test_openai_server.py 47.2 KB
Newer Older
1
2
3
# imports for guided decoding tests
import json
import re
4
from typing import List
5

6
7
import jsonschema
import openai  # use the official client for correctness check
8
import pytest
9
10
11
# using Ray for overall ease of process management, parallel requests,
# and debugging.
import ray
12
import requests
13
import torch
14
15
# downloading lora to test lora requests
from huggingface_hub import snapshot_download
16
from openai import BadRequestError
17

18
19
from vllm.transformers_utils.tokenizer import get_tokenizer

20
from ..utils import RemoteOpenAIServer
21

22
23
24
25
26
# 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"
27

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
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"]
}

67
68
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)")
69
70
71
72
73
74

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

75
pytestmark = pytest.mark.openai
76
77
78


@pytest.fixture(scope="session")
79
80
81
82
def zephyr_lora_files():
    return snapshot_download(repo_id=LORA_NAME)


83
@pytest.fixture(scope="module")
84
def ray_ctx():
85
    ray.init()
86
87
88
89
90
91
92
    yield
    ray.shutdown()


@pytest.fixture(scope="module")
def server(zephyr_lora_files, ray_ctx):
    return RemoteOpenAIServer([
93
94
        "--model",
        MODEL_NAME,
95
        # use half precision for speed and memory savings in CI environment
96
        "--dtype",
97
        "bfloat16",
98
        "--max-model-len",
99
100
        "8192",
        "--enforce-eager",
101
102
103
104
105
106
107
108
109
110
        # 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",
111
        "128",
112
113
114
    ])


115
@pytest.fixture(scope="module")
116
117
def client(server):
    return server.get_async_client()
118
119


120
async def test_check_models(client: openai.AsyncOpenAI):
121
122
123
124
125
126
127
128
129
130
    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"


131
@pytest.mark.asyncio
132
133
134
135
136
@pytest.mark.parametrize(
    # first test base model, then test loras
    "model_name",
    [MODEL_NAME, "zephyr-lora", "zephyr-lora2"],
)
137
async def test_single_completion(client: openai.AsyncOpenAI, model_name: str):
138
    completion = await client.completions.create(model=model_name,
139
140
141
142
143
144
                                                 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
145
146
147
148

    choice = completion.choices[0]
    assert len(choice.text) >= 5
    assert choice.finish_reason == "length"
149
150
151
    assert completion.usage == openai.types.CompletionUsage(
        completion_tokens=5, prompt_tokens=6, total_tokens=11)

152
153
154
155
156
157
158
    # test using token IDs
    completion = await client.completions.create(
        model=MODEL_NAME,
        prompt=[0, 0, 0, 0, 0],
        max_tokens=5,
        temperature=0.0,
    )
159
    assert len(completion.choices[0].text) >= 5
160

161

162
163
164
165
166
167
@pytest.mark.asyncio
@pytest.mark.parametrize(
    # first test base model, then test loras
    "model_name",
    [MODEL_NAME, "zephyr-lora", "zephyr-lora2"],
)
168
async def test_no_logprobs(client: openai.AsyncOpenAI, model_name: str):
169
170
171
172
173
174
175
176
177
178
179
180
    # 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=None,
    )
    choice = completion.choices[0]
    assert choice.logprobs is None


181
@pytest.mark.asyncio
182
@pytest.mark.parametrize(
183
    # just test 1 lora hereafter
184
    "model_name",
185
    [MODEL_NAME, "zephyr-lora"],
186
)
187
async def test_zero_logprobs(client: openai.AsyncOpenAI, model_name: str):
188
189
190
191
192
193
194
195
196
197
198
    # 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
199
    assert choice.logprobs.top_logprobs is not None
200
    assert len(choice.logprobs.top_logprobs[0]) == 1
201
202
203
204
205
206
207


@pytest.mark.asyncio
@pytest.mark.parametrize(
    "model_name",
    [MODEL_NAME, "zephyr-lora"],
)
208
async def test_some_logprobs(client: openai.AsyncOpenAI, model_name: str):
209
210
211
212
213
214
215
216
217
218
219
220
    # 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=5,
    )
    choice = completion.choices[0]
    assert choice.logprobs is not None
    assert choice.logprobs.token_logprobs is not None
    assert choice.logprobs.top_logprobs is not None
221
    assert 5 <= len(choice.logprobs.top_logprobs[0]) <= 6
222
223
224
225
226
227
228


