| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | import asyncio |
| 4 | import os |
| 5 | import threading |
| 6 | import time |
| 7 | import traceback |
| 8 | from typing import cast |
| 9 | |
| 10 | import openai # use the official client for correctness check |
| 11 | import pytest |
| 12 | import pytest_asyncio |
| 13 | import requests |
| 14 | |
| 15 | from tests.utils import RemoteOpenAIServer |
| 16 | from tests.v1.utils import check_request_balancing |
| 17 | from vllm.platforms import current_platform |
| 18 | |
| 19 | MODEL_NAME = "ibm-research/PowerMoE-3b" |
| 20 | |
| 21 | # Number of data parallel ranks for multi-node internal LB testing |
| 22 | DP_SIZE = int(os.getenv("DP_SIZE", "2")) |
| 23 | # Default tensor parallel size to use |
| 24 | TP_SIZE = int(os.getenv("TP_SIZE", "1")) |
| 25 | |
| 26 | # Number of nodes to simulate |
| 27 | NUM_NODES = 2 |
| 28 | |
| 29 | |
| 30 | async def _make_completion_request( |
| 31 | client: openai.AsyncOpenAI, |
| 32 | model_name: str, |
| 33 | ) -> openai.types.Completion: |
| 34 | """Make a single completion request and validate the response. |
| 35 | |
| 36 | Uses temperature=1.0 to ensure diverse outputs across concurrent |
| 37 | requests for realistic load balancer testing. |
| 38 | """ |
| 39 | completion = await client.completions.create( |
| 40 | model=model_name, |
| 41 | prompt="Hello, my name is", |
| 42 | max_tokens=5, |
| 43 | temperature=1.0, |
| 44 | ) |
| 45 | |
| 46 | assert completion.id is not None, ( |
| 47 | f"Expected non-None completion id. usage={completion.usage!r}" |
| 48 | ) |
| 49 | assert completion.choices is not None and len(completion.choices) == 1, ( |
| 50 | f"Expected 1 choice, got " |
| 51 | f"{len(completion.choices) if completion.choices else 'None'}" |
| 52 | ) |
| 53 | |
| 54 | choice = completion.choices[0] |
| 55 | # With temperature=1.0, the model may emit a stop token immediately, |
| 56 | # producing empty text with finish_reason='stop'. This is valid |
| 57 | # model behavior - the test's purpose is load balancing, not output |
| 58 | # quality. |
| 59 | assert choice.finish_reason in ("length", "stop"), ( |
| 60 | f"Expected finish_reason 'length' or 'stop', " |
| 61 | f"got {choice.finish_reason!r}. text={choice.text!r}" |
| 62 | ) |
| 63 | if choice.finish_reason == "length": |
| 64 | assert len(choice.text) >= 1, ( |
| 65 | f"Expected non-empty text with finish_reason='length', got {choice.text!r}" |
| 66 | ) |
| 67 | |
| 68 | assert completion.usage.prompt_tokens > 0, ( |
| 69 | f"Expected positive prompt_tokens, got {completion.usage.prompt_tokens}" |
| 70 | ) |
| 71 | assert completion.usage.total_tokens > 0, ( |
| 72 | f"Expected positive total_tokens, got {completion.usage.total_tokens}" |
| 73 | ) |
| 74 | return completion |
| 75 | |
| 76 | |
| 77 | async def _run_request_bursts( |
| 78 | client: openai.AsyncOpenAI, |
| 79 | model_name: str, |
| 80 | num_requests: int = 200, |
| 81 | num_bursts: int = 2, |
| 82 | ): |
| 83 | """Send multiple bursts of completion requests and validate all succeed.""" |
| 84 | for burst in range(num_bursts): |
| 85 | all_tasks = [] |
| 86 | for _ in range(num_requests): |
| 87 | all_tasks.append( |
| 88 | asyncio.create_task(_make_completion_request(client, model_name)) |
| 89 | ) |
| 90 | await asyncio.sleep(0.01) |
| 91 | |
| 92 | results = await asyncio.gather(*all_tasks, return_exceptions=True) |
