v2.0.0 — MAJOR Version Change Stable Release

Released: April 2026

Upgrading from v2.0.0rc1

No user-facing API changes. This stable release closes out the v2 release candidate cycle. All CI matrices (Linux, macOS, Windows × Python 3.10/3.11/3.12) are green at 94 % test coverage.

Bug Fixes

  • Bug Fix Fixed TypeError: can't convert mps:0 device type tensor to numpy in CRPSLoss._crps_mixture_mc on Apple Silicon (MPS). The entire mixture MC computation now moves inputs to CPU at entry when an MPS device is detected, because keras.ops.mean/sum/abs internally call NumPy in the CI-runner version of Keras, which fails for MPS tensors.

  • Bug Fix Fixed ValueError: operands could not be broadcast together with shapes (2,3,2,1) (2,3,1) in CRPSLoss._crps_mixture_mc on Linux CI. Weight normalisation now uses tf_expand_dims(tf_reduce_sum(w, axis=2), axis=2) instead of tf_reduce_sum(w, axis=2, keepdims=True) for non-torch backends, making the shape reconstruction explicit and independent of keepdims behaviour in the installed Keras version.

  • Bug Fix Fixed TypeError in SqueezeExcite1D.call and drop_path when the output of keras.ops landed on a different device than the input tensor (MPS vs. CPU mismatch); added .to(x.device) guards matching the pattern already used in layer_utils.py.

  • Bug Fix Fixed TypeError from np.asarray(mps_tensor) calls in test_cov_components_keras.py; replaced with the _to_numpy helper (which first calls .cpu() on torch tensors) in all five affected assertion sites.

  • Bug Fix Fixed monkey-patch lambda np.asarray(x).shape in the test_query_mask_builds_attention_mask test — replaced with _to_numpy(x).shape to survive MPS tensors.

  • Bug Fix Fixed _get_static_value in base_attentive/__init__.py: TensorFlow’s get_static_value returns the input object unchanged for non-tensor Python objects; added a result is not value guard so the function correctly returns None in that case.

  • Bug Fix Fixed _tf_build_positional_encoding in components/misc.py to build the angle matrix entirely in NumPy and only materialise a backend tensor at the end via tf_cast, avoiding MPS in-place slice-assignment failures (mps_tensor[:, 0::2] = …).

  • Bug Fix Fixed broadcast_like in layer_utils.py: added a _scalar_to_int helper that calls .cpu().item() on each rep scalar before tf_tile, preventing keras.src.backend.torch.numpy.tile from calling repeats.int().numpy() on an MPS tensor.

Infrastructure / CI

  • Internal Test suite now passes the full Linux × macOS × Windows CI matrix at 94 % coverage (up from ~88 % at rc0).

  • Internal test_cov_losses_extended.py — test_mixture_mode_dict mock refactored across three rounds to minimise monkey-patching: only tf_gather is mocked (batch_dims=1 cross-backend compat); real Keras ops are used for all other computations.

Documentation

  • Internal Version badge in docs/index.rst updated to 2.0.0.

  • Internal Citation in README.md updated to version = {2.0.0}.

  • Internal pyproject.toml classifier promoted from Development Status :: 4 - Beta to Development Status :: 5 - Production/Stable.