docker-scripts

Docker launch script conventions

When creating or editing docker_run.sh under any model's examples/ directory:

Required patterns

1. AMD Container Toolkit detection — Auto-detect --runtime=amd vs raw device passthrough:

   GPU_ARGS=()
   if docker info 2>/dev/null | grep -qi "amd"; then
       GPU_ARGS=(--runtime=amd -e AMD_VISIBLE_DEVICES=all)
   else
       GPU_ARGS=(--device=/dev/kfd)
       for dev in /dev/dri/renderD*; do
           [[ -e "$dev" ]] && GPU_ARGS+=(--device="$dev")
       done
       GPU_ARGS+=(--group-add video)
   fi
   

2. Existing container check — Before creating a new container, check if one already exists:

   if docker ps -a --format '{{.Names}}' | grep -qx "$CONTAINER_NAME"; then
       echo "Container '$CONTAINER_NAME' already exists. Attach with: docker exec -it $CONTAINER_NAME bash"
       exit 0
   fi
   

3. Auto-clone upstream — If the model depends on an upstream repo, clone it automatically if absent:

   if [[ ! -d "$WORKSPACE_DIR/<repo>" ]]; then
       git clone https://github.com/<org>/<repo>.git "$WORKSPACE_DIR/<repo>"
   fi
   

4. Mode argument — Support at least inference, shell, and stop modes via a positional argument.

5. Env vars: never clobber via -e — Only use docker run -e for variables you are introducing. Append to LD_LIBRARY_PATH, PYTHONPATH, PATH inside the container script, not via -e.

6. Use the canonical ROCm image — rocm/pytorch:rocm7.2.2_ubuntu24.04_py3.12_pytorch_release_2.10.0 unless upstream requires a specific version.