Release Notes

24.8.0 - August 1st 2024

This is a major release containing many substantial changes from previous releases.

Major changes

• Changed versioning scheme to CalVer: Year.Month.Micro
• Transitioned to more of a functional approach in SDK usage, allowing the user to integrate the CLIKA Compression logic into existing code using torch.compile
• Deploying models can now be done using a simple torch.onnx.export call.
• Implemented CLIKA as a backend for torch.compile
• Added initial support for Qualcomm's QNN. Please let us know if you encounter any issues!

Improvements

• Addition of improved model parsing capabilities
• ONNX export of ClikaModule now exports newer ops instead of resorting to the oldest version supported, if possible to do so

Bug fixes

• Fixed an issue with FloorDivide module in which the 'Floor' operation was not properly applied after deserialization of a ClikaModule.

Supported operations

Added support for more operations: sort, argsort, topk, masked_fill, scatter_nd, randint, unbind, sym_int, sym_float, tensor_split, amax, amin, builtin.max, builtin.min, normalize

Please note that we generally add support for both torch.function_name and Tensor.function_name but if you encounter a case where only one or the other is supported, please let us know at support@clika.io.

Future features

• Full Float16/BFloat16 training - the parameters will be in Float16/BFloat16. This will help reduce memory requirement.