| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import json |
| 5 | import subprocess |
| 6 | import tempfile |
| 7 | import threading |
| 8 | from collections.abc import Generator |
| 9 | from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer |
| 10 | from unittest.mock import AsyncMock, MagicMock, patch |
| 11 | |
| 12 | import pytest |
| 13 | |
| 14 | import vllm.envs as envs |
| 15 | from vllm.assets.audio import AudioAsset |
| 16 | from vllm.connections import HTTPConnection |
| 17 | from vllm.entrypoints.launchers.run_batch import ( |
| 18 | BatchRequestOutput, |
| 19 | BatchTranscriptionRequest, |
| 20 | download_bytes_from_url, |
| 21 | make_transcription_wrapper, |
| 22 | upload_data, |
| 23 | ) |
| 24 | from vllm.entrypoints.serve.engine.protocol import ErrorResponse |
| 25 | from vllm.exceptions import VLLMValidationError |
| 26 | from vllm.utils.mem_constants import MiB_bytes |
| 27 | |
| 28 | CHAT_MODEL_NAME = "hmellor/tiny-random-LlamaForCausalLM" |
| 29 | EMBEDDING_MODEL_NAME = "intfloat/multilingual-e5-small" |
| 30 | RERANKER_MODEL_NAME = "BAAI/bge-reranker-v2-m3" |
| 31 | REASONING_MODEL_NAME = "Qwen/Qwen3-0.6B" |
| 32 | SPEECH_LARGE_MODEL_NAME = "openai/whisper-large-v3" |
| 33 | SPEECH_SMALL_MODEL_NAME = "openai/whisper-small" |
| 34 | |
| 35 | INPUT_BATCH = "\n".join( |
| 36 | json.dumps(req) |
| 37 | for req in [ |
| 38 | { |
| 39 | "custom_id": "request-1", |
| 40 | "method": "POST", |
| 41 | "url": "/v1/chat/completions", |
| 42 | "body": { |
| 43 | "model": CHAT_MODEL_NAME, |
| 44 | "messages": [ |
| 45 | { |
| 46 | "role": "system", |
| 47 | "content": "You are a helpful assistant.", |
| 48 | }, |
| 49 | {"role": "user", "content": "Hello world!"}, |
| 50 | ], |
| 51 | "max_tokens": 1000, |
| 52 | }, |
| 53 | }, |
| 54 | { |
| 55 | "custom_id": "request-2", |
| 56 | "method": "POST", |
| 57 | "url": "/v1/chat/completions", |
| 58 | "body": { |
| 59 | "model": CHAT_MODEL_NAME, |
| 60 | "messages": [ |
| 61 | { |
| 62 | "role": "system", |
| 63 | "content": "You are an unhelpful assistant.", |
| 64 | }, |
| 65 | {"role": "user", "content": "Hello world!"}, |
| 66 | ], |
| 67 | "max_tokens": 1000, |
| 68 | }, |
| 69 | }, |
| 70 | { |
| 71 | "custom_id": "request-3", |
| 72 | "method": "POST", |
| 73 | "url": "/v1/chat/completions", |
| 74 | "body": { |
| 75 | "model": "NonExistModel", |
| 76 | "messages": [ |
| 77 | { |
| 78 | "role": "system", |
| 79 | "content": "You are an unhelpful assistant.", |
| 80 | }, |
| 81 | {"role": "user", "content": "Hello world!"}, |
| 82 | ], |
| 83 | "max_tokens": 1000, |
| 84 | }, |
| 85 | }, |
| 86 | { |
| 87 | "custom_id": "request-4", |
| 88 | "method": "POST", |
| 89 | "url": "/bad_url", |
| 90 | "body": { |
| 91 | "model": CHAT_MODEL_NAME, |
| 92 | "messages": [ |
| 93 | { |
| 94 | "role": "system", |
| 95 | "content": "You are an unhelpful assistant.", |
| 96 | }, |
| 97 | {"role": "user", "content": "Hello world!"}, |
| 98 | ], |
| 99 | "max_tokens": 1000, |
| 100 | }, |
| 101 | }, |
| 102 | { |
| 103 | "custom_id": "request-5", |
| 104 | "method": "POST", |
| 105 | "url": "/v1/chat/completions", |
| 106 | "body": { |
| 107 | "stream": "True", |
| 108 | "model": CHAT_MODEL_NAME, |
| 109 | "messages": [ |
| 110 | { |
| 111 | "role": "system", |
| 112 | "content": "You are an unhelpful assistant.", |
| 113 | }, |
| 114 | {"role": "user", "content": "Hello world!"}, |
| 115 | ], |
| 116 | "max_tokens": 1000, |
| 117 | }, |
| 118 | }, |
| 119 | ] |
| 120 | ) |
| 121 | |
| 122 | INVALID_INPUT_BATCH = "\n".join( |
| 123 | json.dumps(req) |
| 124 | for req in [ |
| 125 | { |
| 126 | "invalid_field": "request-1", |
| 127 | "method": "POST", |
| 128 | "url": "/v1/chat/completions", |
| 129 | "body": { |
| 130 | "model": CHAT_MODEL_NAME, |
| 131 | "messages": [ |
| 132 | {"role": "system", "content": "You are a helpful assistant."}, |
| 133 | {"role": "user", "content": "Hello world!"}, |
| 134 | ], |
| 135 | "max_tokens": 1000, |
| 136 | }, |
| 137 | }, |
| 138 | { |
| 139 | "custom_id": "request-2", |
| 140 | "method": "POST", |
| 141 | "url": "/v1/chat/completions", |
| 142 | "body": { |
| 143 | "model": CHAT_MODEL_NAME, |
| 144 | "messages": [ |
| 145 | {"role": "system", "content": "You are an unhelpful assistant."}, |
| 146 | {"role": "user", "content": "Hello world!"}, |
| 147 | ], |
| 148 | "max_tokens": 1000, |
| 149 | }, |
| 150 | }, |
| 151 | ] |
| 152 | ) |
| 153 | |
| 154 | INPUT_EMBEDDING_BATCH = "\n".join( |
| 155 | json.dumps(req) |
| 156 | for req in [ |
| 157 | { |
| 158 | "custom_id": "request-1", |
| 159 | "method": "POST", |
| 160 | "url": "/v1/embeddings", |
| 161 | "body": { |
| 162 | "model": EMBEDDING_MODEL_NAME, |
