vLLM reached 25,000 tokens per second per GPU serving Qwen3.5 on Blackwell GPUs via a disaggregated prefill-decode architecture. Three optimizations drove the result: a Blackwell-optimized GDN prefill kernel delivering up to 5.92x improvement, hybrid cache management transferring both attention and state-machine layer data between workers, and race-condition fixes enabling asynchronous scheduling. Tests on GB200 NVL72 systems with the 397-billion-parameter model showed 88% accuracy on GSM8K while sustaining peak throughput across varying concurrency levels.