@pytest.mark.asyncio
@pytest.mark.parametrize(
    "model_name",
    [MODEL_NAME, "zephyr-lora"],
)
229
async def test_too_many_completion_logprobs(client: openai.AsyncOpenAI,
230
231
232
233
234
235
236
237
238
                                            model_name: str):

    with pytest.raises(
        (openai.BadRequestError, openai.APIError)):  # test using token IDs
        await client.completions.create(
            model=MODEL_NAME,
            prompt=[0, 0, 0, 0, 0],
            max_tokens=5,
            temperature=0.0,
239
240
241
            # vLLM has higher default max_logprobs (20 instead of 5) to support
            # both Completion API and Chat Completion API
            logprobs=21,
242
243
244
245
246
247
248
249
250
        )
        ...
    with pytest.raises(
        (openai.BadRequestError, openai.APIError)):  # test using token IDs
        stream = await client.completions.create(
            model=MODEL_NAME,
            prompt=[0, 0, 0, 0, 0],
            max_tokens=5,
            temperature=0.0,
251
252
253
            # vLLM has higher default max_logprobs (20 instead of 5) to support
            # both Completion API and Chat Completion API
            logprobs=30,
254
255
256
257
258
259
260
261
262
263
264
265
            stream=True,
        )
        async for chunk in stream:
            ...

    # the server should still work afterwards
    completion = await client.completions.create(
        model=model_name,
        prompt=[0, 0, 0, 0, 0],
        max_tokens=5,
        temperature=0.0,
    )
266
    assert len(completion.choices[0].text) >= 0
267
268


269
270
271
272
273
274
@pytest.mark.asyncio
@pytest.mark.parametrize(
    # first test base model, then test loras
    "model_name",
    [MODEL_NAME, "zephyr-lora", "zephyr-lora2"],
)
275
async def test_no_logprobs_chat(client: openai.AsyncOpenAI, model_name: str):
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
    messages = [{
        "role": "system",
        "content": "you are a helpful assistant"
    }, {
        "role": "user",
        "content": "what is 1+1?"
    }]

    chat_completion = await client.chat.completions.create(model=model_name,
                                                           messages=messages,
                                                           max_tokens=5,
                                                           temperature=0.0,
                                                           logprobs=False)

    choice = chat_completion.choices[0]
    assert choice.logprobs is None


@pytest.mark.asyncio
@pytest.mark.parametrize(
    # just test 1 lora hereafter
    "model_name",
    [MODEL_NAME, "zephyr-lora"],
)
300
async def test_zero_logprobs_chat(client: openai.AsyncOpenAI, model_name: str):
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
    messages = [{
        "role": "system",
        "content": "you are a helpful assistant"
    }, {
        "role": "user",
        "content": "what is 1+1?"
    }]

    chat_completion = await client.chat.completions.create(model=model_name,
                                                           messages=messages,
                                                           max_tokens=5,
                                                           temperature=0.0,
                                                           logprobs=True,
                                                           top_logprobs=0)

    choice = chat_completion.choices[0]
    assert choice.logprobs is not None
    assert choice.logprobs.content is not None
319
    assert len(choice.logprobs.content[0].top_logprobs) == 0
320
321
322
323
324
325
326


@pytest.mark.asyncio
@pytest.mark.parametrize(
    "model_name",
    [MODEL_NAME, "zephyr-lora"],
)
327
async def test_some_logprobs_chat(client: openai.AsyncOpenAI, model_name: str):
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
    messages = [{
        "role": "system",
        "content": "you are a helpful assistant"
    }, {
        "role": "user",
        "content": "what is 1+1?"
    }]

    chat_completion = await client.chat.completions.create(model=model_name,
                                                           messages=messages,
                                                           max_tokens=5,
                                                           temperature=0.0,
                                                           logprobs=True,
                                                           top_logprobs=5)

    choice = chat_completion.choices[0]
    assert choice.logprobs is not None
    assert choice.logprobs.content is not None
346
    assert len(choice.logprobs.content[0].top_logprobs) == 5
347
348