| 163 | "input": "You are a helpful assistant.", |
| 164 | }, |
| 165 | }, |
| 166 | { |
| 167 | "custom_id": "request-2", |
| 168 | "method": "POST", |
| 169 | "url": "/v1/embeddings", |
| 170 | "body": { |
| 171 | "model": EMBEDDING_MODEL_NAME, |
| 172 | "input": "You are an unhelpful assistant.", |
| 173 | }, |
| 174 | }, |
| 175 | { |
| 176 | "custom_id": "request-3", |
| 177 | "method": "POST", |
| 178 | "url": "/v1/embeddings", |
| 179 | "body": { |
| 180 | "model": EMBEDDING_MODEL_NAME, |
| 181 | "input": "Hello world!", |
| 182 | }, |
| 183 | }, |
| 184 | { |
| 185 | "custom_id": "request-4", |
| 186 | "method": "POST", |
| 187 | "url": "/v1/embeddings", |
| 188 | "body": { |
| 189 | "model": "NonExistModel", |
| 190 | "input": "Hello world!", |
| 191 | }, |
| 192 | }, |
| 193 | ] |
| 194 | ) |
| 195 | |
| 196 | _SCORE_RERANK_DOCUMENTS = [ |
| 197 | "The capital of Brazil is Brasilia.", |
| 198 | "The capital of France is Paris.", |
| 199 | ] |
| 200 | |
| 201 | INPUT_SCORE_BATCH = "\n".join( |
| 202 | json.dumps(req) |
| 203 | for req in [ |
| 204 | { |
| 205 | "custom_id": "request-1", |
| 206 | "method": "POST", |
| 207 | "url": "/score", |
| 208 | "body": { |
| 209 | "model": RERANKER_MODEL_NAME, |
| 210 | "queries": "What is the capital of France?", |
| 211 | "documents": _SCORE_RERANK_DOCUMENTS, |
| 212 | }, |
| 213 | }, |
| 214 | { |
| 215 | "custom_id": "request-2", |
| 216 | "method": "POST", |
| 217 | "url": "/v1/score", |
| 218 | "body": { |
| 219 | "model": RERANKER_MODEL_NAME, |
| 220 | "queries": "What is the capital of France?", |
| 221 | "documents": _SCORE_RERANK_DOCUMENTS, |
| 222 | }, |
| 223 | }, |
| 224 | ] |
| 225 | ) |
| 226 | |
| 227 | INPUT_RERANK_BATCH = "\n".join( |
| 228 | json.dumps(req) |
| 229 | for req in [ |
| 230 | { |
| 231 | "custom_id": "request-1", |
| 232 | "method": "POST", |
| 233 | "url": "/rerank", |
| 234 | "body": { |
| 235 | "model": RERANKER_MODEL_NAME, |
| 236 | "query": "What is the capital of France?", |
| 237 | "documents": _SCORE_RERANK_DOCUMENTS, |
| 238 | }, |
| 239 | }, |
| 240 | { |
| 241 | "custom_id": "request-2", |
| 242 | "method": "POST", |
| 243 | "url": "/v1/rerank", |
| 244 | "body": { |
| 245 | "model": RERANKER_MODEL_NAME, |
| 246 | "query": "What is the capital of France?", |
| 247 | "documents": _SCORE_RERANK_DOCUMENTS, |
| 248 | }, |
| 249 | }, |
| 250 | { |
| 251 | "custom_id": "request-2", |
| 252 | "method": "POST", |
| 253 | "url": "/v2/rerank", |
| 254 | "body": { |
| 255 | "model": RERANKER_MODEL_NAME, |
| 256 | "query": "What is the capital of France?", |
| 257 | "documents": _SCORE_RERANK_DOCUMENTS, |
| 258 | }, |
| 259 | }, |
| 260 | ] |
| 261 | ) |
| 262 | |
| 263 | INPUT_REASONING_BATCH = "\n".join( |
| 264 | json.dumps(req) |
| 265 | for req in [ |
| 266 | { |
| 267 | "custom_id": "request-1", |
| 268 | "method": "POST", |
| 269 | "url": "/v1/chat/completions", |
| 270 | "body": { |
| 271 | "model": REASONING_MODEL_NAME, |
| 272 | "messages": [ |
| 273 | {"role": "system", "content": "You are a helpful assistant."}, |
| 274 | {"role": "user", "content": "Solve this math problem: 2+2=?"}, |
| 275 | ], |
| 276 | }, |
| 277 | }, |
| 278 | { |
| 279 | "custom_id": "request-2", |
| 280 | "method": "POST", |
| 281 | "url": "/v1/chat/completions", |
| 282 | "body": { |
| 283 | "model": REASONING_MODEL_NAME, |
| 284 | "messages": [ |
| 285 | {"role": "system", "content": "You are a helpful assistant."}, |
| 286 | {"role": "user", "content": "What is the capital of France?"}, |
| 287 | ], |
| 288 | }, |
| 289 | }, |
| 290 | ] |
| 291 | ) |
| 292 | |
| 293 | MINIMAL_WAV_BASE64 = "UklGRigAAABXQVZFZm10IBAAAAABAAEAgD4AAAB9AAACABAAZGF0YQQAAAAAAP9/" |
| 294 | _EXACT_LIMIT_AUDIO = b"a" * MiB_bytes |
| 295 | _OVERSIZED_BASE64_AUDIO = "A" * (4 * ((MiB_bytes + 2) // 3)) |
| 296 | _OVERSIZED_HTTP_AUDIO = b"b" * (MiB_bytes + 1) |
| 297 | |
| 298 | |
| 299 | class _BatchHTTPServer(ThreadingHTTPServer): |
| 300 | request_paths: list[str] |
| 301 | |
| 302 | |
| 303 | class _BatchHTTPHandler(BaseHTTPRequestHandler): |
| 304 | def do_GET(self) -> None: |
| 305 | server = self.server |
| 306 | assert isinstance(server, _BatchHTTPServer) |
| 307 | server.request_paths.append(self.path) |
| 308 | |
| 309 | if self.path == "/exact-limit": |
| 310 | self._send_body(_EXACT_LIMIT_AUDIO) |
| 311 | return |
| 312 | |
| 313 | if self.path == "/oversized-chunked": |
| 314 | self.send_response(200) |
| 315 | self.end_headers() |
| 316 | self._write_body(_OVERSIZED_HTTP_AUDIO) |
| 317 | return |
| 318 | |
| 319 | self.send_error(404, "Unknown batch HTTP test path") |
| 320 | |
| 321 | def _send_body(self, body: bytes) -> None: |
| 322 | self.send_response(200) |
