MELLM

KO | EN

gitlite — search

#hacktoberfest #llm #go #android #machine-learning #rust #deep-learning #nodejs #java #windows #awesome #cli #vue #security

(★ 33)

Lightweight Modular AI Routing Engine for Local LLMs — Run specialised experts efficiently on consumer GPUs using smart Mixture-of-Experts routing.

MELLM 최신버젼 다운로드

최종 버전 다운로드 (.zip)

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

<p align="center">
  <h1 align="center">🧠 MELLM — Lightweight Modular AI Routing Engine</h1>
  <p align="center">
    Run multiple small, domain-specialized LLMs instead of one massive general model.<br/>
    Better answers. Less VRAM. Faster responses.
  </p>
  <p align="center">
    <img src="https://img.shields.io/badge/python-3.10%2B-blue" alt="Python 3.10+"/>
    <img src="https://img.shields.io/badge/backend-llama--cpp--python-green" alt="llama-cpp-python"/>
    <img src="https://img.shields.io/badge/format-GGUF-orange" alt="GGUF"/>
    <img src="https://img.shields.io/badge/GPU-NVIDIA_CUDA-76B900" alt="CUDA"/>
    <img src="https://img.shields.io/badge/license-MIT-lightgrey" alt="MIT License"/>
    <img src="https://img.shields.io/github/stars/Rahul-14507/MELLM" alt="GitHub Stars"/>
    <img src="https://img.shields.io/github/issues/Rahul-14507/MELLM" alt="GitHub Issues"/>
    <img src="https://img.shields.io/github/last-commit/Rahul-14507/MELLM" alt="Last Commit"/>
  </p>
  <img src="mellm-demo.gif" alt="MELLM Demo" width="100%"/>
</p>

## ⚡ TL;DR

MELLM is a lightweight AI router that:
- **Uses a small model** to classify queries with high precision.
- **Routes them to domain-specific experts** (code, math, medical, legal).
- **Conserves VRAM** by keeping only one expert model active at a time.

👉 **Better accuracy** than small general models  
👉 **Runs on 6GB GPUs** (RTX 3050, GTX 1060)  
👉 **No more waiting** for massive 70B models to layer-stream

---

## 🔍 What is MELLM?

Large general-purpose models are trained to know everything — which means they're not optimised for anything in particular. A 70B general model answering a medical question is overkill on resources and often outperformed by a 7B model that was fine-tuned specifically on medical literature.

Meanwhile, running models larger than 14B is completely out of reach on consumer hardware like a 6GB GPU. You're stuck choosing between a small general model that gives mediocre answers, or a large specialist model you can't run.

**MELLM solves this differently.**

Instead of running one large model, MELLM runs a tiny **router model** that reads your query, identifies the domain, and loads the right **specialist model** — a small model fine-tuned specifically for that task. When you ask a medical question, you get a medical model. When you ask for code, you get a code model. When your query spans multiple domains, MELLM decomposes it and runs each part through the appropriate specialist simultaneously.

The result: **expert-level answers from models that fit on your GPU**, at a fraction of the compute cost of running a monolithic large model.
```
User query → Router (1.5B) → identifies domain → loads specialist (1.5–7B)
                                                 → generates response
                                                 → unloads, ready for next query
```

This architecture isn't just a consumer hardware workaround — it's a fundamentally more efficient approach. The same principle applies at production scale: routing to purpose-built specialists is cheaper and more accurate than throwing a 70B general model at every query.

### The numbers

| Approach | Model needed | VRAM | Query time | Domain accuracy |
|----------|-------------|------|------------|-----------------|
| General monolithic | 14B–70B | 16–80 GB | minutes | moderate |
| **MELLM specialist routing** | **1.5B–7B** | **6 GB** | **8–20s** | **high** |

A domain-specific 7B model can often outperform a larger general model on tasks within its specialty — while using significantly less compute.

## 🧠 Traditional vs MELLM

## 💡 Why this matters

Running large LLMs locally is expensive and often impractical. MELLM shows that:
- **You don’t need a single massive model** to get expert answers.
- **You can orchestrate experts** to achieve better results with less hardware.
- **Efficient routing** beats brute-force scaling every time.

