test_openai_server.py 45.6 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 torch
13
14
# downloading lora to test lora requests
from huggingface_hub import snapshot_download
15
from openai import BadRequestError
16

17
18
from vllm.transformers_utils.tokenizer import get_tokenizer

19
from ..utils import RemoteOpenAIServer
20

21
22
23
24
25
# 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"
26

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
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"]
}

66
67
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)")
68
69
70
71
72
73

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

74
pytestmark = pytest.mark.openai
75
76
77


@pytest.fixture(scope="session")
78
79
80
81
def zephyr_lora_files():
    return snapshot_download(repo_id=LORA_NAME)


82
@pytest.fixture(scope="module")
83
def ray_ctx():
84
    ray.init()
85
86
87
88
89
90
91
    yield
    ray.shutdown()


@pytest.fixture(scope="module")
def server(zephyr_lora_files, ray_ctx):
    return RemoteOpenAIServer([
92
93
        "--model",
        MODEL_NAME,
94
        # use half precision for speed and memory savings in CI environment
95
        "--dtype",
96
        "bfloat16",
97
        "--max-model-len",
98
99
        "8192",
        "--enforce-eager",
100
101
102
103
104
105
106
107
108
109
        # 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",
110
        "128",
111
112
113
    ])


114
@pytest.fixture(scope="module")
115
116
def client(server):
    return server.get_async_client()
117
118


119
async def test_check_models(client: openai.AsyncOpenAI):
120
121
122
123
124
125
126
127
128
129
    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"


130
@pytest.mark.asyncio
131
132
133
134
135
@pytest.mark.parametrize(
    # first test base model, then test loras
    "model_name",
    [MODEL_NAME, "zephyr-lora", "zephyr-lora2"],
)
136
async def test_single_completion(client: openai.AsyncOpenAI, model_name: str):
137
    completion = await client.completions.create(model=model_name,
138
139
140
141
142
143
                                                 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
144
145
146
147

    choice = completion.choices[0]
    assert len(choice.text) >= 5
    assert choice.finish_reason == "length"
148
149
150
    assert completion.usage == openai.types.CompletionUsage(
        completion_tokens=5, prompt_tokens=6, total_tokens=11)

151
152
153
154
155
156
157
    # test using token IDs
    completion = await client.completions.create(
        model=MODEL_NAME,
        prompt=[0, 0, 0, 0, 0],
        max_tokens=5,
        temperature=0.0,
    )
158
    assert len(completion.choices[0].text) >= 5
159

160

161
162
163
164
165
166
@pytest.mark.asyncio
@pytest.mark.parametrize(
    # first test base model, then test loras
    "model_name",
    [MODEL_NAME, "zephyr-lora", "zephyr-lora2"],
)
167
async def test_no_logprobs(client: openai.AsyncOpenAI, model_name: str):
168
169
170
171
172
173
174
175
176
177
178
179
    # 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


180
@pytest.mark.asyncio
181
@pytest.mark.parametrize(
182
    # just test 1 lora hereafter
183
    "model_name",
184
    [MODEL_NAME, "zephyr-lora"],
185
)
186
async def test_zero_logprobs(client: openai.AsyncOpenAI, model_name: str):
187
188
189
190
191
192
193
194
195
196
197
    # 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
198
    assert choice.logprobs.top_logprobs is not None
199
    assert len(choice.logprobs.top_logprobs[0]) == 1
200
201
202
203
204
205
206


@pytest.mark.asyncio
@pytest.mark.parametrize(
    "model_name",
    [MODEL_NAME, "zephyr-lora"],
)
207
async def test_some_logprobs(client: openai.AsyncOpenAI, model_name: str):
208
209
210
211
212
213
214
215
216
217
218
219
    # 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
220
    assert 5 <= len(choice.logprobs.top_logprobs[0]) <= 6
221
222
223
224
225
226
227


@pytest.mark.asyncio
@pytest.mark.parametrize(
    "model_name",
    [MODEL_NAME, "zephyr-lora"],
)
228
async def test_too_many_completion_logprobs(client: openai.AsyncOpenAI,
229
230
231
232
233
234
235
236
237
                                            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,
238
239
240
            # vLLM has higher default max_logprobs (20 instead of 5) to support
            # both Completion API and Chat Completion API
            logprobs=21,
241
242
243
244
245
246
247
248
249
        )
        ...
    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,
250
251
252
            # vLLM has higher default max_logprobs (20 instead of 5) to support
            # both Completion API and Chat Completion API
            logprobs=30,
253
254
255
256
257
258
259
260
261
262
263
264
            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,
    )
265
    assert len(completion.choices[0].text) >= 0
266
267


