Repository navigation

[CUDA] Branchless NF4/FP4 kDequantizeBlockwise kernel for faster dequantization - #1746

Merged

matthewdouglas merged 3 commits into

bitsandbytes-foundation:mainfrom

Mhmd-Hisham:cuda-branchless-dequantization-float32-lut

Sep 18, 2025

Commits on Sep 4, 2025
- Added branchless LUT-based dequantization for FP4 and NF4
  Mhmd-Hisham
  committed
  ffc6bab
- Added extra command line options to control reproducibility
  Mhmd-Hisham
  committed
  4424b73
Commits on Sep 6, 2025
- Restore FP4 quantization/dequantization order
  Mhmd-Hisham
  committed
  2c4927d