Skip to content

Hook proposal#1774

Draft
andrija-s wants to merge 38 commits intoTHUDM:mainfrom
GymPod:hook-proposal
Draft

Hook proposal#1774
andrija-s wants to merge 38 commits intoTHUDM:mainfrom
GymPod:hook-proposal

Conversation

@andrija-s
Copy link
Copy Markdown

Proposal for hooks that inject callbacks into slime training.

hleehlee-amazon and others added 30 commits February 5, 2026 19:43
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Documents the branch synchronization workflow and development practices
for contributing to upstream while maintaining a clean sync branch.

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
fix workflow syncing

fix workflow syncing
github-actions bot and others added 8 commits March 26, 2026 02:26
Add slime/hooks.py with an Op enum and hook() context manager that
wraps key training loop operations. When no callbacks are registered,
hooks are near-zero-cost no-ops.

This enables downstream consumers (e.g. OTel tracing) to register
pre/post callbacks without monkey-patching internal functions.

See: THUDM#1728

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
NODE_INIT fires once per Ray actor process during init, enabling
downstream consumers to start background tasks (e.g. GPU/CPU metrics
polling) on every node in a multi-node run.

Uses SLIME_NODE_NAME env var for portable node identification.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
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.

3 participants