[trainer] refactor: code refactor for diffusion training

[zhtmike]

What does this PR do?

Major change:

• Decouple LLM and diffusion configs, losses, etc., for easier integration of diffusion models/algorithms/engines in the future. (resolve comment from #5951)
  • Diffusion configs are now all placed in verl/trainer/config/diffusion, including diffusion_actor, dp_diffusion_actor, diffusion_fsdp, diffusion_rollout, etc.
  • Diffusion config classes are now all placed in verl/workers/config/diffusion, including actor, model, rollout.
  • Clean LLM-specific configs such as use_remove_padding, or unused configs. The content of _generated_diffusion_trainer.yaml has now been reduced by ~60%.

Other changes:

• Clean the diffusion agent loop output; remove dead code originally borrowed from the LLM agent loop, such as input_id, attention_mask, etc., which are not used in diffusion training currently.
• Refactor and remove extra_configs in diffusion configs. It is too loose and may cause confusion for users. (resolve comment from #5951)
• Decouple LLM and diffusion losses to make diffusion losses clearer.
• Code cleanup for diffusion trainer, diffusion agent loop, etc.
• Change the vllm-omni rollout import to registration, following the diffusion trainer pattern. (resolve comment from #5951)

Add concise overview of what this PR aims to achieve or accomplish. Reference related GitHub issues and PRs that help with the review.

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, vllm_omni, rollout, trainer, ci, training_utils, recipe, hardware, deployment, ray, worker, single_controller, misc, perf, model, algo, env, tool, ckpt, doc, data, cfg, reward, fully_async, one_step_off
  • 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 refactors the diffusion training pipeline, reorganizing configurations into a dedicated directory and introducing a registry for custom vLLM-Omni pipelines. It also adds several CI sanity checks for PR metadata and code quality, along with a pyproject.toml file for project management. A review comment correctly identified that the PR title validation script is missing the new diffusion and omni module tags, which would result in CI failures for PRs using these tags.

[SamitHuang]

I think we can add doc ci after doc system is built

[zhtmike]

dropped.

[SamitHuang]

can drop requirements* if we use pyproject.toml?

[zhtmike]

it is required by sanity check CI, etc.