fix(ci): extend SGLang GPU test timeout

[xianlubird]

Summary

• Increase the SGLang single-GPU test stage timeout from the reusable workflow default of 30 minutes to 45 minutes.
• Keep the VRAM-aware scheduler and individual pytest timeouts unchanged.
• Preserve the outer timeout as protection against a genuinely stuck test stage.

Background

Full CI for PR #12746 exposed how little headroom the current SGLang timeout has. In the failed run, the parallel GPU step started at 00:56:52 UTC and was terminated at 01:26:52 UTC, exactly at its 30-minute limit. The runner reported Executing the custom container implementation failed; this was not a pytest assertion failure.

The last scheduled test, tests/router/test_router_e2e_with_sglang.py::test_sglang_kv_router_basic[tcp], did not start until 01:25:22 UTC. Its own timeout was 270 seconds, but the outer step limit stopped the container after roughly 90 seconds. Individual test timeouts already begin when each subprocess starts, so queued time does not consume the test's own budget; the problem is that the workflow-level timeout includes collection, queueing, execution, and cleanup.

This is also close to occurring on successful runs. A recent successful SGLang parallel step took 29 minutes and 33 seconds, leaving only 27 seconds of headroom, while another completed in 13 minutes and 49 seconds. That variation makes a 30-minute stage limit brittle even when every test is healthy.

Implementation

Set gpu_test_timeout_minutes: 45 explicitly for the SGLang test job. This matches the existing vLLM single-GPU timeout and gives a queued test enough time to use its own timeout budget. The value remains an upper bound: successful jobs finish immediately and do not wait for the full 45 minutes.

This does not retry the GPU suite, ignore failures, or change any per-test timeout.

Validation

• Parsed .github/workflows/pr.yaml with PyYAML.
• Ran git diff --check.
• Compared the failed SGLang parallel step with two recent successful runs to verify the timeout boundary and runtime variance.

Summary by CodeRabbit

• Tests
  • Extended the GPU test job timeout to 45 minutes, allowing longer-running test suites to complete.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[dynamo-ops]

/ok to test 0945b8d

[devin-ai-integration]

Devin Review found 1 potential issue.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: e0d75225-9bc1-46a1-b20c-b7b04b5e16fe

📥 Commits

Reviewing files that changed from the base of the PR and between 6a4b074 and 0945b8d.

📒 Files selected for processing (1)

• .github/workflows/pr.yaml

---

Walkthrough

The pull request adds a 45-minute timeout to the sglang-test GPU test job in the pull request workflow.

Changes

GPU test timeout

Layer / File(s)	Summary
Configure GPU test timeout .github/workflows/pr.yaml	The sglang-test job now has a 45-minute GPU test timeout.

Estimated code review effort: 1 (Trivial) | ~2 minutes

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Description check	⚠️ Warning	The description covers motivation, implementation, and validation, but omits the required Related Issues section and reviewer-start guidance.	Add the required Related Issues section with an issue link or checked no-related-issue confirmation, and identify where reviewers should start.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title clearly summarizes the change to the SGLang GPU test timeout.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

---

_{Comment @coderabbitai help to get the list of available commands.}

[datadog-official]

🎯 Code Coverage (details)
• Patch Coverage: 100.00%
• Overall Coverage: 45.14% (-6.19%)

_{This comment will be updated automatically if new data arrives.
🔗 Commit SHA: 0945b8d | Docs | Datadog PR Page | Give us feedback!}

[dmitry-tokarev-nv]

This chang moved qualifying tests from gpu_1 to gpu_0 decreasing test runtime significantly: #12718
If PRs are still experiencing long SGL GPU parallel test runs - they may just have to rebase on / merge main
Example recent run with short SGL test runtime https://github.com/ai-dynamo/dynamo/actions/runs/31057166523/job/92478147343
I recommend closing this PR so we don't bloat our SGL tests runtime by allowing longer running test runs