Binaries for PyTorch 1.9.0

[dimzog]

🚀 Feature

Pre built binaries for new PyTorch version (1.9.0) and CUDA 11.1

Motivation & Examples

I'm having problems working with 3xxx nvidia series and cuda 11.1.
I rolled back to torch 1.7.1 and cuda 10.2, but VRAM caching seems to be a tremendous issue.

Hopefully, new releases fix that.

[ParadoxRobotics]

Hi,
Did you build Detectron 2 from source ?
What kind of issue ?
I personally have have a core dump when I import anything from Detectron 2.

[dimzog]

I did not build from source, as wheels fails to bundle everything. What i suspect is nvcc version of the 3070.
My 1060ti works just fine, so the issue must be with the 3xxx series.

What i end up using in 3070 is
pip3 install torch==1.7.1+cu110 torchvision==0.8.2+cu110 torchaudio===0.7.2 -f https://download.pytorch.org/whl/torch_stable.html

and

python3 -m pip install detectron2 -f https://dl.fbaipublicfiles.com/detectron2/wheels/cu110/torch1.7/index.html

which runs, but VRAM quickly fills up.

Issues I'm having are similar to #2837 and pytorch/pytorch#55027

[dimzog]

I managed to fix the 3xxx issues. Posting incase anyone needs this info.

What i did was follow the tutorial for installing CuDNN 8.1 (link), but

• Instead of CUDA 11.2 Toolkit, search the archives and find CUDA 11.3 that matches your driver version.
  https://developer.nvidia.com/cuda-toolkit-archive

• Instead of CuDNN 8.1, install CuDNN 8.2 (23/4/2021)
  https://developer.nvidia.com/rdp/cudnn-archive

I followed instructions provided in the link, until conda installation. I do not run conda, therefore it is of no use to me.

Next, install PyTorch 1.9.0 with CUDA 11.1.

pip3 install torch==1.9.0+cu111 torchvision==0.10.0+cu111 torchaudio==0.9.0 -f https://download.pytorch.org/whl/torch_stable.html

Then i simply installed detectron2 from source. Everything seems to work like a charm for the time.