Major new features: Drop support for C. Only C++ and Python are now supported. Python and C++ code now throw more detailed exceptions allowing fine-grained control, especially for executions on GPU. Memory handling on GPU: memory is now automatically freed in case of an error. This allows for catching the Exception in case of error and continuing the computation without a forced exit of the process. Improvements: build the CUDA code and import the Python library without errors, even without a GPU. A runtime error is thrown if code is attempted to be executed on the GPU. cuBLAS handle is now created only upon request, not by default easier to navigate documentation