[CI] feat: add mypy to pre-commit

[frrad]

What does this PR do?

Adds mypy typechecking to pre-commit. Only one module is whitelisted and it is configured to be very permissive.

Test

this is a CI / testing change

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

• Once your PR is ready for CI, send a message in the ci-request channel in the verl Slack workspace.

[gemini-code-assist]

Code Review

This pull request introduces mypy for static type checking, initially enabling it for the verl.trainer.ppo.core_algos module. The changes include adding the mypy pre-commit hook, configuring it in pyproject.toml, and fixing type errors in the target module. My review focuses on a critical issue with the pre-commit configuration that will break the CI, and a high-severity issue regarding the use of typing.Any which undermines the goal of adding static type analysis. Addressing these will ensure the CI works as expected and improve the long-term maintainability of the typed code.

[zhaochenyang20]

@frrad CI is good. I can help to merge if haibin approve.

[frrad]

copying comment here from Lark

@eric-haibin-lin

i wonder why the type is corrected to "Any". in fact we pass DictConfig or AlgoConfig.
Does the mypy make the correction itself, or the change has to be done manually?

@frrad

I made that change manually. the AlgoConfig type is incorrect - none of the fields referenced on L939

verl/verl/trainer/ppo/core_algos.py

Line 939 in 0fdbe54

kl_cov_ratio = config.policy_loss.kl_cov_ratio if config.policy_loss.kl_cov_ratio is not None else 0.0002

actually exist on the datatype

verl/verl/trainer/config/algorithm.py

Lines 105 to 114 in f252da3

	gamma: float = 1.0
	lam: float = 1.0
	adv_estimator: str = "gae"
	norm_adv_by_std_in_grpo: bool = True
	use_kl_in_reward: bool = False
	kl_penalty: str = "kl"
	kl_ctrl: KLControlConfig = field(default_factory=KLControlConfig)
	use_pf_ppo: bool = False
	pf_ppo: Optional[PFPPOConfig] = None
	filter_groups: Optional[FilterGroupsConfig] = None

I was not sure what the real type was so I changed it to Any. I can change it to DictConfig instead if you prefer.

[eric-haibin-lin]

Let's change it to Optional AlgoConfig|DictConfig]

[frrad]

please take another look

[frrad]

In cc2972c I have added an additional path to the mypy whitelist which make it necessary to change AlgoConfig to add the fields referenced from core_algos. Happy to revert that commit to keep the scope of this first PR as small as possible if you want.

edit: originally I had linked the wrong commit in this message

[zhaochenyang20]

I think this PR is almost done. Let me help on the merge! @frrad

[eric-haibin-lin]

There are two issues. One is that verl already has actor_rollout_ref.actor.policy_loss and actor_rollout_ref.actor.clip_ratio_high/low, duplicating them without changing existing example script would be confusing.
Also restructuring algo config & actor config is not a high priority for now and moving them around and breaking others' recipes may lead to bad experience. If we do want to make breaking changes, we usually make announcement for deprecations, keep it for one version before breaking it..

I do not think it's necessary to couple mypy setup with breaking changes ..

[frrad]

I have reverted that commit.

[frrad]

this is probably a better fix. please take another look.

[zhaochenyang20]

@frrad Let me merge it after the CI