diff --git a/vllm_ascend/eplb/core/eplb_device_transfer_loader.py b/vllm_ascend/eplb/core/eplb_device_transfer_loader.py
index 9a8a323..67e4d56 100644
--- a/vllm_ascend/eplb/core/eplb_device_transfer_loader.py
+++ b/vllm_ascend/eplb/core/eplb_device_transfer_loader.py
@@ -50,10 +50,6 @@ class D2DExpertWeightLoader:
             )
             return
 
-        # If neither send nor receive task is needed for this layer on this rank, return
-        if not (expert_send_info or expert_recv_info):
-            return
-
         self.updated_expert_map = updated_expert_map
 
         self.layer_id = layer_id
diff --git a/vllm_ascend/ops/expert_load_balancer.py b/vllm_ascend/ops/expert_load_balancer.py
index 5b03ffa..08c0106 100644
--- a/vllm_ascend/ops/expert_load_balancer.py
+++ b/vllm_ascend/ops/expert_load_balancer.py
@@ -48,7 +48,7 @@ class ExpertLoadBalancer(object):
 
     def generate_expert_placement_map(self):
         expert_placement_map = torch.full(
-            (self.layers_num, self.ranks_num, self.global_expert_num),
+            (self.layers_num, self.ranks_num, self.num_experts),
             -1,
             dtype=torch.int32,
         )
@@ -71,7 +71,7 @@ class ExpertLoadBalancer(object):
                 result_dict[key] = []
             result_dict[key].append(idx)
 
-        log2phy_map = torch.full((self.ranks_num, self.global_expert_num),
+        log2phy_map = torch.full((self.ranks_num, self.num_experts),
                                  -1,
                                  dtype=torch.int32)
         for rank in range(self.ranks_num):
diff --git a/vllm_ascend/ops/moe/moe_mlp.py b/vllm_ascend/ops/moe/moe_mlp.py
index 5ee7d70..2c73421 100644
--- a/vllm_ascend/ops/moe/moe_mlp.py
+++ b/vllm_ascend/ops/moe/moe_mlp.py
@@ -105,6 +105,9 @@ def quant_apply_mlp(hidden_states: torch.Tensor,
                 group_list=group_list,
                 output_dtype=torch.int32)[0]
             # act_fn: swiglu
+            group_diff = torch.diff(group_list)
+            new_group = torch.cat([group_list[0].unsqueeze(0), group_diff],
+                                  dim=0)
             hidden_states, swiglu_out_scale = torch_npu.npu_dequant_swiglu_quant(
                 x=hidden_states,
                 weight_scale=w1_scale,
@@ -112,7 +115,7 @@ def quant_apply_mlp(hidden_states: torch.Tensor,
                 bias=None,
                 quant_scale=None,
                 quant_offset=None,
-                group_index=group_list,
+                group_index=new_group,
                 activate_left=True,
                 quant_mode=1,
             )
diff --git a/vllm_ascend/ops/moe/token_dispatcher.py b/vllm_ascend/ops/moe/token_dispatcher.py
index da9b2be..a6e1f55 100644
--- a/vllm_ascend/ops/moe/token_dispatcher.py
+++ b/vllm_ascend/ops/moe/token_dispatcher.py
@@ -122,18 +122,14 @@ class TokenDispatcherWithMC2(MoETokenDispatcher):
         expert_map: torch.Tensor,
         global_redundant_expert_num: int = 0,
     ):
-        if self.with_quant:
-            quant_mode = 2
-            moe_expert_num = len(expert_map)
-        else:
-            quant_mode = 0
-            moe_expert_num = len(expert_map)
+        quant_mode = 2 if self.with_quant else 0
+        self.moe_expert_num = len(expert_map) + global_redundant_expert_num
         kwargs_mc2 = {
             "x": hidden_states,
             "expert_ids": topk_ids,
             "expert_shard_type": 0,
             "shared_expert_rank_num": 0,
-            "moe_expert_num": moe_expert_num,
+            "moe_expert_num": self.moe_expert_num,
             "global_bs": 0,
             "expert_token_nums_type": 0,
         }
@@ -229,7 +225,6 @@ class TokenDispatcherWithMC2(MoETokenDispatcher):
         assert self.topk_weights is not None
         assert self.topk_ids is not None
         assert self.output is not None
-        moe_expert_num = len(self.expert_map)
         # moeCombine
         kwargs_mc2 = {
             "expand_x": hidden_states,
@@ -237,7 +232,7 @@ class TokenDispatcherWithMC2(MoETokenDispatcher):
             "expert_scales": self.topk_weights.to(torch.float32),
             "expert_shard_type": 0,
             "shared_expert_rank_num": 0,
-            "moe_expert_num": moe_expert_num,
+            "moe_expert_num": self.moe_expert_num,
             "global_bs": 0,
         }
         if self.with_quant:
@@ -359,7 +354,7 @@ class TokenDispatcherWithAllGather(MoETokenDispatcher):
             hidden_states = hidden_states * \
                 topk_weights.to(hidden_states.dtype)
         if expert_map is not None:
-            global_num_experts = len(expert_map)
+            global_num_experts = len(expert_map) + global_redundant_expert_num
             mask = (expert_map[topk_ids] != -1)
             self.topk_weights = topk_weights * mask
             first_expert_idx = get_ep_group(
diff --git a/vllm_ascend/quantization/w8a8_dynamic.py b/vllm_ascend/quantization/w8a8_dynamic.py
index 59b0479..709f488 100644
--- a/vllm_ascend/quantization/w8a8_dynamic.py
+++ b/vllm_ascend/quantization/w8a8_dynamic.py
@@ -250,7 +250,7 @@ class AscendW8A8DynamicFusedMoEMethod:
         return moe_comm_method.fused_experts(
             hidden_states=x,
             w1=layer.w13_weight,
-            w1_scale=layer.w13_weight_scale_fp32,
+            w1_scale=layer.w13_weight_scale.to(torch.float32),
             w2=layer.w2_weight,
             w2_scale=layer.w2_weight_scale,
             topk_weights=topk_weights,