Can't install pytorch-directml with Poetry

[davidgilbertson]

I use Poetry for dependency management.

I've created a new environment with Python 3.8 and installed torchvision==0.9.0, but poetry won't allow removing torch as the readme suggests, because it's a requirement of torchvision.

Some questions:

• Is install with Poetry possible?
• Is it likely that pytorch-directml will always be a few versions behind for torch, torchvision and Python itself? It's rather difficult to take an established app on Python 3.10 and the latest version of a bunch of packages and make it all work with old versions. I don't mean to sound pushy, I just want to get my expectations right; if there will always be a lag I'll look for other ways to get PyTorch + Windows + AMD working.
• The MSDN docs say that "only" torch version 1.8 is supported, this seems to contradict the readme which says 1.8 will cause collisions. Which is correct?

[shashankp]

Try recreating env by first installing pytorch-directml & then torchvision
You havent mentioned the exact error you see.

[zhangxiang1993]

Answers to your questions:

1. We never try with poetry.

   • Maybe try install pytorch_directml first and then install torchvision with no-dependencies.
   • Or, install pytorch_directml first, and build torchvision from source on release/0.9.0 branch

2. Yes, the one we released is behind torch, and we target to build with the latest torch today.

3. The MSDN docs might be misleading. The collision will be caused if you have both torch==1.8.0 and pytorch_directml installed.

[smk2007]

Hi @davidgilbertson, torch-directml has moved to a plugin model. Check out the latest release here: torch-directml. The plugin model resolved many of the dependency issues that could cause issues like you saw above. Please take a look and reactivate if you still face any issues using torch.nn with DirectML.

Thanks!