Extensive experiments show weight quantization bottoms out around 10.45 bits
AlpinDale · x · 2026-11-07
The author shares first-hand results from extensive weight-width compression experiments:
- Tested widths from 8 to 12 bits; lower widths make it nearly impossible to write fast inference kernels.
- 8-11 bit variants suffer too many escape groups, making them larger than BF16: pure NF11 has 13.3% of weights escaping vs 0.3% for NF12, with each escaped group effectively FP28.
- A hybrid mechanism letting the encoder pick the dtype (NF8-NF12) with fewest escape groups showed little improvement.
- Explored exponent palettes, compact exception patches, local contexts, Huffman coding, ANS entropy coding, and neighbor-weight predictors — the lowest width achieved consistently was 10.45 bits.
Related event: Developer open-sources nf12: lossless 12-bit compression for BF16 models(10 posts)→
More from Infra
- US AI scale now hinges on coordinating power, components, agentic workflows and humanoid pilots — ingliguori · 2026-11-07
- Fudan's D3D diffusion drafter enables async speculative decoding, up to 7.48x speedup on Qwen3 — pmttyji · 2026-11-07
- Study: data centers have only marginal impact on residential electricity prices — TinfoilTricorn · 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