Remove contiguous before Flashinfer groupwise fp8 gemm (#6804)

python/sglang/srt/layers/quantization/fp8_utils.py

@@ -166,11 +166,13 @@ def flashinfer_gemm_w8a8_block_fp8_linear(
         input_2d, block_size[1], column_major_scales=False
     )
 
-    x_scale_input = x_scale.T.contiguous()
-    weight_scale_input = weight_scale.T.contiguous()
-
     output = gemm_fp8_nt_groupwise(
-        q_input, weight, x_scale_input, weight_scale_input, out_dtype=input_2d.dtype
+        q_input,
+        weight,
+        x_scale,
+        weight_scale,
+        scale_major_mode="K",
+        out_dtype=input_2d.dtype,
     )
 
     if bias is not None: