utils.py 41.5 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

4
import asyncio
5
import contextlib
6
import copy
7
import functools
8
import importlib
9
import itertools
10
import json
11
import os
12
import random
13
import signal
14
15
import subprocess
import sys
16
import tempfile
17
import time
18
import warnings
19
from collections.abc import Callable, Iterable
20
from contextlib import ExitStack, contextmanager, suppress
21
from multiprocessing import Process
22
from pathlib import Path
23
from typing import Any, Literal
24
from unittest.mock import patch
25

26
import anthropic
27
import cloudpickle
28
import httpx
29
import openai
30
import pytest
31
import requests
32
import torch
33
import torch.nn.functional as F
34
from openai.types.completion import Completion
35
from typing_extensions import ParamSpec
36

37
import vllm.envs as envs
38
from tests.models.utils import TextTextLogprobs
39
40
41
42
from vllm.distributed import (
    ensure_model_parallel_initialized,
    init_distributed_environment,
)
43
from vllm.engine.arg_utils import AsyncEngineArgs
44
from vllm.entrypoints.cli.serve import ServeSubcommand
45
from vllm.model_executor.model_loader import get_model_loader
46
from vllm.platforms import current_platform
47

48
from vllm.tokenizers import get_tokenizer
Cyrus Leung's avatar
Cyrus Leung committed
49
from vllm.utils.argparse_utils import FlexibleArgumentParser
50
from vllm.utils.mem_constants import GB_bytes
51
from vllm.utils.network_utils import get_open_port
52
from vllm.utils.torch_utils import cuda_device_count_stateless
53
54
55
56
import vllm.envs as envs
import os

models_path_prefix = os.getenv('VLLM_OPTEST_MODELS_PATH') or os.getenv("OPTEST_MODELS_PATH")
57
urls_port = int(os.getenv('VLLM_OPTEST_URLS_PORT', '8000'))
58

59
if current_platform.is_rocm():
60
61
62
63
64
65
    from amdsmi import (
        amdsmi_get_gpu_vram_usage,
        amdsmi_get_processor_handles,
        amdsmi_init,
        amdsmi_shut_down,
    )
66
67
68
69
70
71
72
73

    @contextmanager
    def _nvml():
        try:
            amdsmi_init()
            yield
        finally:
            amdsmi_shut_down()
74
elif current_platform.is_cuda():
75
76
77
78
79
80
    from vllm.third_party.pynvml import (
        nvmlDeviceGetHandleByIndex,
        nvmlDeviceGetMemoryInfo,
        nvmlInit,
        nvmlShutdown,
    )
81
82
83
84
85
86
87
88

    @contextmanager
    def _nvml():
        try:
            nvmlInit()
            yield
        finally:
            nvmlShutdown()
89
90
91
92
93
else:

    @contextmanager
    def _nvml():
        yield
94

95

96
97
VLLM_PATH = Path(__file__).parent.parent
"""Path to root of the vLLM repository."""
98
99


100
101
class RemoteOpenAIServer:
    DUMMY_API_KEY = "token-abc123"  # vLLM's OpenAI server does not need API key
102

103
    def _start_server(
104
        self, model: str, vllm_serve_args: list[str], env_dict: dict[str, str] | None
105
106
    ) -> None:
        """Subclasses override this method to customize server process launch"""
107
108
109
        env = os.environ.copy()
        # the current process might initialize cuda,
        # to be safe, we should use spawn method
110
        env["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
111
112
        if env_dict is not None:
            env.update(env_dict)
113
114
        serve_cmd = ["vllm", "serve", model, *vllm_serve_args]
        print(f"Launching RemoteOpenAIServer with: {' '.join(serve_cmd)}")
115
        print(f"Environment variables: {env}")
116
        self.proc: subprocess.Popen = subprocess.Popen(
117
            serve_cmd,
118
119
120
121
122
            env=env,
            stdout=sys.stdout,
            stderr=sys.stderr,
        )

123
124
125
126
127
    def __init__(
        self,
        model: str,
        vllm_serve_args: list[str],
        *,
128
        env_dict: dict[str, str] | None = None,
129
        seed: int = 0,
130
        auto_port: bool = True,
131
132
        max_wait_seconds: float | None = None,
        override_hf_configs: dict[str, Any] | None = None,
133
    ) -> None:
134
        if auto_port:
135
            if "-p" in vllm_serve_args or "--port" in vllm_serve_args:
136
137
138
                raise ValueError(
                    "You have manually specified the port when `auto_port=True`."
                )
139

140
141
142
            # No need for a port if using unix sockets
            if "--uds" not in vllm_serve_args:
                # Don't mutate the input args
143
                vllm_serve_args = vllm_serve_args + ["--port", str(get_open_port())]
144
145
        if seed is not None:
            if "--seed" in vllm_serve_args:
146
147
148
                raise ValueError(
                    f"You have manually specified the seed when `seed={seed}`."
                )
149
150

            vllm_serve_args = vllm_serve_args + ["--seed", str(seed)]
151

152
153
154
        if override_hf_configs is not None:
            vllm_serve_args = vllm_serve_args + [
                "--hf-overrides",
155
                json.dumps(override_hf_configs),
156
157
            ]

158
        parser = FlexibleArgumentParser(description="vLLM's remote OpenAI server.")
159
160
        subparsers = parser.add_subparsers(required=False, dest="subparser")
        parser = ServeSubcommand().subparser_init(subparsers)
161
        args = parser.parse_args(["--model", model, *vllm_serve_args])
162
163
164
165
166
        self.uds = args.uds
        if args.uds:
            self.host = None
            self.port = None
        else:
167
            self.host = str(args.host or "127.0.0.1")
168
            self.port = int(args.port)
169

170
        self.show_hidden_metrics = args.show_hidden_metrics_for_version is not None
171

172
173
174
175
        # download the model before starting the server to avoid timeout
        is_local = os.path.isdir(model)
        if not is_local:
            engine_args = AsyncEngineArgs.from_cli_args(args)
176
177
178
179
180
            model_config = engine_args.create_model_config()
            load_config = engine_args.create_load_config()

            model_loader = get_model_loader(load_config)
            model_loader.download_model(model_config)
181

182
        self._start_server(model, vllm_serve_args, env_dict)
183
        max_wait_seconds = max_wait_seconds or 240
184
        self._wait_for_server(url=self.url_for("health"), timeout=max_wait_seconds)
185
186
187
188
189
190

    def __enter__(self):
        return self

    def __exit__(self, exc_type, exc_value, traceback):
        self.proc.terminate()
191
        try:
192
            self.proc.wait(8)
193
194
195
        except subprocess.TimeoutExpired:
            # force kill if needed
            self.proc.kill()
196

197
    def _poll(self) -> int | None:
198
199
200
        """Subclasses override this method to customize process polling"""
        return self.proc.poll()

201
202
203
    def _wait_for_server(self, *, url: str, timeout: float):
        # run health check
        start = time.time()
204
205
206
207
208
        client = (
            httpx.Client(transport=httpx.HTTPTransport(uds=self.uds))
            if self.uds
            else requests
        )
209
210
        while True:
            try:
211
                if client.get(url).status_code == 200:
212
                    break
213
214
215
216
217
            except Exception:
                # this exception can only be raised by requests.get,
                # which means the server is not ready yet.
                # the stack trace is not useful, so we suppress it
                # by using `raise from None`.
218
                result = self._poll()
219
                if result is not None and result != 0:
220
                    raise RuntimeError("Server exited unexpectedly.") from None
221
222
223

                time.sleep(0.5)
                if time.time() - start > timeout:
224
                    raise RuntimeError("Server failed to start in time.") from None
225
226
227

    @property
    def url_root(self) -> str:
228
229
230
231
232
        return (
            f"http://{self.uds.split('/')[-1]}"
            if self.uds
            else f"http://{self.host}:{self.port}"
        )
233
234
235
236

    def url_for(self, *parts: str) -> str:
        return self.url_root + "/" + "/".join(parts)

237
238
239
    def get_client(self, **kwargs):
        if "timeout" not in kwargs:
            kwargs["timeout"] = 600
240
241
242
        return openai.OpenAI(
            base_url=self.url_for("v1"),
            api_key=self.DUMMY_API_KEY,
243
244
            max_retries=0,
            **kwargs,
245
246
        )

247
    def get_async_client(self, **kwargs):
248
249
        if "timeout" not in kwargs:
            kwargs["timeout"] = 600
250
251
252
253
254
255
        return openai.AsyncOpenAI(
            base_url=self.url_for("v1"),
            api_key=self.DUMMY_API_KEY,
            max_retries=0,
            **kwargs,
        )
256

257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
    def get_client_anthropic(self, **kwargs):
        if "timeout" not in kwargs:
            kwargs["timeout"] = 600
        return anthropic.Anthropic(
            base_url=self.url_for(),
            api_key=self.DUMMY_API_KEY,
            max_retries=0,
            **kwargs,
        )

    def get_async_client_anthropic(self, **kwargs):
        if "timeout" not in kwargs:
            kwargs["timeout"] = 600
        return anthropic.AsyncAnthropic(
            base_url=self.url_for(), api_key=self.DUMMY_API_KEY, max_retries=0, **kwargs
        )
273
274


275
276
277
class RemoteOpenAIServerCustom(RemoteOpenAIServer):
    """Launch test server with custom child process"""

278
    def _start_server(
279
        self, model: str, vllm_serve_args: list[str], env_dict: dict[str, str] | None
280
    ) -> None:
281
        self.proc: Process = Process(
282
283
            target=self.child_process_fxn, args=(env_dict, model, vllm_serve_args)
        )  # type: ignore[assignment]
284
285
        self.proc.start()

286
287
288
289
    def __init__(
        self,
        model: str,
        vllm_serve_args: list[str],
290
        child_process_fxn: Callable[[dict[str, str] | None, str, list[str]], None],
291
        *,
292
        env_dict: dict[str, str] | None = None,
293
        seed: int = 0,
294
        auto_port: bool = True,
295
        max_wait_seconds: float | None = None,
296
    ) -> None:
297
298
299
        """Store custom child process function then invoke superclass
        constructor which will indirectly launch it."""
        self.child_process_fxn = child_process_fxn
300
301
302
303
304
305
306
307
        super().__init__(
            model=model,
            vllm_serve_args=vllm_serve_args,
            env_dict=env_dict,
            seed=seed,
            auto_port=auto_port,
            max_wait_seconds=max_wait_seconds,
        )
308

309
    def _poll(self) -> int | None:
310
311
312
313
314
315
316
317
318
319
        return self.proc.exitcode

    def __exit__(self, exc_type, exc_value, traceback):
        self.proc.terminate()
        self.proc.join(8)
        if self.proc.is_alive():
            # force kill if needed
            self.proc.kill()


320
321
322
323
def _test_completion(
    client: openai.OpenAI,
    model: str,
    prompt: str,
324
    token_ids: list[int],
325
326
327
328
):
    results = []