268
269
270
271
272
273
@pytest.mark.asyncio
@pytest.mark.parametrize(
    # first test base model, then test loras
    "model_name",
    [MODEL_NAME, "zephyr-lora", "zephyr-lora2"],
)
274
async def test_no_logprobs_chat(client: openai.AsyncOpenAI, model_name: str):
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
    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"],
)
299
async def test_zero_logprobs_chat(client: openai.AsyncOpenAI, model_name: str):
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
    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
318
    assert len(choice.logprobs.content[0].top_logprobs) == 0
319
320
321
322
323
324
325


@pytest.mark.asyncio
@pytest.mark.parametrize(
    "model_name",
    [MODEL_NAME, "zephyr-lora"],
)
326
async def test_some_logprobs_chat(client: openai.AsyncOpenAI, model_name: str):
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
    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
345
    assert len(choice.logprobs.content[0].top_logprobs) == 5
346
347


348
@pytest.mark.asyncio
349
350
351
352
@pytest.mark.parametrize(
    "model_name",
    [MODEL_NAME, "zephyr-lora"],
)
353
async def test_too_many_chat_logprobs(client: openai.AsyncOpenAI,
354
                                      model_name: str):
355
356
357
358
359
360
361
362
    messages = [{
        "role": "system",
        "content": "you are a helpful assistant"
    }, {
        "role": "user",
        "content": "what is 1+1?"
    }]

363
    # Default max_logprobs is 20, so this should raise an error
364
365
366
367
368
    with pytest.raises((openai.BadRequestError, openai.APIError)):
        stream = await client.chat.completions.create(model=model_name,
                                                      messages=messages,
                                                      max_tokens=10,
                                                      logprobs=True,
369
                                                      top_logprobs=21,
370
371
372
373
374
375
376
377
378
                                                      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,
379
                                             top_logprobs=30,
380
381
382
383
384
385
386
387
388
                                             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
389
390


391
@pytest.mark.asyncio
392
@pytest.mark.parametrize(
393
394
395
    "model_name",
    [MODEL_NAME, "zephyr-lora"],
)
396
async def test_single_chat_session(client: openai.AsyncOpenAI,
397
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
                                   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(
438
439
440
    "model_name",
    [MODEL_NAME, "zephyr-lora"],
)
441
async def test_completion_streaming(client: openai.AsyncOpenAI,
442
                                    model_name: str):
443
444
445
    prompt = "What is an LLM?"

    single_completion = await client.completions.create(
446
        model=model_name,
447
448
449
450
451
        prompt=prompt,
        max_tokens=5,
        temperature=0.0,
    )
    single_output = single_completion.choices[0].text
452
453
454
455
456
    stream = await client.completions.create(model=model_name,
                                             prompt=prompt,
                                             max_tokens=5,
                                             temperature=0.0,
                                             stream=True)
457
    chunks: List[str] = []
458
    finish_reason_count = 0
459
460
    async for chunk in stream:
        chunks.append(chunk.choices[0].text)
461
462
463
464
        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
465
    assert chunk.choices[0].finish_reason == "length"
466
    assert chunk.choices[0].text
467
468
469
    assert "".join(chunks) == single_output


470
@pytest.mark.asyncio
471
472
473
474
475
@pytest.mark.parametrize(
    # just test 1 lora hereafter
    "model_name",
    [MODEL_NAME, "zephyr-lora"],
)
476
async def test_chat_streaming(client: openai.AsyncOpenAI, model_name: str):
477
478
479
480
481
482
483
484
485
486
    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(
487
        model=model_name,
488
489
490
491
492
493
494
495
496
        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(
497
        model=model_name,
498
499
500
501
502
        messages=messages,
        max_tokens=10,
        temperature=0.0,
        stream=True,
    )
503
    chunks: List[str] = []
504
    finish_reason_count = 0
505
506
507
508
509
510
    async for chunk in stream:
        delta = chunk.choices[0].delta
        if delta.role:
            assert delta.role == "assistant"
        if delta.content:
            chunks.append(delta.content)
511
512
513
514
        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
515
    assert chunk.choices[0].finish_reason == stop_reason
516
    assert delta.content
517
518
519
    assert "".join(chunks) == output


520
521
522
523
524
@pytest.mark.asyncio
@pytest.mark.parametrize(
    "model_name",
    ["HuggingFaceH4/zephyr-7b-beta", "zephyr-lora"],
)
525
async def test_chat_completion_stream_options(client: openai.AsyncOpenAI,
526
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
                                              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"],
)
595
async def test_completion_stream_options(client: openai.AsyncOpenAI,
596
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
                                         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})


651
@pytest.mark.asyncio
652
653
654
655
656
@pytest.mark.parametrize(
    # just test 1 lora hereafter
    "model_name",
    [MODEL_NAME, "zephyr-lora"],
)
657
async def test_batch_completions(client: openai.AsyncOpenAI, model_name: str):
658
659
660
661
662
663
664
665
666
667
668
    # 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
669

