Together with @lkdvos , we are pleased to announce v5 of TensorOperations.
In this version, we have further improved the AD support and further extended the support for different backends and allocation strategies to deal with temporaries in contracting tensor networks. In particular, via a package extension we can now use Bumper.jl for allocating any temporary object created during the evaluation of a complicated tensor network. Finally, we now support cuTENSOR v2 and with that the more recent versions of CUDA.jl.
All of this required a number of breaking changes to the interface, which we now hope is sufficiently robust for quite some time, while the package is further extended. While we hope that the transition for everyone is smooth (the basic user interface of the package has not changed), we are ready to address any hickups or bugs that you may discover.