[bug] Fix "Current vLLM config is not set." warnings when FlashInfer attention is used

[nvpohanh]

Purpose

VLLM config is set only during initialization stage, not during runtime stage. Therefore, we should not call get_current_vllm_config() during dunrime stage. Instead, cache the config we want during initialization stage and reuse it during runtime stage.

This was caused by #26315 by @MatthewBonanni .

This fixes #30240

Test Plan

On H200:

python3 -m vllm.entrypoints.openai.api_server --model Qwen/Qwen3-32B-FP8 --kv-cache-dtype auto --attention-config.backend FLASHINFER &
# wait until server is ready
lm_eval --model local-completions --tasks gsm8k --model_args base_url=http://127.0.0.1:8000/v1/completions,model=Qwen/Qwen3-32B-FP8,tokenized_requests=False,tokenizer_backend=None,num_concurrent=128,timeout=120,max_retries=1

Test Result

local-completions (base_url=http://127.0.0.1:8000/v1/completions,model=Qwen/Qwen3-32B-FP8,tokenized_requests=False,tokenizer_backend=None,num_concurrent=128,timeout=120,max_retries=1), gen_kwargs: (None), limit: None, num_fewshot: None, batch_size: 1
|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value |   |Stderr|
|-----|------:|----------------|-----:|-----------|---|-----:|---|-----:|
|gsm8k|      3|flexible-extract|     5|exact_match|↑  |0.6679|±  |0.0130|
|     |       |strict-match    |     5|exact_match|↑  |0.7672|±  |0.0116

without all the warnings.

---

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.
• (Optional) Release notes update. If your change is user facing, please update the release notes draft in the Google Doc.

[gemini-code-assist]

Code Review

This pull request correctly addresses a bug causing "Current vLLM config is not set." warnings by ensuring that the vLLM configuration is accessed only during the initialization stage. The approach of caching the force_use_trtllm_attention setting in FlashInferMetadataBuilder during initialization and then passing it down during the runtime stage is sound and effectively resolves the issue. The changes are well-contained and logical. I have no further comments as the implementation is solid.

[hmellor]

vllm_config already exists in the scope that you're using force_use_trtllm_attention, so it should just be accessed directly

[MatthewBonanni]

Thanks, sorry I missed this

[nvpohanh]

@hmellor could you review again? thanks