remove torch.cat and replace it by TensorList[0]

Signed-off-by: huangxialu <huangxialu1@huawei.com>

Author: loukong33

vllm_ascend/ops/fused_moe.py

@@ -207,11 +207,9 @@ def fused_experts_with_mc2(
         group_list_type=1,
         group_type=0,
         group_list=group_list,
-    )
+    )[0]
 
-    # TODO: Remove this in the future.
-    gate_up_out = torch.cat(gate_up_out_list, dim=0)
-    gate_up_out = torch_npu.npu_swiglu(gate_up_out)
+    gate_up_out = torch_npu.npu_swiglu(gate_up_out_list)
 
     w2 = w2.transpose(1, 2)
     down_out_list = torch_npu.npu_grouped_matmul(
@@ -221,9 +219,8 @@ def fused_experts_with_mc2(
         group_list_type=1,
         group_type=0,
         group_list=group_list,
-    )
+    )[0]
 
-    down_out_list = torch.cat(down_out_list, dim=0)
 
     # moeCombine
     kwargs_mc2 = {
@@ -314,9 +311,8 @@ def apply_mlp(
         group_list_type=group_list_type,
         group_type=0,
         group_list=group_list,
-    )
+    )[0]
 
-    hidden_states = torch.cat(hidden_states, dim=0)
     hidden_states = torch_npu.npu_swiglu(hidden_states)
 
     w2 = w2.transpose(1, 2)
@@ -327,9 +323,8 @@ def apply_mlp(
         group_list_type=group_list_type,
         group_type=0,
         group_list=group_list,
-    )
+    )[0]
 
-    hidden_states = torch.cat(hidden_states, dim=0)
     return hidden_states
 
 
@@ -419,23 +414,20 @@ def fused_experts_with_all2all(
         group_list_type=0,
         group_type=0,
         group_list=expert_tokens,
-    )
+    )[0]
 
-    # TODO: Remove this in the future.
-    hidden_states = torch.cat(gate_up_out_list, dim=0)
-    hidden_states = torch_npu.npu_swiglu(hidden_states)
+    hidden_states = torch_npu.npu_swiglu(gate_up_out_list)
 
     w2 = w2.transpose(1, 2)
-    down_out_list = torch_npu.npu_grouped_matmul(
+    hidden_states = torch_npu.npu_grouped_matmul(
         x=[hidden_states],
         weight=[w2],
         split_item=2,
         group_list_type=0,
         group_type=0,
         group_list=expert_tokens,
-    )
+    )[0]
 
-    hidden_states = torch.cat(down_out_list, dim=0)
 
     if expert_map is not None:
         resorted_idx = torch.argsort(sorted_idx)
@@ -825,11 +817,9 @@ def fused_experts(
         group_list_type=0,
         group_type=0,
         group_list=expert_tokens,
-    )
+    )[0]
 
-    # TODO: Remove this in the future.
-    gate_up_out = torch.cat(gate_up_out_list, dim=0)
-    gate_up_out = torch_npu.npu_swiglu(gate_up_out)
+    gate_up_out = torch_npu.npu_swiglu(gate_up_out_list)
 
     w2 = w2.transpose(1, 2)
     down_out_list = torch_npu.npu_grouped_matmul(
@@ -839,9 +829,8 @@ def fused_experts(
         group_list_type=0,
         group_type=0,
         group_list=expert_tokens,
-    )
+    )[0]
 
-    down_out_list = torch.cat(down_out_list, dim=0)
 
     if expert_map is not None:
         weighted_down_out = down_out_list * sorted_weights.unsqueeze(1)