Skip to content
Permalink

Comparing changes

Choose two branches to see what’s changed or to start a new pull request. If you need to, you can also or learn more about diff comparisons.

Open a pull request

Create a new pull request by comparing changes across two branches. If you need to, you can also . Learn more about diff comparisons here.
base repository: huggingface/transformers
Failed to load repositories. Confirm that selected base ref is valid, then try again.
Loading
base: main
Choose a base ref
...
head repository: huggingface/transformers
Failed to load repositories. Confirm that selected head ref is valid, then try again.
Loading
compare: debug_gc_old_base
Choose a head ref
Checking mergeability… Don’t worry, you can still create the pull request.
  • 14 commits
  • 3 files changed
  • 2 contributors

Commits on Aug 26, 2026

  1. debug gc

    @ydshieh
    ydshieh committed Aug 26, 2026
    Configuration menu
    Copy the full SHA
    b80b4a3 View commit details
    Browse the repository at this point in the history

  2. [debug] BigBirdPegasus gradient checkpointing flakiness investigation 

    Add debug logging to check_training_gradient_checkpointing override in
BigBirdPegasusModelTest to track how often non_zero_grads_gradcp differs
from non_zero_grads_normal across 1000 iterations. Instead of failing on
the first mismatch, collect all mismatches and report the count at the end.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
    @ydshieh @claude
    ydshieh and claude committed Aug 26, 2026
    Configuration menu
    Copy the full SHA
    e444c50 View commit details
    Browse the repository at this point in the history

  3. [debug] Skip flaky gradient checkpointing reentrant tests in BigBirdP… 

    …egasus

test_training_gradient_checkpointing_use_reentrant_true and _false are
flaky on CI: qa_outputs.bias shows non-zero grad in the normal run but
zero in the GC run (or vice versa), due to floating-point rounding
differences between the two backward paths.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
    @ydshieh @claude
    ydshieh and claude committed Aug 26, 2026
    Configuration menu
    Copy the full SHA
    5794ecc View commit details
    Browse the repository at this point in the history

  4. Revert test_modeling_common.py to main (restore removed GC test methods) 

    Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
    @ydshieh @claude
    ydshieh and claude committed Aug 26, 2026
    Configuration menu
    Copy the full SHA
    d01ba35 View commit details
    Browse the repository at this point in the history

  5. debug gc

    @ydshieh
    ydshieh committed Aug 26, 2026
    Configuration menu
    Copy the full SHA
    e7b2595 View commit details
    Browse the repository at this point in the history

  6. Add missing import unittest.mock

    @ydshieh
    ydshieh committed Aug 26, 2026
    Configuration menu
    Copy the full SHA
    6b2b93b View commit details
    Browse the repository at this point in the history

  7. Point PR CI caller to transformers-ci debug_gc branch

    @ydshieh
    ydshieh committed Aug 26, 2026
    Configuration menu
    Copy the full SHA
    879e744 View commit details
    Browse the repository at this point in the history

  8. Remove extra JOB_TO_TEST_FILE entries kept during cherry-pick

    @ydshieh
    ydshieh committed Aug 26, 2026
    Configuration menu
    Copy the full SHA
    1362826 View commit details
    Browse the repository at this point in the history

  9. [debug] Log max grad_sum on MISMATCH for normal and gradcp runs

    @ydshieh
    ydshieh committed Aug 26, 2026
    Configuration menu
    Copy the full SHA
    7902406 View commit details
    Browse the repository at this point in the history

  10. [debug] Use 1e-6 epsilon for non-zero grad check to ignore fp noise f… 

    …rom GC recomputation
    @ydshieh
    ydshieh committed Aug 26, 2026
    Configuration menu
    Copy the full SHA
    cb297b6 View commit details
    Browse the repository at this point in the history

  11. Revert "[debug] Use 1e-6 epsilon for non-zero grad check to ignore fp… 

    … noise from GC recomputation"

This reverts commit cb297b6.
    @ydshieh
    ydshieh committed Aug 26, 2026
    Configuration menu
    Copy the full SHA
    6301d81 View commit details
    Browse the repository at this point in the history

  12. [debug] Use 1e-6 epsilon for non-zero grad check to ignore fp noise f… 

    …rom GC recomputation
    @ydshieh
    ydshieh committed Aug 26, 2026
    Configuration menu
    Copy the full SHA
    60d4994 View commit details
    Browse the repository at this point in the history

  13. [debug] Lower epsilon to 1e-7 to avoid false mismatches at ~1e-6 grad… 

    …ient values
    @ydshieh
    ydshieh committed Aug 26, 2026
    Configuration menu
    Copy the full SHA
    2e710a7 View commit details
    Browse the repository at this point in the history

  14. [debug] Ignore fp noise: only suppress mismatch when one side is 0.0 … 

    …and other <= 1e-6
    @ydshieh
    ydshieh committed Aug 26, 2026
    Configuration menu
    Copy the full SHA
    1e8f22f View commit details
    Browse the repository at this point in the history
Loading