[Feature]Moe alltoallv communication optimization for unquantized RL training sence & alltoallv support dpo

weijinqian_v1 · weijinqian_v1 · commit 47439e82b7d8 · 2025-07-01T13:32:51.000+08:00
Signed-off-by: weijinqian_v1 &lt;weijinqian@huawei.com&gt;
diff --git a/vllm_ascend/ops/fused_moe.py b/vllm_ascend/ops/fused_moe.py
@@ -1233,6 +1233,7 @@ def forward(self,
             global_redundant_expert_num=self.global_redundant_expert_num,
             shared_experts=shared_experts if self.torchair_graph_enabled
             and self.enable_multistream_moe and not is_prefill else None,
+            token_dispatcher=self.token_dispatcher
         )
 
         if shared_experts:

Original file line number	Diff line number	Diff line change
`@@ -1233,6 +1233,7 @@ def forward(self,`
`1233`	`1233`	`global_redundant_expert_num=self.global_redundant_expert_num,`
`1234`	`1234`	`shared_experts=shared_experts if self.torchair_graph_enabled`
`1235`	`1235`	`and self.enable_multistream_moe and not is_prefill else None,`
	`1236`	`+ token_dispatcher=self.token_dispatcher`
`1236`	`1237`	`)`
`1237`	`1238`
`1238`	`1239`	`if shared_experts:`