# Built-in DSpark: 1.92–2.16x output throughput in FP8 and 2.00–2.23x in NVFP4, versus the matching baseline.

- Source: vLLM
- Published: 2026-10-29T15:13:47.000Z
- Bad Influence: https://badinfluence.news/items/t2xu0z3k8yeddkox23xiwb3k1
- Original: https://x.com/vllm_project/status/2115824423953694873

## Summary

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 tests used TP8, 200 real prompts, greedy decoding, and up to 512 output tokens. The post also cites cross-layer KV sharing, context parallelism and NIXL P/D serving for long requests.

## Why it matters

The post gives measured throughput ranges for FP8 and NVFP4 against matching baselines, along with the test setup, so readers can check the claimed speedup.

## Post

Built-in DSpark: 1.92–2.16x output throughput in FP8 and 2.00–2.23x in NVFP4, versus the matching baseline.

TP8 · 200 real prompts · greedy decoding · ≤512 output tokens.

Cross-layer KV sharing, context parallelism and NIXL P/D serving handle long requests.

![](https://badinfluence.news/api/img-proxy?u=https%3A%2F%2Fpbs.twimg.com%2Fmedia%2FHVzrGhLbQAA2rID.jpg&mode=thumb&exp=1793491200&sig=711283208d5b2666)