Skip to content

fix: fail closed for unsatisfied GPU experiment profiles - #202

Open
Lluvia-C wants to merge 1 commit into
sapientinc:mainfrom
Lluvia-C:fix/issue-83-accelerator-preflight
Open

Lluvia-C wants to merge 1 commit into
sapientinc:mainfrom
Lluvia-C:fix/issue-83-accelerator-preflight

Conversation

@Lluvia-C

Copy link
Copy Markdown

Summary

  • Validate the declared default resource profile before the central scheduler publishes its endpoint or starts workers.
  • Fail closed when a GPU-only task has no structurally usable accelerator, while preserving transient resource-pressure queueing.
  • Reject generations that never admitted an evaluation-capable job before writing generation results or committing the next boundary.
  • Add offline regression contracts for scheduler preflight, cohort zero-evaluation handling, and custom topology generation handling.

Validation

  • 132 targeted and related unit tests passed.
  • Ruff check and format check passed for all changed files.
  • Targeted Pyrefly check passed with 0 errors.
  • git diff --check passed.
  • No real GPU, API key, network task, or external task checkout is required by these tests.

The broader local unit run also passed 2915 tests; remaining failures were environment-dependent checks involving missing optional dependencies, macOS process-group behavior, and generated bytecode/cache integrity.

Validate the default accelerator profile before publishing the scheduler endpoint and reject generations that never admitted an evaluation job. Preserve transient resource queueing and add offline regression contracts for issue sapientinc#83.

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant