Any updates for CUDA?

[fsbarros98]

Is there any way to install this with CUDA 11.8 and PyTorch 2.2? I've followed the tips in #6 and was able to install but when running I get the following error:

    match = emd_cuda.approxmatch_forward(xyz1, xyz2)

     RuntimeError: Unknown layout

[stefanengelmann]

Hi, I am able to run with CUDA 11.8, PyTorch 2.2.1 and Python 3.11.8. However, others that seem to have the same error with PyTorch 2.2.1 suggest to downgrade to PyTorch 2.1.0 as it appears to be a torch problem, see e.g. Vahe1994/AQLM#20.

Hope it helps!

[xiaozhi-alan-zhu]

Hi, I am able to run with CUDA 11.8, PyTorch 2.2.1 and Python 3.11.8. However, others that seem to have the same error with PyTorch 2.2.1 suggest to downgrade to PyTorch 2.1.0 as it appears to be a torch problem, see e.g. Vahe1994/AQLM#20.

Hope it helps!

Hello @stefanengelmann, thanks for your information. I was wondering what is your gcc version?

[xiaozhi-alan-zhu]

I also followed the suggestion in #6, but still get the error.

Here is the error message:

Loading extension module emd_ext...
load emd_ext time: 25.744s
Traceback (most recent call last):
  File "/home/pandu/Documents/Research-Projects/diffusion_models/LION/test_func.py", line 18, in <module>
    res = emd_cuda.approxmatch_forward(a, b)
          ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
RuntimeError: CUDA error: an illegal memory access was encountered

Following are the code snippet that caused the error:

import torch
from third_party.PyTorchEMD.backend import emd_cuda_dynamic as emd_cuda 


dim = 3
num_samples = 1
num_points = 100
a = torch.rand((num_samples, num_points, dim)).cuda()
b = torch.rand((num_samples, num_points, dim)).cuda()
res = emd_cuda.approxmatch_forward(a, b)

[Xunmenggod]

Does this repo support for cuda 12.1?