A Docker image based on rocm/pytorch with support for gfx803(Polaris 20-21 (XT/PRO/XL); RX580; RX570; RX560) and Python 3.8

For ROCM 5.4.1

After creating container, you will be logged as root with activated python3.8 environment.

This is a rootless version, which means that your podman/docker needs to be run as rootless and with the files you write within mounted volume as root inside container, on the host will have user ownership and permission.

Supports docker and podman rootless

torch, torchvision and tensorflow-rocm packages are downloaded from https://github.com/xuhuisheng/rocm-gfx803

Every downloaded package should be in /packages.

podman build -t rocm-pytorch-gfx803 ./rocm-pytorch-gfx803-docker

podman run -it --device=/dev/kfd --device=/dev/dri --net=host localhost/rocm-pytorch-gfx803

Not required, but recommended is -v /host:/container parameter. It is easier to work when you have easy access to files already in your system. /host can be any path on host and /container any path on container.