pimm.save_pretrained

save_pretrained(model_or_checkpoint: Any, save_directory: str | Path, *, cfg: Any | None = None, config_path: str | Path | None = None, training_config: Mapping[str, Any] | None = None, safe_serialization: bool = True, device: str = 'cpu', model_card: str | None = None) → Path

Save model weights to a directory as bare serialized tensors.

Writes model.safetensors (or model.bin when safe_serialization=False) and, optionally, training_config.json (provenance) and README.md (model card). No model config is written, so loading requires the architecture from elsewhere (a fine-tune config for warm-start, or model_config/config_path for from_pretrained). model_or_checkpoint may be an nn.Module, checkpoint mapping, or path.