vision_language.py 57.7 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3
"""
Cyrus Leung's avatar
Cyrus Leung committed
4
5
This example shows how to use vLLM for running offline inference with
the correct prompt format on vision language models for text generation.
6
7
8
9

For most models, the prompt format should follow corresponding examples
on HuggingFace model repository.
"""
10

11
import os
12
import random
13
from contextlib import contextmanager
14
from dataclasses import asdict
15
from typing import NamedTuple
16

17
from huggingface_hub import snapshot_download
18
19
from transformers import AutoTokenizer

20
from vllm import LLM, EngineArgs, SamplingParams
21
from vllm.assets.image import ImageAsset
22
from vllm.assets.video import VideoAsset
23
from vllm.lora.request import LoRARequest
24
from vllm.multimodal.image import convert_image_mode
25
26
from vllm.utils import FlexibleArgumentParser

27
28
29
30

class ModelRequestData(NamedTuple):
    engine_args: EngineArgs
    prompts: list[str]
31
32
    stop_token_ids: list[int] | None = None
    lora_requests: list[LoRARequest] | None = None
33
34


35
36
37
38
# NOTE: The default `max_num_seqs` and `max_model_len` may result in OOM on
# lower-end GPUs.
# Unless specified, these settings have been tested to work on a single L4.

39

40
# Aria
41
def run_aria(questions: list[str], modality: str) -> ModelRequestData:
42
43
44
    assert modality == "image"
    model_name = "rhymes-ai/Aria"

45
    # NOTE: Need L40 (or equivalent) to avoid OOM
46
47
48
49
50
    engine_args = EngineArgs(
        model=model_name,
        max_model_len=4096,
        max_num_seqs=2,
        dtype="bfloat16",
51
        limit_mm_per_prompt={modality: 1},
52
    )
53

54
55
56
57
58
59
60
    prompts = [
        (
            f"<|im_start|>user\n<fim_prefix><|img|><fim_suffix>{question}"
            "<|im_end|>\n<|im_start|>assistant\n"
        )
        for question in questions
    ]
61
62

    stop_token_ids = [93532, 93653, 944, 93421, 1019, 93653, 93519]
63
64
65
66
67
68

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
        stop_token_ids=stop_token_ids,
    )
69
70


Jennifer Zhao's avatar
Jennifer Zhao committed
71
72
73
74
75
76
77
78
79
80
# Aya Vision
def run_aya_vision(questions: list[str], modality: str) -> ModelRequestData:
    assert modality == "image"
    model_name = "CohereForAI/aya-vision-8b"

    engine_args = EngineArgs(
        model=model_name,
        max_model_len=2048,
        max_num_seqs=2,
        mm_processor_kwargs={"crop_to_patches": True},
81
        limit_mm_per_prompt={modality: 1},
Jennifer Zhao's avatar
Jennifer Zhao committed
82
83
84
85
86
87
88
89
90
91
92
    )
    prompts = [
        f"<|START_OF_TURN_TOKEN|><|USER_TOKEN|><image>{question}<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>"
        for question in questions
    ]
    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )


93
# BLIP-2
94
def run_blip2(questions: list[str], modality: str) -> ModelRequestData:
95
96
97
98
    assert modality == "image"

    # BLIP-2 prompt format is inaccurate on HuggingFace model repository.
    # See https://huggingface.co/Salesforce/blip2-opt-2.7b/discussions/15#64ff02f3f8cf9e4f5b038262 #noqa
99
    prompts = [f"Question: {question} Answer:" for question in questions]
100
    engine_args = EngineArgs(
101
        model="Salesforce/blip2-opt-2.7b",
102
        limit_mm_per_prompt={modality: 1},
103
104
105
106
107
108
    )

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
109
110
111


# Chameleon
112
def run_chameleon(questions: list[str], modality: str) -> ModelRequestData:
113
114
    assert modality == "image"

115
    prompts = [f"{question}<image>" for question in questions]
116
117
118
119
    engine_args = EngineArgs(
        model="facebook/chameleon-7b",
        max_model_len=4096,
        max_num_seqs=2,
120
        limit_mm_per_prompt={modality: 1},
121
122
123
124
125
126
    )

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
127
128


Roger Wang's avatar
Roger Wang committed
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
# Dots-OCR
def run_dots_ocr(questions: list[str], modality: str) -> ModelRequestData:
    assert modality == "image"

    prompts = [f"<|img|><|imgpad|><|endofimg|>{question}" for question in questions]
    engine_args = EngineArgs(
        model="rednote-hilab/dots.ocr",
        limit_mm_per_prompt={modality: 1},
        trust_remote_code=True,
    )

