[ShengDev]

Fixes vllm-project#9208.

Contributor intent (required human answers)

1. What outcome do you want to achieve?

Stop the READY CI flake in tests/diffusion/test_omni_diffusion_config_master_port.py::TestOmniDiffusionConfigMasterPort::test_master_port_env_takes_precedence_over_kwarg (AssertionError: assert 34510 == 34473 on Buildkite build 17710) and land that fix on upstream main.

2. What is the highest-leverage variable for that outcome?

Whether the equality assertion depends on a port from get_open_port() still being free when settle_port calls is_port_available. That window is a TOCTOU flake. The +37 step is the intended fallback, not a production bug.

3. How will this work move that variable?

Change only the unit test. Stub vllm_omni.diffusion.data.is_port_available so env precedence, an explicit port, the busy-port +37 step, and the default path are deterministic. Also stub test_default_master_port_is_available, which still re-bound a live port in the first version of vllm-project#9210. Leave OmniDiffusionConfig._resolve_master_port and settle_port unchanged.

4. Whose customer or user needs does this address, and what are those needs?

Contributors whose pull requests run the Simple · Engine&Entrypoints&Diffusion unit-test step. They need that step to fail only when master-port precedence is wrong, not when another process takes an unreserved ephemeral port.

Purpose

Before: the equality tests called get_open_port(), which binds a port and then releases it, and then required OmniDiffusionConfig.master_port to equal that value. settle_port checks availability again and, when the port is busy, adds 37. Buildkite build 17710 logged Port 34473 was unavailable, using port 34510 instead and failed assert 34510 == 34473 (34473 + 37 = 34510). test_default_master_port_is_available had the same window: it resolved a port and then bound it again.

After: those tests control vllm_omni.diffusion.data.is_port_available and assert the selected port. MASTER_PORT is not compared against a live bind, the kwarg is not consulted when the env var is set, a busy candidate steps by 37, and the default path accepts the first probed port. Production resolution is unchanged.

vllm-project#9210 used this stub and was closed without merge. This PR re-applies that approach and includes the default-path stub from the start. vllm-project#9215 is an open change to the same file and does not reference vllm-project#9208.

docs/design/module/diffusion/diffusion_runtime.md owns vllm_omni/diffusion/data.py and does not specify master-port selection. No user guide or example documents this path, so documentation is unchanged.

Test Plan

vLLM Version: 0.32.0 (CPU unit test; no GPU)

vLLM-Omni Commit: c9f15bc63 on cursor/stabilize-master-port-tests-1619 (base upstream 0278901f8)

Regression command:

pytest tests/diffusion/test_omni_diffusion_config_master_port.py -q

Test Result

Pending local pytest on this commit. GPU and the full Buildkite shard are out of scope: the diff is the unit test only.

AI assistance: Cursor Grok 4.7 drafted the test change from the closed vllm-project#9210 approach and this description. The production resolver was checked in vllm_omni/diffusion/data.py and left unchanged.