Prototype NVFP4 with FP8 UE5M3 block scales - #3325

Open

timmoon10 wants to merge 33 commits into

NVIDIA:mainfrom

timmoon10:nvfp4-ue5m3-prototype

Commits on Aug 7, 2026

Prototype NVFP4 with UE5M3 scales

committed
25d5b4b
[pre-commit.ci] auto fixes from pre-commit.com hooks
pre-commit-ci[bot]
committed
dede7c7

Commits on Aug 13, 2026

[PyTorch] Enable e5m3 fused GEMM kernels from cuDNN (#2)

authored
343c4bd
Merge branch 'main' into nvfp4-ue5m3-prototype
timmoon10
committed
8feb5e9

Commits on Aug 14, 2026

Use custom recipe for NVFP4-UE5M3 tests
timmoon10
committed
ab3a9b3
[pre-commit.ci] auto fixes from pre-commit.com hooks
pre-commit-ci[bot]
committed
92b1063
Add grouped MLP kernel for GGEMM+SwiGLU+RHT+quant
timmoon10
committed
3d25d20
[pre-commit.ci] auto fixes from pre-commit.com hooks
pre-commit-ci[bot]
committed
f030e40
Merge branch 'main' into nvfp4-ue5m3-prototype
timmoon10
committed
535ea4f
[pre-commit.ci] auto fixes from pre-commit.com hooks
pre-commit-ci[bot]
committed
e366e88

Commits on Aug 15, 2026

Debug integration with GGEMM+GLU+RHT+quant
timmoon10
committed
252ae94

Commits on Aug 18, 2026

Remove scale max helper functions from NVFP4 cast utils
timmoon10
committed
f2b9c98
[pre-commit.ci] auto fixes from pre-commit.com hooks
pre-commit-ci[bot]
committed
8685556
Fix compile error
timmoon10
committed
803dd24

Commits on Aug 21, 2026

Fix NVFP4 scale dtype ABI

tdophung
authored and
timmoon10
committed
27f29d0
remove redundant output alloc

kainzhong
authored and
timmoon10
committed
c9dad3e
no need to pad N now

kainzhong
authored and
timmoon10
committed
2bf42eb
fix linting errors

kainzhong
authored and
timmoon10
committed
b6913b2
[pre-commit.ci] auto fixes from pre-commit.com hooks

pre-commit-ci[bot]
authored and
timmoon10
committed
f2c26bd
Disable cuDNN GGEMM+GLU+RHT+quant kernel
timmoon10
committed
030e9af
Guard NVFP4 alpha scaling by scaling mode

tdophung
authored and
timmoon10
committed
9162e98
Restore UE5M3 NVFP4 cast support

tdophung
authored and
timmoon10
committed
db0957b
Localize NVFP4 4over6 scale policy

tdophung
authored and
timmoon10
committed
982be7b
Tweak arg order in C API functions
timmoon10
committed
01f4edc
Rename cuDNN GGEMM helper functions for general_gemm
timmoon10
committed
3a63623
Merge branch 'main' into HEAD
timmoon10
committed
26db2ee
Fix compilation error in C++ test
timmoon10
committed
8e229f1
Treat nvfp4_e4m3_max=0 as unset value

timmoon10
and
codex
committed
c8e6ced
Debug torch.compile test failure
timmoon10
committed
6e84816
Debug test failures
timmoon10
committed
0c2de5e
[pre-commit.ci] auto fixes from pre-commit.com hooks
pre-commit-ci[bot]
committed
dd90509
Revert accidental CUTLASS commit change
timmoon10
committed
d42ccbe
Enable cuDNN GGEMM+GLU+RHT+quant kernel
timmoon10
committed
b47e1f8