YA-GPT

Yet another GPT implementation from scratch!

Development

Enter the Nix development shell, then sync the Python environment with uv:

nix develop
uv sync

The shell provides Python 3.13, uv, the C++ runtime required by PyTorch, and the NixOS NVIDIA driver library path and CUDA compiler tools used by Triton.

With direnv installed, allow the project once to enter the shell and load local variables from .env automatically:

direnv allow
uv sync

update for cpu inference write inference engine in c++ (llama.cpp and nanochat inspiration) also create own binary model format like GGUF

ToDO: train SLP with numpy dump matrix to binary format write binary file parser in c++ write matrix mult in c++ write sigmoid in c++ prediction = sigmoid(x * W + b)

read weight matrix from file perform prediction calculation + output it