[Installation] <title>can't manually download the wheels

[liuzhihua-web]

Is there an existing issue for this?

• I have searched the existing issues

Have you followed all the steps in the FAQ?

• I have tried the steps in the FAQ.

Current Behavior

No response

Error Line

I can't install torchsparse through pip,so I want manually download the wheels. I clicked on the website you provided.It shows
{"detail":"Not Found"}

Environment

- GCC:
- NVCC:
- PyTorch:
- PyTorch CUDA:

Full Error Log

Error Log

[PUT YOUR ERROR LOG HERE]

[twistalex]

have you solved this problem?can't install this package too

[pphuangyi]

Hi there, I have a successful experience I want to share here. It is a limited method and I cannot guarantee it will work on your platform, but it may be worth a try.

1. Create a conda environment by conda create -n torchsparse python=3.11 (choose a better name if you like);
2. Fork the repo, clone your fork, and cd into the folder;
3. [optional] Open the setup.py and install the requirements manually (I like to do this, but it might not be necessary), EXCEPT for PyTorch related packages.
4. Install CUDA toolkit version 11.8 by sudo apt install cuda-11-8. If you don't have sudo, try to ask the administrator to install it for you.
5. Switch to CUDA-11.8 following instruction here (it was my way of installing MinkowskiEngine). If you are on a server, you may load the module by module load cuda[version].
6. Install a version of PyTorch for CUDA-11.8 using commands provided here (for older versions of PyTorch, I used PyTorch 2.5.1)
7. Finally run python setup.py install.
8. You may need to switch back to your old CUDA after you are done.

NOTE: As you might have noticed, for me, the trick is to use slightly older versions of CUDA. My platform has g++ support for CUDA 11.8 but not CUDA 12.6 and that was why I failed to install torchsparse before. So

I think if you can achieve the following:

1. Get a CUDA version that has g++ support;
2. Get a PyTorch version compiled with the CUDA version in use.

Then you should be fine.

Please let me know your experience! Happy exploring sparse convolution with TorchSparse!