Continual Inference Versions Save

Fixed

• Ensure state_buffer remains on the same device after clean_state.

Fixed

• Option to use strings to specify transformer activation.

Added

• Onnx as dev requirement.

Changed

• Allow torch>=2.0.

Added

• Skip module.
• "leading" mode in Residual.

### Added

• Description of state handling to README.

Fixed

• Documentation formatting for co.Identity() examples.
• Horovod check for newer pytorch lightning versions.

### Added

• query_index argument to SingleOutputTransformerEncoderLayer.

Fixed

• Residual centred residual and Delay auto_delay forward_step.

### Added

• Support for GroupNorm and InstanceNorm

### Added

• append function to co.Sequential.
• Production-ready docstrings for public functions.
• reduce_max to Reduce.

### Changed

• Rename Unity to Identity to follow torch.nn.
• Major overhaul of README, improving descriptions and adding benchmark.
• Major overhaul of docs, improving descriptions and adding benchmark.
• MHA warnings to only log once.

Removed

• Unused parameters batch_first and bidirectional for RNN, GRU, and LSTM.

Fixed

• co.Conditional onnx support for single-option config.

Fixed

• co.Conditional onnx support.