test_vision_openai_server.py 19.9 KB
Newer Older
1
2
3
"""
Usage:
python3 -m unittest test_vision_openai_server.TestOpenAIVisionServer.test_mixed_batch
4
python3 -m unittest test_vision_openai_server.TestOpenAIVisionServer.test_multi_images_chat_completion
5
6
"""

7
8
import base64
import io
Ying Sheng's avatar
Ying Sheng committed
9
import json
10
import os
Ying Sheng's avatar
Ying Sheng committed
11
import unittest
12
from concurrent.futures import ThreadPoolExecutor
Ying Sheng's avatar
Ying Sheng committed
13

14
import numpy as np
Ying Sheng's avatar
Ying Sheng committed
15
import openai
16
17
import requests
from PIL import Image
Ying Sheng's avatar
Ying Sheng committed
18

19
from sglang.srt.utils import kill_process_tree
20
21
22
23
24
from sglang.test.test_utils import (
    DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
    DEFAULT_URL_FOR_TEST,
    popen_launch_server,
)
Ying Sheng's avatar
Ying Sheng committed
25

26
27
28
29
30
31
32
33
34
35
36
# image
IMAGE_MAN_IRONING_URL = "https://raw.githubusercontent.com/sgl-project/sgl-test-files/refs/heads/main/images/man_ironing_on_back_of_suv.png"
IMAGE_SGL_LOGO_URL = "https://raw.githubusercontent.com/sgl-project/sgl-test-files/refs/heads/main/images/sgl_logo.png"

# video
VIDEO_JOBS_URL = "https://raw.githubusercontent.com/sgl-project/sgl-test-files/refs/heads/main/videos/jobs_presenting_ipod.mp4"

# audio
AUDIO_TRUMP_SPEECH_URL = "https://raw.githubusercontent.com/sgl-project/sgl-test-files/refs/heads/main/audios/Trump_WEF_2018_10s.mp3"
AUDIO_BIRD_SONG_URL = "https://raw.githubusercontent.com/sgl-project/sgl-test-files/refs/heads/main/audios/bird_song.mp3"

Ying Sheng's avatar
Ying Sheng committed
37
38
39
40

class TestOpenAIVisionServer(unittest.TestCase):
    @classmethod
    def setUpClass(cls):
41
        cls.model = "lmms-lab/llava-onevision-qwen2-0.5b-ov"
42
        cls.base_url = DEFAULT_URL_FOR_TEST
Ying Sheng's avatar
Ying Sheng committed
43
44
45
46
        cls.api_key = "sk-123456"
        cls.process = popen_launch_server(
            cls.model,
            cls.base_url,
47
            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
Ying Sheng's avatar
Ying Sheng committed
48
49
50
            api_key=cls.api_key,
            other_args=[
                "--chat-template",
51
                "chatml-llava",
52
                # "--log-requests",
Ying Sheng's avatar
Ying Sheng committed
53
54
55
56
57
58
            ],
        )
        cls.base_url += "/v1"

    @classmethod
    def tearDownClass(cls):
59
        kill_process_tree(cls.process.pid)
Ying Sheng's avatar
Ying Sheng committed
60

61
    def test_single_image_chat_completion(self):
Ying Sheng's avatar
Ying Sheng committed
62
63
64
65
66
67
68
69
70
71
        client = openai.Client(api_key=self.api_key, base_url=self.base_url)

        response = client.chat.completions.create(
            model="default",
            messages=[
                {
                    "role": "user",
                    "content": [
                        {
                            "type": "image_url",
72
                            "image_url": {"url": IMAGE_MAN_IRONING_URL},
Ying Sheng's avatar
Ying Sheng committed
73
                        },
Ying Sheng's avatar
Ying Sheng committed
74
75
76
77
                        {
                            "type": "text",
                            "text": "Describe this image in a very short sentence.",
                        },
Ying Sheng's avatar
Ying Sheng committed
78
79
80
81
82
83
84
                    ],
                },
            ],
            temperature=0,
        )

