feat: upgrade to ROCm 7.2.2 / PyTorch 2.10.0 with native RDNA4 support

README.md

@@ -2,29 +2,30 @@
 
 🔥 **ComfyUI with AMD ROCm support** - Run ComfyUI on AMD GPUs with optimized ROCm-compatible dependencies.
 
-[![Docker Pulls](https://img.shields.io/docker/pulls/corundex/comfyui-rocm)](https://hub.docker.com/r/corundex/comfyui-rocm) [![ROCm](https://img.shields.io/badge/ROCm-6.4+-green)](https://rocm.docs.amd.com/) [![AMD GPU](https://img.shields.io/badge/AMD-RX%206000%2B-red)](https://www.amd.com/en/products/graphics/desktops/radeon.html)
+[![Docker Pulls](https://img.shields.io/docker/pulls/corundex/comfyui-rocm)](https://hub.docker.com/r/corundex/comfyui-rocm) [![ROCm](https://img.shields.io/badge/ROCm-7.2+-green)](https://rocm.docs.amd.com/) [![AMD GPU](https://img.shields.io/badge/AMD-RX%206000%2B%20%7C%20RDNA4-red)](https://www.amd.com/en/products/graphics/desktops/radeon.html)
 
 ![ComfyUI Interface](Screenshot.png)
 *ComfyUI running on AMD ROCm with sample workflow and generated landscape image*
 
 ## 📋 Version Information
 
-- **Base Image**: `rocm/pytorch:rocm6.4.1_ubuntu24.04_py3.12_pytorch_release_2.6.0`
-- **Python**: 3.12.10
-- **PyTorch**: 2.6.0+git684f6f2
-- **ROCm**: 6.4.43483-a187df25c
-- **ComfyUI**: v0.3.43 (e18f53c, 2025-06-27)
+- **Base Image**: `rocm/pytorch:rocm7.2.2_ubuntu24.04_py3.12_pytorch_release_2.10.0`
+- **Python**: 3.12.x
+- **PyTorch**: 2.10.0
+- **ROCm**: 7.2.2
+- **ComfyUI**: v0.19.3
 
 ## ✨ Key Features
 
 - 🎨 **Node-based AI workflow** - Visual interface for creating complex AI pipelines
-- 🔥 **AMD ROCm optimized** - Native AMD GPU acceleration with ROCm 6.4+
+- 🔥 **AMD ROCm optimized** - Native AMD GPU acceleration with ROCm 7.2+
+- 🆕 **RDNA4 native support** - RX 9060 XT / RX 9070 XT (gfx1200/gfx1201) work without workarounds
+- 🔊 **torchaudio included** - Ships pre-built in the AMD ROCm base image (2.10.0+rocm7.2.2, enables audio/video custom nodes)
 - 📦 **Smart model management** - Automatic downloads with configurable model sets
 - 🧪 **Tested compatibility** - All dependencies verified on real AMD hardware
-- 🎯 **Ready to use** - Pre-configured with sample workflows
+- 🎯 **Ready to use** - Pre-configured with ComfyUI-Manager bundled
 - 💾 **Persistent storage** - Models and outputs preserved across restarts
 
-
 ## 🚀 Quick Start
 
 ```bash
@@ -43,31 +44,34 @@ Access ComfyUI at: **http://localhost:8188**
 
 ## 📋 Requirements
 
-| Component  | Requirement                                |
-| ---------- | ------------------------------------------ |
-| **GPU**    | AMD RX 6000/7000+ series with ROCm support |
-| **VRAM**   | 8GB minimum (16GB+ recommended)            |
-| **OS**     | Linux (Ubuntu 24.04+ recommended)          |
-| **Docker** | Latest version with GPU support            |
-| **ROCm**   | Drivers 6.4+ installed on host             |
+| Component  | Requirement                                              |
+| ---------- | -------------------------------------------------------- |
+| **GPU**    | AMD RX 6000/7000/9000 series (RDNA2/3/4) with ROCm support |
+| **VRAM**   | 8GB minimum (16GB+ recommended)                         |
+| **OS**     | Linux (Ubuntu 24.04+ recommended)                        |
+| **Docker** | Latest version with GPU support                          |
+| **ROCm**   | Drivers 7.2+ installed on host                           |
 
 ## 🔧 Setup Instructions
 
 ### 1. Install ROCm Drivers
+
 ```bash
 # Ubuntu/Debian
 curl -fsSL https://repo.radeon.com/rocm/rocm.gpg.key | sudo gpg --dearmor -o /etc/apt/keyrings/rocm.gpg
-echo "deb [arch=amd64 signed-by=/etc/apt/keyrings/rocm.gpg] https://repo.radeon.com/rocm/apt/6.4 jammy main" | sudo tee /etc/apt/sources.list.d/rocm.list
+echo "deb [arch=amd64 signed-by=/etc/apt/keyrings/rocm.gpg] https://repo.radeon.com/rocm/apt/7.2 noble main" | sudo tee /etc/apt/sources.list.d/rocm.list
 sudo apt update && sudo apt install rocm-dkms
 sudo usermod -a -G render,video $USER
 ```
 
 ### 2. Verify ROCm Installation
+
 ```bash
 rocm-smi  # Should show your AMD GPU(s)
 ```
 
 ### 3. Run ComfyUI
+
 ```bash
 docker run -d \
   --name comfyui-rocm \
@@ -82,15 +86,15 @@ docker run -d \
 
