Handle triton kernel import exception

[hjh0119]

The PR at #25319 introduced an ImportError catch to prevent failures when importing triton. However, after upgrading vLLM from 0.11.0 to 0.11.1 via:

pip install -U vllm

Running vllm serve Qwen/Qwen3-8B results in the following error:

AttributeError: module 'triton.language' has no attribute 'constexpr_function'
  File "/usr/local/lib/python3.11/site-packages/triton_kernels/numerics_details/flexpoint.py", line 55, in <module>
    @tl.constexpr_function
     ^^^^^^^^^^^^^^^^^^^^^

This suggests a compatibility issue between the upgraded triton version and triton-kernels, where constexpr_function was either removed or renamed.

While the original PR handled ImportError, this case involves an AttributeError due to API changes in triton.language. Adding a guard for AttributeError (in addition to ImportError) could help gracefully handle such mismatches and improve robustness during dependency upgrades.

This change proposes catching AttributeError when accessing tl.constexpr_function to prevent breakage in environments where triton and triton-kernels are mismatched.

[gemini-code-assist]

Code Review

This pull request addresses a runtime AttributeError that occurs due to version incompatibilities between triton and triton-kernels. The change extends an existing try...except block to also catch AttributeError during the import of Triton kernels, making the error handling more robust. Additionally, the exception is now logged, which will aid in debugging similar issues in the future. The change is correct and improves the project's stability in diverse environments.

[LucasWilkinson]

LGTM