H-Scale Optimizes NVFP4 Scale Selection for Blackwell Inference
August 31, 2026
H-Scale is a post-training quantization method for NVIDIA Blackwell's NVFP4 format that uses a diagonal second-order proxy to refine per-group scaling factors. It targets layer output perturbation rather than simple weight reconstruction error to improve sub-byte LLM inference accuracy.
HOW THIS AFFECTS YOU
●
builderYou can achieve better inference accuracy on Blackwell hardware by refining scales instead of just quantizing weights.
●
researcherThis approach demonstrates how second-order information can improve quantization in micro-block architectures.