Environment (WP31)
- Laptop (development, Phase 4): Apple M5 Max, 18 CPU cores, 40 GPU cores, 48 GB unified memory, macOS 27.0, Python 3.12 (uv), PyTorch 2.15.0 (MPS + CPU), MLX 0.32 available but not used by the reference implementation.
- MacLustr nodes used (existing passwordless SSH on the LAN; read-only audit in
hardware/device_inventory.json): M3U96a / M3U96b (Apple M3 Ultra, 32 cores, 60 GPU cores, 96 GB) for Stage A pretraining; M2U64 (M2 Ultra, 64 GB), M4M64a/c (M4 Max, 64 GB) for editor/controller runs. Each node has its own venv built byscripts/cluster_setup.sh(same lock file). No CUDA device was available; no cloud or paid service was used. - Precision: fp32 everywhere (bf16 autocast gave no throughput gain on MPS: 15.3k vs 15.0k tok/s).
- Determinism: seeds fixed (
training.common.seed_all); MPS kernels are not bit-deterministic across devices, so cross-device equivalence tests use tolerances (1e-4) while bypass identity tests are bitwise on CPU. - Dependency lock:
requirements-lock.txt(uv pip freeze, 2026-10-29).