Benchmark training speed

--|--------|--------|-------|---------|-------------| | Shakespeare char | ~10M | 6 | 6 | 384 | 1 GPU, 3 min | | GPT-2 small | 124M | 12 | 12 | 768 | 8×A100, 4 days | | GPT-2 medium | 350M | 24 | 16 | 1024 | Modify config | | GPT-2 large | 774M | 36 | 20 | 1280 | Modify config | | GPT-2 XL | 1.5B | 48 | 25 | 1600 | Multi-node |

10. Evaluation and Benchmarks

# Benchmark training speed
python train.py config/train_gpt2.py --eval_only

# Loss tracking
# nanoGPT-reported GPT-2 (124M) on OpenWebText: val loss 2.85
# OpenAI GPT-2 (124M) on WebText: val loss ~3.11
# (domain gap between WebText and OpenWebText accounts for difference)

# Custom benchmarks
python train.py config/train_shakespeare_char.py --eval_iters=200

Key Patterns

1. Configs are Python files, not YAML — maximum flexibility, easy to diff
2. init_from='scratch'|'resume'|'gpt2*' — switch between training modes
3. Always use --compile=True on GPU for ~2x speedup
4. Gradient accumulation with gradient_accumulation_steps emulates larger batch sizes
5. Data is raw uint16 .bin files — tokenized once, loaded via memmap
6. Meta device init for large models — doesn't allocate until needed
7. Weight tying between wte (embedding) and lm_head — standard GPT practice
8. Use torchrun for multi-GPU — not python -m torch.distributed.launch

References

• nanoGPT Repository
• nchat (successor)
• minGPT (predecessor, more educational)
• GPT-2 Paper
• Attention Is All You Need