This opens the door to local AI systems on consumer hardware, modular architectures that grow with your needs, and lower-cost deployment at scale.

## 🎯 MELLM vs Mixture-of-Experts — What's Different?

You may have heard of **Mixture-of-Experts (MoE)** — the architecture behind models like Mixtral and GPT-4. The concept is similar: route inputs to specialized sub-networks. But production MoE has critical limitations that MELLM directly addresses.

### 🧩 Think: "App Store for LLM Experts"
MELLM's open architecture lets anyone contribute a new specialist:
- Add a **finance model** for stock analysis
- Add a **science model** for research queries
- Add a **language model** for translation tasks

No retraining, no new architecture — just register it in `GGUF_REGISTRY` and it's immediately available as a domain expert.

### 🔧 LLM-as-Router — A Cutting-Edge Twist
Most open-source routers use keyword matching or fine-tuned classifiers. MELLM uses an LLM in JSON-output mode to both **classify the domain** and **rewrite the prompt** for optimal specialist performance. This "LLM as router" approach is an active research direction — MELLM brings it to a practical, runnable implementation.

> **Honest positioning:** MELLM is not a new concept. But *open, modular, low-hardware MoE routing* is genuinely underserved in the open-source ecosystem. That's the gap this fills.

### Key Features

- 🔀 **Intelligent Routing** — A 1.5B router classifies queries into 5 domains with >90% accuracy and rewrites prompts for optimal specialist output
- 🧩 **Multi-Agent Composition** — Cross-domain queries are automatically decomposed, each part routed to the right specialist, and merged into one response
- 🧬 **Domain Specialists** — Fine-tuned models for medical, legal, math, code, and general knowledge — each optimised for its task
- ⚡ **Persistent Router** — Router stays resident in VRAM; zero routing overhead after startup
- 🔥 **Hot Specialist Cache** — Active specialist stays loaded between same-domain queries; only swapped on domain switch
- 🧠 **Conversation Context** — 3-turn history window so follow-up queries like "Now in Python?" work correctly
- 🎯 **Domain Continuity** — Short follow-ups inherit the current domain automatically
- 🖥️ **Interactive Setup Wizard** — Hardware-aware onboarding detects your GPU and recommends appropriate model sizes
- 🌐 **REST API** — FastAPI endpoint so any app can use MELLM as a backend
- ⬇️ **Auto-Download** — Models download from Hugging Face on first use, cached locally

---

## 🏗️ Architecture

```
                      ┌──────────────────────────────────────┐
                      │   STARTUP — once per session         │
                      │   Router (1.5B Qwen) loads into VRAM │
                      └──────────────────┬───────────────────┘
                                         │ stays resident ↓
┌──────────────┐    ┌─────────────┐    ┌─────────────────────┐
│  User Query  │───▶│  Add context│───▶│  Router: Classify   │
│              │    │  (last 3    │    │  domain + rewrite   │
│              │    │   turns)    │    │  prompt (JSON mode) │
└──────────────┘    └─────────────┘    └──────────┬──────────┘
                                                   │
                               domain continuity   │
                               bias applied here   │
                                                   ▼
              ┌────────────────────────────────────────────────┐
              │  Hot Cache Check                               │
              │  Same domain? → reuse loaded specialist (0s)   │
              │  New domain?  → unload old, load new (1-6s)    │
              └───────────────────┬────────────────────────────┘
                                   │
          ┌────────────────────────▼──────────────────────────┐
          │  Specialist generates response (stays in VRAM)    │
          │  History appended → available for next query      │
          └───────────────────────────────────────────────────┘
```

### VRAM Strategy

MELLM uses a **two-tier residency** model to maximize responsiveness within 6GB VRAM:

| Model | Residency | Why |
|-------|-----------|-----|
| **Router** (1.5B) | Always in VRAM | ~1 GB, permanent resident; no per-query overhead |
| **Active Specialist** | In VRAM until domain changes | Only one specialist at a time; swapped on domain switch |
| **Other Specialists** | On disk (GGUF cache) | Loaded on-demand in 1-6s |