| 323 | self.send_header("Content-Type", "application/octet-stream") |
| 324 | self.send_header("Content-Length", str(len(body))) |
| 325 | self.end_headers() |
| 326 | self._write_body(body) |
| 327 | |
| 328 | def _write_body(self, body: bytes) -> None: |
| 329 | try: |
| 330 | for offset in range(0, len(body), 64 * 1024): |
| 331 | self.wfile.write(body[offset : offset + 64 * 1024]) |
| 332 | except (BrokenPipeError, ConnectionResetError): |
| 333 | self.close_connection = True |
| 334 | |
| 335 | def log_message(self, fmt: str, *args: object) -> None: |
| 336 | return |
| 337 | |
| 338 | |
| 339 | @pytest.fixture |
| 340 | def batch_http_server() -> Generator[_BatchHTTPServer, None, None]: |
| 341 | server = _BatchHTTPServer(("127.0.0.1", 0), _BatchHTTPHandler) |
| 342 | server.request_paths = [] |
| 343 | thread = threading.Thread(target=server.serve_forever, daemon=True) |
| 344 | thread.start() |
| 345 | try: |
| 346 | yield server |
| 347 | finally: |
| 348 | server.shutdown() |
| 349 | thread.join() |
| 350 | server.server_close() |
| 351 | |
| 352 | |
| 353 | def _batch_http_url(server: _BatchHTTPServer, path: str) -> str: |
| 354 | return f"http://localhost:{server.server_port}{path}" |
| 355 | |
| 356 | |
| 357 | async def _close_async_connection(connection: HTTPConnection) -> None: |
| 358 | if connection._async_client is not None: |
| 359 | await connection._async_client.close() |
| 360 | |
| 361 | |
| 362 | INPUT_TRANSCRIPTION_BATCH = ( |
| 363 | json.dumps( |
| 364 | { |
| 365 | "custom_id": "request-1", |
| 366 | "method": "POST", |
| 367 | "url": "/v1/audio/transcriptions", |
| 368 | "body": { |
| 369 | "model": SPEECH_LARGE_MODEL_NAME, |
| 370 | "file_url": f"data:audio/wav;base64,{MINIMAL_WAV_BASE64}", |
| 371 | "response_format": "json", |
| 372 | }, |
| 373 | } |
| 374 | ) |
| 375 | + "\n" |
| 376 | ) |
| 377 | |
| 378 | INPUT_TRANSCRIPTION_HTTP_BATCH = ( |
| 379 | json.dumps( |
| 380 | { |
| 381 | "custom_id": "request-1", |
| 382 | "method": "POST", |
| 383 | "url": "/v1/audio/transcriptions", |
| 384 | "body": { |
| 385 | "model": SPEECH_LARGE_MODEL_NAME, |
| 386 | "file_url": AudioAsset("mary_had_lamb").url, |
| 387 | "language": "en", |
| 388 | "response_format": "json", |
| 389 | }, |
| 390 | } |
| 391 | ) |
| 392 | + "\n" |
| 393 | ) |
| 394 | |
| 395 | INPUT_TRANSLATION_BATCH = ( |
| 396 | json.dumps( |
| 397 | { |
| 398 | "custom_id": "request-1", |
| 399 | "method": "POST", |
| 400 | "url": "/v1/audio/translations", |
| 401 | "body": { |
| 402 | "model": SPEECH_SMALL_MODEL_NAME, |
| 403 | "file_url": AudioAsset("mary_had_lamb").url, |
| 404 | "response_format": "text", |
| 405 | "language": "it", |
| 406 | "to_language": "en", |
| 407 | "temperature": 0.0, |
| 408 | }, |
| 409 | } |
| 410 | ) |
| 411 | + "\n" |
| 412 | ) |
| 413 | |
| 414 | WEATHER_TOOL = { |
| 415 | "type": "function", |
| 416 | "function": { |
| 417 | "name": "get_current_weather", |
| 418 | "description": "Get the current weather in a given location", |
| 419 | "parameters": { |
| 420 | "type": "object", |
| 421 | "properties": { |
| 422 | "location": { |
| 423 | "type": "string", |
| 424 | "description": "The city and state, e.g. San Francisco, CA", |
| 425 | }, |
| 426 | "unit": { |
| 427 | "type": "string", |
| 428 | "enum": ["celsius", "fahrenheit"], |
| 429 | }, |
| 430 | }, |
| 431 | "required": ["location"], |
| 432 | }, |
| 433 | }, |
| 434 | } |
| 435 | |
| 436 | INPUT_TOOL_CALLING_BATCH = json.dumps( |
| 437 | { |
| 438 | "custom_id": "request-1", |
| 439 | "method": "POST", |
| 440 | "url": "/v1/chat/completions", |
| 441 | "body": { |
| 442 | "model": REASONING_MODEL_NAME, |
| 443 | "messages": [ |
| 444 | {"role": "user", "content": "What is the weather in San Francisco?"}, |
| 445 | ], |
| 446 | "tools": [WEATHER_TOOL], |
| 447 | "tool_choice": "required", |
| 448 | "max_tokens": 1000, |
| 449 | }, |
| 450 | } |
| 451 | ) |
| 452 | |
| 453 | |
| 454 | def test_empty_file(): |
| 455 | with ( |
| 456 | tempfile.NamedTemporaryFile("w") as input_file, |
| 457 | tempfile.NamedTemporaryFile("r") as output_file, |
| 458 | ): |
| 459 | input_file.write("") |
| 460 | input_file.flush() |
| 461 | proc = subprocess.Popen( |
| 462 | [ |
| 463 | "vllm", |
| 464 | "run-batch", |
| 465 | "-i", |
| 466 | input_file.name, |
| 467 | "-o", |
| 468 | output_file.name, |
| 469 | "--model", |
| 470 | EMBEDDING_MODEL_NAME, |
| 471 | ], |
| 472 | ) |
| 473 | proc.communicate() |
| 474 | proc.wait() |
| 475 | assert proc.returncode == 0, f"{proc=}" |
| 476 | |
| 477 | contents = output_file.read() |
| 478 | assert contents.strip() == "" |
| 479 | |
| 480 | |
| 481 | def test_completions(): |
| 482 | with ( |
| 483 | tempfile.NamedTemporaryFile("w") as input_file, |
