NVFP4 low-precision RL validated at trillion-token scale on ~1,000 GPUs at SGLang Summit

At SGLang Summit, Ziang Li of humans& lab shared production practices for large-scale low-precision RL, covering NVFP4 and MXFP8 training and w4a16 decoding, validated on 1,000 GPUs with trillions of tokens, with benchmarks to be released in a blog and July talk.

2026-11-14 ~ 2026-11-14 · 2 related posts

• SGLang Summit talk details battle-tested NVFP4 4-bit RL across thousands of GPUs — ericzelikman · 2026-11-14

1 near-duplicate retellings: gharik