---

## 📋 Model Registry

All models use the **GGUF** quantized format for efficient inference via `llama-cpp-python`.

| Role | Domain | Model | GGUF File | Size | Context |
|------|--------|-------|-----------|------|---------|
| **Router** | All | Qwen2.5-**1.5B**-Instruct | `qwen2.5-1.5b-instruct-q4_k_m.gguf` | ~1 GB | 4096 |
| Specialist | **Code** | Qwen2.5-Coder-1.5B-Instruct | `qwen2.5-coder-1.5b-instruct-q4_k_m.gguf` | ~1.1 GB | 4096 |
| Specialist | **Math** | Qwen2.5-Math-1.5B-Instruct | `Qwen2.5-Math-1.5B-Instruct-Q4_K_M.gguf` | ~986 MB | 4096 |
| Specialist | **Medical** | BioMistral-7B-DARE | `ggml-model-Q2_K.gguf` | ~2.3 GB | 1024 |
| Specialist | **Legal** | Magistrate-3.2-3B-IT | `magistrate-3.2-3b-it.Q4_K_M.gguf` | ~1.8 GB | 1024 |
| Specialist | **General** | Qwen2.5-1.5B-Instruct | `qwen2.5-1.5b-instruct-q4_k_m.gguf` | ~1.1 GB | 4096 |

> **Note:** Larger models (7B+) automatically use a reduced context window (1024 tokens) to stay within 6GB VRAM limits. Smaller models (≤1.5B) use the full 4096 context.

---

## 📁 Project Structure

```
MELLM/
├── cli.py                    # Rich terminal interface with session efficiency UI
├── api.py                    # FastAPI REST server
├── orchestrator.py           # Core pipeline: persistent router, hot cache, conversation history
├── config.yaml               # Model IDs, token limits, and specialist configuration
├── requirements.txt          # Python dependencies
├── .env                      # Environment variables (HF_TOKEN)
│
├── loader/
│   ├── __init__.py
│   └── airllm_loader.py      # GGUF model loader with auto-download and VRAM management
│
├── router/
│   ├── __init__.py
│   ├── classifier.py         # LLM-based query classifier (JSON output mode)
│   └── prompt_optimizer.py   # Rule-based fallback prompt templates
│
├── webui/
│   └── index.html            # Single-file web UI (served by FastAPI at /)
│
└── specialists/
    ├── __init__.py
    ├── base_specialist.py    # Abstract base class for all specialists
    ├── code.py               # Code generation specialist (temp=0.1)
    ├── math_specialist.py    # Math problem solver (temp=0.1)
    ├── medical.py            # Medical Q&A with system prompt (temp=0.7)
    ├── legal.py              # Legal information specialist
    └── general.py            # General knowledge fallback
```

---

## ⚙️ Hardware Requirements

| Component | Minimum | Recommended |
|-----------|---------|-------------|
| **GPU** | NVIDIA GTX 1060 (6GB) | NVIDIA RTX 3050+ (6GB+) |
| **VRAM** | 6 GB | 8 GB+ |
| **RAM** | 8 GB | 16 GB+ |
| **Storage** | 10 GB free | 20 GB+ free |
| **CUDA** | 11.7+ | 12.1+ |
| **Python** | 3.10+ | 3.11+ |
| **OS** | Linux (Ubuntu 22.04+) | Linux with NVIDIA drivers |

### Hardware Scalability

MELLM works with any GGUF model on HuggingFace. The setup wizard automatically recommends model sizes based on your GPU. Users with larger GPUs can run 13B, 34B, or 70B specialists by simply updating `GGUF_REGISTRY` and `config.yaml` — no code changes needed.

---

## 🚀 Getting Started

### Linux

#### 1. Clone the Repository

```bash
git clone https://github.com/Rahul-14507/MELLM
cd MELLM
```

#### 2. Create a Virtual Environment

```bash
python -m venv .venv
source .venv/bin/activate
```

#### 3. Install llama-cpp-python (with CUDA) + Remaining Dependencies

Use the provided setup script — it auto-detects your CUDA version, installs the matching pre-built wheel, and falls back to a source build if the wheel is incompatible with your toolkit:

```bash
chmod +x setup.sh && ./setup.sh
```

> **Why not just `pip install llama-cpp-python`?**
> Pre-built CUDA wheels are compiled for a specific CUDA minor version (e.g. `cu121`). Installing the wrong wheel causes a silent `SIGILL` crash the first time a model is loaded — not at import time. The setup script detects your toolkit version, picks the nearest known wheel, verifies it actually loads without crashing, and falls back to a source build if needed.

