test_vision.py 14.3 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 os
9
import pytest_asyncio
pansicheng's avatar
pansicheng committed
10
11
12
import requests
from PIL import Image
from transformers import AutoProcessor
13

14
from vllm.multimodal.utils import encode_image_base64, fetch_image
15

16
from ...utils import RemoteOpenAIServer, models_path_prefix, urls_port
17

18
MODEL_NAME = os.path.join(models_path_prefix, "microsoft/Phi-3.5-vision-instruct")
19
MAXIMUM_IMAGES = 2
20

zhuwenwen's avatar
zhuwenwen committed
21
22


23
# Test different image extensions (JPG/PNG) and formats (gray/RGB/RGBA)
zhuwenwen's avatar
zhuwenwen committed
24
25
26
27
28
29
# TEST_IMAGE_URLS = [
#     "https://upload.wikimedia.org/wikipedia/commons/thumb/d/dd/Gfp-wisconsin-madison-the-nature-boardwalk.jpg/2560px-Gfp-wisconsin-madison-the-nature-boardwalk.jpg",
#     "https://upload.wikimedia.org/wikipedia/commons/f/fa/Grayscale_8bits_palette_sample_image.png",
#     "https://upload.wikimedia.org/wikipedia/commons/thumb/9/91/Venn_diagram_rgb.svg/1280px-Venn_diagram_rgb.svg.png",
#     "https://upload.wikimedia.org/wikipedia/commons/0/0b/RGBA_comp.png",
# ]
30
TEST_IMAGE_URLS = [
31
32
33
34
    f"http://localhost:{urls_port}/2560px-Gfp-wisconsin-madison-the-nature-boardwalk.jpg",
    f"http://localhost:{urls_port}/Grayscale_8bits_palette_sample_image.png",
    f"http://localhost:{urls_port}/Venn_diagram_rgb.svg/1280px-Venn_diagram_rgb.svg.png",
    f"http://localhost:{urls_port}/RGBA_comp.png",
35
36
]

37
38
39
40
41
42
43
44
45
46
47
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",
48
        "The image shows a Venn diagram with three intersect",
49
50
51
    ],
    [
        "This image displays a gradient of colors ranging from",
52
        "The image displays a gradient of colors ranging from",
53
54
55
    ],
]

56

57
@pytest.fixture(scope="module")
58
def server():
59
    args = [
60
61
        "--task",
        "generate",
62
63
64
65
66
67
68
        "--max-model-len",
        "2048",
        "--max-num-seqs",
        "5",
        "--enforce-eager",
        "--trust-remote-code",
        "--limit-mm-per-prompt",
69
        json.dumps({"image": MAXIMUM_IMAGES}),
70
71
72
    ]

    with RemoteOpenAIServer(MODEL_NAME, args) as remote_server:
73
        yield remote_server
74
75


76
77
78
79
@pytest_asyncio.fixture
async def client(server):
    async with server.get_async_client() as async_client:
        yield async_client
80
81


82
@pytest.fixture(scope="session")
83
def base64_encoded_image() -> dict[str, str]:
84
    return {
85
        image_url: encode_image_base64(fetch_image(image_url))
86
87
88
89
        for image_url in TEST_IMAGE_URLS
    }


pansicheng's avatar
pansicheng committed
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
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}",
    }]
    images = [Image.open(requests.get(image_url, stream=True).raw)]

    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]


109
110
111
@pytest.mark.asyncio
@pytest.mark.parametrize("model_name", [MODEL_NAME])
@pytest.mark.parametrize("image_url", TEST_IMAGE_URLS)
112
async def test_single_chat_session_image(client: openai.AsyncOpenAI,
113
                                         model_name: str, image_url: str):
pansicheng's avatar
pansicheng committed
114
    content_text = "What's in this image?"
115
116
117
118
119
120
121
122
123
124
125
126
    messages = [{
        "role":
        "user",
        "content": [
            {
                "type": "image_url",
                "image_url": {
                    "url": image_url
                }
            },
            {
                "type": "text",
pansicheng's avatar
pansicheng committed
127
                "text": content_text
128
129
130
131
            },
        ],
    }]

pansicheng's avatar
pansicheng committed
132
    max_completion_tokens = 10
133
    # test single completion
134
135
136
    chat_completion = await client.chat.completions.create(
        model=model_name,
        messages=messages,
pansicheng's avatar
pansicheng committed
137
        max_completion_tokens=max_completion_tokens,
138
        logprobs=True,
139
        temperature=0.0,
140
        top_logprobs=5)
141
142
143
144
    assert len(chat_completion.choices) == 1

    choice = chat_completion.choices[0]
    assert choice.finish_reason == "length"
pansicheng's avatar
pansicheng committed
145
146
    hf_prompt_tokens = get_hf_prompt_tokens(model_name, content_text,
                                            image_url)
147
    assert chat_completion.usage == openai.types.CompletionUsage(
pansicheng's avatar
pansicheng committed
148
149
150
        completion_tokens=max_completion_tokens,
        prompt_tokens=hf_prompt_tokens,
        total_tokens=hf_prompt_tokens + max_completion_tokens)
151
152
153
154
155
156
157
158
159
160
161
162

    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,
163
        max_completion_tokens=10,
164
165
166
167
168
    )
    message = chat_completion.choices[0].message
    assert message.content is not None and len(message.content) >= 0


169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
@pytest.mark.asyncio
@pytest.mark.parametrize("model_name", [MODEL_NAME])
@pytest.mark.parametrize("image_url", TEST_IMAGE_URLS)
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)


