fix(docker) rocm 6.3 based image

docker/Dockerfile

@@ -31,7 +31,8 @@ RUN --mount=type=cache,target=/var/cache/apt \
     libglx-mesa0 \
     build-essential \
     libopencv-dev \
-    libstdc++-10-dev
+    libstdc++-10-dev \
+    wget
 
 ENV \
     PYTHONUNBUFFERED=1 \
@@ -44,7 +45,6 @@ ENV \
     UV_MANAGED_PYTHON=1 \
     UV_LINK_MODE=copy \
     UV_PROJECT_ENVIRONMENT=/opt/venv \
-    UV_INDEX="https://download.pytorch.org/whl/cu124" \
     INVOKEAI_ROOT=/invokeai \
     INVOKEAI_HOST=0.0.0.0 \
     INVOKEAI_PORT=9090 \
@@ -54,6 +54,10 @@ ENV \
 
 ARG GPU_DRIVER=cuda
 
+ARG CUDA_TORCH="https://download.pytorch.org/whl/cu124"
+ARG CPU_TORCH="https://download.pytorch.org/whl/cpu"
+ARG ROCM_TORCH="https://download.pytorch.org/whl/rocm6.2.4"
+
 # Install `uv` for package management
 COPY --from=ghcr.io/astral-sh/uv:0.6.9 /uv /uvx /bin/
 
@@ -72,23 +76,41 @@ WORKDIR ${INVOKEAI_SRC}
 # x86_64/CUDA is the default
 RUN --mount=type=cache,target=/root/.cache/uv \
     --mount=type=bind,source=pyproject.toml,target=pyproject.toml \
-    --mount=type=bind,source=uv.lock,target=uv.lock \
+    # Cannot use uv sync and uv.lock as that is locked to CUDA version packages, which breaks rocm...
+    # --mount=type=bind,source=uv.lock,target=uv.lock \
     # this is just to get the package manager to recognize that the project exists, without making changes to the docker layer
     --mount=type=bind,source=invokeai/version,target=invokeai/version \
-    if [ "$TARGETPLATFORM" = "linux/arm64" ] || [ "$GPU_DRIVER" = "cpu" ]; then UV_INDEX="https://download.pytorch.org/whl/cpu"; \
-    elif [ "$GPU_DRIVER" = "rocm" ]; then UV_INDEX="https://download.pytorch.org/whl/rocm6.2"; \
+    ulimit -n 30000 && \
+    if [ "$TARGETPLATFORM" = "linux/arm64" ] || [ "$GPU_DRIVER" = "cpu" ]; then export UV_INDEX="$CPU_TORCH"; \
+    elif [ "$GPU_DRIVER" = "rocm" ]; then export UV_INDEX="$ROCM_TORCH"; \
+    else export UV_INDEX="$CUDA_TORCH"; \
     fi && \
-    uv sync --frozen
+    uv venv --python 3.12 && \
+    # Use the public version to install existing known dependencies but using the UV_INDEX, not the hardcoded URLs within the uv.lock
+    uv pip install invokeai
+
+RUN --mount=type=cache,target=/var/cache/apt \
+    --mount=type=cache,target=/var/lib/apt \
+    if [ "$GPU_DRIVER" = "rocm" ]; then \
+    wget -O /tmp/amdgpu-install.deb \
+    https://repo.radeon.com/amdgpu-install/6.2.4/ubuntu/noble/amdgpu-install_6.2.60204-1_all.deb && \
+    apt install -y /tmp/amdgpu-install.deb && \
+    apt update && \
+    amdgpu-install --usecase=rocm -y && \
+    apt-get autoclean && \
+    apt clean && \
+    rm -rf /tmp/* /var/tmp/* && \
+    usermod -a -G render ubuntu && \
+    usermod -a -G video ubuntu && \
+    echo "\\n/opt/rocm/lib\\n/opt/rocm/lib64" >> /etc/ld.so.conf.d/rocm.conf && \
+    ldconfig && \
+    update-alternatives --auto rocm; \
+    fi
 
 # build patchmatch
 RUN cd /usr/lib/$(uname -p)-linux-gnu/pkgconfig/ && ln -sf opencv4.pc opencv.pc
 RUN python -c "from patchmatch import patch_match"
 
-# Link amdgpu.ids for ROCm builds
-# contributed by https://github.com/Rubonnek
-RUN mkdir -p "/opt/amdgpu/share/libdrm" &&\
-    ln -s "/usr/share/libdrm/amdgpu.ids" "/opt/amdgpu/share/libdrm/amdgpu.ids"
-
 RUN mkdir -p ${INVOKEAI_ROOT} && chown -R ${CONTAINER_UID}:${CONTAINER_GID} ${INVOKEAI_ROOT}
 
 COPY docker/docker-entrypoint.sh ./
@@ -105,9 +127,12 @@ COPY invokeai ${INVOKEAI_SRC}/invokeai
 # in a previous layer
 RUN --mount=type=cache,target=/root/.cache/uv \
     --mount=type=bind,source=pyproject.toml,target=pyproject.toml \
-    --mount=type=bind,source=uv.lock,target=uv.lock \
-    if [ "$TARGETPLATFORM" = "linux/arm64" ] || [ "$GPU_DRIVER" = "cpu" ]; then UV_INDEX="https://download.pytorch.org/whl/cpu"; \
-    elif [ "$GPU_DRIVER" = "rocm" ]; then UV_INDEX="https://download.pytorch.org/whl/rocm6.2"; \
+    # Cannot use the uv.lock as that is locked to CUDA version packages, which breaks rocm...
+    # --mount=type=bind,source=uv.lock,target=uv.lock \
+    ulimit -n 30000 && \
+    if [ "$TARGETPLATFORM" = "linux/arm64" ] || [ "$GPU_DRIVER" = "cpu" ]; then export UV_INDEX="$CPU_TORCH"; \
+    elif [ "$GPU_DRIVER" = "rocm" ]; then export UV_INDEX="$ROCM_TORCH"; \
+    else export UV_INDEX="$CUDA_TORCH"; \
     fi && \
     uv pip install -e .
 

docker/run.sh

@@ -13,7 +13,7 @@ run() {
 
   # parse .env file for build args
   build_args=$(awk '$1 ~ /=[^$]/ && $0 !~ /^#/ {print "--build-arg " $0 " "}' .env) &&
-  profile="$(awk -F '=' '/GPU_DRIVER/ {print $2}' .env)"
+  profile="$(awk -F '=' '/GPU_DRIVER=/ {print $2}' .env)"
 
   # default to 'cuda' profile
   [[ -z "$profile" ]] && profile="cuda"
@@ -30,7 +30,7 @@ run() {
 
   printf "%s\n" "starting service $service_name"
   docker compose --profile "$profile" up -d "$service_name"
-  docker compose logs -f
+  docker compose --profile "$profile" logs -f
 }
 
 run