 --- a/tests/test_shardformer/test_model/test_shard_blip2.py +++ b/tests/test_shardformer/test_model/test_shard_blip2.py @@ -45,19 +45,51 @@ def check_forward_backward(org_model, sharded_model, data_gen_fn, output_transfo "qformer.encoder.layer[0].attention.output.dense", "language_model.model.decoder.layers[0].self_attn.out_proj", ] - check_grad(blip2, sharded_blip2, col_layer_for_check, atol=1e-6, rtol=1e-5, dim=0, verbose=False) - check_grad(blip2, sharded_blip2, row_layer_for_check, atol=1e-6, rtol=1e-5, dim=1, verbose=False) + check_grad( + blip2, + sharded_blip2, + col_layer_for_check, + atol=1e-6, + rtol=1e-5, + dim=0, + verbose=False, + ) + check_grad( + blip2, + sharded_blip2, + row_layer_for_check, + atol=1e-6, + rtol=1e-5, + dim=1, + verbose=False, + ) @parameterize("enable_fused_normalization", [True, False]) @parameterize("enable_tensor_parallelism", [True, False]) @parameterize("enable_flash_attention", [True, False]) @parameterize("enable_jit_fused", [True, False]) -def run_blip2_test(enable_fused_normalization, enable_tensor_parallelism, enable_flash_attention, enable_jit_fused): +def run_blip2_test( + enable_fused_normalization, + enable_tensor_parallelism, + enable_flash_attention, + enable_jit_fused, +): sub_model_zoo = model_zoo.get_sub_registry("transformers_blip2") - for name, (model_fn, data_gen_fn, output_transform_fn, loss_fn, _) in sub_model_zoo.items(): + for name, ( + model_fn, + data_gen_fn, + output_transform_fn, + loss_fn, + _, + ) in sub_model_zoo.items(): org_model, sharded_model = build_model( - model_fn, enable_fused_normalization, enable_tensor_parallelism, enable_flash_attention, enable_jit_fused + model_fn, + enable_fused_normalization, + enable_tensor_parallelism, + enable_flash_attention, + enable_jit_fused, + dtype=torch.float, ) check_forward_backward(org_model, sharded_model, data_gen_fn, output_transform_fn, loss_fn) @@ -66,7 +98,14 @@ def run_blip2_test(enable_fused_normalization, enable_tensor_parallelism, enable def check_blip2(rank, world_size, port): disable_existing_loggers() - colossalai.launch(config={}, rank=rank, world_size=world_size, host="localhost", port=port, backend="nccl") + colossalai.launch( + config={}, + rank=rank, + world_size=world_size, + host="localhost", + port=port, + backend="nccl", + ) run_blip2_test() diff --git a/tests/test_shardformer/test_model/test_shard_bloom.py b/tests/test_shardformer/test_model/test_shard_bloom.py index b70cba8b4a53..cc0786618853 100644 --- a/tests/test_shardformer/test_model/test_shard_bloom.py +++ b/tests/test_shardformer/test_model/test_shard_bloom.py @@ -99,6 +99,28 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn, @parameterize( "test_config", [ + { + "tp_size": 4, + "pp_size": 1, + "num_microbatches": 1, + "enable_sequence_parallelism": True, + "sequence_parallelism_mode": "ring", + "enable_flash_attention": False, + "use_lazy_init": True, + "precision": "fp32", + "initial_scale": 1, + }, + { + "tp_size": 4, + "pp_size": 1, + "num_microbatches": 1, + "enable_sequence_parallelism": True, + "sequence_parallelism_mode": "split_gather", + "enable_flash_attention": False, + "use_lazy_init": True, + "precision": "fp16", + "initial_scale": 1, + }, { "tp_size": 2, "pp_size": 2, diff --git a/tests/test_shardformer/test_model/test_shard_chatglm2.py b/tests/test_shardformer/test_model/test_shard_chatglm2.py index 29d3592bf34e..405ceba328df 100644 --- a/tests/test_shardformer/test_model/test_shard_chatglm2.py +++ b/tests/test_shardformer/test_model/test_shard_chatglm2.py @@ -11,7 +11,6 @@ build_model_from_hybrid_plugin, check_all_grad_tensors, check_loss, - check_output_hidden_state, check_weight, get_grad_tensors_for_check, run_forward_backward_with_hybrid_plugin, @@ -25,7 +24,13 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn, ) org_loss, org_output, sharded_loss, sharded_output = run_forward_backward_with_hybrid_plugin( - org_model, sharded_model, sharded_optimizer, data_gen_fn, output_transform_fn, criterion, booster + org_model, + sharded_model, + sharded_optimizer, + data_gen_fn, + output_transform_fn, + criterion, + booster, ) stage_manager = booster.plugin.stage_manager @@ -36,7 +41,10 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn, shard_chatglm_model = unwrap_model(sharded_model, "ChatGLMModel", "transformer") norm_layer_for_check = ["encoder.layers[0].input_layernorm"] - row_layer_for_check = ["encoder.layers[0].self_attention.query_key_value", "embedding.word_embeddings"] + row_layer_for_check = [ + "encoder.layers[0].self_attention.query_key_value", + "embedding.word_embeddings", + ] col_layer_for_check = ["encoder.layers[0].self_attention.dense"] # Save gradient tensors for comparison between the original model and the sharded model. @@ -94,8 +102,9 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn, else: atol, rtol = 5e-3, 5e-3 - if org_model.__class__.__name__ == "ChatGLMModel": - check_output_hidden_state(org_output, sharded_output, stage_manager, atol=atol, rtol=rtol, dim=1) + # TODO: ChatGLMModel output is [S, B, H], merging batch of pipeline is wrong + # if org_model.__class__.__name__ == "ChatGLMModel": + # check_output_hidden_state(org_output, sharded_output, stage_manager, atol=atol, rtol=rtol, dim=1) check_loss(org_loss, sharded_loss, atol=atol, rtol=rtol) @@ -126,6 +135,28 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn, @parameterize( "test_config", [ + { + "tp_size": 4, + "pp_size": 1, + "num_microbatches": 1, + "enable_sequence_parallelism": True, + "sequence_parallelism_mode": "ring", + "enable_flash_attention": False, + "use_lazy_init": True, + "precision": "fp32", + "initial_scale": 1, + }, + { + "tp_size": 4, + "pp_size": 1, + "num_microbatches": 1, + "enable_sequence_parallelism": True, + "sequence_parallelism_mode": "split_gather", + "enable_flash_attention": False, + "use_lazy_init": True, + "precision": "fp16", + "initial_scale": 1, + }, { "tp_size": 2, "pp_size": 2, @@ -143,8 +174,20 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn, "use_lazy_init": False, "precision": "fp32", }, - {"tp_size": 4, "pp_size": 1, "enable_all_optimization": True, "use_lazy_init": False, "precision": "fp32"}, - {"tp_size": 2, "pp_size": 1, "enable_all_optimization": True, "use_lazy_init": False, "precision": "fp32"}, + { + "tp_size": 4, + "pp_size": 1, + "enable_all_optimization": True, + "use_lazy_init": False, + "precision": "fp32", + }, + { + "tp_size": 2, + "pp_size": 1, + "enable_all_optimization": True, + "use_lazy_init": False, + "precision": "fp32", + }, { "tp_size": 2, "pp_size": 1, @@ -159,7 +202,13 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn, def run_chatglm_test(test_config): sub_model_zoo = model_zoo.get_sub_registry("transformers_chatglm") - for name, (model_fn, data_gen_fn, output_transform_fn, loss_fn, _) in sub_model_zoo.items(): + for name, ( + model_fn, + data_gen_fn, + output_transform_fn, + loss_fn, + _, + ) in sub_model_zoo.items(): check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn, test_config) clear_layout_converter() @@ -193,7 +242,13 @@ def run_chatglm_test(test_config): def run_chatglm_3d_test(test_config): sub_model_zoo = model_zoo.get_sub_registry("transformers_chatglm") - for name, (model_fn, data_gen_fn, output_transform_fn, loss_fn, _) in sub_model_zoo.items(): + for name, ( + model_fn, + data_gen_fn, + output_transform_fn, + loss_fn, + _, + ) in sub_model_zoo.items(): check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn, test_config) clear_layout_converter() @@ -202,13 +257,27 @@ def run_chatglm_3d_test(test_config): def check_chatglm(rank, world_size, port): disable_existing_loggers() - colossalai.launch(config={}, rank=rank, world_size=world_size, host="localhost", port=port, backend="nccl") + colossalai.launch( + config={}, + rank=rank, + world_size=world_size, + host="localhost", + port=port, + backend="nccl", + ) run_chatglm_test() def check_chatglm_3d(rank, world_size, port): disable_existing_loggers() - colossalai.launch(config={}, rank=rank, world_size=world_size, host="localhost", port=port, backend="nccl") + colossalai.launch( + config={}, + rank=rank, + world_size=world_size, + host="localhost", + port=port, + backend="nccl", + ) run_chatglm_3d_test() diff --git a/tests/test_shardformer/test_model/test_shard_gpt2.py b/tests/test_shardformer/test_model/test_shard_gpt2.py index 3155420f1cf2..4aac7f3d4ed7 100644 --- a/tests/test_shardformer/test_model/test_shard_gpt2.py +++ b/tests/test_shardformer/test_model/test_shard_gpt2.py @@ -25,7 +25,13 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn, ) org_loss, org_output, sharded_loss, sharded_output = run_forward_backward_with_hybrid_plugin( - org_model, sharded_model, sharded_optimizer, data_gen_fn, output_transform_fn, criterion, booster + org_model, + sharded_model, + sharded_optimizer, + data_gen_fn, + output_transform_fn, + criterion, + booster, ) stage_manager = booster.plugin.stage_manager @@ -47,10 +53,24 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn, else: atol, rtol = 5e-3, 5e-3 col_layer_grads = get_grad_tensors_for_check( - gpt2, sharded_gpt2, col_layer_for_check, tp_group, atol=atol, rtol=rtol, dim=1, verbose=False + gpt2, + sharded_gpt2, + col_layer_for_check, + tp_group, + atol=atol, + rtol=rtol, + dim=1, + verbose=False, ) row_layer_grads = get_grad_tensors_for_check( - gpt2, sharded_gpt2, row_layer_for_check, tp_group, atol=atol, rtol=rtol, dim=0, verbose=False + gpt2, + sharded_gpt2, + row_layer_for_check, + tp_group, + atol=atol, + rtol=rtol, + dim=0, + verbose=False, ) norm_layer_grads = get_grad_tensors_for_check( @@ -90,7 +110,16 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn, atol, rtol = 5e-3, 1e-3 else: atol, rtol = 5e-3, 5e-3 - check_weight(gpt2, sharded_gpt2, col_layer_for_check, tp_group, atol=atol, rtol=rtol, dim=1, verbose=False) + check_weight( + gpt2, + sharded_gpt2, + col_layer_for_check, + tp_group, + atol=atol, + rtol=rtol, + dim=1, + verbose=False, + ) # check grads check_all_grad_tensors(grads_to_check) @@ -102,6 +131,28 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn, @parameterize( "test_config", [ + { + "tp_size": 4, + "pp_size": 1, + "num_microbatches": 1, + "enable_sequence_parallelism": True, + "sequence_parallelism_mode": "ring", + "enable_flash_attention": False, + "use_lazy_init": True, + "precision": "fp32", + "initial_scale": 1, + }, + { + "tp_size": 4, + "pp_size": 1, + "num_microbatches": 1, + "enable_sequence_parallelism": True, + "sequence_parallelism_mode": "split_gather", + "enable_flash_attention": False, + "use_lazy_init": True, + "precision": "fp16", + "initial_scale": 1, + }, { "tp_size": 2, "pp_size": 2, @@ -123,14 +174,14 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn, { "tp_size": 4, "pp_size": 1, - "enable_all_optimization": True, + "enable_all_optimization": False, "use_lazy_init": False, "precision": "fp32", }, { "tp_size": 2, "pp_size": 1, - "enable_all_optimization": True, + "enable_all_optimization": False, "use_lazy_init": False, "precision": "fp32", }, @@ -138,7 +189,7 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn, "tp_size": 2, "pp_size": 2, "num_microbatches": 4, - "enable_all_optimization": True, + "enable_all_optimization": False, "use_lazy_init": True, "precision": "fp32", }, @@ -167,7 +218,13 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn, def run_gpt2_test(test_config): sub_model_zoo = model_zoo.get_sub_registry("transformers_gpt", exclude="transformers_gptj") - for name, (model_fn, data_gen_fn, output_transform_fn, loss_fn, _) in sub_model_zoo.items(): + for name, ( + model_fn, + data_gen_fn, + output_transform_fn, + loss_fn, + _, + ) in sub_model_zoo.items(): check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn, test_config) clear_layout_converter() @@ -202,7 +259,13 @@ def run_gpt2_test(test_config): def run_gpt2_3d_test(test_config): sub_model_zoo = model_zoo.get_sub_registry("transformers_gpt", exclude="transformers_gptj") - for name, (model_fn, data_gen_fn, output_transform_fn, loss_fn, _) in sub_model_zoo.items(): + for name, ( + model_fn, + data_gen_fn, + output_transform_fn, + loss_fn, + _, + ) in sub_model_zoo.items(): check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn, test_config) clear_layout_converter() @@ -211,13 +274,27 @@ def run_gpt2_3d_test(test_config): def check_gpt2(rank, world_size, port): disable_existing_loggers() - colossalai.launch(config={}, rank=rank, world_size=world_size, host="localhost", port=port, backend="nccl") + colossalai.launch( + config={}, + rank=rank, + world_size=world_size, + host="localhost", + port=port, + backend="nccl", + ) run_gpt2_test() def check_gpt2_3d(rank, world_size, port): disable_existing_loggers() - colossalai.launch(config={}, rank=rank, world_size=world_size, host="localhost", port=port, backend="nccl") + colossalai.launch( + config={}, + rank=rank, + world_size=world_size, + host="localhost", + port=port, + backend="nccl", + ) run_gpt2_3d_test() diff --git a/tests/test_shardformer/test_model/test_shard_gptj.py b/tests/test_shardformer/test_model/test_shard_gptj.py index 0bf9669808fa..009202a0da7a 100644 --- a/tests/test_shardformer/test_model/test_shard_gptj.py +++ b/tests/test_shardformer/test_model/test_shard_gptj.py @@ -25,7 +25,13 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn, ) org_loss, org_output, sharded_loss, sharded_output = run_forward_backward_with_hybrid_plugin( - org_model, sharded_model, sharded_optimizer, data_gen_fn, output_transform_fn, criterion, booster + org_model, + sharded_model, + sharded_optimizer, + data_gen_fn, + output_transform_fn, + criterion, + booster, ) stage_manager = booster.plugin.stage_manager @@ -46,11 +52,25 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn, else: atol, rtol = 5e-3, 5e-3 col_layer_grads = get_grad_tensors_for_check( - gptj, sharded_gptj, col_layer_for_check, tp_group, atol=atol, rtol=rtol, dim=0, verbose=False + gptj, + sharded_gptj, + col_layer_for_check, + tp_group, + atol=atol, + rtol=rtol, + dim=0, + verbose=False, ) row_layer_grads = get_grad_tensors_for_check( - gptj, sharded_gptj, row_layer_for_check, tp_group, atol=atol, rtol=rtol, dim=1, verbose=False + gptj, + sharded_gptj, + row_layer_for_check, + tp_group, + atol=atol, + rtol=rtol, + dim=1, + verbose=False, ) grads_to_check.update(col_layer_grads) grads_to_check.update(row_layer_grads) @@ -77,7 +97,16 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn, atol, rtol = 5e-3, 1e-3 else: atol, rtol = 5e-3, 5e-3 - check_weight(gptj, sharded_gptj, col_layer_for_check, tp_group, atol=atol, rtol=rtol, dim=0, verbose=False) + check_weight( + gptj, + sharded_gptj, + col_layer_for_check, + tp_group, + atol=atol, + rtol=rtol, + dim=0, + verbose=False, + ) # check grads check_all_grad_tensors(grads_to_check) @@ -110,14 +139,14 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn, { "tp_size": 4, "pp_size": 1, - "enable_all_optimization": True, + "enable_all_optimization": False, "use_lazy_init": False, "precision": "fp32", }, { "tp_size": 2, "pp_size": 1, - "enable_all_optimization": True, + "enable_all_optimization": False, "use_lazy_init": False, "precision": "fp32", }, @@ -125,7 +154,7 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn, "tp_size": 2, "pp_size": 2, "num_microbatches": 4, - "enable_all_optimization": True, + "enable_all_optimization": False, #'use_lazy_init': True, "precision": "fp32", }, @@ -154,7 +183,13 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn, def run_gptj_test(test_config): sub_model_zoo = model_zoo.get_sub_registry("transformers_gptj") - for name, (model_fn, data_gen_fn, output_transform_fn, loss_fn, _) in sub_model_zoo.items(): + for name, ( + model_fn, + data_gen_fn, + output_transform_fn, + loss_fn, + _, + ) in sub_model_zoo.items(): check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn, test_config) clear_layout_converter() @@ -189,7 +224,13 @@ def run_gptj_test(test_config): def run_gptj_3d_test(test_config): sub_model_zoo = model_zoo.get_sub_registry("transformers_gptj") - for name, (model_fn, data_gen_fn, output_transform_fn, loss_fn, _) in sub_model_zoo.items(): + for name, ( + model_fn, + data_gen_fn, + output_transform_fn, + loss_fn, + _, + ) in sub_model_zoo.items(): check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn, test_config) clear_layout_converter() @@ -198,13 +239,27 @@ def run_gptj_3d_test(test_config): def check_gptj(rank, world_size, port): disable_existing_loggers() - colossalai.launch(config={}, rank=rank, world_size=world_size, host="localhost", port=port, backend="nccl") + colossalai.launch( + config={}, + rank=rank, + world_size=world_size, + host="localhost", + port=port, + backend="nccl", + ) run_gptj_test() def check_gptj_3d(rank, world_size, port): disable_existing_loggers() - colossalai.launch(config={}, rank=rank, world_size=world_size, host="localhost", port=port, backend="nccl") + colossalai.launch( + config={}, + rank=rank, + world_size=world_size, + host="localhost", + port=port, + backend="nccl", + ) run_gptj_3d_test() diff --git a/tests/test_shardformer/test_model/test_shard_llama.py b/tests/test_shardformer/test_model/test_shard_llama.py index c7edcfb3510c..27f904292597 100644 --- a/tests/test_shardformer/test_model/test_shard_llama.py +++ b/tests/test_shardformer/test_model/test_shard_llama.py @@ -2,9 +2,12 @@ import pytest import torch +import torch.distributed as dist +from torch.testing import assert_close import colossalai from colossalai.logging import disable_existing_loggers +from colossalai.shardformer import PipelineGradientCheckpointConfig from colossalai.shardformer.layer.utils import Randomizer from colossalai.tensor.d_tensor.api import clear_layout_converter from colossalai.testing import clear_cache_before_run, parameterize, rerun_if_address_is_in_use, spawn @@ -24,9 +27,13 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn, test_config): + enable_gradient_checkpointing = test_config.pop("enable_gradient_checkpointing", False) org_model, org_optimizer, sharded_model, sharded_optimizer, criterion, booster = build_model_from_hybrid_plugin( model_fn, loss_fn, test_config ) + if enable_gradient_checkpointing: + org_model.gradient_checkpointing_enable() + sharded_model.unwrap().gradient_checkpointing_enable() org_loss, org_output, sharded_loss, sharded_output = run_forward_backward_with_hybrid_plugin( org_model, sharded_model, sharded_optimizer, data_gen_fn, output_transform_fn, criterion, booster @@ -41,6 +48,26 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn, row_layer_for_check = ["layers[0].self_attn.q_proj", "embed_tokens"] col_layer_for_check = ["layers[0].self_attn.o_proj"] + # Here we check the grad of layernorm because an all-reduce operation should be performed during sequence parallelism + norm_layer_for_check = ["layers[0].input_layernorm", "layers[0].post_attention_layernorm"] + + # During pipeline parallelism, we cannot get the grad of norm layer during first stage, so we only check this when pp is not enbaled + if stage_manager is None: + norm_layer_for_check.append("norm") + + # Check the grad when using ZeRO-1 and ZeRO-2 + if ( + booster.plugin.zero_stage in [1, 2] + and booster.plugin.shard_config.enable_sequence_parallelism + and booster.plugin.shard_config.sequence_parallelism_mode == "all_to_all" + ): + for p1, p2 in zip(llama_model.parameters(), sharded_optimizer._master_param_groups_of_current_rank[0]): + working_p = sharded_optimizer._param_store.master_to_working_param[id(p2)] + grads = sharded_optimizer._grad_store.get_partitioned_gradients_by_param_id(0, id(working_p)) + grad_index = 0 if sharded_optimizer._partition_grads else sharded_optimizer._local_rank + grad = grads[grad_index] + sharded_grad = p1.grad.view(-1).chunk(dist.get_world_size())[dist.get_rank()] + assert_close(sharded_grad, grad[: sharded_grad.shape[0]], atol=5e-3, rtol=5e-3, check_dtype=False) # Save gradient tensors for comparison between the original model and the sharded model before optimizer step. grads_to_check = {} @@ -55,8 +82,19 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn, col_layer_grads = get_grad_tensors_for_check( llama_model, shard_llama_model, col_layer_for_check, tp_group, atol=atol, rtol=rtol, dim=1, verbose=False ) + norm_layer_grads = get_grad_tensors_for_check( + llama_model, + shard_llama_model, + norm_layer_for_check, + tp_group, + atol=atol, + rtol=rtol, + dim=1, + verbose=False, + ) grads_to_check.update(col_layer_grads) grads_to_check.update(row_layer_grads) + grads_to_check.update(norm_layer_grads) # optimizer executes step org_optimizer.step() @@ -93,6 +131,74 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn, @parameterize( "test_config", [ + { + "tp_size": 2, + "pp_size": 1, + "num_microbatches": 1, + "enable_sequence_parallelism": True, + "sequence_parallelism_mode": "ring", + "enable_flash_attention": True, + "use_lazy_init": True, + "zero_stage": 2, + "precision": "fp16", + "initial_scale": 1, + }, + { + "tp_size": 4, + "pp_size": 1, + "num_microbatches": 1, + "enable_sequence_parallelism": True, + "sequence_parallelism_mode": "ring", + "enable_flash_attention": False, + "use_lazy_init": True, + "precision": "fp32", + "initial_scale": 1, + }, + { + "tp_size": 4, + "pp_size": 1, + "num_microbatches": 1, + "enable_sequence_parallelism": True, + "sequence_parallelism_mode": "split_gather", + "enable_flash_attention": False, + "use_lazy_init": True, + "precision": "fp16", + "initial_scale": 1, + }, + { + "tp_size": 1, + "pp_size": 1, + "sp_size": 2, + "num_microbatches": 1, + "enable_sequence_parallelism": True, + "sequence_parallelism_mode": "all_to_all", + "use_lazy_init": True, + "precision": "fp16", + "initial_scale": 1, + }, + { + "tp_size": 1, + "pp_size": 1, + "sp_size": 2, + "num_microbatches": 1, + "enable_sequence_parallelism": True, + "sequence_parallelism_mode": "all_to_all", + "use_lazy_init": True, + "zero_stage": 2, + "precision": "fp16", + "initial_scale": 1, + }, + { + "tp_size": 1, + "pp_size": 1, + "num_microbatches": 1, + "enable_sequence_parallelism": True, + "sequence_parallelism_mode": "all_to_all", + "enable_flash_attention": False, + "use_lazy_init": True, + "precision": "fp16", + "initial_scale": 1, + }, { "tp_size": 2, "pp_size": 2, @@ -101,6 +207,8 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn, "use_lazy_init": True, "precision": "fp16", "initial_scale": 1, + "enable_gradient_checkpointing": True, + "gradient_checkpoint_config": PipelineGradientCheckpointConfig(gradient_checkpointing_ratio=0.5), }, { "tp_size": 1, @@ -108,11 +216,15 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn, "num_microbatches": 4, "use_lazy_init": False, "precision": "fp32", + "enable_gradient_checkpointing": True, + "gradient_checkpoint_config": PipelineGradientCheckpointConfig( + num_stages=2, num_model_chunks=1, num_model_layers=8, num_ckpt_layers_per_stage=[4, 0] + ), }, { "tp_size": 4, "pp_size": 1, - "enable_all_optimization": True, + "enable_all_optimization": False, "use_lazy_init": False, "precision": "fp32", }, @@ -124,7 +236,7 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn, "use_lazy_init": False, "precision": "fp32", }, - {"tp_size": 2, "pp_size": 1, "enable_all_optimization": True, "use_lazy_init": False, "precision": "fp32"}, + {"tp_size": 2, "pp_size": 1, "enable_all_optimization": False, "use_lazy_init": False, "precision": "fp32"}, { "tp_size": 2, "pp_size": 1, @@ -189,6 +301,13 @@ def run_llama_test(test_config): "precision": "fp16", "zero_stage": 1, "initial_scale": 1, + "enable_gradient_checkpointing": True, + "gradient_checkpoint_config": PipelineGradientCheckpointConfig( + num_stages=2, + num_model_chunks=2, + num_model_layers=8, + num_ckpt_layers_per_stage=[0, 1, 2, 2], + ), }, ], ) diff --git a/tests/test_shardformer/test_model/test_shard_opt.py b/tests/test_shardformer/test_model/test_shard_opt.py index d21ab264d8ab..523ed879bcf7 100644 --- a/tests/test_shardformer/test_model/test_shard_opt.py +++ b/tests/test_shardformer/test_model/test_shard_opt.py @@ -29,7 +29,13 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn, ) org_loss, org_output, sharded_loss, sharded_output = run_forward_backward_with_hybrid_plugin( - org_model, sharded_model, sharded_optimizer, data_gen_fn, output_transform_fn, criterion, booster + org_model, + sharded_model, + sharded_optimizer, + data_gen_fn, + output_transform_fn, + criterion, + booster, ) stage_manager = booster.plugin.stage_manager @@ -39,7 +45,10 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn, opt_model = unwrap_model(org_model, "OPTModel", "model") shard_opt_model = unwrap_model(sharded_model, "OPTModel", "model") - row_layer_for_check = ["decoder.layers[0].self_attn.q_proj", "decoder.embed_tokens"] # 'decoder.embed_tokens' + row_layer_for_check = [ + "decoder.layers[0].self_attn.q_proj", + "decoder.embed_tokens", + ] # 'decoder.embed_tokens' col_layer_for_check = ["decoder.layers[0].self_attn.out_proj"] # Save gradient tensors for comparison between the original model and the sharded model. @@ -50,10 +59,24 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn, else: atol, rtol = 4e-2, 4e-2 row_layer_grads = get_grad_tensors_for_check( - opt_model, shard_opt_model, row_layer_for_check, tp_group, atol=atol, rtol=rtol, dim=0, verbose=False + opt_model, + shard_opt_model, + row_layer_for_check, + tp_group, + atol=atol, + rtol=rtol, + dim=0, + verbose=False, ) col_layer_grads = get_grad_tensors_for_check( - opt_model, shard_opt_model, col_layer_for_check, tp_group, atol=atol, rtol=rtol, dim=1, verbose=False + opt_model, + shard_opt_model, + col_layer_for_check, + tp_group, + atol=atol, + rtol=rtol, + dim=1, + verbose=False, ) grads_to_check.update(col_layer_grads) grads_to_check.update(row_layer_grads) @@ -80,7 +103,14 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn, else: atol, rtol = 5e-3, 5e-3 check_weight( - opt_model, shard_opt_model, col_layer_for_check, tp_group, atol=atol, rtol=rtol, dim=1, verbose=False + opt_model, + shard_opt_model, + col_layer_for_check, + tp_group, + atol=atol, + rtol=rtol, + dim=1, + verbose=False, ) # check grads @@ -110,8 +140,20 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn, "use_lazy_init": False, "precision": "fp32", }, - {"tp_size": 4, "pp_size": 1, "enable_all_optimization": True, "use_lazy_init": False, "precision": "fp32"}, - {"tp_size": 2, "pp_size": 1, "enable_all_optimization": True, "use_lazy_init": False, "precision": "fp32"}, + { + "tp_size": 4, + "pp_size": 1, + "enable_all_optimization": False, + "use_lazy_init": False, + "precision": "fp32", + }, + { + "tp_size": 2, + "pp_size": 1, + "enable_all_optimization": False, + "use_lazy_init": False, + "precision": "fp32", + }, { "tp_size": 2, "pp_size": 1, @@ -135,7 +177,13 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn, ) def run_opt_test(test_config): sub_model_zoo = model_zoo.get_sub_registry("transformers_opt") - for name, (model_fn, data_gen_fn, output_transform_fn, loss_fn, _) in sub_model_zoo.items(): + for name, ( + model_fn, + data_gen_fn, + output_transform_fn, + loss_fn, + _, + ) in sub_model_zoo.items(): check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn, test_config) clear_layout_converter() @@ -169,7 +217,13 @@ def run_opt_test(test_config): def run_opt_3d_test(test_config): sub_model_zoo = model_zoo.get_sub_registry("transformers_opt") - for name, (model_fn, data_gen_fn, output_transform_fn, loss_fn, _) in sub_model_zoo.items(): + for name, ( + model_fn, + data_gen_fn, + output_transform_fn, + loss_fn, + _, + ) in sub_model_zoo.items(): check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn, test_config) clear_layout_converter() @@ -178,13 +232,27 @@ def run_opt_3d_test(test_config): def check_OPTModel(rank, world_size, port): disable_existing_loggers() - colossalai.launch(config={}, rank=rank, world_size=world_size, host="localhost", port=port, backend="nccl") + colossalai.launch( + config={}, + rank=rank, + world_size=world_size, + host="localhost", + port=port, + backend="nccl", + ) run_opt_test() def check_opt_3d(rank, world_size, port): disable_existing_loggers() - colossalai.launch(config={}, rank=rank, world_size=world_size, host="localhost", port=port, backend="nccl") + colossalai.launch( + config={}, + rank=rank, + world_size=world_size, + host="localhost", + port=port, + backend="nccl", + ) run_opt_3d_test() diff --git a/tests/test_shardformer/test_model/test_shard_t5.py b/tests/test_shardformer/test_model/test_shard_t5.py index 22c201458ad4..9b22d54d7d31 100644 --- a/tests/test_shardformer/test_model/test_shard_t5.py +++ b/tests/test_shardformer/test_model/test_shard_t5.py @@ -25,7 +25,13 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn, ) org_loss, org_output, sharded_loss, sharded_output = run_forward_backward_with_hybrid_plugin( - org_model, sharded_model, sharded_optimizer, data_gen_fn, output_transform_fn, criterion, booster + org_model, + sharded_model, + sharded_optimizer, + data_gen_fn, + output_transform_fn, + criterion, + booster, ) stage_manager = booster.plugin.stage_manager @@ -71,7 +77,16 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn, else: atol, rtol = 5e-3, 5e-3 if stage_manager is None or stage_manager.is_first_stage(): - check_weight(t5, sharded_t5, row_layer_for_check, tp_group, atol=atol, rtol=rtol, dim=0, verbose=False) + check_weight( + t5, + sharded_t5, + row_layer_for_check, + tp_group, + atol=atol, + rtol=rtol, + dim=0, + verbose=False, + ) # check grads check_all_grad_tensors(grads_to_check) @@ -104,7 +119,7 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn, { "tp_size": 4, "pp_size": 1, - "enable_all_optimization": True, + "enable_all_optimization": False, "use_lazy_init": False, "precision": "fp32", }, @@ -117,7 +132,6 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn, "use_lazy_init": False, "precision": "fp32", }, - {"tp_size": 2, "pp_size": 1, "enable_all_optimization": True, "use_lazy_init": False, "precision": "fp32"}, { "tp_size": 2, "pp_size": 1, @@ -144,7 +158,13 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn, def run_t5_test(test_config): sub_model_zoo = model_zoo.get_sub_registry("transformers_t5") - for name, (model_fn, data_gen_fn, output_transform_fn, loss_fn, _) in sub_model_zoo.items(): + for name, ( + model_fn, + data_gen_fn, + output_transform_fn, + loss_fn, + _, + ) in sub_model_zoo.items(): # skip 4-stage pp test for t5_encoder if test_config["pp_size"] > 2 and name == "transformers_t5_encoder_model": continue @@ -185,7 +205,13 @@ def run_t5_test(test_config): def run_t5_3d_test(test_config): sub_model_zoo = model_zoo.get_sub_registry("transformers_t5") - for name, (model_fn, data_gen_fn, output_transform_fn, loss_fn, _) in sub_model_zoo.items(): + for name, ( + model_fn, + data_gen_fn, + output_transform_fn, + loss_fn, + _, + ) in sub_model_zoo.items(): check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn, test_config) clear_layout_converter() @@ -194,13 +220,27 @@ def run_t5_3d_test(test_config): def check_t5(rank, world_size, port): disable_existing_loggers() - colossalai.launch(config={}, rank=rank, world_size=world_size, host="localhost", port=port, backend="nccl") + colossalai.launch( + config={}, + rank=rank, + world_size=world_size, + host="localhost", + port=port, + backend="nccl", + ) run_t5_test() def check_t5_3d(rank, world_size, port): disable_existing_loggers() - colossalai.launch(config={}, rank=rank, world_size=world_size, host="localhost", port=port, backend="nccl") + colossalai.launch( + config={}, + rank=rank, + world_size=world_size, + host="localhost", + port=port, + backend="nccl", + ) run_t5_3d_test() diff --git a/tests/test_tensor/test_dtensor/test_layout_converter.py b/tests/test_tensor/test_dtensor/test_layout_converter.py index 4e65401bf7b4..3bface1d286f 100644 --- a/tests/test_tensor/test_dtensor/test_layout_converter.py +++ b/tests/test_tensor/test_dtensor/test_layout_converter.py @@ -123,8 +123,15 @@ def check_layout_converting(rank, world_size, port): assert comm_action_sequence[2].logical_process_axis == 1 # checkout chached_spec_pairs_transform_path - assert layout_converter.cached_solution[("[R, S01, R]", "[S01, R, R]")][0] == transform_path - assert layout_converter.cached_solution[("[R, S01, R]", "[S01, R, R]")][1] == comm_action_sequence + src_shape = source_layout.get_sharded_shape_per_device() + dst_shape = target_layout.get_sharded_shape_per_device() + assert ( + layout_converter.cached_solution[(("[R, S01, R]", src_shape), ("[S01, R, R]", dst_shape))][0] == transform_path + ) + assert ( + layout_converter.cached_solution[(("[R, S01, R]", src_shape), ("[S01, R, R]", dst_shape))][1] + == comm_action_sequence + ) comm_cost = layout_converter.get_total_comm_cost(source_layout, target_layout) diff --git a/tests/test_utils/test_flash_attention.py b/tests/test_utils/test_flash_attention.py deleted file mode 100644 index 3ec1700045e3..000000000000 --- a/tests/test_utils/test_flash_attention.py +++ /dev/null @@ -1,167 +0,0 @@ -import math - -import pytest -import torch -from einops import rearrange - -from colossalai.kernel.extensions.flash_attention import HAS_FLASH_ATTN, HAS_MEM_EFF_ATTN -from colossalai.testing import clear_cache_before_run, parameterize - -if HAS_MEM_EFF_ATTN or HAS_FLASH_ATTN: - from colossalai.nn.layer.colo_attention import AttnMaskType, ColoAttention - -DTYPE = [torch.float16, torch.bfloat16, torch.float32] - - -def attention_ref(q, k, v, attn_mask=None, causal=False): - """ - attention output of the control group - """ - dtype_og = q.dtype - seqlen_q, seqlen_k = q.shape[1], k.shape[1] - d = q.shape[-1] - scale = 1.0 / math.sqrt(d) - scores = torch.einsum("bthd,bshd->bhts", q * scale, k) - - if attn_mask is not None: - scores.masked_fill_(rearrange(~attn_mask, "b s -> b 1 1 s"), float("-inf")) - if causal: - causal_mask = torch.triu(torch.ones(seqlen_q, seqlen_k, dtype=torch.bool, device=q.device), 1) - scores.masked_fill_(causal_mask, float("-inf")) - attention = torch.softmax(scores, dim=-1) - - output = torch.einsum("bhts,bshd->bthd", attention, v) - output = rearrange(output, "b s h d -> b s (h d)") - - # Modify the data at the positions of the mask to 0 - if attn_mask is not None: - output.masked_fill_(rearrange(~attn_mask, "b s -> b s 1"), 0.0) - - return output.to(dtype=dtype_og) - - -@pytest.mark.skipif(not HAS_MEM_EFF_ATTN and not HAS_FLASH_ATTN, reason="xformers is not available") -@clear_cache_before_run() -@parameterize("proj_shape", [(6, 8, 4, 16)]) -@parameterize("dtype", DTYPE) -@parameterize("dropout", [0.0]) -def test_attention_gpt(proj_shape, dtype, dropout): - (B, S, H, D_HEAD) = proj_shape - D = H * D_HEAD - - q = torch.randn((B, S, H, D_HEAD), dtype=dtype, device="cuda", requires_grad=True) - k = torch.randn((B, S, H, D_HEAD), dtype=dtype, device="cuda", requires_grad=True) - v = torch.randn((B, S, H, D_HEAD), dtype=dtype, device="cuda", requires_grad=True) - - mask = [torch.ones(S - i, dtype=torch.bool, device="cuda") for i in range(B)] - mask = torch.nn.utils.rnn.pad_sequence(mask, batch_first=True) - - attn = ColoAttention(D, H, dropout=dropout) - y = attn(q, k, v, attn_mask=mask, attn_mask_type=AttnMaskType.paddedcausal) - - assert list(y.shape) == [B, S, D] - - out_ref = attention_ref(q, k, v, mask, causal=True) - - # check gradients - dy = torch.rand_like(y) - grad_q, grad_k, grad_v = torch.autograd.grad(y, (q, k, v), dy) - grad_ref_q, grad_ref_k, grad_ref_v = torch.autograd.grad(out_ref, (q, k, v), dy) - - torch.allclose(y, out_ref, atol=1e-7), f"{(y - out_ref).abs().max()}" - torch.allclose(grad_q, grad_ref_q, atol=1e-7), f"{(grad_q - grad_ref_q).abs().max()}" - torch.allclose(grad_k, grad_ref_k, atol=1e-7), f"{(grad_k - grad_ref_k).abs().max()}" - torch.allclose(grad_v, grad_ref_v, atol=1e-7), f"{(grad_v - grad_ref_v).abs().max()}" - - -@pytest.mark.skipif(not HAS_MEM_EFF_ATTN and not HAS_FLASH_ATTN, reason="xformers is not available") -@clear_cache_before_run() -@parameterize("proj_shape", [(6, 8, 4, 16)]) -@parameterize("dtype", DTYPE) -@parameterize("dropout", [0.0]) -def