[Kernel][Performance] Add FlashInfer cutedsl NVFP4 GEMM backend (#42235)

Signed-off-by: Mohammad Miadh Angkad <176301910+mmangkad@users.noreply.github.com>
This commit is contained in:
Mohammad Miadh Angkad
2026-06-22 16:17:18 -04:00
committed by GitHub
parent 2b4a7491ec
commit d1a38c2762
6 changed files with 93 additions and 1 deletions
@@ -956,6 +956,15 @@ class AsyncTPPass(VllmFusionPatternMatcherPass):
a_scale_view=a_scale_view,
)
)
self.register(
FlashInferAllGatherFP4Pattern(
self.model_dtype,
self.device,
"cute-dsl",
use_8x4_sf_layout=False,
a_scale_view="float8",
)
)
# NVFP4 reduce-scatter does not need scale communication: FP4
# scales are consumed by the local GEMM and only BF16 partial
# outputs are reduced. Keep this PR scoped to the all-gather