Hi, I would like to generate the source code for a module without having access to the target, for example generating CUDA source code without running TVM on a device that is enabled for CUDA. Is this possible? Otherwise is there a straightforward way that I could modify the TVM source to make it possible?
Thank you
For many targets, TVM already did that. For example, we allow you to generate opencl on a machine without opencl. For the case of CUDA, you only need nvcc compiler toolchain on your machine(for PTX) but do not necessarily need a CUDA runtime.
But it we want to tune model it really need a machine enabled with CUDA and specify the gpu arch, is that right?
yes, but you can use TVM RPC if you need to compile the model on a separate machine and then send the compiled artifact to a remote machine to time it.