    # test with text prompt
329
330
331
332
333
334
335
336
337
338
339
340
    completion = client.completions.create(
        model=model, prompt=prompt, max_tokens=5, temperature=0.0
    )

    results.append(
        {
            "test": "single_completion",
            "text": completion.choices[0].text,
            "finish_reason": completion.choices[0].finish_reason,
            "usage": completion.usage,
        }
    )
341
342
343
344
345
346
347
348
349

    # test using token IDs
    completion = client.completions.create(
        model=model,
        prompt=token_ids,
        max_tokens=5,
        temperature=0.0,
    )

350
351
352
353
354
355
356
357
    results.append(
        {
            "test": "token_ids",
            "text": completion.choices[0].text,
            "finish_reason": completion.choices[0].finish_reason,
            "usage": completion.usage,
        }
    )
358
359

    # test seeded random sampling
360
361
362
363
364
365
366
367
368
369
370
371
    completion = client.completions.create(
        model=model, prompt=prompt, max_tokens=5, seed=33, temperature=1.0
    )

    results.append(
        {
            "test": "seeded_sampling",
            "text": completion.choices[0].text,
            "finish_reason": completion.choices[0].finish_reason,
            "usage": completion.usage,
        }
    )
372
373

    # test seeded random sampling with multiple prompts
374
375
376
377
378
379
380
381
382
383
384
385
    completion = client.completions.create(
        model=model, prompt=[prompt, prompt], max_tokens=5, seed=33, temperature=1.0
    )

    results.append(
        {
            "test": "seeded_sampling",
            "text": [choice.text for choice in completion.choices],
            "finish_reason": [choice.finish_reason for choice in completion.choices],
            "usage": completion.usage,
        }
    )
386
387
388
389
390
391
392
393
394

    # test simple list
    batch = client.completions.create(
        model=model,
        prompt=[prompt, prompt],
        max_tokens=5,
        temperature=0.0,
    )

395
396
397
398
399
400
401
    results.append(
        {
            "test": "simple_list",
            "text0": batch.choices[0].text,
            "text1": batch.choices[1].text,
        }
    )
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417

    # test streaming
    batch = client.completions.create(
        model=model,
        prompt=[prompt, prompt],
        max_tokens=5,
        temperature=0.0,
        stream=True,
    )

    texts = [""] * 2
    for chunk in batch:
        assert len(chunk.choices) == 1
        choice = chunk.choices[0]
        texts[choice.index] += choice.text

418
419
420
421
422
423
    results.append(
        {
            "test": "streaming",
            "texts": texts,
        }
    )
424
425
426
427

    return results


428
429
430
431
432
433
434
435
def _test_completion_close(
    client: openai.OpenAI,
    model: str,
    prompt: str,
):
    results = []

    # test with text prompt
436
437
438
    completion = client.completions.create(
        model=model, prompt=prompt, max_tokens=1, logprobs=5, temperature=0.0
    )
439

440
441
    logprobs = completion.choices[0].logprobs.top_logprobs[0]
    logprobs = {k: round(v, 2) for k, v in logprobs.items()}
442

443
444
445
446
447
448
    results.append(
        {
            "test": "completion_close",
            "logprobs": logprobs,
        }
    )
449
450
451
452

    return results


453
454
455
456
457
458
459
def _test_chat(
    client: openai.OpenAI,
    model: str,
    prompt: str,
):
    results = []

460
    messages = [{"role": "user", "content": [{"type": "text", "text": prompt}]}]
461
462

    # test with text prompt
463
464
465
466
467
468
469
470
471
472
473
474
    chat_response = client.chat.completions.create(
        model=model, messages=messages, max_tokens=5, temperature=0.0
    )

    results.append(
        {
            "test": "completion_close",
            "text": chat_response.choices[0].message.content,
            "finish_reason": chat_response.choices[0].finish_reason,
            "usage": chat_response.usage,
        }
    )
475
476
477
478

    return results


479
480
481
482
483
484
485
486
487
488
489
490
491
492
def _test_embeddings(
    client: openai.OpenAI,
    model: str,
    text: str,
):
    results = []

    # test with text input
    embeddings = client.embeddings.create(
        model=model,
        input=text,
        encoding_format="float",
    )

493
494
495
496
497
498
499
    results.append(
        {
            "test": "single_embedding",
            "embedding": embeddings.data[0].embedding,
            "usage": embeddings.usage,
        }
    )
500
501
502
503

    return results


504
505
506
507
508
509
510
511
def _test_image_text(
    client: openai.OpenAI,
    model_name: str,
    image_url: str,
):
    results = []

    # test pure text input
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
    messages = [
        {
            "role": "user",
            "content": [
                {"type": "text", "text": "How do you feel today?"},
            ],
        }
    ]

    chat_completion = client.chat.completions.create(
        model=model_name,
        messages=messages,
        temperature=0.0,
        max_tokens=1,
        logprobs=True,
        top_logprobs=5,
    )
529
530
531
532
533
    top_logprobs = chat_completion.choices[0].logprobs.content[0].top_logprobs

    for x in top_logprobs:
        x.logprob = round(x.logprob, 2)

534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
    results.append(
        {
            "test": "pure_text",
            "logprobs": top_logprobs,
        }
    )

    messages = [
        {
            "role": "user",
            "content": [
                {"type": "image_url", "image_url": {"url": image_url}},
                {"type": "text", "text": "What's in this image?"},
            ],
        }
    ]

    chat_completion = client.chat.completions.create(
        model=model_name,
        messages=messages,
        temperature=0.0,
        max_tokens=1,
        logprobs=True,
        top_logprobs=5,
    )
559
560
    top_logprobs = chat_completion.choices[0].logprobs.content[0].top_logprobs

561
562
563
564
565
566
    results.append(
        {
            "test": "text_image",
            "logprobs": top_logprobs,
        }
    )
567
568

    return results
569
570


571
572
573
574
def compare_two_settings(
    model: str,
    arg1: list[str],
    arg2: list[str],
575
576
    env1: dict[str, str] | None = None,
    env2: dict[str, str] | None = None,
577
578
    *,
    method: str = "generate",
579
    max_wait_seconds: float | None = None,
580
) -> None:
581
    """
582
583
584
585
586
587
588
589
590
    Launch API server with two different sets of arguments/environments
    and compare the results of the API calls.

