Back to articles
AI & Cloud

Self-Host Private Hugging Face Mirror on VPS Cloud Vietnam: High-Speed AI Solution

Detailed guide to building a private Hugging Face mirror on a VPS Cloud in Vietnam for fast, stable AI model downloads without international connection drops. Effective solution for...

DPTCloud Engineering4 min read
Self-Host Private Hugging Face Mirror on VPS Cloud Vietnam: High-Speed AI Solution

Why Use a Private Hugging Face Mirror on VPS Cloud Vietnam?

Downloading AI models from Hugging Face often suffers from slow speeds or disconnections due to reliance on international network connections. For businesses and developers in Vietnam, downloading large models like Llama, Mistral, or Stable Diffusion can take hours or fail entirely due to bandwidth limits, network interruptions, or internet censorship. The effective solution is to self-host a private Hugging Face mirror on a VPS Cloud in Vietnam—offering stable international connectivity, high bandwidth, and optimal geographic positioning.

This private mirror acts as an internal cache: the first download of a model stores it locally on your VPS, and subsequent downloads retrieve it from the local source—significantly faster and independent of international network conditions.

Requirements and Environment Setup

To deploy a private Hugging Face mirror, you will need:

  • 1 VPS Cloud in Vietnam with at least 32GB RAM and 500GB SSD (optimized for storing large models)
  • Operating system: Ubuntu 22.04 LTS
  • Docker and Docker Compose installed
  • SSH access to the VPS

Configuring the Private Mirror

We’ll use huggingface-mirror—an automated tool to download and sync models from Hugging Face to a local server. Follow these steps:

# 1. Install Docker and Docker Compose
sudo apt update
sudo apt install -y docker.io docker-compose

# 2. Create storage directory
mkdir -p ~/huggingface-mirror
cd ~/huggingface-mirror

# 3. Create docker-compose.yml file
cat < docker-compose.yml
version: '3.8'

services:
  mirror:
    image: ghcr.io/huggingface/huggingface-mirror:latest
    container_name: huggingface-mirror
    ports:
      - "8080:8080"
    volumes:
      - ./cache:/cache
    environment:
      - HF_MIRROR_URL=https://huggingface.co
      - HF_CACHE_DIR=/cache
    restart: unless-stopped
EOF

# 4. Start the service
docker-compose up -d

# 5. Check status
docker-compose ps

After starting, the private mirror will run on port 8080. You can access http://[IP-VPS]:8080 to verify.

Downloading AI Models via the Private Mirror

To download models from your private mirror, configure your environment to use your VPS URL instead of the official Hugging Face endpoint. For example, in Python:

from huggingface_hub import snapshot_download
import os

# Configure private mirror URL
os.environ["HF_ENDPOINT"] = "http://[IP-VPS]:8080"
os.environ["HF_MIRROR"] = "http://[IP-VPS]:8080"

# Download model
model_name = "meta-llama/Llama-2-7b-chat-hf"
snapshot_download(model_name, local_dir="./models")

Replace [IP-VPS] with your VPS’s public IP address. The first download will take time, but subsequent downloads will be much faster.

Verification and Troubleshooting

Troubleshooting:

  • If downloads fail: Check if port 8080 is open on the VPS (firewall, security group)
  • If models are missing: Verify the /cache directory on the VPS
  • If latency is high: Ensure the VPS has stable international connectivity

Verification:

  • Run docker logs huggingface-mirror to check download logs
  • Visit http://[IP-VPS]:8080 to confirm models appear in the list
  • Test downloading a small model like bert-base-uncased for validation

Performance Optimization

To enhance performance:

  • Run multiple mirror containers if downloading multiple models simultaneously
  • Host the VPS in Vietnam to minimize network latency
  • Use NVMe SSD instead of HDD for faster read/write speeds
  • Set up a cron job to automatically sync popular models weekly

References

DPTCloud Tier III Infrastructure

High-RAM Cloud VPS Plans Optimized for AI, Databases & Cache

Enterprise RAID 10 NVMe SSDs, CMC & FPT District 9 uplinks with Anycast Anti-DDoS and 99.99% SLA.

View 22 VPS Plans
ECO · vn-02Recommended

VPS-4E

297.000 ₫ / mo

10% VAT included

  • CPU: 4 vCPU
  • RAM: 8 GB
  • SSD: 50 GB NVMe
ECO · vn-02

VPS-5E

614.900 ₫ / mo

10% VAT included

  • CPU: 8 vCPU
  • RAM: 16 GB
  • SSD: 100 GB NVMe
ADVANCED · vn-03Recommended

VPS-4T

548.900 ₫ / mo

10% VAT included

  • CPU: 4 vCPU
  • RAM: 8 GB
  • SSD: 50 GB NVMe

Related Articles

View all
ZaloZalo SupportDiscord0384 753 478