feat(mps): Apple Silicon MPS port — Flash Attention 3 → PyTorch SDPA

[Rensp]

Summary

Port autoresearch to run fully on Apple Silicon (MPS) — no CUDA required.

• Replace Flash Attention 3 (CUDA/Hopper-only) with PyTorch F.scaled_dot_product_attention
• Remove kernels dependency; switch torch source from CUDA 12.8 index to standard PyPI
• Port prepare.py: MPS device detection, remove unsupported pin_memory=True
• Rewrite MuonAdamW: Python float scalars instead of CPU tensors (fixes MPS cross-device errors); disable torch.compile (not stable on MPS)
• Add read_result.py: helper to parse run.log with \r carriage returns
• Add program.md: MPS-specific experiment loop instructions for autonomous Claude Code research

Motivation

This port enables overnight autonomous ML experiments on Apple M-series hardware. The 5-minute training budget maps well to MPS throughput (~11.5M param model, batch=4).

Baseline result (Apple M-series, MPS)

config	val_bpb	params	steps
depth=4, SDPA, batch=4	1.5258	11.5M	318

Test plan

• uv run python prepare.py --num-shards 2 completes on MPS
• uv run python train.py runs 300s and prints val_bpb:
• No CUDA or kernels imports needed

🤖 Generated with Claude Code