[Perf][Kernel] Optimize FP4 quantization kernels (SM100F) (#32520)

Signed-off-by: LopezCastroRoberto <rocastro@redhat.com>

vllm/compilation/activation_quant_fusion.py

@@ -152,6 +152,7 @@ class SiluMulNvfp4QuantPattern(ActivationQuantPattern):
                 input=result_silu_mul,
                 output_scale=output_scale,
                 input_scale=scale,
+                is_sf_swizzled_layout=True,
             )
             return at[1], at[2]
 

vllm/compilation/collective_fusion.py

@@ -946,6 +946,7 @@ class AllReduceFusedRMSNormStaticQuantNVFP4Pattern(BasePattern):
                 input=rms,
                 output_scale=output_scale,
                 input_scale=input_global_scale,
+                is_sf_swizzled_layout=True,
             )
 
             # quant_out, allreduce_output, output_scale
@@ -1043,6 +1044,7 @@ class AllReduceFusedAddRMSNormStaticQuantNVFP4Pattern(BasePattern):
                 input=rms,
                 output_scale=output_scale,
                 input_scale=input_global_scale,
+                is_sf_swizzled_layout=True,
             )
 
             # quant_out, allreduce_output, output_scale

vllm/compilation/fusion_attn.py

@@ -248,6 +248,7 @@ class AttentionNvfp4QuantPattern(AttentionQuantPattern):
                 input=attn_out_view,
                 output_scale=output_scale,
                 input_scale=input_scale,
+                is_sf_swizzled_layout=True,
             )
             output_scale_view = torch.ops.aten.view.dtype(at2[2], FP8_DTYPE)
             return at2[1], output_scale_view