Fix issue #39191 respect accelerate config to disable torch.dynamo compilation

[bonpiedlaroute]

Description

Fixes #39191 where transformers ignored accelerate configuration to disable torch.dynamo, leading to unexpected compilation and FailOnRecompileLimitHit errors.

Problem

When users configure accelerate to disable torch.dynamo, transformers' automatic compilation in _valid_auto_compile_criteria() was not respecting this setting, causing:

• Unwanted torch.compile activation
• Excessive recompilations
• FailOnRecompileLimitHit crashes in distributed training scenarios

Solution

• Added _is_dynamo_compilation_disabled() method that checks standard environment variables
• Modified _valid_auto_compile_criteria() to respect these environment variables before enabling compilation

Environment Variables Supported

• TORCHDYNAMO_DISABLE=1

Tests done

• Compilation disabled when env var set
• Normal behavior preserved when no env var
• Backward compatible - no breaking changes

Usage

Users experiencing the issue can now use:

export TORCHDYNAMO_DISABLE=1
python training_script.py

Who can review?

@SunMarc @gante @zach-huggingface and @qgallouedec

[gante]

Thank you for opening the PR!

LGTM, added a question/small request to clear out before merging :)

[gante]

This is not a standard env var, correct? If not, then let's remove it 🤗 (let's stick to the known flags and avoid adding more noise, which may have unintended consequences)

[bonpiedlaroute]

You're absolutely right! After double checking pytorch source code, the official flag is: TORCHDYNAMO_DISABLE
PR updated.
Thanks,
Noël

[bonpiedlaroute]

Please note that the failing test tests/models/perceiver/test_modeling_perceiver.py::PerceiverModelTest::test_pipeline_image_classification_fp16 - PIL.UnidentifiedImageError: cannot identify image file
is unrelated to this change. It seems to be an image file corruption or download issue, no link with torch.dynamo compilation fix.
All other pipeline tests (148) passed successfully
Thanks,
Noël

[SunMarc]

Please note that the failing test tests/models/perceiver/test_modeling_perceiver.py::PerceiverModelTest::test_pipeline_image_classification_fp16 - PIL.UnidentifiedImageError: cannot identify image file
is unrelated to this change. It seems to be an image file corruption or download issue, no link with torch.dynamo compilation fix.

Yeah don't worry ! I'll rerun the CI to merge this

[SunMarc]

Thanks !

[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.

[SunMarc]

@bot /style

[github-actions]

Style fix is beginning .... View the workflow run here.

[github-actions]

View the CircleCI Test Summary for this PR:

https://huggingface.co/spaces/transformers-community/circle-ci-viz?pr=39683&sha=824bdc