AI
PrismQuant: Optimal Null-Space Rotations for Grouped Quantizers
Publication: Hugging Face Daily Papers
Author: Not available in collected metadata
Source published: 25/09/2026, 20:00:00 UTC
Briefing created: 30/09/2026, 05:05:25 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 30/09/2026, 04:18:25 UTC.
Briefings
PrismQuant introduces a quantizer-aware rotation framework using compact Householder transformations to align activation eigenspaces with constant group subspace in grouped INT4 quantization. The method formulates rotation design as a Ky Fan trace maximization problem with a provably optimal closed-form solution. Experiments on Llama, Qwen, and Mistral models up to 70B parameters show PrismQuant achieves state-of-the-art performance on Llama-3.2-3B and 3.85 perplexity on Llama-3.1-70B, with minimal overhead in deployment compared to FP16 baselines.
Key points
- PrismQuant uses compact Householder transformations and their compact-WY representation to enable gradient-free construction and efficient application at both foldable and online sites.
- The framework formulates rotation design as a Ky Fan trace maximization problem and derives a closed-form solution that is provably optimal for aligning activation eigenspace with the constant group subspace.
Context and limitations
Evidence covers dense models up to 70B parameters and a 30B mixture-of-experts model; deployment study on Llama-3.1-8B shows 1.51x prefill and 1.22x CUDA Graph decode speedups over matched FP16 baselines.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 30/09/2026, 05:05:25 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeigk4ix6bmtckknys4bleprdzrbtu4yueeextcxysfokieb3x6sueu