<details>
<summary>Manual install (advanced)</summary>

If you prefer to install manually, match the wheel tag to your CUDA version (`nvcc --version`):

```bash
# CUDA 12.1
pip install llama-cpp-python --extra-index-url https://abetlen.github.io/llama-cpp-python/whl/cu121

# CUDA 11.8
pip install llama-cpp-python --extra-index-url https://abetlen.github.io/llama-cpp-python/whl/cu118

# Any CUDA version — build from source (slower but always correct)
CMAKE_ARGS="-DGGML_CUDA=on" pip install llama-cpp-python --no-cache-dir
```

Verify the wheel actually works (import alone is not sufficient — CUDA code only runs on model load):

```bash
python -c "import llama_cpp; print('llama-cpp-python', llama_cpp.__version__, 'OK')"
```

Then install the rest of the dependencies:

```bash
pip install -r requirements.txt
```
</details>

#### 5. Set Up Environment Variables

```bash
cp .env.example .env
# Or create .env manually:
echo "HF_TOKEN=your_huggingface_token_here" > .env
```

You need a [Hugging Face token](https://huggingface.co/settings/tokens) to download models. Some models may require accepting their license on the Hugging Face model page first.

#### 6. Run MELLM

```bash
# Interactive CLI mode
python cli.py

# Pre-download all models at once
python cli.py --preload all

# Pre-download a specific domain
python cli.py --preload medical

# Web UI Mode
python cli.py --web
# Opens http://localhost:8000 in your browser automatically
```

**Interactive Web Features:**
- **Live Streaming**: Tokens appear instantly with polished animations.
- **Expert Panels**: Multi-agent results stream into dedicated parallel panels.
- **Domain Indicators**: Visual badges identify which specialist is responding.
- **Stop Generation**: Interrupt long responses instantly with the "Stop" button.
- **One-Click Regenerate**: Easily retry any query to get a fresh response.
- **Robust Code Rendering**: Proper syntax highlighting with headers and one-click "Copy" buttons.
- **Session Dashboard**: Live sidebar showing VRAM status, cache hits, and time saved.

---

### Windows (via WSL2)

> MELLM runs natively on Linux. Windows users can run it with full GPU acceleration
> through WSL2 (Windows Subsystem for Linux 2), which gives you a real Ubuntu
> environment with direct access to your NVIDIA GPU.

**Prerequisites:**
- Windows 10 (Build 19041+) or Windows 11
- NVIDIA GPU with 6GB+ VRAM
- NVIDIA driver **470.76+** installed on Windows (not inside WSL — the Windows driver handles the bridge)

#### Step 1 — Enable WSL2

Open PowerShell as Administrator and run:

```powershell
wsl --install
```

This installs WSL2 with Ubuntu 22.04 by default. Restart when prompted.

If you already have WSL1, upgrade to WSL2:
```powershell
wsl --set-default-version 2
wsl --install -d Ubuntu-22.04
```

Verify WSL2 is active:
```powershell
wsl --list --verbose
# Should show VERSION 2
```

#### Step 2 — Install NVIDIA CUDA inside WSL2

> **Important:** Do NOT install the NVIDIA driver inside WSL2. The Windows driver already
> handles GPU access. You only need the CUDA toolkit inside WSL2.

Open your Ubuntu WSL2 terminal and run:

```bash
# Remove any existing CUDA GPG keys to avoid conflicts
sudo apt-key del 7fa2af80

# Add NVIDIA's WSL2-specific CUDA repository
wget https://developer.download.nvidia.com/compute/cuda/repos/wsl-ubuntu/x86_64/cuda-wsl-ubuntu.pin
sudo mv cuda-wsl-ubuntu.pin /etc/apt/preferences.d/cuda-repository-pin-600
wget https://developer.download.nvidia.com/compute/cuda/12.1.0/local_installers/cuda-repo-wsl-ubuntu-12-1-local_12.1.0-1_amd64.deb
sudo dpkg -i cuda-repo-wsl-ubuntu-12-1-local_12.1.0-1_amd64.deb
sudo cp /var/cuda-repo-wsl-ubuntu-12-1-local/cuda-*-keyring.gpg /usr/share/keyrings/
sudo apt-get update
sudo apt-get install -y cuda-toolkit-12-1
```

Verify CUDA is working:
```bash
nvidia-smi
# Should show your GPU name and VRAM
```

#### Step 3 — Install Python 3.11

```bash
sudo apt update
sudo apt install software-properties-common -y
sudo add-apt-repository ppa:deadsnakes/ppa -y
sudo apt update
sudo apt install python3.11 python3.11-venv python3.11-dev -y

# Verify
python3.11 --version
```

#### Step 4 — Clone and set up MELLM

```bash
# Navigate to a good location (WSL2 filesystem, not /mnt/c — see note below)
cd ~
git clone https://github.com/Rahul-14507/MELLM
cd MELLM

# Create venv
python3.11 -m venv .venv
source .venv/bin/activate

# Install llama-cpp-python + remaining dependencies
chmod +x setup.sh && ./setup.sh
```

#### Step 5 — Run MELLM

```bash
source .venv/bin/activate
python cli.py
```

The setup wizard will detect your GPU and recommend appropriate model sizes.

#### ⚠️ Important Notes for Windows Users

**Store your project on the WSL2 filesystem, not Windows drives.**

```bash
# Good — fast WSL2 filesystem
~/MELLM/

# Avoid — slow cross-filesystem access, causes I/O bottlenecks
/mnt/c/Users/YourName/MELLM/
/mnt/d/Projects/MELLM/
```

Model downloads go to `~/.cache/huggingface/` inside WSL2 by default. This is fine.
If you want them on a Windows drive with more space, symlink it:

```bash
mkdir -p /mnt/d/mellm-models
ln -s /mnt/d/mellm-models ~/.cache/mellm_gguf
```

**VRAM headroom on Windows is ~500MB–1GB less than native Linux** because Windows
desktop processes and WSL2 overhead occupy some GPU memory. On a 6GB card:

| Setup | Effective VRAM for MELLM |
|-------|--------------------------|
| Native Linux | ~5.5 GB |
| WSL2 on Windows | ~4.5–5.0 GB |

If you hit `Failed to create llama_context` errors, switch to smaller/more quantized
models in the setup wizard (choose Q2_K variants for 7B models, or stick to 1.5B–3B
specialists).

**Accessing MELLM's API from Windows browsers/apps:**

The FastAPI server inside WSL2 is accessible from Windows at `http://localhost:8000`
by default — WSL2 automatically bridges the network. No extra configuration needed.

```bash
# Inside WSL2
python api.py

# From Windows browser or app
curl http://localhost:8000/health
```

#### WSL2 Troubleshooting

**`nvidia-smi` not found inside WSL2:**
Update your Windows NVIDIA driver to 470.76+. The WSL2 GPU bridge requires a recent
driver on the Windows side.

**`CUDA error: no kernel image is available` during inference:**
Your CUDA toolkit version doesn't match your driver. Re-run Step 2 with the CUDA
version that matches your driver (check `nvidia-smi` on Windows for the CUDA version).

**Slow model downloads:**
WSL2 DNS can be slow. Add Google's DNS to `/etc/resolv.conf`:
```bash
echo "nameserver 8.8.8.8" | sudo tee /etc/resolv.conf
```

**Out of memory after loading medical model:**
Windows is using more VRAM than expected. Close GPU-heavy Windows apps (Chrome with
hardware acceleration, games, etc.) before running MELLM, or switch to the Q2_K
medical model in `user_config.yaml`.

---

## 💻 Usage

## ⚡ Quick Demo

```bash
python cli.py
```

**Try these queries to see the router in action:**
1. "Explain binary search in Java"
2. "Now in Python"
3. "What is the time complexity?"

### CLI Mode

```bash
python cli.py
```

On startup, MELLM loads the router model and displays a **Model Availability Dashboard** showing which specialists are cached. The router shows **`Loaded (persistent)`** — it's already in VRAM before your first query.

```
Query: Binary Search in Java

Domain: CODE
╭─── Response (Specialist: code) ───╮
│ Here is a Java implementation...  │
╰───────────────────────────────────╯
Metrics: Router: resident (0s) | Specialist Load: 1.63s | Inference: 8.42s | Context: 1 turns
╭──────────────── ⚡ Efficiency ────────────────╮
│ Queries this session : 1                       │
│ Specialist cache hits: 0/1 (0%)               │
│ Router loads saved   : 0 (~0.0s saved)        │
│ Active specialist    : CODE (freshly loaded)  │
│ Context turns active : 1/3                    │
╰───────────────────────────────────────────────╯

Query: Now in Python?
  → Domain continuity: short follow-up, keeping 'code'
  → Cache hit — reusing code specialist (0s)
╭─── Response (Specialist: code) ───╮
│ Here is the Python equivalent...  │
╰───────────────────────────────────╯
Metrics: Router: resident (0s) | Specialist Load: 0s | Inference: 7.1s | Context: 2 turns
```

**CLI Commands:**

| Command | Effect |
|---------|--------|
| *(any query)* | Routes through MoE pipeline |
| `clear` | Wipes conversation history, starts fresh |
| `exit` / `quit` | Cleanly unloads all models from VRAM and exits |

### Multi-Agent Queries

MELLM automatically detects when a query spans multiple domains and routes sub-tasks to the appropriate specialists, then merges the results into a single coherent response:

```
Query: "Explain binary search AND give Java code AND analyse its complexity"

→ GENERAL specialist: conceptual explanation
→ CODE specialist:    Java implementation
→ MATH specialist:    O(log n) complexity analysis
→ Merged into a single response
```

This happens transparently — just ask your cross-domain question and MELLM handles the rest.

### API Mode

```bash
python api.py
# Server starts at http://0.0.0.0:8000
```

**Endpoints:**

| Method | Endpoint | Description |
|--------|----------|-------------|
| `GET` | `/health` | Health check and router status |
| `POST` | `/query` | Process a query through the router pipeline |

**Example Request:**
```bash
curl -X POST http://localhost:8000/query \
  -H "Content-Type: application/json" \
  -d '{"prompt": "What are the symptoms of appendicitis?"}'
```

**Example Response:**
```json
{
  "original_prompt": "What are the symptoms of appendicitis?",
  "domain": "medical",
  "confidence": 0.95,
  "rewritten_prompt": "...",
  "response": "The common symptoms of appendicitis include...",
  "router_load_time": 0.0,
  "specialist_load_time": 10.81,
  "inference_time_seconds": 5.23,
  "cache_hit": false,
  "context_turns": 1
}
```

### Preloading Models

Pre-download models so they're ready instantly:

```bash
# Download ALL specialist models
python cli.py --preload all

# Download only the medical specialist
python cli.py --preload medical
```

---

## ⚡ Performance & Benchmarks

Benchmarked on **NVIDIA GeForce RTX 3050 6GB Laptop GPU** · CUDA 12.1 · Python 3.11

### Model Load & Inference

| Stage | 1.5B Models | 3B Models | 7B Models (Q2_K) |
|-------|-------------|-----------|-------------------|
| Model Load (first time) | ~1-2s | ~3-4s | ~5-6s |
| Model Load (cache hit) | **0s** | **0s** | **0s** |
| Inference Speed | 15-25 tok/s | 10-15 tok/s | 8-12 tok/s |
| VRAM Usage | ~1.5 GB | ~2.5 GB | ~3-4 GB |
| Context Window | 4096 tokens | 1024 tokens | 1024 tokens |

| End-to-End | 1st Query | 2nd Query (Same Domain) | 2nd Query (Domain Switch) |
|------------|-----------|-------------------------|---------------------------|
| Router overhead | 0s (persistent) | 0s (persistent) | 0s (persistent) |
| Specialist load | 1-6s | **0s (hot cache)** | 1-6s (new domain) |
| Inference (1.5B) | ~5-15s | ~5-15s | ~5-15s |
| **Total typical query** | **~6-20s** | **~5-15s** | **~6-20s** |

