You can directly install with pip and set TORCH_CUDA_ARCH_LIST to specify the cuda architecture if needed.

export TORCH_CUDA_ARCH_LIST="5.2 6.0 6.1 7.0 7.5 8.0 8.6+PTX"

pip install -v .

import torch
from vis4d_cuda_ops import ms_deform_attn_forward, ms_deform_attn_backward
...

1. Add cuda and cpu ops.
2. Delcare its Python interface in src/vision.cpp.