This project provides a comprehensive build system for AMD's ROCm machine learning stack and related AI tools, targeting consumer-grade GPUs on Linux. It simplifies the setup of a development environment by compiling ROCm from source, applying necessary patches, and integrating popular ML frameworks and applications like PyTorch, Stable Diffusion, and Llama.cpp, making advanced GPU computing accessible to a wider audience.

How It Works

The system utilizes a bash-scripting framework (babs.sh) to manage the entire build process. It downloads source code for ROCm libraries and AI tools, applies project-specific patches, and compiles them into an isolated directory structure (/opt/rocm_sdk_612). This approach allows for customization and fine-grained control over the build, ensuring compatibility with a wide range of AMD GPUs and Linux distributions.

Quick Start & Requirements

• Install: Clone the repository and run ./install_deps.sh followed by ./babs.sh -i and ./babs.sh -b.
• Prerequisites: Linux distribution (Fedora 40, Ubuntu 24.04/22.04, Arch, etc.), Git, standard build tools. Specific GPU drivers are required for runtime.
• Setup Time: 5-10 hours for a full build.
• Docs: https://github.com/lamikr/rocm_sdk_builder

Highlighted Details

• Supports a broad range of AMD GPUs, including RDNA3, RDNA2, RDNA1, and CDNA architectures, as well as experimental NPU support.
• Integrates key AI tools: llama.cpp, VLLM, and Stable Diffusion WebUI.
• Includes example applications and benchmarks for learning and performance testing.
• Offers Docker images for users who prefer not to build from source.

Maintenance & Community

The project is actively maintained by Mika Laitio. Community testing and contributions are tracked via GitHub issues.

Licensing & Compatibility

Some components are licensed under LGPL 2.1, while others use the COFFEEWARE license. This may impose restrictions on commercial use or linking with closed-source applications.

Limitations & Caveats

The build process is time-intensive and resource-heavy. Experimental support for newer hardware (e.g., XDNA/XDNA2 NPU) requires manual kernel and firmware patching, which may be complex. Older GPUs with limited memory might not run all benchmarks.