### Routing Accuracy

Tested across 25 queries (5 per domain):

| Domain | Correct | Accuracy |
|--------|---------|----------|
| Code | 4/5 | 80% |
| Math | 5/5 | 100% |
| Medical | 5/5 | 100% |
| Legal | 5/5 | 100% |
| General | 4/5 | 80% |
| **Overall** | **22/25** | **88%** |

Misclassified queries were genuinely ambiguous (e.g. "Debug this segmentation fault in C++" → classified as general; "Explain Occam's Razor" → classified as math). Core domain classification is 100% accurate for unambiguous queries.

### End-to-End Latency (per domain)

| Domain | Model | Cold Load | Hot Cache | Inference | Total (cold) |
|--------|-------|-----------|-----------|-----------|--------------|
| **Code** | Qwen2.5-Coder-1.5B | ~3.4s | **0.0s** | ~7.2s | ~17s |
| **Math** | Qwen2.5-Math-1.5B | ~3.8s | **0.0s** | ~9.5s | ~20s |
| **Medical** | BioMistral-7B Q2 | ~6.3s | **0.0s** | ~18.6s | ~32s |
| **Legal** | Magistrate-3B | ~5.8s | **0.0s** | ~18.5s | ~31s |
| **General** | Qwen2.5-1.5B | ~0.0s* | **0.0s** | ~2.3s | ~10s |

\* General specialist was already cached from router (same base model)

**Router overhead: 0s** — persistent in VRAM, never reloaded between queries.

**Hot cache = 0s load** — same-domain follow-up queries skip loading entirely.
Consecutive queries in a coding session: first query ~17s, all subsequent ~12s.

---

## 🔧 Configuration

All model and generation settings are in `config.yaml`:

```yaml
router:
  model_id: "Qwen/Qwen2.5-1.5B-Instruct"
  max_new_tokens: 256

specialists:
  code:
    model_id: "Qwen/Qwen2.5-Coder-1.5B-Instruct"
    max_new_tokens: 2048
  math:
    model_id: "Qwen/Qwen2.5-Math-1.5B-Instruct"
    max_new_tokens: 2048
  medical:
    model_id: "BioMistral/BioMistral-7B-DARE-GGUF"
    max_new_tokens: 2048
  legal:
    model_id: "AdaptLLM/law-LLM"
    max_new_tokens: 2048
  general:
    model_id: "Qwen/Qwen2.5-1.5B-Instruct"
    max_new_tokens: 2048
```

### Adding a New Specialist

1. **Add the model to `GGUF_REGISTRY`** in `loader/airllm_loader.py`:
```python
GGUF_REGISTRY = {
    # ... existing models ...
    "your-org/your-model": (
        "gguf-repo-id/your-model-GGUF",
        "your-model.Q4_K_M.gguf"
    ),
}
```

2. **Create a specialist class** in `specialists/your_domain.py`:
```python
from .base_specialist import BaseSpecialist

class YourSpecialist(BaseSpecialist):
    def generate(self, prompt: str) -> str:
        response = self.model.create_chat_completion(
            messages=[{"role": "user", "content": prompt}],
            max_tokens=self.max_new_tokens,
            temperature=0.7,
        )
        return self._postprocess(
            response["choices"][0]["message"]["content"]
        )
```

3. **Register it in `orchestrator.py`**:
```python
from specialists.your_domain import YourSpecialist

SPECIALIST_MAP = {
    # ... existing specialists ...
    "your_domain": YourSpecialist,
}
```

4. **Add it to `config.yaml`**:
```yaml
specialists:
  your_domain:
    model_id: "your-org/your-model"
    max_new_tokens: 2048
```

5. **Update the router's domain list** in `router/classifier.py` to include your new domain.

