Changelog

Version 1.0.6 [12 Jun 2026]

In TrainableModule, move tensors with non_blocking=True in all code paths.
Explicitly pass weights_only=True to torch.load for consistency and safety in older PyTorch installations.

If predict_step is a generator function, it is assumed to generate individual predicted items. This is a convenient alternative to generating full batches in predict_step and overriding unpack_batch.
Add sort_keys option to Logger.log_config method.
Add EarlyStopping callback that stops training if a monitored metric fails to improve for a given number of epochs.
Add patience option to KeepBestWeights and SaveBestWeights callbacks.
Add baseline option to the SaveBestWeights callback.
Dynamically resize progress bar on window resize.
Export StopTraining type to replace the invalid Literal[STOP_TRAINING].
Fix a crash in TrainableModule.get_tb_writer by using a correct method name.
Fix producing logs more frequently than MINNT_REPORT_EACH when more then 10 seconds have elapsed between updates.
Fix incorrect recursive calls in unpack_batch.
When overriding dataset limit, do not create a Subset dataset.
When wrapping a module by passing it to TrainableModule constructor, correctly pass all arguments (both positional and keyword) in forward.
Do not consider norm layer subclasses in initializers_override.
Override also InstanceNorm epsilon in global_keras_initializers.
The WandBLogger.log_graph method no longer tries to log the graph to WandB.

Improve format_logdir to correctly handle environments without __file__ and calls with no kwargs.

Require setuptools < 80.9 as a temporal workaround for removed pkg_resources in setuptools == 82.0.
Improve typing (Metric is now just a protocol, together with Loss they are more generic, fit and evaluate return explicitly dict[str, float]).

Support lightweight profiling mode with reduced overhead collecting only basic information.
In PyTorch 2.10, allocator settings are set via a new generic interface, not via the original CUDA-specific interface (which now generates a deprecation warning). Therefore, start using the generic allocator settings interface when available.
In PyTorch 2.10, avoid false-positive warning about acc_events when profiling starts.