Pytorch support for sm120

I apologize if this is the incorrect place to ask… I have an rtx5080 that I was lucky enough to grab… but I can not get it to work with pytorch… I keep getting an error because pytorch does not support cm120… I’ve tried everything to work around it but I am getting stuck… Could someone please help

Blackwell (sm_100 and sm_120) is supported already if you are building PyTorch from source.
We are also working on enabling nightly binaries and first builds are already successful. We are currently verifying these and you could grab the very early build from the build pipelines, e.g.: from here (select the desired Python version and use the CUDA 12.8 builds).
Executed this wheel on a Blackwell GPU:

python -c "import torch; print(torch.__version__); print(torch.cuda.get_arch_list()); print(torch.randn(1).cuda())"
2.7.0.dev20250131+cu128
['sm_50', 'sm_60', 'sm_70', 'sm_75', 'sm_80', 'sm_86', 'sm_90', 'sm_100', 'sm_120', 'compute_120']
tensor([-0.0819], device='cuda:0')

Tomorrow’s nightly binaries should also be ready and we will confirm it once the builds are ready (already triggered), so instead of downloading test wheels you can wait until tomorrow to check if the builds went through.

1 Like

pip install --pre torch --index-url https://download.pytorch.org/whl/nightly/cu128 works now.
torchvision and torchaudio is still WIP.

2 Likes

Not able to download
ERROR: Could not find a version that satisfies the requirement torch (from versions: none)
ERROR: No matching distribution found for torch
thanks.

I should have added that these binaries are currently built for Python 3.9-3.13 and Linux x86 only. I don’t know which Python version, OS, or architecture you are using.

Thank you for your reply.
I think I have figured out the issue.
I am using Python 3.10 on AMD WIN64

is the nightly binaries out for python 3.9.13 and win64?

No, not yet. The binaries are available for Linux on x86 and ARM, so you could use WSL2 in the meantime.

1 Like

Also a user of Python 3.10, WIN64 and CUDA 12.8. I will be watching this thread for the good news of the nightly when it’s ready!

1 Like

Are there any other places we should be watching so that we’ll know the minute there’s a Pytorch that supports sm120? I keep checking this page and I don’t want to miss it!

PyTorch binaries using CUDA 12.8 are already available as nightly binaries for Linux (x86 and SBSA). If you don’t want to use WSL and are looking for native Windows support you could check when the binaries show up here but I will also update this thread once they are available.

3 Likes

Any updates? Is this something that might take another few months or weeks? I’m sure a million people are asking this, but I really have no idea how often to refresh this page, hah

The nightly wheels for Windows are available as of now, e.g. for Python==3.12 you can find the download here.

3 Likes

I just checked, and I think it’s finally out!! https://download.pytorch.org/whl/nightly/torch/

2 Likes

What about torchvision and torchaudio? When might that arrive?

Will this only work for python 3.12+? Currently running an environment that only works with 3.10.13

Yeah I am currently getting this error when trying to update:

ERROR: Cannot install torch==2.7.0.dev20250220+cu128 and torchvision==0.19.1+cu124 because these package versions have conflicting dependencies.

The conflict is caused by:
The user requested torch==2.7.0.dev20250220+cu128
torchvision 0.19.1+cu124 depends on torch==2.4.1+cu124

Need torchvision and torchaudio updates it seems.

2 Likes

Yes, 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. However, the current version of Torchaudio is only 2.6.0, so it seems that I need to wait for a while longer :pensive:

2 Likes

Training with RTX 5090 + Windows 11 Pro + Pytorch latest nightly with cu128 is ~10% slower than when using the RTX 4090 on the same script / same model (resnet f32) / machine configuration / same envrionment. Anyone can confirm this issue? Is the 5090 only faster on Linux or is the 5090 throttled for AI workloads?

So no torchvision and torchaudio yet, damn

2 Likes