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

Signed-off-by: BWAAEEEK <jooho414@gmail.com>

vllm/model_executor/kernels/linear/mxfp4/flashinfer.py

@@ -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,

vllm/utils/flashinfer.py

@@ -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