|
| 1 | +""" |
| 2 | +Cliente para la API de vLLM integrada con ModelSync |
| 3 | +""" |
| 4 | + |
| 5 | +import requests |
| 6 | +import json |
| 7 | +from typing import List, Optional, Dict, Any |
| 8 | +from datetime import datetime |
| 9 | + |
| 10 | +class VLLMClient: |
| 11 | + """Cliente para interactuar con la API de vLLM integrada con ModelSync""" |
| 12 | + |
| 13 | + def __init__(self, base_url: str = "http://localhost:8001"): |
| 14 | + self.base_url = base_url.rstrip('/') |
| 15 | + self.session = requests.Session() |
| 16 | + |
| 17 | + def health_check(self) -> Dict[str, Any]: |
| 18 | + """Verificar el estado del servicio""" |
| 19 | + response = self.session.get(f"{self.base_url}/health") |
| 20 | + response.raise_for_status() |
| 21 | + return response.json() |
| 22 | + |
| 23 | + def list_models(self) -> List[Dict[str, Any]]: |
| 24 | + """Listar modelos disponibles""" |
| 25 | + response = self.session.get(f"{self.base_url}/models") |
| 26 | + response.raise_for_status() |
| 27 | + return response.json() |
| 28 | + |
| 29 | + def generate( |
| 30 | + self, |
| 31 | + prompt: str, |
| 32 | + max_tokens: int = 100, |
| 33 | + temperature: float = 0.7, |
| 34 | + top_p: float = 0.9, |
| 35 | + top_k: int = 50, |
| 36 | + stop: Optional[List[str]] = None, |
| 37 | + model_name: Optional[str] = None, |
| 38 | + save_to_version_control: bool = True |
| 39 | + ) -> Dict[str, Any]: |
| 40 | + """Generar texto con integración ModelSync""" |
| 41 | + payload = { |
| 42 | + "prompt": prompt, |
| 43 | + "max_tokens": max_tokens, |
| 44 | + "temperature": temperature, |
| 45 | + "top_p": top_p, |
| 46 | + "top_k": top_k, |
| 47 | + "stream": False, |
| 48 | + "save_to_version_control": save_to_version_control |
| 49 | + } |
| 50 | + |
| 51 | + if stop: |
| 52 | + payload["stop"] = stop |
| 53 | + if model_name: |
| 54 | + payload["model_name"] = model_name |
| 55 | + |
| 56 | + response = self.session.post(f"{self.base_url}/generate", json=payload) |
| 57 | + response.raise_for_status() |
| 58 | + return response.json() |
| 59 | + |
| 60 | + def generate_batch( |
| 61 | + self, |
| 62 | + prompts: List[str], |
| 63 | + max_tokens: int = 100, |
| 64 | + temperature: float = 0.7, |
| 65 | + top_p: float = 0.9, |
| 66 | + top_k: int = 50, |
| 67 | + stop: Optional[List[str]] = None, |
| 68 | + save_to_version_control: bool = True |
| 69 | + ) -> List[Dict[str, Any]]: |
| 70 | + """Generar texto para múltiples prompts en lote""" |
| 71 | + requests_data = [] |
| 72 | + for prompt in prompts: |
| 73 | + request_data = { |
| 74 | + "prompt": prompt, |
| 75 | + "max_tokens": max_tokens, |
| 76 | + "temperature": temperature, |
| 77 | + "top_p": top_p, |
| 78 | + "top_k": top_k, |
| 79 | + "stream": False, |
| 80 | + "save_to_version_control": save_to_version_control |
| 81 | + } |
| 82 | + if stop: |
| 83 | + request_data["stop"] = stop |
| 84 | + requests_data.append(request_data) |
| 85 | + |
| 86 | + response = self.session.post(f"{self.base_url}/generate/batch", json=requests_data) |
| 87 | + response.raise_for_status() |
| 88 | + return response.json() |
| 89 | + |
| 90 | + def get_modelsync_status(self) -> Dict[str, Any]: |
| 91 | + """Obtener estado de la integración ModelSync""" |
| 92 | + response = self.session.get(f"{self.base_url}/modelsync/status") |
| 93 | + response.raise_for_status() |
| 94 | + return response.json() |
| 95 | + |
| 96 | + def init_modelsync( |
| 97 | + self, |
| 98 | + user_name: str = "vLLM API User", |
| 99 | + user_email: str = "vllm@modelsync.local" |
| 100 | + ) -> Dict[str, Any]: |
| 101 | + """Inicializar repositorio ModelSync""" |
| 102 | + response = self.session.post( |
| 103 | + f"{self.base_url}/modelsync/init", |
| 104 | + params={"user_name": user_name, "user_email": user_email} |
| 105 | + ) |
| 106 | + response.raise_for_status() |
| 107 | + return response.json() |
| 108 | + |
| 109 | + def get_metrics(self) -> Dict[str, Any]: |
| 110 | + """Obtener métricas del servicio""" |
| 111 | + response = self.session.get(f"{self.base_url}/metrics") |
| 112 | + response.raise_for_status() |
| 113 | + return response.json() |
| 114 | + |
| 115 | +class VLLMExperimentManager: |
| 116 | + """Gestor de experimentos para vLLM con ModelSync""" |
| 117 | + |
| 118 | + def __init__(self, client: VLLMClient): |
| 119 | + self.client = client |
| 120 | + |
| 121 | + def run_experiment( |
| 122 | + self, |
| 123 | + experiment_name: str, |
| 124 | + prompts: List[str], |
| 125 | + parameters: Dict[str, Any], |
| 126 | + description: str = "" |