    Args:
        model: The model to test.
        arg1: The first set of arguments to pass to the API server.
        arg2: The second set of arguments to pass to the API server.
        env1: The first set of environment variables to pass to the API server.
        env2: The second set of environment variables to pass to the API server.
591
592
    """

593
594
595
596
597
598
599
600
601
    compare_all_settings(
        model,
        [arg1, arg2],
        [env1, env2],
        method=method,
        max_wait_seconds=max_wait_seconds,
    )


602
603
604
def compare_all_settings(
    model: str,
    all_args: list[list[str]],
605
    all_envs: list[dict[str, str] | None],
606
607
    *,
    method: str = "generate",
608
    max_wait_seconds: float | None = None,
609
) -> None:
610
611
612
613
614
615
616
617
618
    """
    Launch API server with several different sets of arguments/environments
    and compare the results of the API calls with the first set of arguments.
    Args:
        model: The model to test.
        all_args: A list of argument lists to pass to the API server.
        all_envs: A list of environment dictionaries to pass to the API server.
    """

619
    trust_remote_code = False
620
    for args in all_args:
621
622
623
624
625
        if "--trust-remote-code" in args:
            trust_remote_code = True
            break

    tokenizer_mode = "auto"
626
    for args in all_args:
627
628
629
630
631
632
633
634
635
        if "--tokenizer-mode" in args:
            tokenizer_mode = args[args.index("--tokenizer-mode") + 1]
            break

    tokenizer = get_tokenizer(
        model,
        trust_remote_code=trust_remote_code,
        tokenizer_mode=tokenizer_mode,
    )
636

637
638
639
640
641
642
    can_force_load_format = True

    for args in all_args:
        if "--load-format" in args:
            can_force_load_format = False
            break
643
644

    prompt = "Hello, my name is"
645
    token_ids = tokenizer(prompt).input_ids
646
    ref_results: list = []
647
    for i, (args, env) in enumerate(zip(all_args, all_envs)):
648
649
650
651
652
653
654
655
656
        if can_force_load_format:
            # we are comparing the results and
            # usually we don't need real weights.
            # we force to use dummy weights by default,
            # and it should work for most of the cases.
            # if not, we can use VLLM_TEST_FORCE_LOAD_FORMAT
            # environment variable to force the load format,
            # e.g. in quantization tests.
            args = args + ["--load-format", envs.VLLM_TEST_FORCE_LOAD_FORMAT]
657
        compare_results: list = []
658
        results = ref_results if i == 0 else compare_results
659
660
661
        with RemoteOpenAIServer(
            model, args, env_dict=env, max_wait_seconds=max_wait_seconds
        ) as server:
662
663
664
665
666
667
            client = server.get_client()

            # test models list
            models = client.models.list()
            models = models.data
            served_model = models[0]
668
669
670
671
672
673
674
            results.append(
                {
                    "test": "models_list",
                    "id": served_model.id,
                    "root": served_model.root,
                }
            )
675

676
677
            if method == "generate":
                results += _test_completion(client, model, prompt, token_ids)
678
679
            elif method == "generate_close":
                results += _test_completion_close(client, model, prompt)
680
681
            elif method == "generate_chat":
                results += _test_chat(client, model, prompt)
682
683
            elif method == "generate_with_image":
                results += _test_image_text(
684
685
                    client,
                    model,
686
                    "https://vllm-public-assets.s3.us-west-2.amazonaws.com/vision_model_images/RGBA_comp.png",
687
                )
688
689
690
            elif method == "encode":
                results += _test_embeddings(client, model, prompt)
            else:
691
                raise ValueError(f"Unknown method: {method}")
692

693
694
695
696
697
698
            if i > 0:
                # if any setting fails, raise an error early
                ref_args = all_args[0]
                ref_envs = all_envs[0]
                compare_args = all_args[i]
                compare_envs = all_envs[i]
699
                for ref_result, compare_result in zip(ref_results, compare_results):
700
701
702
                    ref_result = copy.deepcopy(ref_result)
                    compare_result = copy.deepcopy(compare_result)
                    if "embedding" in ref_result and method == "encode":
703
704
705
706
707
708
                        sim = F.cosine_similarity(
                            torch.tensor(ref_result["embedding"]),
                            torch.tensor(compare_result["embedding"]),
                            dim=0,
                        )
                        assert sim >= 0.999, (
709
                            f"Embedding for {model=} are not the same.\n"
710
711
                            f"cosine_similarity={sim}\n"
                        )
712
713
                        del ref_result["embedding"]
                        del compare_result["embedding"]
714
715
716
717
718
                    assert ref_result == compare_result, (
                        f"Results for {model=} are not the same.\n"
                        f"{ref_args=} {ref_envs=}\n"
                        f"{compare_args=} {compare_envs=}\n"
                        f"{ref_result=}\n"
719
720
                        f"{compare_result=}\n"
                    )
721
722