 Control model downloading with the `MODEL_DOWNLOAD` environment variable:
 
-| Mode             | Description                   | Models Included                                       |
-| ---------------- | ----------------------------- | ----------------------------------------------------- |
-| `default`        | Essential starter (4GB)       | SD 1.5                                                |
-| `common`         | Comprehensive set (~30GB)     | SD 1.5, SDXL, ControlNets, upscalers, VAE, embeddings |
-| `realistic`      | Photo-realistic models (~8GB) | Realistic Vision, DreamShaper, VAE                    |
-| `photorealistic` | SDXL realistic (~12GB)        | Juggernaut XL, RealVisXL                              |
-| `artistic`       | Creative/stylized (~2GB)      | Deliberate v2                                         |
-| `all`            | Everything (~100GB)           | All model sets combined                               |
-| `none`           | Skip downloads                | Use existing models only                              |
+| Mode             | Description                    | Models Included                                        |
+| ---------------- | ------------------------------ | ------------------------------------------------------ |
+| `default`        | Essential starter (4GB)        | SD 1.5                                                 |
+| `common`         | Comprehensive set (~30GB)      | SD 1.5, SDXL, ControlNets, upscalers, VAE, embeddings |
+| `realistic`      | Photo-realistic models (~8GB)  | Realistic Vision, DreamShaper, VAE                     |
+| `photorealistic` | SDXL realistic (~12GB)         | Juggernaut XL, RealVisXL                               |
+| `artistic`       | Creative/stylized (~2GB)       | Deliberate v2                                          |
+| `all`            | Everything (~100GB)            | All model sets combined                                |
+| `none`           | Skip downloads                 | Use existing models only                               |
 
 ### Usage Examples
 
@@ -145,43 +149,64 @@ Run with: `docker compose up -d`
 ## ⚡ Performance & Hardware
 
 ### Tested Hardware
-- **AMD Radeon RX 9060 XT** (16GB VRAM) ✅
+
+- **AMD Radeon RX 9060 XT** (16GB VRAM, gfx1200/RDNA4) ✅
+- **AMD Radeon RX 9070 XT** (16GB VRAM, gfx1201/RDNA4) ✅
+- **AMD Radeon RX 7900 XTX** (24GB VRAM, gfx1100/RDNA3) ✅
+
+### RDNA4 Support (RX 9000 Series)
+
+ROCm 7.2.2 adds **native** support for RDNA4 GPUs (gfx1200 = RX 9060 XT, gfx1201 = RX 9070 XT).
+No workarounds or `HSA_OVERRIDE_GFX_VERSION` flags are needed with this image.
+
+If you are running ROCm < 7.x on RDNA4 hardware, you can set the following environment variable
+as a workaround (forces RDNA3 kernel paths):
+
+```bash
+-e HSA_OVERRIDE_GFX_VERSION=11.0.0
+```
 
 ### Performance Metrics
+
 - **Generation Time**: ~30-60s for 512x512 images
-- **VRAM Usage**: 4-8GB for basic operations  
+- **VRAM Usage**: 4-8GB for basic operations
 - **Model Loading**: ~30-60s first time, cached afterward
 - **Batch Processing**: Multiple images supported
 
 ### Tips
+
 - Mount persistent volumes to avoid re-downloading models
 - Start with `default` models, upgrade to larger sets as needed
 - Use fast SSD storage for optimal performance
 
