Add static per block MSE for NVFP4 weight#613

Merged

Fridah-nv merged 10 commits intomainfrom

fridah/block-mse

Jan 13, 2026

Commits on Nov 27, 2025

tmp update for per block mse NVFP4 and INT4
Fridah-nv
committed

Commits on Dec 5, 2025

improvements: even steps for mse amax search;calibrate weight quant once; quant scale to FP8; rename static kernel
Fridah-nv
committed

Commits on Dec 10, 2025

tmp: config update for experiments
Fridah-nv
committed

Commits on Dec 18, 2025

clear up cached data after calibration, remove tmp configs
Fridah-nv
committed

Commits on Jan 6, 2026

add unit tests, address reviewer comments
Fridah-nv
committed

Commits on Jan 7, 2026

fix gpu unit test
Fridah-nv
committed

Commits on Jan 9, 2026

use step_size instead of num_steps; move FP8 scale quant into FP4 kernel launch func
Fridah-nv
committed

Commits on Jan 10, 2026

minor: fix calibrator test
Fridah-nv
committed

Commits on Jan 13, 2026