| 484 | tempfile.NamedTemporaryFile("r") as output_file, |
| 485 | ): |
| 486 | input_file.write(INPUT_BATCH) |
| 487 | input_file.flush() |
| 488 | proc = subprocess.Popen( |
| 489 | [ |
| 490 | "vllm", |
| 491 | "run-batch", |
| 492 | "-i", |
| 493 | input_file.name, |
| 494 | "-o", |
| 495 | output_file.name, |
| 496 | "--model", |
| 497 | CHAT_MODEL_NAME, |
| 498 | ], |
| 499 | ) |
| 500 | proc.communicate() |
| 501 | proc.wait() |
| 502 | assert proc.returncode == 0, f"{proc=}" |
| 503 | |
| 504 | contents = output_file.read() |
| 505 | for line in contents.strip().split("\n"): |
| 506 | # Ensure that the output format conforms to the openai api. |
| 507 | # Validation should throw if the schema is wrong. |
| 508 | BatchRequestOutput.model_validate_json(line) |
| 509 | |
| 510 | |
| 511 | def test_completions_invalid_input(): |
| 512 | """Ensure that we fail when the input doesn't conform to the openai api.""" |
| 513 | with ( |
| 514 | tempfile.NamedTemporaryFile("w") as input_file, |
| 515 | tempfile.NamedTemporaryFile("r") as output_file, |
| 516 | ): |
| 517 | input_file.write(INVALID_INPUT_BATCH) |
| 518 | input_file.flush() |
| 519 | proc = subprocess.Popen( |
| 520 | [ |
| 521 | "vllm", |
| 522 | "run-batch", |
| 523 | "-i", |
| 524 | input_file.name, |
| 525 | "-o", |
| 526 | output_file.name, |
| 527 | "--model", |
| 528 | CHAT_MODEL_NAME, |
| 529 | ], |
| 530 | ) |
| 531 | proc.communicate() |
| 532 | proc.wait() |
| 533 | assert proc.returncode != 0, f"{proc=}" |
| 534 | |
| 535 | |
| 536 | def test_embeddings(): |
| 537 | with ( |
| 538 | tempfile.NamedTemporaryFile("w") as input_file, |
| 539 | tempfile.NamedTemporaryFile("r") as output_file, |
| 540 | ): |
| 541 | input_file.write(INPUT_EMBEDDING_BATCH) |
| 542 | input_file.flush() |
| 543 | proc = subprocess.Popen( |
| 544 | [ |
| 545 | "vllm", |
| 546 | "run-batch", |
| 547 | "-i", |
| 548 | input_file.name, |
| 549 | "-o", |
| 550 | output_file.name, |
| 551 | "--model", |
| 552 | EMBEDDING_MODEL_NAME, |
| 553 | ], |
| 554 | ) |
| 555 | proc.communicate() |
| 556 | proc.wait() |
| 557 | assert proc.returncode == 0, f"{proc=}" |
| 558 | |
| 559 | contents = output_file.read() |
| 560 | for line in contents.strip().split("\n"): |
| 561 | # Ensure that the output format conforms to the openai api. |
| 562 | # Validation should throw if the schema is wrong. |
| 563 | BatchRequestOutput.model_validate_json(line) |
| 564 | |
| 565 | |
| 566 | @pytest.mark.parametrize("input_batch", [INPUT_SCORE_BATCH, INPUT_RERANK_BATCH]) |
| 567 | def test_score(input_batch): |
| 568 | with ( |
| 569 | tempfile.NamedTemporaryFile("w") as input_file, |
| 570 | tempfile.NamedTemporaryFile("r") as output_file, |
| 571 | ): |
| 572 | input_file.write(input_batch) |
| 573 | input_file.flush() |
| 574 | proc = subprocess.Popen( |
| 575 | [ |
| 576 | "vllm", |
| 577 | "run-batch", |
| 578 | "-i", |
| 579 | input_file.name, |
| 580 | "-o", |
| 581 | output_file.name, |
| 582 | "--model", |
| 583 | RERANKER_MODEL_NAME, |
| 584 | ], |
| 585 | ) |
| 586 | proc.communicate() |
| 587 | proc.wait() |
| 588 | assert proc.returncode == 0, f"{proc=}" |
| 589 | |
| 590 | contents = output_file.read() |
| 591 | for line in contents.strip().split("\n"): |
| 592 | # Ensure that the output format conforms to the openai api. |
| 593 | # Validation should throw if the schema is wrong. |
| 594 | BatchRequestOutput.model_validate_json(line) |
| 595 | |
| 596 | # Ensure that there is no error in the response. |
| 597 | line_dict = json.loads(line) |
| 598 | assert isinstance(line_dict, dict) |
| 599 | assert line_dict["error"] is None |
| 600 | |
| 601 | |
| 602 | def test_reasoning_parser(): |
| 603 | """Test that reasoning_parser parameter works correctly in run_batch.""" |
| 604 | with ( |
| 605 | tempfile.NamedTemporaryFile("w") as input_file, |
| 606 | tempfile.NamedTemporaryFile("r") as output_file, |
| 607 | ): |
| 608 | input_file.write(INPUT_REASONING_BATCH) |
| 609 | input_file.flush() |
| 610 | proc = subprocess.Popen( |
| 611 | [ |
| 612 | "vllm", |
| 613 | "run-batch", |
| 614 | "-i", |
| 615 | input_file.name, |
| 616 | "-o", |
| 617 | output_file.name, |
| 618 | "--model", |
| 619 | REASONING_MODEL_NAME, |
| 620 | "--reasoning-parser", |
| 621 | "qwen3", |
| 622 | ], |
| 623 | ) |
| 624 | proc.communicate() |
| 625 | proc.wait() |
| 626 | assert proc.returncode == 0, f"{proc=}" |
| 627 | |
| 628 | contents = output_file.read() |
| 629 | for line in contents.strip().split("\n"): |
| 630 | # Ensure that the output format conforms to the openai api. |
| 631 | # Validation should throw if the schema is wrong. |
| 632 | BatchRequestOutput.model_validate_json(line) |
