MLJAR AutoML 1.4.0b1

mljar-supervised 1.4.0b1

First beta release for testing the 1.4.0 installation and prediction workflows.
Requires Python 3.11+; Python 3.11 and 3.12 are tested. Install the exact
version explicitly: beta versions sort before the previously published RCs.

python -m pip install "mljar-supervised==1.4.0b1"

Feature discovery is included in this base installation. Choose a PyTorch build
before adding optional neural/TFM extras; use [apps] to run exported apps.
See the installation guide for CPU/CUDA commands.

• Add [apps] with mercury>=3.2.12,<4, keeping [all] limited to optional model backends. Export requirements for the selected prediction graph, pin MLJAR/joblib/scikit-learn versions, and bundle shared discovery transformations plus TFM weights/licenses for portable predictions.

• Require Python 3.11+ for the base package, all extras, and newly generated prediction apps. Test Python 3.11 and 3.12; older releases remain available for older interpreters.

• Include features-goldmine>=1.1,<1.2 in the base installation. Keep [feature-discovery] as an empty compatibility alias and preserve existing mode/validation rules and feature_discovery=False. Neural/TFM models and the app runtime remain optional.

• Install structured-data-models==0.2.1 from PyPI through [tfm], [kumo], [tabiclv2], or [all]. No separate Git backend installation is needed for beta 1.

• Keep the exact backend-version check and existing pinned model checkpoints. Weights still download on first use.

• The published 1.4.0a1 retains its original Git backend requirement; use a fresh environment when upgrading to beta 1.

• Add a macOS compatibility probe before mixing PyTorch with the standard native ML libraries. All probe cases share a three-second deadline. Automatic selection can skip optional Torch models; explicit requests and incompatible saved models raise NativeCompatibilityError before execution. Decisions appear in run reports. Training stays in the main process; Linux and Windows retain their existing paths.

• Correct saved-model test metadata, quoted app-path expectations on Windows, and simulated GPU allocation diagnostics. Preserve test reports and separate app CI stages for easier diagnosis.

Known limits: optional Torch models can be refused on incompatible macOS native
runtimes; the guard detects unsafe execution rather than repairing the installed
libraries. Apple MPS is not supported by these integrations. Local CUDA checks
cover Linux, an RTX 3070, PyTorch 2.11.0+cu128, and SDM 0.2.1, including all three
Kumo sizes. Python versions newer than 3.12 are not yet validated.