Repository navigation

[PyTorch][Common] Add MXFP8 and NVFP4 QDQ for mixed-precision GEMM - #3633

Open

negvet wants to merge 11 commits into

NVIDIA:mainfrom

negvet:nvfp4_qdq

Open

[PyTorch][Common] Add MXFP8 and NVFP4 QDQ for mixed-precision GEMM#3633
negvet wants to merge 11 commits into
NVIDIA:mainfrom
negvet:nvfp4_qdq

Commits on Oct 6, 2026
- Refactor output dtype handling for Identity and Hybrid quantizers
  negvet
  committed
  f543cd3
- Add experimental NVFP4 weight QDQ quantization
  negvet
  committed
  359f6d3
Commits on Oct 7, 2026
- Merge branch 'main' into nvfp4_qdq
  negvet
  authored
  ecbf527
Commits on Oct 8, 2026
- Add MXFP8 QDQ and shared QDQ quantizer base
  negvet
  committed
  7ee5281
- [pre-commit.ci] auto fixes from pre-commit.com hooks
  pre-commit-ci[bot]
  committed
  432e4d4
Commits on Oct 12, 2026
- Clarify QDQ operand labels and factory naming
  negvet
  committed
  df8269c
Commits on Oct 13, 2026
- Add per-token outer scale to the fused nvfp4 path
  negvet
  committed
  1171a05
- Fix HybridQuantizer dtype validation and column source precision
  negvet
  committed
  f51212d
- Merge branch 'main' into nvfp4_qdq
  negvet
  authored
  9e09c01
Commits on Oct 14, 2026
- Enable NVFP4 and MXFP8 QDQ on Ampere+
  negvet
  committed
  a82d551
Commits on Oct 16, 2026
- Allow CustomRecipe without FP8 hardware and fix QDQ capability gates
  negvet
  committed
  7f5e16f