| 633 | |
| 634 | # Ensure that there is no error in the response. |
| 635 | line_dict = json.loads(line) |
| 636 | assert isinstance(line_dict, dict) |
| 637 | assert line_dict["error"] is None |
| 638 | |
| 639 | # Check that reasoning is present and not empty |
| 640 | reasoning = line_dict["response"]["body"]["choices"][0]["message"][ |
| 641 | "reasoning" |
| 642 | ] |
| 643 | assert reasoning is not None |
| 644 | assert len(reasoning) > 0 |
| 645 | |
| 646 | |
| 647 | def test_transcription(): |
| 648 | with ( |
| 649 | tempfile.NamedTemporaryFile("w") as input_file, |
| 650 | tempfile.NamedTemporaryFile("r") as output_file, |
| 651 | ): |
| 652 | input_file.write(INPUT_TRANSCRIPTION_BATCH) |
| 653 | input_file.flush() |
| 654 | proc = subprocess.Popen( |
| 655 | [ |
| 656 | "vllm", |
| 657 | "run-batch", |
| 658 | "-i", |
| 659 | input_file.name, |
| 660 | "-o", |
| 661 | output_file.name, |
| 662 | "--model", |
| 663 | SPEECH_LARGE_MODEL_NAME, |
| 664 | ], |
| 665 | ) |
| 666 | proc.communicate() |
| 667 | proc.wait() |
| 668 | assert proc.returncode == 0, f"{proc=}" |
| 669 | |
| 670 | contents = output_file.read() |
| 671 | print(f"\n\ncontents: {contents}\n\n") |
| 672 | for line in contents.strip().split("\n"): |
| 673 | BatchRequestOutput.model_validate_json(line) |
| 674 | |
| 675 | line_dict = json.loads(line) |
| 676 | assert isinstance(line_dict, dict) |
| 677 | assert line_dict["error"] is None |
| 678 | |
| 679 | response_body = line_dict["response"]["body"] |
| 680 | assert response_body is not None |
| 681 | assert "text" in response_body |
| 682 | assert "usage" in response_body |
| 683 | |
| 684 | |
| 685 | def test_transcription_http_url(): |
| 686 | with ( |
| 687 | tempfile.NamedTemporaryFile("w") as input_file, |
| 688 | tempfile.NamedTemporaryFile("r") as output_file, |
| 689 | ): |
| 690 | input_file.write(INPUT_TRANSCRIPTION_HTTP_BATCH) |
| 691 | input_file.flush() |
| 692 | proc = subprocess.Popen( |
| 693 | [ |
| 694 | "vllm", |
| 695 | "run-batch", |
| 696 | "-i", |
| 697 | input_file.name, |
| 698 | "-o", |
| 699 | output_file.name, |
| 700 | "--model", |
| 701 | SPEECH_LARGE_MODEL_NAME, |
| 702 | ], |
| 703 | ) |
| 704 | proc.communicate() |
| 705 | proc.wait() |
| 706 | assert proc.returncode == 0, f"{proc=}" |
| 707 | |
| 708 | contents = output_file.read() |
| 709 | for line in contents.strip().split("\n"): |
| 710 | BatchRequestOutput.model_validate_json(line) |
| 711 | |
| 712 | line_dict = json.loads(line) |
| 713 | assert isinstance(line_dict, dict) |
| 714 | assert line_dict["error"] is None |
| 715 | |
| 716 | response_body = line_dict["response"]["body"] |
| 717 | assert response_body is not None |
| 718 | assert "text" in response_body |
| 719 | assert "usage" in response_body |
| 720 | |
| 721 | transcription_text = response_body["text"] |
| 722 | assert "Mary had a little lamb" in transcription_text |
| 723 | |
| 724 | |
| 725 | def test_translation(): |
| 726 | with ( |
| 727 | tempfile.NamedTemporaryFile("w") as input_file, |
| 728 | tempfile.NamedTemporaryFile("r") as output_file, |
| 729 | ): |
| 730 | input_file.write(INPUT_TRANSLATION_BATCH) |
| 731 | input_file.flush() |
| 732 | proc = subprocess.Popen( |
| 733 | [ |
| 734 | "vllm", |
| 735 | "run-batch", |
| 736 | "-i", |
| 737 | input_file.name, |
| 738 | "-o", |
| 739 | output_file.name, |
| 740 | "--model", |
| 741 | SPEECH_SMALL_MODEL_NAME, |
| 742 | ], |
| 743 | ) |
| 744 | proc.communicate() |
| 745 | proc.wait() |
| 746 | assert proc.returncode == 0, f"{proc=}" |
| 747 | |
| 748 | contents = output_file.read() |
| 749 | for line in contents.strip().split("\n"): |
| 750 | BatchRequestOutput.model_validate_json(line) |
| 751 | |
| 752 | line_dict = json.loads(line) |
| 753 | assert isinstance(line_dict, dict) |
| 754 | assert line_dict["error"] is None |
| 755 | |
| 756 | response_body = line_dict["response"]["body"] |
| 757 | assert response_body is not None |
| 758 | assert "text" in response_body |
| 759 | |
| 760 | translation_text = response_body["text"] |
| 761 | translation_text_lower = str(translation_text).strip().lower() |
| 762 | assert "mary" in translation_text_lower or "lamb" in translation_text_lower |
| 763 | |
| 764 | |
| 765 | def test_tool_calling(): |
| 766 | """Test that tool calling works correctly in run_batch. |
| 767 | Verifies that requests with tools return tool_calls in the response. |
| 768 | """ |
| 769 | with ( |
| 770 | tempfile.NamedTemporaryFile("w") as input_file, |
| 771 | tempfile.NamedTemporaryFile("r") as output_file, |
| 772 | ): |
| 773 | input_file.write(INPUT_TOOL_CALLING_BATCH) |
| 774 | input_file.flush() |
| 775 | proc = subprocess.Popen( |
| 776 | [ |
