P/D disaggregation · ML4 NVFP4 + DSpark · 1P1D · 4+4 GB300

Throughput vs interactivity, 1P1D

DSpark depth 3; NIXL; vllm bench serve, real-text prompts, greedy, ignore_eos. Solid: concurrency sweep (labels = concurrency). Dotted, hollow: fixed-QPS sweep. Hover a point for its metrics.