| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import multiprocessing |
| 5 | import socket |
| 6 | import threading |
| 7 | import time |
| 8 | from types import SimpleNamespace |
| 9 | from unittest.mock import patch |
| 10 | |
| 11 | import pytest |
| 12 | import zmq |
| 13 | |
| 14 | from tests.entrypoints.launchers.api_server._api_server_spawn_workers import ( |
| 15 | exit_before_report_worker, |
| 16 | ) |
| 17 | from vllm.utils.network_utils import make_zmq_socket, split_zmq_path |
| 18 | from vllm.v1.engine.admission_control import SharedAdmissionStats |
| 19 | from vllm.v1.utils import ( |
| 20 | APIServerProcessManager, |
| 21 | get_engine_client_zmq_addr, |
| 22 | wait_for_completion_or_failure, |
| 23 | ) |
| 24 | |
| 25 | # Global variables to control worker behavior |
| 26 | WORKER_RUNTIME_SECONDS = 0.5 |
| 27 | |
| 28 | |
| 29 | # Mock implementation of run_api_server_worker |
| 30 | def mock_run_api_server_worker(listen_address, sock, args, client_config=None): |
| 31 | """Mock run_api_server_worker that runs for a specific time.""" |
| 32 | print(f"Mock worker started with client_config: {client_config}") |
| 33 | time.sleep(WORKER_RUNTIME_SECONDS) |
| 34 | print("Mock worker completed successfully") |
| 35 | |
| 36 | |
| 37 | def update_admission_stats_worker(listen_address, sock, args, client_config): |
| 38 | stats = SharedAdmissionStats( |
| 39 | client_config, |
| 40 | client_config["client_count"], |
| 41 | client_config["client_index"], |
| 42 | ) |
| 43 | value = client_config["client_index"] + 1 |
| 44 | stats.set_num_requests(value) |
| 45 | args.barrier.wait() |
| 46 | args.results.put(stats.get_num_requests()) |
| 47 | |
| 48 | |
| 49 | # Module-level stub for the gather_actual_addresses test. Must be |
| 50 | # importable by `multiprocessing.spawn` (no closures, no nesting). |
| 51 | def defer_addresses_stub_worker(listen_address, sock, args, client_config): |
| 52 | """Bind ROUTER/PULL with a kernel-assigned port, report the actual |
| 53 | endpoints back via the pipe, then exit.""" |
| 54 | ctx = zmq.Context() |
| 55 | try: |
| 56 | in_sock = make_zmq_socket( |
| 57 | ctx, client_config["input_address"], zmq.ROUTER, bind=True |
| 58 | ) |
| 59 | out_sock = make_zmq_socket( |
| 60 | ctx, client_config["output_address"], zmq.PULL, bind=True |
| 61 | ) |
| 62 | try: |
| 63 | pipe = client_config["actual_address_pipe"] |
| 64 | try: |
| 65 | pipe.send( |
| 66 | { |
| 67 | "input_address": in_sock.getsockopt(zmq.LAST_ENDPOINT).decode(), |
| 68 | "output_address": out_sock.getsockopt( |
| 69 | zmq.LAST_ENDPOINT |
| 70 | ).decode(), |
| 71 | } |
| 72 | ) |
| 73 | finally: |
| 74 | pipe.close() |
| 75 | finally: |
| 76 | in_sock.close(linger=0) |
| 77 | out_sock.close(linger=0) |
| 78 | finally: |
| 79 | ctx.term() |
| 80 | |
| 81 | |
| 82 | @pytest.fixture |
| 83 | def api_server_args(): |
| 84 | """Fixture to provide arguments for APIServerProcessManager.""" |
| 85 | sock = socket.socket() |
| 86 | return { |
| 87 | "target_server_fn": mock_run_api_server_worker, |
| 88 | "listen_address": "localhost:8000", |
| 89 | "sock": sock, |
| 90 | "args": "test_args", # Simple string to avoid pickling issues |
| 91 | "num_servers": 3, |
| 92 | "input_addresses": [ |
| 93 | "tcp://127.0.0.1:5001", |
| 94 | "tcp://127.0.0.1:5002", |
| 95 | "tcp://127.0.0.1:5003", |
| 96 | ], |
| 97 | "output_addresses": [ |
| 98 | "tcp://127.0.0.1:6001", |
| 99 | "tcp://127.0.0.1:6002", |
| 100 | "tcp://127.0.0.1:6003", |
| 101 | ], |
| 102 | "stats_update_address": "tcp://127.0.0.1:7000", |
| 103 | } |
| 104 | |
| 105 | |