| 93 | assert len(results) == num_requests, ( |
| 94 | f"Burst {burst}: expected {num_requests} results, got {len(results)}" |
| 95 | ) |
| 96 | |
| 97 | for result in results: |
| 98 | if isinstance(result, BaseException): |
| 99 | raise result |
| 100 | |
| 101 | assert all(completion is not None for completion in results), ( |
| 102 | f"Burst {burst}: some completions were None" |
| 103 | ) |
| 104 | |
| 105 | await asyncio.sleep(0.5) |
| 106 | |
| 107 | |
| 108 | class MultinodeInternalLBServerManager: |
| 109 | """Manages multi-node data parallel vLLM server instances for internal |
| 110 | load balancer testing using --headless mode.""" |
| 111 | |
| 112 | def __init__( |
| 113 | self, |
| 114 | model_name: str, |
| 115 | dp_size: int, |
| 116 | api_server_count: int, |
| 117 | base_server_args: list, |
| 118 | dp_per_node: int = 1, |
| 119 | tp_size: int = TP_SIZE, |
| 120 | ): |
| 121 | self.model_name = model_name |
| 122 | self.dp_size = dp_size |
| 123 | self.dp_per_node = dp_per_node |
| 124 | self.tp_size = tp_size |
| 125 | self.api_server_count = api_server_count |
| 126 | self.base_server_args = base_server_args |
| 127 | self.servers: list[tuple[RemoteOpenAIServer, list[str]] | None] = [None] * ( |
| 128 | dp_size // dp_per_node |
| 129 | ) |
| 130 | self.server_threads: list[threading.Thread] = [] |
| 131 | |
| 132 | def __enter__(self) -> list[tuple[RemoteOpenAIServer, list[str]]]: |
| 133 | """Start all server instances for multi-node internal LB mode.""" |
| 134 | for server_idx, rank in enumerate(range(0, self.dp_size, self.dp_per_node)): |
| 135 | # Create server args for this specific rank |
| 136 | server_args = self.base_server_args.copy() |
| 137 | |
| 138 | if rank == 0: |
| 139 | # Head node - runs API server and first DP rank |
| 140 | server_args.extend( |
| 141 | [ |
| 142 | "--data-parallel-size", |
| 143 | str(self.dp_size), |
| 144 | "--data-parallel-size-local", |
| 145 | str(self.dp_per_node), |
| 146 | "--tensor-parallel-size", |
| 147 | str(self.tp_size), |
| 148 | "--port", |
| 149 | "8000", # Single endpoint for all requests |
| 150 | "--api-server-count", |
| 151 | str(self.api_server_count), |
| 152 | "--data-parallel-address", |
| 153 | "127.0.0.1", |
| 154 | "--data-parallel-rpc-port", |
| 155 | "13345", |
| 156 | ] |
| 157 | ) |
| 158 | else: |
| 159 | # Secondary nodes - run in headless mode |
| 160 | server_args.extend( |
| 161 | [ |
| 162 | "--headless", |
| 163 | "--data-parallel-size", |
| 164 | str(self.dp_size), |
| 165 | "--data-parallel-size-local", |
| 166 | str(self.dp_per_node), |
| 167 | "--data-parallel-start-rank", |
| 168 | str(rank), |
| 169 | "--tensor-parallel-size", |
| 170 | str(self.tp_size), |
| 171 | "--data-parallel-address", |
| 172 | "127.0.0.1", |
| 173 | "--data-parallel-rpc-port", |
| 174 | "13345", |
| 175 | ] |
| 176 | ) |
| 177 | |
| 178 | # Use a thread to start each server to allow parallel initialization |
| 179 | def start_server(sidx: int, r: int, sargs: list[str]): |
| 180 | gpus_per_node = self.tp_size * self.dp_per_node |
| 181 | try: |
| 182 | # Start the server |
| 183 | server = RemoteOpenAIServer( |
| 184 | self.model_name, |
| 185 | sargs, |
| 186 | auto_port=False, |
| 187 | max_wait_seconds=int( |
| 188 | os.getenv("VLLM_ENGINE_READY_TIMEOUT_S", "480") |
| 189 | ), |
| 190 | env_dict={ |
| 191 | "VLLM_SERVER_DEV_MODE": "1", |
