v2.0.1 — PATCH Version Change Bug Fix Release

Released: April 2026

Upgrading from v2.0.0

Drop-in replacement. No API changes. Only the MPS device-conflict bug described below is addressed.

Bug Fixes

  • Bug Fix Definitive MPS fix for CRPSLoss._crps_mixture_mc on Apple Silicon (Torch + MPS backend).

    Previous releases (including v2.0.0) attempted to work around keras.ops failing on MPS tensors by moving inputs to CPU at the start of the function. This was not sufficient: keras.ops.* calls such as cumsum, sigmoid, cast, reshape, and gather can silently re-dispatch CPU tensors back to the default Keras device (MPS on Apple Silicon), causing mu_s / sig_s to land on MPS while eps (explicitly moved to CPU) remained on CPU — producing:

    RuntimeError: Expected all tensors to be on the same device,
but found at least two devices, mps:0 and cpu!

    The fix introduces a pure native-PyTorch early-return path for the MPS case. When _on_mps is True, the entire sampling / gather / CRPS computation runs with torch.cumsum, torch.randn, torch.gather, etc. — bypassing all keras.ops entirely and keeping every tensor on CPU throughout. The non-MPS path (TensorFlow, JAX, CPU-Torch) is unchanged.

Documentation

  • Internal Expanded Architecture Guide (docs/architecture_guide.rst). The v2.0.0 Registry / Resolver / Assembly system is now documented in depth:

    • What changed from v1.0.0 and why

    • Breaking changes with migration examples

    • ComponentRegistry and ModelRegistry — how to register and resolve custom components

    • BaseAttentiveSpec / BaseAttentiveComponentSpec field reference

    • End-to-end build-from-spec example

    • Updated data-flow diagram reflecting the v2 assembly pipeline

  • Internal README badges synchronised with docs/index.rst: added version, semver, Keras ≥ 3.0, backends, last-commit, open-issues, PRs-welcome, and platform badges.

Examples

  • Feature Two new Jupyter notebooks added to examples/ and rendered as HTML pages within the documentation (see Example Notebooks):

    • 06_crps_probabilistic_forecasting.ipynb — covers all three CRPSLoss modes ("quantile", "gaussian", "mixture"), Monte Carlo sample-count guidance, output-shape interpretation per mode, and a sketch of a custom GradientTape-style training loop.

    • 07_v2_spec_registry.ipynb — end-to-end walkthrough of the v2 declarative configuration system: BaseAttentiveSpec creation, JSON round-trip serialisation, BaseAttentiveComponentSpec, registering a custom encoder via ComponentRegistry, ModelRegistry, and BaseAttentiveV2Assembly.

  • Internal Notebook rendering via nbsphinx integrated into the Sphinx build pipeline:

    • nbsphinx added to extensions in docs/conf.py with nbsphinx_execute = "never" so ReadTheDocs renders notebooks without requiring an ML runtime.

    • docs/conf.py auto-copies examples/*.ipynb → docs/notebooks/ at build time, keeping examples/ as the single source of truth.

    • New Examples section added to the documentation sidebar (docs/notebooks/index.rst) with a summary table of all seven notebooks and a Binder badge for interactive execution.

    • nbsphinx>=0.9.0 and ipython>=8.0.0 added to the [docs] optional dependency group in pyproject.toml.