A pure CUDA implementation of the LLaMA model for high-performance inference and educational purposes. Supports LLaMA 1, 2, and 3 architectures.

This repository demonstrates how to run LLaMA inference using CUDA C++, making it ideal for learning GPU acceleration techniques and understanding transformer internals with minimal dependencies.

• Pure CUDA Implementation – Direct CUDA kernels for maximum performance without heavy ML frameworks
• Optimized Matrix Operations – Custom CUDA kernels for matrix multiplication and attention mechanisms
• Educational – Clean, readable CUDA code with inline documentation for learning GPU programming

make
./llama stories15M.bin

The examples use small models trained by Andrej Karpathy for demonstration.

Requires NVIDIA CUDA Toolkit (11.0 or later):

make

Or using CMake:

mkdir build && cd build
cmake ..
make

• Implement FP16 (float16) version for better memory efficiency and performance
• Add Flash Attention for faster attention computation

If you're interested in LLaMA implementations in other languages:

• llama.go – Pure Go implementation
• llama.np – NumPy-based implementation

Inspired by llama2.c, llama3.cuda and the broader LLaMA community. This project aims to provide a GPU-accelerated alternative for educational purposes.

MIT