ci-workflow-guide

To debug a failing SGLang CI job, analyze the fast-fail layers, retry classification, and stage health checks to identify failure propagation and timeout symptoms. Check workflow dependencies and whether the failure stems from a specific test suite or runner timeout.

LPT-based test partitioning in SGLang CI distributes GPU testing loads across matrix runners by registering test suites and applying the LPT heuristic to balance execution times. This determines runner selection and expected job counts for each stage.

Adding a GPU test stage requires updating the GitHub Actions workflow job, registering the test suite, configuring matrix partitioning, maintaining wait-for-jobs gating counts, and including final result aggregation to ensure proper stage health checks.

Yes, SGLang CI supports targeted stage reruns alongside PR, scheduled, and standard execution modes. You can rerun specific failed jobs while respecting workflow dependencies, concurrency rules, and stage gating requirements.

SGLang CI stages time out or skip due to misconfigured wait-for-jobs gating, unmet stage dependencies, or exceeded runner limits. Examine the stage health checks, timeout configurations, and workflow concurrency settings to isolate the cause.

SGLang CI orchestrates workflows using GitHub Actions stage ordering, PR gating, concurrency controls, scheduled runs, and targeted stage reruns. These mechanisms ensure test pipelines execute reliably across different execution modes.