SGLang Summit: production NVFP4 low-precision RL battle-tested on thousands of GPUs
gharik · x · 2026-11-14
At the SGLang Summit, zianglih presented production low-precision RL with NVFP4 and MXFP8, plus w4a16 decode. Yuchen He relayed that full details, illustrations and benchmarks for 4-bit (NVFP4) RL are on the humans& blog, all validated via large-scale long-horizon RL across thousands of GPUs and trillions of tokens, with the team contributing results back to the open-source community via sglang, radixark and NVIDIA AI Dev.
Notable blog posts include "The 4-bitter Lesson" (balancing stability and performance in NVFP4 reinforcement learning) and "Faster-than-light NVFP4 decode" on the surprisingly reasonable effectiveness of 16-bit NVFP4 decode — practical references for teams running large-scale LLM RL training.
More from Infra
- Run Gemma 4 locally with one Ollama command: 7-10GB, offline, no subscription caps — nikola_mr64990 · 2026-11-14
- MiniMax H3 I2V on a 5090: cu130 torch fix kills OOM, 0.98MP x 20s in ~21 minutes — Realistic-Fennel-190 · 2026-11-14
- Inkling Small 2 fits on 4x RTX Pro 6000 with 7.5M token KV cache at 120 tok/s — LiTianleli · 2026-11-14
- Perplexity team teases minimalist LLM trainer built with coding agents, open source soon — BanghuaZ · 2026-11-14
- Serving many fine-tuned variants of a 27B model on one GPU: 4-bit CQ format and 55x smaller adapters — textclf · 2026-11-14
- Musk: Future Starship Flights Will Carry Over 50 Tb/s of Starlink Bandwidth to Orbit — elonmusk · 2026-11-14