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )


146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
def run_command_a_vision(questions: list[str], modality: str) -> ModelRequestData:
    assert modality == "image"

    model_name = "CohereLabs/command-a-vision-07-2025"

    engine_args = EngineArgs(
        model=model_name,
        max_model_len=32768,
        tensor_parallel_size=4,
        limit_mm_per_prompt={modality: 1},
    )

    prompts = [
        f"<|START_OF_TURN_TOKEN|><|USER_TOKEN|><|IMG_PATCH|>{question}<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>"
        for question in questions
    ]

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )


169
# Deepseek-VL2
170
def run_deepseek_vl2(questions: list[str], modality: str) -> ModelRequestData:
171
172
    assert modality == "image"

173
    model_name = "deepseek-ai/deepseek-vl2-tiny"
174

175
176
177
178
179
    engine_args = EngineArgs(
        model=model_name,
        max_model_len=4096,
        max_num_seqs=2,
        hf_overrides={"architectures": ["DeepseekVLV2ForCausalLM"]},
180
        limit_mm_per_prompt={modality: 1},
181
    )
182

183
    prompts = [
184
        f"<|User|>: <image>\n{question}\n\n<|Assistant|>:" for question in questions
185
    ]
186
187
188
189
190

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
191
192


193
194
195
196
197
198
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
# Ernie4.5-VL
def run_ernie45_vl(questions: list[str], modality: str) -> ModelRequestData:
    model_name = "baidu/ERNIE-4.5-VL-28B-A3B-PT"

    engine_args = EngineArgs(
        model=model_name,
        max_model_len=4096,
        max_num_seqs=5,
        limit_mm_per_prompt={modality: 1},
        trust_remote_code=True,
    )

    if modality == "image":
        placeholder = "Picture 1:<|IMAGE_START|><|image@placeholder|><|IMAGE_END|>"
    elif modality == "video":
        placeholder = "Video 1:<|VIDEO_START|><|video@placeholder|><|VIDEO_END|>"

    prompts = [
        (
            f"<|begin_of_sentence|>User: {question}{placeholder}\n"
            "Assistant: <think></think>"
        )
        for question in questions
    ]

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )


224
# Fuyu
225
def run_fuyu(questions: list[str], modality: str) -> ModelRequestData:
226
227
    assert modality == "image"

228
    prompts = [f"{question}\n" for question in questions]
229
230
231
232
    engine_args = EngineArgs(
        model="adept/fuyu-8b",
        max_model_len=2048,
        max_num_seqs=2,
233
        limit_mm_per_prompt={modality: 1},
234
235
236
237
238
239
    )

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
240
241


242
# Gemma 3
243
def run_gemma3(questions: list[str], modality: str) -> ModelRequestData:
244
245
246
    assert modality == "image"
    model_name = "google/gemma-3-4b-it"

247
    engine_args = EngineArgs(
248
249
250
        model=model_name,
        max_model_len=2048,
        max_num_seqs=2,
251
252
        # TODO: Support this in transformers backend
        # mm_processor_kwargs={"do_pan_and_scan": True},
253
        limit_mm_per_prompt={modality: 1},
254
    )
255

256
257
258
259
260
261
262
263
    prompts = [
        (
            "<bos><start_of_turn>user\n"
            f"<start_of_image>{question}<end_of_turn>\n"
            "<start_of_turn>model\n"
        )
        for question in questions
    ]
Nicolò Lucchesi's avatar
Nicolò Lucchesi committed
264
265
266
267
268
    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )

269

Nicolò Lucchesi's avatar
Nicolò Lucchesi committed
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
# Gemma3N
def run_gemma3n(questions: list[str], modality: str) -> ModelRequestData:
    assert modality == "image"
    model_name = "google/gemma-3n-E2B-it"

    engine_args = EngineArgs(
        model=model_name,
        max_model_len=2048,
        max_num_seqs=2,
        limit_mm_per_prompt={modality: 1},
        enforce_eager=True,
    )

    prompts = [
        (
            "<start_of_turn>user\n"
            f"<image_soft_token>{question}<end_of_turn>\n"
            "<start_of_turn>model\n"
        )
        for question in questions
    ]
291
292
293
294
    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
295
296


297
# GLM-4v
298
def run_glm4v(questions: list[str], modality: str) -> ModelRequestData:
299
    assert modality == "image"
300
    model_name = "zai-org/glm-4v-9b"
301

