[Performance] Use CuTe-DSL for FlashInfer MXFP4 quantization (#48417)

Signed-off-by: BWAAEEEK <jooho414@gmail.com>
This commit is contained in:
JooHo Lee
2026-07-17 06:53:48 -07:00
committed by GitHub
parent f38f3d11fb
commit ce2aecc4dc
2 changed files with 6 additions and 2 deletions
+3 -1
View File
@@ -618,14 +618,16 @@ if has_flashinfer():
)
def flashinfer_mxfp4_quantize(
a: torch.Tensor,
backend: str,
) -> tuple[torch.Tensor, torch.Tensor]:
from flashinfer import mxfp4_quantize as _mxfp4_quantize
return _mxfp4_quantize(a)
return _mxfp4_quantize(a, backend=backend)
@torch.library.register_fake("vllm::flashinfer_mxfp4_quantize")
def flashinfer_mxfp4_quantize_fake(
a: torch.Tensor,
backend: str,
) -> tuple[torch.Tensor, torch.Tensor]:
m, k = a.shape
sf_vec_size = 32