Fix OmniDiffusionConfig master_port selection for parallel launches (#3803)

This commit is contained in:
Samit authored and GitHub committed 2026-05-22 11:12:29 +08:00
1 parent 076c34abfb
commit 8e937400dc
2 files changed
+69 -3

No files matched your search

@@ -0,0 +1,49 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""Unit tests for OmniDiffusionConfig master_port resolution (issue #3794)."""
import socket
import pytest
from vllm_omni.diffusion.data import OmniDiffusionConfig
pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
class TestOmniDiffusionConfigMasterPort:
def test_honors_master_port_env(self, monkeypatch: pytest.MonkeyPatch) -> None:
monkeypatch.setenv("MASTER_PORT", "43210")
config = OmniDiffusionConfig(model="test")
assert config.master_port == 43210
def test_honors_explicit_master_port(self, monkeypatch: pytest.MonkeyPatch) -> None:
monkeypatch.delenv("MASTER_PORT", raising=False)
config = OmniDiffusionConfig(model="test", master_port=40000)
assert config.master_port == 40000
def test_master_port_env_takes_precedence_over_kwarg(self, monkeypatch: pytest.MonkeyPatch) -> None:
monkeypatch.setenv("MASTER_PORT", "43210")
config = OmniDiffusionConfig(model="test", master_port=40000)
assert config.master_port == 43210
def test_explicit_master_port_is_stable_without_jitter(self, monkeypatch: pytest.MonkeyPatch) -> None:
monkeypatch.delenv("MASTER_PORT", raising=False)
ports = {OmniDiffusionConfig(model="test", master_port=40123).master_port for _ in range(5)}
assert ports == {40123}
def test_default_master_port_is_available(self, monkeypatch: pytest.MonkeyPatch) -> None:
monkeypatch.delenv("MASTER_PORT", raising=False)
config = OmniDiffusionConfig(model="test")
assert config.master_port is not None
with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as sock:
sock.setsockopt(socket.SOL_SOCKET, socket.SO_REUSEADDR, 1)
sock.bind(("", config.master_port))
def test_settle_port_steps_when_port_is_busy(self, monkeypatch: pytest.MonkeyPatch) -> None:
monkeypatch.delenv("MASTER_PORT", raising=False)
with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as sock:
sock.bind(("", 0))
busy_port = sock.getsockname()[1]
config = OmniDiffusionConfig(model="test", master_port=busy_port)
assert config.master_port == busy_port + 37
+20 -3
View File
@@ -606,6 +606,25 @@ class OmniDiffusionConfig:
return False
def _resolve_master_port(self) -> int:
"""Resolve torch.distributed master port without unnecessary random jitter.
Precedence:
1. ``MASTER_PORT`` environment variable (set by orchestrators for multi-replica launch).
2. Explicit ``master_port`` passed at construction time.
3. An OS-assigned ephemeral port when neither is provided.
"""
from vllm.utils.network_utils import get_open_port
from vllm_omni.diffusion import envs
env_port = envs.MASTER_PORT
if env_port is not None:
return self.settle_port(env_port, port_inc=37)
if self.master_port is not None:
return self.settle_port(self.master_port, port_inc=37)
return self.settle_port(get_open_port(), port_inc=37)
def settle_port(self, port: int, port_inc: int = 42, max_attempts: int = 100) -> int:
"""
Find an available port with retry logic.
@@ -642,9 +661,7 @@ class OmniDiffusionConfig:
)
def __post_init__(self):
# TODO: remove hard code
initial_master_port = (self.master_port or 30005) + random.randint(0, 100)
self.master_port = self.settle_port(initial_master_port, 37)
self.master_port = self._resolve_master_port()
if isinstance(self.profiler_config, dict):
from vllm.config import ProfilerConfig