[recipe, megatron, fsdp] fix: checkpoint-engine fix trainer param offload in fully-async mode

[zpltys]

… fully-async mode

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).)

[gemini-code-assist]

Code Review

This pull request introduces support for parameter offloading with the checkpoint-engine in fully-asynchronous mode. It correctly adds logic to load the model to the GPU before synchronization and offload it to the CPU afterward when parameter offloading is enabled. The PR also adds timing metrics for these load/offload operations. My review identifies an opportunity to improve the accuracy of these new performance measurements to ensure they correctly reflect the duration of the operations.

[gemini-code-assist]

The current implementation for measuring load_duration is imprecise. When self._is_actor and self._is_offload_param is false, it measures the time for the conditional check, resulting in a small non-zero value instead of zero. When true, it includes the overhead of the check. To measure the duration accurately, the timing logic should be contained entirely within the conditional block, and load_duration should be initialized to 0.0.

Suggested change

	load_start_time = time.time()
	if self._is_actor and self._is_offload_param:
	load_fsdp_model_to_gpu(self.actor_module_fsdp)
	load_duration = time.time() - load_start_time
	load_duration = 0.0
	if self._is_actor and self._is_offload_param:
	load_start_time = time.time()
	load_fsdp_model_to_gpu(self.actor_module_fsdp)
	load_duration = time.time() - load_start_time

[gemini-code-assist]

Similar to load_duration, the measurement for offload_duration is imprecise. It should be calculated only when the offload operation is performed to avoid including the conditional check overhead and to ensure it is zero when no offload occurs.

Suggested change

	offload_start_time = time.time()
	if self._is_actor and self._is_offload_param:
	offload_fsdp_model_to_cpu(self.actor_module_fsdp)
	offload_duration = time.time() - offload_start_time
	offload_duration = 0.0
	if self._is_actor and self._is_offload_param:
	offload_start_time = time.time()
	offload_fsdp_model_to_cpu(self.actor_module_fsdp)
	offload_duration = time.time() - offload_start_time

[gemini-code-assist]

The measurement for load_duration is imprecise. It includes the overhead of the conditional check and will not be zero if the condition is false. For accurate profiling, the timing should only wrap the load_megatron_model_to_gpu call, and load_duration should be initialized to 0.0.

Suggested change

	load_start_time = time.time()
	if self._is_actor and self._is_offload_param:
	load_megatron_model_to_gpu(self.actor_module)
	load_duration = time.time() - load_start_time
	load_duration = 0.0
	if self._is_actor and self._is_offload_param:
	load_start_time = time.time()
	load_megatron_model_to_gpu(self.actor_module)
	load_duration = time.time() - load_start_time

[gemini-code-assist]

The measurement for offload_duration is imprecise for the same reasons as load_duration. To ensure correctness, the timing logic should be moved inside the conditional block.

Suggested change

	offload_start_time = time.time()
	if self._is_actor and self._is_offload_param:
	offload_megatron_model_to_cpu(self.actor_module)
	offload_duration = time.time() - offload_start_time
	offload_duration = 0.0
	if self._is_actor and self._is_offload_param:
	offload_start_time = time.time()
	offload_megatron_model_to_cpu(self.actor_module)
	offload_duration = time.time() - offload_start_time

[ZSL98]

Hi, it is impressive to see your contribution on fully_async and the integration of checkpoint engine. Could you please leave some comments on #4654. We would like to contribute more to fully_async mode but do not know if it is a good or even correct starting point. @zpltys Also cc @ArronHZG