From 22a175921c4718f195d1eaddf5a26c9aaa07a326 Mon Sep 17 00:00:00 2001 From: aoshen02 Date: Sat, 8 Aug 2026 12:23:26 +0800 Subject: [PATCH] [BugFix] Reject invalid data-parallel RPC ports (#51469) Signed-off-by: aoshen02 Co-authored-by: OpenAI Codex --- tests/test_config.py | 15 +++++++++++++++ vllm/config/parallel.py | 4 ++-- vllm/engine/arg_utils.py | 3 ++- vllm/v1/engine/utils.py | 10 +++++----- 4 files changed, 24 insertions(+), 8 deletions(-) diff --git a/tests/test_config.py b/tests/test_config.py index 9b51aab857..227fc08c93 100644 --- a/tests/test_config.py +++ b/tests/test_config.py @@ -380,6 +380,21 @@ def test_async_scheduling_with_pipeline_parallelism_is_allowed(): assert cfg.scheduler_config.async_scheduling is True +def test_data_parallel_rpc_port_has_fixed_default(): + assert ParallelConfig().data_parallel_rpc_port == 29550 + + +@pytest.mark.parametrize("port", [1, 29550, 65535]) +def test_data_parallel_rpc_port_accepts_valid_ports(port: int): + assert ParallelConfig(data_parallel_rpc_port=port).data_parallel_rpc_port == port + + +@pytest.mark.parametrize("port", [-1, 0, 65536]) +def test_data_parallel_rpc_port_rejects_invalid_ports(port: int): + with pytest.raises(ValidationError): + ParallelConfig(data_parallel_rpc_port=port) + + def test_reconfigure_for_independent_dp_rank_on_multinode_dense_model(): parallel_config = ParallelConfig( tensor_parallel_size=8, diff --git a/vllm/config/parallel.py b/vllm/config/parallel.py index 37b3349a5b..83c4e6fe53 100644 --- a/vllm/config/parallel.py +++ b/vllm/config/parallel.py @@ -140,8 +140,8 @@ class ParallelConfig: """Local rank of the data parallel group, set only in SPMD mode.""" data_parallel_master_ip: str = "127.0.0.1" """IP of the data parallel master.""" - data_parallel_rpc_port: int = 29550 - """Port for data parallel messaging.""" + data_parallel_rpc_port: int = Field(default=29550, ge=1, le=65535) + """Fixed port for data parallel messaging, shared by all nodes.""" data_parallel_master_port: int = 29500 """Port of the data parallel master.""" data_parallel_backend: DataParallelBackend = "mp" diff --git a/vllm/engine/arg_utils.py b/vllm/engine/arg_utils.py index 55a712e0f2..da4359f9ad 100644 --- a/vllm/engine/arg_utils.py +++ b/vllm/engine/arg_utils.py @@ -1100,7 +1100,8 @@ class EngineArgs: "--data-parallel-rpc-port", "-dpp", type=int, - help="Port for data parallel RPC communication.", + help="Fixed port for data parallel RPC communication. All nodes " + "must use the same port.", ) parallel_group.add_argument( "--data-parallel-backend", diff --git a/vllm/v1/engine/utils.py b/vllm/v1/engine/utils.py index be440e9741..7736062d99 100644 --- a/vllm/v1/engine/utils.py +++ b/vllm/v1/engine/utils.py @@ -1165,11 +1165,11 @@ def launch_core_engines( if parallel_config.enable_elastic_ep: handshake_local_only = False - # Preserve "port=0 means auto-pick" for the handshake address, which - # is consumed by engines spawned in this process and so cannot defer - # port resolution to bind time. - rpc_port = parallel_config.data_parallel_rpc_port or get_open_port() - handshake_address = get_engine_client_zmq_addr(handshake_local_only, host, rpc_port) + handshake_address = get_engine_client_zmq_addr( + handshake_local_only, + host, + parallel_config.data_parallel_rpc_port, + ) if local_engines_only and dp_rank > 0: assert not handshake_local_only