| 777 | "vllm", |
| 778 | "run-batch", |
| 779 | "-i", |
| 780 | input_file.name, |
| 781 | "-o", |
| 782 | output_file.name, |
| 783 | "--model", |
| 784 | REASONING_MODEL_NAME, |
| 785 | "--enable-auto-tool-choice", |
| 786 | "--tool-call-parser", |
| 787 | "hermes", |
| 788 | ], |
| 789 | ) |
| 790 | proc.communicate() |
| 791 | proc.wait() |
| 792 | assert proc.returncode == 0, f"{proc=}" |
| 793 | |
| 794 | contents = output_file.read() |
| 795 | for line in contents.strip().split("\n"): |
| 796 | if not line.strip(): # Skip empty lines |
| 797 | continue |
| 798 | # Ensure that the output format conforms to the openai api. |
| 799 | # Validation should throw if the schema is wrong. |
| 800 | BatchRequestOutput.model_validate_json(line) |
| 801 | |
| 802 | # Ensure that there is no error in the response. |
| 803 | line_dict = json.loads(line) |
| 804 | assert isinstance(line_dict, dict) |
| 805 | assert line_dict["error"] is None |
| 806 | |
| 807 | # Check that tool_calls are present in the response |
| 808 | # With tool_choice="required", the model must call a tool |
| 809 | response_body = line_dict["response"]["body"] |
| 810 | assert response_body is not None |
| 811 | message = response_body["choices"][0]["message"] |
| 812 | assert "tool_calls" in message |
| 813 | tool_calls = message.get("tool_calls") |
| 814 | # With tool_choice="required", tool_calls must be present and non-empty |
| 815 | assert tool_calls is not None |
| 816 | assert isinstance(tool_calls, list) |
| 817 | assert len(tool_calls) > 0 |
| 818 | # Verify tool_calls have the expected structure |
| 819 | for tool_call in tool_calls: |
| 820 | assert "id" in tool_call |
| 821 | assert "type" in tool_call |
| 822 | assert tool_call["type"] == "function" |
| 823 | assert "function" in tool_call |
| 824 | assert "name" in tool_call["function"] |
| 825 | assert "arguments" in tool_call["function"] |
| 826 | # Verify the tool name matches our tool definition |
| 827 | assert tool_call["function"]["name"] == "get_current_weather" |
| 828 | |
| 829 | |
| 830 | # --------------------------------------------------------------------------- |
| 831 | # Unit tests for download_bytes_from_url SSRF protection |
| 832 | # --------------------------------------------------------------------------- |
| 833 | |
| 834 | |
| 835 | def _make_aiohttp_mocks(response_data: bytes = b"fake-data", status: int = 200): |
| 836 | """Create mock objects that simulate aiohttp.ClientSession context managers.""" |
| 837 | |
| 838 | async def iter_chunked(chunk_size: int): |
| 839 | del chunk_size |
| 840 | yield response_data |
| 841 | |
| 842 | mock_resp = MagicMock() |
| 843 | mock_resp.status = status |
| 844 | mock_resp.content_length = len(response_data) |
| 845 | mock_resp.content.iter_chunked = iter_chunked |
| 846 | mock_resp.read = AsyncMock(return_value=response_data) |
| 847 | mock_resp.__aenter__ = AsyncMock(return_value=mock_resp) |
| 848 | mock_resp.__aexit__ = AsyncMock(return_value=False) |
| 849 | |
| 850 | mock_session = MagicMock() |
| 851 | mock_session.get = MagicMock(return_value=mock_resp) |
| 852 | mock_session.__aenter__ = AsyncMock(return_value=mock_session) |
| 853 | mock_session.__aexit__ = AsyncMock(return_value=False) |
| 854 | return mock_session |
| 855 | |
| 856 | |
| 857 | @pytest.mark.asyncio |
| 858 | async def test_download_bytes_data_url_bypasses_domain_check(): |
| 859 | """data: URLs must work regardless of the domain allowlist.""" |
| 860 | data_url = f"data:audio/wav;base64,{MINIMAL_WAV_BASE64}" |
| 861 | result = await download_bytes_from_url( |
| 862 | data_url, allowed_media_domains=["example.com"] |
| 863 | ) |
| 864 | assert isinstance(result, bytes) |
| 865 | assert len(result) > 0 |
| 866 | |
| 867 | |
| 868 | @pytest.mark.asyncio |
| 869 | async def test_download_bytes_rejects_disallowed_domain(): |
| 870 | """HTTP URLs whose hostname is not in the allowlist must be rejected.""" |
| 871 | url = "https://evil.internal/secret" |
| 872 | with pytest.raises(VLLMValidationError, match="allowed domains") as exc_info: |
| 873 | await download_bytes_from_url(url, allowed_media_domains=["example.com"]) |
| 874 | # URL validation failures carry structured metadata for the frontend. |
| 875 | assert exc_info.value.parameter == "url" |
| 876 | assert exc_info.value.value == "evil.internal" |
| 877 | |
| 878 | |
| 879 | @pytest.mark.asyncio |
| 880 | async def test_download_bytes_rejects_unsupported_scheme(): |
| 881 | """Unsupported URL schemes are rejected with structured metadata.""" |
| 882 | with pytest.raises(VLLMValidationError, match="Unsupported URL scheme") as exc_info: |
| 883 | await download_bytes_from_url("ftp://example.com/file") |
