test_openai_server.py 30.1 KB
Newer Older
1
2
# imports for guided decoding tests
import json
3
import os
4
import re
5
import subprocess
6
import sys
7
import time
8

9
10
import jsonschema
import openai  # use the official client for correctness check
11
import pytest
12
13
14
# using Ray for overall ease of process management, parallel requests,
# and debugging.
import ray
15
import requests
16
17
# downloading lora to test lora requests
from huggingface_hub import snapshot_download
18
from openai import BadRequestError
19

20
21
from vllm.transformers_utils.tokenizer import get_tokenizer

22
MAX_SERVER_START_WAIT_S = 600  # wait for server to start for 60 seconds
23
24
25
26
27
# any model with a chat template should work here
MODEL_NAME = "HuggingFaceH4/zephyr-7b-beta"
# technically this needs Mistral-7B-v0.1 as base, but we're not testing
# generation quality here
LORA_NAME = "typeof/zephyr-7b-beta-lora"
28

29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
TEST_SCHEMA = {
    "type": "object",
    "properties": {
        "name": {
            "type": "string"
        },
        "age": {
            "type": "integer"
        },
        "skills": {
            "type": "array",
            "items": {
                "type": "string",
                "maxLength": 10
            },
            "minItems": 3
        },
        "work history": {
            "type": "array",
            "items": {
                "type": "object",
                "properties": {
                    "company": {
                        "type": "string"
                    },
                    "duration": {
                        "type": "string"
                    },
                    "position": {
                        "type": "string"
                    }
                },
                "required": ["company", "position"]
            }
        }
    },
    "required": ["name", "age", "skills", "work history"]
}

68
69
TEST_REGEX = (r"((25[0-5]|(2[0-4]|1\d|[1-9]|)\d)\.){3}"
              r"(25[0-5]|(2[0-4]|1\d|[1-9]|)\d)")
70
71
72
73
74
75

TEST_CHOICE = [
    "Python", "Java", "JavaScript", "C++", "C#", "PHP", "TypeScript", "Ruby",
    "Swift", "Kotlin"
]

76
77
78
79
80
81
82
pytestmark = pytest.mark.asyncio


@ray.remote(num_gpus=1)
class ServerRunner:

    def __init__(self, args):
83
84
        env = os.environ.copy()
        env["PYTHONUNBUFFERED"] = "1"
85
86
        self.proc = subprocess.Popen(
            ["python3", "-m", "vllm.entrypoints.openai.api_server"] + args,
87
            env=env,
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
            stdout=sys.stdout,
            stderr=sys.stderr,
        )
        self._wait_for_server()

    def ready(self):
        return True

    def _wait_for_server(self):
        # run health check
        start = time.time()
        while True:
            try:
                if requests.get(
                        "http://localhost:8000/health").status_code == 200:
                    break
            except Exception as err:
                if self.proc.poll() is not None:
                    raise RuntimeError("Server exited unexpectedly.") from err

                time.sleep(0.5)
                if time.time() - start > MAX_SERVER_START_WAIT_S:
                    raise RuntimeError(
                        "Server failed to start in time.") from err

    def __del__(self):
        if hasattr(self, "proc"):
            self.proc.terminate()


@pytest.fixture(scope="session")
119
120
121
122
123
124
def zephyr_lora_files():
    return snapshot_download(repo_id=LORA_NAME)


@pytest.fixture(scope="session")
def server(zephyr_lora_files):
125
126
127
128
    ray.init()
    server_runner = ServerRunner.remote([
        "--model",
        MODEL_NAME,
129
        # use half precision for speed and memory savings in CI environment
130
        "--dtype",
131
        "bfloat16",
132
        "--max-model-len",
133
134
        "8192",
        "--enforce-eager",
135
136
137
138
139
140
141
142
143
144
        # lora config below
        "--enable-lora",
        "--lora-modules",
        f"zephyr-lora={zephyr_lora_files}",
        f"zephyr-lora2={zephyr_lora_files}",
        "--max-lora-rank",
        "64",
        "--max-cpu-loras",
        "2",
        "--max-num-seqs",
145
        "128",
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
    ])
    ray.get(server_runner.ready.remote())
    yield server_runner
    ray.shutdown()


