| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | import contextlib |
| 4 | import os |
| 5 | import signal |
| 6 | import subprocess |
| 7 | import time |
| 8 | from types import TracebackType |
| 9 | |
| 10 | import requests |
| 11 | from typing_extensions import Self |
| 12 | |
| 13 | |
| 14 | class ServerProcess: |
| 15 | VLLM_RESET_CACHE_ENDPOINTS = [ |
| 16 | "/reset_prefix_cache", |
| 17 | "/reset_mm_cache", |
| 18 | "/reset_encoder_cache", |
| 19 | ] |
| 20 | |
| 21 | def __init__( |
| 22 | self, |
| 23 | server_cmd: list[str], |
| 24 | after_bench_cmd: list[str], |
| 25 | *, |
| 26 | show_stdout: bool, |
| 27 | ) -> None: |
| 28 | super().__init__() |
| 29 | |
| 30 | self.server_cmd = server_cmd |
| 31 | self.after_bench_cmd = after_bench_cmd |
| 32 | self.show_stdout = show_stdout |
| 33 | |
| 34 | def __enter__(self) -> Self: |
| 35 | self.start() |
| 36 | return self |
| 37 | |
| 38 | def __exit__( |
| 39 | self, |
| 40 | exc_type: type[BaseException] | None, |
| 41 | exc_value: BaseException | None, |
| 42 | exc_traceback: TracebackType | None, |
| 43 | ) -> None: |
| 44 | self.stop() |
| 45 | |
| 46 | def start(self): |
| 47 | # Create new process for clean termination |
| 48 | self._server_process = subprocess.Popen( |
| 49 | self.server_cmd, |
| 50 | start_new_session=True, |
| 51 | stdout=None if self.show_stdout else subprocess.DEVNULL, |
| 52 | # Need `VLLM_SERVER_DEV_MODE=1` for `_reset_caches` |
| 53 | env=os.environ | {"VLLM_SERVER_DEV_MODE": "1"}, |
| 54 | ) |
| 55 | |
| 56 | def stop(self): |
| 57 | server_process = self._server_process |
| 58 | |
| 59 | if server_process.poll() is None: |
| 60 | # In case only some processes have been terminated |
| 61 | with contextlib.suppress(ProcessLookupError): |
| 62 | # We need to kill both API Server and Engine processes |
| 63 | os.killpg(os.getpgid(server_process.pid), signal.SIGKILL) |
| 64 | |
| 65 | def run_subcommand(self, cmd: list[str]): |
| 66 | return subprocess.run( |
| 67 | cmd, |
| 68 | stdout=None if self.show_stdout else subprocess.DEVNULL, |
| 69 | check=True, |
| 70 | ) |
| 71 | |
| 72 | def after_bench(self) -> None: |
| 73 | if not self.after_bench_cmd: |
| 74 | self.reset_caches() |
| 75 | return |
| 76 | |
| 77 | self.run_subcommand(self.after_bench_cmd) |
| 78 | |
| 79 | def _get_vllm_server_address(self) -> str: |
| 80 | server_cmd = self.server_cmd |
| 81 | |
| 82 | for host_key in ("--host",): |
| 83 | if host_key in server_cmd: |
| 84 | host = server_cmd[server_cmd.index(host_key) + 1] |
| 85 | break |
| 86 | else: |
| 87 | host = "localhost" |
| 88 | |
| 89 | for port_key in ("-p", "--port"): |
| 90 | if port_key in server_cmd: |
| 91 | port = int(server_cmd[server_cmd.index(port_key) + 1]) |
| 92 | break |
| 93 | else: |
| 94 | port = 8000 # The default value in vllm serve |
| 95 | |
| 96 | return f"http://{host}:{port}" |
| 97 | |
| 98 | def is_server_ready(self) -> bool: |
| 99 | server_address = self._get_vllm_server_address() |
| 100 | try: |
| 101 | response = requests.get(f"{server_address}/health") |
| 102 | return response.status_code == 200 |
| 103 | except requests.RequestException: |
| 104 | return False |
| 105 | |
| 106 | def wait_until_ready(self, timeout: int) -> None: |
| 107 | start_time = time.monotonic() |
| 108 | while not self.is_server_ready(): |
| 109 | # Check if server process has crashed |
| 110 | if self._server_process.poll() is not None: |
| 111 | returncode = self._server_process.returncode |
| 112 | raise RuntimeError( |
| 113 | f"Server process crashed with return code {returncode}" |
| 114 | ) |
| 115 | if time.monotonic() - start_time > timeout: |
| 116 | raise TimeoutError( |
| 117 | f"Server failed to become ready within {timeout} seconds." |
| 118 | ) |
| 119 | time.sleep(1) |
| 120 | |
| 121 | def reset_caches(self) -> None: |
| 122 | server_cmd = self.server_cmd |
| 123 | |
| 124 | # Use `.endswith()` to match `/bin/...` |
| 125 | if server_cmd[0].endswith("vllm"): |
| 126 | server_address = self._get_vllm_server_address() |
| 127 | print(f"Resetting caches at {server_address}") |
| 128 | |
| 129 | for endpoint in self.VLLM_RESET_CACHE_ENDPOINTS: |
| 130 | res = requests.post(server_address + endpoint) |
| 131 | res.raise_for_status() |
| 132 | elif server_cmd[0].endswith("infinity_emb"): |
| 133 | if "--vector-disk-cache" in server_cmd: |
| 134 | raise NotImplementedError( |
| 135 | "Infinity server uses caching but does not expose a method " |
| 136 | "to reset the cache" |
| 137 | ) |
| 138 | else: |
| 139 | raise NotImplementedError( |
| 140 | f"No implementation of `reset_caches` for `{server_cmd[0]}` server. " |
| 141 | "Please specify a custom command via `--after-bench-cmd`." |
| 142 | ) |