RTX 3090 runs Qwen2.5-72B at 2,000 tok/s prefill
iamMess · reddit · 2026-09-01
A developer optimized Qwen2.5-72B (referred to as Qwen3.8-27B) on an RTX 3090, achieving 2,000 tokens/s prefill and 132 tokens/s decode. The core improvement is a custom kernel that matches fp32 quality with 0.99997 similarity at int8. The author believes decode speed is currently maxed out and has provided a GitHub repo for testing.
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