| 192 | current_platform.device_control_env_var: ",".join( |
| 193 | str(current_platform.device_id_to_physical_device_id(i)) |
| 194 | for i in range(r, r + gpus_per_node) |
| 195 | ), |
| 196 | }, |
| 197 | ) |
| 198 | server.__enter__() |
| 199 | if r == 0: |
| 200 | print( |
| 201 | f"Head node (rank {r}) started successfully with " |
| 202 | f"{self.api_server_count} API servers" |
| 203 | ) |
| 204 | else: |
| 205 | print(f"Headless node (rank {r}) started successfully") |
| 206 | self.servers[sidx] = (server, sargs) |
| 207 | except Exception as e: |
| 208 | print(f"Failed to start server rank {r}: {e}") |
| 209 | traceback.print_exc() |
| 210 | raise |
| 211 | |
| 212 | thread = threading.Thread( |
| 213 | target=start_server, args=(server_idx, rank, server_args) |
| 214 | ) |
| 215 | thread.start() |
| 216 | |
| 217 | self.server_threads.append(thread) |
| 218 | |
| 219 | # Wait for all servers to start |
| 220 | for thread in self.server_threads: |
| 221 | thread.join() |
| 222 | |
| 223 | # Give servers additional time to fully initialize and coordinate |
| 224 | time.sleep(3) |
| 225 | |
| 226 | if not all(self.servers): |
| 227 | raise Exception("Servers failed to start") |
| 228 | |
| 229 | return cast(list[tuple[RemoteOpenAIServer, list[str]]], self.servers) |
| 230 | |
| 231 | def __exit__(self, exc_type, exc_val, exc_tb): |
| 232 | """Stop all server instances.""" |
| 233 | servers = [entry[0] for entry in self.servers if entry is not None] |
| 234 | self.servers.clear() |
| 235 | try: |
| 236 | RemoteOpenAIServer.shutdown_many(servers) |
| 237 | except Exception as e: |
| 238 | print(f"Error stopping servers: {e}") |
| 239 | traceback.print_exc() |
| 240 | |
| 241 | |
| 242 | class APIOnlyServerManager: |
| 243 | """Manages API-only server (Node 0) and headless engines server (Node 1) |
| 244 | for testing separated API server and engine configuration.""" |
| 245 | |
| 246 | def __init__( |
| 247 | self, |
| 248 | model_name: str, |
| 249 | dp_size: int, |
| 250 | api_server_count: int, |
| 251 | base_server_args: list, |
| 252 | tp_size: int = TP_SIZE, |
| 253 | ): |
| 254 | self.model_name = model_name |
| 255 | self.dp_size = dp_size |
| 256 | self.tp_size = tp_size |
| 257 | self.api_server_count = api_server_count |
| 258 | self.base_server_args = base_server_args |
| 259 | self.servers: list[tuple[RemoteOpenAIServer, list[str]] | None] = [None] * 2 |
| 260 | self.server_threads: list[threading.Thread] = [] |
| 261 | |
| 262 | def __enter__(self) -> list[tuple[RemoteOpenAIServer, list[str]]]: |
| 263 | """Start API-only server and headless engines server.""" |
| 264 | # Start API-only server (Node 0) - no engines, only API server |
| 265 | api_server_args = self.base_server_args.copy() |
| 266 | api_server_args.extend( |
| 267 | [ |
| 268 | "--data-parallel-size", |
| 269 | str(self.dp_size), |
| 270 | "--data-parallel-size-local", |
| 271 | "0", # No engines on this node |
| 272 | "--tensor-parallel-size", |
| 273 | str(self.tp_size), |
| 274 | "--port", |
| 275 | "8000", |
| 276 | "--api-server-count", |
| 277 | str(self.api_server_count), |
| 278 | "--data-parallel-address", |
| 279 | "127.0.0.1", |
| 280 | "--data-parallel-rpc-port", |
| 281 | "13345", |
| 282 | ] |
| 283 | ) |
| 284 | |
| 285 | # Start headless engines server (Node 1) - all engines, no API server |
| 286 | engines_server_args = self.base_server_args.copy() |
