Nvidia-Docker for GPU support when upload?

[kefhuang]

Hello, I have a question regarding the evaluation platform, specifically related to GPU support. According to the website, it mentions that the platform has 12GB GPU support. However, I noticed that the Docker template provided is based on a python:3.10-slim image, rather than an nvidia/cuda base image.

In our local setup, we are utilizing an nvidia/cudagl base image, which requires nvidia-docker to be installed along with the --gpus all parameter when running the Docker container. These configurations, however, don't seem to be present in the current testing script.

Could you please confirm if using nvidia-docker is supported on your evaluation platform? If so, could you recommend the best CUDA version to use? For our local testing environment, we have found that the nvidia/cudagl:11.4.2-devel-ubuntu20.04 image works well, supporting both PyTorch and GPU-accelerated graphics rendering.

[YoruCathy]

You can submit the code using nvidia/cudagl:11.4.2-devel-ubuntu20.04 and contain the RCareWorld code in it. I think you can go ahead and do a test submission with some test code using that environment and see if we are able to run it on our end.