[WIP][data] feat: TransferQueue - integrate TransferQueue into main codebase

[0oshowero0]

What does this PR do?

This PR integrates the experimental TransferQueue feature into the main codebase.

Key Changes

1. New Trainer Implementation

• Added RayPPOTrainerTransferQueue which inherits from RayPPOTrainer.
• Implemented BatchMeta-based operations to replace the distribution of real data objects (DataProto/TensorDict).

2. Configuration

• Updated ppo_trainer.yaml and ppo_megatron_trainer.yaml to include TransferQueue-related configurations.

3. Integration & Entry Point

• Modified RayPPOTrainerTransferQueue to adapt to the new features.
• Added switch logic in main_ppo.py to enable the TransferQueue via config.\
• Update CI & scripts of TQ

Checklist Before Starting

• Search for similar PRs. Paste at least one query link here: ...
• Format the PR title as [{modules}] {type}: {description} (This will be checked by the CI)
  • {modules} include fsdp, megatron, veomni, sglang, vllm, rollout, trainer, ci, training_utils, recipe, hardware, deployment, ray, worker, single_controller, misc, perf, model, algo, env, tool, ckpt, doc, data, cfg, reward
  • If this PR involves multiple modules, separate them with , like [megatron, fsdp, doc]
  • {type} is in feat, fix, refactor, chore, test
  • If this PR breaks any API (CLI arguments, config, function signature, etc.), add [BREAKING] to the beginning of the title.
  • Example: [BREAKING][fsdp, megatron] feat: dynamic batching

Test

For changes that can not be tested by CI (e.g., algorithm implementation, new model support), validate by experiment(s) and show results like training curve plots, evaluation results, etc.

API and Usage Example

Demonstrate how the API changes if any, and provide usage example(s) if possible.

# Add code snippet or script demonstrating how to use this

Design & Code Changes

Demonstrate the high-level design if this PR is complex, and list the specific changes.

Checklist Before Submitting

Important

Please check all the following items before requesting a review, otherwise the reviewer might deprioritize this PR for review.

• Read the Contribute Guide.
• Apply pre-commit checks: pre-commit install && pre-commit run --all-files --show-diff-on-failure --color=always
• Add / Update the documentation.
• Add unit or end-to-end test(s) to the CI workflow to cover all the code. If not feasible, explain why: ...
• Once your PR is ready for CI, send a message in the ci-request channel in the verl Slack workspace. (If not accessible, please try the Feishu group (飞书群).)
• If your PR is related to the recipe submodule, please also update the reference to the submodule commit via git submodule update --remote or cd recipe && git pull origin main.

[gemini-code-assist]

Code Review

This pull request integrates the experimental TransferQueue feature into the main codebase, which is a significant refactoring. The changes involve moving and deleting several files, updating configurations, and modifying the training loop to support BatchMeta-based operations. My review focuses on critical correctness and performance issues. I've identified a critical bug that could lead to a crash in the training loop and a high-severity performance issue related to inefficient data fetching from the TransferQueue. Addressing these points will improve the stability and performance of the new feature.

[Copilot]

Pull request overview

This PR integrates the experimental TransferQueue feature into the main codebase, enabling efficient data transfer in distributed PPO training. The integration includes a new trainer implementation, configuration updates, and modifications to support BatchMeta-based operations.

Changes:

• Added RayPPOTrainerTransferQueue as a new trainer implementation inheriting from RayPPOTrainer
• Updated configuration files to include TransferQueue-related settings (storage backend, batch management)
• Modified entry point (main_ppo.py) to conditionally enable TransferQueue based on configuration
• Integrated TransferQueue client creation in worker initialization and decorator functions

Reviewed changes

Copilot reviewed 25 out of 25 changed files in this pull request and generated 7 comments.

Show a summary per file

File	Description
verl/trainer/ppo/ray_trainer_tq.py	New trainer implementation with TransferQueue support
verl/trainer/main_ppo.py	Added conditional logic to select trainer based on TransferQueue config
verl/utils/transferqueue_utils.py	Enhanced utility functions for BatchMeta/TensorDict conversion
verl/workers/engine_workers.py	Added TransferQueue client initialization in workers
verl/single_controller/base/decorator.py	Extended register decorator with TransferQueue parameters
verl/workers/config/engine.py	Added TransferQueueConfig dataclass
Configuration files	Updated YAML configs with TransferQueue settings
CI/Scripts	Updated test scripts and workflow to use new entry point

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

In line 1094, there's a typo in the field name check: 'resampled_idx' should be 'pf_ppo_reweight_idx' based on the field being selected in line 1095. This inconsistency will cause the condition to never be True.

Suggested change

	if "resampled_idx" in batch_meta.field_names and self.config.transferqueue.enable:
	if "pf_ppo_reweight_idx" in batch_meta.field_names and self.config.transferqueue.enable: