test_vision.py 18.6 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

4
5
import json

6
7
import openai
import pytest
8
import pytest_asyncio
pansicheng's avatar
pansicheng committed
9
from transformers import AutoProcessor
10

11
from vllm.multimodal.base import MediaWithBytes
12
from vllm.multimodal.utils import encode_image_url, fetch_image
13

14
from ...utils import RemoteOpenAIServer
15

16
17
MODEL_NAME = "microsoft/Phi-3.5-vision-instruct"
MAXIMUM_IMAGES = 2
18

19
# Test different image extensions (JPG/PNG) and formats (gray/RGB/RGBA)
20
TEST_IMAGE_ASSETS = [
21
22
23
24
    "2560px-Gfp-wisconsin-madison-the-nature-boardwalk.jpg",  # "https://vllm-public-assets.s3.us-west-2.amazonaws.com/vision_model_images/2560px-Gfp-wisconsin-madison-the-nature-boardwalk.jpg"
    "Grayscale_8bits_palette_sample_image.png",  # "https://vllm-public-assets.s3.us-west-2.amazonaws.com/vision_model_images/Grayscale_8bits_palette_sample_image.png",
    "1280px-Venn_diagram_rgb.svg.png",  # "https://vllm-public-assets.s3.us-west-2.amazonaws.com/vision_model_images/1280px-Venn_diagram_rgb.svg.png",
    "RGBA_comp.png",  # "https://vllm-public-assets.s3.us-west-2.amazonaws.com/vision_model_images/RGBA_comp.png",
25
26
]

27
28
29
30
31
32
33
34
35
36
37
EXPECTED_MM_BEAM_SEARCH_RES = [
    [
        "The image shows a wooden boardwalk leading through a",
        "The image shows a wooden boardwalk extending into a",
    ],
    [
        "The image shows two parrots perched on",
        "The image shows two birds perched on a cur",
    ],
    [
        "The image shows a Venn diagram with three over",
38
        "The image displays a Venn diagram with three over",
39
40
41
    ],
    [
        "This image displays a gradient of colors ranging from",
42
        "This image displays a gradient of colors forming a spectrum",
43
44
45
    ],
]

46

47
@pytest.fixture(scope="module")
48
def server():
49
    args = [
50
        "--runner",
51
        "generate",
52
53
54
55
56
57
58
        "--max-model-len",
        "2048",
        "--max-num-seqs",
        "5",
        "--enforce-eager",
        "--trust-remote-code",
        "--limit-mm-per-prompt",
59
        json.dumps({"image": MAXIMUM_IMAGES}),
60
61
62
    ]

    with RemoteOpenAIServer(MODEL_NAME, args) as remote_server:
63
        yield remote_server
64
65


66
67
68
69
@pytest_asyncio.fixture
async def client(server):
    async with server.get_async_client() as async_client:
        yield async_client
70
71


72
@pytest.fixture(scope="session")
73
def url_encoded_image(local_asset_server) -> dict[str, str]:
74
    return {
75
        image_asset: encode_image_url(local_asset_server.get_image_asset(image_asset))
76
        for image_asset in TEST_IMAGE_ASSETS
77
78
79
    }


80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
def dummy_messages_from_image_url(
    image_urls: str | list[str],
    content_text: str = "What's in this image?",
):
    if isinstance(image_urls, str):
        image_urls = [image_urls]

    return [
        {
            "role": "user",
            "content": [
                *(
                    {"type": "image_url", "image_url": {"url": image_url}}
                    for image_url in image_urls
                ),
                {"type": "text", "text": content_text},
            ],
        }
    ]


pansicheng's avatar
pansicheng committed
101
def get_hf_prompt_tokens(model_name, content, image_url):
102
103
104
    processor = AutoProcessor.from_pretrained(
        model_name, trust_remote_code=True, num_crops=4
    )
pansicheng's avatar
pansicheng committed
105
106

    placeholder = "<|image_1|>\n"
107
108
109
110
111
112
    messages = [
        {
            "role": "user",
            "content": f"{placeholder}{content}",
        }
    ]
113
114
115
116
117
    image = fetch_image(image_url)
    # Unwrap MediaWithBytes if present
    if isinstance(image, MediaWithBytes):
        image = image.media
    images = [image]
pansicheng's avatar
pansicheng committed
118
119

    prompt = processor.tokenizer.apply_chat_template(
120
121
        messages, tokenize=False, add_generation_prompt=True
    )
pansicheng's avatar
pansicheng committed
122
123
124
125
126
    inputs = processor(prompt, images, return_tensors="pt")

    return inputs.input_ids.shape[1]


