Pytorch support for sm120

deployment
21

Any ETA on windows torchvision?

2 Likes
22

You are expecting far too much here

23

Windows torchvision for Blackwell support has already been released. Torchaudio is lagging behind though as far as I know:

1 Like
24

We are so close. Just a little torchaudio update left

1 Like
25

Looks like nightly includes Torchvision for Windows now :slightly_smiling_face:

https://download.pytorch.org/whl/nightly/torchvision/

  1. cu128-cp39-cp39-win_amd64
  2. cu128-cp310-cp310-win_amd64
  3. cu128-cp311-cp311-win_amd64
  4. cu128-cp312-cp312-win_amd64
  5. cu128-cp313-cp313-win_amd64
  6. cu128-cp313-cp313t-win_amd64
2 Likes
26

What’s the latest? torchaudio? Mainly, I’m looking to use ComfyUI with the 5090 on Windows and waiting for full compatibility

27

All audio builds fail on Windows due to: [binary builds] torchaudio Windows cuda nightly builds are failing since 01.29.2025 · Issue #3877 · pytorch/audio · GitHub

28

May I ask when you will probably be able to resolve this issue?

2 Likes
29

Any update on torchaudio builds for Windows?

30

Build issues were resolved already and nightly binaries are now supported on all platforms as of last week. Sorry, but too many cross-posts were created and I missed to update this thread.

2 Likes
31

I have also encountered the same issue. I have been keeping an eye on this post and waiting for PyTorch CUDA 12.8 to support Windows.

32

Take a look at my post just above yours: Pytorch support for sm120 - #30 by ptrblck

33

hi! im using a 5070ti and i cant figure out how to make it work ..should i biuld it from source? or there s already a version?
someone please help

34

Install the nightly binaries with CUDA 12.8 from the install matrix on our website.

35

Even when updated with the nightly binaries, it’s not working for me on a rtx5080. AMD/Win11.

“UserWarning: NVIDIA GeForce RTX 5080 with CUDA capability sm_120 is not compatible with the current PyTorch installation”

36

What does torch.__version__ return in this environment?

37

torch.__version__ returns 2.6.0.dev20241112+cu121

Still getting this warning:
UserWarning: NVIDIA GeForce RTX 5080 with CUDA capability sm_120 is not compatible with the current PyTorch installation.

And torch.cuda.get_arch_list() returns:
['sm_50', 'sm_60', 'sm_61', 'sm_70', 'sm_75', 'sm_80', 'sm_86', 'sm_90']

Running on Windows 11 (AMD build), Python 3.10.11.

38

Your PyTorch binary ships with CUDA 12.1 which does not support Blackwell architectures. Install the latest nightly binary with CUDA 12.8 ad explained before and it will work.

39

Hey ptrblck,

I’ve installed the latest PyTorch nightly with CUDA 12.8 (2.8.0.dev20250405+cu128), and everything initializes fine. However, during training, I’m now getting this error:

RuntimeError: CUDA error: no kernel image is available for execution on the device

It’s being triggered inside torch.nn.functional.embedding. Looks like some kernels (like embedding) still haven’t been compiled for sm_120, even in the latest nightly.

Just wanted to flag this in case it’s still being rolled out — happy to test again as new builds drop.

Thanks again for your help so far!

40

WE are compiling all native kernels for sm_100 and sm_120 in the nightly binaries with CUDA 12.8. What does cuda-gdb return if you rerun your workload and use set cuda api_failures stop inside cuda-gdb before running your workload?