302
303
304
305
306
307
308
    engine_args = EngineArgs(
        model=model_name,
        max_model_len=2048,
        max_num_seqs=2,
        trust_remote_code=True,
        enforce_eager=True,
        hf_overrides={"architectures": ["GLM4VForCausalLM"]},
309
        limit_mm_per_prompt={modality: 1},
310
    )
311

312
    prompts = [
313
314
315
316
        (
            "<|user|>\n<|begin_of_image|><|endoftext|><|end_of_image|>"
            f"{question}<|assistant|>"
        )
317
        for question in questions
318
    ]
319

320
    stop_token_ids = [151329, 151336, 151338]
321
322
323
324
325
326

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
        stop_token_ids=stop_token_ids,
    )
327
328


329
330
# GLM-4.1V
def run_glm4_1v(questions: list[str], modality: str) -> ModelRequestData:
331
    model_name = "zai-org/GLM-4.1V-9B-Thinking"
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364

    engine_args = EngineArgs(
        model=model_name,
        max_model_len=4096,
        max_num_seqs=2,
        mm_processor_kwargs={
            "size": {"shortest_edge": 12544, "longest_edge": 47040000},
            "fps": 1,
        },
        limit_mm_per_prompt={modality: 1},
        enforce_eager=True,
    )

    if modality == "image":
        placeholder = "<|begin_of_image|><|image|><|end_of_image|>"
    elif modality == "video":
        placeholder = "<|begin_of_video|><|video|><|end_of_video|>"

    prompts = [
        (
            "[gMASK]<sop><|system|>\nYou are a helpful assistant.<|user|>\n"
            f"{placeholder}"
            f"{question}<|assistant|>assistant\n"
        )
        for question in questions
    ]

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )


365
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
395
396
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
438
# GLM-4.5V
def run_glm4_5v(questions: list[str], modality: str) -> ModelRequestData:
    model_name = "zai-org/GLM-4.5V"

    engine_args = EngineArgs(
        model=model_name,
        max_model_len=4096,
        max_num_seqs=2,
        mm_processor_kwargs={
            "size": {"shortest_edge": 12544, "longest_edge": 47040000},
            "fps": 1,
        },
        limit_mm_per_prompt={modality: 1},
        enforce_eager=True,
        tensor_parallel_size=4,
    )

    if modality == "image":
        placeholder = "<|begin_of_image|><|image|><|end_of_image|>"
    elif modality == "video":
        placeholder = "<|begin_of_video|><|video|><|end_of_video|>"

    prompts = [
        (
            "[gMASK]<sop><|system|>\nYou are a helpful assistant.<|user|>\n"
            f"{placeholder}"
            f"{question}<|assistant|>assistant\n"
        )
        for question in questions
    ]

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )


# GLM-4.5V-FP8
def run_glm4_5v_fp8(questions: list[str], modality: str) -> ModelRequestData:
    model_name = "zai-org/GLM-4.5V-FP8"

    engine_args = EngineArgs(
        model=model_name,
        max_model_len=4096,
        max_num_seqs=2,
        mm_processor_kwargs={
            "size": {"shortest_edge": 12544, "longest_edge": 47040000},
            "fps": 1,
        },
        limit_mm_per_prompt={modality: 1},
        enforce_eager=True,
        tensor_parallel_size=4,
    )

    if modality == "image":
        placeholder = "<|begin_of_image|><|image|><|end_of_image|>"
    elif modality == "video":
        placeholder = "<|begin_of_video|><|video|><|end_of_video|>"

    prompts = [
        (
            "[gMASK]<sop><|system|>\nYou are a helpful assistant.<|user|>\n"
            f"{placeholder}"
            f"{question}<|assistant|>assistant\n"
        )
        for question in questions
    ]

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )


439
# H2OVL-Mississippi
440
def run_h2ovl(questions: list[str], modality: str) -> ModelRequestData:
441
442
    assert modality == "image"

443
    model_name = "h2oai/h2ovl-mississippi-800m"
444

445
    engine_args = EngineArgs(
446
447
448
        model=model_name,
        trust_remote_code=True,
        max_model_len=8192,
449
        limit_mm_per_prompt={modality: 1},
450
451
    )

452
453
454
455
456
457
458
    tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
    messages = [
        [{"role": "user", "content": f"<image>\n{question}"}] for question in questions
    ]
    prompts = tokenizer.apply_chat_template(
        messages, tokenize=False, add_generation_prompt=True
    )
459
460