127
128
@pytest.mark.asyncio
@pytest.mark.parametrize("model_name", [MODEL_NAME])
129
@pytest.mark.parametrize("image_url", TEST_IMAGE_ASSETS, indirect=True)
130
131
132
async def test_single_chat_session_image(
    client: openai.AsyncOpenAI, model_name: str, image_url: str
):
pansicheng's avatar
pansicheng committed
133
    content_text = "What's in this image?"
134
    messages = dummy_messages_from_image_url(image_url, content_text)
135

pansicheng's avatar
pansicheng committed
136
    max_completion_tokens = 10
137
    # test single completion
138
139
140
    chat_completion = await client.chat.completions.create(
        model=model_name,
        messages=messages,
pansicheng's avatar
pansicheng committed
141
        max_completion_tokens=max_completion_tokens,
142
        logprobs=True,
143
        temperature=0.0,
144
145
        top_logprobs=5,
    )
146
147
148
149
    assert len(chat_completion.choices) == 1

    choice = chat_completion.choices[0]
    assert choice.finish_reason == "length"
150
    hf_prompt_tokens = get_hf_prompt_tokens(model_name, content_text, image_url)
151
    assert chat_completion.usage == openai.types.CompletionUsage(
pansicheng's avatar
pansicheng committed
152
153
        completion_tokens=max_completion_tokens,
        prompt_tokens=hf_prompt_tokens,
154
155
        total_tokens=hf_prompt_tokens + max_completion_tokens,
    )
156
157
158
159
160
161
162
163
164
165
166
167

    message = choice.message
    message = chat_completion.choices[0].message
    assert message.content is not None and len(message.content) >= 10
    assert message.role == "assistant"
    messages.append({"role": "assistant", "content": message.content})

    # test multi-turn dialogue
    messages.append({"role": "user", "content": "express your result in json"})
    chat_completion = await client.chat.completions.create(
        model=model_name,
        messages=messages,
168
        max_completion_tokens=10,
169
170
171
172
173
    )
    message = chat_completion.choices[0].message
    assert message.content is not None and len(message.content) >= 0


174
175
@pytest.mark.asyncio
@pytest.mark.parametrize("model_name", [MODEL_NAME])
176
@pytest.mark.parametrize("image_url", TEST_IMAGE_ASSETS, indirect=True)
177
178
179
async def test_error_on_invalid_image_url_type(
    client: openai.AsyncOpenAI, model_name: str, image_url: str
):
180
    content_text = "What's in this image?"
181
182
183
184
185
186
187
188
189
    messages = [
        {
            "role": "user",
            "content": [
                {"type": "image_url", "image_url": image_url},
                {"type": "text", "text": content_text},
            ],
        }
    ]
190
191
192

    # image_url should be a dict {"url": "some url"}, not directly a string
    with pytest.raises(openai.BadRequestError):
193
194
195
196
197
198
        _ = await client.chat.completions.create(
            model=model_name,
            messages=messages,
            max_completion_tokens=10,
            temperature=0.0,
        )
199
200


