Llama.ttf: A Font That Runs a Large Language Model

If you thought fonts were just for displaying text, think again. llama.ttf is a font file that doubles as a large language model (LLM) and its inference engine. This innovative hack leverages the WebAssembly (Wasm) capabilities of the HarfBuzz font shaping engine to run an LLM directly from a font file.

What Changed?

Traditionally, fonts are used for rendering text in applications like web browsers and word processors. However, llama.ttf takes this concept to a new level by embedding a complete LLM and its inference engine within the font file itself. This means you can use the font to generate text based on prompts, all while running locally and without any additional software installations.

Why Does It Matter?

For developers and enthusiasts, llama.ttf opens up a novel way to integrate AI capabilities into existing applications. Instead of waiting for vendors to add "Copilot"-like features, you can use this font to get text generation in any Wasm-enabled HarfBuzz-based application. This includes your favorite text editor, email client, or even image editing software.

How Does It Work?

• HarfBuzz and Wasm: HarfBuzz is a widely used font shaping engine that supports WebAssembly (Wasm). The Wasm shaper allows arbitrary code to be executed during the text shaping process.
• LLM Embedding: llama.ttf embeds an entire LLM inference engine, specifically the Llama model, along with its trained parameters. This is achieved by treating certain text sequences as "magic ligatures" that trigger the LLM initialization and text generation.
• Local Execution: Since everything runs locally, there's no need for network connectivity or external APIs. The font file itself contains all the necessary components to perform inference.

Technical Details

• Font Size: llama.ttf is a 60 MB download, which includes the 15M parameter TinyStories-based model.
• Requirements: To use llama.ttf, you need an application that uses HarfBuzz with Wasm support. This can be achieved by:
  • Building HarfBuzz with -Dwasm=enabled.
  • Building wasm-micro-runtime with -DWAMR_BUILD_REF_TYPES=1.
  • Adding the resulting shared libraries (libharfbuzz.so.0.60811.0 and libiwasm.so) to the LD_PRELOAD environment variable.
• No Recompilation Needed: Existing applications like gedit or GIMP can use llama.ttf without recompilation, making it a seamless integration.

In Action

To see llama.ttf in action, check out this YouTube video by Søren Fuglede Jørgensen. Skip to 6:09 for a demonstration of the font generating text based on user prompts.

Usage

1. Download llama.ttf: Get the 60 MB font file from the project's GitHub page.
2. Build HarfBuzz and wasm-micro-runtime:
   • Clone the HarfBuzz repository: git clone https://github.com/harfbuzz/harfbuzz.git
   • Build with Wasm support: cmake -Dwasm=enabled . && make
   • Clone the wasm-micro-runtime repository: git clone https://github.com/bytecodealliance/wasm-micro-runtime.git
   • Build with reference types: cmake -DWAMR_BUILD_REF_TYPES=1 . && make
3. Set Up Environment:
   • Add the shared libraries to LD_PRELOAD: export LD_PRELOAD=/path/to/libharfbuzz.so.0.60811.0:/path/to/libiwasm.so
4. Use the Font: Open a HarfBuzz-based application and select llama.ttf as your font.

More Fun with Fonts

If this project piques your interest, consider exploring other creative uses of fonts and Wasm. The possibilities are