Docker container for onnxruntime-gpu

[shashwat623]

I have an application running on docker container which uses onnxruntime-gpu for inference. I would like to use the onxruntime cuda base docker image for the required cuda, cudnn and onnxruntime-gpu dependencies. I would want to know which version of onnxruntime-gpu does the latest base image use. I ask this because I am using onnxruntime-gpu python package for inference and only the latest release of onnxruntime-gpu(1.8.1) provides inference flexibility for both GPU and CPU, so this is the minimum version I would need. Referring to the ONNX Runtime Docker Image tags HERE, it shows that the latest image uses CUDA10.2, which makes me feel that the onnxruntime-gpu version used would be 1.5-1.6 as mentioned HERE. Can someone please clarify this?

[harshithapv]

You would need to upgrade to CUDA 11.1 to use onnxruntime-gpu 1.8.1. The lastest-cuda image uses an CUDA 10.2 and that's why the version is 1.6. You can try building this dockerfile for onnxruntime-gpu 1.8.1.

[snnn]

Sorry, it should be 11.0 for onnxruntime-gpu 1.8.x.

[snnn]

You can open https://github.com/microsoft/onnxruntime/blob/rel-1.8.1/tools/ci_build/github/azure-pipelines/templates/py-packaging-stage.yml and search "cuda_version"