@pytest.fixture(scope="session")
def client():
    client = openai.AsyncOpenAI(
        base_url="http://localhost:8000/v1",
        api_key="token-abc123",
    )
    yield client


161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
async def test_check_models(server, client: openai.AsyncOpenAI):
    models = await client.models.list()
    models = models.data
    served_model = models[0]
    lora_models = models[1:]
    assert served_model.id == MODEL_NAME
    assert all(model.root == MODEL_NAME for model in models)
    assert lora_models[0].id == "zephyr-lora"
    assert lora_models[1].id == "zephyr-lora2"


@pytest.mark.parametrize(
    # first test base model, then test loras
    "model_name",
    [MODEL_NAME, "zephyr-lora", "zephyr-lora2"],
)
async def test_single_completion(server, client: openai.AsyncOpenAI,
                                 model_name: str):
    completion = await client.completions.create(model=model_name,
180
181
182
183
184
185
186
187
188
189
190
191
                                                 prompt="Hello, my name is",
                                                 max_tokens=5,
                                                 temperature=0.0)

    assert completion.id is not None
    assert completion.choices is not None and len(completion.choices) == 1
    assert completion.choices[0].text is not None and len(
        completion.choices[0].text) >= 5
    assert completion.choices[0].finish_reason == "length"
    assert completion.usage == openai.types.CompletionUsage(
        completion_tokens=5, prompt_tokens=6, total_tokens=11)

192
193
194
195
196
197
198
199
200
201
    # test using token IDs
    completion = await client.completions.create(
        model=MODEL_NAME,
        prompt=[0, 0, 0, 0, 0],
        max_tokens=5,
        temperature=0.0,
    )
    assert completion.choices[0].text is not None and len(
        completion.choices[0].text) >= 5

202

203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
@pytest.mark.parametrize(
    # first test base model, then test loras
    "model_name",
    [MODEL_NAME, "zephyr-lora", "zephyr-lora2"],
)
async def test_zero_logprobs(server, client: openai.AsyncOpenAI,
                             model_name: str):
    # test using token IDs
    completion = await client.completions.create(
        model=MODEL_NAME,
        prompt=[0, 0, 0, 0, 0],
        max_tokens=5,
        temperature=0.0,
        logprobs=0,
    )
    choice = completion.choices[0]
    assert choice.logprobs is not None
    assert choice.logprobs.token_logprobs is not None
    assert choice.logprobs.top_logprobs is None


224
225
226
227
228
229
230
@pytest.mark.parametrize(
    # just test 1 lora hereafter
    "model_name",
    [MODEL_NAME, "zephyr-lora"],
)
async def test_single_chat_session(server, client: openai.AsyncOpenAI,
                                   model_name: str):
231
232
233
234
235
236
237
238
239
    messages = [{
        "role": "system",
        "content": "you are a helpful assistant"
    }, {
        "role": "user",
        "content": "what is 1+1?"
    }]

    # test single completion
240
241
242
243
    chat_completion = await client.chat.completions.create(model=model_name,
                                                           messages=messages,
                                                           max_tokens=10,
                                                           logprobs=True,
244
                                                           top_logprobs=5)
245
246
247
248
    assert chat_completion.id is not None
    assert chat_completion.choices is not None and len(
        chat_completion.choices) == 1
    assert chat_completion.choices[0].message is not None
249
250
    assert chat_completion.choices[0].logprobs is not None
    assert chat_completion.choices[0].logprobs.top_logprobs is not None
251
    assert len(chat_completion.choices[0].logprobs.top_logprobs[0]) == 5
252
253
254
255
256
257
258
259
    message = chat_completion.choices[0].message
    assert message.content is not None and len(message.content) >= 10
    assert message.role == "assistant"
    messages.append({"role": "assistant", "content": message.content})

    # test multi-turn dialogue
    messages.append({"role": "user", "content": "express your result in json"})
    chat_completion = await client.chat.completions.create(
260
        model=model_name,
261
262
263
264
265
        messages=messages,
        max_tokens=10,
    )
    message = chat_completion.choices[0].message
    assert message.content is not None and len(message.content) >= 0
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320


@pytest.mark.parametrize("model_name", [MODEL_NAME])
async def test_too_many_logprobs(server, client: openai.AsyncOpenAI,
                                 model_name: str):
    messages = [{
        "role": "system",
        "content": "you are a helpful assistant"
    }, {
        "role": "user",
        "content": "what is 1+1?"
    }]

