Skip to content

feat: support RAO grouped agent rollouts - #1610

Open
MaxLEAF3824 wants to merge 3 commits into
areal-project:mainfrom
MaxLEAF3824:feat/rao-grouped-rollout-contract
Open

feat: support RAO grouped agent rollouts#1610
MaxLEAF3824 wants to merge 3 commits into
areal-project:mainfrom
MaxLEAF3824:feat/rao-grouped-rollout-contract

Conversation

@MaxLEAF3824

Copy link
Copy Markdown

Summary

  • preserve RAO session, node, group, and reward metadata through agent rollout export
  • add FA4 attention support and retain tree-attention handling for packed trajectories
  • compute pre-merge group credit while keeping valid sibling trajectories when one child fails
  • reject invalid RAO exports without aborting the rollout worker
  • bind unset Agent proxy token ceilings to the active train or eval generation context

Tests

  • python -m pytest -q tests/test_eval_agent_token_limit.py tests/test_grouped_rollout_workflow.py — 6 passed
  • python -m ruff check ... on the six changed/new files — passed
  • python -m ruff format --check ... on the six changed/new files — passed
  • python -m pre_commit run --all-files — formatting, whitespace, YAML/JSON, large-file, private-key, lockfile, SPDX, CODEOWNERS, and notebook hooks passed; two environment hooks could not run because the hook environment lacks tomllib and an executable named python

Hardware

The added tests are CPU unit tests. No distributed or GPU integration run was performed for this PR.

🤖 Generated with Claude Code

MaxLEAF3824 and others added 3 commits August 11, 2026 12:56
Carry recursive rollout metadata through the OpenAI proxy and training data path so downstream RAO workloads can preserve node identity and depth. Add the FA4 backend required by the team's B200 environment.

Key changes:
- Preserve session, parent, depth, and interaction metadata
- Export recursive rollout structure for downstream training
- Register FlashAttention 4 as a supported FSDP backend
- Refresh generated CLI references

Co-Authored-By: Claude <noreply@anthropic.com>
Allow agent workflows to attach node-level rollout metadata and process complete rollout groups before flattening, enabling strict root LOO and zero-variance rejection without changing PPO.\n\nCo-Authored-By: Claude <noreply@anthropic.com>
Keep valid siblings when one rollout fails, propagate stable group and rejection metadata, and bind Agent proxy limits to the active train or eval generation context.

Key changes:
- isolate per-child grouped rollout failures
- reject invalid RAO exports without aborting the worker
- preserve group metadata through tensor export
- apply role-specific Agent token ceilings

Co-Authored-By: Claude <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.

1 participant