修复all2all报维度不匹配问题

vllm/model_executor/layers/fused_moe/ep_moe/ep_moe_utlis.py

@@ -327,8 +327,7 @@ def all_to_all(group, input, output_split_sizes, input_split_sizes):
         output = input.new_empty(
             size=[sum(output_split_sizes)] + list(input.size()[1:]),
             dtype=input.dtype,
-            #device=torch.cuda.current_device(),
-            device=input.device,
+            device=torch.cuda.current_device()
         )
 
     torch.distributed.all_to_all_single(
@@ -336,8 +335,7 @@ def all_to_all(group, input, output_split_sizes, input_split_sizes):
         input,
         output_split_sizes=output_split_sizes,
         input_split_sizes=input_split_sizes,
-        group=group,
-        async_op=True
+        group=group
     )
     return output
 

vllm/model_executor/layers/fused_moe/ep_moe/layer.py

@@ -180,7 +180,7 @@ class EPMoE(FusedMoE):
         routed_scaling_factor: Optional[float] = None,
         enable_eplb: bool = False,
         num_redundant_experts: int = 0,
-        moe_permute_fusion: bool = True,
+        moe_permute_fusion: bool = False,
         moe_shared_expert_overlap: bool = False
     ):
         super().__init__(num_experts, top_k, hidden_size,