[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
@@ -56,7 +56,9 @@ class FlashInferMxFp4LinearKernel(MxFp4LinearKernel):
out_shape = x.shape[:-1] + (layer.output_size_per_partition,)
x_2d = x.reshape(-1, x.shape[-1])
x_fp4, x_scale = flashinfer_mxfp4_quantize(x_2d.contiguous())
x_fp4, x_scale = flashinfer_mxfp4_quantize(
x_2d.contiguous(), backend="cute-dsl"
)
out = flashinfer_scaled_fp4_mm(
x_fp4,
weight,