test_attention_bert(proj_shape, dtype, dropout): - (B, S, H, D_HEAD) = proj_shape - D = H * D_HEAD - - q = torch.randn((B, S, H, D_HEAD), dtype=dtype, device="cuda", requires_grad=True) - k = torch.randn((B, S, H, D_HEAD), dtype=dtype, device="cuda", requires_grad=True) - v = torch.randn((B, S, H, D_HEAD), dtype=dtype, device="cuda", requires_grad=True) - - # attention mask of shape [B, S] with zero padding to max length S - mask = torch.randint(0, 2, (B, S), dtype=torch.bool, device="cuda") - - attn = ColoAttention(D, H, dropout=dropout) - y = attn(q, k, v, attn_mask=mask, attn_mask_type=AttnMaskType.padding) - - assert list(y.shape) == [B, S, D] - - out_ref = attention_ref(q, k, v, mask, causal=False) - - dy = torch.rand_like(y) - grad_q, grad_k, grad_v = torch.autograd.grad(y, (q, k, v), dy) - grad_ref_q, grad_ref_k, grad_ref_v = torch.autograd.grad(out_ref, (q, k, v), dy) - - torch.allclose(y, out_ref, atol=1e-7), f"{(y - out_ref).abs().max()}" - torch.allclose(grad_q, grad_ref_q, atol=1e-7), f"{(grad_q - grad_ref_q).abs().max()}" - torch.allclose(grad_k, grad_ref_k, atol=1e-7), f"{(grad_k - grad_ref_k).abs().max()}" - torch.allclose(grad_v, grad_ref_v, atol=1e-7), f"{(grad_v - grad_ref_v).abs().max()}" - - -@pytest.mark.skipif(not HAS_MEM_EFF_ATTN and not HAS_FLASH_ATTN, reason="xformers is not available") -@clear_cache_before_run() -@parameterize("proj_shape", [(6, 8, 4, 16)]) -@parameterize("dtype", DTYPE) -@parameterize("dropout", [0.0]) -def test_attention_no_mask(proj_shape, dtype, dropout): - (B, S, H, D_HEAD) = proj_shape - D = H * D_HEAD - - q = torch.randn((B, S, H, D_HEAD), dtype=dtype, device="cuda", requires_grad=True) - k = torch.randn((B, S, H, D_HEAD), dtype=dtype, device="cuda", requires_grad=True) - v = torch.randn((B, S, H, D_HEAD), dtype=dtype, device="cuda", requires_grad=True) - - attn = ColoAttention(D, H, dropout=dropout) - y = attn(q, k, v) - - assert list(y.shape) == [B, S, D] - - out_ref = attention_ref(q, k, v, None, causal=False) - - dy = torch.rand_like(y) - grad_q, grad_k, grad_v = torch.autograd.grad(y, (q, k, v), dy) - grad_ref_q, grad_ref_k, grad_ref_v = torch.autograd.grad(out_ref, (q, k, v), dy) - - torch.allclose(y, out_ref, atol=1e-7), f"{(y - out_ref).abs().max()}" - torch.allclose(grad_q, grad_ref_q, atol=1e-7), f"{(grad_q - grad_ref_q).abs().max()}" - torch.allclose(grad_k, grad_ref_k, atol=1e-7), f"{(grad_k - grad_ref_k).abs().max()}" - torch.allclose(grad_v, grad_ref_v, atol=1e-7), f"{(grad_v - grad_ref_v).abs().max()}" - - -@pytest.mark.skipif(not HAS_MEM_EFF_ATTN and not HAS_FLASH_ATTN, reason="xformers is not available") -@clear_cache_before_run() -@parameterize("proj_shape", [(6, 24, 8, 4, 16)]) -@parameterize("dtype", DTYPE) -@parameterize("dropout", [0.0]) -def test_cross_attention(proj_shape, dtype, dropout): - (B, S, T, H, D_HEAD) = proj_shape - D = H * D_HEAD - - q = torch.randn((B, T, H, D_HEAD), dtype=dtype, device="cuda", requires_grad=True) - k = torch.randn((B, S, H, D_HEAD), dtype=dtype, device="cuda", requires_grad=True) - v = torch.randn((B, S, H, D_HEAD), dtype=dtype, device="cuda", requires_grad=True) - - attn = ColoAttention(D, H, dropout=dropout) - y = attn(q, k, v, attn_mask_type=AttnMaskType.causal) - - assert list(y.shape) == [B, T, D] - - out_ref = attention_ref(q, k, v, None, causal=True) - - dy = torch.rand_like(y) - grad_q, grad_k, grad_v = torch.autograd.grad(y, (q, k, v), dy) - grad_ref_q, grad_ref_k, grad_ref_v = torch.autograd.grad(out_ref, (q, k, v), dy) - - torch.allclose(y, out_ref, atol=1e-18), f"{(y - out_ref).abs().max()}" - torch.allclose(grad_q, grad_ref_q, atol=1e-7), f"{(grad_q - grad_ref_q).abs().max()}" - torch.allclose(grad_k, grad_ref_k, atol=1e-7), f"{(grad_k - grad_ref_k).abs().max()}" - torch.allclose(grad_v, grad_ref_v, atol=1e-7), f"{(grad_v - grad_ref_v).abs().max()}" From 6911f61445f0401cac1c88d106f0d33626c80d86 Mon Sep 17 00:00:00 2001 From: flybird11111 <1829166702@qq.com> Date: Wed, 17 Apr 2024 15:24:03 +0800 Subject: [PATCH 4/7] [shardformer]enable padding vocabulary size. (#5489) * padding vocab_size when using pipeline parallellism padding vocab_size when using pipeline parallellism fix fix * fix * fix fix fix * fix gather output * fix * fix * fix fix resize embedding fix resize embedding * fix resize embedding fix * revert * revert * revert * padding vocab * padding vocabe * fix * fix * fxi * test ci * fix fix fix fix * fix fix * fix * fix * Update hybrid_parallel_plugin.py fix fix fix * fix fix * fix fix * fix * resolve super init resolve super init resolve super init resolve super init * resolve comments * fix * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * vocab checkpointio * padding vocab_size when using pipeline parallellism padding vocab_size when using pipeline parallellism fix fix * fix fix fix * fix * fix fix resize embedding fix resize embedding * fix resize embedding fix * revert * revert * padding vocab * fix * fix fix * fix fix * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * fix ci * fix * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * fix * cherry-pick * revert moe modify * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * fix fix fix fix fix fix fix fix * resolve comments resolve comments resolve comments resolve comments resolve comments * ptensor ptensor resolve comments fix fix fix fix fix resolve comments resolve comments resolve comments resolve comments resolve comments --------- Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com> Co-authored-by: Hongxin Liu --- colossalai/booster/plugin/gemini_plugin.py | 6 +- .../booster/plugin/hybrid_parallel_plugin.py | 11 +- .../hybrid_parallel_checkpoint_io.py | 29 ++- colossalai/checkpoint_io/utils.py | 9 + colossalai/shardformer/layer/__init__.py | 7 +- colossalai/shardformer/layer/embedding.py | 111 +++++++-- colossalai/shardformer/layer/linear.py | 220 +++++++++++++++++- colossalai/shardformer/layer/loss.py | 32 ++- .../shardformer/layer/parallel_module.py | 192 ++++++++++++++- colossalai/shardformer/modeling/gpt2.py | 13 +- colossalai/shardformer/modeling/llama.py | 11 +- .../shardformer/policies/base_policy.py | 9 + colossalai/shardformer/policies/bert.py | 52 +++-- colossalai/shardformer/policies/blip2.py | 62 +++-- colossalai/shardformer/policies/bloom.py | 47 ++-- colossalai/shardformer/policies/chatglm2.py | 40 ++-- colossalai/shardformer/policies/falcon.py | 49 ++-- colossalai/shardformer/policies/gpt2.py | 71 ++++-- colossalai/shardformer/policies/gptj.py | 55 +++-- colossalai/shardformer/policies/llama.py | 53 +++-- colossalai/shardformer/policies/mistral.py | 57 +++-- colossalai/shardformer/policies/opt.py | 64 +++-- colossalai/shardformer/policies/t5.py | 101 ++++++-- colossalai/shardformer/policies/whisper.py | 45 ++-- colossalai/shardformer/shard/shard_config.py | 3 +- .../tensor/d_tensor/layout_converter.py | 17 +- colossalai/tensor/padded_tensor/__init__.py | 3 + colossalai/tensor/padded_tensor/api.py | 128 ++++++++++ colossalai/testing/comparison.py | 2 +- colossalai/zero/gemini/gemini_ddp.py | 22 ++ colossalai/zero/gemini/gemini_optimizer.py | 45 +++- ...st_hybrid_parallel_plugin_checkpoint_io.py | 3 +- tests/test_optimizer/test_nvme.py | 3 - .../test_vocab_parallel_embedding_1d.py | 2 +- tests/test_shardformer/test_model/_utils.py | 11 +- .../test_model/test_shard_t5.py | 3 +- tests/test_tensor/test_padded_tensor.py | 46 ++++ 37 files changed, 1347 insertions(+), 287 deletions(-) create mode 100644 colossalai/tensor/padded_tensor/__init__.py create mode 100644 colossalai/tensor/padded_tensor/api.py create mode 100644 tests/test_tensor/test_padded_tensor.py diff --git a/colossalai/booster/plugin/gemini_plugin.py b/colossalai/booster/plugin/gemini_plugin.py index 6c503377326a..442ac4a8da06 100644 --- a/colossalai/booster/plugin/gemini_plugin.py +++ b/colossalai/booster/plugin/gemini_plugin.py @@ -44,10 +44,10 @@ def get_param_info(optim: Optimizer): # Get a backup of necessary information of parameters for future use, which includes: # 1. A mapping from integer param_id to param32 shape. - if optim is None: return {} param_info = {"id2shape": {}} + start_index = 0 for group in optim.param_groups: for param_id, param in enumerate(group["params"], start_index): @@ -527,7 +527,7 @@ def configure( dataloader: Optional[DataLoader] = None, lr_scheduler: Optional[LRScheduler] = None, ) -> Tuple[nn.Module, OptimizerWrapper, Callable, DataLoader, LRScheduler]: - optimizer_params_info = get_param_info(optimizer) + params_info = get_param_info(optimizer) if not isinstance(model, ModelWrapper): # convert model to sync bn # FIXME(ver217): gemini does not support sync bn @@ -558,7 +558,7 @@ def configure( **self.zero_optim_config, **self.optim_kwargs, tp_group=self.tp_group, - optimizer_params_info=optimizer_params_info, + params_info=params_info, verbose=self.verbose, ) diff --git a/colossalai/booster/plugin/hybrid_parallel_plugin.py b/colossalai/booster/plugin/hybrid_parallel_plugin.py index 29cec7cfd146..8d12eb80621d 100644 --- a/colossalai/booster/plugin/hybrid_parallel_plugin.py +++ b/colossalai/booster/plugin/hybrid_parallel_plugin.py @@ -213,12 +213,7 @@ def get_param_info(optim: Optimizer): if optim is None: return {} - param_info = { - "param_groups": [], - "param2id": {}, - "id2param": {}, - "param2shape": {}, - } + param_info = {"param_groups": [], "param2id": {}, "id2param": {}, "param2shape": {}} start_index = 0 for group in optim.param_groups: packed_group = {k: v for k, v in group.items() if k != "params"} @@ -947,6 +942,8 @@ class HybridParallelPlugin(PipelinePluginBase): num_model_chunks (int, optional): The number of model chunks for interleaved pipeline parallelism. Defaults to 1. gradient_checkpoint_config (GradientCheckpointConfig, optional): Configuration for gradient checkpointing. Defaults to None. enable_metadata_cache (bool, optional): Whether to enable metadata cache for pipeline parallelism. Defaults to True. + make_vocab_size_divisible_by (int, optional): it's used when padding the vocabulary size, to make it choose an faster kenel. Default to 64. + """ def __init__( @@ -989,6 +986,7 @@ def __init__( num_model_chunks: int = 1, gradient_checkpoint_config: Optional[GradientCheckpointConfig] = None, enable_metadata_cache: bool = True, + make_vocab_size_divisible_by: int = 64, ) -> None: super().__init__() assert ( @@ -1095,6 +1093,7 @@ def __init__( sequence_parallelism_mode=sequence_parallelism_mode, enable_sequence_overlap=enable_sequence_overlap, parallel_output=parallel_output, + make_vocab_size_divisible_by=make_vocab_size_divisible_by, gradient_checkpoint_config=gradient_checkpoint_config, ) self.amp_config = dict( diff --git a/colossalai/checkpoint_io/hybrid_parallel_checkpoint_io.py b/colossalai/checkpoint_io/hybrid_parallel_checkpoint_io.py index 80822724982e..7946d9b9c197 100644 --- a/colossalai/checkpoint_io/hybrid_parallel_checkpoint_io.py +++ b/colossalai/checkpoint_io/hybrid_parallel_checkpoint_io.py @@ -14,6 +14,12 @@ from colossalai.cluster import DistCoordinator from colossalai.interface import ModelWrapper, OptimizerWrapper +from colossalai.tensor.padded_tensor import ( + init_as_padded_tensor, + is_padded_tensor, + to_padded_tensor, + to_unpadded_tensor, +) from colossalai.utils import get_current_device from .general_checkpoint_io import GeneralCheckpointIO @@ -32,6 +38,7 @@ save_param_groups, save_state_dict, save_state_dict_shards, + search_padding_dim, search_tp_partition_dim, sharded_optimizer_loading_epilogue, ) @@ -89,6 +96,8 @@ def _model_sharder( if param is None: continue # Gather tensor pieces when using tensor parallel. + if is_padded_tensor(param): + param = to_unpadded_tensor(param) param_ = gather_distributed_param(param, keep_vars=False) block, block_size = state_dict_sharder.append_param(prefix + name, param_) if block is not None: @@ -231,7 +240,6 @@ def save_sharded_model( # When pipeline is used, each stage produces its own shard files and index files. # Index files belonging to each stage are saved under a temporary folder ./tmp_index_files/ # After all the state_dicts have been saved, the master rank integrates all the index files into one final index file and deletes the tmp folder. - final_index_file_path = copy.deepcopy(save_index_file) tmp_index_file_folder = os.path.join(checkpoint, "tmp_index_files") Path(tmp_index_file_folder).mkdir(parents=True, exist_ok=True) @@ -251,6 +259,7 @@ def save_sharded_model( use_safetensors=use_safetensors, use_pp_format=True, ) + if control_saving: assert ( self.dp_rank == 0 and self.tp_rank == 0 @@ -867,6 +876,11 @@ def gather_from_sharded_optimizer_state( dist.all_gather(gather_tensor, v, group=tp_group) v = torch.cat(gather_tensor, dim=partition_dim) + padding_dim = search_padding_dim(v.shape, original_shape) + if padding_dim is not None: + v = init_as_padded_tensor(v, v.shape[padding_dim], original_shape[padding_dim], padding_dim) + v = to_unpadded_tensor(v) + state_[k] = v.detach().clone().to(device) return state_ @@ -899,6 +913,19 @@ def shard_from_complete_optimizer_state( if isinstance(v, torch.Tensor) and k != "step": # Shard state along tensor parallel group. partition_dim = search_tp_partition_dim(current_shape, original_shape, self.tp_size) + global_shape = current_shape + if partition_dim is not None: + # pad embedding params + global_shape = ( + *current_shape[:partition_dim], + current_shape[partition_dim] * self.tp_size, + *current_shape[partition_dim + 1 :], + ) + + padding_dim = search_padding_dim(global_shape, original_shape) + if padding_dim is not None: + v = to_padded_tensor(v, global_shape[padding_dim], padding_dim) + if partition_dim is not None: slice_size = current_shape[partition_dim] v = v.split(slice_size, dim=partition_dim)[self.tp_rank] diff --git a/colossalai/checkpoint_io/utils.py b/colossalai/checkpoint_io/utils.py index 2a1d4de9b036..6197be9d1c8d 100644 --- a/colossalai/checkpoint_io/utils.py +++ b/colossalai/checkpoint_io/utils.py @@ -120,6 +120,15 @@ def search_tp_partition_dim(current_shape: torch.Size, original_shape: torch.Siz return partition_dim +def search_padding_dim(global_shape: torch.Size, original_shape: torch.Size) -> Optional[int]: + padding_dim = None + for dim, length in enumerate(global_shape): + if length > original_shape[dim]: + padding_dim = dim + break + return padding_dim + + # ====================================== # Helper classes and functions for saving shard file # ====================================== diff --git a/colossalai/shardformer/layer/__init__.py b/colossalai/shardformer/layer/__init__.py index 7b8aa53800f0..f17fad1b6606 100644 --- a/colossalai/shardformer/layer/__init__.py +++ b/colossalai/shardformer/layer/__init__.py @@ -1,8 +1,8 @@ from ._operation import all_to_all_comm from .attn import AttnMaskType, ColoAttention from .dropout import DropoutForParallelInput, DropoutForReplicatedInput -from .embedding import Embedding1D, VocabParallelEmbedding1D -from .linear import Linear1D_Col, Linear1D_Row +from .embedding import Embedding1D, PaddingEmbedding, VocabParallelEmbedding1D +from .linear import Linear1D_Col, Linear1D_Row, PaddingLMHead, VocabParallelLMHead1D from .loss import cross_entropy_1d from .normalization import FusedLayerNorm, FusedRMSNorm, LayerNorm, RMSNorm from .parallel_module import ParallelModule @@ -25,6 +25,9 @@ "FusedRMSNorm", "FusedLinear1D_Col", "ParallelModule", + "PaddingEmbedding", + "PaddingLMHead", + "VocabParallelLMHead1D", "AttnMaskType", "ColoAttention", "all_to_all_comm", diff --git a/colossalai/shardformer/layer/embedding.py b/colossalai/shardformer/layer/embedding.py index d081b204093b..cb7eceae4d25 100644 --- a/colossalai/shardformer/layer/embedding.py +++ b/colossalai/shardformer/layer/embedding.py @@ -21,10 +21,10 @@ ) from ._operation import gather_forward_split_backward, reduce_forward -from .parallel_module import ParallelModule +from .parallel_module import PaddingParallelModule, ParallelModule from .utils import create_randomizer_with_offset -__all__ = ["Embedding1D", "VocabParallelEmbedding1D"] +__all__ = ["Embedding1D", "VocabParallelEmbedding1D", "PaddingEmbedding"] class Embedding1D(ParallelModule): @@ -161,7 +161,80 @@ def forward(self, input_: Tensor) -> Tensor: return output_parallel -class VocabParallelEmbedding1D(ParallelModule): +class PaddingEmbedding(PaddingParallelModule): + def __init__( + self, + num_embeddings: int, + embedding_dim: int, + padding_idx: int = None, + dtype: torch.dtype = None, + device: torch.device = None, + weight: Optional[nn.Parameter] = None, + make_vocab_size_divisible_by: int = 64, + *args, + **kwargs, + ): + self.num_embeddings = num_embeddings + self.embedding_dim = embedding_dim + self.embed_args = args + self.embed_kwargs = kwargs + self.padding_idx = padding_idx + if num_embeddings % make_vocab_size_divisible_by != 0: + self.num_embeddings = ( + num_embeddings + make_vocab_size_divisible_by - (num_embeddings % make_vocab_size_divisible_by) + ) + # create weight and bias + if weight is None: + factory_kwargs = {"device": device, "dtype": dtype} + weight = nn.Parameter(torch.empty((num_embeddings, self.embedding_dim), **factory_kwargs)) + else: + weight.data = weight.data.to(device=device, dtype=dtype) + + super().__init__(self.num_embeddings, num_embeddings, weight) + + if weight is None: + self.reset_parameters() + + def reset_parameters(self) -> None: + init.normal_(self.weight) + self._fill_padding_idx_with_zero() + + def _fill_padding_idx_with_zero(self) -> None: + if self.padding_idx is not None: + with torch.no_grad(): + self.weight[self.padding_idx].fill_(0) + + def forward(self, input: Tensor) -> Tensor: + return F.embedding(input, self.weight, self.padding_idx, *self.embed_args, **self.embed_kwargs) + + @staticmethod + def from_native_module( + module: nn.Embedding, process_group: Union[ProcessGroup, List[ProcessGroup]], *args, **kwargs + ) -> PaddingParallelModule: + r""" + Convert a native pytorch embedding module to a parallel module. + """ + LazyInitContext.materialize(module) + # get the origin attributes + num_embeddings = module.num_embeddings + embedding_dim = module.embedding_dim + padding_idx = module.padding_idx + device = module.weight.device + # create the parallel module + padding_embedding = PaddingEmbedding( + num_embeddings=num_embeddings, + embedding_dim=embedding_dim, + padding_idx=padding_idx, + device=device, + weight=module.weight, + *args, + **kwargs, + ) + + return padding_embedding + + +class VocabParallelEmbedding1D(PaddingParallelModule): r"""Embedding parallelized in the vocabulary dimension. Args: @@ -201,10 +274,10 @@ def __init__( process_group: ProcessGroup = None, weight: Optional[nn.Parameter] = None, weight_initializer: Callable = init.normal_(), + make_vocab_size_divisible_by: int = 64, *args, **kwargs, ): - super().