Simple question: should I build PyTorch from sources? I know that for other frameworks (Tensorflow), the pre-compiled library might lack support to some CPU vector instructions (AVX, FMA, …) to broaden the compatibility to multiple devices/generations.
What does PyTorch pre-compiled library might differ from building from sources?
Building from source does support broader architectures, e.g., older GPUs, CUDA on macos, etc. But for CPU vector instructions, I think the binaries include them as well. That said, we are actively developing on this front, so building from source very possibly will give you newer and better optimized kernels.