Skip to content

docs: explain partial rollout checkpoint states - #1

Merged
ananthsub merged 1 commit into
mainfrom
ananthsub/checkpoint-design-foundation
Sep 1, 2026
Merged

docs: explain partial rollout checkpoint states#1
ananthsub merged 1 commit into
mainfrom
ananthsub/checkpoint-design-foundation

Conversation

@ananthsub

Copy link
Copy Markdown
Owner

Summary

  • distinguish logical rollout, agent execution, checkpoint phase, and model admission state
  • explain failure handling and normal-path versus checkpoint-time overhead
  • compare restart, drain, replay, central storage, uncoordinated snapshots, and process snapshots with the participant design
  • align the API inventory with the Gym stack ending at feat(checkpoint): adapt initial stateful environments NVIDIA-NeMo/Gym#2946 and call out remaining production gaps

Test plan

  • parse the HTML with Python's standard HTML parser
  • parse all embedded JavaScript with Node.js
  • verify the document diff has no whitespace errors

The Pages workflow is intentionally not triggered by this PR. Publication remains a manual workflow dispatch.

Clarify participant state machines, failure handling, runtime overhead, architectural alternatives, and the limits of the current Gym API stack.

Signed-off-by: Ananth Subramaniam <ansubramania@nvidia.com>
@ananthsub
ananthsub merged commit 9c61f0f into main Sep 1, 2026
@ananthsub
ananthsub deleted the ananthsub/checkpoint-design-foundation branch September 1, 2026 22:36
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