 ## 🔍 Troubleshooting
 
-| Issue                     | Solution                                                                                                           |
-| ------------------------- | ------------------------------------------------------------------------------------------------------------------ |
-| **Container won't start** | Check ROCm drivers: `rocm-smi`                                                                                     |
-| **No GPU detected**       | Verify container GPU access: `docker exec comfyui-rocm python -c "import torch; print(torch.cuda.is_available())"` |
-| **Model download fails**  | Check internet connection, disk space, and logs                                                                    |
-| **Out of memory**         | Reduce batch size, use smaller models, ensure 8GB+ VRAM                                                            |
-| **Models not found**      | Verify downloads completed and file permissions                                                                    |
+| Issue                          | Solution                                                                                                           |
+| ------------------------------ | ------------------------------------------------------------------------------------------------------------------ |
+| **Container won't start**      | Check ROCm drivers: `rocm-smi`                                                                                     |
+| **No GPU detected**            | Verify container GPU access: `docker exec comfyui-rocm python -c "import torch; print(torch.cuda.is_available())"` |
+| **Model download fails**       | Check internet connection, disk space, and logs                                                                    |
+| **Out of memory**              | Reduce batch size, use smaller models, ensure 8GB+ VRAM                                                            |
+| **Models not found**           | Verify downloads completed and file permissions                                                                    |
+| **`torchaudio` not found** | The base image ships torchaudio pre-built — if missing, verify your base image tag is correct |
 
 ## 📄 License & Credits
 
 This project is licensed under GPL-3.0. See the [LICENSE](LICENSE) file for details.
 
 ### Third-Party Components
+
 - **ComfyUI**: GPL-3.0 - [ComfyUI](https://github.com/comfyanonymous/ComfyUI)
 - **PyTorch**: BSD 3-Clause - [PyTorch](https://pytorch.org/)
 - **ROCm**: Various OSS licenses - [AMD ROCm](https://rocm.docs.amd.com/)
 
 **Acknowledgments:**
+
 - [ComfyUI](https://github.com/comfyanonymous/ComfyUI) - Node-based AI workflow interface
-- [AMD ROCm](https://rocm.docs.amd.com/) - Open source GPU computing platform  
+- [AMD ROCm](https://rocm.docs.amd.com/) - Open source GPU computing platform
+- [Griznah (Ole-Magnus Sæther)](https://github.com/Griznah) - ROCm 7.x update groundwork (PR #5)
 - ROCm community for AMD GPU AI support
 
 ---
 
-🔗 **Links:** [Docker Hub](https://hub.docker.com/r/corundex/comfyui-rocm) | [GitHub](https://github.com/corundex/comfyui-rocm) | [ComfyUI](https://github.com/comfyanonymous/ComfyUI)
+🔗 **Links:** [Docker Hub](https://hub.docker.com/r/corundex/comfyui-rocm) | [GitHub](https://github.com/corundex/comfyui-rocm) | [ComfyUI](https://github.com/comfyanonymous/ComfyUI)

build.sh

@@ -6,12 +6,12 @@ set -e
 # Configuration
 IMAGE_NAME="comfyui-rocm"
 VERSION="${1:-latest}"
-DETAILED_TAG="comfyui_0.3.43__rocm6.4.1_ubuntu24.04_py3.12_pytorch_2.6.0"
+DETAILED_TAG="comfyui_0.19.3__rocm7.2.2_ubuntu24.04_py3.12_pytorch_2.10.0"
 
 echo "🐳 Building ComfyUI ROCm Docker image..."
 echo "📦 Tags: ${IMAGE_NAME}:${VERSION}, ${IMAGE_NAME}:${DETAILED_TAG}"
 echo ""
 
-# Build the image with both tags
-echo "🔨 Building Docker image..."
-docker build -f docker/Dockerfile -t "${IMAGE_NAME}:${VERSION}" -t "${IMAGE_NAME}:${DETAILED_TAG}" --progress=plain .
+# Build the image with both tags using BuildKit (buildx)
+echo "🔨 Building Docker image with BuildKit..."
+docker buildx build -f docker/Dockerfile -t "${IMAGE_NAME}:${VERSION}" -t "${IMAGE_NAME}:${DETAILED_TAG}" --progress=plain --load .

docker/Dockerfile

@@ -1,7 +1,7 @@
 # ROCm ComfyUI Dockerfile
 # Based on AMD's official ROCm PyTorch container
 
-FROM rocm/pytorch:rocm6.4.1_ubuntu24.04_py3.12_pytorch_release_2.6.0
+FROM rocm/pytorch:rocm7.2.2_ubuntu24.04_py3.12_pytorch_release_2.10.0
 
 # Set working directory
 WORKDIR /workspace
@@ -11,33 +11,49 @@ RUN apt-get update && apt-get install -y \
   git \
   wget \
   curl \
+  ffmpeg \
+  libgl1 \
+  libglib2.0-0 \
   && apt-get clean \
   && rm -rf /var/lib/apt/lists/*
 
