nvalchemi.training.BaseSpec

pydantic model nvalchemi.training.BaseSpec

Base class for JSON-serializable, no-pickle hyperparameter specs.

Concrete spec classes are built dynamically by create_model_spec() via pydantic.create_model(); each carries one field per __init__ kwarg of its target class plus the two metadata fields defined here.

Notes

revalidate_instances="never" is deliberate: specs are immutable records of past state; revalidating on access would reject any already-typed field values (e.g. rehydrated torch.Tensor objects) that were stored through a BeforeValidator.

field cls_path: str [Required]

Dotted import path of the target callable.

Constraints:

• func = <function _ensure_importable at 0xeb0b2465ec00>

field timestamp: str [Required]

ISO-8601 UTC timestamp of spec creation.