[trainer] feat: Implemented VeomniEngine as a alternative training backend

[A1waysBeenHere]

What does this PR do?

This PR introduces a implementation of a VeOmniEngine for VERL, providing an alternative to the existing FSDP engine.

We plan to integrate the Veomni Engine in two phases. The first phase (as part of this PR) is to complete the engine code development and conduct basic validation via SFT. The second phase is to finish the integration of the RL workflow and supplemnt the relevant document.

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, sglang, vllm, rollout, trainer, ci, training_utils, recipe, hardware, deployment, ray, worker, single_controller, misc, perf, model, algo, env, tool, ckpt, doc, data
  • 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 VeomniEngine as a new training backend, which is a significant addition. The changes include new configuration files, modifications to existing configuration dataclasses, and the core implementation of the VeomniEngine. The implementation is still a draft with some commented-out code and NotImplementedErrors. I've identified a few critical issues that would cause runtime errors and a high-severity issue related to model evaluation mode. Addressing these will be crucial for making the engine functional.

[wuxibin89]

Can we reuse fsdp optimizer config: verl/trainer/config/optim/fsdp.yaml?

[A1waysBeenHere]

Can we reuse fsdp optimizer config: verl/trainer/config/optim/fsdp.yaml?

Directly use FSDP optimizer cannot work with the EP case, need to figure out why. Based on the error msg, it seems that there are 2 device_mesh when we enable EP. EP and (DP_shard, SP), they cannot share same device_mesh, and the optimizer cannot work in this suitiation. So, I might keep using VeOmni's optimizer for EP adaptation.

[ji-huazhong]

In Veomni, the parameter optimization for EP and non-EP parameters is managed separately by distinct optimizers, as seen here: https://github.com/ByteDance-Seed/VeOmni/blob/889cb3379a1143f4aa178ff55dbb3b1bcb788135/veomni/optim/optimizer.py#L311 @wuxibin89

[vermouth1992]

Great and beautiful work!!!

[vermouth1992]

Why dp_rank is torch.distributed.get_rank() when ulysses_size==1?

[vermouth1992]

Do we support EP/PP?

[A1waysBeenHere]

Do we support EP/PP?

only SP and EP works right now

[A1waysBeenHere]

Why dp_rank is torch.distributed.get_rank() when ulysses_size==1?

Lmao, I copyed this func with FSDPEngine style, and forgot to refine it. And yeah, I can simply return parallel_state.get_parallel_state().device_mesh.get_local_rank("dp"). Will fix it later