vLLM prefill speed vastly outperforms other inference engines, sparking technical inquiry
dangerous_inference · reddit · 2026-09-02
User reports vLLM prefill speeds (5000-7500 pp) on 4x4090 far exceed llama.cpp/ikllama. Questions the technical obstacles preventing similar performance in other engines.
More from Infra
- SGLang Summit sponsors grow from 11 to 36 in one week; Lip-Bu Tan, Lilian Weng to speak — BanghuaZ · 2026-10-18
- Yacine jokes his datacenter KPI is gigawatt-hours: get paged if not burning enough power — yacineMTB · 2026-10-18
- 1.68x inference speedup drops, directly accelerating GRPO post-training — yacineMTB · 2026-10-18
- OpenRouter is only a fraction of the total addressable market, poster argues — sloppenheimer · 2026-10-18
- Developer builds change alerts for LLM prices and retirements across Bedrock, Azure and Vertex — jsandhu0780 · 2026-10-18
- 12.3M tokens in 27 minutes for $0.14: agent session with 99.6% cache hit — teortaxesTex · 2026-10-18