doppeltilde · jonafeucht · Nov 9, 2024 · Nov 9, 2024
diff --git a/.github/workflows/prod.opencl.yml b/.github/workflows/prod.opencl.yml
@@ -0,0 +1,64 @@
+name: OpenCL Production CI/CD
+
+on:
+  push:
+    tags:
+      - "v*"
+    branches:
+      - "main"
+
+env:
+  REGISTRY: "ghcr.io"
+  REGISTRY_USER: ${{ github.actor }}
+  REGISTRY_SECRET: ${{ secrets.GITHUB_TOKEN }}
+  IMAGE_NAME: ${{ github.repository }}
+
+permissions:
+  contents: read
+  packages: write
+
+jobs:
+    build_and_publish:
+        runs-on: ubuntu-latest
+        strategy:
+          matrix:
+            include:
+              - docker_file: Dockerfile.opencl.nvidia
+                label: "latest-opencl-nvidia"
+              - docker_file: Dockerfile.opencl.amd.rusticl
+                label: "latest-opencl-amd-rusticl"
+              - docker_file: Dockerfile.opencl.amd.clover
+                label: "latest-opencl-amd-clover"
+
+        steps:
+            - uses: actions/checkout@v4
+
+            - name: Meta
+              id: meta
+              uses: docker/metadata-action@v5
+              with:
+                images: ${{ env.REGISTRY }}/${{ env.IMAGE_NAME }}
+                tags: |
+                  type=ref,event=branch
+                  type=ref,event=pr
+                  type=ref,event=tag
+                  type=sha,prefix=sha-
+                  type=raw,value=latest,enable=${{ matrix.label == 'latest' }}
+                labels: |
+                  type=raw,value=image-type=${{ matrix.label }}
+
+            - uses: docker/login-action@v3
+              with:
+                registry: ${{ env.REGISTRY }}
+                username: ${{ env.REGISTRY_USER }}
+                password: ${{ env.REGISTRY_SECRET }}
+
+            - name: Build and push
+              uses: docker/build-push-action@v5
+              with:
+                context: .
+                file: ${{ matrix.docker_file }}
+                push: ${{ github.event_name != 'pull_request' }}
+                tags: ${{ steps.meta.outputs.tags }}
+                labels: ${{ steps.meta.outputs.labels }}
+                platforms: linux/amd64
diff --git a/.github/workflows/publish.yml b/.github/workflows/publish.yml
@@ -22,7 +22,7 @@ jobs:
           matrix:
             include:
               - docker_file: Dockerfile
-                label: "latest"
+                label: "latest-cpu"
               - docker_file: Dockerfile.cuda
                 label: "latest-cuda"
         steps:
@@ -41,6 +41,7 @@ jobs:
                   type=ref,event=pr
                   type=ref,event=tag
                   type=sha,prefix=sha-
+                  type=raw,value=latest,enable=${{ matrix.label == 'latest' }}
                 labels: |
                   type=raw,value=image-type=${{ matrix.label }}
 
