Implementation:Vllm project Vllm Snapshot Download LoRA

Knowledge Sources	vLLM vLLM Docs HuggingFace Hub Docs
Domains	LLM Serving, Model Adaptation, Weight Management
Last Updated	2026-02-08 13:00 GMT

Overview

Concrete tool for downloading LoRA adapter weights from HuggingFace Hub to local filesystem provided by the huggingface_hub library.

Description

The snapshot_download function from huggingface_hub downloads all files from a specified HuggingFace repository to a local cache directory. In the multi-LoRA serving workflow, it is used to fetch LoRA adapter weight files (such as adapter_config.json and adapter_model.safetensors) so they can be referenced by LoRARequest objects when submitting inference requests. The function returns the local filesystem path to the downloaded snapshot, which is passed directly as the lora_path argument when constructing LoRA requests.

Usage

Use this function to download LoRA adapter repositories from HuggingFace Hub before starting inference. This is typically called once during setup, before constructing LoRARequest objects. The allow_patterns parameter can be used to filter which files are downloaded, reducing bandwidth and storage when only specific files are needed.

Code Reference

Source Location

• Repository: vllm
• File: examples/offline_inference/multilora_inference.py (usage), huggingface_hub library (source)

Signature

def snapshot_download(
    repo_id: str,
    repo_type: str | None = None,
    revision: str | None = None,
    cache_dir: str | Path | None = None,
    local_dir: str | Path | None = None,
    allow_patterns: list[str] | str | None = None,
    ignore_patterns: list[str] | str | None = None,
    token: str | bool | None = None,
) -> str

Import

from huggingface_hub import snapshot_download

I/O Contract

Inputs

Name	Type	Required	Description
repo_id	str	Yes	The HuggingFace repository ID containing LoRA adapter weights (e.g., "jeeejeee/llama32-3b-text2sql-spider")
repo_type	str or None	No	Type of repository ("model", "dataset", "space"). Defaults to "model".
revision	str or None	No	Git revision (branch, tag, or commit hash) to download. Defaults to main branch.
cache_dir	str, Path, or None	No	Path to the cache directory. Defaults to HuggingFace default cache.
local_dir	str, Path, or None	No	If set, files are downloaded to this directory instead of the cache.
allow_patterns	list[str], str, or None	No	Glob patterns to filter which files to download (e.g., ["*.safetensors", "*.json"]).
ignore_patterns	list[str], str, or None	No	Glob patterns to exclude files from downloading.
token	str, bool, or None	No	Authentication token for private repositories. True uses the stored token.

Outputs

Name	Type	Description
snapshot_path	str	Local filesystem path to the downloaded snapshot directory containing all adapter files

Usage Examples

Download LoRA Adapter for Multi-LoRA Serving

from huggingface_hub import snapshot_download

# Download the LoRA adapter weights to local cache
lora_path = snapshot_download(repo_id="jeeejeee/llama32-3b-text2sql-spider")

# lora_path is now a local directory path like:
# "/home/user/.cache/huggingface/hub/models--jeeejeee--llama32-3b-text2sql-spider/snapshots/abc123"
# This path is passed to LoRARequest when creating inference requests

Download with Filtered Patterns

from huggingface_hub import snapshot_download

# Download only the essential adapter files
lora_path = snapshot_download(
    repo_id="jeeejeee/llama32-3b-text2sql-spider",
    allow_patterns=["*.safetensors", "*.json"],
)

Related Pages

Implements Principle

• Principle:Vllm_project_Vllm_LoRA_Adapter_Preparation