نماذج ⁨LLM⁩ المحلية بلا واجهة (⁨headless⁩): تشغيل النماذج دون واجهة مستخدم (⁨2026⁩)

# Run Ollama headless
ollama serve

# This starts the API at http://localhost:11434/v1
# No chat UI, just a background service

# Use the API from Python
from openai import OpenAI
client = OpenAI(base_url="http://localhost:11434/v1", api_key="ollama")
response = client.chat.completions.create(
  model="llama3.2:3b",
  messages=[{"role": "user", "content": "Hello"}]
)
print(response.choices[0].message.content)

# Or from curl
curl http://localhost:11434/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{{"model": "llama3.2:3b", "messages": [{{"role": "user", "content": "Hello"}}]}}'

# Install vLLM
pip install vllm

# Run headless with API
vllm serve llama-3.1-8b-instruct \
  --host 0.0.0.0 \
  --port 8000 \
  --gpu-memory-utilization 0.9

# Access at http://localhost:8000/v1
# Supports 50+ concurrent requests

# Use from Python (same as Ollama)
from openai import OpenAI
client = OpenAI(base_url="http://localhost:8000/v1", api_key="anything")
response = client.chat.completions.create(
  model="meta-llama/Llama-2-7b-chat-hf",
  messages=[{"role": "user", "content": "Hello"}]
)
print(response.choices[0].message.content)

# Example: Deploy vLLM on a server via Docker
docker run --gpus all -p 8000:8000 \
  --env VLLM_API_KEY="your-secret-key" \
  vllm/vllm-openai:latest \
  --model meta-llama/Llama-2-13b-chat-hf \
  --tensor-parallel-size 2  # Use 2 GPUs

# Nginx reverse proxy config (optional)
# server {
#   listen 80;
#   location / {
#     proxy_pass http://localhost:8000;
#     proxy_set_header Authorization "Bearer $http_authorization";
#   }
# }

# Monitor GPU usage (nvidia-smi)
watch nvidia-smi  # Updates every 2 seconds

# Monitor request latency
# Add logging to your client code
import time
start = time.time()
response = client.chat.completions.create(...)
latency = time.time() - start
print(f"Request took {latency:.2f} seconds")

# Monitor vLLM logs
docker logs -f <container_id>

# Check error rates
# Parse logs for errors or use a monitoring tool (Prometheus + Grafana)