Speed up macOS smoke test

[mgoin]

Purpose

Use a smaller model to fix timeout issues

Test Plan

Test Result

Confirmed https://github.com/vllm-project/vllm/actions/runs/19479854049/job/55748976858 works fine manually

The main issue is that it seems distributed init takes 15 minutes to finish

(EngineCore_DP0 pid=5220) INFO 11-18 20:15:25 [parallel_state.py:1208] world_size=1 rank=0 local_rank=0 distributed_init_method=tcp://192.168.64.24:49303 backend=gloo
(EngineCore_DP0 pid=5220) INFO 11-18 20:31:10 [parallel_state.py:1394] rank 0 in world size 1 is assigned as DP rank 0, PP rank 0, TP rank 0, EP rank 0

I think the likely bottleneck is torch.distributed.new_group() in GroupCoordinator.__init__(). For world_size=1, initialize_model_parallel() creates 5 GroupCoordinator instances (TP, DCP, PP, DP, EP), each creating 2 groups (device + CPU), totaling 10 new_group() calls. Even for single-process groups, PyTorch may still perform slow initialization. The main optimization would be to skip or optimize group creation for single-process cases, but that's a larger change

---

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.
• (Optional) Release notes update. If your change is user facing, please update the release notes draft in the Google Doc.

[gemini-code-assist]

Note

Gemini is unable to generate a review for this pull request due to the file types involved not being currently supported.

[chatgpt-codex-connector]

💡 Codex Review

vllm/.github/workflows/macos-smoke-test.yml

Line 72 in 3a311da

"model": "Qwen/Qwen3-0.6B",

Update completion request to match served model

The smoke test now launches the server with trl-internal-testing/tiny-random-LlamaForCausalLM, but the completion request still posts "model": "Qwen/Qwen3-0.6B". With OpenAI-compatible APIs, a request for an unloaded model returns an error, so this curl check will consistently fail even though the server is running, breaking the workflow on every run. The request should target the model that was actually started.

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[mgoin]

Confirmed https://github.com/vllm-project/vllm/actions/runs/19479854049/job/55748976858 works fine manually

The main issue is that it seems distributed init takes 15 minutes to finish

(EngineCore_DP0 pid=5220) INFO 11-18 20:15:25 [parallel_state.py:1208] world_size=1 rank=0 local_rank=0 distributed_init_method=tcp://192.168.64.24:49303 backend=gloo
(EngineCore_DP0 pid=5220) INFO 11-18 20:31:10 [parallel_state.py:1394] rank 0 in world size 1 is assigned as DP rank 0, PP rank 0, TP rank 0, EP rank 0

I think the likely bottleneck is torch.distributed.new_group() in GroupCoordinator.__init__(). For world_size=1, initialize_model_parallel() creates 5 GroupCoordinator instances (TP, DCP, PP, DP, EP), each creating 2 groups (device + CPU), totaling 10 new_group() calls. Even for single-process groups, PyTorch may still perform slow initialization. The main optimization would be to skip or optimize group creation for single-process cases, but that's a larger change