`xlm.utils.hf_hub`

Utilities for Hugging Face Hub integration.

`repo_id_from_hf_path(path)`

Extract repo_id from HF Hub path (URL or org/repo). Returns None if invalid.

Download model weights from Hugging Face Hub.

Tries model.safetensors first, then pytorch_model.bin.

Parameters:

Name	Type	Description	Default
`repo_id`	`str`	Hugging Face repository ID (e.g., "org/model").	required
`revision`	`str`	Git revision (branch, tag, or commit). Defaults to "main".	`'main'`
`token`	`str \| None`	HF token for private repos. Uses HF_HUB_KEY env if None.	`None`

Returns:

Type	Description
`str`	Path to the downloaded weights file.

Raises:

Type	Description
`ValueError`	If neither model.safetensors nor pytorch_model.bin exists.

Load model state dict from a checkpoint file (safetensors or pickle).

For .safetensors uses load_file directly (no metadata validation). For .bin/.pt uses torch.load.

Parameters:

Name	Type	Description	Default
`checkpoint_path`	`str`	Path to model.safetensors or pytorch_model.bin.	required
`map_location`	`str`	Device to load tensors to.	`'cpu'`
`weights_only`	`bool`	If True, use weights_only for pickle (PyTorch >= 1.13).	`True`

Returns:

Type	Description
	State dict for model.load_state_dict().

Load weights from checkpoint into model. Aligns with harness and hub_mixin.

For .safetensors uses safetensors.torch.load_model (handles tensor sharing). For .bin/.pt uses model.load_state_dict(torch.load(...)).

Parameters:

Name	Type	Description	Default
`model`	`Module`	The model to load weights into.	required
`checkpoint_path`	`str`	Path to model.safetensors or pytorch_model.bin.	required
`map_location`	`str`	Device to load tensors to.	`'cpu'`
`strict`	`bool`	Whether to enforce exact key match (safetensors) or pass to load_state_dict.	`True`
`weights_only`	`bool`	If True, use weights_only for pickle (PyTorch >= 1.13).	`True`