[FonaTech]

Summary

• Native kernel (chronos/model/kernel.py) replaces minimind: no clone, no sys.path change, offline import. Logits are bit-identical on out/chronos_256_moe.pth and the 50M en→zh checkpoint (full, padded, prefill+decode). One shared FP32 RoPE table (3.50 → 0.50 MiB at 50M, 216 → 24 MiB at 768H).
• Fixes that came with the migration: dense GQA mode (crashed on Torch and MLX), BF16 release checkpoint rewritten as FP16, trainer logs missing from the Web UI.
• Tool-call tags: decode_visible keeps <tool_call>/<tools> when a custom HF tokenizer registers them as special tokens.
• MLX training: compiled micro step on by default (measured 1.08× SFT, 1.18× pretrain on the 50M config), and activation recomputation (peak 5.62 → 3.50 GB SFT, 6.07 → 3.62 GB pretrain; off by default at 50M).
  • mlx 0.32.4 drops the gradient of a repeated checkpoint output, and a full-length cache slice is the same graph node without being the same Python object (37–40% wrong gradients). Region outputs are packed per dtype.
  • Exactness is tested per micro step on the same state inside the LongRun loop.
• Versions 0.3.2a0 (kernel and chronos-translate pin). README (en/zh) and CHANGELOG updated.

Test plan

• MLX-CPU suite: 531 passed, 1 skipped (six-stage deselected; run on GPU: passed)
• No-MLX CI simulation (py3.11, transformers 5.22.1): 395 passed, 67 skipped
• ruff 0.17.4 clean
• Wheel build + fresh-venv import with HOME empty: version 0.3.2a0, legacy tokenizer bundled, no minimind, nothing downloaded
• Real-data A/B (en→zh shards, 50M config): loss curves of eager / compiled / recompute / both agree within 2e-3 (SFT, 16 steps) and 2e-4 (pretrain, 12 steps)

🤖 Generated with Claude Code