Backend.AI GO Engine Registry

Official engine package registry for Backend.AI GO.

This registry provides pre-built inference engine packages that Backend.AI GO can automatically download and install.

Available Engines

llama.cpp

High-performance GGUF model inference engine with support for various hardware accelerators.

• Supported model formats: GGUF

Platform	Architecture	Accelerator	Notes
macOS	arm64	Metal	Apple Silicon optimized
Linux	arm64	CPU	ARM64 servers
Linux	x64	CPU	Generic x86_64
Linux	x64	Vulkan	AMD/Intel GPU via Vulkan
Windows	x64	CPU	Generic x86_64
Windows	x64	CUDA 12	NVIDIA GPU (CUDA 12.x)
Windows	x64	CUDA 13	NVIDIA GPU (CUDA 13.x)
Windows	x64	HIP	AMD GPU (ROCm)
Windows	x64	SYCL	Intel GPU (oneAPI)
Windows	x64	Vulkan	Cross-vendor GPU

stable-diffusion.cpp

High-performance Stable Diffusion image generation engine.

• Supported model formats: SafeTensors, GGUF

Platform	Architecture	Accelerator	Notes
macOS	arm64	Metal	Apple Silicon optimized
Linux	x64	CPU	Generic x86_64
Linux	x64	Vulkan	AMD/Intel GPU via Vulkan
Windows	x64	CPU	Generic x86_64
Windows	x64	CUDA 12	NVIDIA GPU (CUDA 12.x)
Windows	x64	Vulkan	Cross-vendor GPU

whisper.cpp

High-performance speech-to-text inference engine.

• Supported model formats: GGUF, bin

Platform	Architecture	Accelerator	Notes
macOS	arm64	Metal	Apple Silicon optimized
Linux	x64	CPU	Generic x86_64
Windows	x64	CPU	Generic x86_64
Windows	x64	CUDA 12	NVIDIA GPU (CUDA 12.x)

MLXcel

An experimental modular model server for Apple Silicon and CUDA GPUs, built with Rust and native MLX C++ bindings.

• Supported model formats: SafeTensors

Platform	Architecture	Accelerator	Notes
macOS	arm64	Metal	Apple Silicon optimized
Linux	arm64	CUDA 13	NVIDIA GB10 (variant: gb10)
Linux	arm64	CUDA 13	NVIDIA GH200 (variant: gh200)

Available Runtimes

Runtime packages provide GPU acceleration libraries required by engine packages.

Runtime	Platform	Description
cuda12-runtime	Windows	NVIDIA CUDA 12 libraries (cuBLAS, cuBLASLt, cuDNN)
cuda13-runtime	Windows	NVIDIA CUDA 13 libraries (cuBLAS, cuBLASLt, cuDNN)
hip-runtime	Windows	AMD HIP runtime libraries (rocBLAS, hipBLASLt)

Usage

Backend.AI GO automatically fetches the package index from this registry and downloads appropriate engine packages based on your system configuration.

Registry URL

https://raw.githubusercontent.com/lablup/backend.ai-go-engine-registry/main/packages.json

Manual Download

Engine packages can also be downloaded directly from GitHub Releases.

Package Format

Engine Packages (.baiengine)

Tar.gz archives containing:

package/
├── manifest.json     # Package metadata
├── <engine-binary>   # Server binary (or .exe on Windows)
└── libs/             # (Optional) Bundled libraries

Runtime Packages (.bairuntime)

Tar.gz archives containing GPU runtime libraries required by engine packages.

Offline Installation

1. Download the .baiengine file from GitHub Releases
2. Place it in the engines directory:
   • macOS: ~/Library/Application Support/Backend.AI GO/engines/incoming/
   • Windows: %APPDATA%\Backend.AI GO\engines\incoming\
   • Linux: ~/.config/Backend.AI GO/engines/incoming/
3. The app will detect and offer to import the package

License

Engine packages are distributed under their respective upstream licenses:

• llama.cpp: MIT License
• stable-diffusion.cpp: MIT License
• whisper.cpp: MIT License
• MLXcel: Lablup Proprietary License