mixa3607/ML-gfx906

ML software (llama.cpp, ComfyUI, Strata, Maya, ...) builds for AMD gfx906 GPUs, e.g. Radeon VII / MI50 / MI60

What it solves

This project provides software support and optimized builds for AMD GFX906 architecture GPUs (such as the MI50/Vega 20), which are officially deprecated by modern ROCm releases. It enables users to continue running machine learning workloads and AI applications on this older hardware.

How it works

It acts as a repository of pre-built packages, container images, and tools specifically compiled for the GFX906 architecture. By adding this repository to an Ubuntu 24.04 system, users gain access to compatible versions of the ROCm stack, PyTorch, llama.cpp, and ComfyUI, bypassing the compatibility issues typically encountered when using newer software on legacy hardware.

Who it’s for

This project is for researchers, developers, and hobbyists who own AMD GFX906-based hardware and want to run modern AI inference and training tasks without upgrading to newer GPU architectures.

Highlights

  • Provides custom ROCm builds and validation tools for legacy hardware.
  • Includes pre-built container images for popular AI frameworks like PyTorch, llama.cpp, and ComfyUI.
  • Offers telemetry tools like the Vega20 Metrics Exporter for monitoring GPU performance.
  • Simplifies installation via standard APT repository integration.

Related

  • Project
    noonghunna/club-3090

    A collection of optimized Docker recipes and benchmarks for running LLMs locally on NVIDIA RTX 3090 and similar consumer GPUs, providing an OpenAI-compatible API for easy integration.

  • Project
    amd/RyzenAI-SW

    A suite of tools and runtime libraries for optimizing and deploying AI inference on AMD Ryzen AI-based PCs to maximize hardware acceleration.

  • Project
    ROCm/AMDMIGraphX

    AMD MIGraphX is a graph inference engine that accelerates machine learning model inference on AMD hardware.

  • Project
    DaoyuanLi2816/can-i-finetune-this

    A tool to estimate the GPU VRAM memory budget required to fine-tune an LLM and generate runnable training recipes to avoid OOM errors.

  • Project
    GPUOpen-Tools/radeon_gpu_profiler

    A low-level GPU optimization tool from AMD that uses hardware thread-tracing to provide detailed timing and visualizations of GPU workloads.