test_vision.py 14.2 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.utils import encode_image_base64, fetch_image
12

13
from ...utils import RemoteOpenAIServer
14

15
16
MODEL_NAME = "microsoft/Phi-3.5-vision-instruct"
MAXIMUM_IMAGES = 2
17

18
# Test different image extensions (JPG/PNG) and formats (gray/RGB/RGBA)
19
20
21
22
23
TEST_IMAGE_ASSETS = [
    "2560px-Gfp-wisconsin-madison-the-nature-boardwalk.jpg",  # "https://upload.wikimedia.org/wikipedia/commons/thumb/d/dd/Gfp-wisconsin-madison-the-nature-boardwalk.jpg/2560px-Gfp-wisconsin-madison-the-nature-boardwalk.jpg"
    "Grayscale_8bits_palette_sample_image.png",  # "https://upload.wikimedia.org/wikipedia/commons/f/fa/Grayscale_8bits_palette_sample_image.png",
    "1280px-Venn_diagram_rgb.svg.png",  # "https://upload.wikimedia.org/wikipedia/commons/thumb/9/91/Venn_diagram_rgb.svg/1280px-Venn_diagram_rgb.svg.png",
    "RGBA_comp.png",  # "https://upload.wikimedia.org/wikipedia/commons/0/0b/RGBA_comp.png",
24
25
]

26
27
28
29
30
31
32
33
34
35
36
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",
37
        "The image shows a Venn diagram with three intersect",
38
39
40
    ],
    [
        "This image displays a gradient of colors ranging from",
41
        "The image displays a gradient of colors ranging from",
42
43
44
    ],
]

45

46
@pytest.fixture(scope="module")
47
def server():
48
    args = [
49
        "--runner",
50
        "generate",
51
52
53
54
55
56
57
        "--max-model-len",
        "2048",
        "--max-num-seqs",
        "5",
        "--enforce-eager",
        "--trust-remote-code",
        "--limit-mm-per-prompt",
58
        json.dumps({"image": MAXIMUM_IMAGES}),
59
60
61
    ]

    with RemoteOpenAIServer(MODEL_NAME, args) as remote_server:
62
        yield remote_server
63
64


65
66
67
68
@pytest_asyncio.fixture
async def client(server):
    async with server.get_async_client() as async_client:
        yield async_client
69
70


71
@pytest.fixture(scope="session")
72
def base64_encoded_image(local_asset_server) -> dict[str, str]:
73
    return {
74
75
76
        image_asset:
        encode_image_base64(local_asset_server.get_image_asset(image_asset))
        for image_asset in TEST_IMAGE_ASSETS
77
78
79
    }


pansicheng's avatar
pansicheng committed
80
81
82
83
84
85
86
87
88
89
def get_hf_prompt_tokens(model_name, content, image_url):
    processor = AutoProcessor.from_pretrained(model_name,
                                              trust_remote_code=True,
                                              num_crops=4)

    placeholder = "<|image_1|>\n"
    messages = [{
        "role": "user",
        "content": f"{placeholder}{content}",
    }]
90
    images = [fetch_image(image_url)]
pansicheng's avatar
pansicheng committed
91
92
93
94
95
96
97
98

    prompt = processor.tokenizer.apply_chat_template(
        messages, tokenize=False, add_generation_prompt=True)
    inputs = processor(prompt, images, return_tensors="pt")

    return inputs.input_ids.shape[1]


99
100
@pytest.mark.asyncio
@pytest.mark.parametrize("model_name", [MODEL_NAME])
101
@pytest.mark.parametrize("image_url", TEST_IMAGE_ASSETS, indirect=True)
102
async def test_single_chat_session_image(client: openai.AsyncOpenAI,
103
                                         model_name: str, image_url: str):
pansicheng's avatar
pansicheng committed
104
    content_text = "What's in this image?"
105
106
107
108
109
110
111
112
113
114
115
116
    messages = [{
        "role":
        "user",
        "content": [
            {
                "type": "image_url",
                "image_url": {
                    "url": image_url
                }
            },
            {
                "type": "text",
pansicheng's avatar
pansicheng committed
117
                "text": content_text
118
119
120
121
            },
        ],
    }]