__init__() self.num_embeddings = num_embeddings self.embedding_dim = embedding_dim self.embed_args = args @@ -214,8 +287,23 @@ def __init__( tensor_parallel_size = dist.get_world_size(group=process_group) tensor_parallel_rank = dist.get_rank(group=process_group) - self.num_embeddings_per_partition = divide(num_embeddings, tensor_parallel_size) - self.num_embeddings = self.num_embeddings_per_partition + # generate weight and bias + if weight is None: + factory_kwargs = {"device": device, "dtype": dtype} + weight = nn.Parameter(torch.empty((num_embeddings, self.embedding_dim), **factory_kwargs)) + else: + weight.data = weight.data.to(device=device, dtype=dtype) + + # calculate new padding size + multiple = make_vocab_size_divisible_by * tensor_parallel_size + if num_embeddings % multiple != 0: + self.num_embeddings = num_embeddings + multiple - (num_embeddings % multiple) + + # resize vocabulary size + super().__init__(self.num_embeddings, num_embeddings, weight) + + # deal with tensor parallelism + self.num_embeddings_per_partition = divide(self.num_embeddings, tensor_parallel_size) self.vocab_start_index = tensor_parallel_rank * self.num_embeddings_per_partition self.vocab_end_index = self.vocab_start_index + self.num_embeddings_per_partition @@ -226,13 +314,6 @@ def __init__( seed = torch.random.initial_seed() self.randomizer = create_randomizer_with_offset(seed, process_group=self.process_group) - # parameter - if weight is None: - factory_kwargs = {"device": device, "dtype": dtype} - self.weight = nn.Parameter(torch.empty((num_embeddings, self.embedding_dim), **factory_kwargs)) - else: - weight.data = weight.data.to(device=device, dtype=dtype) - self.weight = weight if not is_distributed_tensor(self.weight): sharded_weight = shard_rowwise(self.weight.data, process_group) sharded_tensor_to_existing_param(sharded_weight, self.weight) @@ -243,7 +324,7 @@ def __init__( @staticmethod def from_native_module( module: nn.Embedding, process_group: Union[ProcessGroup, List[ProcessGroup]], *args, **kwargs - ) -> ParallelModule: + ) -> PaddingParallelModule: r""" Convert a native pytorch embedding module to a parallel module. """ @@ -303,11 +384,9 @@ def forward(self, input_: Tensor) -> Tensor: # Mask the input. masked_input = input_.clone() - self.vocab_start_index masked_input[input_mask] = 0 - output_parallel = F.embedding( masked_input, self.weight, self.padding_idx, *self.embed_args, **self.embed_kwargs ) - # Mask the output embedding. embedding_output = output_parallel.clone() embedding_output[input_mask, :] = 0.0 diff --git a/colossalai/shardformer/layer/linear.py b/colossalai/shardformer/layer/linear.py index 7c8619ad8f5c..37c7542416f6 100644 --- a/colossalai/shardformer/layer/linear.py +++ b/colossalai/shardformer/layer/linear.py @@ -32,7 +32,7 @@ reducescatter_forward_gather_backward, split_forward_gather_backward, ) -from .parallel_module import ParallelModule +from .parallel_module import PaddingParallelModule, ParallelModule from .utils import create_randomizer_with_offset __all__ = ["Linear1D_Col", "Linear1D_Row"] @@ -84,8 +84,9 @@ def __init__( bias_: Optional[Parameter] = None, weight_initializer: Callable = init.kaiming_uniform_(a=math.sqrt(5)), bias_initializer: Callable = init.xavier_uniform_(a=1, scale=1), + **kwargs, ): - super().__init__() + super().__init__(weight=weight, bias_=bias_, **kwargs) # Keep input parameters self.in_features = in_features @@ -118,6 +119,7 @@ def __init__( else: weight.data = weight.data.to(device=device, dtype=dtype) self.weight = weight + if not is_distributed_tensor(self.weight): sharded_weight = shard_rowwise(self.weight.data, self.process_group) sharded_tensor_to_existing_param(sharded_weight, self.weight) @@ -140,7 +142,7 @@ def __init__( @staticmethod def from_native_module( - module: nn.Linear, process_group: Union[ProcessGroup, List[ProcessGroup]], *args, **kwargs + module: nn.Linear, process_group: Union[ProcessGroup, List[ProcessGroup]], **kwargs ) -> ParallelModule: r""" Convert a native PyTorch linear layer to a parallelized linear layer. @@ -173,7 +175,6 @@ def from_native_module( process_group=process_group, weight=module.weight, bias_=module.bias, - *args, **kwargs, ) @@ -322,7 +323,7 @@ def __init__( @staticmethod def from_native_module( - module: nn.Linear, process_group: Union[ProcessGroup, List[ProcessGroup]], *args, **kwargs + module: nn.Linear, process_group: Union[ProcessGroup, List[ProcessGroup]], **kwargs ) -> ParallelModule: r""" Convert a native PyTorch linear layer to a parallelized linear layer. @@ -356,7 +357,6 @@ def from_native_module( process_group=process_group, weight=module.weight, bias_=module.bias, - *args, **kwargs, ) @@ -439,3 +439,211 @@ def forward(self, input_: Tensor) -> Tensor: return output else: return output, self.bias + + +class PaddingLMHead(PaddingParallelModule): + def __init__( + self, + in_features: int, + out_features: int, + bias: bool = True, + dtype: torch.dtype = None, + device: torch.device = None, + weight: Optional[Parameter] = None, + bias_: Optional[Parameter] = None, + make_vocab_size_divisible_by: int = 64, + weight_initializer: Callable = init.kaiming_uniform_(a=math.sqrt(5)), + bias_initializer: Callable = init.xavier_uniform_(a=1, scale=1), + ): + # Keep input parameters + self.in_features = in_features + self.out_features = out_features + + if out_features % make_vocab_size_divisible_by != 0: + self.out_features = ( + out_features + make_vocab_size_divisible_by - (out_features % make_vocab_size_divisible_by) + ) + if weight is None: + factory_kwargs = {"device": device, "dtype": dtype} + weight = Parameter(torch.empty(out_features, self.in_features, **factory_kwargs)) + else: + weight.data = weight.data.to(device=device, dtype=dtype) + + if bias: + if bias_ is None: + self.bias = Parameter(torch.empty(out_features, **factory_kwargs)) + else: + bias_.data = bias_.data.to(device=device, dtype=dtype) + else: + bias_ = None + + # resize embeddings + super().__init__(self.out_features, out_features, weight, bias_) + + if weight is None: + self.reset_parameters(weight_initializer, bias_initializer) + + def reset_parameters(self, weight_initializer, bias_initializer) -> None: + fan_in, fan_out = self.in_features, self.out_features + weight_initializer(self.weight, fan_in=fan_in, fan_out=fan_out) + if self.bias is not None: + bias_initializer(self.bias, fan_in=fan_in) + + @staticmethod + def from_native_module( + module: nn.Linear, process_group: Union[ProcessGroup, List[ProcessGroup]], **kwargs + ) -> PaddingParallelModule: + r""" + Convert a native PyTorch linear layer to a parallelized linear layer. + """ + LazyInitContext.materialize(module) + # get the attributes + in_features = module.in_features + out_features = module.out_features + bias = module.bias is not None + device = module.weight.device + # ensure only one process group is passed + + lm_head_linear = PaddingLMHead( + in_features=in_features, + out_features=out_features, + bias=bias, + device=device, + weight=module.weight, + bias_=module.bias, + **kwargs, + ) + + return lm_head_linear + + def forward(self, input: Tensor) -> Tensor: + output = F.linear(input, self.weight, self.bias) + output = output[..., : self.old_num_embeddings] + return output + + +class VocabParallelLMHead1D(Linear1D_Col, PaddingParallelModule): + r"""Linear layer with column parallelism. + + The linear layer is defined as :math:`Y = XA + b`. A is parallelized along + its second dimension as :math:`A = [A_1, ..., A_p]`. + + Args: + in_features (int): size of each input sample. + out_features (int): size of each output sample. + bias (bool, optional): If set to ``False``, the layer will not learn an additive bias, defaults to ``True``. + dtype (`torch.dtype`): The dtype of parameters, defaults to None. + device (`torch.device`): The device of parameters, defaults to None. + process_group (`torch.distributed.ProcessGroup`): The process group to be used for weight sharding and communication, defaults to None. + gather_output (bool, optional): If true, call all-gather on output and make Y available + to all GPUs, otherwise, every GPU will have its output + which is :math:`Y_i = XA_i`, defaults to False + seq_parallel (`bool`): If set to ``True``, it will use sequence parallel, defaults to False. + overlap (`bool`): If set to ``True``, it will overlap input all-gather with gradient computation during backward, defaults to False. + skip_bias_add (bool): If set to ``True``, it will skip bias add for linear layer, + which is preserved for kernel fusion, defaults to False + weight_initializer (`typing.Callable`): + The initializer of weight, defaults to kaiming uniform initializer. + bias_initializer (`typing.Callable`): + The initializer of bias, defaults to xavier uniform initializer. + + More details about ``initializer`` please refer to + `init `_. + """ + + def __init__( + self, + in_features: int, + out_features: int, + bias: bool = True, + dtype: torch.dtype = None, + device: torch.device = None, + process_group: ProcessGroup = None, + weight: Optional[Parameter] = None, + bias_: Optional[Parameter] = None, + make_vocab_size_divisible_by: int = 64, + **kwargs, + ): + # create weight and bias + if weight is None: + factory_kwargs = {"device": device, "dtype": dtype} + weight = Parameter(torch.empty(out_features, self.in_features, **factory_kwargs)) + if bias: + if bias_ is None: + bias_ = Parameter(torch.empty(out_features, **factory_kwargs)) + else: + bias_ = None + + # calculate new vocab size + self.tensor_parallel_size = dist.get_world_size(group=process_group) + new_out_features = out_features + multiple = make_vocab_size_divisible_by * self.tensor_parallel_size + if out_features % multiple != 0: + new_out_features = out_features + multiple - (out_features % multiple) + + super().__init__( + in_features=in_features, + out_features=new_out_features, + bias=bias, + device=device, + process_group=process_group, + weight=weight, + bias_=bias_, + **kwargs, + new_num_embeddings=new_out_features, + old_num_embeddings=out_features, + ) + # get the length of valid embeddings + tp_rank = dist.get_rank(process_group) + partition_size = self.new_num_embeddings // dist.get_world_size(process_group) + if self.old_num_embeddings >= (tp_rank + 1) * partition_size: + self.num_valid_embeddings_local = partition_size + elif self.old_num_embeddings >= tp_rank * partition_size: + self.num_valid_embeddings_local = self.old_num_embeddings - tp_rank * partition_size + else: + self.num_valid_embeddings_local = 0 + + @staticmethod + def from_native_module( + module: nn.Linear, process_group: Union[ProcessGroup, List[ProcessGroup]], **kwargs + ) -> PaddingParallelModule: + r""" + Convert a native PyTorch linear layer to a parallelized linear layer. + """ + LazyInitContext.materialize(module) + # get the attributes + in_features = module.in_features + out_features = module.out_features + bias = module.bias is not None + device = module.weight.device + + lm_head_linear = VocabParallelLMHead1D( + in_features=in_features, + out_features=out_features, + bias=bias, + device=device, + process_group=process_group, + weight=module.weight, + bias_=module.bias, + **kwargs, + ) + + return lm_head_linear + + def forward(self, input_: Tensor) -> Tuple[Tensor, Tensor]: + # get forward output + if self.skip_bias_add: + output, bias = super().forward(input_) + else: + output = super().forward(input_) + + # delete the padding of output + if self.gather_output: + output = output[..., : self.old_num_embeddings] + else: + output = output[..., : self.num_valid_embeddings_local] + + # return + if self.skip_bias_add: + return output, bias + return output diff --git a/colossalai/shardformer/layer/loss.py b/colossalai/shardformer/layer/loss.py index c4cf3fb8517c..6d99efc19bbf 100644 --- a/colossalai/shardformer/layer/loss.py +++ b/colossalai/shardformer/layer/loss.py @@ -15,7 +15,14 @@ class DistCrossEntropy(Function): """ @staticmethod - def forward(ctx, vocab_logits: torch.Tensor, target: torch.Tensor, ignore_index: int, process_group: ProcessGroup): + def forward( + ctx, + vocab_logits: torch.Tensor, + target: torch.Tensor, + ignore_index: int, + process_group: ProcessGroup, + vocab_size: int, + ): r""" Calculate the cross entropy loss before gather, the origin loss function is as follows: loss = -log(exp(x[class])/sum(exp(x[i])) @@ -41,15 +48,21 @@ def forward(ctx, vocab_logits: torch.Tensor, target: torch.Tensor, ignore_index: vocab_logits = vocab_logits - logits_max.unsqueeze(dim=-1) # mask the target in the local device - partition_vocab_size = vocab_logits.size()[-1] rank = dist.get_rank(group=process_group) world_size = dist.get_world_size(group=process_group) - global_vocab_size = partition_vocab_size * world_size + if vocab_size == None: + partition_vocab_size = vocab_logits.size()[-1] + global_vocab_size = partition_vocab_size * world_size + else: + global_vocab_size = vocab_size + partition_vocab_size = global_vocab_size // world_size # [down, up) => false, other device and -100 => true delta = (global_vocab_size + world_size - 1) // world_size down_threshold = rank * delta up_threshold = down_threshold + delta + if up_threshold > global_vocab_size: + up_threshold = global_vocab_size mask = (target < down_threshold) | (target >= up_threshold) masked_target = target.clone() - down_threshold masked_target[mask] = 0 @@ -57,7 +70,8 @@ def forward(ctx, vocab_logits: torch.Tensor, target: torch.Tensor, ignore_index: # reshape the logits and target # reshape the vocab_logits to [bath_size * seq_len, vocab_size] # reshape the labels to [bath_size * seq_len] - logits_2d = vocab_logits.view(-1, partition_vocab_size) + self_vocab_size = vocab_logits.size()[-1] + logits_2d = vocab_logits.view(-1, self_vocab_size) masked_target_1d = masked_target.view(-1) # extract the x[class] and set the x[other device] to zero @@ -104,10 +118,14 @@ def backward(ctx, grad_output): grad_logits_2d[torch.arange(0, grad_logits_2d.shape[0]), masked_target_1d] -= update grad_logits.mul_(grad_output.unsqueeze(dim=-1)) - return grad_logits, None, None, None + return grad_logits, None, None, None, None def cross_entropy_1d( - vocab_logits: torch.Tensor, labels: torch.Tensor, ignore_index: int = -100, process_group: ProcessGroup = None + vocab_logits: torch.Tensor, + labels: torch.Tensor, + ignore_index: int = -100, + process_group: ProcessGroup = None, + vocab_size: int = None, ) -> torch.Tensor: - return DistCrossEntropy.apply(vocab_logits, labels, ignore_index, process_group) + return DistCrossEntropy.apply(vocab_logits, labels, ignore_index, process_group, vocab_size) diff --git a/colossalai/shardformer/layer/parallel_module.py b/colossalai/shardformer/layer/parallel_module.py index 6c0d83cc7a20..11ef73538c36 100644 --- a/colossalai/shardformer/layer/parallel_module.py +++ b/colossalai/shardformer/layer/parallel_module.py @@ -3,7 +3,7 @@ import itertools from abc import ABC, abstractmethod -from typing import List, Union +from typing import List, Optional, Union import torch import torch.nn as nn @@ -20,11 +20,15 @@ is_distributed_tensor, sharded_tensor_to_param, ) +from colossalai.tensor.padded_tensor import is_padded_tensor, to_padded_tensor, to_unpadded_tensor __all__ = ["ParallelModule"] class ParallelModule(nn.Module, ABC): + def __init__(self, **kwargs): + super().__init__() + @abstractmethod def from_native_module( module: nn.Module, process_group: Union[ProcessGroup, List[ProcessGroup]] = None @@ -54,7 +58,7 @@ def _save_to_state_dict(self, destination, prefix, keep_vars): """ for name, param in self._parameters.items(): if param is not None: - destination[prefix + name] = gather_distributed_param(param, keep_vars=keep_vars) + destination[prefix + name] = gather_distributed_param(param, keep_vars=keep_vars).data for name, buf in self._buffers.items(): if buf is not None and name not in self._non_persistent_buffers_set: @@ -171,3 +175,187 @@ def _load_from_state_dict( input_name = input_name.split(".", 1)[0] # get the name of param/buffer/child if input_name not in self._modules and input_name not in local_state: unexpected_keys.append(key) + + +class PaddingParallelModule(ParallelModule): + def __init__( + self, + new_num_embeddings: int, + old_num_embeddings: int, + weight: Optional[nn.Parameter], + bias_: Optional[nn.Parameter] = None, + **kwargs, + ) -> None: + super().