修复all2all报维度不匹配问题

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 = False,
+        moe_permute_fusion: bool = True,
         moe_shared_expert_overlap: bool = False
     ):
         super().__init__(num_experts, top_k, hidden_size,

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

@@ -212,7 +212,6 @@ class MoEAlltoAllTokenDispatcher(MoETokenDispatcher):
             if self.use_all_gather:
                 # Gather is not supported for some devices such as TPUs.
                 # Use all-gather instead.
-
                 num_global_tokens_per_expert = expert_parallel_all_gather(num_local_tokens_per_expert) \
                     .reshape(self.ep_size, self.tp_size, self.num_experts) \
                     .transpose(0, 1)
@@ -319,7 +318,7 @@ class MoEAlltoAllTokenDispatcher(MoETokenDispatcher):
             num_out_tokens=self.num_out_tokens,
             fused=self.config.moe_permute_fusion
         )
-        
+
         # Perform expert parallel AlltoAll communication
         # tokens_per_expert = self._maybe_dtoh_and_synchronize(
         #     "before_ep_alltoall", tokens_per_expert
@@ -327,7 +326,7 @@ class MoEAlltoAllTokenDispatcher(MoETokenDispatcher):
 
         ###test##############
         #cuda_dtoh_stream.synchronize()
-        #cuda_dtoh_sync_event.synchronize()
+        cuda_dtoh_sync_event.synchronize()
         ###test##############
 
         global_input_tokens = all_to_all(
@@ -462,10 +461,10 @@ class MoEAlltoAllTokenDispatcher(MoETokenDispatcher):
                         self.num_global_tokens_per_local_expert, record_stream=on_side_stream
                     )
 
-                #cuda_dtoh_sync_event.record()
+                cuda_dtoh_sync_event.record()
 
-        if point == self.cuda_sync_point:
-            # Synchronize with the dtoh stream at self.cuda_sync_point.
-            cuda_dtoh_stream.synchronize()
+        # if point == self.cuda_sync_point:
+        #     # Synchronize with the dtoh stream at self.cuda_sync_point.
+        #     cuda_dtoh_stream.synchronize()
 
         return tokens_per_expert
\ No newline at end of file