@@ -56,8 +57,6 @@ jobs:
                 context: .
                 file: ${{ matrix.docker_file }}
                 push: ${{ github.event_name != 'pull_request' }}
-                tags: |
-                  ${{ steps.meta.outputs.tags }}
-                  ${{ env.REGISTRY }}/${{ env.IMAGE_NAME }}:latest${{ matrix.label =='latest-cuda' && '-cuda' || '' }}
-                # labels: ${{ steps.meta.outputs.labels }}
+                tags: ${{ steps.meta.outputs.tags }}
+                labels: ${{ steps.meta.outputs.labels }}
                 platforms: linux/amd64,linux/arm64
diff --git a/Dockerfile.opencl.amd.rusticl b/Dockerfile.opencl.amd.rusticl
@@ -0,0 +1,94 @@
+FROM ubuntu:noble
+
+ARG DEBIAN_FRONTEND=noninteractive
+
+USER root
+
+ARG PYTORCH_OCL_VERSION=0.2.0
+ARG TORCH_VERSION=2.4
+ARG PYTHON_VERSION=cp312
+ARG PLATFORM=linux_x86_64
+
+ENV WHL_FILE=pytorch_ocl-${PYTORCH_OCL_VERSION}+torch${TORCH_VERSION}-${PYTHON_VERSION}-none-${PLATFORM}.whl
+ENV WHL_URL=https://github.com/artyom-beilis/pytorch_dlprim/releases/download/${PYTORCH_OCL_VERSION}/${WHL_FILE}
+
+WORKDIR /app
+COPY . /app
+
+RUN apt-get update && \
+    apt install software-properties-common -y && \
+    add-apt-repository ppa:kisak/kisak-mesa -y && \
+    apt-get update -y && \
+    apt dist-upgrade -y && \
+    apt-get install -y \
+    wget \
+    software-properties-common \
+    apt-transport-https \
+    ca-certificates \
+    curl \
+    gpg && \
+    rm -rf /var/lib/apt/lists/*
+
+# Install OpenCL dependencies
+RUN apt-get update && apt-get install -y \
+    ocl-icd-opencl-dev \
+    opencl-headers \
+    ocl-icd-libopencl1 \
+    clinfo \
+    mesa-opencl-icd \
+    && rm -rf /var/lib/apt/lists/*
+
+# Install Python 3.12
+RUN add-apt-repository ppa:deadsnakes/ppa -y && \
+    apt-get update && \
+    apt-get install -y \
+    python3.12 \
+    python3.12-venv \
+    python3.12-dev \
+    python3-pip \
+    && rm -rf /var/lib/apt/lists/*
+
+# Set up Python virtual environment
+RUN python3.12 -m venv /app/venv
+ENV PATH="/app/venv/bin:$PATH"
+RUN pip install --upgrade pip
+
+# Install AMD GPU drivers
+RUN apt update -y && \
+    wget https://repo.radeon.com/amdgpu-install/6.2.3/ubuntu/noble/amdgpu-install_6.2.60203-1_all.deb && \
+    apt install ./amdgpu-install_6.2.60203-1_all.deb -y && \
+    rm -rf ./amdgpu-install_6.2.60203-1_all.deb 
+
+# Configure OpenCL vendors
+RUN mkdir -p /etc/OpenCL/vendors && \
+    dpkg-divert --divert '/etc/OpenCL/vendors/mesa.icd.disabled' --rename '/etc/OpenCL/vendors/mesa.icd' && \
+    echo "libamdocl64.so" > /etc/OpenCL/vendors/amdocl64.icd && \
+    chmod 755 /etc/OpenCL/vendors
+
+RUN groupadd -f render && groupadd -f video
+RUN usermod -a -G render,video root
+
+RUN mkdir -p /dev/dri && \
+    touch /dev/dri/renderD128 && \
+    chmod 666 /dev/dri/renderD128
+
+RUN chmod 666 /dev/dri/render* || true
+RUN chmod 666 /dev/dri/card* || true
+
+# Set environment variables for OpenCL
+ENV OCL_ICD_VENDORS="/etc/OpenCL/vendors"
+ENV OPENCL_VENDOR_PATH="/etc/OpenCL/vendors"
+
+ENV RUSTICL_ENABLE=radeonsi
+ENV RUSTICL_CL_VERSION=2.1
+ENV RUSTICL_FEATURES=fp16,fp64
+
+COPY requirements.txt /app/requirements.txt
+RUN /app/venv/bin/pip install --no-cache-dir -r requirements.txt --extra-index-url https://download.pytorch.org/whl/cpu
+
+# Install OpenCL backend for PyTorch
+RUN wget ${WHL_URL} && \
+    /app/venv/bin/pip install ${WHL_FILE} && \
+    rm ${WHL_FILE}
+
+CMD ["/app/venv/bin/python", "-m", "fastapi", "run", "main.py", "--proxy-headers", "--host", "0.0.0.0", "--port", "8000"]
diff --git a/README.md b/README.md
@@ -1,4 +1,4 @@
-# Image & Video Classification for NSFW and SFW media.
+# Vision Classification for NSFW and SFW media.
 
 ## Stack:
 - [FastAPI](https://fastapi.tiangolo.com)
@@ -9,15 +9,16 @@
 
 - For ease of use it's recommended to use the provided [docker-compose.yml](https://github.com/doppeltilde/image_video_classification/blob/main/docker-compose.yml).
 
-**CPU Support:** Use the `latest` tag.
+### **CPU Support (AMD64 & ARM64)**
+Use the `latest` tag.
 ```yml
 services:
   image_video_classification:
     image: ghcr.io/doppeltilde/image_video_classification:latest
     ports:
       - "8000:8000"
     volumes:
-      - models:/root/.cache/huggingface/hub:rw
+      - ./models:/root/.cache/huggingface/hub:rw
     environment:
       - DEFAULT_MODEL_NAME
       - BATCH_SIZE
@@ -26,20 +27,46 @@ services:
       - USE_API_KEYS
       - API_KEYS
     restart: unless-stopped
-
-volumes:
-  models:
 ```
 
