[https://nvbugs/5670793][fix] Solve trtllm-serve launch_disaggregated…

tensorrt_llm/commands/serve.py

@@ -25,7 +25,7 @@
                                               parse_disagg_config_file,
                                               parse_metadata_server_config_file)
 from tensorrt_llm.llmapi.llm_utils import update_llm_args_with_extra_dict
-from tensorrt_llm.llmapi.mpi_session import find_free_port
+from tensorrt_llm.llmapi.mpi_session import find_free_ipc_addr
 from tensorrt_llm.llmapi.reasoning_parser import ReasoningParserFactory
 from tensorrt_llm.logger import logger, severity_map
 from tensorrt_llm.serve import OpenAIDisaggServer, OpenAIServer
@@ -641,10 +641,10 @@ def _launch_disaggregated_leader(sub_comm, instance_idx: int, config_file: str,
 
     # This mimics the behavior of trtllm-llmapi-launch
     # TODO: Make the port allocation atomic
-    free_port = find_free_port()
+    free_ipc_addr = find_free_ipc_addr()
     os.environ[LlmLauncherEnvs.TLLM_SPAWN_PROXY_PROCESS] = "1"
-    os.environ[LlmLauncherEnvs.TLLM_SPAWN_PROXY_PROCESS_IPC_ADDR.
-               value] = f"tcp://127.0.0.1:{free_port}"
+    os.environ[
+        LlmLauncherEnvs.TLLM_SPAWN_PROXY_PROCESS_IPC_ADDR.value] = free_ipc_addr
     os.environ[DisaggLauncherEnvs.TLLM_DISAGG_RUN_REMOTE_MPI_SESSION_CLIENT.
                value] = "1"
     os.environ[DisaggLauncherEnvs.TLLM_DISAGG_INSTANCE_IDX] = str(instance_idx)

tensorrt_llm/llmapi/mpi_session.py

@@ -544,6 +544,13 @@ def find_free_port() -> int:
         return s.getsockname()[1]
 
 
+def find_free_ipc_addr() -> str:
+    import os
+    import tempfile
+    import uuid
+    return f'ipc://{os.path.join(tempfile.gettempdir(), "rpc_" + str(uuid.uuid4()))}'
+
+
 def get_mpi_world_size() -> int:
     # avoid cyclic import
     from ..executor.utils import get_spawn_proxy_process_env