This would address the concern in https://github.com/pop-os/tensorman/issues/21, making Tensorman more secure and easier (not needing to have the docker daemon running and be in the docker group).
Simple test command:
tensorman run --gpu --podman -- python -c "import tensorflow; print(tensorflow.config.list_physical_devices('GPU'))"
Then there is the question of whether this should be the default (Podman is packaged in Ubuntu starting in groovy), and how to make it easy to transition. And probably testing it more thoroughly than my test that it is able to see the GPU.
SicLuceatLux
commented
2 years ago
This would address the concern in https://github.com/pop-os/tensorman/issues/21, making Tensorman more secure and easier (not needing to have the docker daemon running and be in the
dockergroup).Simple test command:
This requires https://github.com/pop-os/nvidia-container-toolkit/pull/4, which is currently blocked by an upstream issue without an obvious solution.
Then there is the question of whether this should be the default (Podman is packaged in Ubuntu starting in groovy), and how to make it easy to transition. And probably testing it more thoroughly than my test that it is able to see the GPU.