Tech Stack: CUDA C++ · PyTorch (for reference) · Jupyter Notebook

miniGPT is a minimal, CUDA-accelerated implementation of a transformer-based language model inspired by GPT architectures. This project demonstrates the core components of transformer inference, including attention, feedforward layers, and positional encoding, all implemented from scratch in CUDA C++.

Project Requirements

• CUDA Toolkit (version 11.0 or higher recommended)
• NVIDIA GPU with compute capability 6.0+
• g++ (for compiling C++/CUDA code)
• Python 3.x (for running notebooks and reference scripts)
• PyTorch (for reference and comparison, optional)
• Jupyter Notebook (for exploration and demonstration)

How to Build and Run

1. Clone the repository:

   git clone https://github.com/yourusername/cuda-miniGPT.git
   cd cuda-miniGPT

2. Compile the CUDA source code:

   bash scripts/compile.sh

3. Compile the CUDA source code:

   ./build/minigptinference

4. (Optional) Explore the Jupyter Notebook:

   jupyter notebook miniGPT.ipynb