PyTorch 2.0 support

[Looong01]

PyTorch 2.0 has been released.
But when I try to install torch-directml, it removed torch-2.0.0 and installed torch-1.13.1
We need it because it has fixed the problem of #400

Besides, can you set the version checking as >=version (Like 1.13.1) rather than ==version (Like 1.13.1) in the pip installing processing?

[lshqqytiger]

I think there should be more fixes and adaptations to support PyTorch 2.0.
Not just installing PyTorch 2.0.0 instead of 1.13.1.
Nevertheless, it will be nice to support PyTorch 2.0.
I'm also waiting for it.

[iperov]

forget about 2.0
There are too many bugs in current version.

[zhangxiang1993]

We have torch-directml 0.2.0 published today, which is built with torch 2.0.0.

[znsoftm]

I have test 0.2.0, it works well with transformers. Thank you guys.

[Looong01]

I have test 0.2.0, it works well with transformers. Thank you guys.

How? I still get this:

import torch
import torch_directml
dml = torch_directml.device()
tensor1 = torch. Tensor([1]).to(dml)
tensor2 = torch. Tensor([2]).to(dml)
tensor1 = tensor1.new(tensor1.shape[0]).fill_(0)
tensor2 = tensor2.new(tensor2.shape[0]).fill_(0)
print("sum:", (tensor1 + tensor2).item())


---------------------------------------------------------------------------
RuntimeError                              Traceback (most recent call last)
Cell In[24], line 8
      5 tensor1 = torch.tensor([1]).to(dml)
      6 tensor2 = torch.tensor([2]).to(dml)
----> 8 tensor1 = tensor1.new(tensor1.shape[0]).fill_(0)
      9 tensor2 = tensor2.new(tensor2.shape[0]).fill_(0)
     11 print("sum:", (tensor1 + tensor2).item())

RuntimeError: new(): expected key in DispatchKeySet(CPU, CUDA, HIP, XLA, MPS, IPU, XPU, HPU, Lazy, Meta) but got: PrivateUse1

[Lovinity]

As of pyTorch 2.0 and 2.0.1 (Python 3.11; which is supposed to work with 2.0.1), pip is still erroring and cannot install torch-directml basically saying there are no compatible versions (implying it still won't work with pyTorch 2 / 2.0.1).

Looking in the samples, I see there are none for pyTorch 2, so I don't know how you can say it works and close the issue when it clearly doesn't and does not have the samples to do so.

This is also a big problem because the latest supported version, pyTorch 1.13, can no longer be installed via conventional means like pip.

Thus, torch-directml is basically broken.

EDIT: Looks like the problem is actually Python... seems directml does not yet support torch 2.0.1 / Python 3.11, but it will work on torch 2.0.0 / Python 3.10.6. So to summarize:

• PyTorch 2 / Python 3.10.6: Works
• PyTorch 2 / Python 3.11: Does not work (expected; PyTorch 2 does not support python 3.11 until 2.0.1)
• PyTorch 2.0.1 / Python 3.10.6: Untested. I expect it to work.
• PyTorch 2.0.1 / Python 3.11: Does not work (unexpected; maybe directml does not support Python 3.11)

[lshqqytiger]

Every release of torch-directml is bounded to a specific version of torch. (Because torch-directml-native should be built with torch)
torch-directml==0.2.0.dev230426, which is the latest one, is bounded to torch==2.0.0.
So only Python 3.10 + torch 2.0.0 will work.

[aljohnston112]

Docs are very out of date. This issue is the only "recent" source of version compatibilities I could find.

[kovtcharov]

any plans for supporting torch-directml going forward (torch 2.2.1 and 3.12.2)?