Repository navigation

Quantize NVFP4 1D from registers instead of through TMA staging - #3808

Draft

tdophung wants to merge 3 commits into

NVIDIA:mainfrom

tdophung:teddy/nvfp4-register-quantize-1d

Draft

Quantize NVFP4 1D from registers instead of through TMA staging#3808
tdophung wants to merge 3 commits into
NVIDIA:mainfrom
tdophung:teddy/nvfp4-register-quantize-1d

Commits on Oct 16, 2026
- Keep several 32-byte loads in flight in the amax kernel
  tdophung
  committed
  ac1bf61
Commits on Oct 20, 2026
- Quantize NVFP4 1D from registers instead of through TMA staging
  tdophung
  committed
  1fae46f
Commits on Nov 3, 2026
- Merge branch 'NVIDIA:main' into teddy/nvfp4-register-quantize-1d
  tdophung
  authored
  036e6a6