| 287 | engines_server_args.extend( |
| 288 | [ |
| 289 | "--headless", |
| 290 | "--data-parallel-size", |
| 291 | str(self.dp_size), |
| 292 | "--data-parallel-size-local", |
| 293 | str(self.dp_size), # All engines on this node |
| 294 | "--tensor-parallel-size", |
| 295 | str(self.tp_size), |
| 296 | "--data-parallel-address", |
| 297 | "127.0.0.1", |
| 298 | "--data-parallel-rpc-port", |
| 299 | "13345", |
| 300 | ] |
| 301 | ) |
| 302 | |
| 303 | # Use threads to start both servers in parallel |
| 304 | def start_api_server(): |
| 305 | try: |
| 306 | server = RemoteOpenAIServer( |
| 307 | self.model_name, |
| 308 | api_server_args, |
| 309 | auto_port=False, |
| 310 | env_dict={ |
| 311 | "VLLM_SERVER_DEV_MODE": "1", |
| 312 | # No GPUs needed for API-only server |
| 313 | }, |
| 314 | ) |
| 315 | server.__enter__() |
| 316 | print( |
| 317 | f"API-only server started successfully with " |
| 318 | f"{self.api_server_count} API servers" |
| 319 | ) |
| 320 | self.servers[0] = (server, api_server_args) |
| 321 | except Exception as e: |
| 322 | print(f"Failed to start API-only server: {e}") |
| 323 | raise |
| 324 | |
| 325 | def start_engines_server(): |
| 326 | try: |
| 327 | server = RemoteOpenAIServer( |
| 328 | self.model_name, |
| 329 | engines_server_args, |
| 330 | auto_port=False, |
| 331 | env_dict={ |
| 332 | current_platform.device_control_env_var: ",".join( |
| 333 | str(current_platform.device_id_to_physical_device_id(i)) |
| 334 | for i in range(self.dp_size * self.tp_size) |
| 335 | ), |
| 336 | }, |
| 337 | ) |
| 338 | server.__enter__() |
| 339 | print( |
| 340 | f"Headless engines server started successfully with " |
| 341 | f"{self.dp_size} engines" |
| 342 | ) |
| 343 | self.servers[1] = (server, engines_server_args) |
| 344 | except Exception as e: |
| 345 | print(f"Failed to start headless engines server: {e}") |
| 346 | raise |
| 347 | |
| 348 | # Start API server first |
| 349 | api_thread = threading.Thread(target=start_api_server) |
| 350 | api_thread.start() |
| 351 | self.server_threads.append(api_thread) |
| 352 | |
| 353 | # Start engines server second |
| 354 | engines_thread = threading.Thread(target=start_engines_server) |
| 355 | engines_thread.start() |
| 356 | self.server_threads.append(engines_thread) |
| 357 | |
| 358 | # Wait for both servers to start |
| 359 | for thread in self.server_threads: |
| 360 | thread.join() |
| 361 | |
| 362 | # Give servers additional time to fully initialize and coordinate |
| 363 | time.sleep(3) |
| 364 | |
| 365 | if not all(self.servers): |
| 366 | raise Exception("Both servers failed to start") |
| 367 | |
| 368 | return cast(list[tuple[RemoteOpenAIServer, list[str]]], self.servers) |
| 369 | |
| 370 | def __exit__(self, exc_type, exc_val, exc_tb): |
| 371 | """Stop both server instances.""" |
| 372 | servers = [entry[0] for entry in self.servers if entry is not None] |
| 373 | self.servers.clear() |
| 374 | try: |
| 375 | RemoteOpenAIServer.shutdown_many(servers) |
| 376 | except Exception as e: |
| 377 | print(f"Error stopping servers: {e}") |
| 378 | traceback.print_exc() |
| 379 | |
| 380 | |
| 381 | @pytest.fixture(scope="module") |
| 382 | def default_server_args(): |
| 383 | return [ |
| 384 | # use half precision for speed and memory savings in CI environment |
| 385 | "--dtype", |
| 386 | "bfloat16", |
| 387 | "--max-model-len", |
| 388 | "2048", |
