Fix Zamba2MambaMixer ignoring use_mamba_kernels=False

[sergiopaniego]

What does this PR do?

Zamba2MambaMixer.__init__ calls lazy_load_kernel("mamba-ssm") and lazy_load_kernel("causal-conv1d") unconditionally. Models that inherit from it (like NemotronH) and set use_mamba_kernels=False in their config have the flag ignored, causing failures when the kernels package is installed but causal-conv1d CUDA kernels are not available.

Fix: Gate the lazy_load_kernel calls behind getattr(config, "use_mamba_kernels", True) in the Zamba2 modular.

Related to: huggingface/trl#5278

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline,
  Pull Request section?
• Was this discussed/approved via a Github issue or the forum? Please add a link
  to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
• Did you write any new necessary tests?

Who can review?

@ArthurZucker

[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.

[albertvillanova]

Thanks for addressing the underlying issue! 🤗

Just one question: the attribute use_mamba_kernels is only defined in NemotronHConfig, why do you use it in zamba2?

Additionally, you modified modeling_nemotron_h.py, but, as explained in that file header, that file was automatically generated from modular_nemotron_h.py. I think you should modify modular_nemotron_h.py instead.

CC: some maintainers that modified these files recently are @ydshieh, @Cyrilvallez

[sergiopaniego]

Just one question: the attribute use_mamba_kernels is only defined in NemotronHConfig, why do you use it in zamba2?

The problem comes fork Zamba2MambaMixer.__init__, which calls lazy_load_kernel unc onditionally.NemotronHMamba2Mixer calls super().__init__(), so the kernel loading already happens before NemotronH's own __init__ runs 😓

Additionally, you modified modeling_nemotron_h.py, but, as explained in that file header, that file was automatically generated from modular_nemotron_h.py. I think you should modify modular_nemotron_h.py instead.

Same idea as before

[albertvillanova]

Thanks for the fixing commit 93840ed: more clear now! 🤗

Let's wait for the opinion of some maintainers.

[github-actions]

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

run-slow: nemotron_h, zamba2

[Cyrilvallez]

causing failures when the kernels package is installed but causal-conv1d CUDA kernels are not available.

In this case, lazy_load_kernel should just grab it from hub no? Looking at lazy_load_kernel, I don't really see how it can crash?

[sergiopaniego]

it downloads it from the hub but the current version is not compatible with transformers (check here).
This solution is defensive against that. Updating it on kernels would also resolve the problem.

cc @danieldk

[Cyrilvallez]

Ha I see! Thanks!