349
@pytest.mark.asyncio
350
351
352
353
@pytest.mark.parametrize(
    "model_name",
    [MODEL_NAME, "zephyr-lora"],
)
354
async def test_too_many_chat_logprobs(client: openai.AsyncOpenAI,
355
                                      model_name: str):
356
357
358
359
360
361
362
363
    messages = [{
        "role": "system",
        "content": "you are a helpful assistant"
    }, {
        "role": "user",
        "content": "what is 1+1?"
    }]

364
    # Default max_logprobs is 20, so this should raise an error
365
366
367
368
369
    with pytest.raises((openai.BadRequestError, openai.APIError)):
        stream = await client.chat.completions.create(model=model_name,
                                                      messages=messages,
                                                      max_tokens=10,
                                                      logprobs=True,
370
                                                      top_logprobs=21,
371
372
373
374
375
376
377
378
379
                                                      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,
380
                                             top_logprobs=30,
381
382
383
384
385
386
387
388
389
                                             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
390
391


392
@pytest.mark.asyncio
393
@pytest.mark.parametrize(
394
395
396
    "model_name",
    [MODEL_NAME, "zephyr-lora"],
)
397
async def test_single_chat_session(client: openai.AsyncOpenAI,
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
                                   model_name: str):
    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(model=model_name,
                                                           messages=messages,
                                                           max_tokens=10,
                                                           logprobs=True,
                                                           top_logprobs=5)
    assert chat_completion.id is not None
    assert len(chat_completion.choices) == 1

    choice = chat_completion.choices[0]
    assert choice.finish_reason == "length"
    assert chat_completion.usage == openai.types.CompletionUsage(
        completion_tokens=10, prompt_tokens=37, total_tokens=47)

    message = choice.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(
        model=model_name,
        messages=messages,
        max_tokens=10,
    )
    message = chat_completion.choices[0].message
    assert message.content is not None and len(message.content) >= 0


@pytest.mark.asyncio
@pytest.mark.parametrize(
439
440
441
    "model_name",
    [MODEL_NAME, "zephyr-lora"],
)
442
async def test_completion_streaming(client: openai.AsyncOpenAI,
443
                                    model_name: str):
444
445
446
    prompt = "What is an LLM?"

    single_completion = await client.completions.create(
447
        model=model_name,
448
449
450
451
452
        prompt=prompt,
        max_tokens=5,
        temperature=0.0,
    )
    single_output = single_completion.choices[0].text
453
454
455
456
457
    stream = await client.completions.create(model=model_name,
                                             prompt=prompt,
                                             max_tokens=5,
                                             temperature=0.0,
                                             stream=True)
458
    chunks: List[str] = []
459
    finish_reason_count = 0
460
461
    async for chunk in stream:
        chunks.append(chunk.choices[0].text)
462
463
464
465
        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
466
    assert chunk.choices[0].finish_reason == "length"
467
    assert chunk.choices[0].text
468
469
470
    assert "".join(chunks) == single_output


471
@pytest.mark.asyncio
472
473
474
475
476
@pytest.mark.parametrize(
    # just test 1 lora hereafter
    "model_name",
    [MODEL_NAME, "zephyr-lora"],
)
477
async def test_chat_streaming(client: openai.AsyncOpenAI, model_name: str):
478
479
480
481
482
483
484
485
486
487
    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(
488
        model=model_name,
489
490
491
492
493
494
495
496
497
        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(
498
        model=model_name,
499
500
501
502
503
        messages=messages,
        max_tokens=10,
        temperature=0.0,
        stream=True,
    )
504
    chunks: List[str] = []
505
    finish_reason_count = 0
506
507
508
509
510
511
    async for chunk in stream:
        delta = chunk.choices[0].delta
        if delta.role:
            assert delta.role == "assistant"
        if delta.content:
            chunks.append(delta.content)
512
513
514
515
        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
516
    assert chunk.choices[0].finish_reason == stop_reason
517
    assert delta.content
518
519
520
    assert "".join(chunks) == output


521
522
523
524
525
@pytest.mark.asyncio
@pytest.mark.parametrize(
    "model_name",
    ["HuggingFaceH4/zephyr-7b-beta", "zephyr-lora"],
)
526
async def test_chat_completion_stream_options(client: openai.AsyncOpenAI,
527
528
529
530
531
532
533
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
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
                                              model_name: str):
    messages = [{
        "role": "system",
        "content": "You are a helpful assistant."
    }, {
        "role": "user",
        "content": "What is the capital of France?"
    }]