199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
@pytest.mark.asyncio
@pytest.mark.parametrize("model_name", [MODEL_NAME])
@pytest.mark.parametrize("image_url", TEST_IMAGE_URLS)
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,
226
        max_completion_tokens=10,
227
228
229
230
231
232
233
234
        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


235
236
237
238
@pytest.mark.asyncio
@pytest.mark.parametrize("model_name", [MODEL_NAME])
@pytest.mark.parametrize("image_url", TEST_IMAGE_URLS)
async def test_single_chat_session_image_base64encoded(
239
        client: openai.AsyncOpenAI, model_name: str, image_url: str,
240
        base64_encoded_image: dict[str, str]):
241

pansicheng's avatar
pansicheng committed
242
    content_text = "What's in this image?"
243
244
245
246
247
248
249
250
251
252
253
254
255
    messages = [{
        "role":
        "user",
        "content": [
            {
                "type": "image_url",
                "image_url": {
                    "url":
                    f"data:image/jpeg;base64,{base64_encoded_image[image_url]}"
                }
            },
            {
                "type": "text",
pansicheng's avatar
pansicheng committed
256
                "text": content_text
257
258
259
260
            },
        ],
    }]

pansicheng's avatar
pansicheng committed
261
    max_completion_tokens = 10
262
    # test single completion
263
264
265
    chat_completion = await client.chat.completions.create(
        model=model_name,
        messages=messages,
pansicheng's avatar
pansicheng committed
266
        max_completion_tokens=max_completion_tokens,
267
        logprobs=True,
268
        temperature=0.0,
269
        top_logprobs=5)
270
271
272
273
    assert len(chat_completion.choices) == 1

    choice = chat_completion.choices[0]
    assert choice.finish_reason == "length"
pansicheng's avatar
pansicheng committed
274
275
    hf_prompt_tokens = get_hf_prompt_tokens(model_name, content_text,
                                            image_url)
276
    assert chat_completion.usage == openai.types.CompletionUsage(
pansicheng's avatar
pansicheng committed
277
278
279
        completion_tokens=max_completion_tokens,
        prompt_tokens=hf_prompt_tokens,
        total_tokens=hf_prompt_tokens + max_completion_tokens)
280
281
282
283
284
285
286
287
288
289
290
291

    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,
292
        max_completion_tokens=10,
293
        temperature=0.0,
294
295
296
297
298
    )
    message = chat_completion.choices[0].message
    assert message.content is not None and len(message.content) >= 0


299
300
@pytest.mark.asyncio
@pytest.mark.parametrize("model_name", [MODEL_NAME])
301
@pytest.mark.parametrize("image_idx", list(range(len(TEST_IMAGE_URLS))))
302
async def test_single_chat_session_image_base64encoded_beamsearch(
303
        client: openai.AsyncOpenAI, model_name: str, image_idx: int,
304
        base64_encoded_image: dict[str, str]):
305
306
307
    # NOTE: This test also validates that we pass MM data through beam search
    image_url = TEST_IMAGE_URLS[image_idx]
    expected_res = EXPECTED_MM_BEAM_SEARCH_RES[image_idx]
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329

    messages = [{
        "role":
        "user",
        "content": [
            {
                "type": "image_url",
                "image_url": {
                    "url":
                    f"data:image/jpeg;base64,{base64_encoded_image[image_url]}"
                }
            },
            {
                "type": "text",
                "text": "What's in this image?"
            },
        ],
    }]
    chat_completion = await client.chat.completions.create(
        model=model_name,
        messages=messages,
        n=2,
330
        max_completion_tokens=10,
331
        temperature=0.0,
332
333
        extra_body=dict(use_beam_search=True))
    assert len(chat_completion.choices) == 2
334
335
    for actual, expected_str in zip(chat_completion.choices, expected_res):
        assert actual.message.content == expected_str
336
337


338
339
340
@pytest.mark.asyncio
@pytest.mark.parametrize("model_name", [MODEL_NAME])
@pytest.mark.parametrize("image_url", TEST_IMAGE_URLS)
341
async def test_chat_streaming_image(client: openai.AsyncOpenAI,
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
                                    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,
364
        max_completion_tokens=10,
365
366
367
368
369
370
371
372
373
        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,
374
        max_completion_tokens=10,
375
376
377
        temperature=0.0,
        stream=True,
    )
378
    chunks: list[str] = []
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
    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])
397
398
399
@pytest.mark.parametrize(
    "image_urls",
    [TEST_IMAGE_URLS[:i] for i in range(2, len(TEST_IMAGE_URLS))])
400
async def test_multi_image_input(client: openai.AsyncOpenAI, model_name: str,
401
                                 image_urls: list[str]):
402
403
404
405
406

    messages = [{
        "role":
        "user",
        "content": [
407
            *({
408
409
410
411
                "type": "image_url",
                "image_url": {
                    "url": image_url
                }
412
            } for image_url in image_urls),
413
414
415
416
417
418
419
            {
                "type": "text",
                "text": "What's in this image?"
            },
        ],
    }]

420
421
422
423
424
    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,
425
                max_completion_tokens=10,
426
427
428
429
430
431
432
433
434
435
436
437
438
439
                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(
440
441
            model=model_name,
            messages=messages,
442
            max_completion_tokens=10,
443
444
            temperature=0.0,
        )
445
446
        message = chat_completion.choices[0].message
        assert message.content is not None and len(message.content) >= 0