monkeypatch.kernelize_fixes
monkeypatch.kernelize_fixes
Repairs for transformers’ model.kernelize() under use_kernels=True.
kernelize() swaps each module’s _hidden_kernels entries for hub
kernels, but two upstream defects crash it (transformers 5.10): ~30
architectures (gemma4, qwen3.5, glm4, olmo, …) stash a bare function it
refuses to register, and gpt-oss’s rotary Func keeps a deprecated
position_ids parameter that fails the kernels library’s signature check
against kernels-community/rotary.
Wraps PreTrainedModel.kernelize to repair the stashes right before the
swap: bare functions are dropped (the model’s forward still calls them
directly) and position_ids is removed from rotary signature metadata
(call behavior unchanged). Both repairs no-op once fixed upstream.