feat(rollout-skip): add dump_steps list parameter to specify particular dump steps

[zyang6]

What does this PR do?

• Add a new parameter:rollout.skip.dump_steps(a list of integers, 1-based step indices).
• Keep the existing parameter: rollout.skip.max_dump_step (dumps/loads for steps in the range ([1, max_dump_step]).).

follow up #5556

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
actor_rollout_ref.rollout.skip.enable=true \ # Enable the feature
actor_rollout_ref.rollout.skip.dump_dir=/path/skip_rollout/ \ # Path for saving data
actor_rollout_ref.rollout.skip.dump_steps=[1,2,5] \ # dump/load only at steps 1, 2, and 5.

Design & Code Changes

• If dump_steps is set and non-empty:

Enable dump/load only when curr_train_step is in dump_steps.
dump_steps overrides the max_dump_step window.

• If dump_steps is null or an empty list ([]):

Fall back to the existing window logic: dump/load while curr_train_step <= max_dump_step.

• Compatibility

Existing configs that rely on max_dump_step continue to work unchanged.
dump_steps is optional; leaving it null (default) preserves current behavior.

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 introduces the dump_steps configuration parameter, allowing users to specify an explicit list of training steps for rollout data dumping and loading, which takes precedence over the max_dump_step window. The changes include updates to documentation, configuration YAML files, and the RolloutSkip utility, along with new unit tests and a typo fix in a warning message. Feedback was provided regarding the robustness of dump_steps parsing, specifically suggesting the use of json.loads to handle string-encoded lists passed via the command line to prevent potential runtime errors.

[gemini-code-assist]

The current implementation for parsing dump_steps is not robust enough to handle string inputs, which can occur when the parameter is passed via the command line (e.g., dump_steps='[1,2,5]'). The documentation in docs/advance/rollout_skip.rst even provides an example with a string value. The current code [int(x) for x in raw_steps] will iterate over the characters of the string and raise a ValueError, causing a crash.

The code should be updated to handle string-encoded lists, for example by using json.loads.

Suggested change

	if raw_steps is not None:
	step_list = [int(x) for x in raw_steps]
	if len(step_list) > 0:
	self._dump_step_set = frozenset(step_list)
	if raw_steps is not None:
	if isinstance(raw_steps, str):
	try:
	raw_steps = json.loads(raw_steps)
	except json.JSONDecodeError:
	print(f"{self.print_mark}\033[31mWarning: Could not parse 'dump_steps' from string: '{raw_steps}'. It will be ignored.\033[0m")
	raw_steps = None
	if raw_steps:
	try:
	step_list = [int(x) for x in raw_steps]
	if step_list:
	self._dump_step_set = frozenset(step_list)
	except (ValueError, TypeError):
	print(f"{self.print_mark}\033[31mWarning: 'dump_steps' contains non-integer values: '{raw_steps}'. It will be ignored.\033[0m")

[Moocharr]

This line is code; no annotation symbols(#) are needed. Simply state in the comments that it's optional.

[zyang6]

ok