Local_LLM_Running_Ollama.md

Skill: Local LLM Running (Ollama)

Purpose

To run Large Language Models locally using Ollama, ensuring data privacy, zero API costs, and offline capability while providing a standard REST API for applications.

When to Use

• When data privacy is paramount (medical, legal, personal data).
• For development and testing without incurring API costs.
• When you need to experiment with open-source models (Llama 3, Mistral, etc.).

Procedure

1. Installation & Model Setup

Install Ollama and pull the desired model.

# Install (macOS/Linux)
curl -fsSL https://ollama.com/install.sh | sh

# Pull a model
ollama pull llama3:8b

2. Basic Usage (CLI)

Interact with the model directly in your terminal.

ollama run llama3:8b "Why is the sky blue?"

3. Programmatic Integration (Node.js)

Use the Ollama REST API or official library to integrate into your app.

npm install ollama

import ollama from 'ollama';

async function chat() {
  const response = await ollama.chat({
    model: 'llama3:8b',
    messages: [{ role: 'user', content: 'Explain quantum physics to a 5-year old' }],
    stream: true,
  });

  for await (const part of response) {
    process.stdout.write(part.message.content);
  }
}

4. Customizing Models (Modelfile)

Create a specialized version of a model with custom system prompts.

1. Create a file named Modelfile:

FROM llama3:8b

# Set parameters
PARAMETER temperature 0.1
PARAMETER top_p 0.9

# Set system message
SYSTEM """
You are a senior TypeScript developer. 
You provide concise, high-performance code snippets.
Always use ESM syntax.
"""

2. Create the model:

ollama create ts-expert -f Modelfile

5. Running as a Service (Docker)

Run Ollama in a container for consistent deployment.

docker run -d -v ollama:/root/.ollama -p 11434:11434 --name ollama ollama/ollama

Constraints

• VRAM Requirements:
  • 7B/8B models: ~8GB RAM/VRAM.
  • 13B models: ~16GB RAM/VRAM.
  • 70B models: ~48GB+ RAM/VRAM.
• Latency: Local models are significantly slower than GPT-4o unless running on a high-end GPU (RTX 3090/4090 or Apple M2/M3 Max).
• Quantization: Most Ollama models are 4-bit quantized (Q4_K_M) by default, which slightly reduces reasoning capability but saves memory.

Expected Output

A locally running LLM service accessible via a REST API on localhost:11434.