[ci] breaks down V1 Test into 3 groups of approx 30 minutes runtime

[jeanschmidt]

This PR is part of the CI sprint initiative, with the goal of reduce CI latency for tests.

This is following the first step from #23668

It breaks down 'V1 Tests' in 3 groups, this is based on the analysis of the wallclock for test runs:

[2025-08-27T02:20:48Z] + pytest -v -s v1/core - 00:03 
[2025-08-27T02:23:19Z] + pytest -v -s v1/engine - 00:13
[2025-08-27T02:36:32Z] + pytest -v -s v1/entrypoints - 00:36
[2025-08-27T03:04:35Z] + pytest -v -s v1/executor - 00:02
[2025-08-27T03:06:21Z] + pytest -v -s v1/sample - 00:07
[2025-08-27T03:13:35Z] + pytest -v -s v1/logits_processors - 00:05
[2025-08-27T03:18:39Z] + pytest -v -s v1/worker - 00:01
[2025-08-27T03:19:29Z] + pytest -v -s v1/structured_output - 00:00
[2025-08-27T03:19:44Z] + pytest -v -s v1/spec_decode - 00:08
[2025-08-27T03:27:46Z] + pytest -v -s v1/kv_connector/unit - 00:04
[2025-08-27T03:31:17Z] + pytest -v -s v1/metrics - 00:01
[2025-08-27T03:32:25Z] + pytest -v -s v1/test_serial_utils.py - 00:00
[2025-08-27T03:32:39Z] + pytest -v -s v1/test_utils.py - 00:00
[2025-08-27T03:32:56Z] + pytest -v -s v1/test_oracle.py - 00:03
[2025-08-27T03:35:06Z] + pytest -v -s v1/test_metrics_reader.py - 00:00
# TODO: accuracy does not match, whether setting
# VLLM_USE_FLASHINFER_SAMPLER or not on H100.
[2025-08-27T03:35:20Z] + pytest -v -s v1/e2e - 00:20
# Integration test for streaming correctness (requires special branch).
[2025-08-27T03:55:34Z] + pip install -U git+https://github.com/robertgshaw2-redhat/lm-evaluation-harness.git@streaming-api - 00:06
[2025-08-27T03:55:48Z] + pytest -v -s entrypoints/openai/correctness/test_lmeval.py::test_lm_eval_accuracy_v1_engine
[2025-08-27T04:01:13Z] 33m=============================== warnings summary ===============================

as a important note, there are 3 outliers:

• v1/e2e - 00:20
• v1/entrypoints - 00:36
• v1/engine - 00:13

it seems the most efficient approach would be breaking into the following groups:

• 1. v1/e2e - 00:20
• 2. v1/entrypoints - 00:36
• 3. ...everything else

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.

💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels.

Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run fastcheck CI which starts running only a small and essential subset of CI tests to quickly catch errors.

You ask your reviewers to trigger select CI tests on top of fastcheck CI.

Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.

To run CI, PR reviewers can either: Add ready label to the PR or enable auto-merge.

If you have any questions, please reach out to us on Slack at https://slack.vllm.ai.

🚀

[gemini-code-assist]

Code Review

This pull request aims to reduce CI latency by splitting the 'V1 Tests' job into three parallel jobs. The implementation correctly separates the longest-running test suites, v1/e2e and v1/entrypoints, into their own jobs. However, there are two key areas for improvement. First, the new 'V1 Test (others)' job is missing a necessary source file dependency, which could cause it to not run when it should. Second, the distribution of tests among the new jobs is unbalanced, leaving one job with a runtime of nearly an hour. A more balanced distribution could further reduce the overall CI wall-clock time.