Having trouble getting the mps version of torch to install

Mac OS X
1

I’m on Sonoma 14.7.1, using conda. A few weeks ago, I followed this guide: Accelerated PyTorch training on Mac - Metal - Apple Developer

I used the conda installation method: conda install pytorch torchvision torchaudio -c pytorch-nightly, at least as far as i remember, and according to conda’s list, installed: pytorch=2.3.1=gpu_mps_py310h7b7e308_100

Then, yesterday, I broke my installation when installing a requirements.txt file of a codebase i was working with. Stupid, i know, but then I tried to recreate the conda env from scratch. After doing the above, my installed pytorch shows:
pytorch=2.6.0.dev20241112=py3.10_0

I have ended up in multi day troubleshooting holes because of conda and pip issues with torch, so i’m hoping someone may be able to clearly explain why this even happened? Specifically, how did i get the mps-specific torch installed before, and why isn’t it installing now?

2

Hello, i installed it from this fork and it “kinda” works. GitHub - LalithShiyam/pytorch-mps: Tensors and Dynamic neural networks in Python with strong GPU acceleration, the version is 2.3.0 (torch base version). There are a LOT of things that are not supported in torchmps right now, like transpose convolutions etc etc. This package seems to have various implementations, however there are things that it will crash hard.

tl;dr pip3 install git+https://github.com/LalithShiyam/pytorch-mps.git