Panduan Kimi K2.5 Ollama: Akses Cloud + Catatan Self-Hosted

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

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

# Verifikasi instalasi
ollama --version

# Pull tag cloud Ollama saat ini
ollama pull kimi-k2.5:cloud

# Jalankan query uji
ollama run kimi-k2.5:cloud "Hello from Ollama cloud mode"

# Modelfile untuk alur kerja Kimi K2.5 self-hosted
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."""

# Penyetelan parameter
PARAMETER temperature 0.7
PARAMETER top_p 0.9
PARAMETER top_k 40
PARAMETER num_ctx 65536  # Sesuaikan dengan VRAM yang tersedia
PARAMETER num_predict 4096
PARAMETER repeat_penalty 1.1

# Parameter spesifik engine berbeda per backend (vLLM/SGLang/KTransformers)

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

# Cek VRAM yang tersedia
nvidia-smi

# Jalankan dengan alokasi GPU tertentu
CUDA_VISIBLE_DEVICES=0,1,2,3 ollama run kimi-local

# Batasi context window untuk menekan penggunaan VRAM
# Di Modelfile: PARAMETER num_ctx 32768

# Mode interaktif
ollama run kimi-k2.5:cloud

# Prompt tunggal
ollama run kimi-k2.5:cloud "Explain quantum computing"

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

import requests
import json

# Endpoint API Ollama
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"]

# Contoh penggunaan
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.")

# Konfigurasi Ollama untuk banyak GPU
export OLLAMA_NUM_PARALLEL=4
export OLLAMA_MAX_LOADED_MODELS=1
export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7

# Jalankan server Ollama
ollama serve

# Modelfile berperforma tinggi
FROM /path/to/Kimi-K2.5

# Optimasi untuk kecepatan
PARAMETER num_ctx 32768  # Seimbangkan antara kapasitas dan kecepatan
PARAMETER num_gpu 100     # Gunakan semua layer yang tersedia
PARAMETER batch_size 512  # Tingkatkan pemrosesan batch

# Turunkan presisi untuk inferensi lebih cepat
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"
    }
  ]
}

# Asisten kode lokal
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 penggunaan GPU
watch -n 1 nvidia-smi

# Cek log Ollama
journalctl -u ollama -f

# Monitor waktu respons
ollama run kimi-k2.5:cloud --verbose "Test query"

# Update ke versi terbaru
ollama pull kimi-k2.5:cloud

# Daftar versi yang tersedia
ollama list

# Hapus versi lama
ollama rm kimi-k2.5:cloud

# Kurangi context window
# Di Modelfile: PARAMETER num_ctx 16384

# Coba pull ulang untuk tag cloud
ollama pull kimi-k2.5:cloud

# Tambah GPU layer
PARAMETER num_gpu 100

# Cek utilisasi GPU
nvidia-smi dmon

# Lanjutkan unduhan yang terputus
ollama pull kimi-k2.5:cloud

# Cek ruang disk
df -h