miku-discord/QUICK_REFERENCE.md

Quick Reference: Ollama → Llama.cpp Migration

Environment Variables

Old (Ollama)	New (llama.cpp)	Purpose
OLLAMA_URL	LLAMA_URL	Server endpoint
OLLAMA_MODEL	TEXT_MODEL	Text generation model name
N/A	VISION_MODEL	Vision model name

API Endpoints

Purpose	Old (Ollama)	New (llama.cpp)
Text generation	/api/generate	/v1/chat/completions
Vision	/api/generate	/v1/chat/completions
Health check	GET /	GET /health
Model management	Manual switch_model()	Automatic via llama-swap

Function Changes

Old Function	New Function	Status
query_ollama()	query_llama()	Aliased for compatibility
analyze_image_with_qwen()	analyze_image_with_vision()	Aliased for compatibility
switch_model()	Removed	llama-swap handles automatically

Request Format

Text Generation

Before (Ollama):

payload = {
    "model": "llama3.1",
    "prompt": "Hello world",
    "system": "You are Miku",
    "stream": False
}
await session.post(f"{OLLAMA_URL}/api/generate", json=payload)

After (OpenAI):

payload = {
    "model": "llama3.1",
    "messages": [
        {"role": "system", "content": "You are Miku"},
        {"role": "user", "content": "Hello world"}
    ],
    "stream": False
}
await session.post(f"{LLAMA_URL}/v1/chat/completions", json=payload)

Vision Analysis

Before (Ollama):

await switch_model("moondream")  # Manual switch!
payload = {
    "model": "moondream",
    "prompt": "Describe this image",
    "images": [base64_img],
    "stream": False
}
await session.post(f"{OLLAMA_URL}/api/generate", json=payload)

After (OpenAI):

# No manual switch needed!
payload = {
    "model": "moondream",  # llama-swap auto-switches
    "messages": [{
        "role": "user",
        "content": [
            {"type": "text", "text": "Describe this image"},
            {"type": "image_url", "image_url": {"url": f"data:image/jpeg;base64,{base64_img}"}}
        ]
    }],
    "stream": False
}
await session.post(f"{LLAMA_URL}/v1/chat/completions", json=payload)

Response Format

Before (Ollama):

{
  "response": "Hello! I'm Miku!",
  "model": "llama3.1"
}

After (OpenAI):

{
  "choices": [{
    "message": {
      "role": "assistant",
      "content": "Hello! I'm Miku!"
    }
  }],
  "model": "llama3.1"
}

Docker Services

Before:

services:
  ollama:
    image: ollama/ollama
    ports: ["11434:11434"]
    volumes: ["ollama_data:/root/.ollama"]
    
  bot:
    environment:
      - OLLAMA_URL=http://ollama:11434
      - OLLAMA_MODEL=llama3.1

After:

services:
  llama-swap:
    image: ghcr.io/mostlygeek/llama-swap:cuda
    ports: ["8080:8080"]
    volumes:
      - ./models:/models
      - ./llama-swap-config.yaml:/app/config.yaml
    
  bot:
    environment:
      - LLAMA_URL=http://llama-swap:8080
      - TEXT_MODEL=llama3.1
      - VISION_MODEL=moondream

Model Management

Feature	Ollama	llama.cpp + llama-swap
Model loading	Manual ollama pull	Download GGUF files to /models
Model switching	Manual switch_model() call	Automatic based on request
Model unloading	Manual or never	Automatic after TTL (30m text, 15m vision)
VRAM management	Always loaded	Load on demand, unload when idle
Storage format	Ollama format	GGUF files
Location	Docker volume	Host directory ./models/

Configuration Files

File	Purpose	Format
docker-compose.yml	Service orchestration	YAML
llama-swap-config.yaml	Model configs, TTL settings	YAML
models/llama3.1.gguf	Text model weights	Binary GGUF
models/moondream.gguf	Vision model weights	Binary GGUF
models/moondream-mmproj.gguf	Vision projector	Binary GGUF

Monitoring

Tool	URL	Purpose
llama-swap Web UI	http://localhost:8080/ui	Monitor models, logs, timers
Health endpoint	http://localhost:8080/health	Check if server is ready
Running models	http://localhost:8080/running	List currently loaded models
Metrics	http://localhost:8080/metrics	Prometheus-compatible metrics

Common Commands

# Check what's running
curl http://localhost:8080/running

# Check health
curl http://localhost:8080/health

# Manually unload all models
curl -X POST http://localhost:8080/models/unload

# View logs
docker-compose logs -f llama-swap

# Restart services
docker-compose restart

# Check model files
ls -lh models/

Quick Troubleshooting

Issue	Solution
"Model not found"	Verify files in ./models/ match config
CUDA errors	Check: docker run --rm --gpus all nvidia/cuda:12.0-base nvidia-smi
Slow responses	First load is slow; subsequent loads use cache
High VRAM usage	Models will auto-unload after TTL expires
Bot can't connect	Check: curl http://localhost:8080/health

---

Remember: The migration maintains backward compatibility. Old function names are aliased, so existing code continues to work!