I built a module on Nvidia K40m GPU and used export_library to save the module. When I try to run this module on Nvidia K80m, it gave an error, whilst I can run module built for K40m on K80m. How to choose GPU Type to K80m when I build a module on K40m please?
Are you currently only compiling in source mode, or are you building the module fully? My understanding is that only compiling to source is more flexible—though it may come with some additional initialization cost as the kernel must be JIT compiled before execution.
Try seeing what happens when you compile without building TVM with explicit CUDA support.