Merge branch 'v0.9.2-dev' into v0.9.2-step3v

vllm/distributed/kv_transfer/kv_connector/v1/p2p/p2p_nccl_connector.py

@@ -214,7 +214,7 @@ class P2pNcclConnector(KVConnectorBase_V1):
                 kv_cache_layer = kv_cache[ \
                     forward_context.virtual_engine]
 
-                if self.p2p_nccl_engine.tensor_split_num == P2pNcclEngine.TENSOR_SPLIT_OFF:
+                if not envs.VLLM_P2P_ASYNC:
                     kv_cache = self.p2p_nccl_engine.recv_tensor(
                         request.request_id + "#" + layer_name)
 

vllm/distributed/kv_transfer/kv_connector/v1/p2p/p2p_nccl_engine.py

@@ -63,8 +63,6 @@ def set_p2p_nccl_context(num_channels: str):
 
 class P2pNcclEngine:
 
-    TENSOR_SPLIT_OFF = 0
-
     def __init__(self,
                  local_rank: int,
                  config: KVTransferConfig,
@@ -368,8 +366,6 @@ class P2pNcclEngine:
                     tensor_id = data["tensor_id"]
                     if "tensor_split_num" in data:
                         self.tensor_split_num = data["tensor_split_num"]
-                    else:
-                        self.tensor_split_num= self.TENSOR_SPLIT_OFF
                     try:
                         with torch.cuda.stream(self.recv_stream):
                             tensor = torch.empty(data["shape"],