Kimi K2.5 Ollama Rehberi: Cloud Erişimi + Self-Hosted Notları

# macOS
curl -fsSL https://ollama.com/install.sh | sh

# Linux
curl -fsSL https://ollama.com/install.sh | sh

# Verify installation
ollama --version

# Pull the current Ollama cloud tag
ollama pull kimi-k2.5:cloud

# Run a test query
ollama run kimi-k2.5:cloud "Hello from Ollama cloud mode"

# Modelfile for self-hosted Kimi K2.5 workflow
FROM /path/to/Kimi-K2.5

# System prompt
SYSTEM """You are Kimi K2.5, running in a self-hosted deployment.
You provide helpful, accurate, and detailed responses."""

# Parameter tuning
PARAMETER temperature 0.7
PARAMETER top_p 0.9
PARAMETER top_k 40
PARAMETER num_ctx 65536  # Adjust based on available VRAM
PARAMETER num_predict 4096
PARAMETER repeat_penalty 1.1

# Engine-specific parameters vary by backend (vLLM/SGLang/KTransformers)

ollama create kimi-local -f Modelfile
ollama run kimi-local

# Check available VRAM
nvidia-smi

# Run with specific GPU allocation
CUDA_VISIBLE_DEVICES=0,1,2,3 ollama run kimi-local

# Limit context window for lower VRAM usage
# In Modelfile: PARAMETER num_ctx 32768

# Interactive mode
ollama run kimi-k2.5:cloud

# Single prompt
ollama run kimi-k2.5:cloud "Explain quantum computing"

# With system prompt
ollama run kimi-k2.5:cloud --system "You are a code assistant" "Write Python for fibonacci"

import requests
import json

# Ollama API endpoint
OLLAMA_URL = "http://localhost:11434/api/generate"

def query_kimi(prompt, system=None):
    payload = {
        "model": "kimi-k2.5:cloud",
        "prompt": prompt,
        "system": system or "You are a helpful assistant.",
        "stream": False,
        "options": {
            "temperature": 0.7,
            "num_ctx": 65536,
            "num_predict": 4096
        }
    }

    response = requests.post(OLLAMA_URL, json=payload)
    return response.json()["response"]

# Example usage
result = query_kimi(
    "Write a function to sort a list",
    system="You are a Python expert"
)
print(result)

async function queryKimi(prompt: string, system?: string) {
  const response = await fetch('http://localhost:11434/api/generate', {
    method: 'POST',
    headers: { 'Content-Type': 'application/json' },
    body: JSON.stringify({
      model: 'kimi-k2.5:cloud',
      prompt,
      system: system || 'You are a helpful assistant.',
      stream: false,
      options: {
        temperature: 0.7,
        num_ctx: 65536,
      },
    }),
  });

  const data = await response.json();
  return data.response;
}

import requests

def stream_kimi(prompt):
    payload = {
        "model": "kimi-k2.5:cloud",
        "prompt": prompt,
        "stream": True
    }

    response = requests.post(
        "http://localhost:11434/api/generate",
        json=payload,
        stream=True
    )

    for line in response.iter_lines():
        if line:
            data = json.loads(line)
            if "response" in data:
                print(data["response"], end="", flush=True)
            if data.get("done"):
                break

stream_kimi("Tell me a story about AI.")

# Configure Ollama for multiple GPUs
export OLLAMA_NUM_PARALLEL=4
export OLLAMA_MAX_LOADED_MODELS=1
export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7

# Start Ollama server
ollama serve

# High-performance Modelfile
FROM /path/to/Kimi-K2.5

# Optimize for speed
PARAMETER num_ctx 32768  # Balance between capacity and speed
PARAMETER num_gpu 100     # Use all available layers
PARAMETER batch_size 512  # Increase batch processing

# Reduce precision for faster inference
PARAMETER f16_kv true

# docker-compose.yml
version: '3.8'
services:
  ollama:
    image: ollama/ollama:latest
    container_name: ollama-kimi
    volumes:
      - ollama:/root/.ollama
    ports:
      - '11434:11434'
    environment:
      - OLLAMA_NUM_PARALLEL=4
      - CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: 8
              capabilities: [gpu]

volumes:
  ollama:

// settings.json
{
  "ollama.model": "kimi-k2.5:cloud",
  "ollama.apiUrl": "http://localhost:11434",
  "ollama.parameters": {
    "temperature": 0.7,
    "num_ctx": 65536
  }
}

// config.json
{
  "models": [
    {
      "title": "Kimi K2.5 (Ollama Cloud)",
      "provider": "ollama",
      "model": "kimi-k2.5:cloud",
      "apiBase": "http://localhost:11434"
    }
  ]
}

# Local code assistant
def local_code_review(code):
    prompt = f"""Review this code for:
    1. Security issues
    2. Performance optimizations
    3. Best practices

    Code:
    {code}
    """
    return query_kimi(prompt, system="You are a senior software engineer.")

# Monitor GPU usage
watch -n 1 nvidia-smi

# Check Ollama logs
journalctl -u ollama -f

# Monitor response times
ollama run kimi-k2.5:cloud --verbose "Test query"

# Update to latest version
ollama pull kimi-k2.5:cloud

# List available versions
ollama list

# Remove old versions
ollama rm kimi-k2.5:cloud

# Reduce context window
# In Modelfile: PARAMETER num_ctx 16384

# Retry pull for the cloud tag
ollama pull kimi-k2.5:cloud

# Increase GPU layers
PARAMETER num_gpu 100

# Check GPU utilization
nvidia-smi dmon

# Resume interrupted download
ollama pull kimi-k2.5:cloud

# Check disk space
df -h