        assert response.choices[0].message.role == "assistant"
Ying Sheng's avatar
Ying Sheng committed
85
86
        text = response.choices[0].message.content
        assert isinstance(text, str)
87
88
        # `driver` is for gemma-3-it
        assert "man" in text or "person" or "driver" in text, text
89
90
        assert "cab" in text or "taxi" in text or "SUV" in text, text
        assert "iron" in text, text
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
        assert response.id
        assert response.created
        assert response.usage.prompt_tokens > 0
        assert response.usage.completion_tokens > 0
        assert response.usage.total_tokens > 0

    def test_multi_turn_chat_completion(self):
        client = openai.Client(api_key=self.api_key, base_url=self.base_url)

        response = client.chat.completions.create(
            model="default",
            messages=[
                {
                    "role": "user",
                    "content": [
                        {
                            "type": "image_url",
108
                            "image_url": {"url": IMAGE_MAN_IRONING_URL},
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
                        },
                        {
                            "type": "text",
                            "text": "Describe this image in a very short sentence.",
                        },
                    ],
                },
                {
                    "role": "assistant",
                    "content": [
                        {
                            "type": "text",
                            "text": "There is a man at the back of a yellow cab ironing his clothes.",
                        }
                    ],
                },
                {
                    "role": "user",
                    "content": [
                        {"type": "text", "text": "Repeat your previous answer."}
                    ],
                },
            ],
            temperature=0,
        )

        assert response.choices[0].message.role == "assistant"
        text = response.choices[0].message.content
        assert isinstance(text, str)
138
        assert "man" in text or "cab" in text, text
Ying Sheng's avatar
Ying Sheng committed
139
140
141
142
        assert response.id
        assert response.created
        assert response.usage.prompt_tokens > 0
        assert response.usage.completion_tokens > 0
143
144
        assert response.usage.total_tokens > 0

145
    def test_multi_images_chat_completion(self):
146
147
148
149
150
151
152
153
154
155
156
        client = openai.Client(api_key=self.api_key, base_url=self.base_url)

        response = client.chat.completions.create(
            model="default",
            messages=[
                {
                    "role": "user",
                    "content": [
                        {
                            "type": "image_url",
                            "image_url": {
157
                                "url": "https://raw.githubusercontent.com/sgl-project/sglang/main/test/lang/example_image.png"
158
                            },
159
                            "modalities": "multi-images",
160
161
162
                        },
                        {
                            "type": "image_url",
163
                            "image_url": {"url": IMAGE_SGL_LOGO_URL},
164
                            "modalities": "multi-images",
165
166
167
                        },
                        {
                            "type": "text",
168
169
                            "text": "I have two very different images. They are not related at all. "
                            "Please describe the first image in one sentence, and then describe the second image in another sentence.",
170
171
172
173
174
175
176
177
178
179
                        },
                    ],
                },
            ],
            temperature=0,
        )

        assert response.choices[0].message.role == "assistant"
        text = response.choices[0].message.content
        assert isinstance(text, str)
180
        print(f"LLM response: {text}")
Mick's avatar
Mick committed
181
        assert "man" in text or "cab" in text or "SUV" in text or "taxi" in text, text
182
        assert "logo" in text or '"S"' in text or "SG" in text, text
183
184
185
186
        assert response.id
        assert response.created
        assert response.usage.prompt_tokens > 0
        assert response.usage.completion_tokens > 0
Ying Sheng's avatar
Ying Sheng committed
187
188
        assert response.usage.total_tokens > 0

189
    def prepare_video_messages(self, video_path):
190
191
        # the memory consumed by the Vision Attention varies a lot, e.g. blocked qkv vs full-sequence sdpa
        # the size of the video embeds differs from the `modality` argument when preprocessed
192
193
194
195
196
197
198

        # We import decord here to avoid a strange Segmentation fault (core dumped) issue.
        # The following import order will cause Segmentation fault.
        # import decord
        # from transformers import AutoTokenizer
        from decord import VideoReader, cpu