    # Stop tokens for H2OVL-Mississippi
461
    # https://huggingface.co/h2oai/h2ovl-mississippi-800m
462
    stop_token_ids = [tokenizer.eos_token_id]
463
464
465
466
467
468

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
        stop_token_ids=stop_token_ids,
    )
469
470


471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
# naver-hyperclovax/HyperCLOVAX-SEED-Vision-Instruct-3B
def run_hyperclovax_seed_vision(
    questions: list[str], modality: str
) -> ModelRequestData:
    model_name = "naver-hyperclovax/HyperCLOVAX-SEED-Vision-Instruct-3B"
    tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)

    engine_args = EngineArgs(
        model=model_name,
        trust_remote_code=True,
        max_model_len=8192 if modality == "image" else 16384,
        limit_mm_per_prompt={modality: 1},
    )

    messages = list()
    for question in questions:
        if modality == "image":
            """
489
490
            ocr: List the words in the image in raster order.
                Even if the word order feels unnatural for reading,
491
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
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
                the model will handle it as long as it follows raster order.
                e.g. "Naver, CLOVA, bigshane"
            lens_keywords: List the entity names in the image.
                e.g. "iPhone"
            lens_local_keywords: List the entity names with quads in the image.
                e.g. "[0.07, 0.21, 0.92, 0.90] iPhone"
            """
            messages.append(
                [
                    {
                        "role": "user",
                        "content": [
                            {
                                "type": "image",
                                "ocr": "",
                                "lens_keywords": "",
                                "lens_local_keywords": "",
                            },
                            {
                                "type": "text",
                                "text": question,
                            },
                        ],
                    }
                ]
            )
        elif modality == "video":
            messages.append(
                [
                    {
                        "role": "user",
                        "content": [
                            {
                                "type": "video",
                            },
                            {
                                "type": "text",
                                "text": question,
                            },
                        ],
                    }
                ]
            )
        else:
            raise ValueError(f"Unsupported modality: {modality}")

    prompts = tokenizer.apply_chat_template(
        messages,
        tokenize=False,
        add_generation_prompt=True,
    )

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
        stop_token_ids=None,
    )


550
# Idefics3-8B-Llama3
551
def run_idefics3(questions: list[str], modality: str) -> ModelRequestData:
552
553
554
    assert modality == "image"
    model_name = "HuggingFaceM4/Idefics3-8B-Llama3"

555
    engine_args = EngineArgs(
556
557
558
559
560
561
562
        model=model_name,
        max_model_len=8192,
        max_num_seqs=2,
        enforce_eager=True,
        # if you are running out of memory, you can reduce the "longest_edge".
        # see: https://huggingface.co/HuggingFaceM4/Idefics3-8B-Llama3#model-optimizations
        mm_processor_kwargs={
563
            "size": {"longest_edge": 3 * 364},
564
        },
565
        limit_mm_per_prompt={modality: 1},
566
    )
567
568
569
570
    prompts = [
        (f"<|begin_of_text|>User:<image>{question}<end_of_utterance>\nAssistant:")
        for question in questions
    ]
571
572
573
574
575

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
576
577


Lyu Han's avatar
Lyu Han committed
578
579
# Intern-S1
def run_interns1(questions: list[str], modality: str) -> ModelRequestData:
580
    model_name = "internlm/Intern-S1-mini"
Lyu Han's avatar
Lyu Han committed
581
582
583
584
585
586
587
588
589
590

    engine_args = EngineArgs(
        model=model_name,
        trust_remote_code=True,
        max_model_len=8192,
        max_num_seqs=2,
        limit_mm_per_prompt={modality: 1},
        enforce_eager=True,
    )

591
592
593
594
595
    if modality == "image":
        placeholder = "<IMG_CONTEXT>"
    elif modality == "video":
        placeholder = "<video>"

Lyu Han's avatar
Lyu Han committed
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
    tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
    messages = [
        [{"role": "user", "content": f"{placeholder}\n{question}"}]
        for question in questions
    ]
    prompts = tokenizer.apply_chat_template(
        messages, tokenize=False, add_generation_prompt=True
    )

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )


611
# InternVL
612
def run_internvl(questions: list[str], modality: str) -> ModelRequestData:
613
    model_name = "OpenGVLab/InternVL3-2B"
614

615
    engine_args = EngineArgs(
616
617
        model=model_name,
        trust_remote_code=True,
618
        max_model_len=8192,
619
        limit_mm_per_prompt={modality: 1},
620
621
    )

