Skip to content

Add compiled local partition lifecycle - #94

Open
insujang wants to merge 1 commit into
refactorfrom
refactor-oobleck
Open

Add compiled local partition lifecycle#94
insujang wants to merge 1 commit into
refactorfrom
refactor-oobleck

Conversation

@insujang

Copy link
Copy Markdown
Collaborator

Summary

  • split process-group-independent ownership compilation from partition activation
  • add explicit uneven pipeline stage specifications and stable logical state manifests
  • preserve the reusable model blueprint across partition swaps
  • add idempotent context retirement while keeping homogeneous materialize() compatibility

Why

Oobleck must choose rank-local ownership before WORLD exists and repeatedly replace heterogeneous pipeline partitions across membership generations. These generic lifecycle hooks keep elasticity policy in Oobleck while making Cornstarch partitions reusable.

Validation

  • pytest -q tests/distributed/test_compiled_plan.py tests/distributed/test_compiled_plan_direct_import.py (5 passed)
  • exercised by Oobleck real synthetic Llama one-step integration

This PR intentionally targets refactor and should remain open for upstream review.

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.

2 participants