199
        max_frames_num = 20
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
        vr = VideoReader(video_path, ctx=cpu(0))
        total_frame_num = len(vr)
        uniform_sampled_frames = np.linspace(
            0, total_frame_num - 1, max_frames_num, dtype=int
        )
        frame_idx = uniform_sampled_frames.tolist()
        frames = vr.get_batch(frame_idx).asnumpy()

        base64_frames = []
        for frame in frames:
            pil_img = Image.fromarray(frame)
            buff = io.BytesIO()
            pil_img.save(buff, format="JPEG")
            base64_str = base64.b64encode(buff.getvalue()).decode("utf-8")
            base64_frames.append(base64_str)

        messages = [{"role": "user", "content": []}]
        frame_format = {
            "type": "image_url",
            "image_url": {"url": "data:image/jpeg;base64,{}"},
220
            "modalities": "video",
221
222
223
224
225
226
227
228
229
230
231
232
233
        }

        for base64_frame in base64_frames:
            frame_format["image_url"]["url"] = "data:image/jpeg;base64,{}".format(
                base64_frame
            )
            messages[0]["content"].append(frame_format.copy())

        prompt = {"type": "text", "text": "Please describe the video in detail."}
        messages[0]["content"].append(prompt)

        return messages

234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
    def prepare_video_messages_video_direct(self, video_path):
        messages = [
            {
                "role": "user",
                "content": [
                    {
                        "type": "image_url",
                        "image_url": {"url": f"video:{video_path}"},
                        "modalities": "video",
                    },
                    {"type": "text", "text": "Please describe the video in detail."},
                ],
            },
        ]
        return messages

250
    def get_or_download_file(self, url: str) -> str:
251
        cache_dir = os.path.expanduser("~/.cache")
252
253
254
255
        if url is None:
            raise ValueError()
        file_name = url.split("/")[-1]
        file_path = os.path.join(cache_dir, file_name)
256
257
258
259
260
261
262
263
        os.makedirs(cache_dir, exist_ok=True)

        if not os.path.exists(file_path):
            response = requests.get(url)
            response.raise_for_status()

            with open(file_path, "wb") as f:
                f.write(response.content)
264
265
266
267
268
        return file_path

    def test_video_chat_completion(self):
        url = VIDEO_JOBS_URL
        file_path = self.get_or_download_file(url)
269
270
271

        client = openai.Client(api_key=self.api_key, base_url=self.base_url)

272
        # messages = self.prepare_video_messages_video_direct(file_path)
273
274
275
276
277
278
279
280
281
        messages = self.prepare_video_messages(file_path)

        video_request = client.chat.completions.create(
            model="default",
            messages=messages,
            temperature=0,
            max_tokens=1024,
            stream=True,
        )
282

283
284
285
286
287
288
        print("-" * 30)
        video_response = ""
        for chunk in video_request:
            if chunk.choices[0].delta.content is not None:
                content = chunk.choices[0].delta.content
                video_response += content
289
                print(content, end="", flush=True)
290
291
292
        print("-" * 30)

        # Add assertions to validate the video response
Mick's avatar
Mick committed
293
294
295
296
297
        assert "iPod" in video_response or "device" in video_response, video_response
        assert (
            "man" in video_response
            or "person" in video_response
            or "individual" in video_response
298
            or "speaker" in video_response
Mick's avatar
Mick committed
299
300
301
302
303
        ), video_response
        assert (
            "present" in video_response
            or "examine" in video_response
            or "display" in video_response
304
            or "hold" in video_response
Mick's avatar
Mick committed
305
306
        )
        assert "black" in video_response or "dark" in video_response
307
308
309
        self.assertIsNotNone(video_response)
        self.assertGreater(len(video_response), 0)

Ying Sheng's avatar
Ying Sheng committed
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
    def test_regex(self):
        client = openai.Client(api_key=self.api_key, base_url=self.base_url)

        regex = (
            r"""\{\n"""
            + r"""   "color": "[\w]+",\n"""
            + r"""   "number_of_cars": [\d]+\n"""
            + r"""\}"""
        )

        response = client.chat.completions.create(
            model="default",
            messages=[
                {
                    "role": "user",
                    "content": [
                        {
                            "type": "image_url",
328
                            "image_url": {"url": IMAGE_MAN_IRONING_URL},
Ying Sheng's avatar
Ying Sheng committed
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
                        },
                        {
                            "type": "text",
                            "text": "Describe this image in the JSON format.",
                        },
                    ],
                },
            ],
            temperature=0,
            extra_body={"regex": regex},
        )
        text = response.choices[0].message.content

        try:
            js_obj = json.loads(text)
        except (TypeError, json.decoder.JSONDecodeError):
            print("JSONDecodeError", text)
            raise
        assert isinstance(js_obj["color"], str)
        assert isinstance(js_obj["number_of_cars"], int)

