| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import datetime |
| 5 | import json |
| 6 | import logging |
| 7 | import os |
| 8 | import platform |
| 9 | import time |
| 10 | from enum import Enum |
| 11 | from pathlib import Path |
| 12 | from threading import Thread |
| 13 | from typing import Any |
| 14 | from uuid import uuid4 |
| 15 | |
| 16 | import cpuinfo |
| 17 | import psutil |
| 18 | import requests |
| 19 | import torch |
| 20 | |
| 21 | import vllm.envs as envs |
| 22 | from vllm.logger import init_logger |
| 23 | from vllm.utils.platform_utils import cuda_get_device_properties |
| 24 | from vllm.version import __version__ as VLLM_VERSION |
| 25 | |
| 26 | logger = init_logger(__name__) |
| 27 | |
| 28 | _config_home = envs.VLLM_CONFIG_ROOT |
| 29 | _USAGE_STATS_JSON_PATH = os.path.join(_config_home, "usage_stats.json") |
| 30 | _USAGE_STATS_DO_NOT_TRACK_PATH = os.path.join(_config_home, "do_not_track") |
| 31 | _USAGE_STATS_ENABLED = None |
| 32 | _USAGE_STATS_SERVER = envs.VLLM_USAGE_STATS_SERVER |
| 33 | |
| 34 | _GLOBAL_RUNTIME_DATA = dict[str, str | int | bool]() |
| 35 | |
| 36 | _USAGE_ENV_VARS_TO_COLLECT = [ |
| 37 | "VLLM_USE_MODELSCOPE", |
| 38 | "VLLM_USE_FLASHINFER_SAMPLER", |
| 39 | "VLLM_PP_LAYER_PARTITION", |
| 40 | "VLLM_USE_TRITON_AWQ", |
| 41 | "VLLM_ENABLE_V1_MULTIPROCESSING", |
| 42 | ] |
| 43 | |
| 44 | |
| 45 | def set_runtime_usage_data(key: str, value: str | int | bool) -> None: |
| 46 | """Set global usage data that will be sent with every usage heartbeat.""" |
| 47 | _GLOBAL_RUNTIME_DATA[key] = value |
| 48 | |
| 49 | |
| 50 | def is_usage_stats_enabled(): |
| 51 | """Determine whether or not we can send usage stats to the server. |
| 52 | The logic is as follows: |
| 53 | - By default, it should be enabled. |
| 54 | - Three environment variables can disable it: |
| 55 | - VLLM_DO_NOT_TRACK=1 |
| 56 | - DO_NOT_TRACK=1 |
| 57 | - VLLM_NO_USAGE_STATS=1 |
| 58 | - A file in the home directory can disable it if it exists: |
| 59 | - $HOME/.config/vllm/do_not_track |
| 60 | """ |
| 61 | global _USAGE_STATS_ENABLED |
| 62 | if _USAGE_STATS_ENABLED is None: |
| 63 | do_not_track = envs.VLLM_DO_NOT_TRACK |
| 64 | no_usage_stats = envs.VLLM_NO_USAGE_STATS |
| 65 | do_not_track_file = os.path.exists(_USAGE_STATS_DO_NOT_TRACK_PATH) |
| 66 | |
| 67 | _USAGE_STATS_ENABLED = not (do_not_track or no_usage_stats or do_not_track_file) |
| 68 | return _USAGE_STATS_ENABLED |
| 69 | |
| 70 | |
| 71 | def _get_current_timestamp_ns() -> int: |
| 72 | return int(datetime.datetime.now(datetime.timezone.utc).timestamp() * 1e9) |
| 73 | |
| 74 | |
| 75 | def _detect_cloud_provider() -> str: |
| 76 | # Try detecting through vendor file |
| 77 | vendor_files = [ |
| 78 | "/sys/class/dmi/id/product_version", |
| 79 | "/sys/class/dmi/id/bios_vendor", |
| 80 | "/sys/class/dmi/id/product_name", |
| 81 | "/sys/class/dmi/id/chassis_asset_tag", |
| 82 | "/sys/class/dmi/id/sys_vendor", |
| 83 | ] |
| 84 | # Mapping of identifiable strings to cloud providers |
| 85 | cloud_identifiers = { |
| 86 | "amazon": "AWS", |
| 87 | "microsoft corporation": "AZURE", |
| 88 | "google": "GCP", |
| 89 | "oraclecloud": "OCI", |
| 90 | } |
| 91 | |
| 92 | for vendor_file in vendor_files: |
| 93 | path = Path(vendor_file) |
| 94 | if path.is_file(): |
| 95 | file_content = path.read_text().lower() |
| 96 | for identifier, provider in cloud_identifiers.items(): |