| 389 | "--max-num-seqs", |
| 390 | "128", |
| 391 | "--enforce-eager", |
| 392 | ] |
| 393 | |
| 394 | |
| 395 | @pytest.fixture(scope="module", params=[1, 4]) |
| 396 | def server_manager(request, default_server_args): |
| 397 | api_server_count = request.param |
| 398 | server_manager = MultinodeInternalLBServerManager( |
| 399 | MODEL_NAME, |
| 400 | DP_SIZE, |
| 401 | api_server_count, |
| 402 | default_server_args, |
| 403 | DP_SIZE // NUM_NODES, |
| 404 | TP_SIZE, |
| 405 | ) |
| 406 | |
| 407 | with server_manager: |
| 408 | yield server_manager |
| 409 | |
| 410 | |
| 411 | @pytest.fixture |
| 412 | def servers(server_manager): |
| 413 | return server_manager.servers |
| 414 | |
| 415 | |
| 416 | @pytest.fixture(scope="module", params=[1, 4]) |
| 417 | def api_only_servers(request, default_server_args): |
| 418 | """Fixture for API-only server + headless engines configuration.""" |
| 419 | api_server_count = request.param |
| 420 | with APIOnlyServerManager( |
| 421 | MODEL_NAME, DP_SIZE, api_server_count, default_server_args, TP_SIZE |
| 422 | ) as server_list: |
| 423 | yield server_list |
| 424 | |
| 425 | |
| 426 | @pytest_asyncio.fixture |
| 427 | async def client(servers: list[tuple[RemoteOpenAIServer, list[str]]]): |
| 428 | # For internal LB, we only connect to the head node (rank 0) |
| 429 | # which provides the single API endpoint |
| 430 | head_server = servers[0][0] |
| 431 | async with head_server.get_async_client() as client: |
| 432 | yield client |
| 433 | |
| 434 | |
| 435 | @pytest_asyncio.fixture |
| 436 | async def api_only_client(api_only_servers: list[tuple[RemoteOpenAIServer, list[str]]]): |
| 437 | """Client fixture for API-only server configuration.""" |
| 438 | # Connect to the API-only server (first server in the list) |
| 439 | api_server = api_only_servers[0][0] |
| 440 | async with api_server.get_async_client() as client: |
| 441 | yield client |
| 442 | |
| 443 | |
| 444 | def _get_parallel_config(server: RemoteOpenAIServer): |
| 445 | response = requests.get(server.url_for("server_info?config_format=json")) |
| 446 | response.raise_for_status() |
| 447 | |
| 448 | vllm_config = response.json()["vllm_config"] |
| 449 | return vllm_config["parallel_config"] |
| 450 | |
| 451 | |
| 452 | def test_multinode_dp_server_info(server_manager): |
| 453 | head_server = server_manager.servers[0][0] |
| 454 | api_server_count = server_manager.api_server_count |
| 455 | |
| 456 | # Each request will hit one of the API servers |
| 457 | # `n_reqs` is set so that there is a good chance each server |
| 458 | # receives at least one request |
| 459 | n_reqs = 2 * api_server_count * api_server_count |
| 460 | parallel_configs = [_get_parallel_config(head_server) for _ in range(n_reqs)] |
| 461 | api_process_counts = [c["_api_process_count"] for c in parallel_configs] |
| 462 | api_process_ranks = [c["_api_process_rank"] for c in parallel_configs] |
| 463 | |
| 464 | assert all(c == api_server_count for c in api_process_counts), api_process_counts |
| 465 | assert all(0 <= r < api_server_count for r in api_process_ranks), api_process_ranks |
| 466 | |
| 467 | |
| 468 | @pytest.mark.asyncio |
| 469 | @pytest.mark.parametrize( |
| 470 | "model_name", |
| 471 | [MODEL_NAME], |
| 472 | ) |
| 473 | async def test_multinode_dp_completion( |
| 474 | client: openai.AsyncOpenAI, |
| 475 | servers: list[tuple[RemoteOpenAIServer, list[str]]], |
