| 1 | #!/usr/bin/env python3 |
| 2 | # SPDX-License-Identifier: Apache-2.0 |
| 3 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 4 | # mypy: ignore-errors |
| 5 | """vLLM gRPC Server |
| 6 | |
| 7 | Starts a gRPC server backed by AsyncLLM, using the VllmEngineServicer |
| 8 | from the smg-grpc-servicer package. |
| 9 | |
| 10 | Usage: |
| 11 | python -m vllm.entrypoints.launchers.grpc_server --model <model_path> |
| 12 | |
| 13 | Example: |
| 14 | python -m vllm.entrypoints.launchers.grpc_server \ |
| 15 | --model meta-llama/Llama-2-7b-hf \ |
| 16 | --host 0.0.0.0 \ |
| 17 | --port 50051 |
| 18 | |
| 19 | """ |
| 20 | |
| 21 | import argparse |
| 22 | import asyncio |
| 23 | import signal |
| 24 | import sys |
| 25 | import time |
| 26 | |
| 27 | try: |
| 28 | import grpc |
| 29 | from grpc_health.v1 import health_pb2_grpc |
| 30 | from grpc_reflection.v1alpha import reflection |
| 31 | from smg_grpc_proto import vllm_engine_pb2, vllm_engine_pb2_grpc |
| 32 | from smg_grpc_servicer.vllm.health_servicer import VllmHealthServicer |
| 33 | from smg_grpc_servicer.vllm.servicer import VllmEngineServicer |
| 34 | except ImportError as e: |
| 35 | raise ImportError( |
| 36 | "gRPC mode requires smg-grpc-servicer. " |
| 37 | "If not installed, run: pip install vllm[grpc]. " |
| 38 | "If already installed, there may be a broken import due to a " |
| 39 | "version mismatch — see the chained exception above for details." |
| 40 | ) from e |
| 41 | |
| 42 | import uvloop |
| 43 | |
| 44 | from vllm import envs |
| 45 | from vllm.engine.arg_utils import AsyncEngineArgs |
| 46 | from vllm.entrypoints.serve.utils.api_utils import log_version_and_model |
| 47 | from vllm.logger import configure_logging_from_args, init_logger |
| 48 | from vllm.usage.usage_lib import UsageContext |
| 49 | from vllm.utils.argparse_utils import FlexibleArgumentParser |
| 50 | from vllm.v1.engine.async_llm import AsyncLLM |
| 51 | from vllm.version import __version__ as VLLM_VERSION |
| 52 | |
| 53 | logger = init_logger(__name__) |
| 54 | |
| 55 | |
| 56 | async def serve_grpc(args: argparse.Namespace): |
| 57 | """Main gRPC serving function. |
| 58 | |
| 59 | Args: |
| 60 | args: Parsed command line arguments |
| 61 | |
| 62 | """ |
| 63 | log_version_and_model(logger, VLLM_VERSION, args.model) |
| 64 | logger.info("vLLM gRPC server args: %s", args) |
| 65 | |
| 66 | start_time = time.time() |
| 67 | |
| 68 | # Create engine args |
| 69 | engine_args = AsyncEngineArgs.from_cli_args(args) |
| 70 | |
| 71 | # Build vLLM config |
| 72 | vllm_config = engine_args.create_engine_config( |
| 73 | usage_context=UsageContext.OPENAI_API_SERVER, |
| 74 | ) |
| 75 | |
| 76 | # Create AsyncLLM |
| 77 | async_llm = AsyncLLM.from_vllm_config( |
| 78 | vllm_config=vllm_config, |
| 79 | usage_context=UsageContext.OPENAI_API_SERVER, |
| 80 | enable_log_requests=args.enable_log_requests, |
| 81 | disable_log_stats=args.disable_log_stats, |
| 82 | ) |
| 83 | |
| 84 | # Create servicer |
| 85 | servicer = VllmEngineServicer(async_llm, start_time) |
| 86 | |
| 87 | # Create gRPC server |
| 88 | server = grpc.aio.server( |
| 89 | options=[ |
| 90 | ("grpc.max_send_message_length", -1), |
| 91 | ("grpc.max_receive_message_length", -1), |
| 92 | # A non-streaming Generate writes nothing while the engine |
| 93 | # decodes, so client keepalive and BDP-probe PINGs arriving in |
| 94 | # that window are strikes and eventually earn a |
| 95 | # GOAWAY(ENHANCE_YOUR_CALM, "too_many_pings") that fails every |
| 96 | # in-flight RPC. Disable the strike count to tolerate them. |
| 97 | ("grpc.http2.max_ping_strikes", 0), |
| 98 | # GRPC_ARG_HTTP2_MIN_RECV_PING_INTERVAL_WITHOUT_DATA_MS expands to |
| 99 | # a string with no "recv", and gRPC Core silently ignores unknown |
| 100 | # keys, so the misspelt form left the 300s default in force. |