| 884 | assert exc_info.value.parameter == "url" |
| 885 | assert exc_info.value.value == "ftp" |
| 886 | |
| 887 | |
| 888 | @pytest.mark.asyncio |
| 889 | async def test_download_bytes_rejects_cloud_metadata_ip(): |
| 890 | """Cloud metadata endpoints must be blocked when an allowlist is set.""" |
| 891 | url = "http://169.254.169.254/latest/meta-data/" |
| 892 | with pytest.raises(VLLMValidationError, match="allowed domains"): |
| 893 | await download_bytes_from_url(url, allowed_media_domains=["example.com"]) |
| 894 | |
| 895 | |
| 896 | @pytest.mark.asyncio |
| 897 | async def test_download_bytes_rejects_internal_ip(): |
| 898 | """Private-range IPs must be blocked when an allowlist is set.""" |
| 899 | for internal_url in [ |
| 900 | "http://10.0.0.1/secret", |
| 901 | "http://192.168.1.1/admin", |
| 902 | "http://127.0.0.1:8080/internal", |
| 903 | ]: |
| 904 | with pytest.raises(VLLMValidationError, match="allowed domains"): |
| 905 | await download_bytes_from_url( |
| 906 | internal_url, allowed_media_domains=["example.com"] |
| 907 | ) |
| 908 | |
| 909 | |
| 910 | @pytest.mark.asyncio |
| 911 | async def test_download_bytes_allows_permitted_domain(): |
| 912 | """HTTP URLs whose hostname IS in the allowlist must be fetched.""" |
| 913 | url = "https://example.com/audio.wav" |
| 914 | expected = b"audio-bytes" |
| 915 | mock_session = _make_aiohttp_mocks(expected) |
| 916 | |
| 917 | with patch( |
| 918 | "vllm.entrypoints.launchers.run_batch.aiohttp.ClientSession", |
| 919 | return_value=mock_session, |
| 920 | ): |
| 921 | result = await download_bytes_from_url( |
| 922 | url, allowed_media_domains=["example.com"] |
| 923 | ) |
| 924 | assert result == expected |
| 925 | |
| 926 | |
| 927 | @pytest.mark.asyncio |
| 928 | async def test_download_bytes_no_allowlist_permits_any_domain(): |
| 929 | """Without an allowlist all HTTP URLs must be attempted (backward compat).""" |
| 930 | url = "https://any-domain.example.org/file.wav" |
| 931 | expected = b"some-data" |
| 932 | mock_session = _make_aiohttp_mocks(expected) |
| 933 | |
| 934 | with patch( |
| 935 | "vllm.entrypoints.launchers.run_batch.aiohttp.ClientSession", |
| 936 | return_value=mock_session, |
| 937 | ): |
| 938 | result = await download_bytes_from_url(url, allowed_media_domains=None) |
| 939 | assert result == expected |
| 940 | |
| 941 | |
| 942 | @pytest.mark.asyncio |
| 943 | async def test_download_bytes_empty_allowlist_denies_all(): |
| 944 | """An empty allowlist must deny all HTTP URLs (least privilege).""" |
| 945 | url = "https://any-domain.example.org/file.wav" |
| 946 | with pytest.raises(VLLMValidationError, match="allowed domains"): |
| 947 | await download_bytes_from_url(url, allowed_media_domains=[]) |
| 948 | |
| 949 | |
| 950 | @pytest.mark.asyncio |
| 951 | async def test_download_bytes_unsupported_scheme(): |
| 952 | """Unsupported URL schemes must be rejected regardless of allowlist.""" |
| 953 | with pytest.raises(VLLMValidationError, match="Unsupported URL scheme"): |
| 954 | await download_bytes_from_url("ftp://example.com/file.wav") |
| 955 | |
| 956 | with pytest.raises(VLLMValidationError, match="Unsupported URL scheme"): |
| 957 | await download_bytes_from_url( |
| 958 | "ftp://example.com/file.wav", |
| 959 | allowed_media_domains=["example.com"], |
| 960 | ) |
| 961 | |
| 962 | |
| 963 | @pytest.mark.asyncio |
| 964 | async def test_download_bytes_backslash_bypass(): |
| 965 | """Backslash-@ URL confusion must not bypass the allowlist. |
| 966 | |
| 967 | urllib3.parse_url() and aiohttp/yarl disagree on backslash-before-@. |
| 968 | The fix normalizes through urllib3 before handing to aiohttp. |
| 969 | """ |
| 970 | bypass_url = "http://allowed.example.com\\@evil.internal/secret" |
| 971 | with pytest.raises(VLLMValidationError, match="allowed domains"): |
| 972 | await download_bytes_from_url( |
| 973 | bypass_url, allowed_media_domains=["evil.internal"] |
| 974 | ) |
| 975 | |
| 976 | |
| 977 | @pytest.mark.asyncio |
| 978 | async def test_transcription_wrapper_rejects_oversized_data_url_before_decode( |
| 979 | monkeypatch: pytest.MonkeyPatch, |
| 980 | ): |
| 981 | monkeypatch.setattr(envs, "VLLM_MAX_AUDIO_CLIP_FILESIZE_MB", 1) |
| 982 | handler = AsyncMock() |
| 983 | wrapped_handler = make_transcription_wrapper(is_translation=False)(handler) |
| 984 | request = BatchTranscriptionRequest.model_validate( |
| 985 | { |
| 986 | "model": SPEECH_LARGE_MODEL_NAME, |
| 987 | "file_url": f"data:audio/wav;base64,{_OVERSIZED_BASE64_AUDIO}", |
| 988 | "response_format": "json", |
| 989 | } |
| 990 | ) |
| 991 | |
| 992 | with patch("vllm.entrypoints.launchers.run_batch.base64.b64decode") as decode: |
