[Bug] DeepGemm: Fix Cuda Init Error

[yewentao256]

Purpose

Fixes #21304

Thanks for @robertgshaw2-redhat 's help!

Test

Origin:

ERROR 07-19 10:21:10 [registry.py:388]   File "/home/wentao/.wentao_env/lib/python3.12/site-packages/torch/cuda/__init__.py", line 576, in get_device_properties
ERROR 07-19 10:21:10 [registry.py:388]     _lazy_init()  # will define _get_device_properties
ERROR 07-19 10:21:10 [registry.py:388]     ^^^^^^^^^^^^
ERROR 07-19 10:21:10 [registry.py:388]   File "/home/wentao/.wentao_env/lib/python3.12/site-packages/torch/cuda/__init__.py", line 372, in _lazy_init
ERROR 07-19 10:21:10 [registry.py:388]     torch._C._cuda_init()
ERROR 07-19 10:21:10 [registry.py:388] RuntimeError: CUDA error: initialization error

Now:

vllm (pretrained=Qwen/Qwen3-30B-A3B-FP8,max_model_len=32768,enforce_eager=True,trust_remote_code=True), gen_kwargs: (None), limit: None, num_fewshot: 5, batch_size: auto
|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value |   |Stderr|
|-----|------:|----------------|-----:|-----------|---|-----:|---|-----:|
|gsm8k|      3|flexible-extract|     5|exact_match|↑  |0.8340|±  |0.0102|
|     |       |strict-match    |     5|exact_match|↑  |0.8939|±  |0.0085|

[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 refactors the deep_gemm utility to use lazy initialization to fix a CUDA initialization error. A critical bug was identified in the new lazy initialization logic due to a missing global declaration and an incomplete "fast path" check, which would prevent some of the functionality from working as intended. Suggestions have been provided to fix these issues.

[mgoin]

LGTM, thanks!