A minimal probe into what torch.compile's Inductor backend generates per device (CPU, CUDA, MPS, XPU): Triton, Metal, vectorized C++, or a library dispatch. - View it on GitHub