How NF12 works: escaped groups take 28 bytes, rarely exceeding BF16 size
AlpinDale · x · 2026-11-07
More details on NF12: it isn't exactly 12-bit across the board — escape buffers and a shared exponent base take some space. Weights are grouped in 8s; a normal group packs into 12 bytes (instead of 16). Groups that can't be represented are marked as exceptions, with the original eight BF16 values stored in a separate escape buffer (28 bytes total). If escape buffers pile up, total size could exceed BF16, but this almost never happens with NF12 (it would at NF11/10).
Related event: Developer open-sources nf12: lossless 12-bit compression for BF16 models(10 posts)→
More from Infra
- Fudan's D3D diffusion drafter enables async speculative decoding, up to 7.48x speedup on Qwen3 — pmttyji · 2026-11-07
- YouJiacheng: post-training needs high tps for long-tail rollouts — YouJiacheng · 2026-11-07
- vLLM contributor: lossless KV cache quantization viable but not worth it — AlpinDale · 2026-11-07
- NVIDIA's first co-packaged optics Ethernet switch, Spectrum-X Photonics, enters full production — TinfoilTricorn · 2026-11-07
- This week's open training stack: MoE training handbook, harness adapters, and a harness leaderboard — ariG23498 · 2026-11-07
- Musk: space will be the cheapest place to run superintelligence within 2-3 years — TinfoilTricorn · 2026-11-07