    # Test stream=True, stream_options={"include_usage": False}
    stream = await client.chat.completions.create(
        model=model_name,
        messages=messages,
        max_tokens=10,
        temperature=0.0,
        stream=True,
        stream_options={"include_usage": False})
    async for chunk in stream:
        assert chunk.usage is None

    # Test stream=True, stream_options={"include_usage": True}
    stream = await client.chat.completions.create(
        model=model_name,
        messages=messages,
        max_tokens=10,
        temperature=0.0,
        stream=True,
        stream_options={"include_usage": True})

    async for chunk in stream:
        if chunk.choices[0].finish_reason is None:
            assert chunk.usage is None
        else:
            assert chunk.usage is None
            final_chunk = await stream.__anext__()
            assert final_chunk.usage is not None
            assert final_chunk.usage.prompt_tokens > 0
            assert final_chunk.usage.completion_tokens > 0
            assert final_chunk.usage.total_tokens == (
                final_chunk.usage.prompt_tokens +
                final_chunk.usage.completion_tokens)
            assert final_chunk.choices == []

    # Test stream=False, stream_options={"include_usage": None}
    with pytest.raises(BadRequestError):
        await client.chat.completions.create(
            model=model_name,
            messages=messages,
            max_tokens=10,
            temperature=0.0,
            stream=False,
            stream_options={"include_usage": None})

    # Test stream=False, stream_options={"include_usage": True}
    with pytest.raises(BadRequestError):
        await client.chat.completions.create(
            model=model_name,
            messages=messages,
            max_tokens=10,
            temperature=0.0,
            stream=False,
            stream_options={"include_usage": True})


@pytest.mark.asyncio
@pytest.mark.parametrize(
    "model_name",
    ["HuggingFaceH4/zephyr-7b-beta", "zephyr-lora"],
)
596
async def test_completion_stream_options(client: openai.AsyncOpenAI,
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
                                         model_name: str):
    prompt = "What is the capital of France?"

    # Test stream=True, stream_options={"include_usage": False}
    stream = await client.completions.create(
        model=model_name,
        prompt=prompt,
        max_tokens=5,
        temperature=0.0,
        stream=True,
        stream_options={"include_usage": False})
    async for chunk in stream:
        assert chunk.usage is None

    # Test stream=True, stream_options={"include_usage": True}
    stream = await client.completions.create(
        model=model_name,
        prompt=prompt,
        max_tokens=5,
        temperature=0.0,
        stream=True,
        stream_options={"include_usage": True})
    async for chunk in stream:
        if chunk.choices[0].finish_reason is None:
            assert chunk.usage is None
        else:
            assert chunk.usage is None
            final_chunk = await stream.__anext__()
            assert final_chunk.usage is not None
            assert final_chunk.usage.prompt_tokens > 0
            assert final_chunk.usage.completion_tokens > 0
            assert final_chunk.usage.total_tokens == (
                final_chunk.usage.prompt_tokens +
                final_chunk.usage.completion_tokens)
            assert final_chunk.choices == []

    # Test stream=False, stream_options={"include_usage": None}
    with pytest.raises(BadRequestError):
        await client.completions.create(model=model_name,
                                        prompt=prompt,
                                        max_tokens=5,
                                        temperature=0.0,
                                        stream=False,
                                        stream_options={"include_usage": None})

    # Test stream=False, stream_options={"include_usage": True}
    with pytest.raises(BadRequestError):
        await client.completions.create(model=model_name,
                                        prompt=prompt,
                                        max_tokens=5,
                                        temperature=0.0,
                                        stream=False,
                                        stream_options={"include_usage": True})


652
@pytest.mark.asyncio
653
654
655
656
657
@pytest.mark.parametrize(
    # just test 1 lora hereafter
    "model_name",
    [MODEL_NAME, "zephyr-lora"],
)
658
async def test_batch_completions(client: openai.AsyncOpenAI, model_name: str):
659
660
661
662
663
664
665
666
667
668
669
    # test both text and token IDs
    for prompts in (["Hello, my name is"] * 2, [[0, 0, 0, 0, 0]] * 2):
        # test simple list
        batch = await client.completions.create(
            model=model_name,
            prompt=prompts,
            max_tokens=5,
            temperature=0.0,
        )
        assert len(batch.choices) == 2
        assert batch.choices[0].text == batch.choices[1].text
670

