Skip to main content
Version: 24.9.0

Release Notes

24.9.0 - September 10th 2024​

note

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.

Features​

  • BFloat16 training is now possible - This will help reduce memory requirements.

Improvements​

  • Improved the torch.compile functionality by trying more ways to do so without failing first,

Bug fixes​

  • Fixed an issue with Detach layer. The SDK was unable to serialize it properly.
  • Fixed an issue with Tensor.size() not being correctly connected to successor nodes.
  • Fixed an issue with multiple ClikaModules living concurrently. After torch.compile(..., backend='clika') if you loaded another ClikaModule from serialized it could potentially fail due to global state collision.
  • Significantly improved the 'svg' file generation runtime using ClikaModule.clika_visualize method. On gigantic models it used to hang, seemingly forever. Let us know if there are any further issues.
  • Fixed an issue when deploying Clamp layer to ONNX.
  • Fixed an Input ordering issue when input_names is provided to torch.onnx.export

Future features​

  • Full Float16 training - the parameters will be in Float16
  • More work on Multi-GPU capability and make it more efficient
  • Optimization for VRAM/RAM consumptions
  • Adding a method on the ClikaModule object to get the Quantization Sensitivities sorted by highest first. This will make understanding your model easier in addition to having the 'svg' files.
  • Additional Quantization Algorithm that will reduce the time even further.