ROCmFix: One Python File That Fixes AMD GPU Issues for Local AI
Running local AI models on AMD GPUs has a well-known pain point. Tools like Ollama, LM Studio, llama.cpp, and PyTorch frequently throw HSA_STATUS_ERROR_INVALID_ISA errors or silently fall back to CPU when the HSA_OVERRIDE_GFX_VERSION environment variable is not set correctly. A new open-source tool called ROCmFix eliminates that problem with a single Python file and no installation required.
The tool detects the exact AMD GPU in the system, queries a live community database for the correct override value, and applies it automatically. No pip installs, no virtual environments, no configuration files. Download rocmfix.py and run it.
17+ GPUs, Live Database, Automatic Sync
ROCmFix currently supports RDNA4 (RX 9070 and 9060 series), RDNA3 (RX 7000 series), RDNA2 (RX 6000 series), and AMD integrated graphics. The GPU database updates daily from a cloud endpoint, so newly released cards appear without requiring an app update.
If a GPU is missing, the tool generates a pre-filled GitHub issue link so users can contribute their card's profile to the database. The rocmfix list command shows all supported GPUs at any time.
Doctor, Bench, and One-Click HIP Install
The rocmfix doctor command scans the full stack: hardware detection, Adrenalin driver status, HIP SDK presence, and Vulkan API availability. It reports what is installed, what is missing, and what needs fixing.
For users who want to compare backends, rocmfix bench runs a 10-second race. It temporarily isolates Ollama or LM Studio, runs a generation test through Vulkan, then through HIP, and reports which backend is faster on the specific hardware. This settles the common question of whether to use Vulkan or ROCm on consumer cards without manual benchmarking.
On Windows, rocmfix install-hip downloads and launches the official AMD HIP SDK installer, removing the need to hunt for the correct download page.
Vulkan vs. HIP: When Each Wins
The tool's documentation includes a practical comparison. Vulkan requires zero configuration and works out of the box on Windows. On Linux, HIP and ROCm generally perform better. For consumer RDNA3 cards, Vulkan is often faster than HIP in local LLM inference. PyTorch and Stable Diffusion workflows require HIP, since Vulkan support in those frameworks is limited.
The recommendation is straightforward: running LLMs on Windows means trying Vulkan first and using rocmfix bench to confirm. PyTorch or Stable Diffusion workloads need the HIP SDK installed and the override applied.
Export, Telemetry, and Community
rocmfix export generates a Markdown system report that users can paste into Reddit or Discord when asking for help. The telemetry system is opt-in and sends only GPU model, driver version, OS, and benchmark results to a private collector. Usernames, IPs, file paths, and AI prompts are never collected.
The tool can be installed as a global command with rocmfix install, updated from GitHub with rocmfix update, and its GPU database force-synced with rocmfix sync. The rocmfix undo command safely reverts the last auto-applied override.
The project is MIT licensed and available on GitHub at xanpavle/rocmfix. One file, no dependencies, and a one-command fix for the most common AMD GPU problem in local AI development.