350
351
352
353
354
355
356
357
    def run_decode_with_image(self, image_id):
        client = openai.Client(api_key=self.api_key, base_url=self.base_url)

        content = []
        if image_id == 0:
            content.append(
                {
                    "type": "image_url",
358
                    "image_url": {"url": IMAGE_MAN_IRONING_URL},
359
360
361
362
363
364
                }
            )
        elif image_id == 1:
            content.append(
                {
                    "type": "image_url",
365
                    "image_url": {"url": IMAGE_SGL_LOGO_URL},
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
                }
            )
        else:
            pass

        content.append(
            {
                "type": "text",
                "text": "Describe this image in a very short sentence.",
            }
        )

        response = client.chat.completions.create(
            model="default",
            messages=[
                {"role": "user", "content": content},
            ],
            temperature=0,
        )

        assert response.choices[0].message.role == "assistant"
        text = response.choices[0].message.content
        assert isinstance(text, str)

    def test_mixed_batch(self):
        image_ids = [0, 1, 2] * 4
        with ThreadPoolExecutor(4) as executor:
            list(executor.map(self.run_decode_with_image, image_ids))

Ying Sheng's avatar
Ying Sheng committed
395

396
class TestQwen2VLServer(TestOpenAIVisionServer):
Yineng Zhang's avatar
Yineng Zhang committed
397
398
399
400
401
402
403
404
405
406
407
408
409
    @classmethod
    def setUpClass(cls):
        cls.model = "Qwen/Qwen2-VL-7B-Instruct"
        cls.base_url = DEFAULT_URL_FOR_TEST
        cls.api_key = "sk-123456"
        cls.process = popen_launch_server(
            cls.model,
            cls.base_url,
            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
            api_key=cls.api_key,
            other_args=[
                "--chat-template",
                "qwen2-vl",
410
411
                "--mem-fraction-static",
                "0.4",
Yineng Zhang's avatar
Yineng Zhang committed
412
413
414
415
416
            ],
        )
        cls.base_url += "/v1"


417
class TestQwen2_5_VLServer(TestOpenAIVisionServer):
Mick's avatar
Mick committed
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
    @classmethod
    def setUpClass(cls):
        cls.model = "Qwen/Qwen2.5-VL-7B-Instruct"
        cls.base_url = DEFAULT_URL_FOR_TEST
        cls.api_key = "sk-123456"
        cls.process = popen_launch_server(
            cls.model,
            cls.base_url,
            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
            api_key=cls.api_key,
            other_args=[
                "--chat-template",
                "qwen2-vl",
                "--mem-fraction-static",
                "0.4",
            ],
        )
        cls.base_url += "/v1"


438
class TestVLMContextLengthIssue(unittest.TestCase):
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
    @classmethod
    def setUpClass(cls):
        cls.model = "Qwen/Qwen2-VL-7B-Instruct"
        cls.base_url = DEFAULT_URL_FOR_TEST
        cls.api_key = "sk-123456"
        cls.process = popen_launch_server(
            cls.model,
            cls.base_url,
            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
            api_key=cls.api_key,
            other_args=[
                "--chat-template",
                "qwen2-vl",
                "--context-length",
                "300",
                "--mem-fraction-static=0.80",
            ],
        )
        cls.base_url += "/v1"

    @classmethod
    def tearDownClass(cls):
461
        kill_process_tree(cls.process.pid)
462

463
    def test_single_image_chat_completion(self):
464
465
        client = openai.Client(api_key=self.api_key, base_url=self.base_url)