pansicheng's avatar
pansicheng committed
122
    max_completion_tokens = 10
123
    # test single completion
124
125
126
    chat_completion = await client.chat.completions.create(
        model=model_name,
        messages=messages,
pansicheng's avatar
pansicheng committed
127
        max_completion_tokens=max_completion_tokens,
128
        logprobs=True,
129
        temperature=0.0,
130
        top_logprobs=5)
131
132
133
134
    assert len(chat_completion.choices) == 1

    choice = chat_completion.choices[0]
    assert choice.finish_reason == "length"
pansicheng's avatar
pansicheng committed
135
136
    hf_prompt_tokens = get_hf_prompt_tokens(model_name, content_text,
                                            image_url)
137
    assert chat_completion.usage == openai.types.CompletionUsage(
pansicheng's avatar
pansicheng committed
138
139
140
        completion_tokens=max_completion_tokens,
        prompt_tokens=hf_prompt_tokens,
        total_tokens=hf_prompt_tokens + max_completion_tokens)
141
142
143
144
145
146
147
148
149
150
151
152

    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,
153
        max_completion_tokens=10,
154
155
156
157
158
    )
    message = chat_completion.choices[0].message
    assert message.content is not None and len(message.content) >= 0


159
160
@pytest.mark.asyncio
@pytest.mark.parametrize("model_name", [MODEL_NAME])
161
@pytest.mark.parametrize("image_url", TEST_IMAGE_ASSETS, indirect=True)
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
async def test_error_on_invalid_image_url_type(client: openai.AsyncOpenAI,
                                               model_name: str,
                                               image_url: str):
    content_text = "What's in this image?"
    messages = [{
        "role":
        "user",
        "content": [
            {
                "type": "image_url",
                "image_url": image_url
            },
            {
                "type": "text",
                "text": content_text
            },
        ],
    }]

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


189
190
@pytest.mark.asyncio
@pytest.mark.parametrize("model_name", [MODEL_NAME])
191
@pytest.mark.parametrize("image_url", TEST_IMAGE_ASSETS, indirect=True)
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
async def test_single_chat_session_image_beamsearch(client: openai.AsyncOpenAI,
                                                    model_name: str,
                                                    image_url: str):
    messages = [{
        "role":
        "user",
        "content": [
            {
                "type": "image_url",
                "image_url": {
                    "url": image_url
                }
            },
            {
                "type": "text",
                "text": "What's in this image?"
            },
        ],
    }]

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


225
226
@pytest.mark.asyncio
@pytest.mark.parametrize("model_name", [MODEL_NAME])
227
228
@pytest.mark.parametrize("raw_image_url", TEST_IMAGE_ASSETS)
@pytest.mark.parametrize("image_url", TEST_IMAGE_ASSETS, indirect=True)
229
async def test_single_chat_session_image_base64encoded(
230
231
        client: openai.AsyncOpenAI, model_name: str, raw_image_url: str,
        image_url: str, base64_encoded_image: dict[str, str]):
232

pansicheng's avatar
pansicheng committed
233
    content_text = "What's in this image?"
234
235
236
237
238
239
240
241
    messages = [{
        "role":
        "user",
        "content": [
            {
                "type": "image_url",
                "image_url": {
                    "url":
242
                    f"data:image/jpeg;base64,{base64_encoded_image[raw_image_url]}"
243
244
245
246
                }
            },
            {
                "type": "text",
pansicheng's avatar
pansicheng committed
247
                "text": content_text
248
249
250
251
            },
        ],
    }]

pansicheng's avatar
pansicheng committed
252
    max_completion_tokens = 10
253
    # test single completion
254
255
256
    chat_completion = await client.chat.completions.create(
        model=model_name,
        messages=messages,
pansicheng's avatar
pansicheng committed
257
        max_completion_tokens=max_completion_tokens,
258
        logprobs=True,
259
        temperature=0.0,
260
        top_logprobs=5)
261
262
263
264
    assert len(chat_completion.choices) == 1

    choice = chat_completion.choices[0]
    assert choice.finish_reason == "length"
pansicheng's avatar
pansicheng committed
265
266
    hf_prompt_tokens = get_hf_prompt_tokens(model_name, content_text,
                                            image_url)
267
    assert chat_completion.usage == openai.types.CompletionUsage(
pansicheng's avatar
pansicheng committed
268
269
270
        completion_tokens=max_completion_tokens,
        prompt_tokens=hf_prompt_tokens,
        total_tokens=hf_prompt_tokens + max_completion_tokens)
271
272
273
274
275
276
277
278
279
280
281
282

