MicroGpt-torch

A minimal GPT implementation in 83 lines of Python + PyTorch. Rewrite of microgpt.py by Andrej Karpathy, replacing scalar-level autograd with PyTorch tensor ops + CUDA.

Trains on ~32K English names and generates new ones.

Requirements

• Python >= 3.10
• PyTorch >= 2.4 (for F.rms_norm, F.scaled_dot_product_attention)

Usage

python microgpt-torch.py

The script will:

1. Download the dataset (if not present)
2. Train for 40 epochs (auto-detects GPU)
3. Generate 20 new "hallucinated" names

Files

• microgpt-torch.py — compact version (83 lines)
• microgpt-torch-comments.py — same code with bilingual comments (EN/CN)