622
623
624
625
626
    if modality == "image":
        placeholder = "<image>"
    elif modality == "video":
        placeholder = "<video>"

627
628
629
630
631
632
633
634
    tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
    messages = [
        [{"role": "user", "content": f"{placeholder}\n{question}"}]
        for question in questions
    ]
    prompts = tokenizer.apply_chat_template(
        messages, tokenize=False, add_generation_prompt=True
    )
635
636
637
638
639
640
641

    # Stop tokens for InternVL
    # models variants may have different stop tokens
    # please refer to the model card for the correct "stop words":
    # https://huggingface.co/OpenGVLab/InternVL2-2B/blob/main/conversation.py
    stop_tokens = ["<|endoftext|>", "<|im_start|>", "<|im_end|>", "<|end|>"]
    stop_token_ids = [tokenizer.convert_tokens_to_ids(i) for i in stop_tokens]
642
    stop_token_ids = [token_id for token_id in stop_token_ids if token_id is not None]
643
644
645
646
647
648

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
        stop_token_ids=stop_token_ids,
    )
649
650


651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
# Keye-VL
def run_keye_vl(questions: list[str], modality: str) -> ModelRequestData:
    model_name = "Kwai-Keye/Keye-VL-8B-Preview"

    engine_args = EngineArgs(
        model=model_name,
        max_model_len=8192,
        trust_remote_code=True,
        limit_mm_per_prompt={modality: 1},
    )

    if modality == "image":
        placeholder = "<|image_pad|>"
    elif modality == "video":
        placeholder = "<|video_pad|>"

    prompts = [
        (
            f"<|im_start|>user\n<|vision_start|>{placeholder}<|vision_end|>"
            f"{question}<|im_end|>\n"
            "<|im_start|>assistant\n"
        )
        for question in questions
    ]

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )


682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
# Keye-VL-1.5
def run_keye_vl1_5(questions: list[str], modality: str) -> ModelRequestData:
    model_name = "Kwai-Keye/Keye-VL-1.5-8B"

    engine_args = EngineArgs(
        model=model_name,
        max_model_len=8192,
        trust_remote_code=True,
        limit_mm_per_prompt={modality: 1},
    )

    if modality == "image":
        placeholder = "<|image_pad|>"
    elif modality == "video":
        placeholder = "<|video_pad|>"

    prompts = [
        (
            f"<|im_start|>user\n<|vision_start|>{placeholder}<|vision_end|>"
            f"{question}<|im_end|>\n"
            "<|im_start|>assistant\n"
        )
        for question in questions
    ]

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )


713
714
715
716
717
718
719
# Kimi-VL
def run_kimi_vl(questions: list[str], modality: str) -> ModelRequestData:
    assert modality == "image"

    prompts = [
        "<|im_user|>user<|im_middle|><|media_start|>image<|media_content|>"
        f"<|media_pad|><|media_end|>{question}<|im_end|>"
720
721
        "<|im_assistant|>assistant<|im_middle|>"
        for question in questions
722
723
724
725
726
    ]

    engine_args = EngineArgs(
        model="moonshotai/Kimi-VL-A3B-Instruct",
        trust_remote_code=True,
Cyrus Leung's avatar
Cyrus Leung committed
727
        max_model_len=4096,
728
        limit_mm_per_prompt={modality: 1},
729
730
731
732
733
734
735
736
    )

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )


737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
# LightOnOCR
def run_lightonocr(questions: list[str], modality: str) -> ModelRequestData:
    assert modality == "image"

    prompts = [
        "<|im_start|>system<|im_end|>\n<|im_start|>user\n<|image_pad|><|im_end|>\n<|im_start|>assistant\n"
        for _ in questions
    ]

    engine_args = EngineArgs(
        model="lightonai/LightOnOCR-1B",
        limit_mm_per_prompt={modality: 1},
    )

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )


757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
def run_llama4(questions: list[str], modality: str) -> ModelRequestData:
    assert modality == "image"

    model_name = "meta-llama/Llama-4-Scout-17B-16E-Instruct"

    engine_args = EngineArgs(
        model=model_name,
        max_model_len=8192,
        max_num_seqs=4,
        tensor_parallel_size=8,
        gpu_memory_utilization=0.4,
        limit_mm_per_prompt={modality: 1},
    )

    tokenizer = AutoTokenizer.from_pretrained(model_name)
    messages = [
        [
            {
                "role": "user",
                "content": [{"type": "image"}, {"type": "text", "text": f"{question}"}],
            }
        ]
        for question in questions
    ]
    prompts = tokenizer.apply_chat_template(
        messages, add_generation_prompt=True, tokenize=False
    )
    stop_token_ids = None
    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
        stop_token_ids=stop_token_ids,
    )


