[gpt-oss] guard import when triton kernel is not installed

[zyongye]

No description provided.

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.

💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels.

Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run fastcheck CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your fastcheck build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping simon-mo or khluu to add you in our Buildkite org.

Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.

To run CI, PR reviewers can either: Add ready label to the PR or enable auto-merge.

🚀

[gemini-code-assist]

Code Review

This pull request aims to make triton_kernels an optional dependency by guarding its import. My review focuses on the correctness of this implementation. I've identified a critical issue where the current change is incomplete and will likely cause runtime errors if triton_kernels is not installed. The code that depends on the optional imports also needs to be conditionally defined.

[gemini-code-assist]

While this change correctly guards the imports, the code that uses these imported modules (e.g., swiglu, PrecisionConfig) also needs to be conditionally defined. Without this, if triton_kernels is not installed, the Python interpreter will raise a NameError when it tries to define functions or classes that reference these names.

To fix this, you should move all code that depends on triton_kernels inside this if block. You should also consider adding an else block to define dummy implementations or raise an ImportError for the exported names (like BatchedOAITritonExperts and triton_kernel_moe_forward) to provide a clear error message to users who try to use this functionality without triton_kernels installed.

[WoosukKwon]

@zyongye I think gemini's comment makes sense. Can you add something like this?

if has_triton_kernels():
    import ...
else:
    FnSpecs = None
    FusedActivation = None
...

[zyongye]

But this function will never touch those lines if there's no triton kernel installed. I put an error in the FusedMoe init
https://github.com/vllm-project/vllm/blob/main/vllm/model_executor/layers/fused_moe/layer.py#L729-L730

[zifeitong]

Just tried and I have to use:

-if True:
+if has_triton_kernels():
     import triton_kernels.swiglu
     from triton_kernels.matmul_ogs import (FnSpecs, FusedActivation,
                                            PrecisionConfig, matmul_ogs)
     from triton_kernels.routing import routing
+else:
+    PrecisionConfig = None

Otherwise, I get:

(APIServer pid=5435)   File "/root/vllm/vllm/model_executor/layers/fused_moe/gpt_oss_triton_kernels_moe.py", line 146, in BatchedOAITritonExperts
(APIServer pid=5435)     w1_precision: PrecisionConfig, w2_precision: PrecisionConfig):
(APIServer pid=5435)                   ^^^^^^^^^^^^^^^
(APIServer pid=5435) NameError: name 'PrecisionConfig' is not defined

[zyongye]

ok new change pushed

[WoosukKwon]

@zyongye Actaully, we can just use TYPE_CHECKING guard. I fixed the PR.

[WoosukKwon]

Thanks for the fix. Will merge after it passes the pre-commit and lints.