What it does

AMD-SHARK Studio packages a browser-based Stable Diffusion UI and command-line tools that compile PyTorch models through Torch-MLIR into IREE’s runtime. The result: image generation on AMD RDNA cards via Vulkan, NVIDIA via CUDA, and Apple Silicon via Metal—plus CPU fallback. Download an .exe, point your browser to localhost:8080, and go.

The interesting bit

Most Stable Diffusion tooling assumes CUDA. This project instead bets on Vulkan as a least-common-denominator GPU path, with IREE handling the actual kernel compilation and dispatch. The README even warns that Linux MESA/RADV drivers won’t work with FP16—an oddly specific constraint that suggests they actually tested this on real AMD hardware.

Key highlights

• Prebuilt Windows .exe releases; no Python setup required for basic use
• Web UI and CLI both supported (--ui=web or direct python index.py)
• Device targeting via --device=vulkan, cuda, cpu, or metal
• Benchmarking infrastructure for individual dispatch kernels
• Hundreds of models in a “tank” test suite with pytest integration

Caveats

• Not currently maintained. The README opens with a large “NOTE: This project is not currently maintained” and warns that main is a broken refactor toward IREE-Turbine; you must use the AMDSHARK-1.0 branch or an .exe release for working image generation
• First run compiles and downloads ~5GB of models; the README literally asks for “your patience”
• macOS users are pinned to Vulkan SDK 1.3.216; “newer versions will not work”

Verdict

Worth a look if you’re on AMD hardware and frustrated by CUDA-centric SD tooling, or if you’re studying IREE/MLIR deployment patterns. Skip it if you want actively maintained software with modern model support—this is effectively archived code with a rebuild in indefinite limbo.