| 97 | if identifier in file_content: |
| 98 | return provider |
| 99 | |
| 100 | # Try detecting through environment variables |
| 101 | env_to_cloud_provider = { |
| 102 | "RUNPOD_DC_ID": "RUNPOD", |
| 103 | } |
| 104 | for env_var, provider in env_to_cloud_provider.items(): |
| 105 | if os.environ.get(env_var): |
| 106 | return provider |
| 107 | |
| 108 | return "UNKNOWN" |
| 109 | |
| 110 | |
| 111 | class UsageContext(str, Enum): |
| 112 | UNKNOWN_CONTEXT = "UNKNOWN_CONTEXT" |
| 113 | LLM_CLASS = "LLM_CLASS" |
| 114 | API_SERVER = "API_SERVER" |
| 115 | OPENAI_API_SERVER = "OPENAI_API_SERVER" |
| 116 | OPENAI_BATCH_RUNNER = "OPENAI_BATCH_RUNNER" |
| 117 | ENGINE_CONTEXT = "ENGINE_CONTEXT" |
| 118 | |
| 119 | |
| 120 | class UsageMessage: |
| 121 | """Collect platform information and send it to the usage stats server.""" |
| 122 | |
| 123 | def __init__(self) -> None: |
| 124 | # NOTE: vLLM's server _only_ support flat KV pair. |
| 125 | # Do not use nested fields. |
| 126 | |
| 127 | self.uuid = str(uuid4()) |
| 128 | |
| 129 | # Environment Information |
| 130 | self.provider: str | None = None |
| 131 | self.num_cpu: int | None = None |
| 132 | self.cpu_type: str | None = None |
| 133 | self.cpu_family_model_stepping: str | None = None |
| 134 | self.total_memory: int | None = None |
| 135 | self.architecture: str | None = None |
| 136 | self.platform: str | None = None |
| 137 | self.xpu_runtime: str | None = None |
| 138 | self.cuda_runtime: str | None = None |
| 139 | self.gpu_count: int | None = None |
| 140 | self.gpu_type: str | None = None |
| 141 | self.gpu_memory_per_device: int | None = None |
| 142 | self.env_var_json: str | None = None |
| 143 | |
| 144 | # vLLM Information |
| 145 | self.model_architecture: str | None = None |
| 146 | self.vllm_version: str | None = None |
| 147 | self.context: str | None = None |
| 148 | |
| 149 | # Metadata |
| 150 | self.log_time: int | None = None |
| 151 | self.source: str | None = None |
| 152 | |
| 153 | def report_usage( |
| 154 | self, |
| 155 | model_architecture: str, |
| 156 | usage_context: UsageContext, |
| 157 | extra_kvs: dict[str, Any] | None = None, |
| 158 | ) -> None: |
| 159 | t = Thread( |
| 160 | target=self._report_usage_worker, |
| 161 | args=(model_architecture, usage_context, extra_kvs or {}), |
| 162 | daemon=True, |
| 163 | ) |
| 164 | t.start() |
| 165 | |
| 166 | def _report_usage_worker( |
| 167 | self, |
| 168 | model_architecture: str, |
| 169 | usage_context: UsageContext, |
| 170 | extra_kvs: dict[str, Any], |
| 171 | ) -> None: |
| 172 | self._report_usage_once(model_architecture, usage_context, extra_kvs) |
| 173 | self._report_continuous_usage() |
| 174 | |
| 175 | def _report_tpu_inference_usage(self) -> bool: |
| 176 | try: |
| 177 | from tpu_inference import tpu_info, utils |
| 178 | |
| 179 | self.gpu_count = tpu_info.get_num_chips() |
| 180 | self.gpu_type = tpu_info.get_tpu_type() |
| 181 | self.gpu_memory_per_device = utils.get_device_hbm_limit() |
| 182 | self.cuda_runtime = "tpu_inference" |
| 183 | return True |
| 184 | except Exception: |
| 185 | return False |
| 186 | |
| 187 | def _report_usage_once( |
| 188 | self, |
| 189 | model_architecture: str, |
| 190 | usage_context: UsageContext, |
| 191 | extra_kvs: dict[str, Any], |
| 192 | ) -> None: |
| 193 | # Platform information |
