[algo] Add SVD-LoRA GRPO

[alekseymalakhov11]

What does this PR do?

This PR implements the experimental SVD-LoRA-GRPO method derived by paper ESSA: Evolutionary Strategies for Scalable Alignment

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 an experimental implementation of SVD-LoRA-GRPO. The core logic is in the new verl/utils/experimental/svd_lora.py file, with integrations in FSDP utilities and worker configurations. My review has identified a few critical issues. There's a TypeError in SVDLinear.create_from_weight that will cause a runtime crash. More importantly, the implementation in apply_svd_lora seems to misinterpret the paper by performing SVD on LoRA's A and B matrices separately, instead of on their product W=BA. I've also pointed out a high-severity issue in fsdp_utils.py where reconstructed weights are not moved to the CPU, and a high-severity maintainability issue regarding the use of a deprecated PyTorch function. Please address these points to ensure the correctness and robustness of this new feature.

[alekseymalakhov11]

We’re implementing SVD-LoRA-GRPO on top of the existing SFT-LoRA adapter. Since #3523 is already in progress and touches the same components, we won’t push our implementation here to avoid conflicts. We’ll wait for #3523 to merge, then rebase and update this PR.

[alekseymalakhov11]

Example training run — Qwen2.5-7B on PRM800K with an SFT-LoRA adapter.
The SFT adapter was trained on a different (non-overlapping) PRM800K training subset.
Hyperparameters:

max_prompt_length = 512
max_response_length = 4096
train_batch_size = 512
lr = 0.01
n = 8
loss_agg_mode = "token-mean"
lora_rank = 16
lora_alpha = 32
target_modules = ["q_proj", "k_proj", "v_proj", "o_proj"]
ppo_mini_batch_size = 64
strategy = "fsdp2"
nnodes = 1
n_gpus_per_node = 8