| 476 | model_name: str, |
| 477 | ) -> None: |
| 478 | # Test single request |
| 479 | result = await _make_completion_request(client, model_name) |
| 480 | assert result is not None |
| 481 | print("Multi-node internal LB handled single completion request successfully") |
| 482 | |
| 483 | await asyncio.sleep(0.5) |
| 484 | |
| 485 | # Send multiple bursts - internal LB should distribute across DP ranks |
| 486 | await _run_request_bursts(client, model_name) |
| 487 | |
| 488 | _, server_args = servers[0] |
| 489 | api_server_count = ( |
| 490 | server_args.count("--api-server-count") |
| 491 | and server_args[server_args.index("--api-server-count") + 1] |
| 492 | or 1 |
| 493 | ) |
| 494 | print( |
| 495 | f"Successfully completed multi-node internal LB test with " |
| 496 | f"{len(servers)} DP ranks (API server count: {api_server_count})" |
| 497 | ) |
| 498 | |
| 499 | # Check request balancing via Prometheus metrics |
| 500 | head_server = servers[0][0] |
| 501 | check_request_balancing(head_server, DP_SIZE) |
| 502 | |
| 503 | |
| 504 | @pytest.mark.asyncio |
| 505 | @pytest.mark.parametrize( |
| 506 | "model_name", |
| 507 | [MODEL_NAME], |
| 508 | ) |
| 509 | async def test_multinode_dp_completion_streaming( |
| 510 | client: openai.AsyncOpenAI, |
| 511 | servers: list[tuple[RemoteOpenAIServer, list[str]]], |
| 512 | model_name: str, |
| 513 | ) -> None: |
| 514 | prompt = "What is an LLM?" |
| 515 | |
| 516 | async def make_streaming_request(): |
| 517 | # Perform a non-streaming request to get the expected full output |
| 518 | single_completion = await client.completions.create( |
| 519 | model=model_name, |
| 520 | prompt=prompt, |
| 521 | max_tokens=5, |
| 522 | temperature=0.0, |
| 523 | ) |
| 524 | single_output = single_completion.choices[0].text |
| 525 | |
| 526 | # Perform the streaming request |
| 527 | stream = await client.completions.create( |
| 528 | model=model_name, prompt=prompt, max_tokens=5, temperature=0.0, stream=True |
| 529 | ) |
| 530 | chunks: list[str] = [] |
| 531 | finish_reason_count = 0 |
| 532 | last_chunk = None |
| 533 | async for chunk in stream: |
| 534 | chunks.append(chunk.choices[0].text) |
| 535 | if chunk.choices[0].finish_reason is not None: |
| 536 | finish_reason_count += 1 |
| 537 | last_chunk = chunk # Keep track of the last chunk |
| 538 | |
| 539 | # finish reason should only return in the last block for OpenAI API |
| 540 | assert finish_reason_count == 1, "Finish reason should appear exactly once." |
| 541 | assert last_chunk is not None, "Stream should have yielded at least one chunk." |
| 542 | assert last_chunk.choices[0].finish_reason == "length", ( |
| 543 | "Finish reason should be 'length'." |
| 544 | ) |
| 545 | # Check that the combined text matches the non-streamed version. |
| 546 | assert "".join(chunks) == single_output, ( |
| 547 | "Streamed output should match non-streamed output." |
| 548 | ) |
| 549 | return True # Indicate success for this request |
| 550 | |
| 551 | # Test single streaming request |
| 552 | result = await make_streaming_request() |
| 553 | assert result is not None |
| 554 | print("Multi-node internal LB handled single streaming request successfully") |
| 555 | |
| 556 | await asyncio.sleep(0.5) |
| 557 | |
| 558 | # Send multiple streaming requests - internal LB should distribute across |
| 559 | # DP ranks |
| 560 | num_requests = 200 |