---

## 🗺️ Roadmap

- [x] Persistent router model (no per-query load overhead)
- [x] Hot specialist cache (domain-switch-only unloading)
- [x] Conversation context memory (last 3 turns)
- [x] Domain continuity bias for follow-up queries
- [x] Live session efficiency panel in CLI
- [x] Multi-agent composition (cross-domain queries)
- [x] Domain-aware session history (streak display)
- [x] Interactive setup wizard with hardware detection
- [x] FastAPI REST endpoint
- [x] Web UI
- [x] Evaluation benchmark suite for routing accuracy
- [x] Streaming token output
- [ ] Docker container for easy deployment
- [ ] Support for AMD GPUs (ROCm)

---

## 🐛 Known Limitations

- **First-query latency**: The first query for a new domain takes 1-6s to load the specialist; subsequent same-domain queries are instant (hot cache)
- **Context window**: 7B models are limited to 1024 tokens of context to fit in VRAM
- **Single query at a time**: The system processes one query before accepting the next
- **Context continuity for the API**: The conversation history is session-bound to the `LLMRouter` instance; the REST API resets on each server restart
- **Domain continuity limitations**: Very ambiguous short queries (e.g., "More?") are biased toward the previous domain, which may not always be correct

---

## 🛠️ Tech Stack

| Component | Technology | Purpose |
|-----------|------------|---------|
| **Inference Engine** | [llama-cpp-python](https://github.com/abetlen/llama-cpp-python) | GPU-accelerated GGUF model inference |
| **Model Source** | [Hugging Face Hub](https://huggingface.co) | Auto-download quantized GGUF models |
| **CLI Interface** | [Rich](https://github.com/Textualize/rich) | Beautiful terminal dashboards and panels |
| **REST API** | [FastAPI](https://fastapi.tiangolo.com/) + [Uvicorn](https://www.uvicorn.org/) | Async HTTP server |
| **Configuration** | [PyYAML](https://pyyaml.org/) | YAML-based model and parameter configuration |
| **Environment** | [python-dotenv](https://pypi.org/project/python-dotenv/) | Secure token management via `.env` |
| **Model Format** | [GGUF](https://github.com/ggerganov/ggml) | Quantized model format (Q2_K, Q4_K_M) |

---

## 🤝 Contributing

Contributions are welcome! See [CONTRIBUTING.md](CONTRIBUTING.md) for the full guide covering development setup, how to add a new specialist, code style, and the PR process.

**Quick ideas to get started:**
- 🌍 **New domain specialists** — Add support for science, finance, history, etc.
- 🧪 **Evaluation benchmarks** — Build test suites to measure routing accuracy and specialist quality
- 🖥️ **Web UI** — Build a Gradio or Streamlit frontend
- 📊 **Metrics dashboard** — Track routing accuracy, latency, and VRAM usage over time
- 🏎️ **Performance optimization** — Explore batch inference, speculative decoding, or model caching strategies
- 📝 **Better prompts** — Improve specialist system prompts for higher-quality responses
- 🐛 **Bug fixes** — Check the [Issues](https://github.com/Rahul-14507/MELLM/issues) tab for known bugs

---

## 📄 License

This project is licensed under the **MIT License** — see the [LICENSE](LICENSE) file for details.

> **Note:** The individual models used by MELLM have their own licenses. Please check each model's Hugging Face page for their specific terms before redistribution.

---

## 🙏 Acknowledgements

- [Qwen Team](https://huggingface.co/Qwen) — For the excellent Qwen2.5 series of small instruction-tuned models
- [BioMistral](https://huggingface.co/BioMistral) — For the medical domain-adapted BioMistral-7B-DARE
- [mradermacher](https://huggingface.co/mradermacher) — For the Magistrate legal model GGUF quantization
- [bartowski](https://huggingface.co/bartowski) — For high-quality GGUF quantizations of math models
- [Georgi Gerganov](https://github.com/ggerganov) — For the `ggml` library and GGUF format
- [Andrei Betlen](https://github.com/abetlen) — For `llama-cpp-python`, the Python bindings for llama.cpp

MELLM

MELLM 최신버젼 다운로드

오늘의 추천 프로젝트

Was this content helpful?

Recent Feedback

MELLM

MELLM 최신버젼 다운로드

오늘의 추천 프로젝트

Was this content helpful?

Recent Feedback

Download README