ToyLLM: Learning LLM from Scratch

A hands-on educational project for understanding and implementing Large Language Models (LLMs) from scratch. This project provides implementations of GPT-2 and related techniques, making it an excellent resource for learning about transformer architectures and modern language models.

A clean, educational implementation of GPT-2 with type hints, supporting both training and inference.

An implementation of speculative sampling for faster inference, featuring configurable draft models and performance benchmarking.

A memory-efficient GPT-2 implementation with KV cache optimization for handling longer sequences.

• Python 3.14
• Git and Git LFS (for model files)
• UV (recommended package manager)

1. Clone the repository:

   git clone https://github.com/ai-glimpse/toyllm.git
   cd toyllm

2. Set up the environment:

   # Create the environment and install the project
   uv sync

3. Download model files:

   # Install Git LFS if not already installed
   git lfs install
   
   # Download model files
   git clone https://huggingface.co/MathewShen/toyllm-gpt2 models

   Alternatively, you can manually download the model files from Hugging Face and place them in the models directory.

uv run python -m toyllm.cli.run_gpt2 --help  # View available options
uv run python -m toyllm.cli.run_gpt2         # Run with default settings

uv run python -m toyllm.cli.run_gpt2_kv --help  # View available options
uv run python -m toyllm.cli.run_gpt2_kv         # Run with default settings

uv run python -m toyllm.cli.run_speculative_sampling --help  # View available options
uv run python -m toyllm.cli.run_speculative_sampling         # Run with default settings

uv run python -m toyllm.cli.benchmark.bench_gpt2kv --help  # View available options
uv run python -m toyllm.cli.benchmark.bench_gpt2kv         # Run benchmarks

src/toyllm/
├── cli/                    # Command-line interface modules
├── core/                   # Shared generation primitives
├── gpt2/                   # GPT-2 specific implementations
├── gpt2kv/                 # KV-cache optimized GPT-2
├── sps/                    # Speculative sampling implementations
└── util/                   # Utility functions

This project is inspired by and builds upon the following excellent resources:

• rasbt/LLMs-from-scratch
• neelnanda-io/TransformerLens

Contributions are welcome! Please feel free to submit a Pull Request.