PR #45395: Fix IndexError with DeepSpeed ZeRO-3 when kernels rotary is active

• Repository: huggingface/transformers
• Author: ArthurZucker
• State: closed | merged: False
• Link: https://github.com/huggingface/transformers/pull/45395

Description (problem / solution / changelog)

Summary

Fixes #45137.

Since #41147, attention layers are decorated with @use_kernelized_func(apply_rotary_pos_emb) which attaches a rotary_fn child nn.Module at init when the kernels library is available.

DeepSpeed ZeRO-3's parameter coordinator traces the module graph at init and expects every registered submodule to run during forward. The attention forward still calls the Python apply_rotary_pos_emb, so rotary_fn is never invoked and the parameter-fetch trace desynchronizes, raising:

IndexError: pop from an empty deque
  at deepspeed/runtime/zero/partitioned_param_coordinator.py

on the second forward (reproducible via TRL's RLOO/GRPO trainers under ZeRO-3, see huggingface/trl#4899).

Changed files

• docs/source/en/model_doc/pp_chart2table.md (modified, +1/-1)
• docs/source/en/model_doc/slanext.md (modified, +1/-1)
• docs/source/en/model_doc/uvdoc.md (modified, +1/-1)
• src/transformers/integrations/hub_kernels.py (modified, +9/-0)

PR #45414: Fix IndexError with DeepSpeed ZeRO-3 when kernels rotary is active

• Repository: huggingface/transformers
• Author: ArthurZucker
• State: closed | merged: True
• Link: https://github.com/huggingface/transformers/pull/45414

Description (problem / solution / changelog)

Summary

Fixes #45137. Re-opened from #45395 on a same-repo branch so CI can run.

Since #41147, attention layers are decorated with @use_kernelized_func(apply_rotary_pos_emb) which attaches a rotary_fn child nn.Module at init when the kernels library is available. DeepSpeed ZeRO-3's parameter coordinator traces the module graph at init and expects every registered submodule to fire during forward. The attention forward still calls the plain Python apply_rotary_pos_emb, so rotary_fn is never invoked and the parameter-fetch trace desynchronizes, raising:

IndexError: pop from an empty deque
  at deepspeed/runtime/zero/partitioned_param_coordinator.py

on the second forward (reproducible via TRL's RLOO/GRPO trainers under ZeRO-3, see huggingface/trl#4899).

Fix

Skip attaching the kernelized submodule when is_deepspeed_zero3_enabled() is true. Under ZeRO-3 the Python apply_rotary_pos_emb path is used (same behavior as before #41147). Non-ZeRO-3 users are unaffected.

The second commit refreshes dates on three model cards (pp_chart2table, slanext, uvdoc) that were missing them on main — required for check-repository-consistency to pass.

Test plan

• Reproducer from huggingface/trl#4899 no longer raises IndexError: pop from an empty deque
• Qwen3 forward + kernelize still replaces rotary_fn when not under ZeRO-3
• make style + check-repository-consistency pass

Changed files

• docs/source/en/model_doc/pp_chart2table.md (modified, +1/-1)
• docs/source/en/model_doc/slanext.md (modified, +1/-1)
• docs/source/en/model_doc/uvdoc.md (modified, +1/-1)
• src/transformers/integrations/hub_kernels.py (modified, +9/-0)

PR #5541: Update tests with zero3 for RLOO and GRPO once fixed in transformers 5.5.4

• Repository: huggingface/trl
• Author: albertvillanova
• State: closed | merged: True
• Link: https://github.com/huggingface/trl/pull/5541

Description (problem / solution / changelog)

Update tests with zero3 for RLOO and GRPO once fixed in transformers 5.5.4.

This PR updates the test conditions for ZeRO-3 integration with the transformers library to reflect a recent upstream fix. The tests now only expect failures for a specific range of transformers versions where the issue is known to occur, improving the accuracy of test expectations.

Fix #4899, after the upstream issue in transformers:

• https://github.com/huggingface/transformers/issues/45137

has been fixed by:

• https://github.com/huggingface/transformers/pull/45414

Follow-up to:

• #5420
• #5404
• #4898
• #4899

Changes

Test condition updates:

• In both test_reward and test_rloo in tests/distributed/test_distributed.py, the pytest.mark.xfail condition for the "zero3" parameter is updated to only expect failures when transformers version is greater than or equal to 5.0.0 and less than 5.5.4, reflecting that the issue is fixed in transformers#45414. The reason message is also updated for clarity.

[!NOTE] Low Risk Low risk: only adjusts pytest xfail version gating and messages in distributed tests, with no production code changes.

Overview Updates distributed tests so the zero3 parameter is only marked xfail for transformers versions >= 5.0.0 and < 5.5.4, reflecting that the upstream ZeRO-3 issue is fixed in transformers 5.5.4.

Also updates the associated xfail reason strings (and keeps strict=True) in test_rloo and test_grpo to document the fixed upstream PR/reference.

^{Reviewed by Cursor Bugbot for commit fef7620e6204dafedc6e16fb5c42f619bc7a135b. Bugbot is set up for automated code reviews on this repo. Configure here.}

Changed files

• tests/distributed/test_distributed.py (modified, +4/-4)