mobius

Created: 2026-03-18 · Last updated: 2026-04-23

ONNX model definitions for generative AI architectures using the onnxscript nn API.

Build ONNX models directly from HuggingFace model IDs with automatic weight downloading, dtype casting, and multi-component export.

📊 Model Support Dashboard

User Guide

• Getting Started
  • Installation
  • Quick Start
  • CLI Quick Start
  • Output Format
  • Development
• CLI Reference
  • Usage
  • mobius build
  • mobius build-gguf
  • mobius list
  • mobius info
• Module Architecture Guide
  • Module Overview
  • 1. Registering a New Model Architecture
  • 2. Adding a New Decoder Layer Variant
  • 3. Weight Loading Security Policy
  • 4. Import Conventions
  • Quick Reference: Common Tasks
• Model Catalog
  • Decoder-Only LLMs (RoPE)
  • Decoder-Only LLMs (Absolute Position Embeddings)
  • Mixture-of-Experts (MoE)
  • SSM / State-Space Models
  • Hybrid SSM+Attention
  • Vision-Language (Multimodal)
  • Encoder-Only (BERT Family)
  • Encoder-Decoder (Seq2Seq)
  • Speech & Audio
  • Vision
  • Diffusion (via Diffusers)
  • Quantization Support
  • Summary
• Supported Models
  • Text Generation
  • Mixture of Experts
  • Multimodal
  • Speech-to-Text
  • Audio
  • encoder-only
  • encoder
  • encoder-decoder
  • causal-lm
  • vision
  • Depth Estimation
  • Encoder
  • Hybrid SSM+Attention
  • Object Detection
  • Segmentation
  • Text

Execution Providers

• EP-Aware Building: Quick Start
  • 1. Build for a specific EP
  • 2. See what each EP does
  • 3. Query which EPs are available
  • 4. Register a custom EP
  • 5. Common EP configurations at a glance
  • What to read next
• Execution Provider (EP) Aware Building
  • Quick Start
  • Supported Execution Providers
  • Three-Tier Support Strategy
  • The EpCapabilities Dataclass
  • The Optimization Pipeline
  • The Default EP: Portable ONNX
  • Adding a New Execution Provider
  • Adding EP Support to a New Model
  • Debugging: Trace Mode
  • EP Validation
  • Design Decisions
  • When to Use build_context() vs optimize_model() Directly
  • Build-Time EP Queries

Reference

• API Reference
  • Core Functions
  • Core Classes
• Feature Flags
  • Available flags
  • Setting flags via environment variables
  • Setting flags programmatically (internal code only)
  • Using override_flags() in tests
  • Listing all flags
  • Adding new flags

Design

• Design Documents
  • System Architecture
  • Config System Redesign: mobius
  • DeepSeek-V4-Flash export design
  • GGUF Support Proposal for mobius
  • Golden File Infrastructure Design — L4 & L5 Testing
  • Multi-Tier Testing Strategy
  • Performance Evaluation & Regression Testing Infrastructure
  • Phi-4-Multimodal-Instruct Architecture Breakdown
  • Phi4MM Four-Model Split Design
  • Phi4MM ORT-GenAI Integration Spec
  • ONNX GenAI Models - Test Infrastructure Comprehensive Map

Contributing

• Strategy: AI-Assisted Model Support
  • Motivation
  • Architecture recap
  • Phase 1: Classification
  • Phase 2: Analysis (AI-driven)
  • Phase 3: Generation
  • Phase 4: Validation
  • AI Agent Workflow
  • Scaling strategy
  • Appendix: Component reuse matrix

Research

• Testing Strategy Analysis
  • 1. Testing Tiers: Synthetic Parity as a First-Class Layer
  • 2. Parity Metrics: What “Correct” Means
  • 3. Architecture Variants and Code-Path Coverage
  • 4. Dashboard Design and Communication
  • 5. Developer Experience and CI Strategy
  • 6. Infrastructure Concerns
  • 7. Summary of Recommendations