t0nd3/rocmate

rocmate is a command line tool and compatibility reference for getting AMD graphics cards to work properly with popular AI software. AMD GPUs often provide a lot of video memory for the price, which is attractive for running local AI models, but making them actually work with tools like Ollama or ComfyUI can mean digging through scattered blog posts, GitHub issues, and forum threads that are frequently outdated and specific to one exact chip model. rocmate addresses this by maintaining a curated, community tested database of what configuration works on which AMD chip. After installing it with pip, you can run rocmate doctor to check your system, which detects your GPU, checks whether the ROCm driver stack is installed, verifies permissions, and flags anything missing. The show command displays the exact tested settings and environment variables for a specific AI tool on your detected chip, and the install command applies the correct configuration automatically after showing you a plan and asking for confirmation. The compatibility matrix in the README lists support status for tools including Ollama, ComfyUI, faster-whisper, llama.cpp, Stable Diffusion WebUI, vLLM, Axolotl, and ExLlamaV2, tracked across AMD chip generations labeled gfx1030 through gfx1201, with each combination marked as tested, partial, or unverified. A live version of this matrix is published online and updated on every commit, so the data stays current as more users report their own configurations. The README frames the underlying problem plainly: AMD hardware can be a strong value option for local AI work, but the ecosystem around it has historically required more manual tinkering than the equivalent Nvidia setup, and that tinkering knowledge tends to live in scattered community posts rather than official documentation. rocmate's goal is to turn that scattered, informal knowledge into a single, version controlled, testable source that anyone can query from the command line or contribute back to. The project is written in Python, released under the MIT license, and accepts community contributions of new chip configurations as single YAML files, keeping the barrier to adding support for another AMD chip low.