[weifengpy]

What does this PR do?

for torch==2.6.0, distributed state dict is buggy and can leads to OOM
copy the fixed state dict api from torch==2.7.0 to verl/third_party. It's convinent for users who cannot upgrade to torch==2.7.0

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.

HYDRA_FULL_ERROR=1 CUDA_LAUNCH_BLOCKING=1 python3 -m verl.trainer.main_ppo       algorithm.adv_estimator=grpo       data.train_files=$HOME/data/gsm8k/train.parquet       data.val_files=$HOME/data/gsm8k/test.parquet       data.train_batch_size=512       data.max_prompt_length=1024       data.max_response_length=2048       data.filter_overlong_prompts=True       data.truncation='error'       data.image_key=images       actor_rollout_ref.model.path=Qwen/Qwen2.5-VL-32B-Instruct       actor_rollout_ref.model.use_remove_padding=True       actor_rollout_ref.model.enable_gradient_checkpointing=True       actor_rollout_ref.actor.strategy=fsdp2       actor_rollout_ref.actor.optim.lr=1e-6       actor_rollout_ref.actor.ppo_mini_batch_size=128       actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=10       actor_rollout_ref.actor.use_kl_loss=True       actor_rollout_ref.actor.kl_loss_coef=0.01       actor_rollout_ref.actor.kl_loss_type=low_var_kl       actor_rollout_ref.actor.entropy_coeff=0       actor_rollout_ref.rollout.name=vllm       actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=20       actor_rollout_ref.rollout.tensor_model_parallel_size=4       actor_rollout_ref.rollout.gpu_memory_utilization=0.6       actor_rollout_ref.rollout.enable_chunked_prefill=True       actor_rollout_ref.rollout.enforce_eager=False       actor_rollout_ref.rollout.free_cache_engine=False       actor_rollout_ref.rollout.n=5      actor_rollout_ref.ref.strategy=fsdp2       actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=20       actor_rollout_ref.ref.fsdp_config.param_offload=True       trainer.critic_warmup=0       trainer.logger=['console','tensorboard']       trainer.project_name='verl_grpo_example_geo3k'       trainer.experiment_name='qwen2_5_vl_32b_function_rm'       trainer.n_gpus_per_node=8      trainer.nnodes=1       trainer.save_freq=-1       trainer.test_freq=5       trainer.total_epochs=5

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.

FSDP2 memory snapshot: cpu offloading works and peak memory is slightly lower than FSDP1

[CLAassistant]

All committers have signed the CLA.

[gemini-code-assist]

Code Review

This pull request addresses an Out-Of-Memory issue in torch<2.7 by backporting the state_dict API from torch 2.7. The approach of vendoring the newer code and using it conditionally based on the PyTorch version is sound. However, the changes include debugging code in verl/workers/fsdp_workers.py that must be removed before merging, as it contains a hard breakpoint that will halt execution.

[gemini-code-assist]

This line appears to be for debugging memory usage and uses an internal PyTorch API (_record_memory_history). This should be removed before merging to avoid unintended side effects in production.

[gemini-code-assist]

This try...except block and the subsequent lines for memory snapshotting and breakpointing appear to be debugging code. The torch.distributed.breakpoint() will halt execution, and memory dumps are not intended for production code. This debugging logic should be removed.

            (
                self.actor_module_fsdp,
                self.actor_optimizer,
                self.actor_lr_scheduler,
                self.actor_model_config,
            ) = self._build_model_optimizer(
                model_path=local_path,
                fsdp_config=fsdp_config,
                optim_config=optim_config,
                override_model_config=override_model_config,
                use_remove_padding=use_remove_padding,
                use_fused_kernels=use_fused_kernels,
                enable_gradient_checkpointing=self.config.model.get("enable_gradient_checkpointing", False),
                trust_remote_code=self.config.model.get("trust_remote_code", False),
                use_liger=self.config.model.get("use_liger", False),
                role="actor",
                enable_activation_offload=self.config.model.get("enable_activation_offload", False),
            )

[eric-haibin-lin]

is this code for debugging only?

[weifengpy]

yeap. I will update the PR to undo all the changes to fsdp_workder.py

[weifengpy]

done. the PR is now ready for review

[weifengpy]

fsdp_utils.py has the true change @eric-haibin-lin

third_party/torch are just copy-paste from torch==2.7.0 (state_dict.py and _state_dict_utils.py)

[weifengpy]

@eric-haibin-lin should we kick off CI?

[kisseternity]

Still have oom issue using fsdp2_load_full_state_dict after upgrading to verl-0.5.0 release, using torch 2.6.0. May still need to upgrade torch version?

[weifengpy]

Still have oom issue using fsdp2_load_full_state_dict after upgrading to verl-0.5.0 release, using torch 2.6.0. May still need to upgrade torch version?

probably to install verl from source code directly to include the fix