File size: 11,122 Bytes
5b42a0e | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 | """
vLLM-based model interface for high-performance LLM serving.
"""
import os
import logging
import subprocess
import time
import signal
import requests
from typing import List, Dict, Any, Optional, Union
from dataclasses import dataclass
from .base_model import BaseModel
from ..constants import SUPPORTED_MODELS, MODEL_METADATA, VLLM_DEFAULT_SETTINGS
logger = logging.getLogger(__name__)
@dataclass
class VLLMServerConfig:
"""Configuration for vLLM server."""
host: str = "localhost"
port: int = 8000
model: str = ""
max_model_len: int = 4096
gpu_memory_utilization: float = 0.9
dtype: str = "auto"
tensor_parallel_size: int = 1
trust_remote_code: bool = True
@property
def api_base(self) -> str:
return f"http://{self.host}:{self.port}/v1"
class VLLMServer:
"""
Manages a vLLM server instance for serving LLMs.
Usage:
server = VLLMServer(model_name="mistral-7b-instruct")
server.start()
# Use the server...
server.stop()
Or as context manager:
with VLLMServer(model_name="mistral-7b-instruct") as server:
# Use the server...
"""
def __init__(
self,
model_name: str,
host: str = "localhost",
port: int = 8000,
max_model_len: int = 4096,
gpu_memory_utilization: float = 0.9,
tensor_parallel_size: int = 1,
**kwargs
):
# Resolve model name to HuggingFace ID
if model_name in SUPPORTED_MODELS:
self.hf_model_id = SUPPORTED_MODELS[model_name]
self.model_name = model_name
else:
self.hf_model_id = model_name
self.model_name = model_name
self.config = VLLMServerConfig(
host=host,
port=port,
model=self.hf_model_id,
max_model_len=max_model_len,
gpu_memory_utilization=gpu_memory_utilization,
tensor_parallel_size=tensor_parallel_size,
)
self.process = None
self._started = False
def start(self, wait_for_ready: bool = True, timeout: int = 300) -> bool:
"""
Start the vLLM server.
Args:
wait_for_ready: Wait for server to be ready before returning
timeout: Maximum time to wait for server (seconds)
Returns:
True if server started successfully
"""
if self._started:
logger.warning("Server already started")
return True
cmd = [
"python", "-m", "vllm.entrypoints.openai.api_server",
"--model", self.config.model,
"--host", self.config.host,
"--port", str(self.config.port),
"--max-model-len", str(self.config.max_model_len),
"--gpu-memory-utilization", str(self.config.gpu_memory_utilization),
"--tensor-parallel-size", str(self.config.tensor_parallel_size),
]
if self.config.trust_remote_code:
cmd.append("--trust-remote-code")
logger.info(f"Starting vLLM server with command: {' '.join(cmd)}")
try:
self.process = subprocess.Popen(
cmd,
stdout=subprocess.PIPE,
stderr=subprocess.PIPE,
preexec_fn=os.setsid
)
if wait_for_ready:
return self._wait_for_ready(timeout)
self._started = True
return True
except Exception as e:
logger.error(f"Failed to start vLLM server: {e}")
return False
def _wait_for_ready(self, timeout: int = 300) -> bool:
"""Wait for server to be ready."""
start_time = time.time()
health_url = f"{self.config.api_base}/models"
while time.time() - start_time < timeout:
try:
response = requests.get(health_url, timeout=5)
if response.status_code == 200:
logger.info("vLLM server is ready!")
self._started = True
return True
except requests.exceptions.RequestException:
pass
# Check if process died
if self.process and self.process.poll() is not None:
stderr = self.process.stderr.read().decode() if self.process.stderr else ""
logger.error(f"vLLM server process died: {stderr}")
return False
time.sleep(2)
logger.info("Waiting for vLLM server to start...")
logger.error(f"vLLM server failed to start within {timeout} seconds")
return False
def stop(self):
"""Stop the vLLM server."""
if self.process:
try:
os.killpg(os.getpgid(self.process.pid), signal.SIGTERM)
self.process.wait(timeout=10)
except Exception as e:
logger.warning(f"Error stopping server: {e}")
try:
os.killpg(os.getpgid(self.process.pid), signal.SIGKILL)
except:
pass
finally:
self.process = None
self._started = False
logger.info("vLLM server stopped")
def is_running(self) -> bool:
"""Check if server is running."""
if not self._started:
return False
try:
response = requests.get(f"{self.config.api_base}/models", timeout=5)
return response.status_code == 200
except:
return False
def __enter__(self):
self.start()
return self
def __exit__(self, exc_type, exc_val, exc_tb):
self.stop()
class VLLMModel(BaseModel):
"""
vLLM-based model for LLM inference using OpenAI-compatible API.