671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
        # test n = 2
        batch = await client.completions.create(
            model=model_name,
            prompt=prompts,
            n=2,
            max_tokens=5,
            temperature=0.0,
            extra_body=dict(
                # NOTE: this has to be true for n > 1 in vLLM, but not necessary
                # for official client.
                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(
            model=model_name,
            prompt=prompts,
            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]
705
706


707
@pytest.mark.asyncio
708
async def test_logits_bias(client: openai.AsyncOpenAI):
709
710
711
712
713
714
715
716
717
718
719
720
    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},
721
        seed=42,
722
    )
723
    assert len(completion.choices[0].text) >= 5
724
725
726
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
    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


754
@pytest.mark.asyncio
755
756
@pytest.mark.parametrize("guided_decoding_backend",
                         ["outlines", "lm-format-enforcer"])
757
async def test_guided_json_completion(client: openai.AsyncOpenAI,
758
                                      guided_decoding_backend: str):
759
760
    completion = await client.completions.create(
        model=MODEL_NAME,
761
762
        prompt=f"Give an example JSON for an employee profile "
        f"that fits this schema: {TEST_SCHEMA}",
763
764
765
        n=3,
        temperature=1.0,
        max_tokens=500,
766
767
        extra_body=dict(guided_json=TEST_SCHEMA,
                        guided_decoding_backend=guided_decoding_backend))
768
769

    assert completion.id is not None
770
    assert len(completion.choices) == 3
771
772
773
774
775
    for i in range(3):
        output_json = json.loads(completion.choices[i].text)
        jsonschema.validate(instance=output_json, schema=TEST_SCHEMA)


776
@pytest.mark.asyncio
777
778
@pytest.mark.parametrize("guided_decoding_backend",
                         ["outlines", "lm-format-enforcer"])
779
async def test_guided_json_chat(client: openai.AsyncOpenAI,
780
                                guided_decoding_backend: str):
781
782
783
784
    messages = [{
        "role": "system",
        "content": "you are a helpful assistant"
    }, {
785
786
787
788
789
        "role":
        "user",
        "content":
        f"Give an example JSON for an employee profile that "
        f"fits this schema: {TEST_SCHEMA}"
790
791
792
793
    }]
    chat_completion = await client.chat.completions.create(
        model=MODEL_NAME,
        messages=messages,
794
795
796
        max_tokens=1000,
        extra_body=dict(guided_json=TEST_SCHEMA,
                        guided_decoding_backend=guided_decoding_backend))
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
    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,
812
813
814
        max_tokens=1000,
        extra_body=dict(guided_json=TEST_SCHEMA,
                        guided_decoding_backend=guided_decoding_backend))
815
816
817
818
819
820
821
822
    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"]


823
@pytest.mark.asyncio
824
825
@pytest.mark.parametrize("guided_decoding_backend",
                         ["outlines", "lm-format-enforcer"])
826
async def test_guided_regex_completion(client: openai.AsyncOpenAI,
827
                                       guided_decoding_backend: str):
828
829
830
831
832
833
    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,
834
835
        extra_body=dict(guided_regex=TEST_REGEX,
                        guided_decoding_backend=guided_decoding_backend))
836
837

    assert completion.id is not None
838
    assert len(completion.choices) == 3
839
840
841
842
    for i in range(3):
        assert re.fullmatch(TEST_REGEX, completion.choices[i].text) is not None


843
@pytest.mark.asyncio
844
845
@pytest.mark.parametrize("guided_decoding_backend",
                         ["outlines", "lm-format-enforcer"])
846
async def test_guided_regex_chat(client: openai.AsyncOpenAI,
847
                                 guided_decoding_backend: str):
848
849
850
851
852
853
854
855
856
857
858
859
860
    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,
861
862
        extra_body=dict(guided_regex=TEST_REGEX,
                        guided_decoding_backend=guided_decoding_backend))
863
864
865
866
867
868
869
870
871
872
    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,
873
874
        extra_body=dict(guided_regex=TEST_REGEX,
                        guided_decoding_backend=guided_decoding_backend))
875
876
877
878
879
880
    ip2 = chat_completion.choices[0].message.content
    assert ip2 is not None
    assert re.fullmatch(TEST_REGEX, ip2) is not None
    assert ip1 != ip2