| 106 | @pytest.mark.parametrize("with_stats_update", [True, False]) |
| 107 | def test_api_server_process_manager_init(api_server_args, with_stats_update): |
| 108 | """Test initializing the APIServerProcessManager.""" |
| 109 | # Set the worker runtime to ensure tests complete in reasonable time |
| 110 | global WORKER_RUNTIME_SECONDS |
| 111 | WORKER_RUNTIME_SECONDS = 0.5 |
| 112 | |
| 113 | # Copy the args to avoid mutating them |
| 114 | args = api_server_args.copy() |
| 115 | |
| 116 | if not with_stats_update: |
| 117 | args.pop("stats_update_address") |
| 118 | manager = APIServerProcessManager(**args) |
| 119 | |
| 120 | try: |
| 121 | # Verify the manager was initialized correctly |
| 122 | assert len(manager.processes) == 3 |
| 123 | |
| 124 | # Verify all processes are running |
| 125 | for proc in manager.processes: |
| 126 | assert proc.is_alive() |
| 127 | |
| 128 | print("Waiting for processes to run...") |
| 129 | time.sleep(WORKER_RUNTIME_SECONDS / 2) |
| 130 | |
| 131 | # They should still be alive at this point |
| 132 | for proc in manager.processes: |
| 133 | assert proc.is_alive() |
| 134 | |
| 135 | finally: |
| 136 | # Always clean up the processes |
| 137 | print("Cleaning up processes...") |
| 138 | manager.shutdown() |
| 139 | |
| 140 | # Give processes time to terminate |
| 141 | time.sleep(0.2) |
| 142 | |
| 143 | # Verify all processes were terminated |
| 144 | for proc in manager.processes: |
| 145 | assert not proc.is_alive() |
| 146 | |
| 147 | |
| 148 | @pytest.mark.skip_global_cleanup |
| 149 | def test_api_server_processes_share_admission_stats(api_server_args): |
| 150 | args = api_server_args.copy() |
| 151 | spawn_context = multiprocessing.get_context("spawn") |
| 152 | worker_args = SimpleNamespace( |
| 153 | barrier=spawn_context.Barrier(args["num_servers"]), |
| 154 | results=spawn_context.Queue(), |
| 155 | max_num_queued_reqs=10, |
| 156 | ) |
| 157 | args["args"] = worker_args |
| 158 | args["target_server_fn"] = update_admission_stats_worker |
| 159 | manager = APIServerProcessManager(**args) |
| 160 | |
| 161 | try: |
| 162 | for process in manager.processes: |
| 163 | process.join(timeout=30) |
| 164 | assert process.exitcode == 0 |
| 165 | |
| 166 | assert [worker_args.results.get(timeout=5) for _ in range(3)] == [6] * 3 |
| 167 | finally: |
| 168 | manager.shutdown() |
| 169 | worker_args.results.close() |
| 170 | worker_args.results.join_thread() |
| 171 | |
| 172 | |
| 173 | @patch("vllm.v1.utils.run_api_server_worker_proc", mock_run_api_server_worker) |
| 174 | def test_wait_for_completion_or_failure(api_server_args): |
| 175 | """Test that wait_for_completion_or_failure works with failures.""" |
| 176 | global WORKER_RUNTIME_SECONDS |
| 177 | WORKER_RUNTIME_SECONDS = 1.0 |
| 178 | |
| 179 | # Create the manager |
| 180 | manager = APIServerProcessManager(**api_server_args) |
| 181 | |
| 182 | try: |
| 183 | assert len(manager.processes) == 3 |
| 184 | |
| 185 | # Create a result capture for the thread |
| 186 | result: dict[str, Exception | None] = {"exception": None} |
| 187 | |
| 188 | def run_with_exception_capture(): |
| 189 | try: |
| 190 | wait_for_completion_or_failure(api_server_manager=manager) |
| 191 | except Exception as e: |
| 192 | result["exception"] = e |
| 193 | finally: |
| 194 | manager.shutdown() |
| 195 | |
| 196 | # Start a thread to run wait_for_completion_or_failure |
| 197 | wait_thread = threading.Thread(target=run_with_exception_capture, daemon=True) |
| 198 | wait_thread.start() |
| 199 | |
| 200 | # Let all processes run for a short time |
| 201 | time.sleep(0.2) |
| 202 | |
