DeepSeek-V4.1-Flash · GB300 · DEP2 prefill after a 128K cache hit

Three changes raise new-token prefill throughput by 35.6–43.9%

New tokens per second per GPU (higher is better). Each step adds one change on top of the one before. Y axis starts at 40K.

vllm bench forward on DEP2, one request per rank: 128K cached + N new tokens. Throughput = N / forward time (median of 100 CUDA-graph replays, slowest rank). Excludes spec decode, KV transfer and scheduling.