[ROCm][MXFP4] Infer w4a4 quant method in rocm aiter fused moe (#29775)

Signed-off-by: ZhiweiYan-96 <zhiwei.yan@amd.com>

vllm/model_executor/layers/fused_moe/config.py

@@ -345,6 +345,10 @@ class FusedMoEQuantConfig:
     def use_mxfp4_w4a16(self) -> bool:
         return self._a1.dtype is None and self._w1.dtype == "mxfp4"
 
+    @property
+    def use_mxfp4_w4a4(self) -> bool:
+        return self._a1.dtype == "mxfp4" and self._w1.dtype == "mxfp4"
+
     @property
     def use_nvfp4_w4a4(self) -> bool:
         return self.quant_dtype == "nvfp4"

vllm/model_executor/layers/fused_moe/rocm_aiter_fused_moe.py

@@ -221,8 +221,8 @@ def rocm_aiter_fused_experts(
 
     else:
         quant_method = QuantMethod.NO.value
-        # quark moe for mxfp4 w_dtype
-        if quant_config.use_mxfp4_w4a16:
+        # quark moe for mxfp4 w_dtype mxfp4 a_dtype
+        if quant_config.use_mxfp4_w4a4:
             quant_method = QuantMethod.BLOCK_1X32.value
         # w8a8 block-scaled
         if quant_config.block_shape is not None and quant_config.use_fp8_w8a8: