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

Signed-off-by: weijinqian_v1 <weijinqian@huawei.com>

Author: weijinqian_v1

vllm_ascend/models/deepseek_dbo.py

@@ -172,7 +172,7 @@ def __init__(
             top_k=config.num_experts_per_tok,
             hidden_size=config.hidden_size,
             intermediate_size=config.moe_intermediate_size,
-            reduce_results=True,
+            reduce_results=False,
             renormalize=config.norm_topk_prob,
             quant_config=quant_config,
             use_grouped_topk=True,
@@ -190,7 +190,7 @@ def __init__(
                 intermediate_size=intermediate_size,
                 hidden_act=config.hidden_act,
                 quant_config=quant_config,
-                reduce_results=True,
+                reduce_results=True if not envs_ascend.VLLM_ASCEND_ENABLE_MOE_ALL2ALL_SEQ else False,
                 prefix=f"{prefix}.shared_experts",
             )
         CustomDeepseekDBOMoE.top_k = config.num_experts_per_tok

vllm_ascend/worker/model_runner_v1.py

@@ -603,7 +603,7 @@ def _check_dbo_is_valid(self, query_lens: torch.Tensor,
         ]:
             return False
         # considering the case that one dp rank may enable dbo while others may not
-        if not self.vllm_config.model_config.use_mla or not envs_ascend.VLLM_ASCEND_ENABLE_DBO:
+        if not envs_ascend.VLLM_ASCEND_ENABLE_DBO:
             return False
         # TODO: remove it if token-level microbatch is enabled
         [token_index,