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
        # 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]
704
705


706
@pytest.mark.asyncio
707
async def test_logits_bias(client: openai.AsyncOpenAI):
708
709
710
711
712
713
714
715
716
717
718
719
    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},
720
        seed=42,
721
    )
722
    assert len(completion.choices[0].text) >= 5
723
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
    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


753
@pytest.mark.asyncio
754
755
@pytest.mark.parametrize("guided_decoding_backend",
                         ["outlines", "lm-format-enforcer"])
756
async def test_guided_json_completion(client: openai.AsyncOpenAI,
757
                                      guided_decoding_backend: str):
758
759
    completion = await client.completions.create(
        model=MODEL_NAME,
760
761
        prompt=f"Give an example JSON for an employee profile "
        f"that fits this schema: {TEST_SCHEMA}",
762
763
764
        n=3,
        temperature=1.0,
        max_tokens=500,
765
766
        extra_body=dict(guided_json=TEST_SCHEMA,
                        guided_decoding_backend=guided_decoding_backend))
767
768

    assert completion.id is not None
769
    assert len(completion.choices) == 3
770
771
772
773
774
    for i in range(3):
        output_json = json.loads(completion.choices[i].text)
        jsonschema.validate(instance=output_json, schema=TEST_SCHEMA)


775
@pytest.mark.asyncio
776
777
@pytest.mark.parametrize("guided_decoding_backend",
                         ["outlines", "lm-format-enforcer"])
778
async def test_guided_json_chat(client: openai.AsyncOpenAI,
779
                                guided_decoding_backend: str):
780
781
782
783
    messages = [{
        "role": "system",
        "content": "you are a helpful assistant"
    }, {
784
785
786
787
788
        "role":
        "user",
        "content":
        f"Give an example JSON for an employee profile that "
        f"fits this schema: {TEST_SCHEMA}"
789
790
791
792
    }]
    chat_completion = await client.chat.completions.create(
        model=MODEL_NAME,
        messages=messages,
793
794
795
        max_tokens=1000,
        extra_body=dict(guided_json=TEST_SCHEMA,
                        guided_decoding_backend=guided_decoding_backend))
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
    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,
811
812
813
        max_tokens=1000,
        extra_body=dict(guided_json=TEST_SCHEMA,
                        guided_decoding_backend=guided_decoding_backend))
814
815
816
817
818
819
820
821
    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"]


822
@pytest.mark.asyncio
823
824
@pytest.mark.parametrize("guided_decoding_backend",
                         ["outlines", "lm-format-enforcer"])
825
async def test_guided_regex_completion(client: openai.AsyncOpenAI,
826
                                       guided_decoding_backend: str):
827
828
829
830
831
832
    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,
833
834
        extra_body=dict(guided_regex=TEST_REGEX,
                        guided_decoding_backend=guided_decoding_backend))
835
836

    assert completion.id is not None
837
    assert len(completion.choices) == 3
838
839
840
841
    for i in range(3):
        assert re.fullmatch(TEST_REGEX, completion.choices[i].text) is not None


842
@pytest.mark.asyncio
843
844
@pytest.mark.parametrize("guided_decoding_backend",
                         ["outlines", "lm-format-enforcer"])
845
async def test_guided_regex_chat(client: openai.AsyncOpenAI,
846
                                 guided_decoding_backend: str):
847
848
849
850
851
852
853
854
855
856
857
858
859
    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,
860
861
        extra_body=dict(guided_regex=TEST_REGEX,
                        guided_decoding_backend=guided_decoding_backend))
862
863
864
865
866
867
868
869
870
871
    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,
872
873
        extra_body=dict(guided_regex=TEST_REGEX,
                        guided_decoding_backend=guided_decoding_backend))
874
875
876
877
878
879
    ip2 = chat_completion.choices[0].message.content
    assert ip2 is not None
    assert re.fullmatch(TEST_REGEX, ip2) is not None
    assert ip1 != ip2


880
@pytest.mark.asyncio
881
882
@pytest.mark.parametrize("guided_decoding_backend",
                         ["outlines", "lm-format-enforcer"])
883
async def test_guided_choice_completion(client: openai.AsyncOpenAI,
884
                                        guided_decoding_backend: str):
885
886
887
888
889
890
    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,
891
892
        extra_body=dict(guided_choice=TEST_CHOICE,
                        guided_decoding_backend=guided_decoding_backend))
893
894

    assert completion.id is not None
895
    assert len(completion.choices) == 2
896
897
898
899
    for i in range(2):
        assert completion.choices[i].text in TEST_CHOICE


