PrismQuant: null-space rotations make INT4 near-lossless, only 0.22pp below FP16 on Llama-70B

NanyangTechnologicalUniversity · hf · 2026-09-30

PrismQuant aligns activation eigenspaces with the constant group subspace of asymmetric grouped INT4 via a closed-form Ky Fan trace-maximizing rotation built from compact Householder transforms. Under W4A4KV4 it sets SOTA on Llama-3.2-3B; Llama-3.1-70B reaches 3.85 perplexity and 72.46% zero-shot accuracy (only 0.22pp below FP). Deployed on Llama-3.1-8B it yields 1.51x prefill and 1.22x decode speedups with 56.34% lower peak decode memory. Code released.

Original post →

More from Infra

Infra channel →