    # Default max_logprobs is 5, so this should raise an error
    with pytest.raises((openai.BadRequestError, openai.APIError)):
        stream = await client.chat.completions.create(model=model_name,
                                                      messages=messages,
                                                      max_tokens=10,
                                                      logprobs=True,
                                                      top_logprobs=10,
                                                      stream=True)
        async for chunk in stream:
            ...

    with pytest.raises(openai.BadRequestError):
        await client.chat.completions.create(model=model_name,
                                             messages=messages,
                                             max_tokens=10,
                                             logprobs=True,
                                             top_logprobs=10,
                                             stream=False)

    with pytest.raises((openai.BadRequestError, openai.APIError)):
        stream = await client.completions.create(model=model_name,
                                                 prompt="Test",
                                                 max_tokens=10,
                                                 logprobs=10,
                                                 stream=True)
        async for chunk in stream:
            ...

    with pytest.raises(openai.BadRequestError):
        await client.completions.create(model=model_name,
                                        prompt="Test",
                                        max_tokens=10,
                                        logprobs=10,
                                        stream=False)

    # the server should still work afterwards
    chat_completion = await client.chat.completions.create(model=model_name,
                                                           messages=messages,
                                                           max_tokens=10,
                                                           stream=False)
    message = chat_completion.choices[0].message
    assert message.content is not None and len(message.content) >= 0
321
322


323
324
325
326
327
328
329
@pytest.mark.parametrize(
    # just test 1 lora hereafter
    "model_name",
    [MODEL_NAME, "zephyr-lora"],
)
async def test_completion_streaming(server, client: openai.AsyncOpenAI,
                                    model_name: str):
330
331
332
    prompt = "What is an LLM?"

    single_completion = await client.completions.create(
333
        model=model_name,
334
335
336
337
338
339
340
        prompt=prompt,
        max_tokens=5,
        temperature=0.0,
    )
    single_output = single_completion.choices[0].text
    single_usage = single_completion.usage

341
342
343
344
345
    stream = await client.completions.create(model=model_name,
                                             prompt=prompt,
                                             max_tokens=5,
                                             temperature=0.0,
                                             stream=True)
346
    chunks = []
347
    finish_reason_count = 0
348
349
    async for chunk in stream:
        chunks.append(chunk.choices[0].text)
350
351
352
353
        if chunk.choices[0].finish_reason is not None:
            finish_reason_count += 1
    # finish reason should only return in last block
    assert finish_reason_count == 1
354
    assert chunk.choices[0].finish_reason == "length"
355
    assert chunk.choices[0].text
356
357
358
359
    assert chunk.usage == single_usage
    assert "".join(chunks) == single_output


360
361
362
363
364
365
366
@pytest.mark.parametrize(
    # just test 1 lora hereafter
    "model_name",
    [MODEL_NAME, "zephyr-lora"],
)
async def test_chat_streaming(server, client: openai.AsyncOpenAI,
                              model_name: str):
367
368
369
370
371
372
373
374
375
376
    messages = [{
        "role": "system",
        "content": "you are a helpful assistant"
    }, {
        "role": "user",
        "content": "what is 1+1?"
    }]

    # test single completion
    chat_completion = await client.chat.completions.create(
377
        model=model_name,
378
379
380
381
382
383
384
385
386
        messages=messages,
        max_tokens=10,
        temperature=0.0,
    )
    output = chat_completion.choices[0].message.content
    stop_reason = chat_completion.choices[0].finish_reason

    # test streaming
    stream = await client.chat.completions.create(
387
        model=model_name,
388
389
390
391
392
393
        messages=messages,
        max_tokens=10,
        temperature=0.0,
        stream=True,
    )
    chunks = []
394
    finish_reason_count = 0
395
396
397
398
399
400
    async for chunk in stream:
        delta = chunk.choices[0].delta
        if delta.role:
            assert delta.role == "assistant"
        if delta.content:
            chunks.append(delta.content)
401
402
403
404
        if chunk.choices[0].finish_reason is not None:
            finish_reason_count += 1
    # finish reason should only return in last block
    assert finish_reason_count == 1
405
    assert chunk.choices[0].finish_reason == stop_reason
406
    assert delta.content
407
408
409
    assert "".join(chunks) == output


