v2.0.0rc0 — MAJOR Version Change Release Candidate

Released: April 2026

This is the first release candidate of the v2 architecture. It is feature-complete and considered stable for production use; the rc0 tag will be dropped once the CI suite is fully green across all three backends (TensorFlow, JAX, PyTorch).

Upgrading from v1.0.0

No user-facing API changes relative to v1.0.0. The release candidate hardens the internal test infrastructure and continuous-integration pipeline. All constructor signatures, parameter names, and return shapes are identical to v1.0.0.

Bug Fixes

  • Bug Fix Resolved TypeError: metaclass conflict that prevented importing the four Keras Loss sub-classes (MeanSquaredErrorLoss, QuantileLoss, HuberLoss, WeightedLoss) when the NNLearner mixin — whose metaclass is incompatible with TensorFlow’s trackable metaclass — was present in the inheritance chain. Fixed by removing NNLearner from all four loss classes.

  • Bug Fix Resolved ModuleNotFoundError: No module named 'torch' that caused pytest collection to abort entirely on environments where PyTorch is not installed. The bare import torch at module level in test_cov_implementations_gaps.py was replaced with pytest.importorskip("torch", ...) so the file is gracefully skipped rather than failing.

  • Bug Fix Resolved TypeError: function() argument 'code' must be code, not str during test collection on TensorFlow-only environments. Root cause: the Keras-patch helper in test_cov_components_pure.py and test_cov_components_keras.py used lambda *a, **kw: None as the default fallback for unknown KERAS_DEPS attributes (including Loss, Layer, Model). Subclassing a lambda triggers Python’s function type as the metaclass, which then receives the class name as a string in place of a code object. Fixed by:

    • Removing os.environ.setdefault("KERAS_BACKEND", "torch") — the tests now run against whatever backend is available (TensorFlow on CI).

    • Introducing _KerasStub — a genuine Python class — as the default fallback so any unknown Keras base class resolves to a safe, importable stub.

    • Explicitly mapping "Loss", "Layer", and "Model" in _FALLBACKS to _KerasStub, and "register_keras_serializable" to a no-op decorator factory.

Internal

  • Internal Test collection: all 178 tests across test_cov_components_pure and test_cov_components_keras now collect without errors on TensorFlow-only environments.

  • Internal CI matrix continues to target Python 3.10, 3.11, and 3.12 with the TensorFlow backend; JAX and PyTorch jobs are skipped cleanly when the respective backend is absent.