723
724
725
726
727
728
729
def init_test_distributed_environment(
    tp_size: int,
    pp_size: int,
    rank: int,
    distributed_init_port: str,
    local_rank: int = -1,
) -> None:
730
731
732
733
734
735
736
    # Note: This function is often called from Ray worker processes, so we
    # can't rely on pytest fixtures to set the config. We check if the config
    # is already set and only create a default one if needed.
    from vllm.config import (
        VllmConfig,
        get_current_vllm_config_or_none,
        set_current_vllm_config,
737
    )
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757

    distributed_init_method = f"tcp://localhost:{distributed_init_port}"

    if get_current_vllm_config_or_none() is not None:
        # Config already set, use it directly
        init_distributed_environment(
            world_size=pp_size * tp_size,
            rank=rank,
            distributed_init_method=distributed_init_method,
            local_rank=local_rank,
        )
    else:
        # No config set, create a default one for the test
        with set_current_vllm_config(VllmConfig()):
            init_distributed_environment(
                world_size=pp_size * tp_size,
                rank=rank,
                distributed_init_method=distributed_init_method,
                local_rank=local_rank,
            )
758
759
760
    ensure_model_parallel_initialized(tp_size, pp_size)


761
def multi_process_parallel(
762
    monkeypatch: pytest.MonkeyPatch,
763
764
    tp_size: int,
    pp_size: int,
765
    test_target: Any,
766
) -> None:
767
768
    import ray

769
770
    # Using ray helps debugging the error when it failed
    # as compared to multiprocessing.
771
772
    # NOTE: We need to set working_dir for distributed tests,
    # otherwise we may get import errors on ray workers
zhuwenwen's avatar
zhuwenwen committed
773
774
775
776
777
778
    # NOTE: Force ray not to use gitignore file as excluding, otherwise
    # it will not move .so files to working dir.
    # So we have to manually add some of large directories
    os.environ["RAY_RUNTIME_ENV_IGNORE_GITIGNORE"] = "1"
    ray.init(
        runtime_env={
779
            "working_dir": VLLM_PATH,
780
            "excludes": [
781
782
783
784
785
786
787
788
789
                "build",
                ".git",
                "cmake-build-*",
                "shellcheck",
                "dist",
                "ep_kernels_workspace",
            ],
        }
    )
790
791
792
793
794

    distributed_init_port = get_open_port()
    refs = []
    for rank in range(tp_size * pp_size):
        refs.append(
795
796
797
798
799
800
            test_target.remote(
                monkeypatch,
                tp_size,
                pp_size,
                rank,
                distributed_init_port,
801
802
            ),
        )
803
804
805
    ray.get(refs)

    ray.shutdown()
806
807
808


@contextmanager
809
def error_on_warning(category: type[Warning] = Warning):
810
811
    """
    Within the scope of this context manager, tests will fail if any warning
812
    of the given category is emitted.
813
814
    """
    with warnings.catch_warnings():
815
        warnings.filterwarnings("error", category=category)
816
817

        yield
818
819


820
821
822
823
824
825
826
827
828
829
def get_physical_device_indices(devices):
    visible_devices = os.environ.get("CUDA_VISIBLE_DEVICES")
    if visible_devices is None:
        return devices

    visible_indices = [int(x) for x in visible_devices.split(",")]
    index_mapping = {i: physical for i, physical in enumerate(visible_indices)}
    return [index_mapping[i] for i in devices if i in index_mapping]


830
@_nvml()
831
832
833
def wait_for_gpu_memory_to_clear(
    *,
    devices: list[int],
834
835
    threshold_bytes: int | None = None,
    threshold_ratio: float | None = None,
836
837
    timeout_s: float = 120,
) -> None:
838
    assert threshold_bytes is not None or threshold_ratio is not None
839
840
    # Use nvml instead of pytorch to reduce measurement error from torch cuda
    # context.
841
    devices = get_physical_device_indices(devices)
842
843
    start_time = time.time()
    while True:
844
        output: dict[int, str] = {}
845
        output_raw: dict[int, tuple[float, float]] = {}
846
        for device in devices:
847
            if current_platform.is_rocm():
848
849
850
                dev_handle = amdsmi_get_processor_handles()[device]
                mem_info = amdsmi_get_gpu_vram_usage(dev_handle)
                gb_used = mem_info["vram_used"] / 2**10
851
                gb_total = mem_info["vram_total"] / 2**10
852
853
854
855
            else:
                dev_handle = nvmlDeviceGetHandleByIndex(device)
                mem_info = nvmlDeviceGetMemoryInfo(dev_handle)
                gb_used = mem_info.used / 2**30
856
857
                gb_total = mem_info.total / 2**30
            output_raw[device] = (gb_used, gb_total)
858
            output[device] = f"{gb_used:.02f}/{gb_total:.02f}"
859

860
        print("gpu memory used/total (GiB): ", end="")
861
        for k, v in output.items():
862
863
            print(f"{k}={v}; ", end="")
        print("")
864

865
866
        if threshold_bytes is not None:
            is_free = lambda used, total: used <= threshold_bytes / 2**30
867
            threshold = f"{threshold_bytes / 2**30} GiB"
868
869
870
871
        else:
            is_free = lambda used, total: used / total <= threshold_ratio
            threshold = f"{threshold_ratio:.2f}"

