autoresearch-mlx: Apple Silicon MLX Port of Autoresearch
autoresearch-mlx v1.0.0 Apple Silicon MLX port of karpathy/autoresearch for running autonomous AI pretraining research natively on M-series Macs. Highlights Full GPT model with RoPE, value embeddings (ResFormer), sliding window attention, softcap, and squared ReLU AdamW optimizer with 6 parameter groups and dimension-scaled learning rates Linear weight decay schedule for 2-5% BPB improvement over the reference fork NaN loss detection, proper eval batch sizing for 16GB machines, and __main__ guard for agent workflows Tunable --time-budget flag (default 5 min) MFU calculation, FLOP estimation, and detailed config logging Benchmark (M1 Pro 16GB,
Resource information
- Format
- zip
- Original repository
- 18956067
- DOI
- 10.5281/zenodo.18956067
- Credit basis
- Creators in Zenodo record metadata
- Access
- Open files available from the original Zenodo deposit
- Last checked
- Nov 5, 2026
- License
- mit-license
- Last record update
- Nov 5, 2026
About this resource
autoresearch-mlx v1.0.0 Apple Silicon MLX port of karpathy/autoresearch for running autonomous AI pretraining research natively on M-series Macs. Highlights Full GPT model with RoPE, value embeddings (ResFormer), sliding window attention, softcap, and squared ReLU AdamW optimizer with 6 parameter groups and dimension-scaled learning rates Linear weight decay schedule for 2-5% BPB improvement over the reference fork NaN loss detection, proper eval batch sizing for 16GB machines, and __main__ guard for agent workflows Tunable --time-budget flag (default 5 min) MFU calculation, FLOP estimation, and detailed config logging Benchmark (M1 Pro 16GB, 5-min budget) | Metric | Value | |--------|------:| | Starting train loss | 9.012 | | Final train loss | 6.762 | | val_bpb | 2.371 | | Steps | 55 | | Tokens | 3.6M | | Peak Memory | 11.0 GB | | Parameters | 11.5M | Quick start uv sync uv run prepare.py uv run train.py