validate_model

Checkpoint validation through the native AutoModel scoring recipe.

Functions

validate_model

Validate one realized checkpoint and preserve the legacy metrics return shape.
validate_model(args, model=None, tokenizer=None, target_hidden_states_per_batch=None, return_hidden_states=False, calculate_full_score_ablations=False, val_dataloader=None, *, hydra_cfg=None)

Validate one realized checkpoint and preserve the legacy metrics return shape.

Direct preloaded-model and external-hidden-state validation belonged to the deleted stitched runtime. Callers must provide a checkpoint; its stage-local automodel.parallel mesh is compiled into the native AutoModel configuration at launch time.

Parameters:

  • args (DictConfig)

  • return_hidden_states (bool)

  • calculate_full_score_ablations (bool)

  • hydra_cfg (DictConfig | None)