| 203 | # All processes should still be running |
| 204 | assert all(proc.is_alive() for proc in manager.processes) |
| 205 | |
| 206 | # Now simulate a process failure |
| 207 | print("Simulating process failure...") |
| 208 | manager.processes[0].terminate() |
| 209 | |
| 210 | # Wait for the wait_for_completion_or_failure |
| 211 | # to detect and handle the failure |
| 212 | # This should trigger it to terminate all other processes |
| 213 | wait_thread.join(timeout=1.0) |
| 214 | |
| 215 | # The wait thread should have exited |
| 216 | assert not wait_thread.is_alive() |
| 217 | |
| 218 | # Verify that an exception was raised with appropriate error message |
| 219 | assert result["exception"] is not None |
| 220 | assert "died with exit code" in str(result["exception"]) |
| 221 | |
| 222 | # All processes should now be terminated |
| 223 | for i, proc in enumerate(manager.processes): |
| 224 | assert not proc.is_alive(), f"Process {i} should not be alive" |
| 225 | |
| 226 | finally: |
| 227 | manager.shutdown() |
| 228 | time.sleep(0.2) |
| 229 | |
| 230 | |
| 231 | @pytest.mark.timeout(30) |
| 232 | def test_normal_completion(api_server_args): |
| 233 | """Test that wait_for_completion_or_failure works in normal completion.""" |
| 234 | global WORKER_RUNTIME_SECONDS |
| 235 | WORKER_RUNTIME_SECONDS = 0.1 |
| 236 | |
| 237 | # Create the manager |
| 238 | manager = APIServerProcessManager(**api_server_args) |
| 239 | |
| 240 | try: |
| 241 | # Give processes time to terminate |
| 242 | # wait for processes to complete |
| 243 | remaining_processes = manager.processes.copy() |
| 244 | while remaining_processes: |
| 245 | for proc in remaining_processes: |
| 246 | if not proc.is_alive(): |
| 247 | remaining_processes.remove(proc) |
| 248 | time.sleep(0.1) |
| 249 | |
| 250 | # Verify all processes have terminated |
| 251 | for i, proc in enumerate(manager.processes): |
| 252 | assert not proc.is_alive(), f"Process {i} still alive after terminate()" |
| 253 | |
| 254 | # Now call wait_for_completion_or_failure |
| 255 | # since all processes have already |
| 256 | # terminated, it should return immediately |
| 257 | # with no error |
| 258 | try: |
| 259 | wait_for_completion_or_failure(api_server_manager=manager) |
| 260 | finally: |
| 261 | manager.shutdown() |
| 262 | |
| 263 | finally: |
| 264 | # Clean up just in case |
| 265 | manager.shutdown() |
| 266 | time.sleep(0.2) |
| 267 | |
| 268 | |
| 269 | @pytest.mark.timeout(30) |
| 270 | def test_external_process_monitoring(api_server_args): |
| 271 | """Test that wait_for_completion_or_failure handles additional processes.""" |
| 272 | global WORKER_RUNTIME_SECONDS |
| 273 | prev_worker_runtime = WORKER_RUNTIME_SECONDS |
| 274 | WORKER_RUNTIME_SECONDS = 100 |
| 275 | |
| 276 | # Create and start the external process |
| 277 | # (simulates local_engine_manager or coordinator) |
| 278 | spawn_context = multiprocessing.get_context("spawn") |
| 279 | external_proc = spawn_context.Process( |
| 280 | target=mock_run_api_server_worker, name="MockExternalProcess" |
| 281 | ) |
| 282 | external_proc.start() |
| 283 | |
| 284 | # Create the class to simulate a coordinator |
| 285 | class MockCoordinator: |
| 286 | def __init__(self, proc): |
| 287 | self.proc = proc |
| 288 | |
| 289 | def shutdown(self): |
| 290 | if self.proc.is_alive(): |
| 291 | self.proc.terminate() |
| 292 | self.proc.join(timeout=0.5) |
| 293 | |
| 294 | # Create a mock coordinator with the external process |
| 295 | mock_coordinator = MockCoordinator(external_proc) |
| 296 | |
| 297 | # Create the API server manager |
| 298 | manager = APIServerProcessManager(**api_server_args) |