| 194 | from vllm.platforms import current_platform |
| 195 | |
| 196 | if current_platform.is_cuda_alike(): |
| 197 | self.gpu_count = current_platform.device_count() |
| 198 | self.gpu_type, self.gpu_memory_per_device = cuda_get_device_properties( |
| 199 | 0, ("name", "total_memory") |
| 200 | ) |
| 201 | if current_platform.is_cuda(): |
| 202 | self.cuda_runtime = torch.version.cuda |
| 203 | if current_platform.is_xpu(): |
| 204 | self.xpu_runtime = torch.version.xpu |
| 205 | self.gpu_count = torch.xpu.device_count() |
| 206 | self.gpu_type = torch.xpu.get_device_name(0) |
| 207 | self.gpu_memory_per_device = torch.xpu.get_device_properties(0).total_memory |
| 208 | if current_platform.is_tpu(): # noqa: SIM102 |
| 209 | if not self._report_tpu_inference_usage(): |
| 210 | logger.exception("Failed to collect TPU information") |
| 211 | self.provider = _detect_cloud_provider() |
| 212 | self.architecture = platform.machine() |
| 213 | self.platform = platform.platform() |
| 214 | self.total_memory = psutil.virtual_memory().total |
| 215 | |
| 216 | info = cpuinfo.get_cpu_info() |
| 217 | self.num_cpu = info.get("count", None) |
| 218 | self.cpu_type = info.get("brand_raw", "") |
| 219 | self.cpu_family_model_stepping = ",".join( |
| 220 | [ |
| 221 | str(info.get("family", "")), |
| 222 | str(info.get("model", "")), |
| 223 | str(info.get("stepping", "")), |
| 224 | ] |
| 225 | ) |
| 226 | |
| 227 | # vLLM information |
| 228 | self.context = usage_context.value |
| 229 | self.vllm_version = VLLM_VERSION |
| 230 | self.model_architecture = model_architecture |
| 231 | |
| 232 | # Environment variables |
| 233 | self.env_var_json = json.dumps( |
| 234 | {env_var: getattr(envs, env_var) for env_var in _USAGE_ENV_VARS_TO_COLLECT} |
| 235 | ) |
| 236 | |
| 237 | # Metadata |
| 238 | self.log_time = _get_current_timestamp_ns() |
| 239 | self.source = envs.VLLM_USAGE_SOURCE |
| 240 | |
| 241 | data = vars(self) |
| 242 | if extra_kvs: |
| 243 | data.update(extra_kvs) |
| 244 | |
| 245 | self._write_to_file(data) |
| 246 | self._send_to_server(data) |
| 247 | |
| 248 | def _report_continuous_usage(self): |
| 249 | """Report usage every 10 minutes. |
| 250 | |
| 251 | This helps us to collect more data points for uptime of vLLM usages. |
| 252 | This function can also help send over performance metrics over time. |
| 253 | """ |
| 254 | while True: |
| 255 | time.sleep(600) |
| 256 | data = { |
| 257 | "uuid": self.uuid, |
| 258 | "log_time": _get_current_timestamp_ns(), |
| 259 | } |
| 260 | data.update(_GLOBAL_RUNTIME_DATA) |
| 261 | |
| 262 | self._write_to_file(data) |
| 263 | self._send_to_server(data) |
| 264 | |
| 265 | def _send_to_server(self, data: dict[str, Any]) -> None: |
| 266 | try: |
| 267 | with requests.Session() as client: |
| 268 | client.post(_USAGE_STATS_SERVER, json=data) |
| 269 | except requests.exceptions.RequestException: |
| 270 | # silently ignore unless we are using debug log |
| 271 | logging.debug("Failed to send usage data to server") |
| 272 | |
| 273 | def _write_to_file(self, data: dict[str, Any]) -> None: |
| 274 | os.makedirs(os.path.dirname(_USAGE_STATS_JSON_PATH), exist_ok=True) |
| 275 | Path(_USAGE_STATS_JSON_PATH).touch(exist_ok=True) |
| 276 | with open(_USAGE_STATS_JSON_PATH, "a") as f: |
| 277 | json.dump(data, f) |
| 278 | f.write("\n") |
| 279 | |
| 280 | |
| 281 | usage_message = UsageMessage() |