201
202
@pytest.mark.asyncio
@pytest.mark.parametrize("model_name", [MODEL_NAME])
203
@pytest.mark.parametrize("image_url", TEST_IMAGE_ASSETS, indirect=True)
204
205
206
async def test_single_chat_session_image_beamsearch(
    client: openai.AsyncOpenAI, model_name: str, image_url: str
):
207
208
    content_text = "What's in this image?"
    messages = dummy_messages_from_image_url(image_url, content_text)
209
210
211
212
213

    chat_completion = await client.chat.completions.create(
        model=model_name,
        messages=messages,
        n=2,
214
        max_completion_tokens=10,
215
216
        logprobs=True,
        top_logprobs=5,
217
218
        extra_body=dict(use_beam_search=True),
    )
219
    assert len(chat_completion.choices) == 2
220
221
222
223
    assert (
        chat_completion.choices[0].message.content
        != chat_completion.choices[1].message.content
    )
224
225


226
227
@pytest.mark.asyncio
@pytest.mark.parametrize("model_name", [MODEL_NAME])
228
229
@pytest.mark.parametrize("raw_image_url", TEST_IMAGE_ASSETS)
@pytest.mark.parametrize("image_url", TEST_IMAGE_ASSETS, indirect=True)
230
async def test_single_chat_session_image_base64encoded(
231
232
233
234
    client: openai.AsyncOpenAI,
    model_name: str,
    raw_image_url: str,
    image_url: str,
235
    url_encoded_image: dict[str, str],
236
):
pansicheng's avatar
pansicheng committed
237
    content_text = "What's in this image?"
238
    messages = dummy_messages_from_image_url(
239
        url_encoded_image[raw_image_url],
240
241
        content_text,
    )
242

pansicheng's avatar
pansicheng committed
243
    max_completion_tokens = 10
244
    # test single completion
245
246
247
    chat_completion = await client.chat.completions.create(
        model=model_name,
        messages=messages,
pansicheng's avatar
pansicheng committed
248
        max_completion_tokens=max_completion_tokens,
249
        logprobs=True,
250
        temperature=0.0,
251
252
        top_logprobs=5,
    )
253
254
255
256
    assert len(chat_completion.choices) == 1

    choice = chat_completion.choices[0]
    assert choice.finish_reason == "length"
257
    hf_prompt_tokens = get_hf_prompt_tokens(model_name, content_text, image_url)
258
    assert chat_completion.usage == openai.types.CompletionUsage(
pansicheng's avatar
pansicheng committed
259
260
        completion_tokens=max_completion_tokens,
        prompt_tokens=hf_prompt_tokens,
261
262
        total_tokens=hf_prompt_tokens + max_completion_tokens,
    )
263
264
265
266
267
268
269
270
271
272
273
274

    message = choice.message
    message = chat_completion.choices[0].message
    assert message.content is not None and len(message.content) >= 10
    assert message.role == "assistant"
    messages.append({"role": "assistant", "content": message.content})

    # test multi-turn dialogue
    messages.append({"role": "user", "content": "express your result in json"})
    chat_completion = await client.chat.completions.create(
        model=model_name,
        messages=messages,
275
        max_completion_tokens=10,
276
        temperature=0.0,
277
278
279
280
281
    )
    message = chat_completion.choices[0].message
    assert message.content is not None and len(message.content) >= 0


282
283
@pytest.mark.asyncio
@pytest.mark.parametrize("model_name", [MODEL_NAME])
284
@pytest.mark.parametrize("image_idx", list(range(len(TEST_IMAGE_ASSETS))))
285
async def test_single_chat_session_image_base64encoded_beamsearch(
286
287
288
    client: openai.AsyncOpenAI,
    model_name: str,
    image_idx: int,
289
    url_encoded_image: dict[str, str],
290
):
291
    # NOTE: This test also validates that we pass MM data through beam search
292
    raw_image_url = TEST_IMAGE_ASSETS[image_idx]
293
    expected_res = EXPECTED_MM_BEAM_SEARCH_RES[image_idx]
294

295
    messages = dummy_messages_from_image_url(url_encoded_image[raw_image_url])
296

297
298
299
300
    chat_completion = await client.chat.completions.create(
        model=model_name,
        messages=messages,
        n=2,
301
        max_completion_tokens=10,
302
        temperature=0.0,
303
304
        extra_body=dict(use_beam_search=True),
    )
305
    assert len(chat_completion.choices) == 2
306
307
    for actual, expected_str in zip(chat_completion.choices, expected_res):
        assert actual.message.content == expected_str
308
309


