Auto-configure AMD GPUs for local AI workloads

ROCmFix is a zero-install Python diagnostic and configuration tool that detects your AMD GPU model and suggests the correct HSA_OVERRIDE_GFX_VERSION environment variable override for ROCm compatibility. The tool can modify shell configurations (CMD, PowerShell, Bash, Zsh, Fish), run diagnostics on drivers and SDKs, and includes a self-updater that syncs a live community GPU database daily. It includes a benchmarking command that tests Vulkan and HIP backends with llama.cpp-based tools.

Use ROCmFix when you need to configure ROCm environment variables for llama.cpp, Ollama, LM Studio, or PyTorch on AMD RDNA4 (RX 9070/9060), RDNA3 (RX 7000), or RDNA2 (RX 6000) GPUs. Run it before installing the HIP SDK on Windows to check your hardware setup, or after installation to verify configuration. Use rocmfix bench to test Vulkan versus HIP performance for your specific card and workload. Do NOT use this tool if you are running NVIDIA GPUs or if your AMD GPU is already working correctly with ROCm without overrides. Skip the HIP installation step if you only need Vulkan support for LLM inference on Windows.

You provide the Python script (single file, no dependencies) and run commands like python rocmfix.py, rocmfix doctor, or rocmfix bench. The tool automatically detects your GPU via system queries. For benchmarking, it temporarily isolates Ollama or LM Studio and runs a 10-second generation test. The tool can modify your shell configuration with override values, generate diagnostic reports, produce benchmark timing results, and create a Markdown system report for troubleshooting posts.

ROCmFix integrates with llama.cpp-based tools (Ollama, LM Studio) for backend benchmarking, PyTorch for ROCm environment configuration, AMD HIP SDK (downloads and launches the official Windows installer via rocmfix install-hip), and shell environments (CMD, PowerShell, Bash, Zsh, Fish) for persistent override application. It fetches GPU data from a live community database and includes opt-in telemetry functionality.

Download the single Python file with one command:

curl -O https://raw.githubusercontent.com/xanpavle/rocmfix/main/rocmfix.py

On Windows PowerShell:

Invoke-WebRequest -Uri "https://raw.githubusercontent.com/xanpavle/rocmfix/main/rocmfix.py" -OutFile "rocmfix.py"

Run the tool:

python rocmfix.py

The default command detects your GPU, suggests the override, and auto-applies it. Use rocmfix doctor for a full diagnostic scan, rocmfix bench to compare Vulkan and HIP performance, rocmfix undo to revert changes, and rocmfix list to see all 17+ supported GPUs. The tool requires Python 3.10+ and works on Windows and Linux with no pip installs or virtual environments.

ROCmFix is built for local AI enthusiasts and developers running LLMs or Stable Diffusion on consumer AMD GPUs who need ROCm environment configuration. It serves Windows users setting up the HIP SDK, Linux users needing quick override configuration, and anyone working with llama.cpp-based tools on AMD hardware. The tool differs from manual override guides by automating detection, providing a live GPU database, and offering backend benchmarking capabilities.