[megatron] feat: Share actor and ref in LoRA

[HollowMan6]

For compute_ref_log_prob, we can do that by disabling lora layers temporarily for the forward pass, as base weight are frozen and only lora layers are trained.

This has already been supported in FSDP LoRA.

What does this PR do?

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, 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).)

_{✨ Presented to you with Mind Lab - A Lab for Experiential Intelligence.}

[gemini-code-assist]

Code Review

This pull request introduces a valuable feature by enabling LoRA for Megatron models, allowing the actor and reference models to share weights. The implementation correctly uses a context manager to disable LoRA adapters for computing reference log probabilities. My review includes two high-severity suggestions: one to address duplicated code for determining lora_rank across multiple files, and another to fix a performance issue where entropy is unnecessarily calculated and then discarded.

[gemini-code-assist]

Code Review

This pull request refactors the LoRA configuration handling to support a nested structure and introduces a mechanism to share the actor and reference models in megatron_workers.py when using LoRA. This is achieved by temporarily disabling the LoRA adapter to compute reference log probabilities, which is a clean and efficient approach.

My main feedback is regarding the duplicated logic for determining the lora_rank across four different trainer files. I've left comments with suggestions to refactor this into a shared utility function for better maintainability. The rest of the changes look good.

[HollowMan6]

Hi @ISEEKYAN @wuxibin89 ! Since NVIDIA-NeMo/Megatron-Bridge#1811 has been merged into main, I now mark this PR as ready, please let me know if you have any comments!

[Copilot]

Pull request overview

This PR implements the ability to share actor and reference models in LoRA training for the Megatron backend by temporarily disabling LoRA layers during reference log probability computation. This aligns with the existing FSDP LoRA implementation.

Key changes:

• Modified compute_ref_log_prob to redirect to compute_log_prob with LoRA disabled when using LoRA
• Updated compute_log_prob to support a conditional LoRA mode via is_lora flag in meta_info
• Enhanced ray_trainer.py to support both old and new LoRA configuration formats

Reviewed changes

Copilot reviewed 7 out of 7 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
verl/workers/megatron_workers.py	Implements LoRA sharing logic in compute_ref_log_prob and compute_log_prob methods with adapter context management
verl/trainer/ppo/ray_trainer.py	Updates ref_in_actor detection to support nested lora config structure with backward compatibility
examples/grpo_trainer/run_qwen3moe-30b_megatron_lora.sh	Updates recommended Megatron-Bridge commit hash for better MoE LoRA support
examples/grpo_trainer/run_qwen2-7b_math_megatron_lora.sh	Updates recommended Megatron-Bridge commit hash for better MoE LoRA support
docs/advance/ppo_lora.rst	Updates documentation with new recommended Megatron-Bridge commit hash
.github/workflows/e2e_ppo_trainer_megatron_vllm_2.yml	Updates CI to use newer Megatron-Bridge and Megatron-LM commits
.github/workflows/e2e_ppo_trainer_megatron_vllm.yml	Updates CI to use newer Megatron-Bridge and Megatron-LM commits

---

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

[gemini-code-assist]

Code Review

This pull request introduces a significant feature by enabling weight sharing between the actor and reference models for LoRA with the Megatron backend. The implementation is clean, using a context manager to temporarily disable LoRA adapters for computing reference log probabilities. The configuration changes are also consistent across different trainers to support a more structured LoRA setup. I've identified one area of code duplication that could be refactored to improve maintainability, which I've detailed in a specific comment.

[HollowMan6]

Update: It seems to be fixed with 59a3c5e

pip3 install --no-deps --force-reinstall .

I'm not sure why but I can't reproduce the error locally, and it seems to work fine...

I turned this back to draft due to https://github.com/volcengine/verl/actions/runs/20815124533/job/59788776831?pr=4673, pending investigation.

[wuxibin89]

@HollowMan6 We will soon switch default engine from legacy (fsdp_workers.py, megatron_workers.py) to new model engine (engine_workers.py). Could you take lora feature parity between new and legacy model engine?
We will change the use_legacy_worker_impl default value from auto to disable, while user still can switch back to legacy engine by trainer.use_legacy_worker_impl=enable. The legacy engine is going to be removed in v0.9.

[HollowMan6]

Could you take lora feature parity between new and legacy model engine? We will change the use_legacy_worker_impl default value from auto to disable, while user still can switch back to legacy engine by trainer.use_legacy_worker_impl=enable. The legacy engine is going to be removed in v0.9.

Sure! I think I have already tried to make it compatible with the new model engine in #4063, I will test this locally and see if there's any issue

[HollowMan6]

@wuxibin89 The following PRs should ensure features parity for LoRA between Megatron backend and FSDP backend, when we either set trainer.use_legacy_worker_impl to enable or disable:

• [worker] feat: New engine share actor and ref for LoRA #4867
• [worker] fix: new engine saves megatron LoRA adapters checkpoints #4866
• [megatron] feat: LoRA adapter only refit (TensorLoRARequest) #4632

While the last one (#4632) is still pending for NVIDIA-NeMo/Megatron-Bridge#1766, the first 2 should be ready for merge. Please let me know if you have any comments!