[recipe, ckpt] feat: bypass the CPU in the checkpoint engine for fully_async mode.

[ZSL98]

What does this PR do?

Bypass the CPU in the checkpoint engine for fully_async mode.

Currently, in fully_async mode, each time the weights are updated, they are offloaded to the CPU, then copied to the h2d_buffer, and finally copied to the broadcast buffer.

The offload-to-CPU operation seems unnecessary. So this PR adds an additional configuration named bypass_cpu to bypass the CPU cache. Early-stage experiments show benefits.

Checklist Before Starting

• Search for similar PRs. Paste at least one query link here:
  • [WIP][BREAKING][worker, ckpt] support checkpoint engine for sync parameters in hybrid mode #4602
  • [wip][BREAKING][recipe, ckpt]add checkpoint engine for one step off policy #4601
  • [BREAKING][recipe, ckpt] feat: support parameter sync by checkpoint-engine. only for fully_async mode. #4427
• Format the PR title as [{modules}] {type}: {description} (This will be checked by the CI)
  • {modules} include fsdp, megatron, 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).)

[gemini-code-assist]

Code Review

This pull request introduces an optimization to bypass the CPU in the checkpoint engine for fully_async mode, which should improve performance by reducing data copies between host and device. A new configuration bypass_cpu is added to control this behavior. The changes in checkpoint_engine.py and fsdp_workers.py correctly implement this new path. My main feedback is to improve the naming of some functions and variables that have become misleading with this change, to ensure code clarity and maintainability.

[zpltys]

The reason I cache params to cpu is to further support parameters version management. But I think it is a good pr. Can you update the relevant experiments/readme as in the commit.

[ZSL98]

@zpltys Sorry for the delay. I have updated the README and conducted some basic experiments.

With a single-node Ascend A2 server and a 4B model, enabling bypass_cpu can reduce the total sync weight time from 0.84s to 0.64s (a 24% reduction), reduce the cache cost from 0.16s to 0.10s, and reduce the update cost from 0.26s to 0.21s. All numbers are averages over 500 trials. bypass_cpu is set to False by default, so it does not conflict with your original setting. Since we do not have GPU machines, the numbers on larger-scale GPU nodes remain untested. Hope these numbers can be tested by the community.

Besides, fully_async, together with the checkpoint engine, can work well on Ascend NPUs with some adaptations, as in #4658. We may first have a discussion internally and then figure out how to support fully_async and the checkpoint engine more gracefully. Many thanks.

[ZSL98]

@zpltys @ArronHZG Would you please help review the code?