v2.1.0 — Multi-Backend Stabilization and Compatibility Release

Released: April 2026

Upgrading from v2.0.1

This is a backward-compatible feature release focused on Keras 3 multi-backend stabilization, runtime/import hardening, and a new parameter-compatibility migration layer for the evolving BaseAttentive API.

Most projects can upgrade directly from v2.0.1. Existing code using the older GeoPrior-v3 kernel-style parameter names remains supported through explicit compatibility rules and deprecation warnings.

Highlights

• BaseAttentive is now stabilized as a Keras 3 multi-backend package across TensorFlow, Torch, and JAX.

• New compatibility migration layer under base_attentive.compat.versioning centralizes renamed parameters, deprecated aliases, precedence rules, and transitional warnings.

• Runtime bootstrap and imports are significantly lighter. Import-time backend probing and eager TensorFlow/Keras loading were reduced so collection, documentation, and lightweight runtime paths are safer on mixed environments, especially Windows.

• Examples and documentation notebooks were refreshed for the current v2 API and versioned as v2.1.0.

Compatibility and Migration

• Centralized parameter migration rules added in base_attentive.compat.versioning.

  The new compatibility layer provides:

  • declarative ParameterRule definitions,

  • dedicated warning classes,

  • resolve_deprecated_kwargs(...) and resolve_deprecated_config(...) helpers,

  • optional decorator-based application for constructors.

• Legacy GeoPrior-v3 kernel-style names remain accepted and are transparently mapped to the newer public API with warnings. The current rules include:

  • static_input_dim → static_dim

  • dynamic_input_dim → dynamic_dim

  • future_input_dim → future_dim

  • max_window_size → lookback_window

  • attention_levels → attention_stack

• Quantile-count compatibility added: n_quantiles is accepted and converted to evenly spaced quantiles.

• Transitional output-mode compatibility added: output_mode="gaussian" and output_mode="mixture" are accepted in the compatibility layer, with explicit warnings when the requested behavior is not yet wired into the active path.

• Precedence rules are now explicit. When both deprecated and canonical names are provided, the canonical parameter wins and the deprecated one is ignored with a warning.

Runtime and Backend Architecture

• Backend foundation refactored around a cleaner Keras 3 multi-backend runtime surface.

  • Root package imports were lightened.

  • keras_runtime was made lazy.

  • backend initialization was deferred so import base_attentive no longer behaves like full runtime initialization.

• Collection-safe bootstrap improvements reduce accidental TensorFlow/Keras imports during module import and test discovery.

• Explicit runtime imports were adopted in core modules. Runtime symbols such as KERAS_DEPS now come from their defining runtime modules (for example _bootstrap.py) instead of being pulled indirectly through package __init__.

• Torch and JAX resolver coverage expanded across the V2 component surface, including projections, pooling, fusion, decoder-side blocks, and forecast heads.

• Resolver-built V2 components are more serialization-safe. Generic resolver pieces that were previously plain callables are now tracked as serializable Keras layers where appropriate.

• Backend registration and registry aliasing were hardened. Resolver registrations now better tolerate repeated setup and legacy key aliases such as head.quantile.

Model Behavior and Output Contracts

• Forecast-output contract standardized.

  • point forecasts follow BHO

  • quantile forecasts follow BHQO

  This aligns the quantile path with the conventional batch × horizon × quantile × output ordering.

• TensorFlow builder kwargs filtering prevents resolver metadata (such as component_key) from leaking into Keras layer constructors.

• TensorFlow registration idempotence improved so repeated resolver setup no longer fails with duplicate component-registration errors.

• Torch/JAX runtime wrappers hardened with improved builder shims, alias coverage, and backend-safe paths for the V2 stack.

Fallback, Validation, and Numeric Stability

• Fallback numeric operations were hardened for mixed runtime situations and Torch-backed tensors.

  This includes safer conversion paths for:

  • reductions,

  • concatenation,

  • reduce_logsumexp,

  • validation helpers operating on BH/BH1/BHQO-style outputs.

• Temporal multiscale aggregation fallback paths were corrected so sum-based aggregation works consistently in non-native fallback situations.

• Probabilistic head fallback paths were improved so scalar NLL reductions no longer break when backend tensors flow through the safe NumPy fallback route.

Serialization and Configuration

• V2 config round-trips were stabilized.

  • nested spec serialization is normalized,

  • alias paths are cleaned during export,

  • from_config accepts the modern nested-spec form,

  • resolver-built assemblies survive JSON/config round-trips more reliably.

• V2 JSON/config workflows improved with better support for normalized spec export and reconstruction.

Documentation and Examples

• Examples refreshed for the current API and v2.1.0 release.

  Updated notebooks include:

  • 01_quickstart.ipynb

  • 04_standalone_applications.ipynb

  • 05_kernel_robust_networks.ipynb

  • 06_crps_probabilistic_forecasting.ipynb

  • 07_v2_spec_registry.ipynb

• Notebook content now reflects the post-v2.0.1 API, including refreshed constructor arguments, updated registry/spec examples, and versioned install/display snippets for v2.1.0.

• Release-note structure expanded to document the transition from the original TensorFlow/GeoPrior kernel usage toward the new, broader BaseAttentive compatibility and multi-backend design.