    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,
283
        max_completion_tokens=10,
284
        temperature=0.0,
285
286
287
288
289
    )
    message = chat_completion.choices[0].message
    assert message.content is not None and len(message.content) >= 0


290
291
@pytest.mark.asyncio
@pytest.mark.parametrize("model_name", [MODEL_NAME])
292
@pytest.mark.parametrize("image_idx", list(range(len(TEST_IMAGE_ASSETS))))
293
async def test_single_chat_session_image_base64encoded_beamsearch(
294
        client: openai.AsyncOpenAI, model_name: str, image_idx: int,
295
        base64_encoded_image: dict[str, str]):
296
    # NOTE: This test also validates that we pass MM data through beam search
297
    raw_image_url = TEST_IMAGE_ASSETS[image_idx]
298
    expected_res = EXPECTED_MM_BEAM_SEARCH_RES[image_idx]
299
300
301
302
303
304
305
306
307

    messages = [{
        "role":
        "user",
        "content": [
            {
                "type": "image_url",
                "image_url": {
                    "url":
308
                    f"data:image/jpeg;base64,{base64_encoded_image[raw_image_url]}"
309
310
311
312
313
314
315
316
317
318
319
320
                }
            },
            {
                "type": "text",
                "text": "What's in this image?"
            },
        ],
    }]
    chat_completion = await client.chat.completions.create(
        model=model_name,
        messages=messages,
        n=2,
321
        max_completion_tokens=10,
322
        temperature=0.0,
323
324
        extra_body=dict(use_beam_search=True))
    assert len(chat_completion.choices) == 2
325
326
    for actual, expected_str in zip(chat_completion.choices, expected_res):
        assert actual.message.content == expected_str
327
328


329
330
@pytest.mark.asyncio
@pytest.mark.parametrize("model_name", [MODEL_NAME])
331
@pytest.mark.parametrize("image_url", TEST_IMAGE_ASSETS, indirect=True)
332
async def test_chat_streaming_image(client: openai.AsyncOpenAI,
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
                                    model_name: str, image_url: str):
    messages = [{
        "role":
        "user",
        "content": [
            {
                "type": "image_url",
                "image_url": {
                    "url": image_url
                }
            },
            {
                "type": "text",
                "text": "What's in this image?"
            },
        ],
    }]

    # test single completion
    chat_completion = await client.chat.completions.create(
        model=model_name,
        messages=messages,
355
        max_completion_tokens=10,
356
357
358
359
360
361
362
363
364
        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,
365
        max_completion_tokens=10,
366
367
368
        temperature=0.0,
        stream=True,
    )
369
    chunks: list[str] = []
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
    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])
388
389
@pytest.mark.parametrize(
    "image_urls",
390
391
    [TEST_IMAGE_ASSETS[:i] for i in range(2, len(TEST_IMAGE_ASSETS))],
    indirect=True)
392
async def test_multi_image_input(client: openai.AsyncOpenAI, model_name: str,
393
                                 image_urls: list[str]):
394
395
396
397
398

    messages = [{
        "role":
        "user",
        "content": [
399
            *({
400
401
402
403
                "type": "image_url",
                "image_url": {
                    "url": image_url
                }
404
            } for image_url in image_urls),
405
406
407
408
409
410
411
            {
                "type": "text",
                "text": "What's in this image?"
            },
        ],
    }]

412
413
414
415
416
    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,
417
                max_completion_tokens=10,
418
419
420
421
422
423
424
425
426
427
428
429
430
431
                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(
432
433
            model=model_name,
            messages=messages,
434
            max_completion_tokens=10,
435
436
            temperature=0.0,
        )
437
438
        message = chat_completion.choices[0].message
        assert message.content is not None and len(message.content) >= 0