Skip to content

v3.3.2 - New Inference Engine (pre-release)

Pre-release
Pre-release

Choose a tag to compare

View all tags
@saddam213 saddam213 released this 15 Apr 23:11
· 54 commits to master since this release
v3.3.2
a051c6a
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Amuse v3.3.2 - New Inference Engine (pre-release)

This is a development build introducing a complete re-architecture of the inference engine. We are transitioning away from ONNX Runtime as the primary backend to a more modular engine supporting native Safetensors and GGUF.

This build serves as the first technical preview on the roadmap toward Amuse 4.0.

Technical Pivot: From ONNX to Native Weights

The legacy ONNX implementation required significant conversion overhead and lacked immediate compatibility with the latest research. The new engine changes this:

  • Format Support: Added native loaders for .safetensors and .gguf.
  • SOTA Integration: Provides the foundation to run FLUX.2, Z-Image, and LTX-2 without waiting for ONNX-specific optimizations or model conversions.
  • Quantization: Automatic quantization to bfloat16, float8 or NF4 data types, support for GGUF allows for advanced bit-depth control (4-bit, 5-bit, 8-bit, etc.), significantly improving VRAM management for high-parameter models on consumer hardware.

What to Expect in this Dev Build

  • New Backend Implementation: This is a ground-up rebuild. Expect instability as we port features over.
  • Breaking Changes: Model loading logic has been overhauled. Existing ONNX models from your v3.0 install need to be updated to their Safetensors equivalents.
  • Performance: Initial focus is on extensibility and compatibility. Performance tuning for specific AMD/NPU/GPU targets is ongoing.

Installation

This dev build is NOT compatible with previous versions of Amuse or existing ONNX models, full uninstall of Amuse 3.0 is REQUIRED.

1. Installer Version

Best for a standard Windows setup with automatic shortcuts.

  • New Install & Upgrade:
    1. Download and run Amuse_vX.X.X.exe.
    2. Follow the on-screen instructions.

2. Standalone Version

Best for custom drive locations.

New Install:

  1. Download and extract Amuse_vX.X.X.zip to your preferred folder.
  2. Run Amuse.exe.

Note: A fast SSD with plenty of free space is highly recommended, as model downloads can be large.

Device Compatibility

This new engine leverages a python based compute backend. Please note the following hardware requirements for this dev release:

  • NVIDIA: Supports all CUDA compatible devices (CUDA 12.8).
  • AMD: Supports Radeon 7000 series GPUs only (ROCm 7.2.1).
  • Intel: Not Supported.
  • Legacy Hardware: If you are using an incompatible device or older AMD hardware, it is highly recommended to stay on v3.2.0 until broader support is ported to the new engine.

Developers Note

Only install this version if you are interested in testing development builds. As I am a solo developer with very limited testing resources and only a few devices, things may not work, and features may be missing. Each dev build may require re-installs and other manual processes. Please only install if this is okay; otherwise, wait until a stable release is ready.

Full Changelog: v3.2.0...v3.3.2
GitHub Downloads (all assets, specific tag)

Assets 4
Loading