Can't `FORCE_CUDA=1 pip install -e .` detectron2 during docker build on a host without a GPU

[guy4261]

Expected results

FORCE_CUDA=1 pip install -e . on the detectron2 repo, when the builder does not have a GPU (but it will be used on an environment with GPU and nvidia-docker2.

Actual results

RuntimeError: cuda runtime error (38) : no CUDA-capable device is detected at /pytorch/aten/src/THC/THCGeneral.cpp:50

If I don't install and run the image without detectron2 on an instance with a GPU, the installation does work and the demo works perfectly.

Detailed steps to reproduce

The relevant part of the Dockerfile:

RUN CUDA_VISIBLE_DEVICES=0 FORCE_CUDA=1 pip install -e /detectron2

System information

• Operating system: Dockerized Ubuntu 18.04
• Compiler version: -
• CUDA version: 10.0
• cuDNN version: -
• NVIDIA driver version: -
• GPU models (for all devices if they are not all the same): none
• PYTHONPATH environment variable: -
• python --version output: 3.6

[guy4261]

It's for detectron2 😅