| 993 | response = await wrapped_handler(request) |
| 994 | |
| 995 | assert isinstance(response, ErrorResponse) |
| 996 | assert "Maximum file size exceeded" in response.error.message |
| 997 | decode.assert_not_called() |
| 998 | handler.assert_not_awaited() |
| 999 | |
| 1000 | |
| 1001 | @pytest.mark.asyncio |
| 1002 | async def test_download_bytes_allows_http_body_at_audio_limit( |
| 1003 | monkeypatch: pytest.MonkeyPatch, |
| 1004 | batch_http_server: _BatchHTTPServer, |
| 1005 | ): |
| 1006 | monkeypatch.setattr(envs, "VLLM_MAX_AUDIO_CLIP_FILESIZE_MB", 1) |
| 1007 | connection = HTTPConnection() |
| 1008 | monkeypatch.setattr( |
| 1009 | "vllm.entrypoints.launchers.run_batch.global_http_connection", |
| 1010 | connection, |
| 1011 | ) |
| 1012 | |
| 1013 | try: |
| 1014 | result = await download_bytes_from_url( |
| 1015 | _batch_http_url(batch_http_server, "/exact-limit") |
| 1016 | ) |
| 1017 | assert result == _EXACT_LIMIT_AUDIO |
| 1018 | finally: |
| 1019 | await _close_async_connection(connection) |
| 1020 | |
| 1021 | |
| 1022 | @pytest.mark.asyncio |
| 1023 | async def test_transcription_wrapper_rejects_oversized_http_before_handler( |
| 1024 | monkeypatch: pytest.MonkeyPatch, |
| 1025 | batch_http_server: _BatchHTTPServer, |
| 1026 | ): |
| 1027 | monkeypatch.setattr(envs, "VLLM_MAX_AUDIO_CLIP_FILESIZE_MB", 1) |
| 1028 | connection = HTTPConnection() |
| 1029 | monkeypatch.setattr( |
| 1030 | "vllm.entrypoints.launchers.run_batch.global_http_connection", |
| 1031 | connection, |
| 1032 | ) |
| 1033 | handler = AsyncMock() |
| 1034 | wrapped_handler = make_transcription_wrapper(is_translation=False)(handler) |
| 1035 | request = BatchTranscriptionRequest.model_validate( |
| 1036 | { |
| 1037 | "model": SPEECH_LARGE_MODEL_NAME, |
| 1038 | "file_url": _batch_http_url(batch_http_server, "/oversized-chunked"), |
| 1039 | "response_format": "json", |
| 1040 | } |
| 1041 | ) |
| 1042 | |
| 1043 | try: |
| 1044 | response = await wrapped_handler(request) |
| 1045 | finally: |
| 1046 | await _close_async_connection(connection) |
| 1047 | |
| 1048 | assert isinstance(response, ErrorResponse) |
| 1049 | assert "Maximum file size exceeded" in response.error.message |
| 1050 | handler.assert_not_awaited() |
| 1051 | |
| 1052 | |
| 1053 | # --------------------------------------------------------------------------- |
| 1054 | # Unit tests for upload_data retry behavior |
| 1055 | # --------------------------------------------------------------------------- |
| 1056 | |
| 1057 | |
| 1058 | def _make_aiohttp_put_session(status: int = 200, body_text: str = ""): |
| 1059 | """Mock an aiohttp.ClientSession whose PUT returns the given status.""" |
| 1060 | mock_resp = MagicMock() |
| 1061 | mock_resp.status = status |
| 1062 | mock_resp.text = AsyncMock(return_value=body_text) |
| 1063 | mock_resp.__aenter__ = AsyncMock(return_value=mock_resp) |
| 1064 | mock_resp.__aexit__ = AsyncMock(return_value=False) |
| 1065 | |
| 1066 | mock_session = MagicMock() |
| 1067 | mock_session.put = MagicMock(return_value=mock_resp) |
| 1068 | mock_session.__aenter__ = AsyncMock(return_value=mock_session) |
| 1069 | mock_session.__aexit__ = AsyncMock(return_value=False) |
| 1070 | return mock_session |
| 1071 | |
| 1072 | |
| 1073 | @pytest.mark.asyncio |
| 1074 | async def test_upload_data_uploads_once_on_success(): |
| 1075 | """A successful upload must not be retried (regression guard).""" |
| 1076 | session = _make_aiohttp_put_session(status=200) |
| 1077 | with patch( |
| 1078 | "vllm.entrypoints.launchers.run_batch.aiohttp.ClientSession", |
| 1079 | return_value=session, |
| 1080 | ): |
| 1081 | await upload_data( |
| 1082 | "https://example.com/output.jsonl", "payload", from_file=False |
| 1083 | ) |
| 1084 | assert session.put.call_count == 1 |
| 1085 | |
| 1086 | |
| 1087 | @pytest.mark.asyncio |
| 1088 | async def test_upload_data_error_includes_awaited_response_body(): |
| 1089 | """A failed upload must surface the awaited response body, not a coroutine.""" |
| 1090 | session = _make_aiohttp_put_session(status=500, body_text="server-error-detail") |
| 1091 | with ( |
| 1092 | patch( |
| 1093 | "vllm.entrypoints.launchers.run_batch.aiohttp.ClientSession", |
| 1094 | return_value=session, |
| 1095 | ), |
| 1096 | patch( |
| 1097 | "vllm.entrypoints.launchers.run_batch.asyncio.sleep", |
| 1098 | AsyncMock(), |
| 1099 | ), |
| 1100 | pytest.raises(Exception) as exc_info, |
| 1101 | ): |
| 1102 | await upload_data( |
| 1103 | "https://example.com/output.jsonl", "payload", from_file=False |
| 1104 | ) |
| 1105 | message = str(exc_info.value) |
| 1106 | assert "server-error-detail" in message |
| 1107 | assert "coroutine" not in message |