mixa3607/ML-gfx906

What is this?

ML‑gfx906 is a collection of pre‑built packages and Docker images that let you run modern machine‑learning software on the old AMD GFX906 GPU (the “Vega 20/MI50” generation). The repo maintains a small apt‑repository for Ubuntu 24.04 and continuously builds the following pieces for that hardware:

Sub‑project	What it provides	Current state
ROCm	Base ROCm runtime and drivers (deb + container)	Actively built
ROCm Toolkit	Extra ROCm tools (container)	Actively built
ROCm Validation Suite	Test suite for ROCm compliance (deb)	Actively built
ROCm TransferBench	Bandwidth/latency benchmark (deb)	Actively built
AMD Memory Tweak	HBM2 timing adjustment utility (deb)	Actively built
AMD Tuning	General GPU tuning helpers (deb)	Actively built
Vega20 Metrics Exporter	Prometheus exporter for telemetry (deb + container)	Actively built
PyTorch	PyTorch wheel and container compiled for gfx906	Actively built
llama.cpp	ggml‑based LLaMA inference container	Actively built
ComfyUI	Stable‑diffusion UI container (runs on top of the gfx906‑PyTorch image)	Actively built
ROCm Bandwidth Test	Simple bandwidth test (deb)	Paused
vLLM	High‑throughput LLM serving container	Paused
ROCm Tensile	Low‑level GEMM libraries (files only)	Deprecated

The repository also hosts a small apt source (hosted on an S3 bucket) that you can add to an Ubuntu 24.04 system to install the deb packages directly.

---

How to get started

1. Add the apt repository (the README gives a ready‑to‑copy script). It installs the GPG key, creates a gfx906.sources file, and runs apt‑get update.
2. Install the pieces you need, e.g.:

   sudo apt-get install rocm gfx906-llvm rocm-toolkit
   # or install a single framework
   sudo apt-get install pytorch-gfx906
   

3. For container‑based workflows, pull the images that are built on top of an Ubuntu base:

   docker pull mixa3607/rocm-gfx906:latest
   docker pull mixa3607/pytorch-gfx906:latest
   docker pull mixa3607/llama.cpp-gfx906:latest
   

4. Follow the individual sub‑project READMEs (linked in the table) for usage instructions, benchmarks, or integration with Prometheus.

---

Who might use this?

• Researchers or hobbyists who still own legacy Vega 20/MI50 GPUs and want to run current ML workloads without buying new hardware.
• CI/CD pipelines that need reproducible ROC‑compatible environments for older hardware.
• Performance engineers looking to benchmark or tune the GFX906 architecture using the provided validation suite, TransferBench, and memory‑tuning tools.

---

Maintenance & community

• The project is actively building most sub‑projects via GitHub Actions (status badges shown in the table).
• Legacy builds older than the 20260802001858 release are no longer supported; users stuck on those are encouraged to open an issue.
• A Discord channel (gfx906) is linked for real‑time help.
• Documentation lives on a separate wiki site: https://arkprojects.space/wiki/AMD_GFX906.

---

TL;DR

ML‑gfx906 bundles ROCm, PyTorch, llama.cpp, ComfyUI and a handful of GPU‑tuning utilities into deb packages and Docker images that let you run modern AI workloads on the now‑deprecated AMD GFX906 GPU. Installation is as simple as adding an apt source or pulling a Docker image, and the project is kept up‑to‑date via automated builds.