[trainer, fsdp, megatron] feat: support one_step_off_policy on Ascend NPU

[baymax591]

What does this PR do?

Since Ray's collective communication interface does not support the hccl backend, we refer to the example code of vLLM and complete the weight synchronization between actor and rollout.

Checklist Before Starting

• Search for similar PRs. Paste at least one query link here: pr#2924,pr4307
• 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 support for one_step_off_policy on Ascend NPUs by replacing ray.util.collective with a manual torch.distributed process group setup, which is a sound approach to support the hccl backend. The changes are consistent across the modified files. However, I've found a critical race condition in recipe/one_step_off_policy/ray_trainer.py during the weight synchronization group creation. The initialization for the actor worker group is not awaited, which could lead to runtime errors. I've provided a suggestion to fix this issue by ensuring both actor and rollout worker groups are fully initialized before proceeding.

[lzxdjb]

@baymax591
I am using the one-step-off-policy method for multi-machine training, but I encounter the following error during execution:

After investigating, I traced the error to the following code path:

File: ./recipe/one_step_off_policy/distributed_util.py

Line: 61

At this line, the code calls a utility function from vLLM:
File: ./vllm/distributed/utils.py

However, the implementation in vllm/distributed/utils only supports IPv4 addresses. When the training environment uses IPv6, this results in a failure during distributed initialization.

Therefore, I guess the one-step-off-policy distributed training pipeline currently does not support IPv6?