| 299 | |
| 300 | try: |
| 301 | # Verify manager initialization |
| 302 | assert len(manager.processes) == 3 |
| 303 | |
| 304 | # Create a result capture for the thread |
| 305 | result: dict[str, Exception | None] = {"exception": None} |
| 306 | |
| 307 | def run_with_exception_capture(): |
| 308 | try: |
| 309 | wait_for_completion_or_failure( |
| 310 | api_server_manager=manager, |
| 311 | coordinator=mock_coordinator, |
| 312 | ) |
| 313 | except Exception as e: |
| 314 | result["exception"] = e |
| 315 | finally: |
| 316 | manager.shutdown() |
| 317 | mock_coordinator.shutdown() |
| 318 | |
| 319 | # Start a thread to run wait_for_completion_or_failure |
| 320 | wait_thread = threading.Thread(target=run_with_exception_capture, daemon=True) |
| 321 | wait_thread.start() |
| 322 | |
| 323 | # Terminate the external process to trigger a failure |
| 324 | time.sleep(0.2) |
| 325 | external_proc.terminate() |
| 326 | |
| 327 | # Wait for the thread to detect the failure |
| 328 | wait_thread.join(timeout=1.0) |
| 329 | |
| 330 | # The wait thread should have completed |
| 331 | assert not wait_thread.is_alive(), ( |
| 332 | "wait_for_completion_or_failure thread still running" |
| 333 | ) |
| 334 | |
| 335 | # Verify that an exception was raised with appropriate error message |
| 336 | assert result["exception"] is not None, "No exception was raised" |
| 337 | error_message = str(result["exception"]) |
| 338 | assert "died with exit code" in error_message, ( |
| 339 | f"Unexpected error message: {error_message}" |
| 340 | ) |
| 341 | assert "MockExternalProcess" in error_message, ( |
| 342 | f"Error doesn't mention external process: {error_message}" |
| 343 | ) |
| 344 | |
| 345 | # Verify that all API server processes were terminated as a result |
| 346 | for i, proc in enumerate(manager.processes): |
| 347 | assert not proc.is_alive(), f"API server process {i} was not terminated" |
| 348 | |
| 349 | finally: |
| 350 | # Clean up |
| 351 | manager.shutdown() |
| 352 | mock_coordinator.shutdown() |
| 353 | time.sleep(0.2) |
| 354 | WORKER_RUNTIME_SECONDS = prev_worker_runtime |
| 355 | |
| 356 | |
| 357 | @pytest.mark.timeout(60) |
| 358 | def test_gather_actual_addresses_end_to_end(): |
| 359 | """Each child binds ROUTER/PULL with a kernel-picked port and reports |
| 360 | the bound endpoints back via its per-child pipe; the manager surfaces |
| 361 | them via :py:meth:`gather_actual_addresses`.""" |
| 362 | host = "127.0.0.1" |
| 363 | num_servers = 4 |
| 364 | |
| 365 | placeholder_inputs = [ |
| 366 | get_engine_client_zmq_addr(local_only=False, host=host) |
| 367 | for _ in range(num_servers) |
| 368 | ] |
| 369 | placeholder_outputs = [ |
| 370 | get_engine_client_zmq_addr(local_only=False, host=host) |
| 371 | for _ in range(num_servers) |
| 372 | ] |
| 373 | for addr in placeholder_inputs + placeholder_outputs: |
| 374 | assert addr == f"tcp://{host}:0", addr |
| 375 | |
| 376 | sock = socket.socket() |
| 377 | manager = APIServerProcessManager( |
| 378 | listen_address=f"tcp://{host}:0", |
| 379 | sock=sock, |
| 380 | args="test_args", |
| 381 | num_servers=num_servers, |
| 382 | input_addresses=placeholder_inputs, |
| 383 | output_addresses=placeholder_outputs, |
| 384 | target_server_fn=defer_addresses_stub_worker, |
| 385 | ) |
| 386 | |
| 387 | try: |
| 388 | assert len(manager.processes) == num_servers |
| 389 | actual_inputs, actual_outputs = manager.gather_actual_addresses(timeout=15.0) |
| 390 | finally: |
| 391 | manager.shutdown() |
| 392 | time.sleep(0.2) |
| 393 | sock.close() |
| 394 | |
