New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
nvrtc: error: invalid value for --gpu-architecture (-arch) #87595
Comments
Hmm, can you please run |
4090 is sm89 that's explicitly supported only by cuda 11.8, and jiterator probably doesn't handle this correctly, we'd need to set some earlier sm version and hope that it's binary compatible. |
Yeah, but sm_86 should be compatible with sm89, so if we set maxversion to sm_86, jiterator should just work, shouldn't it? |
As @ngimel suggested probably adding few more entires to pytorch/aten/src/ATen/native/cuda/jit_utils.cpp Lines 885 to 900 in 512a3a4
|
Remaining high-pri. Natalia has an attempted fix in #87611, waiting for Nvidia to test. |
Fixes pytorch#87595 (maybe?) Pull Request resolved: pytorch#87611 Approved by: https://github.com/malfet, https://github.com/atalman
Fixes #87595 (maybe?) Pull Request resolved: #87611 Approved by: https://github.com/malfet, https://github.com/atalman Co-authored-by: Natalia Gimelshein <ngimel@fb.com>
@1697427528 We have deployed the fix, should be available in nightlies soon. Should be available on pytorch-test, can be installed using:
Can you please validate and confirm if its working for you |
Works for me on a 11.7 container with an AD102 GPU that repro'd the original bug. |
Fixes pytorch#87595 (maybe?) Pull Request resolved: pytorch#87611 Approved by: https://github.com/malfet, https://github.com/atalman
Fixes pytorch#87595 (maybe?) Pull Request resolved: pytorch#87611 Approved by: https://github.com/malfet, https://github.com/atalman
Have you solved this problem?I'm also having this issue |
This should be fixed in the nightly builds by now: check the nightly tab on https://pytorch.org/ |
is there an easy fix to use torch.compile with sm89 ? cf #90170 |
Fixes pytorch#87595 (maybe?) Pull Request resolved: pytorch#87611 Approved by: https://github.com/malfet, https://github.com/atalman
I having the same problem with rtx4070ti: "nvrtc: error: invalid value for --gpu-architecture (-arch)". Versions The pytorch 1.13.1 nightly from https://pytorch.org/ does not work for me. |
I'm not aware of a |
Sorry, I am referring to 1.13.1 on https://pytorch.org/. I saw stable and preview (nightly) tabs for cuda 11.7. Do they include the fix for this bug? I installed preview (nightly) but this bug still occurred with my 4070ti. |
You would need a nightly recent enough to include the fix; many of the |
the answer is at the end of this thread: #90170 |
I am not able to install Triton on my win10. (python version 3.8) C:\Users\yuanb> pip.exe install triton I also tried installing from source but it gave err too:
|
Is there a way to make this work with pytorch 1.10 and cuda 11.3? Currently I can't train OneFormer on RTX 4090. |
Yeah I have the same issue |
(8,9) |
Hi.I meet the same problem, here is my versions |
I am facing the same problem on RTX 4070ti, have you managed to find a solution to this issue? |
I meet the same problem. |
As RTX 4090 is sm8.9, it would not be supported in CUDA 11.3 (minimum of CUDA 11.8 is required). |
Thanks for your reply :) |
Solved by update |
Solved by update to:
|
🐛 Describe the bug
This problem only occurs when I use RTX4090.
when I use RTX3090, the problem doesn't occur.
Versions
RTX 4090
pytorch 1.14+cu117 nightly
cuda 11.8
cudnn 8.6.0
ubuntu 20.04
cc @ezyang @gchanan @zou3519 @ngimel @mruberry
The text was updated successfully, but these errors were encountered: