[jiqing-feng]

Problem

PR #45420 refactored kernel function registration to use a hidden _hidden_kernels dict + temporary register_module attach/detach cycle. This works correctly when the target function is decorated with @use_kernel_func_from_hub (which wraps it into an nn.Module instance via _create_func_module).

However, gpt_oss uses @use_kernelized_func(apply_rotary_pos_emb) where apply_rotary_pos_emb is a plain function — missing the @use_kernel_func_from_hub("rotary_pos_emb") decorator that other models (deepseek_v3, gemma3, etc.) have. This causes:

1. register_module raises TypeError: ... is not a Module subclass
2. In the finally block, delattr raises AttributeError since attach never succeeded

Fix

1. modeling_gpt_oss.py: Add @use_kernel_func_from_hub("rotary_pos_emb") decorator to apply_rotary_pos_emb, consistent with all other models.
2. modeling_utils.py: Make attach_hidden_kernels defensive — skip non-nn.Module entries with a warning instead of crashing. Use safe _modules dict access for detach.

Reproduce

from transformers import pipeline
generator = pipeline('text-generation', model='openai/gpt-oss-20b', model_kwargs={'use_kernels': True})

error:

......
  File "/opt/venv/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1077, in apply
    fn(self)
  File "/home/jiqing/transformers/src/transformers/modeling_utils.py", line 4581, in detach_hidden_kernels
    delattr(module, name)
  File "/opt/venv/lib/python3.12/site-packages/torch/nn/modules/module.py", line 2086, in __delattr__
    super().__delattr__(name)
AttributeError: 'GptOssAttention' object has no attribute 'apply_rotary_pos_emb'

Regression

Introduced in #45420 (commit 253809c).

[jiqing-feng]

Hi @vasqu . Would you please review this PR? Thanks!

[vasqu]

Can you add a regression test for gpt oss that runs with kernels --> just that it does not crash

[jiqing-feng]

Hi @vasqu . I have fixed the comment and added the test, please review it. Thanks!

[jiqing-feng]

The failed CI is not related to my changes.

[github-actions]

[For maintainers] Suggested jobs to run (before merge)

run-slow: gpt_oss, openai_privacy_filter

[vasqu]

Thanks, I just changed it to raise an ValueError at the end instead. Having a broken registry shouldn't be the case and silently warning will be ignored so I'd rather be more strict here

[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.