Picsart-AI-Research/VideoINR-Continuous-Space-Time-Super-Resolution

Unknown CUDA arch 8.0 or GPU not supported in DCNv2 setup

barakmam opened this issue · 3 comments

barakmam commented

I did the following int your installation guid.
It seems to require cuda arch 8.0 but this torch version (1.4.0) doesn't support that version.

In cpp_extension.py there is not 8.0 valid arch for cuda.

Any help?

Thanks in advance.
image

zychen-ustc commented

Yes. It seems CUDA arch 8.0 is too old. Maybe try a newer version of it.

jakobtroidl commented

I think I solved this problem using a new version of DCNv2. I am using Python 3.8, and Torch 1.11 and CUDA 11.3on Ubuntu. Here's how to reproduce my setup.

Create a VINR conda environment.

conda create -n vinr python=3.8
conda activate vinr
pip install torch==1.11.0+cu113 torchvision==0.12.0+cu113 torchaudio==0.11.0 --extra-index-url https://download.pytorch.org/whl/cu113

Initialize and build the DCNv2 submodule.

git submodule add https://github.com/lucasjinreal/DCNv2_latest ./models/modules/
git submodule init
cd models/modules/DCNv2_latest
python3 setup.py build develop
cd ../..

Run demo

python demo.py --space_scale=4 --time_scale=8 --data_path="YOUR DATA"
  • 👍1