881
@pytest.mark.asyncio
882
883
@pytest.mark.parametrize("guided_decoding_backend",
                         ["outlines", "lm-format-enforcer"])
884
async def test_guided_choice_completion(client: openai.AsyncOpenAI,
885
                                        guided_decoding_backend: str):
886
887
888
889
890
891
    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,
892
893
        extra_body=dict(guided_choice=TEST_CHOICE,
                        guided_decoding_backend=guided_decoding_backend))
894
895

    assert completion.id is not None
896
    assert len(completion.choices) == 2
897
898
899
900
    for i in range(2):
        assert completion.choices[i].text in TEST_CHOICE


901
@pytest.mark.asyncio
902
903
@pytest.mark.parametrize("guided_decoding_backend",
                         ["outlines", "lm-format-enforcer"])
904
async def test_guided_choice_chat(client: openai.AsyncOpenAI,
905
                                  guided_decoding_backend: str):
906
907
908
909
910
911
912
913
914
915
916
917
918
    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,
919
920
        extra_body=dict(guided_choice=TEST_CHOICE,
                        guided_decoding_backend=guided_decoding_backend))
921
922
923
924
925
926
927
928
929
930
931
932
    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,
933
934
        extra_body=dict(guided_choice=TEST_CHOICE,
                        guided_decoding_backend=guided_decoding_backend))
935
936
937
938
939
    choice2 = chat_completion.choices[0].message.content
    assert choice2 in TEST_CHOICE
    assert choice1 != choice2


940
@pytest.mark.asyncio
941
942
@pytest.mark.parametrize("guided_decoding_backend",
                         ["outlines", "lm-format-enforcer"])
943
async def test_guided_decoding_type_error(client: openai.AsyncOpenAI,
944
                                          guided_decoding_backend: str):
945
946
947
948
    with pytest.raises(openai.BadRequestError):
        _ = await client.completions.create(
            model=MODEL_NAME,
            prompt="Give an example JSON that fits this schema: 42",
949
950
            extra_body=dict(guided_json=42,
                            guided_decoding_backend=guided_decoding_backend))
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975

    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))


976
@pytest.mark.asyncio
977
978
@pytest.mark.parametrize("guided_decoding_backend",
                         ["outlines", "lm-format-enforcer"])
979
async def test_guided_choice_chat_logprobs(client: openai.AsyncOpenAI,
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
                                           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))
998
999
1000

    assert chat_completion.choices[0].logprobs is not None
    assert chat_completion.choices[0].logprobs.content is not None
1001
    top_logprobs = chat_completion.choices[0].logprobs.content[0].top_logprobs
1002
1003

    # -9999.0 is the minimum logprob returned by OpenAI
1004
1005
    for item in top_logprobs:
        assert item.logprob >= -9999.0, f"Failed (top_logprobs={top_logprobs})"
1006
1007


1008
1009
1010
@pytest.mark.asyncio
@pytest.mark.parametrize("guided_decoding_backend",
                         ["outlines", "lm-format-enforcer"])
1011
async def test_named_tool_use(client: openai.AsyncOpenAI,
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
                              guided_decoding_backend: str):
    messages = [{
        "role": "system",
        "content": "you are a helpful assistant"
    }, {
        "role":
        "user",
        "content":
        f"Give an example JSON for an employee profile that "
        f"fits this schema: {TEST_SCHEMA}"
    }]

    # non-streaming

    chat_completion = await client.chat.completions.create(
        model=MODEL_NAME,
        messages=messages,
        max_tokens=1000,
        tools=[{
            "type": "function",
            "function": {
                "name": "dummy_function_name",
                "description": "This is a dummy function",
                "parameters": TEST_SCHEMA
            }
        }],
        tool_choice={
            "type": "function",
            "function": {
                "name": "dummy_function_name"
            }
        })
    message = chat_completion.choices[0].message
    assert len(message.content) == 0
    json_string = message.tool_calls[0].function.arguments
    json1 = json.loads(json_string)
    jsonschema.validate(instance=json1, schema=TEST_SCHEMA)

    messages.append({"role": "assistant", "content": json_string})
    messages.append({
        "role":
        "user",
        "content":
        "Give me another one with a different name and age"
    })

