vLLM adds built-in DSpark, reporting 1.92–2.23x output throughput gains

vLLM says its built-in DSpark support raises output throughput 1.92–2.16x in FP8 and 2.00–2.23x in NVFP4 versus each precision's matching baseline.

The figures come from tests on TP8 with 200 real prompts, greedy decoding and up to 512 output tokens.

The same post says cross-layer KV sharing, context parallelism and NIXL P/D serving handle long requests. The post does not report results for long-request workloads.