fix: call warm_up_process to prevent loguru blocking

[binary-husky]

No description provided.

[gemini-code-assist]

Code Review

This pull request introduces a warm_up_process call in main_verl.py. Feedback recommends moving the import inside the run method for better Ray actor performance and resolving the configuration before the call to handle Hydra interpolations correctly.

[gemini-code-assist]

The warm_up_process function uses configuration values such as experiment_dir and experiment_name to initialize logging. These values often contain Hydra interpolations (e.g., ${hydra:runtime.output_dir}) that must be resolved before they can be used as valid file system paths. Calling warm_up_process before OmegaConf.resolve(config) may lead to incorrect log paths or initialization failures. It is recommended to resolve the configuration first. Additionally, importing warm_up_process locally here maintains consistency with other imports in this method. Note that the existing call to OmegaConf.resolve(config) at line 124 will become redundant and should be removed.

Suggested change

	from loguru import logger
	from omegaconf import OmegaConf
	from verl.utils.fs import copy_to_local
	warm_up_process(config)
	from loguru import logger
	from omegaconf import OmegaConf
	from verl.utils.fs import copy_to_local
	from ajet.backbone.warm_up import warm_up_process
	OmegaConf.resolve(config)
	warm_up_process(config)

[gemini-code-assist]

The import of warm_up_process is currently placed between the comment # Create training and validation datasets. and the actual dataset-related imports, which is confusing. Furthermore, for Ray remote classes like TaskRunner, it is a common practice to keep imports local to the methods where they are used to keep the actor definition lightweight and avoid unnecessary side effects during module import on the head node. This import should be moved inside the run method.

Suggested change

	# Create training and validation datasets.
	from ajet.backbone.warm_up import warm_up_process
	from ajet.task_reader import RouterTaskReader, task_to_standard_dataset
	# Create training and validation datasets.
	from ajet.task_reader import RouterTaskReader, task_to_standard_dataset