410
411
412
413
414
415
416
@pytest.mark.parametrize(
    # just test 1 lora hereafter
    "model_name",
    [MODEL_NAME, "zephyr-lora"],
)
async def test_batch_completions(server, client: openai.AsyncOpenAI,
                                 model_name: str):
417
418
    # test simple list
    batch = await client.completions.create(
419
        model=model_name,
420
421
422
423
424
425
426
427
428
        prompt=["Hello, my name is", "Hello, my name is"],
        max_tokens=5,
        temperature=0.0,
    )
    assert len(batch.choices) == 2
    assert batch.choices[0].text == batch.choices[1].text

    # test n = 2
    batch = await client.completions.create(
429
        model=model_name,
430
431
432
433
434
        prompt=["Hello, my name is", "Hello, my name is"],
        n=2,
        max_tokens=5,
        temperature=0.0,
        extra_body=dict(
435
436
            # NOTE: this has to be true for n > 1 in vLLM, but not necessary
            # for official client.
437
438
439
440
441
442
443
444
445
446
447
448
            use_beam_search=True),
    )
    assert len(batch.choices) == 4
    assert batch.choices[0].text != batch.choices[
        1].text, "beam search should be different"
    assert batch.choices[0].text == batch.choices[
        2].text, "two copies of the same prompt should be the same"
    assert batch.choices[1].text == batch.choices[
        3].text, "two copies of the same prompt should be the same"

    # test streaming
    batch = await client.completions.create(
449
        model=model_name,
450
451
452
453
454
455
456
457
458
459
460
461
462
        prompt=["Hello, my name is", "Hello, my name is"],
        max_tokens=5,
        temperature=0.0,
        stream=True,
    )
    texts = [""] * 2
    async for chunk in batch:
        assert len(chunk.choices) == 1
        choice = chunk.choices[0]
        texts[choice.index] += choice.text
    assert texts[0] == texts[1]


463
464
465
466
467
468
469
470
471
472
473
474
475
async def test_logits_bias(server, client: openai.AsyncOpenAI):
    prompt = "Hello, my name is"
    max_tokens = 5
    tokenizer = get_tokenizer(tokenizer_name=MODEL_NAME)

    # Test exclusive selection
    token_id = 1000
    completion = await client.completions.create(
        model=MODEL_NAME,
        prompt=prompt,
        max_tokens=max_tokens,
        temperature=0.0,
        logit_bias={str(token_id): 100},
476
        seed=42,
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
    )
    assert completion.choices[0].text is not None and len(
        completion.choices[0].text) >= 5
    response_tokens = tokenizer(completion.choices[0].text,
                                add_special_tokens=False)["input_ids"]
    expected_tokens = tokenizer(tokenizer.decode([token_id] * 5),
                                add_special_tokens=False)["input_ids"]
    assert all([
        response == expected
        for response, expected in zip(response_tokens, expected_tokens)
    ])

    # Test ban
    completion = await client.completions.create(
        model=MODEL_NAME,
        prompt=prompt,
        max_tokens=max_tokens,
        temperature=0.0,
    )
    response_tokens = tokenizer(completion.choices[0].text,
                                add_special_tokens=False)["input_ids"]
    first_response = completion.choices[0].text
    completion = await client.completions.create(
        model=MODEL_NAME,
        prompt=prompt,
        max_tokens=max_tokens,
        temperature=0.0,
        logit_bias={str(token): -100
                    for token in response_tokens},
    )
    assert first_response != completion.choices[0].text


510
511
512
513
@pytest.mark.parametrize("guided_decoding_backend",
                         ["outlines", "lm-format-enforcer"])
async def test_guided_json_completion(server, client: openai.AsyncOpenAI,
                                      guided_decoding_backend: str):
514
515
    completion = await client.completions.create(
        model=MODEL_NAME,
516
517
        prompt=f"Give an example JSON for an employee profile "
        f"that fits this schema: {TEST_SCHEMA}",
518
519
520
        n=3,
        temperature=1.0,
        max_tokens=500,
521
522
        extra_body=dict(guided_json=TEST_SCHEMA,
                        guided_decoding_backend=guided_decoding_backend))
523
524
525
526
527
528
529
530
531

    assert completion.id is not None
    assert completion.choices is not None and len(completion.choices) == 3
    for i in range(3):
        assert completion.choices[i].text is not None
        output_json = json.loads(completion.choices[i].text)
        jsonschema.validate(instance=output_json, schema=TEST_SCHEMA)


