Repository navigation

CuTeDSL NVFP4 1D quantize kernels - #3809

Draft

tdophung wants to merge 5 commits into

NVIDIA:mainfrom

tdophung:teddy/nvfp4-cutedsl-register-quantize

Draft

CuTeDSL NVFP4 1D quantize kernels#3809
tdophung wants to merge 5 commits into
NVIDIA:mainfrom
tdophung:teddy/nvfp4-cutedsl-register-quantize

Commits on Nov 2, 2026
- Port the CuTeDSL NVFP4 quantize backend to the current CuTeDSL infrastructure
  
  janekb04
  authored and
  tdophung
  committed
  54fd205
Commits on Nov 3, 2026
- Split the CuTeDSL NVFP4 backend into a dispatcher and per-kernel modules
  tdophung
  committed
  0c55146
- Add register-resident CuTeDSL NVFP4 quantize kernels
  tdophung
  committed
  7712fef
Commits on Nov 4, 2026
- Simplify the CuTeDSL NVFP4 backend tests to the MXFP8 test layout
  tdophung
  committed
  bcd15f1
- Share CuTeDSL helpers between the NVFP4 and MXFP8 backends
  tdophung
  committed
  a18f3e9