XLMRoberta with Flash Attention 2

[IvanPy96]

System Info

• transformers version: 4.36.0
• Platform: Linux-4.19.0-22-amd64-x86_64-with-glibc2.31
• Python version: 3.10.13
• Huggingface_hub version: 0.19.4
• Safetensors version: 0.4.0
• Accelerate version: 0.24.1
• Accelerate config: not found
• PyTorch version (GPU?): 2.0.1+cu117 (True)
• Tensorflow version (GPU?): not installed (NA)
• Flax version (CPU?/GPU?/TPU?): not installed (NA)
• Jax version: not installed
• JaxLib version: not installed
• Using GPU in script?:
• Using distributed or parallel set-up in script?:

Who can help?

@ArthurZucker @younesbelkada

Information

• The official example scripts
• My own modified scripts

Tasks

• An officially supported task in the examples folder (such as GLUE/SQuAD, ...)
• My own task or dataset (give details below)

Reproduction

from transformers import AutoModelForSequenceClassification

model = AutoModelForSequenceClassification.from_pretrained("my_model/", attn_implementation="flash_attention_2")

Expected behavior

Ability to use flash attention 2 for inference. Is it possible to add support of flash attention 2 for XLMRoberta model?

[ArthurZucker]

Thanks for opening, will mark as a good second issue 🤗

[mohammedElfatihSalah]

Hi @IvanPy96 & @ArthurZucker I want to work on this issue. Could you please assign it to me?

[ArthurZucker]

Hey, we don't assign issue, feel free to open a PR and link it to this issue 😉

[aikangjun]

Hi, it seems that this issue has not been resolved ,XLMRoberta still cannot use FlashAttention 2.

[ArthurZucker]

Hey! Yes as both PR were closed: see the last comment

@aikangjun This PR wasn't merged - it closed because of inactivity it seems. We've recently merged in other PRs to add SDPA to roberta based models though #30510 which adds it to this model. This isn't part of 4.42 but will be part of the next release