Kernels flash attn

[ArthurZucker]

What does this PR do?

pip install transformers[torch] kernels

from transformers import AutoModelForCausalLM, AutoTokenizer
import time

model = AutoModelForCausalLM.from_pretrained(
    "meta-llama/Llama-3.2-3B-Instruct",
    device_map="auto",
    torch_dtype="auto",
    attn_implementation="flash_attention_2",
).eval()
tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-3.2-3B-Instruct")
tokenizer.pad_token = tokenizer.eos_token
inputs = tokenizer(
    ["Hello, how are you?", "is this life?"],
    padding=True,
    padding_side="left",
    return_tensors="pt",
).to(model.device)


start = time.time()
outputs = model.generate(**inputs, max_new_tokens=50)
print(f"Generation time: {time.time() - start:.2f} seconds")
print(tokenizer.batch_decode(outputs, skip_special_tokens=True))

model.set_attn_implementation("kernels-community/flash-attn3")
start = time.time()
outputs = model.generate(**inputs, max_new_tokens=50)
print(f"Generation time: {time.time() - start:.2f} seconds")
print(tokenizer.batch_decode(outputs, skip_special_tokens=True))

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[kadirnar]

tf install:

uv pip install git+https://github.com/huggingface/transformers.git@kernels-flash-attn

env:

Copy-and-paste the text below in your GitHub issue and FILL OUT the two last points.

- `transformers` version: 4.54.0.dev0
- Platform: Linux-6.11.0-29-generic-x86_64-with-glibc2.40
- Python version: 3.10.16
- Huggingface_hub version: 0.33.4
- Safetensors version: 0.5.3
- Accelerate version: 1.9.0
- Accelerate config:    - compute_environment: LOCAL_MACHINE

Error Message:

[rank0]: ValueError: Specified `attn_implementation="https://huggingface.co/kernels-community/flash-attn3:flash_attention"` is not supported. The only possible arguments are `attn_implementation="eager"` (manual attention implementation), `"attn_implementation=flash_attention_3"` (implementation using flash attention 3), `"attn_implementation=flash_attention_2"` (implementation using flash attention 2), `"attn_implementation=sdpa"` (implementation using torch.nn.functional.scaled_dot_product_attention), `"attn_implementation=flex_attention"` (implementation using torch's flex_attention).

[ArthurZucker]

You can't pass the full http! You need to pass kernels-community/flash-attn3:flash_attention

[kadirnar]

You can't pass the full http! You need to pass kernels-community/flash-attn3:flash_attention

I think writing the URL is silly too. However, since you shared it like this on Twitter, I gave it a try.
https://x.com/art_zucker/status/1945821883858915695

New Error Message:

[rank1]:     cache_position = torch.arange(
[rank1]: RuntimeError: CUDA error: device-side assert triggered
[rank1]: CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect.
[rank1]: For debugging consider passing CUDA_LAUNCH_BLOCKING=1
[rank1]: Compile with `TORCH_USE_CUDA_DSA` to enable device-side assertions.

] Assertion `srcIndex < srcSelectDimSize,0,0: indexSelectLargeIndex:1553: indexSelectLargeIndex` failed.
,0,0: block: [237: indexSelectLargeIndex: block: [569/pytorch/aten/src/ATen/native/cuda/Indexing.cu], thread: [96], thread: [69,0: block: [173,0:1553,0,0,0,0,0: indexSelectLargeIndex,0,0], thread: [49,0], thread: [62: block: [844] Assertion `srcIndex < srcSelectDimSize] Assertion `srcIndex < srcSelectDimSize,0], thread: [45,0,0` failed.
` failed.

Should I wait for you to finish your development?

[ArthurZucker]

Ah that's weird can you share a small reproducer?

[ArthurZucker]

run-slow: llama,mistral,gemma

[github-actions]

This comment contains run-slow, running the specified jobs:

models: ['models/gemma', 'models/llama', 'models/mistral']
quantizations: [] ...

[kadirnar]

@ArthurZucker I tried it with a different LLM model, and it worked. It seems that the dataset of the Qwen model is faulty. I will fix this and provide feedback on the performance.

[ArthurZucker]

Thanks @kadirnar !

[kadirnar]

@ArthurZucker This method only supports LLM models, right? What should we do to add kernel support for speech models?

Example: https://huggingface.co/docs/transformers/main/en/model_doc/dia

[ArthurZucker]

This should be supported by all models as long as they have a the ALL_ATTENTION_FUNCTION refactor + you can set the attention for sub_modules!