I am using MacBook Pro (16-inch, 2019, macOS 10.15.5 (19F96))
GPU
I am trying to use Pytorch with Cuda on my mac.
All of the guides I saw assume that i have Nvidia graphic card.
I found this: https://github.com/pytorch/pytorch/issues/10657 issue, but it looks like I need to install ROCm, and according to their Supported Operating Systems, it only supports Linux.
Is it possible to run Pytorch on GPU using mac and AMD Graphic card?
No.
CUDA works only with supported NVidia GPUs, not with AMD GPUs.
There is an ongoing effort to support acceleration for AMD GPUs with PyTorch (via ROCm, which does not work on MacOS).
CUDA is a framework for GPU computing, that is developed by nVidia, for the nVidia GPUs. Also, the same goes for the CuDNN framework.
At the moment, you cannot use GPU acceleration with PyTorch with AMD GPU, i.e. without an nVidia GPU. The O.S. is not the problem, i.e. it doesn't matter that you have macOS. It is a matter of what GPU you have.
What you can do though, is that you can either purchase an external nVidia GPU or use some cluster. For example, Google Colab offers PyTorch compatibility.
Announcement: https://pytorch.org/blog/introducing-accelerated-pytorch-training-on-mac/
To get started, install the latest nightly build of PyTorch: https://pytorch.org/get-started/locally/
Unfortunately, no GPU acceleration is available when using Pytorch on macOS. CUDA has not available on macOS for a while and it only runs on NVIDIA GPUs. AMDs equivalent library ROCm requires Linux.
If you are working with macOS 12.0 or later and would be willing to use TensorFlow instead, you can use the Mac optimized build of TensorFlow, which supports GPU training using Apple's own GPU acceleration library Metal.
Currently, you need Python 3.8 (<=3.7 and >=3.9 don't work) to run it. To install, run:
pip3 install tensorflow-macos
pip3 install tensorflow-metal
You may need to uninstall existing tensorflow distributions first or work in a virtual environment.
Then you can just
import tensorflow as tf
tf.test.is_gpu_available() # should return True
It will be possible in 4 months, around march 2022. See Soumith reply to this question on GitHub. https://github.com/pytorch/pytorch/issues/47702