can PyTorch run smoothly via Rosetta 2?

@qiangbo1222 it probably can, but it will be slower than naive binary, wouldn't it?

I got an m1 Macbook Air and a 2018 Intel Macbook Air right here. For an identical CNN for MNIST, the Intel Macbook took some 1975 seconds whereas the m1 took 2100 seconds.

So the Apple Silicon Macbook is slower, but on paper, it does 'run smoothly' I suppose.

PS: I'm a complete noob to developing on PyTorch but I'd be glad to do some help here or there since I have one of these Macs if the team can just help point me on what to do.

@mwidjaja1 can you check if https://ossci-macos-build.s3.amazonaws.com/torch-1.8.0a0-cp38-cp38-macosx_11_0_arm64.whl installs on your system, and if it does, how much faster will it be than x86_64 build?

I got an m1 Macbook Air and a 2018 Intel Macbook Air right here. For an identical CNN for MNIST, the Intel Macbook took some 1975 seconds whereas the m1 took 2100 seconds.

So the Apple Silicon Macbook is slower, but on paper, it does 'run smoothly' I suppose.

I tested one of my projects that uses transfomer networks with libtorch and inference time was several times slower under Rosetta 2 on a MacBook Air M1. Note that Rosetta 2 does not support AVX, AVX2, or AVX512 instructions, so it is expected that x86_64 PyTorch will be much slower in many cases on M1 Macs.