792
# LLaVA-1.5
793
def run_llava(questions: list[str], modality: str) -> ModelRequestData:
794
    assert modality == "image"
795

796
    prompts = [f"USER: <image>\n{question}\nASSISTANT:" for question in questions]
797

798
799
800
    engine_args = EngineArgs(
        model="llava-hf/llava-1.5-7b-hf",
        max_model_len=4096,
801
        limit_mm_per_prompt={modality: 1},
802
803
804
805
806
807
    )

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
808
809
810


# LLaVA-1.6/LLaVA-NeXT
811
def run_llava_next(questions: list[str], modality: str) -> ModelRequestData:
812
    assert modality == "image"
813

814
    prompts = [f"[INST] <image>\n{question} [/INST]" for question in questions]
815
816
817
    engine_args = EngineArgs(
        model="llava-hf/llava-v1.6-mistral-7b-hf",
        max_model_len=8192,
818
        limit_mm_per_prompt={modality: 1},
819
820
821
822
823
824
    )

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
825
826
827
828


# LlaVA-NeXT-Video
# Currently only support for video input
829
def run_llava_next_video(questions: list[str], modality: str) -> ModelRequestData:
830
831
    assert modality == "video"

832
    prompts = [f"USER: <video>\n{question} ASSISTANT:" for question in questions]
833
834
835
    engine_args = EngineArgs(
        model="llava-hf/LLaVA-NeXT-Video-7B-hf",
        max_model_len=8192,
836
        max_num_seqs=2,
837
        limit_mm_per_prompt={modality: 1},
838
839
840
841
842
843
    )

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
844
845


846
# LLaVA-OneVision
847
def run_llava_onevision(questions: list[str], modality: str) -> ModelRequestData:
848
    if modality == "video":
849
        prompts = [
850
            f"<|im_start|>user <video>\n{question}<|im_end|><|im_start|>assistant\n"
851
            for question in questions
852
        ]
853
854

    elif modality == "image":
855
        prompts = [
856
            f"<|im_start|>user <image>\n{question}<|im_end|><|im_start|>assistant\n"
857
            for question in questions
858
        ]
859

860
861
862
    engine_args = EngineArgs(
        model="llava-hf/llava-onevision-qwen2-7b-ov-hf",
        max_model_len=16384,
863
        limit_mm_per_prompt={modality: 1},
864
865
866
867
868
869
    )

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
870
871


872
# Mantis
873
def run_mantis(questions: list[str], modality: str) -> ModelRequestData:
874
    assert modality == "image"
875

876
877
    llama3_template = "<|start_header_id|>user<|end_header_id|>\n\n{}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"  # noqa: E501
    prompts = [llama3_template.format(f"{question}\n<image>") for question in questions]
878

879
    engine_args = EngineArgs(
880
        model="TIGER-Lab/Mantis-8B-siglip-llama3",
881
        max_model_len=4096,
882
        hf_overrides={"architectures": ["MantisForConditionalGeneration"]},
883
        limit_mm_per_prompt={modality: 1},
884
    )
885
    stop_token_ids = [128009]
886
887
888
889
890
891

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
        stop_token_ids=stop_token_ids,
    )
892
893
894


# MiniCPM-V
895
def run_minicpmv_base(questions: list[str], modality: str, model_name):
896
897
    assert modality in ["image", "video"]
    # If you want to use `MiniCPM-o-2_6` with audio inputs, check `audio_language.py` # noqa
898
899
900
901
902
903
904

    # 2.0
    # The official repo doesn't work yet, so we need to use a fork for now
    # For more details, please see: See: https://github.com/vllm-project/vllm/pull/4087#issuecomment-2250397630 # noqa
    # model_name = "HwwwH/MiniCPM-V-2"

    # 2.5
905
906
    # model_name = "openbmb/MiniCPM-Llama3-V-2_5"

907
    # 2.6
908
909
910
911
912
913
914
915
916
    # model_name = "openbmb/MiniCPM-V-2_6"
    # o2.6