872
        dur_s = time.time() - start_time
873
        if all(is_free(used, total) for used, total in output_raw.values()):
874
875
876
877
            print(
                f"Done waiting for free GPU memory on devices {devices=} "
                f"({threshold=}) {dur_s=:.02f}"
            )
878
879
880
            break

        if dur_s >= timeout_s:
881
882
883
884
            raise ValueError(
                f"Memory of devices {devices=} not free after "
                f"{dur_s=:.02f} ({threshold=})"
            )
885
886

        time.sleep(5)
887
888


889
890
891
_P = ParamSpec("_P")


892
def fork_new_process_for_each_test(func: Callable[_P, None]) -> Callable[_P, None]:
893
894
895
    """Decorator to fork a new process for each test function.
    See https://github.com/vllm-project/vllm/issues/7053 for more details.
    """
896

897
    @functools.wraps(func)
898
    def wrapper(*args: _P.args, **kwargs: _P.kwargs) -> None:
899
900
901
902
        # Make the process the leader of its own process group
        # to avoid sending SIGTERM to the parent process
        os.setpgrp()
        from _pytest.outcomes import Skipped
903
904
905

        # Create a unique temporary file to store exception info from child
        # process. Use test function name and process ID to avoid collisions.
906
907
        with (
            tempfile.NamedTemporaryFile(
908
                delete=False,
909
                mode="w+b",
910
                prefix=f"vllm_test_{func.__name__}_{os.getpid()}_",
911
912
913
914
                suffix=".exc",
            ) as exc_file,
            ExitStack() as delete_after,
        ):
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
            exc_file_path = exc_file.name
            delete_after.callback(os.remove, exc_file_path)

            pid = os.fork()
            print(f"Fork a new process to run a test {pid}")
            if pid == 0:
                # Parent process responsible for deleting, don't delete
                # in child.
                delete_after.pop_all()
                try:
                    func(*args, **kwargs)
                except Skipped as e:
                    # convert Skipped to exit code 0
                    print(str(e))
                    os._exit(0)
                except Exception as e:
                    import traceback
932

933
934
935
936
937
938
939
                    tb_string = traceback.format_exc()

                    # Try to serialize the exception object first
                    exc_to_serialize: dict[str, Any]
                    try:
                        # First, try to pickle the actual exception with
                        # its traceback.
940
                        exc_to_serialize = {"pickled_exception": e}
941
942
943
944
945
                        # Test if it can be pickled
                        cloudpickle.dumps(exc_to_serialize)
                    except (Exception, KeyboardInterrupt):
                        # Fall back to string-based approach.
                        exc_to_serialize = {
946
947
948
                            "exception_type": type(e).__name__,
                            "exception_msg": str(e),
                            "traceback": tb_string,
949
950
                        }
                    try:
951
                        with open(exc_file_path, "wb") as f:
952
953
954
955
956
957
958
                            cloudpickle.dump(exc_to_serialize, f)
                    except Exception:
                        # Fallback: just print the traceback.
                        print(tb_string)
                    os._exit(1)
                else:
                    os._exit(0)
959
            else:
960
961
962
                pgid = os.getpgid(pid)
                _pid, _exitcode = os.waitpid(pid, 0)
                # ignore SIGTERM signal itself
963
                old_signal_handler = signal.signal(signal.SIGTERM, signal.SIG_IGN)
964
965
966
967
968
969
970
971
                # kill all child processes
                os.killpg(pgid, signal.SIGTERM)
                # restore the signal handler
                signal.signal(signal.SIGTERM, old_signal_handler)
                if _exitcode != 0:
                    # Try to read the exception from the child process
                    exc_info = {}
                    if os.path.exists(exc_file_path):
972
973
974
975
                        with (
                            contextlib.suppress(Exception),
                            open(exc_file_path, "rb") as f,
                        ):
976
977
                            exc_info = cloudpickle.load(f)

978
979
980
                    if (
                        original_exception := exc_info.get("pickled_exception")
                    ) is not None:
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
                        # Re-raise the actual exception object if it was
                        # successfully pickled.
                        assert isinstance(original_exception, Exception)
                        raise original_exception

                    if (original_tb := exc_info.get("traceback")) is not None:
                        # Use string-based traceback for fallback case
                        raise AssertionError(
                            f"Test {func.__name__} failed when called with"
                            f" args {args} and kwargs {kwargs}"
                            f" (exit code: {_exitcode}):\n{original_tb}"
                        ) from None

                    # Fallback to the original generic error
                    raise AssertionError(
                        f"function {func.__name__} failed when called with"
                        f" args {args} and kwargs {kwargs}"
998
999
                        f" (exit code: {_exitcode})"
                    ) from None
1000
1001

    return wrapper
1002
1003


1004
1005
def spawn_new_process_for_each_test(f: Callable[_P, None]) -> Callable[_P, None]:
    """Decorator to spawn a new process for each test function."""
1006
1007
1008
1009

    @functools.wraps(f)
    def wrapper(*args: _P.args, **kwargs: _P.kwargs) -> None:
        # Check if we're already in a subprocess
1010
        if os.environ.get("RUNNING_IN_SUBPROCESS") == "1":
1011
1012
1013
1014
            # If we are, just run the function directly
            return f(*args, **kwargs)

        import torch.multiprocessing as mp
1015

1016
        with suppress(RuntimeError):
1017
            mp.set_start_method("spawn")
1018
1019
1020
1021
1022
1023

        # Get the module
        module_name = f.__module__

        # Create a process with environment variable set
        env = os.environ.copy()
1024
        env["RUNNING_IN_SUBPROCESS"] = "1"
1025
1026
1027
1028
1029
1030
1031

        with tempfile.TemporaryDirectory() as tempdir:
            output_filepath = os.path.join(tempdir, "new_process.tmp")

