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

[fix]解决dp模式mtp卡住问题

See merge request dcutoolkit/deeplearing/vllm!280

vllm/forward_context.py

@@ -135,7 +135,9 @@ def set_forward_context(
     if need_to_track_batchsize:
         forward_start_time = time.perf_counter()
     dp_metadata: Optional[DPMetadata] = None
-    if vllm_config.parallel_config.data_parallel_size > 1 and (
+    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 (
             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

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

vllm/v1/spec_decode/eagle.py

@@ -88,6 +88,9 @@ class EagleProposer:
                                    1,
                                    device=device,
                                    dtype=torch.int32)
+        
+        self.dp_size = vllm_config.parallel_config.data_parallel_size
+        self.enable_expert_parallel = vllm_config.parallel_config.enable_expert_parallel
 
     def propose(
         self,
@@ -529,6 +532,17 @@ class EagleProposer:
                 self.hidden_states[:num_tokens],
             )
 
+        if self.dp_size > 1 and self.enable_expert_parallel and self.num_speculative_tokens > 1:
+            for _ in range(self.num_speculative_tokens - 1):
+                with set_forward_context(attn_metadata,
+                                 self.vllm_config,
+                                 num_tokens=num_tokens):
+                    self.model(
+                        self.input_ids[:num_tokens],
+                        self.positions[:num_tokens],
+                        self.hidden_states[:num_tokens],
+                    )
+
     def validate_same_kv_cache_group(self,
                                      kv_cache_config: KVCacheConfig) -> None:
         """