900
@pytest.mark.asyncio
901
902
@pytest.mark.parametrize("guided_decoding_backend",
                         ["outlines", "lm-format-enforcer"])
903
async def test_guided_choice_chat(client: openai.AsyncOpenAI,
904
                                  guided_decoding_backend: str):
905
906
907
908
909
910
911
912
913
914
915
916
917
    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,
918
919
        extra_body=dict(guided_choice=TEST_CHOICE,
                        guided_decoding_backend=guided_decoding_backend))
920
921
922
923
924
925
926
927
928
929
930
931
    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,
932
933
        extra_body=dict(guided_choice=TEST_CHOICE,
                        guided_decoding_backend=guided_decoding_backend))
934
935
936
937
938
    choice2 = chat_completion.choices[0].message.content
    assert choice2 in TEST_CHOICE
    assert choice1 != choice2


939
@pytest.mark.asyncio
940
941
@pytest.mark.parametrize("guided_decoding_backend",
                         ["outlines", "lm-format-enforcer"])
942
async def test_guided_decoding_type_error(client: openai.AsyncOpenAI,
943
                                          guided_decoding_backend: str):
944
945
946
947
    with pytest.raises(openai.BadRequestError):
        _ = await client.completions.create(
            model=MODEL_NAME,
            prompt="Give an example JSON that fits this schema: 42",
948
949
            extra_body=dict(guided_json=42,
                            guided_decoding_backend=guided_decoding_backend))
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974

    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))


975
@pytest.mark.asyncio
976
977
@pytest.mark.parametrize("guided_decoding_backend",
                         ["outlines", "lm-format-enforcer"])
978
async def test_guided_choice_chat_logprobs(client: openai.AsyncOpenAI,
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
                                           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))
997
998
999

    assert chat_completion.choices[0].logprobs is not None
    assert chat_completion.choices[0].logprobs.content is not None
1000
    top_logprobs = chat_completion.choices[0].logprobs.content[0].top_logprobs
1001
1002

    # -9999.0 is the minimum logprob returned by OpenAI
1003
1004
    for item in top_logprobs:
        assert item.logprob >= -9999.0, f"Failed (top_logprobs={top_logprobs})"
1005
1006


1007
1008
1009
@pytest.mark.asyncio
@pytest.mark.parametrize("guided_decoding_backend",
                         ["outlines", "lm-format-enforcer"])
1010
async def test_named_tool_use(client: openai.AsyncOpenAI,
1011
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
                              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(
1101
        client: openai.AsyncOpenAI, guided_decoding_backend: str):
1102
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
    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(
1147
        client: openai.AsyncOpenAI, guided_decoding_backend: str):
1148
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
    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"
                }
            })


1192
@pytest.mark.asyncio
1193
async def test_response_format_json_object(client: openai.AsyncOpenAI):
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
    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
1206
1207
        assert content is not None

1208
1209
        loaded = json.loads(content)
        assert loaded == {"result": 2}, loaded
1210
1211


1212
@pytest.mark.asyncio
1213
async def test_extra_fields(client: openai.AsyncOpenAI):
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
    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


1228
@pytest.mark.asyncio
1229
async def test_complex_message_content(client: openai.AsyncOpenAI):
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
    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"


1248
@pytest.mark.asyncio
1249
async def test_custom_role(client: openai.AsyncOpenAI):
1250
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
    # 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


1279
@pytest.mark.asyncio
1280
async def test_guided_grammar(client: openai.AsyncOpenAI):
1281
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
    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


1314
@pytest.mark.asyncio
1315
1316
1317
1318
1319
@pytest.mark.parametrize(
    # first test base model, then test loras
    "model_name",
    [MODEL_NAME, "zephyr-lora", "zephyr-lora2"],
)
1320
@pytest.mark.parametrize("logprobs_arg", [1, 0])
1321
async def test_echo_logprob_completion(client: openai.AsyncOpenAI,
1322
                                       model_name: str, logprobs_arg: int):
1323
1324
1325
1326
1327
1328
1329
1330
    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,
1331
                                                     logprobs=logprobs_arg)
1332
1333
1334

        prompt_text = tokenizer.decode(prompt) if isinstance(prompt,
                                                             list) else prompt
1335
        assert re.search(r"^" + prompt_text, completion.choices[0].text)
1336
1337
1338
1339
1340
1341
1342
        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)
1343
1344
1345
        for top_logprobs in logprobs.top_logprobs[1:]:
            assert max(logprobs_arg,
                       1) <= len(top_logprobs) <= logprobs_arg + 1
1346
1347
1348
        assert len(logprobs.tokens) > 5


1349
@pytest.mark.asyncio
1350
async def test_long_seed(client: openai.AsyncOpenAI):
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
    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)


1369
1370
if __name__ == "__main__":
    pytest.main([__file__])