            # `cloudpickle` allows pickling complex functions directly
            input_bytes = cloudpickle.dumps((f, output_filepath))

1032
1033
1034
1035
1036
            repo_root = str(VLLM_PATH.resolve())

            env = dict(env or os.environ)
            env["PYTHONPATH"] = repo_root + os.pathsep + env.get("PYTHONPATH", "")

1037
1038
            cmd = [sys.executable, "-m", f"{module_name}"]

1039
1040
1041
            returned = subprocess.run(
                cmd, input=input_bytes, capture_output=True, env=env
            )
1042
1043
1044
1045
1046
1047

            # check if the subprocess is successful
            try:
                returned.check_returncode()
            except Exception as e:
                # wrap raised exception to provide more information
1048
1049
1050
                raise RuntimeError(
                    f"Error raised in subprocess:\n{returned.stderr.decode()}"
                ) from e
1051
1052
1053
1054
1055

    return wrapper


def create_new_process_for_each_test(
1056
    method: Literal["spawn", "fork"] | None = None,
1057
1058
1059
1060
) -> Callable[[Callable[_P, None]], Callable[_P, None]]:
    """Creates a decorator that runs each test function in a new process.

    Args:
1061
        method: The process creation method. Can be either "spawn" or "fork".
1062
1063
               If not specified, it defaults to "spawn" on ROCm and XPU
               platforms and "fork" otherwise.
1064
1065
1066
1067
1068

    Returns:
        A decorator to run test functions in separate processes.
    """
    if method is None:
1069
1070
        use_spawn = current_platform.is_rocm() or current_platform.is_xpu()
        method = "spawn" if use_spawn else "fork"
1071

1072
    assert method in ["spawn", "fork"], "Method must be either 'spawn' or 'fork'"
1073
1074
1075
1076
1077
1078
1079

    if method == "fork":
        return fork_new_process_for_each_test

    return spawn_new_process_for_each_test


1080
def large_gpu_mark(min_gb: int) -> pytest.MarkDecorator:
1081
    """
1082
1083
    Get a pytest mark, which skips the test if the GPU doesn't meet
    a minimum memory requirement in GB.
1084

1085
1086
    This can be leveraged via `@large_gpu_test` to skip tests in environments
    without enough resources, or called when filtering tests to run directly.
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
    """
    try:
        if current_platform.is_cpu():
            memory_gb = 0
        else:
            memory_gb = current_platform.get_device_total_memory() / GB_bytes
    except Exception as e:
        warnings.warn(
            f"An error occurred when finding the available memory: {e}",
            stacklevel=2,
        )
        memory_gb = 0

1100
    return pytest.mark.skipif(
1101
        memory_gb < min_gb,
1102
        reason=f"Need at least {min_gb}GB GPU memory to run the test.",
1103
1104
    )

1105

1106
1107
1108
1109
1110
1111
1112
requires_fp8 = pytest.mark.skipif(
    not current_platform.supports_fp8(),
    reason="FP8 is not supported on this GPU (requires Hopper or "
    "Ada architecture, compute capability 8.9+)",
)


1113
def large_gpu_test(*, min_gb: int):
1114
    """
1115
1116
1117
1118
1119
    Decorate a test to be skipped if no GPU is available or it does not have
    sufficient memory.

    Currently, the CI machine uses L4 GPU which has 24 GB VRAM.
    """
1120
    mark = large_gpu_mark(min_gb)
1121

1122
    def wrapper(f: Callable[_P, None]) -> Callable[_P, None]:
1123
        return mark(f)
1124
1125
1126
1127

    return wrapper


1128
1129
1130
def multi_gpu_marks(*, num_gpus: int):
    """Get a collection of pytest marks to apply for `@multi_gpu_test`."""
    test_selector = pytest.mark.distributed(num_gpus=num_gpus)
1131
1132
1133
1134
1135
    test_skipif = pytest.mark.skipif(
        cuda_device_count_stateless() < num_gpus,
        reason=f"Need at least {num_gpus} GPUs to run the test.",
    )

1136
1137
1138
1139
1140
1141
1142
1143
1144
    return [test_selector, test_skipif]


def multi_gpu_test(*, num_gpus: int):
    """
    Decorate a test to be run only when multiple GPUs are available.
    """
    marks = multi_gpu_marks(num_gpus=num_gpus)

1145
    def wrapper(f: Callable[_P, None]) -> Callable[_P, None]:
1146
        func = create_new_process_for_each_test()(f)
1147
1148
1149
1150
        for mark in reversed(marks):
            func = mark(func)

        return func
1151
1152
1153
1154

    return wrapper


1155
async def completions_with_server_args(
1156
    prompts: list[str],
1157
    model_name: str,
1158
    server_cli_args: list[str],
1159
    num_logprobs: int | None,
1160
    max_wait_seconds: int = 240,
1161
    max_tokens: int | list = 5,
1162
) -> list[Completion]:
1163
    """Construct a remote OpenAI server, obtain an async client to the
1164
1165
1166
1167
1168
1169
1170
1171
1172
    server & invoke the completions API to obtain completions.

