nanotorch
Project
Description:
nanotorch is a from-scratch reimplementation of core PyTorch internals, hand-built for deep understanding rather than production use. It covers autograd and tensors, RoPE/attention, a pretrained tiny-GPT, KV caching, LoRA and full fine-tuning, and quantization — the full stack of what makes a modern deep learning framework and LLM training loop work.