Implementation:Huggingface Optimum Download Model From HF

Overview

Downloads model artifacts from the Hugging Face Hub to the local filesystem, returning the path to the cached model snapshot. Supports selective downloading to skip weight files when only the model configuration is needed.

Source

Property	Value
File	`optimum/fx/parallelization/utils.py`
Lines	L359-416
Module	`optimum.fx.parallelization.utils`

Signature

def download_model_from_hf(
    model_name_or_path: str,
    cache_dir: Optional[str],
    revision: Optional[str] = None,
    local_files_only: bool = False,
    skip_download_weights: bool = False,
) -> str:

Import

from optimum.fx.parallelization.utils import download_model_from_hf

Input / Output

Direction	Name	Type	Description
Input	model_name_or_path	`str`	Hugging Face Hub model identifier (e.g., meta-llama/Llama-2-7b-hf) or a local filesystem path.
Input	cache_dir	`Optional[str]`	Local directory for caching downloaded files.
Input	revision	`Optional[str]`	Git revision (branch, tag, or commit hash) to download. Defaults to None (latest).
Input	local_files_only	`bool`	If True, only look for files in the local cache. Defaults to False.
Input	skip_download_weights	`bool`	If True, skip downloading safetensors weight files. Defaults to False.
Output	(return)	`str`	Local filesystem path to the downloaded model snapshot directory.

Behavior

The function performs the following steps:

Check for local path: If model_name_or_path is a local directory, return it directly without downloading.
Determine files to ignore: If skip_download_weights is True, build a list of patterns to ignore safetensors weight files during download.
Call huggingface_hub.snapshot_download: Download the model snapshot from the Hub with the specified parameters (cache directory, revision, ignore patterns).
Return snapshot path: Return the local path where the snapshot was cached.

Example Usage

from optimum.fx.parallelization.utils import download_model_from_hf

# Download full model (weights + config)
model_path = download_model_from_hf(
    model_name_or_path="meta-llama/Llama-2-7b-hf",
    cache_dir="/tmp/model_cache",
)

# Download config only (skip weights for meta-device init)
config_path = download_model_from_hf(
    model_name_or_path="meta-llama/Llama-2-7b-hf",
    cache_dir="/tmp/model_cache",
    skip_download_weights=True,
)

Dependencies

huggingface_hub.snapshot_download - Core download functionality from the Hugging Face Hub library.
os.path.isdir - Used to detect local model paths.

Page Connections

Double-click a node to navigate. Hold to expand connections.

Principle

Implementation

Heuristic

Environment