| 395 | assert len(actual_inputs) == num_servers |
| 396 | assert len(actual_outputs) == num_servers |
| 397 | |
| 398 | for addr in actual_inputs + actual_outputs: |
| 399 | scheme, parsed_host, port = split_zmq_path(addr) |
| 400 | assert scheme == "tcp", addr |
| 401 | assert parsed_host == host, addr |
| 402 | assert port and int(port) > 0, addr |
| 403 | |
| 404 | all_addrs = actual_inputs + actual_outputs |
| 405 | assert len(set(all_addrs)) == len(all_addrs), all_addrs |
| 406 | |
| 407 | |
| 408 | @pytest.mark.timeout(30) |
| 409 | def test_gather_actual_addresses_child_crash_before_report(): |
| 410 | """A child that exits before sending its endpoints must surface a |
| 411 | clear ``RuntimeError`` rather than hang or return ``None`` slots.""" |
| 412 | host = "127.0.0.1" |
| 413 | num_servers = 2 |
| 414 | placeholder_inputs = [ |
| 415 | get_engine_client_zmq_addr(local_only=False, host=host) |
| 416 | for _ in range(num_servers) |
| 417 | ] |
| 418 | placeholder_outputs = [ |
| 419 | get_engine_client_zmq_addr(local_only=False, host=host) |
| 420 | for _ in range(num_servers) |
| 421 | ] |
| 422 | |
| 423 | sock = socket.socket() |
| 424 | manager = APIServerProcessManager( |
| 425 | listen_address=f"tcp://{host}:0", |
| 426 | sock=sock, |
| 427 | args="test_args", |
| 428 | num_servers=num_servers, |
| 429 | input_addresses=placeholder_inputs, |
| 430 | output_addresses=placeholder_outputs, |
| 431 | # exit_before_report_worker exits without touching |
| 432 | # ``actual_address_pipe`` — simulates a child that dies before |
| 433 | # reporting its bound addresses. |
| 434 | target_server_fn=exit_before_report_worker, |
| 435 | ) |
| 436 | try: |
| 437 | # Sentinel-first vs pipe-EOF-first both produce "reporting". |
| 438 | with pytest.raises(RuntimeError, match="reporting"): |
| 439 | manager.gather_actual_addresses(timeout=10.0) |
| 440 | finally: |
| 441 | manager.shutdown() |
| 442 | time.sleep(0.2) |
| 443 | sock.close() |
| 444 | |
| 445 | |
| 446 | def test_rust_frontend_launch_log_redacts_credentials(monkeypatch, caplog): |
| 447 | """The Rust frontend command carries every non-default arg as JSON. |
| 448 | Credentials must not reach the log line.""" |
| 449 | import subprocess as subprocess_mod |
| 450 | |
| 451 | from vllm.entrypoints.launchers.cli_args import make_arg_parser |
| 452 | from vllm.utils.argparse_utils import FlexibleArgumentParser |
| 453 | from vllm.v1.utils import RustFrontendProcessManager |
| 454 | |
| 455 | hf_token = "hf_TESTTOKEN123" |
| 456 | api_key = "sk-TESTKEY456" |
| 457 | args = make_arg_parser(FlexibleArgumentParser()).parse_args( |
| 458 | ["--model", "org/model", "--hf-token", hf_token, "--api-key", api_key] |
| 459 | ) |
| 460 | |
| 461 | class _FakeProc: |
| 462 | pid = 4321 |
| 463 | returncode = 0 |
| 464 | |
| 465 | def poll(self): |
| 466 | return 0 |
| 467 | |
| 468 | def wait(self, timeout=None): |
| 469 | return 0 |
| 470 | |
| 471 | sock = socket.socket(socket.AF_INET, socket.SOCK_STREAM) |
| 472 | try: |
| 473 | monkeypatch.setattr(subprocess_mod, "Popen", lambda *a, **kw: _FakeProc()) |
| 474 | with caplog.at_level("INFO", logger="vllm.v1.utils"): |
| 475 | RustFrontendProcessManager( |
| 476 | binary_path="/nonexistent/vllm-rs", |
| 477 | sock=sock, |
| 478 | args=args, |
| 479 | input_address="ipc:///tmp/in", |
| 480 | output_address="ipc:///tmp/out", |
| 481 | engine_start_index=0, |
| 482 | engine_count=1, |
| 483 | data_parallel_size=1, |
| 484 | ) |
| 485 | finally: |
| 486 | sock.close() |
| 487 | |
| 488 | message = caplog.text |
| 489 | assert "Launching Rust frontend:" in message |
