global: scrape_interval: 2s # Intervalo corto para capturar datos de latencia y saturación evaluation_interval: 2s scrape_configs: - job_name: 'vllm' metrics_path: '/metrics' static_configs: - targets: ['vllm:8000'] labels: engine: 'vllm' model: 'qwen2.5-3b-awq' - job_name: 'litellm' metrics_path: '/metrics' static_configs: - targets: ['litellm:4000'] labels: service: 'llm-gateway'