o `۷i @sddlmZddlmZmZmZmZmZddlm Z er"ddl m Z e ddGdddeZ ee ddGd d d Z e ddGd d d Zd S)) dataclass) TYPE_CHECKINGAnyDictOptionalProtocol) PublicAPI) Checkpointalpha) stabilityc@s&eZdZdZdddedefddZdS) ValidationFnz4Protocol for a function that validates a checkpoint. checkpointr kwargsreturncKsdSN)selfr rrrX/home/ubuntu/vllm_env/lib/python3.10/site-packages/ray/train/v2/api/validation_config.py__call__szValidationFn.__call__N)__name__ __module__ __qualname____doc__rrrrrrrr sr c@s2eZdZUdZdZeeeefe d<ddZ dS)ValidationTaskConfigaConfiguration for a specific validation task, passed to report(). Args: fn_kwargs: json-serializable keyword arguments to pass to the validation function. Note that we always pass `checkpoint` as the first argument to the validation function. N fn_kwargscCs|jdur i|_dSdSr)r)rrrr __post_init__s  z"ValidationTaskConfig.__post_init__) rrrrrrrstrr__annotations__rrrrrrs  rc @s<eZdZdZ  ddedeedeeee ffddZ dS) ValidationConfigaeConfiguration for validation, passed to the trainer. Args: fn: The validation function to run on checkpoints. This function should accept a checkpoint as the first argument and return a dictionary of metrics. task_config: Default configuration for validation tasks. The fn_kwargs in this config can be overridden by ValidationTaskConfig passed to report(). ray_remote_kwargs: Keyword arguments to pass to `ray.remote()` for the validation task. This can be used to specify resource requirements, number of retries, etc. Nfn task_configray_remote_kwargscCs:||_|dur t|_n||_|duri|_dS||_dSr)rrr r!)rrr r!rrr__init__3s   zValidationConfig.__init__)NN) rrrrr rrrrrr"rrrrr$srN) dataclassesrtypingrrrrrray.util.annotationsr ray.trainr r rrrrrrs