| 561 | all_tasks = [] |
| 562 | for _ in range(num_requests): |
| 563 | all_tasks.append(asyncio.create_task(make_streaming_request())) |
| 564 | await asyncio.sleep(0.01) |
| 565 | |
| 566 | results = await asyncio.gather(*all_tasks) |
| 567 | assert len(results) == num_requests |
| 568 | assert all(results), "Not all streaming requests completed successfully." |
| 569 | |
| 570 | await asyncio.sleep(0.5) |
| 571 | |
| 572 | # Second burst of streaming requests |
| 573 | all_tasks = [] |
| 574 | for _ in range(num_requests): |
| 575 | all_tasks.append(asyncio.create_task(make_streaming_request())) |
| 576 | await asyncio.sleep(0.01) |
| 577 | |
| 578 | results = await asyncio.gather(*all_tasks) |
| 579 | assert len(results) == num_requests |
| 580 | assert all(results), "Not all streaming requests completed successfully." |
| 581 | |
| 582 | _, server_args = servers[0] |
| 583 | api_server_count = ( |
| 584 | server_args.count("--api-server-count") |
| 585 | and server_args[server_args.index("--api-server-count") + 1] |
| 586 | or 1 |
| 587 | ) |
| 588 | print( |
| 589 | f"Successfully completed multi-node internal LB streaming test with " |
| 590 | f"{len(servers)} DP ranks (API server count: {api_server_count})" |
| 591 | ) |
| 592 | |
| 593 | # Check request balancing via Prometheus metrics |
| 594 | head_server = servers[0][0] |
| 595 | check_request_balancing(head_server, DP_SIZE) |
| 596 | |
| 597 | |
| 598 | @pytest.mark.asyncio |
| 599 | @pytest.mark.parametrize( |
| 600 | "model_name", |
| 601 | [MODEL_NAME], |
| 602 | ) |
| 603 | async def test_api_only_multinode_dp_completion( |
| 604 | api_only_client: openai.AsyncOpenAI, |
| 605 | api_only_servers: list[tuple[RemoteOpenAIServer, list[str]]], |
| 606 | model_name: str, |
| 607 | ) -> None: |
| 608 | """Test API-only server with all engines on separate headless server.""" |
| 609 | # Test single request |
| 610 | result = await _make_completion_request(api_only_client, model_name) |
| 611 | assert result is not None |
| 612 | print("API-only server handled single completion request successfully") |
| 613 | |
| 614 | await asyncio.sleep(0.5) |
| 615 | |
| 616 | # Send multiple bursts - should be distributed across engines on |
| 617 | # headless server |
| 618 | await _run_request_bursts(api_only_client, model_name) |
| 619 | |
| 620 | api_server, api_server_args = api_only_servers[0] |
| 621 | api_server_count = ( |
| 622 | api_server_args.count("--api-server-count") |
| 623 | and api_server_args[api_server_args.index("--api-server-count") + 1] |
| 624 | or 1 |
| 625 | ) |
| 626 | print( |
| 627 | f"Successfully completed API-only multi-node test with {DP_SIZE} " |
| 628 | f"engines on headless server (API server count: {api_server_count})" |
| 629 | ) |
| 630 | |
| 631 | # Check request balancing via Prometheus metrics |
| 632 | check_request_balancing(api_server, DP_SIZE) |
| 633 | |
| 634 | |
| 635 | @pytest.mark.asyncio |
| 636 | @pytest.mark.parametrize( |
| 637 | "model_name", |
| 638 | [MODEL_NAME], |
| 639 | ) |
| 640 | async def test_api_only_multinode_dp_completion_streaming( |
| 641 | api_only_client: openai.AsyncOpenAI, |
| 642 | api_only_servers: list[tuple[RemoteOpenAIServer, list[str]]], |
| 643 | model_name: str, |
| 644 | ) -> None: |
| 645 | """Test API-only server streaming with all engines on separate |
| 646 | headless server.""" |
| 647 | prompt = "What is an LLM?" |