    # streaming

    stream = await client.chat.completions.create(
        model=MODEL_NAME,
        messages=messages,
        max_tokens=1000,
        tools=[{
            "type": "function",
            "function": {
                "name": "dummy_function_name",
                "description": "This is a dummy function",
                "parameters": TEST_SCHEMA
            }
        }],
        tool_choice={
            "type": "function",
            "function": {
                "name": "dummy_function_name"
            }
        },
        stream=True)

    output = []
    finish_reason_count = 0
    async for chunk in stream:
        delta = chunk.choices[0].delta
        if delta.role:
            assert delta.role == "assistant"
        assert delta.content is None or len(delta.content) == 0
        if delta.tool_calls:
            output.append(delta.tool_calls[0].function.arguments)
        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
    json2 = json.loads("".join(output))
    jsonschema.validate(instance=json2, schema=TEST_SCHEMA)
    assert json1["name"] != json2["name"]
    assert json1["age"] != json2["age"]


@pytest.mark.asyncio
@pytest.mark.parametrize("guided_decoding_backend", ["outlines"])
async def test_required_tool_use_not_yet_supported(
1102
        client: openai.AsyncOpenAI, guided_decoding_backend: str):
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
    messages = [{
        "role": "system",
        "content": "you are a helpful assistant"
    }, {
        "role":
        "user",
        "content":
        f"Give an example JSON for an employee profile that "
        f"fits this schema: {TEST_SCHEMA}"
    }]

    with pytest.raises(openai.BadRequestError):
        await client.chat.completions.create(
            model=MODEL_NAME,
            messages=messages,
            max_tokens=1000,
            tools=[{
                "type": "function",
                "function": {
                    "name": "dummy_function_name",
                    "description": "This is a dummy function",
                    "parameters": TEST_SCHEMA
                }
            }],
            tool_choice="required")

    with pytest.raises(openai.BadRequestError):
        await client.chat.completions.create(
            model=MODEL_NAME,
            messages=messages,
            max_tokens=1000,
            tools=[{
                "type": "function",
                "function": {
                    "name": "dummy_function_name",
                    "description": "This is a dummy function",
                    "parameters": TEST_SCHEMA
                }
            }],
            tool_choice="auto")


@pytest.mark.asyncio
@pytest.mark.parametrize("guided_decoding_backend", ["outlines"])
async def test_inconsistent_tool_choice_and_tools(
1148
        client: openai.AsyncOpenAI, guided_decoding_backend: str):
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
    messages = [{
        "role": "system",
        "content": "you are a helpful assistant"
    }, {
        "role":
        "user",
        "content":
        f"Give an example JSON for an employee profile that "
        f"fits this schema: {TEST_SCHEMA}"
    }]

    with pytest.raises(openai.BadRequestError):
        await client.chat.completions.create(model=MODEL_NAME,
                                             messages=messages,
                                             max_tokens=1000,
                                             tool_choice={
                                                 "type": "function",
                                                 "function": {
                                                     "name":
                                                     "dummy_function_name"
                                                 }
                                             })

    with pytest.raises(openai.BadRequestError):
        await client.chat.completions.create(
            model=MODEL_NAME,
            messages=messages,
            max_tokens=1000,
            tools=[{
                "type": "function",
                "function": {
                    "name": "dummy_function_name",
                    "description": "This is a dummy function",
                    "parameters": TEST_SCHEMA
                }
            }],
            tool_choice={
                "type": "function",
                "function": {
                    "name": "nondefined_function_name"
                }
            })


1193
@pytest.mark.asyncio
1194
async def test_response_format_json_object(client: openai.AsyncOpenAI):
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
    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
1207
1208
        assert content is not None

1209
1210
        loaded = json.loads(content)
        assert loaded == {"result": 2}, loaded
1211
1212


1213
@pytest.mark.asyncio
1214
async def test_extra_fields(client: openai.AsyncOpenAI):
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
    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


1229
@pytest.mark.asyncio
1230
async def test_complex_message_content(client: openai.AsyncOpenAI):
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
    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"


