These tools together solve a fraction of the problen. The other parts of the problem are interfacing with classic c, c++ libraries and handling different hardware and different OSes. It is not even funny how tricky it is to use the same GPU/CUDA versions but with different CPU architectures and hopefully most people dont need to be exposed to it. Sometimes parts of the stack depends on a different version of a c++ library than other parts of the stack. Or some require different kernel modules or CUDA driver settings. But I would be happy if there was a standardized way to at least link to the same C++ libraries, hopefully with the same ABI, across different clusters or different OS versions. Python is so far from solved…