Build for pytorch 1.13.1 and support for RTX 4000 series

[drapado]

🚀 Feature

Hi, if you try to run pytorch3D with a RTX 4000 gpu, you run into the following error:

File ~/miniconda3/envs/wm/lib/python3.9/site-packages/torch/nn/modules/module.py:1130, in Module._call_impl(self, *input, **kwargs)
   1126 # If we don't have any hooks, we want to skip the rest of the logic in
   1127 # this function, and just call forward.
   1128 if not (self._backward_hooks or self._forward_hooks or self._forward_pre_hooks or _global_backward_hooks
   1129         or _global_forward_hooks or _global_forward_pre_hooks):
-> 1130     return forward_call(*input, **kwargs)
   1131 # Do not call functions when jit is used
   1132 full_backward_hooks, non_full_backward_hooks = [], []

Which I believes come from the fact that neither pytorch3d nor pytorch were compiled to support the most recent gpus. I believe that compiling also for CUDA arch 8.9 should solve the issue.

Best, David

Motivation

Adding support for RTX 4000 gpus

Pitch

Adding support for torch 1.13.0 and CUDA arch 8.9

[TheKiLLerDz]

Any updates? I'm looking to buy a GPU from the 4000 series, and I really need it to work with PyTorch3D. Thank you in advance!

[bottler]

I think the official PyTorch 1.13.1 and 2.0.0 builds are done with compute capability up to 8.6 I think. So if you need 8.9 (which I'm not sure you do, due to compatibility) you need to be building pytorch and pytorch3d from source.