HuggingFace Accelerated Downloads
Use internal acceleration nodes for fast HuggingFace model downloads
Overview
SynpixCloud GPU instances include built-in HuggingFace acceleration nodes that provide internal high-speed transfers, improving download speeds by 100x or more. Large models that would take hours to download can now be completed in just minutes.
Quick Setup
Run the following command in your GPU instance terminal to enable accelerated downloads:
export HF_ENDPOINT=http://192.168.50.202:18090After configuration, use huggingface-cli to download models:
huggingface-cli download deepseek-ai/DeepSeek-R1Permanent Configuration
Add the environment variable to ~/.bashrc for persistent configuration:
echo 'export HF_ENDPOINT=http://192.168.50.202:18090' >> ~/.bashrc
source ~/.bashrcUsing in Python Code
Set the environment variable in your Python code:
import os
os.environ['HF_ENDPOINT'] = 'http://192.168.50.202:18090'
from transformers import AutoModel, AutoTokenizer
# Downloads will now automatically use the acceleration node
model = AutoModel.from_pretrained("bert-base-uncased")
tokenizer = AutoTokenizer.from_pretrained("bert-base-uncased")Pre-cached Model Repository
We have pre-cached some popular models for even faster downloads. Use this configuration:
export HF_ENDPOINT="http://guest:guest@223.109.239.18:20281/repository/hf/"
export HF_HUB_ENABLE_HF_TRANSFER=0
export HF_HUB_ENABLE_HF_XET=1
export HF_HUB_DOWNLOAD_TIMEOUT=120
export HF_HUB_ETAG_TIMEOUT=1800Available Cached Models
Browse all cached models at the Model Repository.
Some cached models include:
| Model | Description |
|---|---|
| Qwen/Qwen3-235B-A22B | Qwen3 large language model |
| meta-llama/Llama-4-Scout-17B-16E | Meta Llama 4 |
| black-forest-labs/FLUX.1-dev | FLUX.1 image generation |
| google/gemma-3-27b-it | Google Gemma 3 |
| deepseek-ai/DeepSeek-R1 | DeepSeek R1 reasoning model |
Speed Comparison
| Configuration | Download Speed | Recommendation |
|---|---|---|
| HuggingFace Official | ~125 KB/s | Not recommended |
| hf-mirror.com Public Mirror | ~3 MB/s | Alternative |
| Internal Acceleration Node | 50+ MB/s | Recommended |
The internal acceleration node is only available within SynpixCloud GPU instances. For external access, use the public address: http://223.109.239.18:18090
Troubleshooting
Download Failed?
- Verify the environment variable is set correctly:
echo $HF_ENDPOINT - Check network connectivity:
ping 192.168.50.202 - Try downloading again
How to Verify Acceleration is Working?
Download a small model to test:
time huggingface-cli download bert-base-uncased config.jsonIf configured correctly, the download should complete in seconds.
Support
Need help? Contact us at support@synpixcloud.com