532
533
534
535
@pytest.mark.parametrize("guided_decoding_backend",
                         ["outlines", "lm-format-enforcer"])
async def test_guided_json_chat(server, client: openai.AsyncOpenAI,
                                guided_decoding_backend: str):
536
537
538
539
    messages = [{
        "role": "system",
        "content": "you are a helpful assistant"
    }, {
540
541
542
543
544
        "role":
        "user",
        "content":
        f"Give an example JSON for an employee profile that "
        f"fits this schema: {TEST_SCHEMA}"
545
546
547
548
    }]
    chat_completion = await client.chat.completions.create(
        model=MODEL_NAME,
        messages=messages,
549
550
551
        max_tokens=1000,
        extra_body=dict(guided_json=TEST_SCHEMA,
                        guided_decoding_backend=guided_decoding_backend))
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
    message = chat_completion.choices[0].message
    assert message.content is not None
    json1 = json.loads(message.content)
    jsonschema.validate(instance=json1, schema=TEST_SCHEMA)

    messages.append({"role": "assistant", "content": message.content})
    messages.append({
        "role":
        "user",
        "content":
        "Give me another one with a different name and age"
    })
    chat_completion = await client.chat.completions.create(
        model=MODEL_NAME,
        messages=messages,
567
568
569
        max_tokens=1000,
        extra_body=dict(guided_json=TEST_SCHEMA,
                        guided_decoding_backend=guided_decoding_backend))
570
571
572
573
574
575
576
577
    message = chat_completion.choices[0].message
    assert message.content is not None
    json2 = json.loads(message.content)
    jsonschema.validate(instance=json2, schema=TEST_SCHEMA)
    assert json1["name"] != json2["name"]
    assert json1["age"] != json2["age"]


578
579
580
581
@pytest.mark.parametrize("guided_decoding_backend",
                         ["outlines", "lm-format-enforcer"])
async def test_guided_regex_completion(server, client: openai.AsyncOpenAI,
                                       guided_decoding_backend: str):
582
583
584
585
586
587
    completion = await client.completions.create(
        model=MODEL_NAME,
        prompt=f"Give an example IPv4 address with this regex: {TEST_REGEX}",
        n=3,
        temperature=1.0,
        max_tokens=20,
588
589
        extra_body=dict(guided_regex=TEST_REGEX,
                        guided_decoding_backend=guided_decoding_backend))
590
591
592
593
594
595
596
597

    assert completion.id is not None
    assert completion.choices is not None and len(completion.choices) == 3
    for i in range(3):
        assert completion.choices[i].text is not None
        assert re.fullmatch(TEST_REGEX, completion.choices[i].text) is not None


598
599
600
601
@pytest.mark.parametrize("guided_decoding_backend",
                         ["outlines", "lm-format-enforcer"])
async def test_guided_regex_chat(server, client: openai.AsyncOpenAI,
                                 guided_decoding_backend: str):
602
603
604
605
606
607
608
609
610
611
612
613
614
    messages = [{
        "role": "system",
        "content": "you are a helpful assistant"
    }, {
        "role":
        "user",
        "content":
        f"Give an example IP address with this regex: {TEST_REGEX}"
    }]
    chat_completion = await client.chat.completions.create(
        model=MODEL_NAME,
        messages=messages,
        max_tokens=20,
615
616
        extra_body=dict(guided_regex=TEST_REGEX,
                        guided_decoding_backend=guided_decoding_backend))
617
618
619
620
621
622
623
624
625
626
    ip1 = chat_completion.choices[0].message.content
    assert ip1 is not None
    assert re.fullmatch(TEST_REGEX, ip1) is not None

    messages.append({"role": "assistant", "content": ip1})
    messages.append({"role": "user", "content": "Give me a different one"})
    chat_completion = await client.chat.completions.create(
        model=MODEL_NAME,
        messages=messages,
        max_tokens=20,
627
628
        extra_body=dict(guided_regex=TEST_REGEX,
                        guided_decoding_backend=guided_decoding_backend))
629
630
631
632
633
634
    ip2 = chat_completion.choices[0].message.content
    assert ip2 is not None
    assert re.fullmatch(TEST_REGEX, ip2) is not None
    assert ip1 != ip2


