About

Welcome

BitDance: Computing NVFP4 on MXFP6 Hardware

MXFP and NVFP: Block Scales, Two-Level Scaling, and Quantization Error

Welcome to the English edition of YIN Xuanwu's technical blog.

Folding NVFP4 B32 E4M3 scales into MXFP6 to preserve 4-bit weight storage while using native MXFP6 compute, with weight, operator, end-to-end, memory, and performance results.

A comparison of MXFP4, MXFP6, and NVFP4, focusing on block-scale formats, two-level scaling, and weight-only quantization error.

技术分享

中文

Switch to the explicit Chinese locale route so browser language preferences cannot redirect back to English.

LLM Quantization

公告

Technology, Practice, and Reflections by YIN Xuanwu

社区

Category

社区