SGLang's 75% sampler speedup skips Qwen3 chat requests

SGLang v0.5.20 replaces torch.multinomial with a Gumbel-max trick, claiming +75% decode at batch 1 and +23% at batch 32. But the fast path only runs when no request in the batch sets top_k, top_p or min_p; Qwen3's generation_config.json sets top_k and top_p by default on /v1/chat/completions, so that traffic never reaches it.

• Gumbel-max swap speeds decode by 75% at batch 1 and 23% at batch 32
• A single request with top_k, top_p or min_p moves the whole batch to the filtered path
• Batch flags persist after a request ends until the batch fully drains
• SGLANG_OPT_USE_GUMBEL_SAMPLE is enabled by default