Fix random token-generation issue + MP-checkpoint loading/saving#2132
Merged
Fix random token-generation issue + MP-checkpoint loading/saving#2132
Conversation
RezaYazdaniAminabadi
requested review from
ShadenSmith,
arashb,
awan-10,
cli99,
conglongli,
duli2012,
eltonzheng,
jeffra,
minjiaz,
mrwyattii,
samadejacobs,
samyam,
tjruwase,
xiaoxiawu-microsoft and
yaozhewei
as code owners
…ed into ds-inference/bloom-fix
…ed into ds-inference/bloom-fix
Contributor
Author
Just fixed it, please give it a try |
|
Looks like I need to add a new key to my checkpoint.json? Is it mandatory? What value should I put in it for the huggingface checkpoint file list? EDIT: I looked at the code and set it to |
|
After getting past the The My new checkpoints json file looks like this: {"type": "BLOOM-176B",
"base_dir": "/home/ubuntu/.cache/deepspeed/bigscience/bloom",
"checkpoints": ["BLOOM-176B-non-tp.pt", "BLOOM-176B-non-tp.pt", "BLOOM-176B-non-tp.pt", "BLOOM-176B-non-tp.pt", "BLOOM-176B-non-tp.pt", "BLOOM-176B-non-tp.pt", "BLOOM-176B-non-tp.pt", "BLOOM-176B-non-tp.pt", "BLOOM-176B-tp_00.pt", "BLOOM-176B-tp_01.pt", "BLOOM-176B-tp_02.pt", "BLOOM-176B-tp_03.pt", "BLOOM-176B-tp_04.pt", "BLOOM-176B-tp_05.pt", "BLOOM-176B-tp_06.pt", "BLOOM-176B-tp_07.pt"],
"version": 1.0,
"parallelization": "tp",
"mp_size": 8}EDIT: I found the |
zcrypt0
reviewed
Jul 28, 2022
| v in dict(replaced_module.state_dict()).items() | ||
| if transformer_name not in k | ||
| }), | ||
| non_tp_ckpt_name) |
There was a problem hiding this comment.
f'{save_mp_checkpoint_path}/{non_tp_ckpt_name}'
Contributor
Author
There was a problem hiding this comment.
that's true, it's not saved correctly. I am gonna fix it now
Contributor
Author
There was a problem hiding this comment.
okay, I just fixed it, you should see the following files and sizes under the save_path:
Contributor
Author
There was a problem hiding this comment.
@zcrypt0, it also generates a config file under the same path that you can use to run inference with
There was a problem hiding this comment.
EDIT: I just realized the change in the non-tp file size, I will give it a try soon
There was a problem hiding this comment.
@RezaYazdaniAminabadi Just tested and it works without a hitch, nice! 👍
…ed into ds-inference/bloom-fix
jeffra
approved these changes
Jul 29, 2022
Contributor
|
Still getting this error @RezaYazdaniAminabadi running with batch size = 1 |
Contributor
Ran this with CUDA 11.6 with DeepSpeed on master branch. |
Contributor
|
@RezaYazdaniAminabadi how much time is cached tp model loading supposed to take? self.model is loaded using HF AutoModel as in bloom-ds-inference.py |
Contributor
|
nvm |
Contributor
|
@RezaYazdaniAminabadi I am seeing again after updating to master branch and saving without providing checkpoint json |
Collaborator
Just want to double check, did your install include this commit in master? #2237 |
Contributor
|
@jeffra yes I am on the latest commit |
Contributor
|
I use this code: run with : deepspeed --num_gpus 8 scripts/bloom-inference-server/cache_ds_checkpoints.py --model_name bigscience/bloom --dtype fp16 --save_mp_checkpoint_path ../DS_cacheimport argparse
import os
import deepspeed
import torch
from transformers import AutoConfig, AutoModelForCausalLM
def get_args() -> argparse.Namespace:
parser = argparse.ArgumentParser()
group = parser.add_argument_group(title="launch config")
group.add_argument("--local_rank", required=False,
type=int, help="used by dist launchers")
group.add_argument("--save_mp_checkpoint_path", required=True,
type=str, help="MP checkpoints path for DS inference")
group = parser.add_argument_group(title="model")
group.add_argument("--model_name", type=str,
required=True, help="model to use")
group.add_argument("--dtype", type=str, required=True,
choices=["bf16", "fp16"], help="dtype for model")
args = parser.parse_args()
if (args.dtype == "bf16"):
args.dtype = torch.bfloat16
elif (args.dtype == "fp16"):
args.dtype = torch.float16
return args
def main() -> None:
args = get_args()
if (args.local_rank == 0):
print("Loading model...")
world_size = int(os.getenv("WORLD_SIZE", "1"))
# Load model
with deepspeed.OnDevice(dtype=args.dtype, device="meta"):
model = AutoModelForCausalLM.from_config(
AutoConfig.from_pretrained(args.model_name),
torch_dtype=torch.bfloat16
)
model = model.eval()
if (args.dtype == torch.float16):
model = deepspeed.init_inference(
model,
mp_size=world_size,
dtype=args.dtype,
replace_with_kernel_inject=True,
save_mp_checkpoint_path=args.save_mp_checkpoint_path
)
elif (args.dtype == torch.bfloat16):
raise NotImplementedError("bfloat16 is not yet supported")
print("Model loaded")
if (__name__ == "__main__"):
main() |
Contributor
|
@jeffra This issue is blocking bigscience-workshop/Megatron-DeepSpeed#328 |
Contributor
|
@jeffra ^^ |
@mayank31398 Hi, also facing this issue. Is it solved right now? |
Contributor
|
@pai4451 not yet |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
6 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
This PR fixes the token-generation issue with different random seed on several MP ranks. It also adds the ability to load/save MP-partitioned checkpoints to speed up the checkpoint loading for inference.
cc: @stas00 @jeffra