NVFP4 primary weights#2690

Closed

WanZzzzzz wants to merge 47 commits intoNVIDIA:mainfrom

WanZzzzzz:fp4_primary_weights

Commits on Nov 6, 2025

Support for NVFP4 primary weights
qiyuw
committed

Commits on Nov 10, 2025

conflicts
qiyuw
committed

Commits on Nov 18, 2025

partial amax and cast
qiyuw
committed

Commits on Nov 20, 2025

add unit tests
qiyuw
committed

Commits on Nov 21, 2025

Commits on Nov 22, 2025

fix build error
qiyuw
committed

Commits on Nov 25, 2025

fix some issues
qiyuw
committed

Commits on Nov 30, 2025

use larger problem for cublas gemm,fix some issues
qiyuw
committed

Commits on Dec 8, 2025

debug code
qiyuw
committed
Fix partial amax kernel and Add clear transpose cache
kunlunl
committed
Merge branch 'fp4_primary_weights' into 'fp4_primary_weights'
Qiyu Wan
committed
create colwise
qiyuw
committed
colwise usage fix
qiyuw
committed

Commits on Dec 9, 2025

Fix partial cast numerical bug
kunlunl
committed
Merge branch 'fp4_primary_weights' into 'fp4_primary_weights'
Qiyu Wan
committed
nvfp4 transpose
qiyuw
committed
bitwise identical transpose and add multi-gpu test
qiyuw
committed
clean up
qiyuw
committed

Commits on Dec 10, 2025

fix create columnwise and add debugging and testing code
qiyuw
committed

Commits on Dec 11, 2025

fix and debug
qiyuw
committed
fix loss mismatch, add debugging code
qiyuw
committed
clean up debugging code
qiyuw
committed
minor
qiyuw
committed
clean up again
qiyuw
committed

Commits on Dec 12, 2025

more restrict shape in test
qiyuw
committed

Commits on Dec 17, 2025

fix conflict
qiyuw
committed

Commits on Dec 30, 2025

nvfp4 scale transpose
qiyuw
committed

Commits on Jan 5, 2026

partial cast optimize
qiyuw
committed
partial cast optimize patch
qiyuw
committed
fix typo
qiyuw
committed

Commits on Jan 6, 2026

reduce kernel launches
qiyuw
committed

Commits on Jan 7, 2026

fused kernel
qiyuw
committed

Commits on Jan 8, 2026

optimize transpose kernel
qiyuw
committed
multi-tensor apply for transpose
qiyuw
committed
multi-tensor for amax
qiyuw
committed
multi-tensor for partial cast
qiyuw
committed

Commits on Jan 21, 2026

Merge branch 'main' into fp4_primary_weights
qiyuw
committed
clean up
qiyuw
committed

Commits on Jan 27, 2026

fix syntax error
qiyuw
committed

Commits on Jan 28, 2026

bf16 to fp4 quant
qiyuw
committed

Commits on Feb 19, 2026

Revert "bf16 to fp4 quant"
qiyuw
committed
Merge branch 'main' into fp4_primary_weights
qiyuw
committed
file permission and minor
qiyuw
committed
remove prints
qiyuw
committed
[pre-commit.ci] auto fixes from pre-commit.com hooks
pre-commit-ci[bot]
committed