In the world of open-source, I believe nothing is as compelling as a well-contextualized compiler. Here is mine, doing my part to convert the planet to memory-safe Rust, in GPL form.

tiny-gpt

A memory-safe Rust port of tiny-gpt2, with GPT-2 weight download, tokenizer support, inference, and attention visualization.

Original project credit: Stephen Diehl
Source: https://github.com/sdiehl/tiny-gpt2

Features

• Download GPT-2 config and weights from Hugging Face (openai-community/gpt2)
• Download GPT-2 tokenizer assets (encoder.json, vocab.bpe)
• Run greedy text generation from a prompt
• Visualize per-head attention weights as PNG heatmaps
• Keep implementation fully in safe Rust using ndarray
• Core logic is functional and idiomatic, with minimal mutable state

Quick Start

cd /Users/jrule/git/rust/tiny-gpt
cargo run -- generate --prompt "The quick brown fox" --max-tokens 20

cd /Users/jrule/git/rust/tiny-gpt
cargo run -- viz --prompt "The quick brown fox jumps over the lazy dog" --block 0 --head 0 --out attention.png

Attention Visualization Example

Generated with:

cd /Users/jrule/git/rust/tiny-gpt
cargo run -- viz --prompt "The quick brown fox jumped over the lazy dog." --block 0 --head 0 --out attention.png

CLI

• generate: Generate continuation text from a prompt
• tokens: Show token ids and token text pieces
• viz: Save attention heatmap for a specific block/head

Example:

cd /Users/jrule/git/rust/tiny-gpt
cargo run -- tokens --text "hello rust"

License

This repository is licensed under GPL-3.0-or-later. See LICENSE.