Can connect to an existing vLLM server or manage its own.
Usage:
# Connect to existing server
model = VLLMModel(model_name="mistral-7b-instruct", api_base="http://localhost:8000/v1")
# Or with managed server
model = VLLMModel(model_name="mistral-7b-instruct", start_server=True)
"""
def __init__(
self,
model_name: str,
api_base: Optional[str] = None,
api_key: str = "EMPTY",
start_server: bool = False,
server_config: Optional[Dict] = None,
**kwargs
):
super().__init__(model_name)
# Resolve model name
if model_name in SUPPORTED_MODELS:
self.hf_model_id = SUPPORTED_MODELS[model_name]
else:
self.hf_model_id = model_name
self.api_key = api_key
self.server = None
# Start server if requested
if start_server:
config = server_config or {}
self.server = VLLMServer(model_name, **config)
self.server.start()
self.api_base = self.server.config.api_base
else:
self.api_base = api_base or "http://localhost:8000/v1"
# Get model metadata
self.metadata = MODEL_METADATA.get(model_name, {})
def generate(
self,
prompt: str,
max_tokens: int = 512,
temperature: float = 0.7,
top_p: float = 0.95,
stop: Optional[List[str]] = None,
**kwargs
) -> str:
"""Generate a response from the model."""
payload = {
"model": self.hf_model_id,
"prompt": prompt,
"max_tokens": max_tokens,
"temperature": temperature,
"top_p": top_p,
}
if stop:
payload["stop"] = stop
try:
response = requests.post(
f"{self.api_base}/completions",
json=payload,
headers={"Authorization": f"Bearer {self.api_key}"},
timeout=120
)
response.raise_for_status()
result = response.json()
return result["choices"][0]["text"].strip()
except Exception as e:
logger.error(f"Error generating response: {e}")
return ""
def generate_chat(
self,
messages: List[Dict[str, str]],
max_tokens: int = 512,
temperature: float = 0.7,
top_p: float = 0.95,
**kwargs
) -> str:
"""Generate a chat response."""
payload = {
"model": self.hf_model_id,
"messages": messages,
"max_tokens": max_tokens,
"temperature": temperature,
"top_p": top_p,
}
try:
response = requests.post(
f"{self.api_base}/chat/completions",
json=payload,
headers={"Authorization": f"Bearer {self.api_key}"},
timeout=120
)
response.raise_for_status()
result = response.json()
return result["choices"][0]["message"]["content"].strip()
except Exception as e:
logger.error(f"Error generating chat response: {e}")
return ""
def generate_batch(
self,
prompts: List[str],
max_tokens: int = 512,
temperature: float = 0.7,
**kwargs
) -> List[str]:
"""Generate responses for a batch of prompts."""
# vLLM handles batching internally, but we can also send multiple requests
responses = []
for prompt in prompts:
response = self.generate(prompt, max_tokens, temperature, **kwargs)
responses.append(response)
return responses
def get_response(
self,
idx: int,
stage: str,
messages: List[Dict[str, str]],
langcode: Optional[str] = None
) -> tuple:
"""
Get response compatible with the pipeline interface.
Returns:
Tuple of (response_string, cost)
"""
response = self.generate_chat(messages)
return response, 0.0 # vLLM is local, no cost
def __del__(self):
"""Cleanup server if managed."""
if self.server:
self.server.stop()
class VLLMModelFactory:
"""Factory for creating VLLMModel instances."""
@staticmethod
def create(
model_name: str,
api_base: Optional[str] = None,
**kwargs
) -> VLLMModel:
"""Create a VLLMModel instance."""
return VLLMModel(model_name, api_base=api_base, **kwargs)
@staticmethod
def list_models() -> List[str]:
"""List available models."""
return list(SUPPORTED_MODELS.keys())
@staticmethod
def get_model_info(model_name: str) -> Dict:
"""Get model metadata."""
return MODEL_METADATA.get(model_name, {})
|