DualModel — Dual-Model Cooperative Inference Framework

A local model + a cloud API model, working together via Gate routing and Interrupter monitoring.

User Input
   │
   ▼
┌───────┐   Route Decision   ┌───────────────┐
│  Gate  │ ────────────────► │  Local Model   │ ◄── Qwen2.5 (0.5B/1.5B/7B)
│ Router │                   └───────┬───────┘
└───────┘                           │
   │  ◄──── Switch Signal ──────     │ token-by-token
   ▼                                ▼
┌─────────────────┐         ┌───────────────┐
│  Interrupter    │ ◄────── │  Cloud Model  │ ◄── DeepSeek / OpenAI API
│  Monitor        │         └───────────────┘
└────────┬────────┘
         │ Quality degradation → triggers Gate re-route
         ▼
      Model switch (with context compression)

---

Quick Start

1. Install

git clone <repo-url>
cd dualmodel

# Recommended
pip install -e .

# Or just core dependencies
pip install -r requirements.txt

2. Set API Key

export DEEPSEEK_API_KEY="sk-your-key"

# Windows PowerShell:
# $env:DEEPSEEK_API_KEY = "sk-your-key"

No API Key? No problem — the system falls back to mock responses in api mode.

3. Start the Web UI

python -m dualmodel --model api

Open http://localhost:8765 in your browser.

4. Start with a Local Model

# 0.5B model (~1GB, runs on CPU)
python -m dualmodel --model 0.5B

# 1.5B model (~3GB)
python -m dualmodel --model 1.5B

# 7B model (~15GB, requires GPU)
python -m dualmodel --model 7B

CLI Options

Argument	Description	Default
--model, -m	Local model size: api / 0.5B / 1.5B / 7B	api
--port, -p	API server port	8765
--host, -H	Listen address	0.0.0.0
--verbose, -v	Enable debug logging	false

---

Model Download

Models are downloaded on demand from ModelScope (fast in China) when you first use a local model size. You can also manage downloads through the web UI:

http://localhost:8765/model-downloads

The download page shows real-time progress and allows you to delete cached models.

---

Running Modes

Mode	Setting	Description	Best for
API Mode	--model api	Zero download, no GPU needed. Uses cloud API or mock responses.	First try, quick demo
Lightweight	--model 0.5B or 1.5B	Model ~1-3GB, runs on CPU/Mac M-series	Daily Q&A
Standard	--model 7B	Requires GPU (~16GB VRAM)	Local inference

---

How It Works

DualModel's core is the Gate + Interrupter dual-component system:

1. Gate Router — Analyzes user input and selects the best model (local or cloud). Supports keyword rules, semantic embedding routing, and trained model routing.

2. Interrupter Monitor — Monitors each generated token's confidence and entropy. Uses a sliding window to detect quality degradation. Performs semantic coherence checks. Triggers model switch when issues are found.

3. Model Switch — Automatically compresses context during switching to prevent token bloat. Includes a cooldown period to prevent thrashing. Falls back to the other model if one fails.

---

Environment Variables

Variable	Description	Default
DEEPSEEK_API_KEY	Cloud API key	""
LOCAL_MODEL_SIZE	Model size: api / 0.5B / 1.5B / 7B	7B
LOCAL_DEVICE	Device: auto / cuda / mps / cpu	auto
LOCAL_MAX_LENGTH	Max generation length (local)	2048
LOCAL_PRELOAD	Preload model on startup	false
LOCAL_WARMUP	Run warmup inference	true
CLOUD_API_BASE	Cloud API base URL	https://api.deepseek.com/v1
CLOUD_API_KEY	Cloud API key (overrides DEEPSEEK_API_KEY)	""
CLOUD_MODEL_NAME	Cloud model name	deepseek-chat
ENABLE_THINKING	Enable thinking/reasoning mode	true
TOOL_CALL_ENABLED	Enable function calling	true
API_HOST	Server listen address	0.0.0.0
API_PORT	Server port	8765

Full config reference: dualmodel/config.py.

---

Project Structure

dualmodel/
├── __init__.py          # Package entry, version 0.2.0
├── __main__.py          # CLI entry: python -m dualmodel
├── api_server.py        # OpenAI-compatible API server + Web UI
├── config.py            # Configuration (env vars, dataclass)
├── gate.py              # Gate router
├── interrupter.py       # Interrupter monitor
├── pipeline.py          # Main inference pipeline
├── local_model.py       # Local model wrapper (HuggingFace / ModelScope)
├── cloud_model.py       # Cloud API wrapper (OpenAI format)
├── utils.py             # Utilities
├── train/               # Training module
│   ├── generate_data.py
│   ├── train_gate.py
│   └── train_interrupter.py
├── trained_models/      # Pre-trained models
│   ├── gate_model.pt
│   └── interrupter_model.pt
├── tray.py              # System tray (macOS/Windows)
├── app.py               # Gradio Web UI (legacy)
└── demo.py              # CLI demo (legacy)

---

Dependencies

Dependency	Use	Install
torch >= 2.1.0	Deep learning framework	Core
transformers >= 4.36.0	HF model loading	Core
openai >= 1.6.0	Cloud API calls	Core
numpy	Scientific computing	Core
fastapi	API server	pip install dualmodel[api]
uvicorn	ASGI server	pip install dualmodel[api]
modelscope	Model download (China)	Core
sentence-transformers	Semantic encoding (training)	pip install dualmodel[train]
scikit-learn	ML tools (training)	pip install dualmodel[train]

# Install all dependencies
pip install -e .[all]

---

Training Custom Models

DualModel supports training custom Gate and Interrupter models:

# 1. Generate training data
python -m dualmodel.train.generate_data

# 2. Train Gate model
python -m dualmodel.train.train_gate

# 3. Train Interrupter model
python -m dualmodel.train.train_interrupter

Trained models are saved to dualmodel/trained_models/ and loaded automatically on startup.

---

License

Apache 2.0

---

Tips

• First time? Run python -m dualmodel --model api — zero downloads, instant start.
• No GPU? Use --model api — everything goes through the cloud API.
• Slow downloads? The system uses ModelScope by default (fast in China). Set HF_ENDPOINT=https://hf-mirror.com if you need HuggingFace mirror.
• Different API provider? Change CLOUD_API_BASE and CLOUD_MODEL_NAME to work with any OpenAI-compatible API (DeepSeek, OpenAI, Qwen, etc.).