1249
@pytest.mark.asyncio
1250
async def test_custom_role(client: openai.AsyncOpenAI):
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
    # Not sure how the model handles custom roles so we just check that
    # both string and complex message content are handled in the same way

    resp1 = await client.chat.completions.create(
        model=MODEL_NAME,
        messages=[{
            "role": "my-custom-role",
            "content": "what is 1+1?",
        }],  # type: ignore
        temperature=0,
        seed=0)

    resp2 = await client.chat.completions.create(
        model=MODEL_NAME,
        messages=[{
            "role": "my-custom-role",
            "content": [{
                "type": "text",
                "text": "what is 1+1?"
            }]
        }],  # type: ignore
        temperature=0,
        seed=0)

    content1 = resp1.choices[0].message.content
    content2 = resp2.choices[0].message.content
    assert content1 == content2


1280
@pytest.mark.asyncio
1281
async def test_guided_grammar(client: openai.AsyncOpenAI):
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
    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


1315
@pytest.mark.asyncio
1316
1317
1318
1319
1320
@pytest.mark.parametrize(
    # first test base model, then test loras
    "model_name",
    [MODEL_NAME, "zephyr-lora", "zephyr-lora2"],
)
1321
@pytest.mark.parametrize("logprobs_arg", [1, 0])
1322
async def test_echo_logprob_completion(client: openai.AsyncOpenAI,
1323
                                       model_name: str, logprobs_arg: int):
1324
1325
1326
1327
1328
1329
1330
1331
    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,
1332
                                                     logprobs=logprobs_arg)
1333
1334
1335

        prompt_text = tokenizer.decode(prompt) if isinstance(prompt,
                                                             list) else prompt
1336
        assert re.search(r"^" + prompt_text, completion.choices[0].text)
1337
1338
1339
1340
1341
1342
1343
        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)
1344
1345
1346
        for top_logprobs in logprobs.top_logprobs[1:]:
            assert max(logprobs_arg,
                       1) <= len(top_logprobs) <= logprobs_arg + 1
1347
1348
1349
        assert len(logprobs.tokens) > 5


1350
@pytest.mark.asyncio
1351
async def test_long_seed(client: openai.AsyncOpenAI):
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
    for seed in [
            torch.iinfo(torch.long).min - 1,
            torch.iinfo(torch.long).max + 1
    ]:
        with pytest.raises(BadRequestError) as exc_info:
            await client.chat.completions.create(
                model=MODEL_NAME,
                messages=[{
                    "role": "system",
                    "content": "You are a helpful assistant.",
                }],
                temperature=0,
                seed=seed)

        assert ("greater_than_equal" in exc_info.value.message
                or "less_than_equal" in exc_info.value.message)


1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
@pytest.mark.asyncio
@pytest.mark.parametrize(
    "model_name",
    [MODEL_NAME],
)
async def test_tokenize(server, client: openai.AsyncOpenAI, model_name: str):
    base_url = str(client.base_url)[:-3]
    tokenizer = get_tokenizer(tokenizer_name=MODEL_NAME, tokenizer_mode="fast")

    for add_special in [False, True]:
        prompt = "This is a test prompt."
        tokens = tokenizer.encode(prompt, add_special_tokens=add_special)

        response = requests.post(base_url + "/tokenize",
                                 json={
                                     "add_special_tokens": add_special,
                                     "model": model_name,
                                     "prompt": prompt
                                 })
        response.raise_for_status()
        assert response.json() == {
            "tokens": tokens,
            "count": len(tokens),
            "max_model_len": 8192
        }


@pytest.mark.asyncio
@pytest.mark.parametrize(
    "model_name",
    [MODEL_NAME],
)
async def test_detokenize(server, client: openai.AsyncOpenAI, model_name: str):
    base_url = str(client.base_url)[:-3]
    tokenizer = get_tokenizer(tokenizer_name=MODEL_NAME, tokenizer_mode="fast")

    prompt = "This is a test prompt."
    tokens = tokenizer.encode(prompt, add_special_tokens=False)

    response = requests.post(base_url + "detokenize",
                             json={
                                 "model": model_name,
                                 "tokens": tokens
                             })
    response.raise_for_status()
    assert response.json() == {"prompt": prompt}


1418
1419
if __name__ == "__main__":
    pytest.main([__file__])