xlm.commands.lightning_eval

instantiate_model(cfg, datamodule, tokenizer)

Instantiate and load a model for evaluation.

Supports two modes

1. Load a model from full training checkpoint using lightning_module.load_from_checkpoint(cfg.eval.checkpoint_path)
2. Load a model from model only checkpoint using lightning_module.model.load_state_dict(torch.load(cfg.eval.model_only_checkpoint_path))

Parameters:

Name	Type	Description	Default
cfg	DictConfig	Hydra config	required
datamodule	Any	Datamodule	required
tokenizer	Any	Tokenizer	required

Returns:

Type	Description
Harness	Tuple of (lightning_module, ckpt_path) where ckpt_path is the full checkpoint path
Optional[str]	(or None if using model-only checkpoint)