Merge branch 'v0.9.2-dev-wm-1202' into 'v0.9.2-dev'

回退错误修改，避免dp模式漏掉的dispatch和combine通信

See merge request dcutoolkit/deeplearing/vllm!282

vllm/forward_context.py

@@ -136,8 +136,7 @@ def set_forward_context(
         forward_start_time = time.perf_counter()
     dp_metadata: Optional[DPMetadata] = None
     dp_size = vllm_config.parallel_config.data_parallel_size
-    use_navie_ep = envs.VLLM_ALL2ALL_BACKEND == 'naive' and dp_size > 1 and vllm_config.parallel_config.enable_expert_parallel
-    if use_navie_ep and dp_size > 1 and (
+    if dp_size > 1 and (
             attn_metadata is not None or num_tokens is not None):
         dp_metadata = DPMetadata.make(vllm_config.parallel_config,
                                       attn_metadata, num_tokens or 0,

vllm/model_executor/layers/fused_moe/layer.py

@@ -1537,9 +1537,7 @@ class FusedMoE(torch.nn.Module):
 
         do_naive_dispatch_combine: bool = (
             self.dp_size > 1
-            and self.ep_size > 1
-            and envs.VLLM_ALL2ALL_BACKEND == 'naive')
-            #and not self.moe_parallel_config.use_deepep_ht_kernels)
+            and not self.moe_parallel_config.use_deepep_ht_kernels)
         if do_naive_dispatch_combine:
             hidden_states, router_logits = get_ep_group().dispatch(
                 hidden_states, router_logits)