    # modality supports
    # 2.0: image
    # 2.5: image
    # 2.6: image, video
    # o2.6: image, video, audio
    # model_name = "openbmb/MiniCPM-o-2_6"
917
    tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
918
    engine_args = EngineArgs(
919
        model=model_name,
920
921
        max_model_len=4096,
        max_num_seqs=2,
922
        trust_remote_code=True,
923
        limit_mm_per_prompt={modality: 1},
924
    )
925
926
927
928
929
930
931
    # NOTE The stop_token_ids are different for various versions of MiniCPM-V
    # 2.0
    # stop_token_ids = [tokenizer.eos_id]

    # 2.5
    # stop_token_ids = [tokenizer.eos_id, tokenizer.eot_id]

932
    # 2.6 / o2.6
933
    stop_tokens = ["<|im_end|>", "<|endoftext|>"]
934
    stop_token_ids = [tokenizer.convert_tokens_to_ids(i) for i in stop_tokens]
935

936
937
938
939
940
    modality_placeholder = {
        "image": "(<image>./</image>)",
        "video": "(<video>./</video>)",
    }

941
942
    prompts = [
        tokenizer.apply_chat_template(
943
944
945
946
947
948
            [
                {
                    "role": "user",
                    "content": f"{modality_placeholder[modality]}\n{question}",
                }
            ],
949
            tokenize=False,
950
951
952
            add_generation_prompt=True,
        )
        for question in questions
953
    ]
954
955
956
957
958
959

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
        stop_token_ids=stop_token_ids,
    )
960
961


962
def run_minicpmo(questions: list[str], modality: str) -> ModelRequestData:
963
    return run_minicpmv_base(questions, modality, "openbmb/MiniCPM-o-2_6")
964
965


966
def run_minicpmv(questions: list[str], modality: str) -> ModelRequestData:
967
    return run_minicpmv_base(questions, modality, "openbmb/MiniCPM-V-2_6")
968
969


970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
def run_minimax_vl_01(questions: list[str], modality: str) -> ModelRequestData:
    assert modality == "image"

    model_name = "MiniMaxAI/MiniMax-VL-01"

    engine_args = EngineArgs(
        model=model_name,
        max_num_seqs=2,
        limit_mm_per_prompt={modality: 1},
        trust_remote_code=True,
        tensor_parallel_size=8,
    )

    tokenizer = AutoTokenizer.from_pretrained(model_name)
    messages = [
        [
            {
                "role": "user",
                "content": [{"type": "image"}, {"type": "text", "text": question}],
            }
        ]
        for question in questions
    ]
    prompts = tokenizer.apply_chat_template(
        messages, add_generation_prompt=True, tokenize=False
    )

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )


1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
# Mistral-3 HF-format
def run_mistral3(questions: list[str], modality: str) -> ModelRequestData:
    assert modality == "image"

    model_name = "mistralai/Mistral-Small-3.1-24B-Instruct-2503"

    # NOTE: Need L40 (or equivalent) to avoid OOM
    engine_args = EngineArgs(
        model=model_name,
        max_model_len=8192,
        max_num_seqs=2,
        tensor_parallel_size=2,
1015
        limit_mm_per_prompt={modality: 1},
1016
        ignore_patterns=["consolidated.safetensors"],
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
    )

    prompts = [f"<s>[INST]{question}\n[IMG][/INST]" for question in questions]

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )


1027
1028
# Molmo
def run_molmo(questions: list[str], modality: str) -> ModelRequestData:
1029
1030
    assert modality == "image"

1031
    model_name = "allenai/Molmo-7B-D-0924"
1032
1033
1034

    engine_args = EngineArgs(
        model=model_name,
1035
1036
        trust_remote_code=True,
        dtype="bfloat16",
1037
        limit_mm_per_prompt={modality: 1},
1038
1039
    )

1040
    prompts = [
1041
        f"<|im_start|>user <image>\n{question}<|im_end|><|im_start|>assistant\n"
1042
1043
        for question in questions
    ]
1044

1045
1046
1047
1048
1049
1050
    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )


1051
1052
1053
# Nemontron_VL
def run_nemotron_vl(questions: list[str], modality: str) -> ModelRequestData:
    model_name = "nvidia/Llama-3.1-Nemotron-Nano-VL-8B-V1"
1054

1055
    engine_args = EngineArgs(
1056
        model=model_name,
1057
        trust_remote_code=True,
1058
        max_model_len=8192,
1059
        limit_mm_per_prompt={modality: 1},
1060
    )
1061