| 101 | ("grpc.http2.min_ping_interval_without_data_ms", 10000), |
| 102 | # Without this, gRPC Core widens the floor to a fixed two hours |
| 103 | # once no streams are open, striking keepalives between requests. |
| 104 | ("grpc.keepalive_permit_without_calls", True), |
| 105 | ], |
| 106 | ) |
| 107 | |
| 108 | # Add servicer to server |
| 109 | vllm_engine_pb2_grpc.add_VllmEngineServicer_to_server(servicer, server) |
| 110 | |
| 111 | # Add standard gRPC health service for Kubernetes probes |
| 112 | health_servicer = VllmHealthServicer(async_llm) |
| 113 | health_pb2_grpc.add_HealthServicer_to_server(health_servicer, server) |
| 114 | |
| 115 | # Enable reflection for grpcurl and other tools |
| 116 | service_names = ( |
| 117 | vllm_engine_pb2.DESCRIPTOR.services_by_name["VllmEngine"].full_name, |
| 118 | "grpc.health.v1.Health", |
| 119 | reflection.SERVICE_NAME, |
| 120 | ) |
| 121 | reflection.enable_server_reflection(service_names, server) |
| 122 | |
| 123 | # Bind to address |
| 124 | host = args.host or "0.0.0.0" |
| 125 | address = f"{host}:{args.port}" |
| 126 | server.add_insecure_port(address) |
| 127 | |
| 128 | try: |
| 129 | # Start server |
| 130 | await server.start() |
| 131 | logger.info("vLLM gRPC server started on %s", address) |
| 132 | logger.info("Server is ready to accept requests") |
| 133 | |
| 134 | # Start periodic stats logging (mirrors the HTTP server's lifespan task) |
| 135 | if not args.disable_log_stats: |
| 136 | |
| 137 | async def _force_log(): |
| 138 | while True: |
| 139 | await asyncio.sleep(envs.VLLM_LOG_STATS_INTERVAL) |
| 140 | await async_llm.do_log_stats() |
| 141 | |
| 142 | stats_task = asyncio.create_task(_force_log()) |
| 143 | else: |
| 144 | stats_task = None |
| 145 | |
| 146 | # Handle shutdown signals |
| 147 | loop = asyncio.get_running_loop() |
| 148 | stop_event = asyncio.Event() |
| 149 | |
| 150 | def signal_handler(): |
| 151 | logger.info("Received shutdown signal") |
| 152 | stop_event.set() |
| 153 | |
| 154 | for sig in (signal.SIGTERM, signal.SIGINT): |
| 155 | loop.add_signal_handler(sig, signal_handler) |
| 156 | |
| 157 | try: |
| 158 | await stop_event.wait() |
| 159 | except KeyboardInterrupt: |
| 160 | logger.info("Interrupted by user") |
| 161 | finally: |
| 162 | logger.info("Shutting down vLLM gRPC server...") |
| 163 | if stats_task is not None: |
| 164 | stats_task.cancel() |
| 165 | try: |
| 166 | health_servicer.set_not_serving() |
| 167 | except Exception: # broad: must not prevent server.stop() / shutdown() |
| 168 | logger.warning("Failed to set health status to NOT_SERVING", exc_info=True) |
| 169 | await server.stop(grace=5.0) |
| 170 | logger.info("gRPC server stopped") |
| 171 | async_llm.shutdown() |
| 172 | logger.info("AsyncLLM engine stopped") |
| 173 | logger.info("Shutdown complete") |
| 174 | |
| 175 | |
| 176 | def main(): |
| 177 | """Main entry point for python -m vllm.entrypoints.launchers.grpc_server.""" |
| 178 | parser = FlexibleArgumentParser( |
| 179 | description="vLLM gRPC Server", |
| 180 | ) |
| 181 | |
| 182 | # Server args |
| 183 | parser.add_argument( |
| 184 | "--host", |
| 185 | type=str, |
| 186 | default="0.0.0.0", |
| 187 | help="Host to bind gRPC server to", |
| 188 | ) |
| 189 | parser.add_argument( |
| 190 | "--port", |
| 191 | type=int, |
| 192 | default=50051, |
| 193 | help="Port to bind gRPC server to", |
| 194 | ) |
| 195 | parser = AsyncEngineArgs.add_cli_args(parser) |
| 196 | |
| 197 | args = parser.parse_args() |
| 198 | configure_logging_from_args(args) |
| 199 | |
| 200 | # Run server |
| 201 | try: |
| 202 | uvloop.run(serve_grpc(args)) |
| 203 | except Exception as e: |
| 204 | logger.exception("Server failed: %s", e) |
| 205 | sys.exit(1) |
| 206 | |
| 207 | |
| 208 | if __name__ == "__main__": |
| 209 | main() |