ROCmFix and InferBench – AMD Local-LLM Setup and Vulkan vs. Hip Benchmarking

ROCmFix is a single-file Python utility that auto-detects AMD GPUs and applies the correct HSA_OVERRIDE_GFX_VERSION to fix common local-LLM failures (HSA_STATUS_ERROR_INVALID_ISA, silent CPU fallbacks, etc.). It can automatically modify shell configs (CMD, PowerShell, Bash, Zsh, Fish) and undo changes, run a full diagnostic sweep of drivers, HIP SDK and Vulkan, and generate a Markdown system report for troubleshooting. It also provides a simple installer helper for the AMD HIP SDK on Windows, a self-updating live GPU database, and a set of convenient commands (doctor, bench, install-hip, export, test, undo, update, sync, list). Installation is trivial: download the rocmfix.py file and run it with Python - no venv or pip required.

The tool includes a 10-second benchmark that isolates LM Studio or Ollama and runs a short generation on Vulkan and on HIP/ROCm to tell which backend is faster on a given machine. Guidance distilled from testing: on Windows consumer AMD cards, try Vulkan first (zero config) and use rocmfix bench to verify; for PyTorch or Stable Diffusion workflows install HIP and apply the override ROCmFix suggests. Supported cards include RDNA4/3/2 and integrated GPUs. An opt-in telemetry option shares anonymized GPU, driver, OS and benchmark stats to improve the community database. Licensed MIT.