466
467
468
469
470
471
472
473
474
        with self.assertRaises(openai.BadRequestError) as cm:
            client.chat.completions.create(
                model="default",
                messages=[
                    {
                        "role": "user",
                        "content": [
                            {
                                "type": "image_url",
475
                                "image_url": {"url": IMAGE_MAN_IRONING_URL},
476
                            },
477
478
479
480
481
482
483
484
485
                            {
                                "type": "text",
                                "text": "Give a lengthy description of this picture",
                            },
                        ],
                    },
                ],
                temperature=0,
            )
486

487
488
489
490
491
        # context length is checked first, then max_req_input_len, which is calculated from the former
        assert (
            "Multimodal prompt is too long after expanding multimodal tokens."
            in str(cm.exception)
            or "is longer than the model's context length" in str(cm.exception)
492
        )
493
494


495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
class TestMllamaServer(TestOpenAIVisionServer):
    @classmethod
    def setUpClass(cls):
        cls.model = "meta-llama/Llama-3.2-11B-Vision-Instruct"
        cls.base_url = DEFAULT_URL_FOR_TEST
        cls.api_key = "sk-123456"
        cls.process = popen_launch_server(
            cls.model,
            cls.base_url,
            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
            api_key=cls.api_key,
            other_args=[
                "--chat-template",
                "llama_3_vision",
            ],
        )
        cls.base_url += "/v1"

    def test_video_chat_completion(self):
        pass


Mick's avatar
Mick committed
517
518
519
520
521
522
523
524
525
526
527
528
529
530
class TestMinicpmvServer(TestOpenAIVisionServer):
    @classmethod
    def setUpClass(cls):
        cls.model = "openbmb/MiniCPM-V-2_6"
        cls.base_url = DEFAULT_URL_FOR_TEST
        cls.api_key = "sk-123456"
        cls.process = popen_launch_server(
            cls.model,
            cls.base_url,
            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
            other_args=[
                "--trust-remote-code",
                "--chat-template",
                "minicpmv",
531
532
                "--mem-fraction-static",
                "0.4",
Mick's avatar
Mick committed
533
534
535
536
537
            ],
        )
        cls.base_url += "/v1"


538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
class TestDeepseekVL2Server(TestOpenAIVisionServer):
    @classmethod
    def setUpClass(cls):
        cls.model = "deepseek-ai/deepseek-vl2-small"
        cls.base_url = DEFAULT_URL_FOR_TEST
        cls.api_key = "sk-123456"
        cls.process = popen_launch_server(
            cls.model,
            cls.base_url,
            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
            other_args=[
                "--trust-remote-code",
                "--chat-template",
                "deepseek-vl2",
                "--context-length",
                "4096",
            ],
        )
        cls.base_url += "/v1"

    def test_video_chat_completion(self):
        pass


Mick's avatar
Mick committed
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
class TestJanusProServer(TestOpenAIVisionServer):
    @classmethod
    def setUpClass(cls):
        cls.model = "deepseek-ai/Janus-Pro-7B"
        cls.base_url = DEFAULT_URL_FOR_TEST
        cls.api_key = "sk-123456"
        cls.process = popen_launch_server(
            cls.model,
            cls.base_url,
            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
            other_args=[
                "--trust-remote-code",
                "--chat-template",
                "janus-pro",
                "--mem-fraction-static",
                "0.4",
            ],
        )
        cls.base_url += "/v1"

    def test_video_chat_completion(self):
        pass

585
586
587
588
    def test_single_image_chat_completion(self):
        # Skip this test because it is flaky
        pass

Mick's avatar
Mick committed
589

590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
class TestGemma3itServer(TestOpenAIVisionServer):
    @classmethod
    def setUpClass(cls):
        cls.model = "google/gemma-3-4b-it"
        cls.base_url = DEFAULT_URL_FOR_TEST
        cls.api_key = "sk-123456"
        cls.process = popen_launch_server(
            cls.model,
            cls.base_url,
            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
            other_args=[
                "--trust-remote-code",
                "--chat-template",
                "gemma-it",
            ],
        )
        cls.base_url += "/v1"

    def test_video_chat_completion(self):
        pass


Ying Sheng's avatar
Ying Sheng committed
612
if __name__ == "__main__":
Lianmin Zheng's avatar
Lianmin Zheng committed
613
    unittest.main()