[PD] Correct the ip and port env (#2450)

1. rename `VLLM_LLMDD_RPC_PORT` to `VLLM_ASCEND_LLMDD_RPC_PORT` to make
the prefix the same in vllm-ascend
2. enable `VLLM_ASCEND_LLMDD_RPC_IP` env for PD feature.

Signed-off-by: wangxiyuan <[email protected]>

Author: wangxiyuan

vllm_ascend/distributed/llmdatadist_c_mgr_connector.py

@@ -181,7 +181,7 @@ def __init__(self, vllm_config: VllmConfig, engine_id: Optional[str]):
         dp_rank_local = self.vllm_config.parallel_config.data_parallel_rank_local
         tp_size = self.vllm_config.parallel_config.tensor_parallel_size
 
-        self.port = dp_rank_local * tp_size + envs_ascend.VLLM_LLMDD_RPC_PORT if dp_rank_local is not None else tp_size + envs_ascend.VLLM_LLMDD_RPC_PORT
+        self.port = dp_rank_local * tp_size + envs_ascend.VLLM_ASCEND_LLMDD_RPC_PORT if dp_rank_local is not None else tp_size + envs_ascend.VLLM_ASCEND_LLMDD_RPC_PORT
 
         self._reqs_need_recv: dict[str, tuple[Request, list[int]]] = {}
 
@@ -344,8 +344,8 @@ def __init__(self, vllm_config: VllmConfig):
 
     def listen_for_agent_metadata_req(self, event: threading.Event):
         assert self.local_agent_metadata is not None
-        port = envs_ascend.VLLM_LLMDD_RPC_PORT + self.local_dp_rank * self.tp_size + self.tp_rank if self.local_dp_rank is not None else envs_ascend.VLLM_LLMDD_RPC_PORT + self.tp_size + self.tp_rank
-        url = f"tcp://0.0.0.0:{port}"
+        port = envs_ascend.VLLM_ASCEND_LLMDD_RPC_PORT + self.local_dp_rank * self.tp_size + self.tp_rank if self.local_dp_rank is not None else envs_ascend.VLLM_ASCEND_LLMDD_RPC_PORT + self.tp_size + self.tp_rank
+        url = f"tcp://{envs_ascend.VLLM_ASCEND_LLMDD_RPC_IP}:{port}"
         msg_encoder = msgspec.msgpack.Encoder()
         msg_decoder = msgspec.msgpack.Decoder()
         msg_to_send = msg_encoder.encode(self.local_agent_metadata)

vllm_ascend/envs.py

@@ -123,11 +123,11 @@
     # remote worker.
     "VLLM_ASCEND_LLMDD_RPC_IP":
     lambda: os.getenv("VLLM_ASCEND_LLMDD_RPC_IP", "0.0.0.0"),
-    # `LLMDataDistCMgrConnector` required variable. `VLLM_LLMDD_RPC_PORT` is used as the
+    # `LLMDataDistCMgrConnector` required variable. `VLLM_ASCEND_LLMDD_RPC_PORT` is used as the
     # rpc communication listening port, which will be used to receive the agent metadata from the
     # remote worker.
-    "VLLM_LLMDD_RPC_PORT":
-    lambda: int(os.getenv("VLLM_LLMDD_RPC_PORT", 5557)),
+    "VLLM_ASCEND_LLMDD_RPC_PORT":
+    lambda: int(os.getenv("VLLM_ASCEND_LLMDD_RPC_PORT", 5557)),
     # Whether to enable mla_pa for deepseek mla decode, this flag will be removed after its available torch_npu is public accessible
     # and the mla_pa will be the default path of deepseek decode path.
     "VLLM_ASCEND_MLA_PA":