-# Clone ComfyUI
-RUN git clone https://github.com/comfyanonymous/ComfyUI.git
+# Clone ComfyUI and checkout pinned version
+RUN git clone https://github.com/comfyanonymous/ComfyUI.git \
+  && cd ComfyUI \
+  && git checkout v0.19.3
 
 # Set ComfyUI as working directory
 WORKDIR /workspace/ComfyUI
 
-# Create necessary directories with separate commands
-RUN mkdir -p /workspace/ComfyUI/models/checkpoints && \
-  mkdir -p /workspace/ComfyUI/models/vae && \
-  mkdir -p /workspace/ComfyUI/models/loras && \
-  mkdir -p /workspace/ComfyUI/models/embeddings && \
-  mkdir -p /workspace/ComfyUI/models/upscale_models && \
-  mkdir -p /workspace/ComfyUI/models/controlnet && \
-  mkdir -p /workspace/ComfyUI/output && \
-  mkdir -p /workspace/ComfyUI/input && \
-  mkdir -p /workspace/ComfyUI/custom_nodes
+# Create necessary directories
+RUN mkdir -p \
+  /workspace/ComfyUI/models/checkpoints \
+  /workspace/ComfyUI/models/vae \
+  /workspace/ComfyUI/models/loras \
+  /workspace/ComfyUI/models/embeddings \
+  /workspace/ComfyUI/models/upscale_models \
+  /workspace/ComfyUI/models/controlnet \
+  /workspace/ComfyUI/output \
+  /workspace/ComfyUI/input \
+  /workspace/ComfyUI/custom_nodes
+
+# Install onnxruntime_migraphx — ROCm 7.x replacement for onnxruntime_rocm
+# AMD renamed the package starting with ROCm 7.x; provides the same 'onnxruntime' Python namespace
+RUN pip install --no-cache-dir \
+  "https://repo.radeon.com/rocm/manylinux/rocm-rel-7.2.2/onnxruntime_migraphx-1.23.2-cp312-cp312-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl"
 
 # Copy ROCm-tested requirements files
 COPY docker/requirements_rocm.txt /workspace/
 
-# Install Python dependencies using ROCm-compatible requirements
-# These files have been tested on real AMD hardware and filter out packages that break ROCm
+# Install Python dependencies using ROCm-compatible requirements.
+# Note: torchaudio (2.10.0+rocm7.2.2) ships pre-built in the base image — pip
+# satisfies the 'torchaudio' requirement from requirements.txt immediately.
 RUN pip install --no-cache-dir -r /workspace/requirements_rocm.txt
 
+# Install ComfyUI-Manager
+RUN git clone --depth 1 https://github.com/Comfy-Org/ComfyUI-Manager.git custom_nodes/ComfyUI-Manager \
+  && pip install --no-cache-dir -r custom_nodes/ComfyUI-Manager/requirements.txt
+
 # Copy startup script, models config, and sample workflow
 COPY docker/startup.sh /workspace/startup.sh
 COPY docker/download_models.py /workspace/download_models.py
@@ -47,14 +63,26 @@ COPY docker/sample_workflow.json /workspace/ComfyUI/
 # Make startup script executable
 RUN chmod +x /workspace/startup.sh
 
+# Verify ROCm stack is intact after all installs
+RUN python - <<'PY'
+import torch, torchvision
+print(f"torch={torch.__version__} hip_available={torch.cuda.is_available()}")
+print(f"torchvision={torchvision.__version__}")
+try:
+    import torchaudio
+    print(f"torchaudio={torchaudio.__version__}")
+except Exception as e:
+    print(f"torchaudio=unavailable ({e})")
+PY
+
 # Environment variables for model download behavior
 # MODEL_DOWNLOAD options:
 #   "default"   - Download basic SD 1.5 model (default)
 #   "all"       - Download full model set (SD 1.5, SDXL, ControlNet, etc.)
-#   "realistic" - Download realistic photo models  
+#   "realistic" - Download realistic photo models
 #   "none"      - Skip all downloads, use existing models only
 #   Custom      - Any section name from models.conf
-ENV MODEL_DOWNLOAD=default
+ENV MODEL_DOWNLOAD=none
 
 # Expose ComfyUI port
 EXPOSE 8188
@@ -65,4 +93,3 @@ HEALTHCHECK --interval=30s --timeout=10s --start-period=5s --retries=3 \
 
 # Use smart startup script that handles model downloads
 CMD ["/workspace/startup.sh"]
-