| 490 | assert hf_token not in message |
| 491 | assert api_key not in message |
| 492 | assert '"hf_token": "***"' in message |
| 493 | |
| 494 | |
| 495 | def test_rust_frontend_uses_config_model_as_model_tag(monkeypatch, caplog, tmp_path): |
| 496 | """Config-only model selection must be forwarded to the Rust frontend.""" |
| 497 | import subprocess as subprocess_mod |
| 498 | |
| 499 | from vllm.entrypoints.launchers.cli_args import make_arg_parser |
| 500 | from vllm.utils.argparse_utils import FlexibleArgumentParser |
| 501 | from vllm.v1.utils import RustFrontendProcessManager |
| 502 | |
| 503 | config_path = tmp_path / "config.yml" |
| 504 | config_path.write_text("model: org/model\n") |
| 505 | args = make_arg_parser(FlexibleArgumentParser()).parse_args( |
| 506 | ["--config", str(config_path)] |
| 507 | ) |
| 508 | assert args.model == "org/model" |
| 509 | assert args.model_tag is None |
| 510 | |
| 511 | class _FakeProc: |
| 512 | pid = 4321 |
| 513 | returncode = 0 |
| 514 | |
| 515 | def poll(self): |
| 516 | return 0 |
| 517 | |
| 518 | def wait(self, timeout=None): |
| 519 | return 0 |
| 520 | |
| 521 | sock = socket.socket(socket.AF_INET, socket.SOCK_STREAM) |
| 522 | try: |
| 523 | monkeypatch.setattr(subprocess_mod, "Popen", lambda *a, **kw: _FakeProc()) |
| 524 | with caplog.at_level("INFO", logger="vllm.v1.utils"): |
| 525 | RustFrontendProcessManager( |
| 526 | binary_path="/nonexistent/vllm-rs", |
| 527 | sock=sock, |
| 528 | args=args, |
| 529 | input_address="ipc:///tmp/in", |
| 530 | output_address="ipc:///tmp/out", |
| 531 | engine_start_index=0, |
| 532 | engine_count=1, |
| 533 | data_parallel_size=1, |
| 534 | ) |
| 535 | finally: |
| 536 | sock.close() |
| 537 | |
| 538 | assert '"model_tag": "org/model"' in caplog.text |
| 539 | |
| 540 | |
| 541 | def test_rust_frontend_inherits_grpc_listener(monkeypatch): |
| 542 | """The Python-bound gRPC socket is passed to Rust like the HTTP one.""" |
| 543 | import subprocess as subprocess_mod |
| 544 | |
| 545 | from vllm.entrypoints.launchers.cli_args import make_arg_parser |
| 546 | from vllm.utils.argparse_utils import FlexibleArgumentParser |
| 547 | from vllm.v1.utils import RustFrontendProcessManager |
| 548 | |
| 549 | args = make_arg_parser(FlexibleArgumentParser()).parse_args( |
| 550 | ["--model", "org/model", "--grpc-port", "50051"] |
| 551 | ) |
| 552 | |
| 553 | class _FakeProc: |
| 554 | pid = 4321 |
| 555 | returncode = 0 |
| 556 | |
| 557 | def poll(self): |
| 558 | return 0 |
| 559 | |
| 560 | def wait(self, timeout=None): |
| 561 | return 0 |
| 562 | |
| 563 | popen_calls = [] |
| 564 | |
| 565 | def fake_popen(cmd, **kwargs): |
| 566 | popen_calls.append((cmd, kwargs)) |
| 567 | return _FakeProc() |
| 568 | |
| 569 | sock = socket.socket(socket.AF_INET, socket.SOCK_STREAM) |
| 570 | grpc_sock = socket.socket(socket.AF_INET, socket.SOCK_STREAM) |
| 571 | try: |
| 572 | monkeypatch.setattr(subprocess_mod, "Popen", fake_popen) |
| 573 | RustFrontendProcessManager( |
| 574 | binary_path="/nonexistent/vllm-rs", |
| 575 | sock=sock, |
| 576 | args=args, |
| 577 | input_address="ipc:///tmp/in", |
| 578 | output_address="ipc:///tmp/out", |
| 579 | engine_start_index=0, |
| 580 | engine_count=1, |
| 581 | data_parallel_size=1, |
| 582 | grpc_sock=grpc_sock, |
| 583 | ) |
| 584 | [(cmd, kwargs)] = popen_calls |
| 585 | grpc_fd = grpc_sock.fileno() |
| 586 | flag_index = cmd.index("--grpc-listen-fd") |
| 587 | assert cmd[flag_index + 1] == str(grpc_fd) |
| 588 | assert list(kwargs["pass_fds"]) == [sock.fileno(), grpc_fd] |
| 589 | finally: |
| 590 | sock.close() |
| 591 | grpc_sock.close() |