diff --git a/vllm/kernels/linear/nvfp4/cutedsl.py b/vllm/kernels/linear/nvfp4/cutedsl.py
index 4c7ae95e..c8368918 100644
--- a/vllm/kernels/linear/nvfp4/cutedsl.py
+++ b/vllm/kernels/linear/nvfp4/cutedsl.py
@@ -58,7 +58,8 @@ def _cutedsl_nvfp4_linear(
     padded_rows = cutedsl_ceil_div(num_tokens, 128) * 128
     if num_tokens < padded_rows:
         x_fp4_padded = torch.zeros(padded_rows, x_fp4.shape[1],
-                                    dtype=x_fp4.dtype, device=x.device)
+                                    dtype=torch.uint8, device=x.device
+                                    ).view(torch.float4_e2m1fn_x2)
         x_fp4_padded[:num_tokens] = x_fp4
     else:
         x_fp4_padded = x_fp4