635
636
637
638
@pytest.mark.parametrize("guided_decoding_backend",
                         ["outlines", "lm-format-enforcer"])
async def test_guided_choice_completion(server, client: openai.AsyncOpenAI,
                                        guided_decoding_backend: str):
639
640
641
642
643
644
    completion = await client.completions.create(
        model=MODEL_NAME,
        prompt="The best language for type-safe systems programming is ",
        n=2,
        temperature=1.0,
        max_tokens=10,
645
646
        extra_body=dict(guided_choice=TEST_CHOICE,
                        guided_decoding_backend=guided_decoding_backend))
647
648
649
650
651
652
653

    assert completion.id is not None
    assert completion.choices is not None and len(completion.choices) == 2
    for i in range(2):
        assert completion.choices[i].text in TEST_CHOICE


654
655
656
657
@pytest.mark.parametrize("guided_decoding_backend",
                         ["outlines", "lm-format-enforcer"])
async def test_guided_choice_chat(server, client: openai.AsyncOpenAI,
                                  guided_decoding_backend: str):
658
659
660
661
662
663
664
665
666
667
668
669
670
    messages = [{
        "role": "system",
        "content": "you are a helpful assistant"
    }, {
        "role":
        "user",
        "content":
        "The best language for type-safe systems programming is "
    }]
    chat_completion = await client.chat.completions.create(
        model=MODEL_NAME,
        messages=messages,
        max_tokens=10,
671
672
        extra_body=dict(guided_choice=TEST_CHOICE,
                        guided_decoding_backend=guided_decoding_backend))
673
674
675
676
677
678
679
680
681
682
683
684
    choice1 = chat_completion.choices[0].message.content
    assert choice1 in TEST_CHOICE

    messages.append({"role": "assistant", "content": choice1})
    messages.append({
        "role": "user",
        "content": "I disagree, pick another one"
    })
    chat_completion = await client.chat.completions.create(
        model=MODEL_NAME,
        messages=messages,
        max_tokens=10,
685
686
        extra_body=dict(guided_choice=TEST_CHOICE,
                        guided_decoding_backend=guided_decoding_backend))
687
688
689
690
691
    choice2 = chat_completion.choices[0].message.content
    assert choice2 in TEST_CHOICE
    assert choice1 != choice2


692
693
694
695
@pytest.mark.parametrize("guided_decoding_backend",
                         ["outlines", "lm-format-enforcer"])
async def test_guided_decoding_type_error(server, client: openai.AsyncOpenAI,
                                          guided_decoding_backend: str):
696
697
698
699
    with pytest.raises(openai.BadRequestError):
        _ = await client.completions.create(
            model=MODEL_NAME,
            prompt="Give an example JSON that fits this schema: 42",
700
701
            extra_body=dict(guided_json=42,
                            guided_decoding_backend=guided_decoding_backend))
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726

    messages = [{
        "role": "system",
        "content": "you are a helpful assistant"
    }, {
        "role":
        "user",
        "content":
        "The best language for type-safe systems programming is "
    }]
    with pytest.raises(openai.BadRequestError):
        _ = await client.chat.completions.create(model=MODEL_NAME,
                                                 messages=messages,
                                                 extra_body=dict(guided_regex={
                                                     1: "Python",
                                                     2: "C++"
                                                 }))

    with pytest.raises(openai.BadRequestError):
        _ = await client.completions.create(
            model=MODEL_NAME,
            prompt="Give an example string that fits this regex",
            extra_body=dict(guided_regex=TEST_REGEX, guided_json=TEST_SCHEMA))


727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
@pytest.mark.parametrize("guided_decoding_backend",
                         ["outlines", "lm-format-enforcer"])
async def test_guided_choice_chat_logprobs(server, client: openai.AsyncOpenAI,
                                           guided_decoding_backend: str):
    messages = [{
        "role": "system",
        "content": "you are a helpful assistant"
    }, {
        "role":
        "user",
        "content":
        "The best language for type-safe systems programming is "
    }]
    chat_completion = await client.chat.completions.create(
        model=MODEL_NAME,
        messages=messages,
        max_tokens=10,
        logprobs=True,
        top_logprobs=5,
        extra_body=dict(guided_choice=TEST_CHOICE,
                        guided_decoding_backend=guided_decoding_backend))
    top_logprobs = chat_completion.choices[0].logprobs.top_logprobs

