This repository provides a high-speed download script for Facebook's LLaMA 65B parameter model, targeting researchers and individual hackers interested in experimenting with large language models. It offers a significantly faster alternative to torrent downloads, enabling quicker access to the model weights.

How It Works

The project utilizes curl to download model weights directly from a mirrored source, bypassing the slower peer-to-peer torrent method. This approach leverages a direct HTTP download, achieving speeds up to 40MB/s, which is approximately 3.6x faster than the original torrent. The script is designed for simplicity, allowing users to download all model variants (7B, 13B, 30B, 65B) with a single command.

Quick Start & Requirements

• Install/Run: Execute sh llama.sh after downloading the script.
• Prerequisites: Linux or macOS with bash and wget (macOS requires brew install bash wget).
• Disk Space: Requires 219GB for all model weights.
• Docs: GitHub Repository

Highlighted Details

• Downloads all LLaMA model weights (7B, 13B, 30B, 65B).
• Achieves download speeds of approximately 40MB/s.
• Script is version-locked to a specific commit for safety.
• Includes notes on running LLaMA-65B on a single A100 80GB with 8-bit quantization.

Maintenance & Community

• The original download link was shut down by Facebook; the script was updated to point to a mirrored source.
• Author: Shawn Presser (@theshawwn).

Licensing & Compatibility

• LLaMA model weights are released under a "non-commercial bespoke license."
• Use for commercial purposes may be restricted and could lead to legal action from Facebook.

Limitations & Caveats

• The script downloads weights from an unofficial mirror after the original link was disabled.
• Users are advised to exercise caution when running scripts from the internet.
• The LLaMA model itself has a non-commercial license, limiting its use in commercial products.