| 648 | |
| 649 | async def make_streaming_request(): |
| 650 | # Perform a non-streaming request to get the expected full output |
| 651 | single_completion = await api_only_client.completions.create( |
| 652 | model=model_name, |
| 653 | prompt=prompt, |
| 654 | max_tokens=5, |
| 655 | temperature=0.0, |
| 656 | ) |
| 657 | single_output = single_completion.choices[0].text |
| 658 | |
| 659 | # Perform the streaming request |
| 660 | stream = await api_only_client.completions.create( |
| 661 | model=model_name, prompt=prompt, max_tokens=5, temperature=0.0, stream=True |
| 662 | ) |
| 663 | chunks: list[str] = [] |
| 664 | finish_reason_count = 0 |
| 665 | last_chunk = None |
| 666 | async for chunk in stream: |
| 667 | chunks.append(chunk.choices[0].text) |
| 668 | if chunk.choices[0].finish_reason is not None: |
| 669 | finish_reason_count += 1 |
| 670 | last_chunk = chunk # Keep track of the last chunk |
| 671 | |
| 672 | # finish reason should only return in the last block for OpenAI API |
| 673 | assert finish_reason_count == 1, "Finish reason should appear exactly once." |
| 674 | assert last_chunk is not None, "Stream should have yielded at least one chunk." |
| 675 | assert last_chunk.choices[0].finish_reason == "length", ( |
| 676 | "Finish reason should be 'length'." |
| 677 | ) |
| 678 | # Check that the combined text matches the non-streamed version. |
| 679 | assert "".join(chunks) == single_output, ( |
| 680 | "Streamed output should match non-streamed output." |
| 681 | ) |
| 682 | return True # Indicate success for this request |
| 683 | |
| 684 | # Test single streaming request |
| 685 | result = await make_streaming_request() |
| 686 | assert result is not None |
| 687 | print("API-only server handled single streaming request successfully") |
| 688 | |
| 689 | await asyncio.sleep(0.5) |
| 690 | |
| 691 | # Send multiple streaming requests - should be distributed across engines |
| 692 | num_requests = 200 |
| 693 | all_tasks = [] |
| 694 | for _ in range(num_requests): |
| 695 | all_tasks.append(asyncio.create_task(make_streaming_request())) |
| 696 | await asyncio.sleep(0.01) |
| 697 | |
| 698 | results = await asyncio.gather(*all_tasks) |
| 699 | assert len(results) == num_requests |
| 700 | assert all(results), "Not all streaming requests completed successfully." |
| 701 | |
| 702 | await asyncio.sleep(0.5) |
| 703 | |
| 704 | # Second burst of streaming requests |
| 705 | all_tasks = [] |
| 706 | for _ in range(num_requests): |
| 707 | all_tasks.append(asyncio.create_task(make_streaming_request())) |
| 708 | await asyncio.sleep(0.01) |
| 709 | |
| 710 | results = await asyncio.gather(*all_tasks) |
| 711 | assert len(results) == num_requests |
| 712 | assert all(results), "Not all streaming requests completed successfully." |
| 713 | |
| 714 | _, api_server_args = api_only_servers[0] |
| 715 | api_server_count = ( |
| 716 | api_server_args.count("--api-server-count") |
| 717 | and api_server_args[api_server_args.index("--api-server-count") + 1] |
| 718 | or 1 |
| 719 | ) |
| 720 | print( |
| 721 | f"Successfully completed API-only streaming test with {DP_SIZE} " |
| 722 | f"engines on headless server (API server count: {api_server_count})" |
| 723 | ) |
| 724 | |
| 725 | # Check request balancing via Prometheus metrics |
| 726 | api_server = api_only_servers[0][0] |
| 727 | check_request_balancing(api_server, DP_SIZE) |