| 127 | + ) -> Dict[str, Any]: |
| 128 | + """Ejecutar un experimento con diferentes parámetros""" |
| 129 | + results = [] |
| 130 | + |
| 131 | + for i, prompt in enumerate(prompts): |
| 132 | + try: |
| 133 | + response = self.client.generate( |
| 134 | + prompt=prompt, |
| 135 | + **parameters, |
| 136 | + save_to_version_control=True |
| 137 | + ) |
| 138 | + |
| 139 | + results.append({ |
| 140 | + "prompt_index": i, |
| 141 | + "prompt": prompt, |
| 142 | + "response": response["text"], |
| 143 | + "model_version_id": response.get("model_version_id"), |
| 144 | + "usage": response["usage"], |
| 145 | + "created": response["created"] |
| 146 | + }) |
| 147 | + |
| 148 | + except Exception as e: |
| 149 | + results.append({ |
| 150 | + "prompt_index": i, |
| 151 | + "prompt": prompt, |
| 152 | + "error": str(e), |
| 153 | + "created": datetime.now().isoformat() |
| 154 | + }) |
| 155 | + |
| 156 | + # Crear resumen del experimento |
| 157 | + experiment_summary = { |
| 158 | + "experiment_name": experiment_name, |
| 159 | + "description": description, |
| 160 | + "parameters": parameters, |
| 161 | + "total_prompts": len(prompts), |
| 162 | + "successful_generations": len([r for r in results if "error" not in r]), |
| 163 | + "failed_generations": len([r for r in results if "error" in r]), |
| 164 | + "total_tokens": sum(r.get("usage", {}).get("total_tokens", 0) for r in results if "usage" in r), |
| 165 | + "results": results, |
| 166 | + "created": datetime.now().isoformat() |
| 167 | + } |
| 168 | + |
| 169 | + return experiment_summary |
| 170 | + |
| 171 | + def compare_parameters( |
| 172 | + self, |
| 173 | + base_prompts: List[str], |
| 174 | + parameter_sets: List[Dict[str, Any]], |
| 175 | + experiment_name: str = "parameter_comparison" |
| 176 | + ) -> Dict[str, Any]: |
| 177 | + """Comparar diferentes conjuntos de parámetros""" |
| 178 | + comparison_results = [] |
| 179 | + |
| 180 | + for i, params in enumerate(parameter_sets): |
| 181 | + param_name = f"param_set_{i+1}" |
| 182 | + experiment_result = self.run_experiment( |
| 183 | + experiment_name=f"{experiment_name}_{param_name}", |
| 184 | + prompts=base_prompts, |
| 185 | + parameters=params, |
| 186 | + description=f"Parameter set {i+1}: {params}" |
| 187 | + ) |
| 188 | + |
| 189 | + comparison_results.append({ |
| 190 | + "parameter_set": params, |
| 191 | + "parameter_name": param_name, |
| 192 | + "experiment_result": experiment_result |
| 193 | + }) |
| 194 | + |
| 195 | + return { |
| 196 | + "comparison_name": experiment_name, |
| 197 | + "base_prompts": base_prompts, |
| 198 | + "parameter_sets": parameter_sets, |
| 199 | + "results": comparison_results, |
| 200 | + "created": datetime.now().isoformat() |
| 201 | + } |
| 202 | + |
| 203 | +# Ejemplo de uso |
| 204 | +if __name__ == "__main__": |
| 205 | + client = VLLMClient() |
| 206 | + |
| 207 | + # Verificar salud |
| 208 | + health = client.health_check() |
| 209 | + print(f"Estado del servicio: {health}") |
| 210 | + |
| 211 | + # Verificar estado de ModelSync |
| 212 | + modelsync_status = client.get_modelsync_status() |
| 213 | + print(f"Estado ModelSync: {modelsync_status}") |
| 214 | + |
| 215 | + # Inicializar ModelSync si es necesario |
| 216 | + if modelsync_status["status"] == "not_initialized": |
| 217 | + init_result = client.init_modelsync() |
| 218 | + print(f"Inicialización ModelSync: {init_result}") |
| 219 | + |
| 220 | + # Generar texto |
| 221 | + response = client.generate( |
| 222 | + prompt="Hola, ¿cómo estás?", |
| 223 | + max_tokens=50, |
| 224 | + temperature=0.7 |
| 225 | + ) |
| 226 | + print(f"Respuesta: {response['text']}") |
| 227 | + print(f"Versión del modelo: {response.get('model_version_id', 'No guardado')}") |
| 228 | + |
| 229 | + # Ejemplo de experimento |
| 230 | + experiment_manager = VLLMExperimentManager(client) |
| 231 | + |
| 232 | + experiment_result = experiment_manager.run_experiment( |
| 233 | + experiment_name="test_experiment", |
| 234 | + prompts=[ |
| 235 | + "Cuéntame un chiste", |
| 236 | + "Explica qué es la inteligencia artificial", |
| 237 | + "¿Cuál es la capital de Francia?" |
| 238 | + ], |
| 239 | + parameters={ |
| 240 | + "max_tokens": 30, |
| 241 | + "temperature": 0.7, |
| 242 | + "top_p": 0.9 |
| 243 | + }, |
| 244 | + description="Experimento de prueba con diferentes prompts" |
| 245 | + ) |
| 246 | + |
| 247 | + print(f"Experimento completado: {experiment_result['successful_generations']}/{experiment_result['total_prompts']} exitosos") |
0 commit comments