Since I need fast computing and the Theano gpuarray was so fast, I have been maintaining Theano, and have kept it up to date so that it works with latest Numpy (> 1.23.5), Scipy. If anyone is interested, contact me and I will make it available as a tar archive. Every change I made is marked with my initials ‘PMB’. Perhaps someone could clean it up and update the Theano Github, but I have no time for that, but would gladly provide the code.
I have uploaded the archives to Index of /theano. There are two archives: theano_1.0.5-pmb.tgz, and theano_1.0.5-pymc.tgz, corresponding to the latest theano releases I could find from regular and PyMC theano, repspectively. Both have been tested with GPU functionality using Numpy 1.24.2 as well as CUDA-12. Note: all the changes have been marked with ‘PMB’ so can easily be found. This applies as well to some C-code, notably gpuarray/c_code/dnn_fwd.c, gpuarray/c_code/dnn_gw.c, gpuarray/c_code/dnn_gi.c.