I wanted to try PyTorch on M1 Max, but I couldn’t…
Published:
TLDR
- When installing
pytorch
andtorchvision
together,pytorch
version is forced to fall back to latest stable release (1.11) which does not have Apple silicon support. - Same issue happened before. Conda install from pytorch-nightly channel with cpu-only option delivers 1.3.1 version
- New issue submitted. Conda install from pytorch-nightly channel does not install the expected version on macOS
PyTorch has announced support for Apple silicon GPUs for sometime. The official release will be in next v1.12, and is already available in nightly build.
When I was thinking about Friday evening activity for today, a thought came to me to play with PyTorch nightly a bit and see how it performs on my new Mac Studio with M1 Max. Sadly I failed because of some wired bug…
I googled a bit and found the currently best solution for installing native Python on Apple silcon is through miniforge
brew install miniforge
conda init zsh
conda activate
(If you are using other shells instead of zsh, adjust conde init
accordingly.)
Everything looked great and I happily installed pytorch
conda install pytorch torchvision torchaudio -c pytorch-nightly
Here comes the first problem: torchaudio
is not found
PackagesNotFoundError: The following packages are not available from current channels:
- torchaudio
Current channels:
- https://conda.anaconda.org/pytorch/osx-arm64
- https://conda.anaconda.org/pytorch/noarch
- https://conda.anaconda.org/conda-forge/osx-arm64
- https://conda.anaconda.org/conda-forge/noarch
To search for alternate channels that may provide the conda package you're
looking for, navigate to
https://anaconda.org
and use the search bar at the top of the page.
There is actually an active issue for this problem: Unable to install Preview (Nightly) on M1 macOS: “Symbol not found”.
This is not blcking me since I just need the other two.
conda install pytorch torchvision -c pytorch-nightly
Unexpectedly, conda does not pull from pytorch-nightly
channel, but the default one conda-forge
. PyTorch v1.11 got installed. This is the latest stable release without Apple silicon support!
I tried installing pytorch only without torchvision
conda install pytorch -c pytorch-nightly
and this time v1.13 was installed. Problem solved? Not really…
When instaling torchvision following pytorch
conda install torchvision -c pytorch-nightly
pytorch is forced to fall back to v1.11
The following packages will be SUPERSEDED by a higher-priority channel:
pytorch pytorch-nightly::pytorch-1.13.0.dev20~ --> conda-forge::pytorch-1.11.0-cpu_py39h03f923b_1
Now I’m blocked…
Actually the same issue has happened before where torchvision nightly build is bind with earlier version of pytorch: Conda install from pytorch-nightly channel with cpu-only option delivers 1.3.1 version.
I submitted a new issue and let’s hope this gets resolved or v1.12 get released ASAP. Conda install from pytorch-nightly channel does not install the expected version on macOS