__init__(**kwargs) + self.new_num_embeddings = new_num_embeddings + self.old_num_embeddings = old_num_embeddings + self.weight = weight + self.bias = bias_ + + if not (is_distributed_tensor(self.weight) or self.weight.shape[0] == self.new_num_embeddings): + self.resize_embedding_weight() + + if self.bias is not None and not ( + is_distributed_tensor(self.bias) or self.bias.shape[0] == self.new_num_embeddings + ): + self.resize_embedding_bias() + + @abstractmethod + def from_native_module( + module: nn.Module, process_group: Union[ProcessGroup, List[ProcessGroup]] = None + ) -> "PaddingParallelModule": + """ + Convert a native PyTorch module to a parallelized module. + + Args: + module (nn.Module): the module to be converted. + process_group (ProcessGroup or list[ProcessGroup]): the process group(s) to be used for communication. + If this is a list, the process group at the ith index of the list will correspond to the process group + in the ith axis of the device mesh. Defaults to None, which means the global process group. + """ + raise NotImplementedError + + def _save_to_state_dict(self, destination, prefix, keep_vars): + r"""Saves module state to `destination` dictionary, containing a state + of the module, but not its descendants. This is called on every + submodule in :meth:`~torch.nn.Module.state_dict`. + + In rare cases, subclasses can achieve class-specific behavior by + overriding this method with custom logic. + + Args: + destination (dict): a dict where state will be stored + prefix (str): the prefix for parameters and buffers used in this + module + """ + for name, param in self._parameters.items(): + if param is not None: + param = gather_distributed_param(param, keep_vars=keep_vars) + if is_padded_tensor(param): + param = to_unpadded_tensor(param) + destination[prefix + name] = param.data + + for name, buf in self._buffers.items(): + if buf is not None and name not in self._non_persistent_buffers_set: + destination[prefix + name] = buf if keep_vars else buf.detach() + extra_state_key = prefix + _EXTRA_STATE_KEY_SUFFIX + if getattr(self.__class__, "get_extra_state", Module.get_extra_state) is not Module.get_extra_state: + destination[extra_state_key] = self.get_extra_state() + + def _load_from_state_dict( + self, state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs + ): + r"""Copies parameters and buffers from :attr:`state_dict` into only + this module, but not its descendants. This is called on every submodule + in :meth:`~torch.nn.Module.load_state_dict`. Metadata saved for this + module in input :attr:`state_dict` is provided as :attr:`local_metadata`. + For state dicts without metadata, :attr:`local_metadata` is empty. + Subclasses can achieve class-specific backward compatible loading using + the version number at `local_metadata.get("version", None)`. + + .. note:: + :attr:`state_dict` is not the same object as the input + :attr:`state_dict` to :meth:`~torch.nn.Module.load_state_dict`. So + it can be modified. + + Args: + state_dict (dict): a dict containing parameters and + persistent buffers. + prefix (str): the prefix for parameters and buffers used in this + module + local_metadata (dict): a dict containing the metadata for this module. + See + strict (bool): whether to strictly enforce that the keys in + :attr:`state_dict` with :attr:`prefix` match the names of + parameters and buffers in this module + missing_keys (list of str): if ``strict=True``, add missing keys to + this list + unexpected_keys (list of str): if ``strict=True``, add unexpected + keys to this list + error_msgs (list of str): error messages should be added to this + list, and will be reported together in + :meth:`~torch.nn.Module.load_state_dict` + """ + for hook in self._load_state_dict_pre_hooks.values(): + hook(state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs) + + persistent_buffers = {k: v for k, v in self._buffers.items() if k not in self._non_persistent_buffers_set} + local_name_params = itertools.chain(self._parameters.items(), persistent_buffers.items()) + local_state = {k: v for k, v in local_name_params if v is not None} + + for name, param in local_state.items(): + key = prefix + name + + if key in state_dict: + input_param = state_dict[key] + if not torch.overrides.is_tensor_like(input_param): + error_msgs.append( + 'While copying the parameter named "{}", ' + "expected torch.Tensor or Tensor-like object from checkpoint but " + "received {}".format(key, type(input_param)) + ) + continue + + if is_padded_tensor(param): + input_param = to_padded_tensor(input_param, param._current_length, param._padding_dim) + + if is_distributed_tensor(param): + # shard the input param + device_mesh = get_device_mesh(param) + sharding_spec = get_sharding_spec(param) + sharded_tensor = distribute_tensor(input_param, device_mesh, sharding_spec) + input_param = sharded_tensor_to_param(sharded_tensor) + elif is_customized_distributed_tensor(param): + input_param = distribute_tensor_with_customization(input_param, param.shard_fn, param.gather_fn) + + # This is used to avoid copying uninitialized parameters into + # non-lazy modules, since they dont have the hook to do the checks + # in such case, it will error when accessing the .shape attribute. + is_param_lazy = torch.nn.parameter.is_lazy(param) + # Backward compatibility: loading 1-dim tensor from 0.3.* to version 0.4+ + if not is_param_lazy and len(param.shape) == 0 and len(input_param.shape) == 1: + input_param = input_param[0] + + if not is_param_lazy and input_param.shape != param.shape: + # local shape should match the one in checkpoint + error_msgs.append( + "size mismatch for {}: copying a param with shape {} from checkpoint, " + "the shape in current model is {}.".format(key, input_param.shape, param.shape) + ) + continue + + try: + with torch.no_grad(): + param.copy_(input_param) + except Exception as ex: + error_msgs.append( + 'While copying the parameter named "{}", ' + "whose dimensions in the model are {} and " + "whose dimensions in the checkpoint are {}, " + "an exception occurred : {}.".format(key, param.size(), input_param.size(), ex.args) + ) + elif strict: + missing_keys.append(key) + + extra_state_key = prefix + _EXTRA_STATE_KEY_SUFFIX + if getattr(self.__class__, "set_extra_state", Module.set_extra_state) is not Module.set_extra_state: + if extra_state_key in state_dict: + self.set_extra_state(state_dict[extra_state_key]) + elif strict: + missing_keys.append(extra_state_key) + elif strict and (extra_state_key in state_dict): + unexpected_keys.append(extra_state_key) + + if strict: + for key in state_dict.keys(): + if key.startswith(prefix) and key != extra_state_key: + input_name = key[len(prefix) :] + input_name = input_name.split(".", 1)[0] # get the name of param/buffer/child + if input_name not in self._modules and input_name not in local_state: + unexpected_keys.append(key) + + def resize_embedding_weight(self): + self.weight = to_padded_tensor(self.weight, self.new_num_embeddings, 0) + + def resize_embedding_bias(self): + self.bias = to_padded_tensor(self.bias, self.new_num_embeddings, 0) diff --git a/colossalai/shardformer/modeling/gpt2.py b/colossalai/shardformer/modeling/gpt2.py index 1306c8aa6299..26088569a4aa 100644 --- a/colossalai/shardformer/modeling/gpt2.py +++ b/colossalai/shardformer/modeling/gpt2.py @@ -26,7 +26,6 @@ from colossalai.shardformer.shard import ShardConfig from ..layer import cross_entropy_1d -from ..layer._operation import gather_forward_split_backward logger = logging.get_logger(__name__) @@ -397,13 +396,11 @@ def gpt2_lmhead_model_forward( shift_logits, shift_labels, process_group=shard_config.tensor_parallel_process_group, + vocab_size=self.lm_head.out_features, ) else: loss = loss_fct(shift_logits, shift_labels) - if not shard_config.parallel_output: - lm_logits = gather_forward_split_backward(lm_logits, -1, shard_config.tensor_parallel_process_group) - if not return_dict: output = (lm_logits,) + outputs[1:] return ((loss,) + output) if loss is not None else output @@ -1301,12 +1298,12 @@ def forward( shift_logits = shift_logits.view(-1, shift_logits.size(-1)) shift_labels = shift_labels.view(-1) loss = cross_entropy_1d( - shift_logits, shift_labels, process_group=shard_config.tensor_parallel_process_group + shift_logits, + shift_labels, + process_group=shard_config.tensor_parallel_process_group, + vocab_size=self.lm_head.out_features, ) - if not shard_config.parallel_output: - lm_logits = gather_forward_split_backward(lm_logits, -1, shard_config.tensor_parallel_process_group) - if not return_dict: output = (lm_logits,) + transformer_outputs[1:] return ((loss,) + output) if loss is not None else output diff --git a/colossalai/shardformer/modeling/llama.py b/colossalai/shardformer/modeling/llama.py index 0f1b4ad0a5d5..c3b5426c2875 100644 --- a/colossalai/shardformer/modeling/llama.py +++ b/colossalai/shardformer/modeling/llama.py @@ -316,7 +316,10 @@ def llama_for_causal_lm_forward( new_vocab_size = logits.shape[-1] shift_logits = shift_logits.view(-1, new_vocab_size) loss = cross_entropy_1d( - shift_logits, shift_labels, process_group=shard_config.tensor_parallel_process_group + shift_logits, + shift_labels, + process_group=shard_config.tensor_parallel_process_group, + vocab_size=self.lm_head.out_features, ) else: shift_logits = shift_logits.view(-1, self.config.vocab_size) @@ -735,11 +738,13 @@ def forward( shift_labels = shift_labels.view(-1) # Enable model parallelism shift_labels = shift_labels.to(shift_logits.device) - new_vocab_size = logits.shape[-1] shift_logits = shift_logits.view(-1, new_vocab_size) loss = cross_entropy_1d( - shift_logits, shift_labels, process_group=shard_config.tensor_parallel_process_group + shift_logits, + shift_labels, + process_group=shard_config.tensor_parallel_process_group, + vocab_size=self.lm_head.out_features, ) if not return_dict: diff --git a/colossalai/shardformer/policies/base_policy.py b/colossalai/shardformer/policies/base_policy.py index d67ab0a3c6bb..e976672bbfd2 100644 --- a/colossalai/shardformer/policies/base_policy.py +++ b/colossalai/shardformer/policies/base_policy.py @@ -195,3 +195,12 @@ def get_shared_params(self) -> List[Dict[int, Tensor]]: List[Dict[int, Tensor]]: List of parameters that should be shared across stages. E.g. [{0: module.model.embed_tokens.weight, 3: module.lm_head.weight}] """ return [] + + def tie_weight_check(self): + input_embedding = self.model.get_input_embeddings() + output_embedding = self.model.get_output_embeddings() + return ( + input_embedding is not None + and output_embedding is not None + and id(input_embedding.weight) == id(output_embedding.weight) + ) diff --git a/colossalai/shardformer/policies/bert.py b/colossalai/shardformer/policies/bert.py index 0a61d8cff410..d43fc893aedc 100644 --- a/colossalai/shardformer/policies/bert.py +++ b/colossalai/shardformer/policies/bert.py @@ -37,17 +37,7 @@ def config_sanity_check(self): pass def preprocess(self): - # reshape the embedding layer - r""" - Reshape the Embedding layer to make the embedding dimension divisible by world_size - """ - # TODO: - if self.shard_config.enable_tensor_parallelism: - vocab_size = self.model.config.vocab_size - world_size = self.shard_config.tensor_parallel_size - if vocab_size % world_size != 0: - new_vocab_size = vocab_size + world_size - vocab_size % world_size - self.model.resize_token_embeddings(new_vocab_size) + self.tie_weight = self.tie_weight_check() return self.model def module_policy(self): @@ -62,6 +52,13 @@ def module_policy(self): policy = {} + embedding_cls = None + if self.shard_config.enable_tensor_parallelism: + embedding_cls = col_nn.VocabParallelEmbedding1D + else: + if self.tie_weight: + embedding_cls = col_nn.PaddingEmbedding + if self.shard_config.enable_fused_normalization: norm_cls = col_nn.FusedLayerNorm else: @@ -150,10 +147,6 @@ def module_policy(self): policy[BertEmbeddings] = ModulePolicyDescription( sub_module_replacement=[ - SubModuleReplacementDescription( - suffix="word_embeddings", - target_module=col_nn.VocabParallelEmbedding1D, - ), SubModuleReplacementDescription( suffix="dropout", target_module=col_nn.DropoutForReplicatedInput, @@ -168,6 +161,18 @@ def module_policy(self): target_key=BertModel, ) + if embedding_cls is not None: + self.append_or_create_submodule_replacement( + description=[ + SubModuleReplacementDescription( + suffix="word_embeddings", + target_module=embedding_cls, + ) + ], + policy=policy, + target_key=BertEmbeddings, + ) + # optimization configuration # Handle bert layer self.append_or_create_submodule_replacement( @@ -237,8 +242,21 @@ def add_lm_head_policy(self, base_policy): self.append_or_create_submodule_replacement( description=SubModuleReplacementDescription( suffix="decoder", - target_module=col_nn.Linear1D_Col, - kwargs={"gather_output": True}, + target_module=col_nn.VocabParallelLMHead1D, + kwargs={ + "gather_output": True, + "make_vocab_size_divisible_by": self.shard_config.make_vocab_size_divisible_by, + }, + ), + policy=base_policy, + target_key=BertLMPredictionHead, + ) + else: + self.append_or_create_submodule_replacement( + description=SubModuleReplacementDescription( + suffix="decoder", + target_module=col_nn.PaddingLMHead, + kwargs={"make_vocab_size_divisible_by": self.shard_config.make_vocab_size_divisible_by}, ), policy=base_policy, target_key=BertLMPredictionHead, diff --git a/colossalai/shardformer/policies/blip2.py b/colossalai/shardformer/policies/blip2.py index 9be2a1e78073..b845e9336cac 100644 --- a/colossalai/shardformer/policies/blip2.py +++ b/colossalai/shardformer/policies/blip2.py @@ -17,16 +17,7 @@ def config_sanity_check(self): pass def preprocess(self): - # reshape the embedding layer - r""" - Reshape the Embedding layer to make the embedding dimension divisible by world_size - """ - # TODO: - vocab_size = self.model.config.qformer_config.vocab_size - world_size = self.shard_config.tensor_parallel_size - if vocab_size % world_size != 0: - new_vocab_size = vocab_size + world_size - vocab_size % world_size - self.model.resize_token_embeddings(new_vocab_size) + self.tie_weight = self.tie_weight_check() return self.model def module_policy(self): @@ -43,6 +34,13 @@ def module_policy(self): policy = {} + embedding_cls = None + if self.shard_config.enable_tensor_parallelism: + embedding_cls = col_nn.VocabParallelEmbedding1D + else: + if self.tie_weight: + embedding_cls = col_nn.PaddingEmbedding + if self.shard_config.enable_fused_normalization: norm_cls = col_nn.FusedLayerNorm else: @@ -202,22 +200,48 @@ def module_policy(self): ], ) - policy[OPTForCausalLM] = ModulePolicyDescription( - sub_module_replacement=[ + policy[Blip2Attention] = ModulePolicyDescription(method_replacement={"forward": forward_fn()}) + + if embedding_cls is not None: + self.append_or_create_submodule_replacement( + description=[ SubModuleReplacementDescription( suffix="model.decoder.embed_tokens", - target_module=col_nn.VocabParallelEmbedding1D, + target_module=embedding_cls, + kwargs={"make_vocab_size_divisible_by": self.shard_config.make_vocab_size_divisible_by}, ), + ], + policy=policy, + target_key=OPTForCausalLM, + ) + + if self.shard_config.enable_tensor_parallelism: + self.append_or_create_submodule_replacement( + description=[ SubModuleReplacementDescription( suffix="lm_head", - target_module=col_nn.Linear1D_Col, - kwargs={"gather_output": True}, + target_module=col_nn.VocabParallelLMHead1D, + kwargs={ + "gather_output": True, + "make_vocab_size_divisible_by": self.shard_config.make_vocab_size_divisible_by, + }, ), - ] + ], + policy=policy, + target_key=OPTForCausalLM, + ) + else: + self.append_or_create_submodule_replacement( + description=[ + SubModuleReplacementDescription( + suffix="lm_head", + target_module=col_nn.PaddingLMHead, + kwargs={"make_vocab_size_divisible_by": self.shard_config.make_vocab_size_divisible_by}, + ), + ], + policy=policy, + target_key=OPTForCausalLM, ) - - policy[Blip2Attention] = ModulePolicyDescription(method_replacement={"forward": forward_fn()}) - # optimization configuration # Handle Blip2EncoderLayer layer self.append_or_create_submodule_replacement( diff --git a/colossalai/shardformer/policies/bloom.py b/colossalai/shardformer/policies/bloom.py index 2becadc3fb19..953592abc16a 100644 --- a/colossalai/shardformer/policies/bloom.py +++ b/colossalai/shardformer/policies/bloom.py @@ -35,16 +35,7 @@ def config_sanity_check(self): pass def preprocess(self): - # reshape the embedding layer - r""" - Reshape the Embedding layer to make the embedding dimension divisible by world_size - """ - if self.shard_config.enable_tensor_parallelism: - vocab_size = self.model.config.vocab_size - world_size = self.shard_config.tensor_parallel_size - if vocab_size % world_size != 0: - new_vocab_size = vocab_size + world_size - vocab_size % world_size - self.model.resize_token_embeddings(new_vocab_size) + self.tie_weight = self.tie_weight_check() return self.model def module_policy(self): @@ -52,6 +43,13 @@ def module_policy(self): policy = {} + embedding_cls = None + if self.shard_config.enable_tensor_parallelism: + embedding_cls = col_nn.VocabParallelEmbedding1D + else: + if self.tie_weight: + embedding_cls = col_nn.PaddingEmbedding + if self.shard_config.enable_fused_normalization: norm_cls = col_nn.FusedLayerNorm else: @@ -112,12 +110,19 @@ def module_policy(self): method_replacement={ "build_alibi_tensor": build_bloom_alibi_tensor_fn(self.shard_config.tensor_parallel_process_group) }, - sub_module_replacement=[ + ) + + if embedding_cls is not None: + self.append_or_create_submodule_replacement( + description=[ SubModuleReplacementDescription( suffix="word_embeddings", - target_module=col_nn.VocabParallelEmbedding1D, - ) + target_module=embedding_cls, + kwargs={"make_vocab_size_divisible_by": self.shard_config.make_vocab_size_divisible_by}, + ), ], + policy=policy, + target_key=BloomModel, ) # optimization configuration @@ -282,7 +287,21 @@ def module_policy(self): if self.shard_config.enable_tensor_parallelism: self.append_or_create_submodule_replacement( description=SubModuleReplacementDescription( - suffix="lm_head", target_module=col_nn.Linear1D_Col, kwargs=dict(gather_output=True) + suffix="lm_head", + target_module=col_nn.VocabParallelLMHead1D, + kwargs=dict( + gather_output=True, make_vocab_size_divisible_by=self.shard_config.make_vocab_size_divisible_by + ), + ), + policy=policy, + target_key=BloomForCausalLM, + ) + else: + self.append_or_create_submodule_replacement( + description=SubModuleReplacementDescription( + suffix="lm_head", + target_module=col_nn.PaddingLMHead, + kwargs=dict(make_vocab_size_divisible_by=self.shard_config.make_vocab_size_divisible_by), ), policy=policy, target_key=BloomForCausalLM, diff --git a/colossalai/shardformer/policies/chatglm2.py b/colossalai/shardformer/policies/chatglm2.py index dabc14bffc95..f205835e7815 100644 --- a/colossalai/shardformer/policies/chatglm2.py +++ b/colossalai/shardformer/policies/chatglm2.py @@ -25,20 +25,12 @@ def config_sanity_check(self): pass def preprocess(self): - # Resize embedding - if self.shard_config.enable_tensor_parallelism: - vocab_size = self.model.config.padded_vocab_size - world_size = self.shard_config.tensor_parallel_size - - if vocab_size % world_size != 0: - new_vocab_size = vocab_size + world_size - vocab_size % world_size - self.model.resize_token_embeddings(new_vocab_size) - if self.pipeline_stage_manager is not None: # the batch_size_dim is bounded to Model bsz_dim = 1 setattr(self.model, "batch_size_dim", bsz_dim) + self.tie_weight = self.tie_weight_check() return self.model def module_policy(self) -> Dict[Union[str, nn.Module], ModulePolicyDescription]: @@ -46,6 +38,13 @@ def module_policy(self) -> Dict[Union[str, nn.Module], ModulePolicyDescription]: policy = {} + embedding_cls = None + if self.shard_config.enable_tensor_parallelism: + embedding_cls = col_nn.VocabParallelEmbedding1D + else: + if self.tie_weight: + embedding_cls = col_nn.PaddingEmbedding + if self.shard_config.enable_fused_normalization: if self.model.config.rmsnorm: norm_cls = col_nn.FusedRMSNorm @@ -68,16 +67,6 @@ def module_policy(self) -> Dict[Union[str, nn.Module], ModulePolicyDescription]: sp_partial_derived = sp_mode == "split_gather" if self.shard_config.enable_tensor_parallelism: - policy[ChatGLMModel] = ModulePolicyDescription( - attribute_replacement={}, - sub_module_replacement=[ - SubModuleReplacementDescription( - suffix="embedding.word_embeddings", - target_module=col_nn.VocabParallelEmbedding1D, - ) - ], - ) - policy[GLMBlock] = ModulePolicyDescription( attribute_replacement={ "self_attention.num_attention_heads_per_partition": self.model.config.num_attention_heads @@ -114,6 +103,19 @@ def module_policy(self) -> Dict[Union[str, nn.Module], ModulePolicyDescription]: ), ], ) + + if embedding_cls is not None: + self.append_or_create_submodule_replacement( + description=[ + SubModuleReplacementDescription( + suffix="embedding.word_embeddings", + target_module=embedding_cls, + kwargs={"make_vocab_size_divisible_by": self.shard_config.make_vocab_size_divisible_by}, + ), + ], + policy=policy, + target_key=ChatGLMModel, + ) # optimization configuration self.append_or_create_submodule_replacement( description=[ diff --git a/colossalai/shardformer/policies/falcon.py b/colossalai/shardformer/policies/falcon.py index fe61c406fae3..a2f110a41a3c 100644 --- a/colossalai/shardformer/policies/falcon.py +++ b/colossalai/shardformer/policies/falcon.py @@ -32,16 +32,7 @@ def config_sanity_check(self): pass def preprocess(self): - # reshape the embedding layer - r""" - Reshape the Embedding layer to make the embedding dimension divisible by world_size - """ - if self.shard_config.enable_tensor_parallelism: - vocab_size = self.model.config.vocab_size - world_size = self.shard_config.tensor_parallel_size - if vocab_size % world_size != 0: - new_vocab_size = vocab_size + world_size - vocab_size % world_size - self.model.resize_token_embeddings(new_vocab_size) + self.tie_weight = self.tie_weight_check() return self.model def module_policy(self): @@ -58,6 +49,14 @@ def module_policy(self): warnings.warn("Falcon doesn't support sequence parallelism now, will ignore the sequence parallelism flag.") policy = {} + + embedding_cls = None + if self.shard_config.enable_tensor_parallelism: + embedding_cls = col_nn.VocabParallelEmbedding1D + else: + if self.tie_weight: + embedding_cls = col_nn.PaddingEmbedding + if self.shard_config.enable_tensor_parallelism: attn_attribute_replacement = { "self_attention.hidden_size": self.model.config.hidden_size // self.shard_config.tensor_parallel_size, @@ -98,12 +97,19 @@ def module_policy(self): method_replacement={ "build_alibi_tensor": build_falcon_alibi_tensor_fn(self.shard_config.tensor_parallel_process_group) }, - sub_module_replacement=[ + ) + + if embedding_cls is not None: + self.append_or_create_submodule_replacement( + description=[ SubModuleReplacementDescription( suffix="word_embeddings", - target_module=col_nn.VocabParallelEmbedding1D, - ) + target_module=embedding_cls, + kwargs={"make_vocab_size_divisible_by": self.shard_config.make_vocab_size_divisible_by}, + ), ], + policy=policy, + target_key=FalconModel, ) # optimization configuration @@ -232,11 +238,26 @@ def module_policy(self): if self.shard_config.enable_tensor_parallelism: self.append_or_create_submodule_replacement( description=SubModuleReplacementDescription( - suffix="lm_head", target_module=col_nn.Linear1D_Col, kwargs=dict(gather_output=True) + suffix="lm_head", + target_module=col_nn.VocabParallelLMHead1D, + kwargs=dict( + gather_output=True, make_vocab_size_divisible_by=self.shard_config.make_vocab_size_divisible_by + ), ), policy=policy, target_key=FalconForCausalLM, ) + else: + self.append_or_create_submodule_replacement( + description=SubModuleReplacementDescription( + suffix="lm_head", + target_module=col_nn.PaddingLMHead, + kwargs=dict(make_vocab_size_divisible_by=self.shard_config.make_vocab_size_divisible_by), + ), + policy=policy, + target_key=FalconForCausalLM, + ) + if self.pipeline_stage_manager: self.set_pipeline_forward( model_cls=FalconForCausalLM, diff --git a/colossalai/shardformer/policies/gpt2.py b/colossalai/shardformer/policies/gpt2.py index 380a432dc8b8..98db7b948954 100644 --- a/colossalai/shardformer/policies/gpt2.py +++ b/colossalai/shardformer/policies/gpt2.py @@ -34,12 +34,7 @@ def preprocess(self): r""" Reshape the Embedding layer to make the embedding dimension divisible by world_size """ - if self.shard_config.enable_tensor_parallelism: - vocab_size = self.model.config.vocab_size - world_size = self.shard_config.tensor_parallel_size - if vocab_size % world_size != 0: - new_vocab_size = vocab_size + world_size - vocab_size % world_size - self.model.resize_token_embeddings(new_vocab_size) + self.tie_weight = self.tie_weight_check() return self.model def module_policy(self): @@ -47,6 +42,13 @@ def module_policy(self): policy = {} + embedding_cls = None + if self.shard_config.enable_tensor_parallelism: + embedding_cls = col_nn.VocabParallelEmbedding1D + else: + if self.tie_weight: + embedding_cls = col_nn.PaddingEmbedding + if self.shard_config.enable_fused_normalization: norm_cls = col_nn.FusedLayerNorm else: @@ -73,10 +75,6 @@ def module_policy(self): if self.shard_config.enable_tensor_parallelism: policy[GPT2Model] = ModulePolicyDescription( sub_module_replacement=[ - SubModuleReplacementDescription( - suffix="wte", - target_module=col_nn.VocabParallelEmbedding1D, - ), SubModuleReplacementDescription( suffix="drop", target_module=col_nn.DropoutForParallelInput, @@ -137,6 +135,17 @@ def module_policy(self): ), ], ) + if embedding_cls is not None: + # padding vocabulary size when using pp to make it divisible by shard_config.make_vocab_size_divisible_by + self.append_or_create_submodule_replacement( + description=SubModuleReplacementDescription( + suffix="wte", + target_module=embedding_cls, + kwargs={"make_vocab_size_divisible_by": self.shard_config.make_vocab_size_divisible_by}, + ), + policy=policy, + target_key=GPT2Model, + ) # optimization configuration self.append_or_create_submodule_replacement( @@ -298,8 +307,11 @@ def module_policy(self): sub_module_replacement=[ SubModuleReplacementDescription( suffix="lm_head", - target_module=col_nn.Linear1D_Col, - kwargs={"gather_output": not self.shard_config.parallel_output}, + target_module=col_nn.VocabParallelLMHead1D, + kwargs={ + "gather_output": False, + "make_vocab_size_divisible_by": self.shard_config.make_vocab_size_divisible_by, + }, ) ], ) @@ -308,7 +320,19 @@ def module_policy(self): addon_module[GPT2LMHeadModel].method_replacement = { "forward": get_lm_forward_with_dist_cross_entropy(self.shard_config) } - module_policy.update(addon_module) + else: + addon_module = { + GPT2LMHeadModel: ModulePolicyDescription( + sub_module_replacement=[ + SubModuleReplacementDescription( + suffix="lm_head", + target_module=col_nn.PaddingLMHead, + kwargs={"make_vocab_size_divisible_by": self.shard_config.make_vocab_size_divisible_by}, + ) + ] + ) + } + module_policy.update(addon_module) if self.pipeline_stage_manager is not None: self.set_pipeline_forward( @@ -353,13 +377,28 @@ def module_policy(self): sub_module_replacement=[ SubModuleReplacementDescription( suffix="lm_head", - target_module=col_nn.Linear1D_Col, - kwargs={"gather_output": True}, + target_module=col_nn.VocabParallelLMHead1D, + kwargs={ + "gather_output": True, + "make_vocab_size_divisible_by": self.shard_config.make_vocab_size_divisible_by, + }, ) ] ) } - module_policy.update(addon_module) + else: + addon_module = { + GPT2DoubleHeadsModel: ModulePolicyDescription( + sub_module_replacement=[ + SubModuleReplacementDescription( + suffix="lm_head", + target_module=col_nn.PaddingLMHead, + kwargs={"make_vocab_size_divisible_by": self.shard_config.make_vocab_size_divisible_by}, + ) + ] + ) + } + module_policy.update(addon_module) if self.pipeline_stage_manager is not None: self.set_pipeline_forward( diff --git a/colossalai/shardformer/policies/gptj.py b/colossalai/shardformer/policies/gptj.py index eab4c214a41f..4b69137a6892 100644 --- a/colossalai/shardformer/policies/gptj.py +++ b/colossalai/shardformer/policies/gptj.py @@ -29,22 +29,21 @@ def config_sanity_check(self): pass def preprocess(self): - # reshape the embedding layer - r""" - Reshape the Embedding layer to make the embedding dimension divisible by world_size - """ - if self.shard_config.enable_tensor_parallelism: - vocab_size = self.model.config.vocab_size - world_size = self.shard_config.tensor_parallel_size - if vocab_size % world_size != 0: - new_vocab_size = vocab_size + world_size - vocab_size % world_size - self.model.resize_token_embeddings(new_vocab_size) + self.tie_weight = self.tie_weight_check() return self.model def module_policy(self): from transformers.models.gptj.modeling_gptj import GPTJAttention, GPTJBlock, GPTJModel policy = {} + + embedding_cls = None + if self.shard_config.enable_tensor_parallelism: + embedding_cls = col_nn.VocabParallelEmbedding1D + else: + if self.tie_weight: + embedding_cls = col_nn.PaddingEmbedding + if self.shard_config.enable_sequence_parallelism: self.shard_config.enable_sequence_parallelism = False warnings.warn("GPTJ doesn't support sequence parallelism now, will ignore the sequence parallelism flag.") @@ -54,10 +53,6 @@ def module_policy(self): if self.shard_config.enable_tensor_parallelism: policy[GPTJModel] = ModulePolicyDescription( sub_module_replacement=[ - SubModuleReplacementDescription( - suffix="wte", - target_module=col_nn.VocabParallelEmbedding1D, - ), SubModuleReplacementDescription( suffix="drop", target_module=col_nn.DropoutForParallelInput, @@ -126,6 +121,17 @@ def module_policy(self): ], ) + if embedding_cls is not None: + self.append_or_create_submodule_replacement( + description=SubModuleReplacementDescription( + suffix="wte", + target_module=embedding_cls, + kwargs={"make_vocab_size_divisible_by": self.shard_config.make_vocab_size_divisible_by}, + ), + policy=policy, + target_key=GPTJModel, + ) + # optimization configuration if self.shard_config.enable_fused_normalization: self.append_or_create_submodule_replacement( @@ -255,13 +261,28 @@ def module_policy(self): sub_module_replacement=[ SubModuleReplacementDescription( suffix="lm_head", - target_module=col_nn.Linear1D_Col, - kwargs={"gather_output": True}, + target_module=col_nn.VocabParallelLMHead1D, + kwargs={ + "gather_output": True, + "make_vocab_size_divisible_by": self.shard_config.make_vocab_size_divisible_by, + }, + ) + ] + ) + } + else: + addon_module = { + GPTJForCausalLM: ModulePolicyDescription( + sub_module_replacement=[ + SubModuleReplacementDescription( + suffix="lm_head", + target_module=col_nn.PaddingLMHead, + kwargs={"make_vocab_size_divisible_by": self.shard_config.make_vocab_size_divisible_by}, ) ] ) } - policy.update(addon_module) + policy.update(addon_module) if self.pipeline_stage_manager is not None: self.set_pipeline_forward( diff --git a/colossalai/shardformer/policies/llama.py b/colossalai/shardformer/policies/llama.py index bb4551b2c31c..ff686a179553 100644 --- a/colossalai/shardformer/policies/llama.py +++ b/colossalai/shardformer/policies/llama.py @@ -6,7 +6,16 @@ from torch import Tensor from torch.nn import Module -from colossalai.shardformer.layer import FusedRMSNorm, Linear1D_Col, Linear1D_Row, RMSNorm, VocabParallelEmbedding1D +from colossalai.shardformer.layer import ( + FusedRMSNorm, + Linear1D_Col, + Linear1D_Row, + PaddingEmbedding, + PaddingLMHead, + RMSNorm, + VocabParallelEmbedding1D, + VocabParallelLMHead1D, +) from ..modeling.llama import ( LlamaPipelineForwards, @@ -26,15 +35,7 @@ def config_sanity_check(self): pass def preprocess(self): - if self.shard_config.enable_tensor_parallelism: - # Resize embedding - vocab_size = self.model.config.vocab_size - world_size = self.shard_config.tensor_parallel_size - - if vocab_size % world_size != 0: - new_vocab_size = vocab_size + world_size - vocab_size % world_size - self.model.resize_token_embeddings(new_vocab_size) - + self.tie_weight = self.tie_weight_check() return self.model def module_policy(self) -> Dict[Union[str, nn.Module], ModulePolicyDescription]: @@ -42,6 +43,13 @@ def module_policy(self) -> Dict[Union[str, nn.Module], ModulePolicyDescription]: policy = {} + embedding_cls = None + if self.shard_config.enable_tensor_parallelism: + embedding_cls = VocabParallelEmbedding1D + else: + if self.tie_weight: + embedding_cls = PaddingEmbedding + if self.shard_config.enable_fused_normalization: norm_cls = FusedRMSNorm else: @@ -167,10 +175,12 @@ def module_policy(self) -> Dict[Union[str, nn.Module], ModulePolicyDescription]: ], ) + if embedding_cls is not None: self.append_or_create_submodule_replacement( description=SubModuleReplacementDescription( suffix="embed_tokens", - target_module=VocabParallelEmbedding1D, + target_module=embedding_cls, + kwargs={"make_vocab_size_divisible_by": self.shard_config.make_vocab_size_divisible_by}, ), policy=policy, target_key=LlamaModel, @@ -327,8 +337,11 @@ def module_policy(self): sub_module_replacement=[ SubModuleReplacementDescription( suffix="lm_head", - target_module=Linear1D_Col, - kwargs={"gather_output": not self.shard_config.parallel_output}, + target_module=VocabParallelLMHead1D, + kwargs={ + "gather_output": not self.shard_config.parallel_output, + "make_vocab_size_divisible_by": self.shard_config.make_vocab_size_divisible_by, + }, ) ], ) @@ -337,7 +350,19 @@ def module_policy(self): new_item[LlamaForCausalLM].method_replacement = { "forward": get_lm_forward_with_dist_cross_entropy(self.shard_config) } - policy.update(new_item) + else: + new_item = { + LlamaForCausalLM: ModulePolicyDescription( + sub_module_replacement=[ + SubModuleReplacementDescription( + suffix="lm_head", + target_module=PaddingLMHead, + kwargs={"make_vocab_size_divisible_by": self.shard_config.make_vocab_size_divisible_by}, + ) + ], + ) + } + policy.update(new_item) if self.pipeline_stage_manager: # set None as default diff --git a/colossalai/shardformer/policies/mistral.py b/colossalai/shardformer/policies/mistral.py index c0b8b3375836..b225fd2a9632 100644 --- a/colossalai/shardformer/policies/mistral.py +++ b/colossalai/shardformer/policies/mistral.py @@ -3,7 +3,15 @@ import torch.nn as nn -from colossalai.shardformer.layer import FusedRMSNorm, Linear1D_Col, Linear1D_Row, VocabParallelEmbedding1D +from colossalai.shardformer.layer import ( + FusedRMSNorm, + Linear1D_Col, + Linear1D_Row, + PaddingEmbedding, + PaddingLMHead, + VocabParallelEmbedding1D, + VocabParallelLMHead1D, +) from ..modeling.mistral import get_mistral_flash_attention_forward from .base_policy import ModulePolicyDescription, Policy, SubModuleReplacementDescription @@ -16,15 +24,7 @@ def config_sanity_check(self): pass def preprocess(self): - if self.shard_config.enable_tensor_parallelism: - # Resize embedding - vocab_size = self.model.config.vocab_size - world_size = self.shard_config.tensor_parallel_size - - if vocab_size % world_size != 0: - new_vocab_size = vocab_size + world_size - vocab_size % world_size - self.model.resize_token_embeddings(new_vocab_size) - + self.tie_weight = self.tie_weight_check() return self.model def module_policy(self) -> Dict[Union[str, nn.Module], ModulePolicyDescription]: @@ -32,6 +32,13 @@ def module_policy(self) -> Dict[Union[str, nn.Module], ModulePolicyDescription]: policy = {} + embedding_cls = None + if self.shard_config.enable_tensor_parallelism: + embedding_cls = VocabParallelEmbedding1D + else: + if self.tie_weight: + embedding_cls = PaddingEmbedding + if self.shard_config.enable_sequence_parallelism: self.shard_config.enable_sequence_parallelism = False warnings.warn( @@ -80,10 +87,12 @@ def module_policy(self) -> Dict[Union[str, nn.Module], ModulePolicyDescription]: ], ) + if embedding_cls is not None: self.append_or_create_submodule_replacement( description=SubModuleReplacementDescription( suffix="embed_tokens", - target_module=VocabParallelEmbedding1D, + target_module=embedding_cls, + kwargs={"make_vocab_size_divisible_by": self.shard_config.make_vocab_size_divisible_by}, ), policy=policy, target_key=MistralModel, @@ -146,6 +155,8 @@ def module_policy(self): from transformers import MistralForCausalLM policy = super().module_policy() + if self.pipeline_stage_manager: + warnings.warn("Mistral doesn't support pipeline parallelism now.") if self.shard_config.enable_tensor_parallelism: # add a new item for casual lm @@ -153,16 +164,30 @@ def module_policy(self): MistralForCausalLM: ModulePolicyDescription( sub_module_replacement=[ SubModuleReplacementDescription( - suffix="lm_head", target_module=Linear1D_Col, kwargs=dict(gather_output=True) + suffix="lm_head", + target_module=VocabParallelLMHead1D, + kwargs=dict( + gather_output=True, + make_vocab_size_divisible_by=self.shard_config.make_vocab_size_divisible_by, + ), + ) + ] + ) + } + else: + new_item = { + MistralForCausalLM: ModulePolicyDescription( + sub_module_replacement=[ + SubModuleReplacementDescription( + suffix="lm_head", + target_module=PaddingLMHead, + kwargs=dict(make_vocab_size_divisible_by=self.shard_config.make_vocab_size_divisible_by), ) ] ) } - if self.pipeline_stage_manager: - warnings.warn("Mistral doesn't support pipeline parallelism now.") - - policy.update(new_item) + policy.update(new_item) return policy diff --git a/colossalai/shardformer/policies/opt.py b/colossalai/shardformer/policies/opt.py index 98e584be861b..ac78ff6a702f 100644 --- a/colossalai/shardformer/policies/opt.py +++ b/colossalai/shardformer/policies/opt.py @@ -5,7 +5,16 @@ import torch.nn as nn from torch import Tensor, nn -from colossalai.shardformer.layer import FusedLayerNorm, LayerNorm, Linear1D_Col, Linear1D_Row, VocabParallelEmbedding1D +from colossalai.shardformer.layer import ( + FusedLayerNorm, + LayerNorm, + Linear1D_Col, + Linear1D_Row, + PaddingEmbedding, + PaddingLMHead, + VocabParallelEmbedding1D, + VocabParallelLMHead1D, +) from .._utils import getattr_ from ..modeling.jit import get_jit_fused_dropout_add_func @@ -41,16 +50,7 @@ def config_sanity_check(self): pass def preprocess(self): - # reshape the embedding layer - r""" - Reshape the Embedding layer to make the embedding dimension divisible by world_size - """ - if self.shard_config.enable_tensor_parallelism: - vocab_size = self.model.config.vocab_size - world_size = self.shard_config.tensor_parallel_size - if vocab_size % world_size != 0: - new_vocab_size = vocab_size + world_size - vocab_size % world_size - self.model.resize_token_embeddings(new_vocab_size) + self.tie_weight = self.tie_weight_check() return self.model def module_policy(self): @@ -58,6 +58,13 @@ def module_policy(self): policy = {} + embedding_cls = None + if self.shard_config.enable_tensor_parallelism: + embedding_cls = VocabParallelEmbedding1D + else: + if self.tie_weight: + embedding_cls = PaddingEmbedding + if self.shard_config.enable_fused_normalization: norm_cls = FusedLayerNorm else: @@ -68,14 +75,6 @@ def module_policy(self): warnings.warn("OPT doesn't support sequence parallelism now, will ignore the sequence parallelism flag.") if self.shard_config.enable_tensor_parallelism: - policy[OPTDecoder] = ModulePolicyDescription( - sub_module_replacement=[ - SubModuleReplacementDescription( - suffix="embed_tokens", - target_module=VocabParallelEmbedding1D, - ) - ] - ) policy[OPTDecoderLayer] = ModulePolicyDescription( sub_module_replacement=[ SubModuleReplacementDescription( @@ -114,6 +113,17 @@ def module_policy(self): ], ) + if embedding_cls is not None: + self.append_or_create_submodule_replacement( + description=SubModuleReplacementDescription( + suffix="embed_tokens", + target_module=embedding_cls, + kwargs={"make_vocab_size_divisible_by": self.shard_config.make_vocab_size_divisible_by}, + ), + policy=policy, + target_key=OPTDecoder, + ) + # optimization configuration self.append_or_create_submodule_replacement( description=SubModuleReplacementDescription( @@ -253,8 +263,20 @@ def module_policy(self): self.append_or_create_submodule_replacement( description=SubModuleReplacementDescription( suffix="lm_head", - target_module=Linear1D_Col, - kwargs=dict(gather_output=True), + target_module=VocabParallelLMHead1D, + kwargs=dict( + gather_output=True, make_vocab_size_divisible_by=self.shard_config.make_vocab_size_divisible_by + ), + ), + policy=policy, + target_key=OPTForCausalLM, + ) + else: + self.append_or_create_submodule_replacement( + description=SubModuleReplacementDescription( + suffix="lm_head", + target_module=PaddingLMHead, + kwargs=dict(make_vocab_size_divisible_by=self.shard_config.make_vocab_size_divisible_by), ), policy=policy, target_key=OPTForCausalLM, diff --git a/colossalai/shardformer/policies/t5.py b/colossalai/shardformer/policies/t5.py index 0c8ec15fa0a9..3c7e92b47db0 100644 --- a/colossalai/shardformer/policies/t5.py +++ b/colossalai/shardformer/policies/t5.py @@ -13,8 +13,11 @@ FusedRMSNorm, Linear1D_Col, Linear1D_Row, + PaddingEmbedding, + PaddingLMHead, RMSNorm, VocabParallelEmbedding1D, + VocabParallelLMHead1D, ) from colossalai.shardformer.policies.base_policy import ModulePolicyDescription @@ -36,16 +39,7 @@ def config_sanity_check(self): pass def preprocess(self): - # reshape the embedding layer - r""" - Reshape the Embedding layer to make the embedding dimension divisible by world_size - """ - if self.shard_config.enable_tensor_parallelism: - vocab_size = self.model.config.vocab_size - world_size = self.shard_config.tensor_parallel_size - if vocab_size % world_size != 0: - new_vocab_size = vocab_size + world_size - vocab_size % world_size - self.model.resize_token_embeddings(new_vocab_size) + self.tie_weight = self.tie_weight_check() return self.model def module_policy(self): @@ -61,6 +55,13 @@ def module_policy(self): policy = {} + embedding_cls = None + if self.shard_config.enable_tensor_parallelism: + embedding_cls = VocabParallelEmbedding1D + else: + if self.tie_weight: + embedding_cls = PaddingEmbedding + if self.shard_config.enable_fused_normalization: norm_cls = FusedRMSNorm else: @@ -77,10 +78,6 @@ def module_policy(self): suffix="dropout", target_module=DropoutForParallelInput, ), - SubModuleReplacementDescription( - suffix="embed_tokens", - target_module=VocabParallelEmbedding1D, - ), ] ) policy[T5LayerSelfAttention] = ModulePolicyDescription( @@ -176,6 +173,17 @@ def module_policy(self): ] ) + if embedding_cls is not None: + self.append_or_create_submodule_replacement( + description=SubModuleReplacementDescription( + suffix="embed_tokens", + target_module=embedding_cls, + kwargs={"make_vocab_size_divisible_by": self.shard_config.make_vocab_size_divisible_by}, + ), + policy=policy, + target_key=T5Stack, + ) + # optimization configuration self.append_or_create_submodule_replacement( description=SubModuleReplacementDescription( @@ -370,11 +378,19 @@ def module_policy(self): policy = super().module_policy() + embedding_cls = None if self.shard_config.enable_tensor_parallelism: + embedding_cls = VocabParallelEmbedding1D + else: + if self.tie_weight: + embedding_cls = PaddingEmbedding + + if embedding_cls is not None: self.append_or_create_submodule_replacement( description=SubModuleReplacementDescription( suffix="shared", - target_module=VocabParallelEmbedding1D, + target_module=embedding_cls, + kwargs={"make_vocab_size_divisible_by": self.shard_config.make_vocab_size_divisible_by}, ), policy=policy, target_key=T5Model, @@ -406,17 +422,44 @@ def module_policy(self): policy = super().module_policy() + embedding_cls = None if self.shard_config.enable_tensor_parallelism: + embedding_cls = VocabParallelEmbedding1D + else: + if self.tie_weight: + embedding_cls = PaddingEmbedding + + if embedding_cls is not None: self.append_or_create_submodule_replacement( - description=[ - SubModuleReplacementDescription( - suffix="shared", - target_module=VocabParallelEmbedding1D, - ), - SubModuleReplacementDescription( - suffix="lm_head", target_module=Linear1D_Col, kwargs=dict(gather_output=True) - ), - ], + description=SubModuleReplacementDescription( + suffix="shared", + target_module=embedding_cls, + kwargs={"make_vocab_size_divisible_by": self.shard_config.make_vocab_size_divisible_by}, + ), + policy=policy, + target_key=T5ForConditionalGeneration, + ) + + if self.shard_config.enable_tensor_parallelism: + self.append_or_create_submodule_replacement( + description=SubModuleReplacementDescription( + suffix="lm_head", + target_module=VocabParallelLMHead1D, + kwargs={ + "gather_output": True, + "make_vocab_size_divisible_by": self.shard_config.make_vocab_size_divisible_by, + }, + ), + policy=policy, + target_key=T5ForConditionalGeneration, + ) + else: + self.append_or_create_submodule_replacement( + description=SubModuleReplacementDescription( + suffix="lm_head", + target_module=PaddingLMHead, + kwargs={"make_vocab_size_divisible_by": self.shard_config.make_vocab_size_divisible_by}, + ), policy=policy, target_key=T5ForConditionalGeneration, ) @@ -467,11 +510,19 @@ def module_policy(self): policy = super().module_policy() + embedding_cls = None if self.shard_config.enable_tensor_parallelism: + embedding_cls = VocabParallelEmbedding1D + else: + if self.tie_weight: + embedding_cls = PaddingEmbedding + + if embedding_cls is not None: self.append_or_create_submodule_replacement( description=SubModuleReplacementDescription( suffix="shared", - target_module=VocabParallelEmbedding1D, + target_module=embedding_cls, + kwargs={"make_vocab_size_divisible_by": self.shard_config.make_vocab_size_divisible_by}, ), policy=policy, target_key=T5EncoderModel, diff --git a/colossalai/shardformer/policies/whisper.py b/colossalai/shardformer/policies/whisper.py index c63f6d1cc549..0b5114fa64cd 100644 --- a/colossalai/shardformer/policies/whisper.py +++ b/colossalai/shardformer/policies/whisper.py @@ -45,11 +45,7 @@ def preprocess(self): r""" Reshape the Embedding layer to make the embedding dimension divisible by world_size """ - vocab_size = self.model.config.vocab_size - world_size = self.shard_config.tensor_parallel_size - if vocab_size % world_size != 0: - new_vocab_size = vocab_size + world_size - vocab_size % world_size - self.model.resize_token_embeddings(new_vocab_size) + self.tie_weight = self.tie_weight_check() return self.model def module_policy(self): @@ -63,6 +59,13 @@ def module_policy(self): policy = {} + embedding_cls = None + if self.shard_config.enable_tensor_parallelism: + embedding_cls = col_nn.VocabParallelEmbedding1D + else: + if self.tie_weight: + embedding_cls = col_nn.PaddingEmbedding + if self.shard_config.enable_fused_normalization: norm_cls = col_nn.FusedLayerNorm else: @@ -167,13 +170,17 @@ def module_policy(self): ], ) - policy[WhisperDecoder] = ModulePolicyDescription( - sub_module_replacement=[ + if embedding_cls is not None: + self.append_or_create_submodule_replacement( + description=[ SubModuleReplacementDescription( suffix="embed_tokens", - target_module=col_nn.VocabParallelEmbedding1D, + target_module=embedding_cls, + kwargs={"make_vocab_size_divisible_by": self.shard_config.make_vocab_size_divisible_by}, ), - ] + ], + policy=policy, + target_key=WhisperDecoder, ) # optimization configuration @@ -280,8 +287,21 @@ def add_lm_head_policy(self, base_policy): self.append_or_create_submodule_replacement( description=SubModuleReplacementDescription( suffix="proj_out", - target_module=col_nn.Linear1D_Col, - kwargs={"gather_output": True}, + target_module=col_nn.VocabParallelLMHead1D, + kwargs={ + "gather_output": True, + "make_vocab_size_divisible_by": self.shard_config.make_vocab_size_divisible_by, + }, + ), + policy=base_policy, + target_key=WhisperForConditionalGeneration, + ) + else: + self.append_or_create_submodule_replacement( + description=SubModuleReplacementDescription( + suffix="proj_out", + target_module=col_nn.PaddingLMHead, + kwargs={"make_vocab_size_divisible_by": self.shard_config.make_vocab_size_divisible_by}, ), policy=base_policy, target_key=WhisperForConditionalGeneration, @@ -526,9 +546,6 @@ def get_shared_params(self) -> List[Dict[int, Tensor]]: # WhisperForAudioClassification class WhisperForAudioClassificationPolicy(WhisperPolicy): - def preprocess(self): - return self.model - def module_policy(self): from transformers import WhisperForAudioClassification diff --git a/colossalai/shardformer/shard/shard_config.py b/colossalai/shardformer/shard/shard_config.py index 7489873c2ed6..963732543f27 100644 --- a/colossalai/shardformer/shard/shard_config.py +++ b/colossalai/shardformer/shard/shard_config.py @@ -42,10 +42,9 @@ class ShardConfig: sequence_parallelism_mode: str = None enable_sequence_overlap: bool = False parallel_output: bool = True + make_vocab_size_divisible_by: int = 64 gradient_checkpoint_config: Optional[GradientCheckpointConfig] = None extra_kwargs: Dict[str, Any] = field(default_factory=dict) - # TODO padding vocab - # make_vocab_size_divisible_by: int = 128 # pipeline_parallel_size: int # data_parallel_size: int # tensor_parallel_mode: Literal['1d', '2d', '2.5d', '3d'] diff --git a/colossalai/tensor/d_tensor/layout_converter.py b/colossalai/tensor/d_tensor/layout_converter.py index 667a7b78e4f5..c2cf73181345 100644 --- a/colossalai/tensor/d_tensor/layout_converter.py +++ b/colossalai/tensor/d_tensor/layout_converter.py @@ -10,6 +10,7 @@ from colossalai.tensor.d_tensor.comm_spec import * from colossalai.tensor.d_tensor.layout import Layout from colossalai.tensor.d_tensor.misc import LayoutException +from colossalai.tensor.padded_tensor.api import init_as_padded_tensor, is_padded_tensor from colossalai.tensor.utils import all_gather_simulator, all_to_all_simulator, shard_simulator from .sharding_spec import ShardingSpec @@ -607,8 +608,18 @@ def apply(self, tensor: torch.Tensor, source_layout: Layout, target_layout: Layo [3.], [3.]]) """ + _, comm_action_sequence = self.layout_converting(source_layout, target_layout) + + target_tensor = tensor for comm_spec in comm_action_sequence: - tensor = comm_spec.covert_spec_to_action(tensor) - tensor.dist_layout = target_layout - return tensor + target_tensor = comm_spec.covert_spec_to_action(target_tensor) + target_tensor.dist_layout = target_layout + + # restore the padding information + if is_padded_tensor(tensor) and not is_padded_tensor(target_tensor): + target_tensor = init_as_padded_tensor( + target_tensor, tensor._current_length, tensor._origin_length, tensor._padding_dim + ) + + return target_tensor diff --git a/colossalai/tensor/padded_tensor/__init__.py b/colossalai/tensor/padded_tensor/__init__.py new file mode 100644 index 000000000000..353ff35f84ca --- /dev/null +++ b/colossalai/tensor/padded_tensor/__init__.py @@ -0,0 +1,3 @@ +from .api import init_as_padded_tensor, is_padded_tensor, to_padded_tensor, to_unpadded_tensor + +__all__ = ["is_padded_tensor", "to_padded_tensor", "to_unpadded_tensor", "init_as_padded_tensor"] diff --git a/colossalai/tensor/padded_tensor/api.py b/colossalai/tensor/padded_tensor/api.py new file mode 100644 index 000000000000..5b66c016b399 --- /dev/null +++ b/colossalai/tensor/padded_tensor/api.py @@ -0,0 +1,128 @@ +import torch + + +def _hijack_detach_and_clone(ptensor: torch.Tensor) -> torch.Tensor: + """ + Hijack the detach and clone methods of the tensor to make sure the dist_layout is copied. + + Args: + tensor (torch.Tensor): The tensor to be hijacked. + + Returns: + torch.Tensor: The hijacked tensor. + """ + ptensor._unpad_detach = ptensor.detach + ptensor._unpad_clone = ptensor.clone + + def new_detach(self): + t_ = self._unpad_detach() + t_._padding_dim = self._padding_dim + t_._origin_length = self._origin_length + t_._current_length = self._current_length + return t_ + + def new_clone(self, *args, **kwargs): + t_ = self._unpad_clone(*args, **kwargs) + t_._padding_dim = self._padding_dim + t_._origin_length = self._origin_length + t_._current_length = self._current_length + return t_ + + # bind the new methods to the tensor + ptensor.detach = new_detach.__get__(ptensor) + ptensor.clone = new_clone.__get__(ptensor) + return ptensor + + +def _hijack_back_detach_and_clone(ptensor: torch.Tensor) -> torch.Tensor: + """ + Hijack the detach and clone methods of the tensor to make sure the dist_layout is copied. + + Args: + tensor (torch.Tensor): The tensor to be hijacked. + + Returns: + torch.Tensor: The hijacked tensor. + """ + ptensor.detach = ptensor._unpad_detach + ptensor.clone = ptensor._unpad_clone + + delattr(ptensor, "_unpad_detach") + delattr(ptensor, "_unpad_clone") + + return ptensor + + +def is_padded_tensor(tensor: torch.Tensor) -> bool: + """ + Check whether the given tensor is a padding tensor. + + Args: + tensor (torch.Tensor): The tensor to be checked. + + Returns: + bool: Whether the given tensor is a padding tensor. + """ + return hasattr(tensor, "_padding_dim") + + +def to_padded_tensor( + tensor: torch.Tensor, + current_length: int, + padding_dim: int, +) -> torch.Tensor: + assert ( + padding_dim < tensor.dim() + ), f"Please passing a valid padding_dim. the dimension of the tensor is {tensor.dim()}" + + if is_padded_tensor(tensor): + return tensor + + origin_length = tensor.shape[padding_dim] + padding_num = current_length - origin_length + padding_data = torch.zeros( + *tensor.shape[:padding_dim], + padding_num, + *tensor.shape[padding_dim + 1 :], + device=tensor.device, + dtype=tensor.dtype, + ) + tensor.data = torch.cat((tensor.data, padding_data), dim=padding_dim).contiguous() + + tensor._padding_dim = padding_dim + tensor._origin_length = origin_length + tensor._current_length = current_length + + _hijack_detach_and_clone(tensor) + + return tensor + + +def to_unpadded_tensor(ptensor: torch.Tensor): + if not is_padded_tensor(ptensor): + return ptensor + + unpad_slices = [slice(None)] * ptensor.dim() + unpad_slices[ptensor._padding_dim] = slice(None, ptensor._origin_length) + ptensor.data = ptensor.data[tuple(unpad_slices)] + + delattr(ptensor, "_padding_dim") + delattr(ptensor, "_origin_length") + delattr(ptensor, "_current_length") + + _hijack_back_detach_and_clone(ptensor) + + return ptensor + + +def init_as_padded_tensor(tensor: torch.Tensor, current_length: int, origin_length: int, padding_dim: int): + if is_padded_tensor(tensor): + return tensor + + tensor._padding_dim = padding_dim + tensor._origin_length = origin_length + tensor._current_length = current_length + + _hijack_detach_and_clone(tensor) + + return tensor diff --git a/colossalai/testing/comparison.py b/colossalai/testing/comparison.py index e415b5fc3aa3..bdf7b19f39d0 100644 --- a/colossalai/testing/comparison.py +++ b/colossalai/testing/comparison.py @@ -23,7 +23,7 @@ def assert_close_loose(a: Tensor, b: Tensor, rtol: float = 1e-3, atol: float = 1 rtol=rtol, atol=atol, msg=f"Tensor not close, shape: {a.shape} vs {b.shape}, \ - dtype: {a.dtype} vs {b.dtype}", + dtype: {a.dtype} vs {b.dtype}", ) diff --git a/colossalai/zero/gemini/gemini_ddp.py b/colossalai/zero/gemini/gemini_ddp.py index bc6c9d088094..c79422171f1b 100644 --- a/colossalai/zero/gemini/gemini_ddp.py +++ b/colossalai/zero/gemini/gemini_ddp.py @@ -27,6 +27,12 @@ is_customized_distributed_tensor, is_distributed_tensor, ) +from colossalai.tensor.padded_tensor import ( + init_as_padded_tensor, + is_padded_tensor, + to_padded_tensor, + to_unpadded_tensor, +) from colossalai.tensor.param_op_hook import ColoParamOpHookManager from colossalai.utils import _cast_float, free_storage, is_ddp_ignored @@ -460,6 +466,11 @@ def _get_chunk_to_save_data(self, chunk: Chunk, only_rank_0: bool) -> Dict: record_tensor, shard_fn=tensor.shard_fn, gather_fn=tensor.gather_fn ) record_tensor = gather_distributed_param(record_tensor, keep_vars=False).cpu() + if is_padded_tensor(tensor): + record_tensor = init_as_padded_tensor( + record_tensor, tensor._current_length, tensor._origin_length, tensor._padding_dim + ) + record_tensor = to_unpadded_tensor(record_tensor) assert tensor not in chunk_to_save_data chunk_to_save_data[tensor] = record_tensor @@ -520,6 +531,8 @@ def _save_to_state_dict(self, destination, prefix, keep_vars, only_rank_0=True): # deal with ddp ignored parameters destination[prefix + name] = param if keep_vars else param.detach() else: + if is_padded_tensor(p_mapping[param]): + p_mapping[param] = to_unpadded_tensor(p_mapping[param]) destination[prefix + name] = p_mapping[param] del p_mapping del param_to_save_data @@ -627,6 +640,7 @@ def _load_from_state_dict( list, and will be reported together in :meth:`~torch.nn.Module.load_state_dict` """ + for hook in self._load_state_dict_pre_hooks.values(): hook(state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs) @@ -647,6 +661,14 @@ def load( if state_key in state_dict: input_param = state_dict[state_key] + global_shape = dest_tensor.shape + if source_device_mesh is not None and source_sharding_spec is not None: + global_shape = get_global_shape(dest_tensor) + + if is_padded_tensor(dest_tensor): + padding_dim = dest_tensor._padding_dim + input_param = to_padded_tensor(input_param, global_shape[padding_dim], padding_dim) + if source_device_mesh is not None and source_sharding_spec is not None: input_param = distribute_tensor(input_param, source_device_mesh, source_sharding_spec) elif shard_fn is not None and gather_fn is not None: diff --git a/colossalai/zero/gemini/gemini_optimizer.py b/colossalai/zero/gemini/gemini_optimizer.py index 18367af59d80..ae02fe297d88 100644 --- a/colossalai/zero/gemini/gemini_optimizer.py +++ b/colossalai/zero/gemini/gemini_optimizer.py @@ -21,12 +21,19 @@ distribute_tensor, distribute_tensor_with_customization, get_device_mesh, + get_global_shape, get_sharding_spec, init_as_dtensor, init_tensor_as_customization_distributed, is_customized_distributed_tensor, is_distributed_tensor, ) +from colossalai.tensor.padded_tensor import ( + init_as_padded_tensor, + is_padded_tensor, + to_padded_tensor, + to_unpadded_tensor, +) from colossalai.utils import disposable, is_ddp_ignored from .chunk import Chunk, ChunkManager @@ -106,7 +113,7 @@ def __init__( max_norm: float = 0.0, norm_type: float = 2.0, tp_group: ProcessGroup = None, - optimizer_params_info=None, + params_info=None, verbose: bool = False, **defaults: Any, ): @@ -124,7 +131,7 @@ def __init__( self.clipping_flag = max_norm > 0.0 self.max_norm = max_norm self.tp_group = tp_group - self.optimizer_params_info = optimizer_params_info + self.params_info = params_info self.tp_size = dist.get_world_size(tp_group) if tp_group is not None else 1 self.tp_rank = dist.get_rank(tp_group) if tp_group is not None else 0 self.verbose = verbose @@ -459,7 +466,7 @@ def collect_states(self, param_id: int, only_rank_0: bool = True) -> dict: is_customized_distributed = is_customized_distributed_tensor(param) shard_spec = get_sharding_spec(param) if is_dtensor else None device_mesh = get_device_mesh(param) if is_dtensor else None - global_shape = self.optimizer_params_info["id2shape"][param_id] + global_shape = self.params_info["id2shape"][param_id] # If the chunk is kept gathered, # the parameters are treated the same as that of those in strict DDP during training. @@ -477,6 +484,7 @@ def collect_states(self, param_id: int, only_rank_0: bool = True) -> dict: else: state_tensor = states[state_name].detach().clone().to(torch.float32).cpu() if is_dtensor: + global_shape = get_global_shape(param) state_tensor = torch.reshape(state_tensor, param.shape).to(param.device) state_tensor = init_as_dtensor( state_tensor, @@ -490,8 +498,13 @@ def collect_states(self, param_id: int, only_rank_0: bool = True) -> dict: state_tensor, shard_fn=param.shard_fn, gather_fn=param.gather_fn ) state_tensor = gather_distributed_param(state_tensor, keep_vars=False).cpu() - - collected_states[state_name] = state_tensor.reshape(global_shape) + state_tensor = state_tensor.reshape(global_shape) + if is_padded_tensor(param): + state_tensor = init_as_padded_tensor( + state_tensor, param._current_length, param._origin_length, param._padding_dim + ) + state_tensor = to_unpadded_tensor(state_tensor) + collected_states[state_name] = state_tensor return collected_states # Check whether the param with given id is managed by current process. @@ -535,6 +548,7 @@ def collect_states(self, param_id: int, only_rank_0: bool = True) -> dict: if state_tensor.numel() == param.numel(): collected_states[state_name] = torch.reshape(state_tensor, param.shape) if is_dtensor: + global_shape = get_global_shape(param) state_tensor = state_tensor.to(param.device) state_tensor = init_as_dtensor( state_tensor, sharding_spec=shard_spec, device_mesh=device_mesh, global_shape=global_shape @@ -545,6 +559,11 @@ def collect_states(self, param_id: int, only_rank_0: bool = True) -> dict: state_tensor, shard_fn=param.shard_fn, gather_fn=param.gather_fn ) state_tensor = gather_distributed_param(state_tensor, keep_vars=False).cpu() + if is_padded_tensor(param): + state_tensor = init_as_padded_tensor( + state_tensor, param._current_length, param._origin_length, param._padding_dim + ) + state_tensor = to_unpadded_tensor(state_tensor) return collected_states @@ -698,7 +717,7 @@ def load_single_param_states(self, param_id: int, saved_states: dict): Load saved optimizer states into parameter with given id. """ - def cast(param, state_range, value, key=None): + def cast(param, state_range, value, global_shape, origin_shape, key=None): """ Make a copy of the needed segment of value and cast it to device of param. """ @@ -714,7 +733,14 @@ def cast(param, state_range, value, key=None): ) if is_dtensor: - value = torch.reshape(value, global_shape) + global_shape = get_global_shape(real_param) + + if is_padded_tensor(real_param): + value = torch.reshape(value, origin_shape) + padding_dim = real_param._padding_dim + value = to_padded_tensor(value, global_shape[padding_dim], padding_dim) + + if is_dtensor: value = distribute_tensor(value, sharding_spec=shard_spec, device_mesh=device_mesh) elif is_customized_distributed: value = torch.reshape(value, global_shape) @@ -737,10 +763,11 @@ def cast(param, state_range, value, key=None): is_customized_distributed = is_customized_distributed_tensor(real_param) shard_spec = get_sharding_spec(real_param) if is_dtensor else None device_mesh = get_device_mesh(real_param) if is_dtensor else None - global_shape = self.optimizer_params_info["id2shape"][param_id] + global_shape = self.params_info["id2shape"][param_id] + origin_shape = global_shape for k, v in saved_states.items(): - updated_states[k] = cast(fake_param, state_range, v, k) + updated_states[k] = cast(fake_param, state_range, v, global_shape, origin_shape, k) del v # clean loaded states self.optim.state[fake_param].update(updated_states) diff --git a/tests/test_checkpoint_io/test_hybrid_parallel_plugin_checkpoint_io.py b/tests/test_checkpoint_io/test_hybrid_parallel_plugin_checkpoint_io.py index d8a625b98a66..4753ab637f01 100644 --- a/tests/test_checkpoint_io/test_hybrid_parallel_plugin_checkpoint_io.py +++ b/tests/test_checkpoint_io/test_hybrid_parallel_plugin_checkpoint_io.py @@ -81,8 +81,7 @@ def _preprocess_data(data): optimizer.backward(loss) optimizer.step() - for group in optimizer.param_groups: - group["lr"] = 0.1 + optimizer.zero_grad() with shared_tempdir() as tempdir: model_ckpt_path = f"{tempdir}/model" optimizer_ckpt_path = f"{tempdir}/optimizer" diff --git a/tests/test_optimizer/test_nvme.py b/tests/test_optimizer/test_nvme.py index 603b7b6fa325..4ff16bb9b7c9 100644 --- a/tests/test_optimizer/test_nvme.py +++ b/tests/test_optimizer/test_nvme.py @@ -1,4 +1,3 @@ -import pytest import torch from colossalai.nn.optimizer import CPUAdam, HybridAdam @@ -18,8 +17,6 @@ def check_params_equal(model, torch_model): assert torch.allclose(p, torch_p, atol=1e-3), f"diff: {torch.abs(p - torch_p)}" -# TODO Something wrong with ci when running this test. -@pytest.mark.skip(reason="skip because of something wrong with CI") @clear_cache_before_run() @parameterize("nvme_offload_fraction", [0.0, 0.5, 1.0]) @parameterize("nvme_offload_dir", ["./offload", None]) diff --git a/tests/test_shardformer/test_layer/test_vocab_parallel_embedding_1d.py b/tests/test_shardformer/test_layer/test_vocab_parallel_embedding_1d.py index b23a44f2dffa..91cc1a987a29 100644 --- a/tests/test_shardformer/test_layer/test_vocab_parallel_embedding_1d.py +++ b/tests/test_shardformer/test_layer/test_vocab_parallel_embedding_1d.py @@ -21,7 +21,7 @@ def check_vocab_embedding_1d(lazy_init: bool): dist_embedding_1d = VocabParallelEmbedding1D.from_native_module(embedding_copy, process_group=None) assert dist_embedding_1d.weight.shape == torch.Size([64, 32]) - assert dist_embedding_1d.num_embeddings == 64 + assert dist_embedding_1d.num_embeddings == 128 assert dist_embedding_1d.embedding_dim == 32 assert embedding_copy.weight is dist_embedding_1d.weight diff --git a/tests/test_shardformer/test_model/_utils.py b/tests/test_shardformer/test_model/_utils.py index d5fc2c30f294..a77ba39a122c 100644 --- a/tests/test_shardformer/test_model/_utils.py +++ b/tests/test_shardformer/test_model/_utils.py @@ -14,12 +14,14 @@ from colossalai.booster import Booster from colossalai.booster.plugin import HybridParallelPlugin from colossalai.booster.plugin.hybrid_parallel_plugin import HybridParallelModule +from colossalai.checkpoint_io.utils import gather_distributed_param from colossalai.lazy import LazyInitContext from colossalai.pipeline.stage_manager import PipelineStageManager from colossalai.shardformer import ShardConfig, ShardFormer from colossalai.shardformer._utils import getattr_ from colossalai.shardformer.policies.auto_policy import Policy from colossalai.tensor.d_tensor.api import is_customized_distributed_tensor, is_distributed_tensor +from colossalai.tensor.padded_tensor.api import is_padded_tensor, to_unpadded_tensor def build_model( @@ -247,11 +249,10 @@ def check_weight( continue if is_distributed_tensor(sharded_weight) or is_customized_distributed_tensor(sharded_weight): - sharded_weight_list = [ - torch.zeros_like(sharded_weight).to("cuda") for _ in range(dist.get_world_size(tp_group)) - ] - dist.all_gather(sharded_weight_list, sharded_weight, tp_group) - sharded_weight = torch.cat(sharded_weight_list, dim=dim) + sharded_weight = gather_distributed_param(sharded_weight, keep_vars=False) + + if is_padded_tensor(sharded_weight): + sharded_weight = to_unpadded_tensor(sharded_weight) if verbose and dist.get_rank() == 0: print(f"'{suffix}' weight: {org_weight}, {sharded_weight}") diff --git a/tests/test_shardformer/test_model/test_shard_t5.py b/tests/test_shardformer/test_model/test_shard_t5.py index 9b22d54d7d31..a6fe2dd39383 100644 --- a/tests/test_shardformer/test_model/test_shard_t5.py +++ b/tests/test_shardformer/test_model/test_shard_t5.py @@ -73,7 +73,8 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn, # check weights if test_config["precision"] == "fp32": - atol, rtol = 5e-4, 1e-3 + # TODO he precision in weight checking is too significant. + atol, rtol = 1e-3, 1e-3 else: atol, rtol = 5e-3, 5e-3 if stage_manager is None or stage_manager.is_first_stage(): diff --git a/tests/test_tensor/test_padded_tensor.py b/tests/test_tensor/test_padded_tensor.py new file mode 100644 index 000000000000..31a267c15286 --- /dev/null +++ b/tests/test_tensor/test_padded_tensor.py @@ -0,0 +1,46 @@ +import torch + +from colossalai.device.device_mesh import DeviceMesh +from colossalai.initialize import launch +from colossalai.logging import disable_existing_loggers +from colossalai.tensor.d_tensor import ShardingSpec, distribute_tensor, is_distributed_tensor, to_global +from colossalai.tensor.padded_tensor import is_padded_tensor, to_padded_tensor, to_unpadded_tensor +from colossalai.testing import rerun_if_address_is_in_use, spawn + + +def check_padded_tensor(rank, world_size, port): + disable_existing_loggers() + launch(config={}, rank=rank, world_size=world_size, host="localhost", port=port, backend="nccl") + original_tensor = torch.rand(32, 64).to("cuda") + + device_mesh = DeviceMesh(torch.Tensor([0, 1, 2, 3]), (2, 2), init_process_group=True) + target_sharding_spec = ShardingSpec(dim_size=original_tensor.dim(), dim_partition_dict={0: [0]}) + d_tensor = distribute_tensor(original_tensor, device_mesh, target_sharding_spec) + + padded_tensor = to_padded_tensor(d_tensor, current_length=64, padding_dim=0) + assert padded_tensor.dist_layout == d_tensor.dist_layout + + tensor_copy = padded_tensor.clone() + assert is_padded_tensor(tensor_copy) + assert is_distributed_tensor(tensor_copy) + + tensor_detached = padded_tensor.detach() + assert is_padded_tensor(tensor_detached) + assert is_distributed_tensor(tensor_detached) + + unpadded_tensor = to_unpadded_tensor(padded_tensor) + assert unpadded_tensor.shape == d_tensor.shape + assert is_distributed_tensor(unpadded_tensor) + + global_tensor = to_global(unpadded_tensor) + assert global_tensor.shape == original_tensor.shape + + +@rerun_if_address_is_in_use() +def test_padded_tensor(): + world_size = 4 + spawn(check_padded_tensor, world_size) + + +if __name__ == "__main__": + test_padded_tensor() From 67d0d86a3b915ef8e764e19bab7a1d75871fb318 Mon Sep 17 00:00:00 2001 From: "pre-commit-ci[bot]" <66853113+pre-commit-ci[bot]@users.noreply.github.com> Date: Wed, 17 Apr 2024 07:58:27 +0000 Subject: [PATCH 5/7] [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci --- applications/ColossalMoE/README.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/applications/ColossalMoE/README.md b/applications/ColossalMoE/README.md index 41b47c756119..c3c214789f54 100644 --- a/applications/ColossalMoE/README.md +++ b/applications/ColossalMoE/README.md @@ -36,4 +36,4 @@ Then yon can use colossalai run to launch train: ```bash bash train.sh ``` -It requires 16 H100 (80G) to run the training. The number of GPUs should be divided by 8. If you already have downloaded model weights, you can change name to your weights position in `train.sh`. \ No newline at end of file +It requires 16 H100 (80G) to run the training. The number of GPUs should be divided by 8. If you already have downloaded model weights, you can change name to your weights position in `train.sh`. From c204fa69d5b0dfe8ad1e3bd041f3621ccd314d4f Mon Sep 17 00:00:00 2001 From: flybird11111 <1829166702@qq.com> Date: Thu, 18 Apr 2024 06:18:17 +0000 Subject: [PATCH 6/7] fix rebase --- colossalai/shardformer/policies/chatglm2.py | 14 -------------- .../test_model/test_shard_chatglm2.py | 10 +++++----- 2 files changed, 5 insertions(+), 19 deletions(-) diff --git a/colossalai/shardformer/policies/chatglm2.py b/colossalai/shardformer/policies/chatglm2.py index eb5597005d65..f205835e7815 100644 --- a/colossalai/shardformer/policies/chatglm2.py +++ b/colossalai/shardformer/policies/chatglm2.py @@ -66,20 +66,6 @@ def module_policy(self) -> Dict[Union[str, nn.Module], ModulePolicyDescription]: overlap = self.shard_config.enable_sequence_overlap sp_partial_derived = sp_mode == "split_gather" - if self.shard_config.enable_tensor_parallelism: - policy[ChatGLMModel] = ModulePolicyDescription( - attribute_replacement={}, - sub_module_replacement=[ - SubModuleReplacementDescription( - suffix="embedding.word_embeddings", - target_module=col_nn.VocabParallelEmbedding1D, - ) - ], - ) - sp_mode = "split_gather" - overlap = self.shard_config.enable_sequence_overlap - sp_partial_derived = sp_mode == "split_gather" - if self.shard_config.enable_tensor_parallelism: policy[GLMBlock] = ModulePolicyDescription( attribute_replacement={ diff --git a/tests/test_shardformer/test_model/test_shard_chatglm2.py b/tests/test_shardformer/test_model/test_shard_chatglm2.py index 405ceba328df..d2cd019307eb 100644 --- a/tests/test_shardformer/test_model/test_shard_chatglm2.py +++ b/tests/test_shardformer/test_model/test_shard_chatglm2.py @@ -288,11 +288,11 @@ def test_chatglm(): spawn(check_chatglm, 4) -@pytest.mark.largedist -@rerun_if_address_is_in_use() -@clear_cache_before_run() -def test_chatglm_3d(): - spawn(check_chatglm_3d, 8) +# @pytest.mark.largedist +# @rerun_if_address_is_in_use() +# @clear_cache_before_run() +# def test_chatglm_3d(): +# spawn(check_chatglm_3d, 8) if __name__ == "__main__": From c5a48f0e3c9c5616eccd17a400bb9599c7f926f3 Mon Sep 17 00:00:00 2001 From: flybird11111 <1829166702@qq.com> Date: Thu, 18 Apr 2024 06:18:40 +0000 Subject: [PATCH 7/7] fix rebase --- .../test_shardformer/test_model/test_shard_chatglm2.py | 10 +++++----- 1 file changed, 5 insertions(+), 5 deletions(-) diff --git a/tests/test_shardformer/test_model/test_shard_chatglm2.py b/tests/test_shardformer/test_model/test_shard_chatglm2.py index d2cd019307eb..405ceba328df 100644 --- a/tests/test_shardformer/test_model/test_shard_chatglm2.py +++ b/tests/test_shardformer/test_model/test_shard_chatglm2.py @@ -288,11 +288,11 @@ def test_chatglm(): spawn(check_chatglm, 4) -# @pytest.mark.largedist -# @rerun_if_address_is_in_use() -# @clear_cache_before_run() -# def test_chatglm_3d(): -# spawn(check_chatglm_3d, 8) +@pytest.mark.largedist +@rerun_if_address_is_in_use() +@clear_cache_before_run() +def test_chatglm_3d(): + spawn(check_chatglm_3d, 8) if __name__ == "__main__":