    Args:
      prompts: test prompts
      model_name: model to spin up on the vLLM server
      server_cli_args: CLI args for starting the server
      num_logprobs: Number of logprobs to report (or `None`)
      max_wait_seconds: timeout interval for bringing up server.
                        Default: 240sec
1173
1174
1175
      max_tokens: max_tokens value for each of the given input prompts.
        if only one max_token value is given, the same value is used
        for all the prompts.
1176
1177
1178

    Returns:
      OpenAI Completion instance
1179
    """
1180

1181
1182
1183
1184
1185
    if isinstance(max_tokens, int):
        max_tokens = [max_tokens] * len(prompts)

    assert len(max_tokens) == len(prompts)

1186
    outputs = None
1187
1188
1189
    with RemoteOpenAIServer(
        model_name, server_cli_args, max_wait_seconds=max_wait_seconds
    ) as server:
1190
        client = server.get_async_client()
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
        outputs = [
            client.completions.create(
                model=model_name,
                prompt=[p],
                temperature=0,
                stream=False,
                max_tokens=max_tok,
                logprobs=num_logprobs,
            )
            for p, max_tok in zip(prompts, max_tokens)
        ]
1202
1203
        outputs = await asyncio.gather(*outputs)

1204
    assert outputs is not None, "Completion API call failed."
1205
1206
1207
1208

    return outputs


1209
def get_client_text_generations(completions: list[Completion]) -> list[str]:
1210
    """Extract generated tokens from the output of a
1211
    request made to an Open-AI-protocol completions endpoint.
1212
    """
1213
1214
    assert all([len(x.choices) == 1 for x in completions])
    return [x.choices[0].text for x in completions]
1215
1216
1217


def get_client_text_logprob_generations(
1218
1219
1220
    completions: list[Completion],
) -> list[TextTextLogprobs]:
    """Operates on the output of a request made to an Open-AI-protocol
1221
    completions endpoint; obtains top-rank logprobs for each token in
1222
    each {class}`SequenceGroup`
1223
    """
1224
    text_generations = get_client_text_generations(completions)
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
    text = "".join(text_generations)
    return [
        (
            text_generations,
            text,
            (None if x.logprobs is None else x.logprobs.top_logprobs),
        )
        for completion in completions
        for x in completion.choices
    ]
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245


def has_module_attribute(module_name, attribute_name):
    """
    Helper function to check if a module has a specific attribute.
    """
    try:
        module = importlib.import_module(module_name)
        return hasattr(module, attribute_name)
    except ImportError:
        return False
1246
1247
1248
1249


def get_attn_backend_list_based_on_platform() -> list[str]:
    if current_platform.is_cuda():
1250
        return ["FLASH_ATTN", "TRITON_ATTN", "TREE_ATTN"]
1251
    elif current_platform.is_rocm():
1252
        attn_backend_list = ["TRITON_ATTN"]
1253
1254
        try:
            import aiter  # noqa: F401
1255

1256
            attn_backend_list.append("ROCM_AITER_FA")
1257
        except Exception:
1258
            print("Skip ROCM_AITER_FA on ROCm as aiter is not installed")
1259
1260

        return attn_backend_list
1261
1262
    elif current_platform.is_xpu():
        return ["FLASH_ATTN", "TRITON_ATTN"]
1263
1264
    else:
        raise ValueError("Unsupported platform")
1265
1266
1267
1268
1269


@contextmanager
def override_cutlass_fp8_supported(value: bool):
    with patch(
1270
1271
1272
        "vllm.model_executor.layers.quantization.utils.w8a8_utils.cutlass_fp8_supported",
        return_value=value,
    ):
1273
        yield
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292


def prep_prompts(batch_size: int, ln_range: tuple[int, int] = (800, 1100)):
    """
    Generate prompts which a bunch of assignments,
    then asking for the value of one of them.
    The prompt is just under 10k tokens; sliding window is 4k
    so the answer is outside sliding window, but should still be correct.
    Args:
        batch_size: number of prompts to generate
        ln_range: an argument to control the length of the prompt
    """
    prompts: list[str] = []
    answer: list[int] = []
    indices: list[int] = []
    random.seed(1)
    for _ in range(batch_size):
        idx = random.randint(30, 90)
        indices.append(idx)
1293
1294
1295
1296
        prompt = (
            "```python\n# We set a number of variables, "
            + f"x{idx} will be important later\n"
        )
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
        ln = random.randint(*ln_range)
        for k in range(30, ln):
            v = random.randint(10, 99)
            if k == idx:
                answer.append(v)
            prompt += f"x{k} = {v}\n"
        prompt += f"# Now, we check the value of x{idx}:\n"
        prompt += f"assert x{idx} == "
        prompts.append(prompt)
    return prompts, answer, indices


1309
1310
1311
def check_answers(
    indices: list[int], answer: list[int], outputs: list[str], accept_rate: float = 0.7
):
1312
1313
1314
1315
1316
1317
1318
1319
1320
    answer2 = [int(text[0:2].strip()) for text in outputs]
    print(list(zip(indices, zip(answer, answer2))))
    numok = 0
    for a1, a2 in zip(answer, answer2):
        if a1 == a2:
            numok += 1
    frac_ok = numok / len(answer)
    print(f"Num OK: {numok}/{len(answer)} {frac_ok}")
    assert frac_ok >= accept_rate
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340


def flat_product(*iterables: Iterable[Any]):
    """
    Flatten lists of tuples of the cartesian product.
    Useful when we want to avoid nested tuples to allow
    test params to be unpacked directly from the decorator.

    Example:
    flat_product([(1, 2), (3, 4)], ["a", "b"]) ->
    [
      (1, 2, "a"),
      (1, 2, "b"),
      (3, 4, "a"),
      (3, 4, "b"),
    ]
    """
    for element in itertools.product(*iterables):
        normalized = (e if isinstance(e, tuple) else (e,) for e in element)
        yield tuple(itertools.chain(*normalized))