-**NVIDIA GPU Support:** Use the `latest-cuda` tag.
+### **NVIDIA GPU Support**
+> [!IMPORTANT]
+> Nvidia driver >=560 needs to be installed on the host machine.
+
+**CUDA (AMD64 & ARM64):**
 ```yml
 services:
   image_video_classification_cuda:
     image: ghcr.io/doppeltilde/image_video_classification:latest-cuda
     ports:
       - "8000:8000"
     volumes:
-      - models:/root/.cache/huggingface/hub:rw
+      - ./models:/root/.cache/huggingface/hub:rw
+    environment:
+      - DEFAULT_MODEL_NAME
+      - BATCH_SIZE
+      - ACCESS_TOKEN
+      - DEFAULT_SCORE
+      - USE_API_KEYS
+      - API_KEYS
+    restart: unless-stopped
+    deploy:
+      resources:
+        reservations:
+          devices:
+            - driver: nvidia
+              count: all
+              capabilities: [ gpu ]
+```
+**OpenCL (AMD64):**
+```yml
+services:
+  image_video_classification_opencl:
+    image: ghcr.io/doppeltilde/image_video_classification:latest-opencl-nvidia
+    ports:
+      - "8000:8000"
+    volumes:
+      - ./models:/root/.cache/huggingface/hub:rw
     environment:
       - DEFAULT_MODEL_NAME
       - BATCH_SIZE
@@ -55,11 +82,73 @@ services:
             - driver: nvidia
               count: all
               capabilities: [ gpu ]
+```
+
+### **AMD GPU Support**
+> [!WARNING]
+> Unless AMD starts to officially support older cards again (e.g. gfx8), ROCm will not be available.
 
-volumes:
-  models:
+**OpenCL Rusticl (AMD64):**
+> [!CAUTION]
+> While Rusticl works on Polaris cards, there are significant differences in inference performance, with Polaris cards running Clover showing higher accuracy.
+
+```yml
+services:
+  image_video_classification_cuda:
+    image: ghcr.io/doppeltilde/image_video_classification:latest-opencl-amd-rusticl
+    ports:
+      - "8000:8000"
+    volumes:
+      - ./models:/root/.cache/huggingface/hub:rw
+      - /tmp/.X11-unix:/tmp/.X11-unix
+    environment:
+      - DEFAULT_MODEL_NAME
+      - BATCH_SIZE
+      - ACCESS_TOKEN
+      - DEFAULT_SCORE
+      - USE_API_KEYS
+      - API_KEYS
+    restart: unless-stopped
+    devices:
+      - /dev/kfd
+      - /dev/dri
+    security_opt:
+      - seccomp:unconfined
+    group_add:
+      - "video"
+      - "render"
+```
+**OpenCL Clover (AMD64):**
+> [!CAUTION]
+> Clover is only recommended for legacy cards (e.g. gfx8). If you're using gfx10 or newer, use Rusticl instead.
+```yml
+services:
+  image_video_classification_cuda:
+    image: ghcr.io/doppeltilde/image_video_classification:latest-opencl-amd-clover
+    ports:
+      - "8000:8000"
+    volumes:
+      - ./models:/root/.cache/huggingface/hub:rw
+      - /tmp/.X11-unix:/tmp/.X11-unix
+    environment:
+      - DEFAULT_MODEL_NAME
+      - BATCH_SIZE
+      - ACCESS_TOKEN
+      - DEFAULT_SCORE
+      - USE_API_KEYS
+      - API_KEYS
+    restart: unless-stopped
+    devices:
+      - /dev/kfd
+      - /dev/dri
+    security_opt:
+      - seccomp:unconfined
+    group_add:
+      - "video"
+      - "render"
 ```
 
+## Environment Variables
 - Create a `.env` file and set the preferred values.
 ```sh
 DEFAULT_MODEL_NAME=Falconsai/nsfw_image_detection
@@ -75,6 +164,28 @@ USE_API_KEYS=False
 API_KEYS=abc,123,xyz
 ```
 
+## Tested Devices
+Only tested with consumer grade hardware and only on Linux based systems.
+
+#### CPU
+- AMD FX-6300
+- Intel Core i5-12400F
+- AMD Ryzen 5 1600
+- AMD Ryzen 5 3600
+- AMD Ryzen 9 5950X
+
+#### NVIDIA GPU (CUDA & OpenCL)
+- GTX 950
+- RTX 3060 Ti
+
+#### AMD GPU (OpenCL Clover & Rusticl)
+- RX 570 4GB
+- RX 580 8GB
+- RX 6600 XT
+
+#### Intel GPU (OpenCL NEO)
+- None
+
 ## Models
 Any model designed for image classification and compatible with huggingface transformers should work.
 
@@ -168,6 +279,8 @@ curl -X 'POST' \
 > [!TIP]
 > You can find code examples in the [`examples`](./examples/) folder.
 
+> [!IMPORTANT]  
+> Many thanks to Artyom Beilis and company for providing the awesome [OpenCL backend for PyTorch](https://github.com/artyom-beilis/pytorch_dlprim)!
 ---
 
 _Notice:_ _This project was initally created to be used in-house, as such the