Repository navigation

Add k-bit blockwise quantization (K=2-5) with warp-level CUDA kernels - #1858

Open

TimDettmers wants to merge 12 commits into

mainfrom

feature/kbit-quantization

Open

Add k-bit blockwise quantization (K=2-5) with warp-level CUDA kernels#1858
TimDettmers wants to merge 12 commits into
mainfrom
feature/kbit-quantization

Commits on Feb 14, 2026
- Add k-bit quantization kernels (K=2-5, blocksize=32) -- WIP
  
  TimDettmers
  and
  claude
  committed
  c39f791
- Fix RDC device linking: move kernels to ops.cu, all 157 tests pass
  
  TimDettmers
  and
  claude
  committed
  fb649f1
- Complete k-bit quantization: Stages 6-8, Python API, 218 tests pass
  
  TimDettmers
  and
  claude
  committed
  2825890
- Remove implementation progress report
  
  TimDettmers
  and
  claude
  committed
  4b17a2f
- Add vectorized dequant kernel and E4M4 uint8 absmax support
  
  TimDettmers
  and
  claude
  committed
  2973bf5
- Remove scalar dequant kernel, fp32 absmax, and Stage 1-3 scaffolding
  
  TimDettmers
  and
  claude
  committed
  03415e1
- Template dequant kernel on output type, add bf16/fp32 native output
  
  TimDettmers
  and
  claude
  committed
  8a2817e
- Fix lint and formatting issues from CI pre-commit checks
  
  TimDettmers
  and
  claude
  committed
  f52b572
- Fix analytical error bound for K=5 with E4M4 absmax
  
  TimDettmers
  and
  claude
  committed
  f95a7f2
Commits on Feb 22, 2026
- Add out parameter to dequantize_kbit for CUDA graph compatibility
  
  TimDettmers
  and
  claude
  committed
  d1f3d75
- docs: Add kbit design docs, remove spec.md
  
  TimDettmers
  and
  claude
  committed
  10cf922
Commits on Mar 2, 2026
- Apply suggestion from @matthewdouglas
  matthewdouglas
  authored
  ad7f194