vLLM hit 25K total TPS/GPU on Qwen3.5 and published the exact flags.

Using GB200 NVL72 disaggregated serving, 4–8 prefill endpoints at DEP2 plus one decode endpoint at DEP8, 8192-in/1024-out (vLLM). Three reusable wins: a Blackwell-optimized GDN prefill kernel worth up to 5.92x kernel and 1.13x prefill throughput via --gdn-prefill-backend flashinfer; a hybrid cache/state transfer moving both attention KV and SSM state correctly, cutting descriptors from 4,284 to 1,650 for ~7% throughput; and two race-condition fixes that finally made --async-scheduling viable. Full recipe includes VLLM_SSM_CONV_STATE_LAYOUT=DS, --mamba-ssm-cache-dtype bfloat16, --language-model-only, and --max-num-batched-tokens 16384 at 2x input sequence length.