310
311
@pytest.mark.asyncio
@pytest.mark.parametrize("model_name", [MODEL_NAME])
312
@pytest.mark.parametrize("image_url", TEST_IMAGE_ASSETS, indirect=True)
313
314
315
async def test_chat_streaming_image(
    client: openai.AsyncOpenAI, model_name: str, image_url: str
):
316
    messages = dummy_messages_from_image_url(image_url)
317
318
319
320
321

    # test single completion
    chat_completion = await client.chat.completions.create(
        model=model_name,
        messages=messages,
322
        max_completion_tokens=10,
323
324
325
326
327
328
329
330
331
        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(
        model=model_name,
        messages=messages,
332
        max_completion_tokens=10,
333
334
335
        temperature=0.0,
        stream=True,
    )
336
    chunks: list[str] = []
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
    finish_reason_count = 0
    async for chunk in stream:
        delta = chunk.choices[0].delta
        if delta.role:
            assert delta.role == "assistant"
        if delta.content:
            chunks.append(delta.content)
        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
    assert chunk.choices[0].finish_reason == stop_reason
    assert delta.content
    assert "".join(chunks) == output


@pytest.mark.asyncio
@pytest.mark.parametrize("model_name", [MODEL_NAME])
355
356
@pytest.mark.parametrize(
    "image_urls",
357
    [TEST_IMAGE_ASSETS[:i] for i in range(2, len(TEST_IMAGE_ASSETS))],
358
359
360
361
362
    indirect=True,
)
async def test_multi_image_input(
    client: openai.AsyncOpenAI, model_name: str, image_urls: list[str]
):
363
    messages = dummy_messages_from_image_url(image_urls)
364

365
366
367
368
369
    if len(image_urls) > MAXIMUM_IMAGES:
        with pytest.raises(openai.BadRequestError):  # test multi-image input
            await client.chat.completions.create(
                model=model_name,
                messages=messages,
370
                max_completion_tokens=10,
371
372
373
374
375
376
377
378
379
380
381
382
383
384
                temperature=0.0,
            )

        # 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,
        )
        completion = completion.choices[0].text
        assert completion is not None and len(completion) >= 0
    else:
        chat_completion = await client.chat.completions.create(
385
386
            model=model_name,
            messages=messages,
387
            max_completion_tokens=10,
388
389
            temperature=0.0,
        )
390
391
        message = chat_completion.choices[0].message
        assert message.content is not None and len(message.content) >= 0
392
393
394
395
396
397
398


@pytest.mark.asyncio
@pytest.mark.parametrize("model_name", [MODEL_NAME])
@pytest.mark.parametrize(
    "image_urls",
    [TEST_IMAGE_ASSETS[:i] for i in range(2, len(TEST_IMAGE_ASSETS))],
399
400
    indirect=True,
)
401
402
403
404
405
406
407
408
async def test_completions_with_image(
    client: openai.AsyncOpenAI,
    model_name: str,
    image_urls: list[str],
):
    for image_url in image_urls:
        chat_completion = await client.chat.completions.create(
            messages=[
409
                {"role": "system", "content": "You are a helpful assistant."},
410
                {
411
                    "role": "user",
412
413
414
415
416
417
418
419
420
                    "content": [
                        {
                            "type": "text",
                            "text": "Describe this image.",
                        },
                        {
                            "type": "image_url",
                            "image_url": {
                                "url": image_url,
421
                            },
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
                        },
                    ],
                },
            ],
            model=model_name,
        )
        assert chat_completion.choices[0].message.content is not None
        assert isinstance(chat_completion.choices[0].message.content, str)
        assert len(chat_completion.choices[0].message.content) > 0


