Llama_IPFS

Load models directly from IPFS for llama-cpp-python.

Features

• 🌐 Direct integration with local IPFS nodes (preferred method)
• 🔄 Automatic fallback to IPFS gateways when local node isn't available
• 🔍 Simple URI format: ipfs://CID for easy model sharing
• ⚡ Zero configuration required - works automatically once installed
• 🧩 Compatible with any version of llama-cpp-python

Installation

pip install llama-ipfs
llama-ipfs activate

Once installed and activated, the llama_ipfs integration will be loaded automatically whenever you use Python.

Usage

• 
• 🤗 Demo Space

After installation, use llama-cpp-python with IPFS model URIs:

from llama_cpp import Llama

# Load a model directly from IPFS
model = Llama.from_pretrained(
    # Equivalent HuggingFace model: "aisuko/gpt2-117M-gguf"
    repo_id = "ipfs://bafybeie7quk74kmqg34nl2ewdwmsrlvvt6heayien364gtu2x6g2qpznhq",
    filename = "ggml-model-Q4_K_M.gguf"
)

IPFS Node Connectivity

The llama_ipfs package prioritizes connectivity in the following order:

1. Local IPFS Node (Recommended): If you have an IPFS daemon running locally (ipfs daemon), the package will automatically detect and use it. This method:

   • Is much faster for repeated downloads
   • More reliably loads complex model directories
   • Contributes to the IPFS network by providing content to others

2. IPFS Gateway (Fallback): If a local node isn't available, the package will fall back to public gateways. This method:

   • Works without installing IPFS
   • May be less reliable for complex model directories
   • Downloads can be interrupted more easily

Command Line Interface

# Activate the auto-loading
llama-ipfs activate

# Check if the integration is active
llama-ipfs status

# Test the integration
llama-ipfs test

# Deactivate the integration
llama-ipfs deactivate

Dependencies

• Python 3.8+
• llama-cpp-python

License

MIT License