diff --git a/vllm_mindspore/model_executor/models/mf_models/deepseekv3_weight_processor.py b/vllm_mindspore/model_executor/models/mf_models/deepseekv3_weight_processor.py
index cb93a89fed7d09c89bb6eb3fce163f8e313b2d43..f2613adb61575e7a1bc08fa78a7d89e4cac7886f 100644
--- a/vllm_mindspore/model_executor/models/mf_models/deepseekv3_weight_processor.py
+++ b/vllm_mindspore/model_executor/models/mf_models/deepseekv3_weight_processor.py
@@ -346,6 +346,14 @@ class DeepseekV3WeightProcessor(BaseWeightProcessor):
         w1_scale_ms_name = f"model.layers.{layer_id}.feed_forward.routed_experts.ffn.w1._layer.matmul.weight_scale"
         w2_scale_ms_name = f"model.layers.{layer_id}.feed_forward.routed_experts.ffn.w2._layer.matmul.weight_scale"
         w3_scale_ms_name = f"model.layers.{layer_id}.feed_forward.routed_experts.ffn.w3._layer.matmul.weight_scale"
+        if self.ep_method == EPMethod.ALLTOALL:
+            # use dispatch_quant and dequantSwiGluQuant which supports group_list to optimize performance
+            w1_ms_name = w1_ms_name.replace("._layer.weight", ".weight")
+            w2_ms_name = w2_ms_name.replace("._layer.weight", ".weight")
+            w3_ms_name = w3_ms_name.replace("._layer.weight", ".weight")
+            w1_scale_ms_name = w1_scale_ms_name.replace("._layer.matmul.weight_scale", ".weight_scale")
+            w2_scale_ms_name = w2_scale_ms_name.replace("._layer.matmul.weight_scale", ".weight_scale")
+            w3_scale_ms_name = w3_scale_ms_name.replace("._layer.matmul.weight_scale", ".weight_scale")
 
         w1_list, w2_list, w3_list, w1_scale_list, w2_scale_list, w3_scale_list = \
             self.infer_quant_process_moe(src_hf_dir, hf_weight_map, layer_id)
@@ -361,6 +369,13 @@ class DeepseekV3WeightProcessor(BaseWeightProcessor):
         if ffn_concat:
             # w_gate_hidden
             w_gate_hidden_name = f"model.layers.{layer_id}.feed_forward.routed_experts.ffn.w_gate_hidden._layer.weight"
+            # w_scale_gate_hidden
+            w_scale_gate_hidden_name = \
+                f"model.layers.{layer_id}.feed_forward.routed_experts.ffn.w_gate_hidden._layer.matmul.weight_scale"
+            if self.ep_method == EPMethod.ALLTOALL:
+                w_gate_hidden_name = w_gate_hidden_name.replace("._layer.weight", ".weight")
+                w_scale_gate_hidden_name = w_scale_gate_hidden_name.replace("._layer.matmul.weight_scale",
+                                                                            ".weight_scale")
             w_gate_hidden_np = np.concatenate(
                 [w1_ms_stack_param, w3_ms_stack_param], axis=1)
             w_gate_hidden_param = ms.from_numpy(w_gate_hidden_np).permute(
@@ -369,9 +384,6 @@ class DeepseekV3WeightProcessor(BaseWeightProcessor):
                 w_gate_hidden_param,
                 name=w_gate_hidden_name,
                 requires_grad=False)
-            # w_scale_gate_hidden
-            w_scale_gate_hidden_name = \
-                f"model.layers.{layer_id}.feed_forward.routed_experts.ffn.w_gate_hidden._layer.matmul.weight_scale"
 
             w_scale_gate_hidden_np = np.concatenate(
                 [w1_scale_ms_stack_param, w3_scale_ms_stack_param], axis=1)