    # -9999.0 is the minimum logprob returned by OpenAI
    assert all(
        isinstance(logprob, float) and logprob >= -9999.0
        for token_dict in top_logprobs
        for token, logprob in token_dict.items())


757
async def test_response_format_json_object(server, client: openai.AsyncOpenAI):
758
759
760
761
762
763
764
765
766
767
768
769
770
771
    for _ in range(2):
        resp = await client.chat.completions.create(
            model=MODEL_NAME,
            messages=[{
                "role":
                "user",
                "content": ('what is 1+1? please respond with a JSON object, '
                            'the format is {"result": 2}')
            }],
            response_format={"type": "json_object"})

        content = resp.choices[0].message.content
        loaded = json.loads(content)
        assert loaded == {"result": 2}, loaded
772
773


774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
async def test_extra_fields(server, client: openai.AsyncOpenAI):
    with pytest.raises(BadRequestError) as exc_info:
        await client.chat.completions.create(
            model=MODEL_NAME,
            messages=[{
                "role": "system",
                "content": "You are a helpful assistant.",
                "extra_field": "0",
            }],  # type: ignore
            temperature=0,
            seed=0)

    assert "extra_forbidden" in exc_info.value.message


789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
async def test_complex_message_content(server, client: openai.AsyncOpenAI):
    resp = await client.chat.completions.create(
        model=MODEL_NAME,
        messages=[{
            "role":
            "user",
            "content": [{
                "type":
                "text",
                "text":
                "what is 1+1? please provide the result without any other text."
            }]
        }],
        temperature=0,
        seed=0)
    content = resp.choices[0].message.content
    assert content == "2"


808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
async def test_guided_grammar(server, client: openai.AsyncOpenAI):
    simple_sql_grammar = """
start: select_statement

select_statement: "SELECT" column "from" table "where" condition

column: "col_1" | "col_2"
table: "table_1" | "table_2"
condition: column "=" number

number: "1" | "2"
"""

    completion = await client.completions.create(
        model=MODEL_NAME,
        prompt=("Generate a sql state that select col_1 from "
                "table_1 where it is equals to 1"),
        temperature=1.0,
        max_tokens=500,
        extra_body=dict(guided_grammar=simple_sql_grammar))

    content = completion.choices[0].text

    # use Lark to parse the output, and make sure it's a valid parse tree
    from lark import Lark
    parser = Lark(simple_sql_grammar)
    parser.parse(content)

    # remove spaces for comparison b/c we removed them in the grammar
    ground_truth = "SELECT col_1 from table_1 where col_1 = 1".replace(" ", "")

    assert content.strip() == ground_truth


842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
@pytest.mark.parametrize(
    # first test base model, then test loras
    "model_name",
    [MODEL_NAME, "zephyr-lora", "zephyr-lora2"],
)
async def test_echo_logprob_completion(server, client: openai.AsyncOpenAI,
                                       model_name: str):
    tokenizer = get_tokenizer(tokenizer_name=MODEL_NAME)
    # test using text and token IDs
    for prompt in ("Hello, my name is", [0, 0, 0, 0, 0]):
        completion = await client.completions.create(model=model_name,
                                                     prompt=prompt,
                                                     max_tokens=5,
                                                     temperature=0.0,
                                                     echo=True,
                                                     logprobs=1)

        prompt_text = tokenizer.decode(prompt) if isinstance(prompt,
                                                             list) else prompt
        assert (completion.choices[0].text is not None
                and re.search(r"^" + prompt_text, completion.choices[0].text))
        logprobs = completion.choices[0].logprobs
        assert logprobs is not None
        assert len(logprobs.text_offset) > 5
        assert (len(logprobs.token_logprobs) > 5
                and logprobs.token_logprobs[0] is None)
        assert (len(logprobs.top_logprobs) > 5
                and logprobs.top_logprobs[0] is None)
        assert len(logprobs.tokens) > 5


873
874
if __name__ == "__main__":
    pytest.main([__file__])