Auto-configure AMD GPUs for local AI workloads

ROCmFix is a single-file Python tool for people running local AI on AMD GPUs. Local runtimes such as llama.cpp, Ollama, LM Studio and PyTorch often fail on AMD cards with cryptic errors like HSA_STATUS_ERROR_INVALID_ISA, or silently fall back to the CPU. In many cases the missing piece is the HSA_OVERRIDE_GFX_VERSION environment variable. ROCmFix detects your exact AMD GPU, checks a live community database, applies the matching shell environment override and runs health diagnostics so the setup is ready.

Its main features are:

• Auto-apply and undo: it edits CMD, PowerShell, Bash, Zsh or Fish configuration to set the override, and can safely revert the last change.
• A doctor command that scans the hardware, Adrenalin drivers, HIP SDK and Vulkan API to diagnose problems.
• A 10-second backend race that temporarily isolates Ollama or LM Studio, runs a generation test with Vulkan and then HIP, and reports which is faster on your machine.
• An install-hip command for Windows that downloads and launches the official AMD installer when the HIP SDK is missing.
• A live database and self-updater: new GPUs are fetched from the cloud daily, and the tool updates itself from GitHub.
• A Markdown system report you can paste into Reddit or Discord help posts.

Supported hardware covers RDNA4 (RX 9070 and 9060 series), RDNA3 (RX 7000 series), RDNA2 (RX 6000 series) and integrated graphics, with 17 or more GPUs listed in the database. If your card is missing, running the tool generates a pre-filled GitHub issue link so you can contribute it.

The project's own guidance for consumer AMD cards is to try Vulkan first when running LLMs on Windows and to verify with the bench command. In its comparison, Vulkan needs no configuration and is often faster on consumer RDNA3, while HIP or ROCm needs the HIP SDK plus overrides, is better on Linux and has full PyTorch support. For PyTorch or Stable Diffusion, run install-hip and then apply the override that ROCmFix suggests. It is aimed at AMD RDNA hardware on Windows and Linux.

There is no configuration to write. The tool detects the GPU itself. Commands include the default detect and apply flow, doctor, bench, install-hip, export, update, sync, test (a smoke test to see whether ROCm or HIP detects the override), undo, list, telemetry and install (to register rocmfix as a global command). Outputs are the applied override, diagnostics, benchmark results and an exportable report.

Telemetry is opt-in. If enabled, it sends the GPU model, driver version, OS and benchmark results to a private collector. The project states it never collects usernames, IPs, file paths or AI prompts, and points to PRIVACY.md for details.

It targets Ollama, LM Studio, llama.cpp and PyTorch setups, and works with the AMD HIP SDK and Vulkan. It needs Python 3.10 or later, has no pip or venv step, and is MIT licensed. After downloading rocmfix.py, run it directly:

python rocmfix.py

It is for hobbyists and developers running local LLMs or image models on AMD Radeon cards who want a quick, reversible fix and a measured answer to which backend to use, instead of searching forums for the right override value.