[FonaTech]

Summary

Release 0.3.0. The full list is in CHANGELOG.md → [0.3.0].

• v3 recurrent-depth kernel
  • Attention modes: shared_first, per_iteration, first_only.
  • Tensor-only ACT exit.
  • top-2 MoE plus a shared expert, with sorted grouped dispatch.
  • MLA decode paths and int8 KV.
  • Single-dtype precision chain.
• Self-speculative decoding
  • Runs on Torch and MLX, with full DRAM or lazy SSD experts.
  • Greedy output is exact.
  • The Web UI Inference tab gets a switch, draft depth and γ on all four runner paths.
• Task-agnostic Web UI
  • The monitor discovers eval metrics, samples and inspections from whatever a run logs.
  • The Data tab detects chat / instruction / preference / QA / parallel / plain-text schemas in jsonl, json, csv, tsv, parquet and txt.
• New commands: chronos serve (Ollama/OpenAI protocol), tokenizer, corpus, longrun, monitor, bundle, and the planner.
• Fixes
  • The MLX engine forces eval mode. Training-mode expert masking had made greedy output random.
  • The lazy Torch engine uses no_grad. Experts paged in under inference_mode broke later calls.
  • The Inference tab now uses the checkpoint's own tokenizer.
• Packaging and CI
  • torch>=2.6, because transformers 5 disables its torch backend below that version.
  • CI pins tested versions in .github/ci-constraints.txt.
  • Removed the dead token-based publish job. PyPI publishing stays in pypi.yml (trusted publishing).
  • Added .gitignore and stopped tracking __pycache__.

Speculative decoding: measured, not faster at 50M

Measured on the 50M en→zh pretrain checkpoint (step 13,996): greedy, batch 1, FLORES+ dev, draft depth 1, γ 4. Reports are in reports/speculative/.

Runtime	Plain tok/s	Spec tok/s	Ratio	Identical
MLX GPU bf16, full	112.0	107.4	0.96×	22/24
MLX GPU bf16, lazy	43.7	32.0	0.73×	22/24
Torch MPS fp32, full	32.1	25.9	0.81×	12/12
Torch MPS fp32, lazy (sync)	17.0	14.7	0.86×	12/12

The switch therefore stays off by default.

Tested

• Full suite on macOS (M5 Max, torch 2.10, mlx 0.32.4, gradio): 328 passed, 2 skipped. The skips need local shards and the v2 reference checkpoint.
• CI-equivalent environments built with uv, using the same installs as the workflows:
  • Python 3.11: ruff check . clean, flake8 fatal checks 0, pytest tests/ 249 passed, 49 skipped.
  • Python 3.10: the same result, plus bare pytest as python-app.yml runs it.
  • The skips are MLX, gradio and playwright tests, which are unavailable on Linux CI.
• python -m build and twine check pass for the sdist and the wheel. The wheel contains no weights or run data.

Not included

• Model weights, training runs, datasets and bundles. The training-data licences are unknown, so the weights are research-only.

🤖 Generated with Claude Code