@pytest.mark.asyncio
@pytest.mark.parametrize("model_name", [MODEL_NAME])
@pytest.mark.parametrize(
    "image_urls",
    [TEST_IMAGE_ASSETS[:i] for i in range(2, len(TEST_IMAGE_ASSETS))],
438
439
    indirect=True,
)
440
441
442
443
444
445
446
447
async def test_completions_with_image_with_uuid(
    client: openai.AsyncOpenAI,
    model_name: str,
    image_urls: list[str],
):
    for image_url in image_urls:
        chat_completion = await client.chat.completions.create(
            messages=[
448
                {"role": "system", "content": "You are a helpful assistant."},
449
                {
450
                    "role": "user",
451
452
453
454
455
456
457
458
459
460
                    "content": [
                        {
                            "type": "text",
                            "text": "Describe this image.",
                        },
                        {
                            "type": "image_url",
                            "image_url": {
                                "url": image_url,
                            },
461
                            "uuid": image_url,
462
463
464
465
466
467
468
469
470
471
                        },
                    ],
                },
            ],
            model=model_name,
        )
        assert chat_completion.choices[0].message.content is not None
        assert isinstance(chat_completion.choices[0].message.content, str)
        assert len(chat_completion.choices[0].message.content) > 0

472
473
474
        # Second request, with empty image but the same uuid.
        chat_completion_with_empty_image = await client.chat.completions.create(
            messages=[
475
                {"role": "system", "content": "You are a helpful assistant."},
476
                {
477
                    "role": "user",
478
479
480
481
482
                    "content": [
                        {
                            "type": "text",
                            "text": "Describe this image.",
                        },
483
                        {"type": "image_url", "image_url": {}, "uuid": image_url},
484
485
486
487
488
                    ],
                },
            ],
            model=model_name,
        )
489
        assert chat_completion_with_empty_image.choices[0].message.content is not None
490
        assert isinstance(
491
492
493
            chat_completion_with_empty_image.choices[0].message.content, str
        )
        assert len(chat_completion_with_empty_image.choices[0].message.content) > 0
494
495
496
497
498
499
500
501
502
503
504


@pytest.mark.asyncio
@pytest.mark.parametrize("model_name", [MODEL_NAME])
async def test_completions_with_empty_image_with_uuid_without_cache_hit(
    client: openai.AsyncOpenAI,
    model_name: str,
):
    with pytest.raises(openai.BadRequestError):
        _ = await client.chat.completions.create(
            messages=[
505
                {"role": "system", "content": "You are a helpful assistant."},
506
                {
507
                    "role": "user",
508
509
510
511
512
513
514
515
                    "content": [
                        {
                            "type": "text",
                            "text": "Describe this image.",
                        },
                        {
                            "type": "image_url",
                            "image_url": {},
516
                            "uuid": "uuid_not_previously_seen",
517
518
519
520
521
522
523
                        },
                    ],
                },
            ],
            model=model_name,
        )

524
525
526
527
528
529

@pytest.mark.asyncio
@pytest.mark.parametrize("model_name", [MODEL_NAME])
@pytest.mark.parametrize(
    "image_urls",
    [TEST_IMAGE_ASSETS[:i] for i in range(2, len(TEST_IMAGE_ASSETS))],
530
531
    indirect=True,
)
532
533
534
535
536
537
538
539
async def test_completions_with_image_with_incorrect_uuid_format(
    client: openai.AsyncOpenAI,
    model_name: str,
    image_urls: list[str],
):
    for image_url in image_urls:
        chat_completion = await client.chat.completions.create(
            messages=[
540
                {"role": "system", "content": "You are a helpful assistant."},
541
                {
542
                    "role": "user",
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
                    "content": [
                        {
                            "type": "text",
                            "text": "Describe this image.",
                        },
                        {
                            "type": "image_url",
                            "image_url": {
                                "url": image_url,
                                "incorrect_uuid_key": image_url,
                            },
                            "also_incorrect_uuid_key": image_url,
                        },
                    ],
                },
            ],
            model=model_name,
        )
        assert chat_completion.choices[0].message.content is not None
        assert isinstance(chat_completion.choices[0].message.content, str)
        assert len(chat_completion.choices[0].message.content) > 0