1062
1063
1064
1065
1066
1067
    assert modality == "image"
    placeholder = "<image>"

    tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
    messages = [
        [{"role": "user", "content": f"{placeholder}\n{question}"}]
1068
        for question in questions
1069
    ]
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
    prompts = tokenizer.apply_chat_template(
        messages, tokenize=False, add_generation_prompt=True
    )

    # Stop tokens for InternVL
    # models variants may have different stop tokens
    # please refer to the model card for the correct "stop words":
    # https://huggingface.co/OpenGVLab/InternVL2-2B/blob/main/conversation.py
    stop_tokens = ["<|endoftext|>", "<|im_start|>", "<|im_end|>", "<|end|>"]
    stop_token_ids = [tokenizer.convert_tokens_to_ids(i) for i in stop_tokens]
    stop_token_ids = [token_id for token_id in stop_token_ids if token_id is not None]
1081
1082
1083
1084

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
1085
        stop_token_ids=stop_token_ids,
1086
    )
1087
1088


1089
# NVLM-D
1090
def run_nvlm_d(questions: list[str], modality: str) -> ModelRequestData:
1091
1092
1093
1094
1095
    assert modality == "image"

    model_name = "nvidia/NVLM-D-72B"

    # Adjust this as necessary to fit in GPU
1096
    engine_args = EngineArgs(
1097
1098
1099
1100
        model=model_name,
        trust_remote_code=True,
        max_model_len=4096,
        tensor_parallel_size=4,
1101
        limit_mm_per_prompt={modality: 1},
1102
1103
    )

1104
1105
1106
1107
1108
1109
1110
    tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
    messages = [
        [{"role": "user", "content": f"<image>\n{question}"}] for question in questions
    ]
    prompts = tokenizer.apply_chat_template(
        messages, tokenize=False, add_generation_prompt=True
    )
1111
1112
1113
1114
1115

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
1116
1117


1118
1119
# Ovis
def run_ovis(questions: list[str], modality: str) -> ModelRequestData:
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
    assert modality == "image"

    model_name = "AIDC-AI/Ovis2-1B"

    engine_args = EngineArgs(
        model=model_name,
        max_model_len=4096,
        max_num_seqs=2,
        trust_remote_code=True,
        dtype="half",
1130
        limit_mm_per_prompt={modality: 1},
1131
1132
    )

1133
1134
1135
1136
1137
1138
1139
    tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
    messages = [
        [{"role": "user", "content": f"<image>\n{question}"}] for question in questions
    ]
    prompts = tokenizer.apply_chat_template(
        messages, tokenize=False, add_generation_prompt=True
    )
1140
1141
1142
1143
1144
1145
1146

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )


1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
# Ovis2_5
def run_ovis2_5(questions: list[str], modality: str) -> ModelRequestData:
    model_name = "AIDC-AI/Ovis2.5-2B"

    engine_args = EngineArgs(
        model=model_name,
        max_model_len=4096,
        max_num_seqs=2,
        trust_remote_code=True,
        dtype="half",
        limit_mm_per_prompt={modality: 1},
    )
    if modality == "image":
        placeholder = "<image>"
    elif modality == "video":
        placeholder = "<video>"

1164
1165
    prompts = [
        f"<|im_start|>user\n\n{placeholder}\n{question}<|im_end|>\n<|im_start|>assistant\n"
1166
1167
1168
1169
1170
1171
1172
1173
1174
        for question in questions
    ]

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )


1175
# PaliGemma
1176
def run_paligemma(questions: list[str], modality: str) -> ModelRequestData:
1177
    assert modality == "image"
1178

1179
    # PaliGemma has special prompt format for VQA
1180
1181
1182
    prompts = ["caption en" for _ in questions]
    engine_args = EngineArgs(
        model="google/paligemma-3b-mix-224",
1183
        limit_mm_per_prompt={modality: 1},
1184
    )
1185
1186
1187
1188
1189

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
1190
1191


1192
# PaliGemma 2
1193
def run_paligemma2(questions: list[str], modality: str) -> ModelRequestData:
1194
    assert modality == "image"
1195

1196
    # PaliGemma 2 has special prompt format for VQA
1197
1198
1199
    prompts = ["caption en" for _ in questions]
    engine_args = EngineArgs(
        model="google/paligemma2-3b-ft-docci-448",
1200
        limit_mm_per_prompt={modality: 1},
1201
    )
1202
1203
1204
1205
1206

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
1207
1208


1209
# Phi-3-Vision
1210
def run_phi3v(questions: list[str], modality: str) -> ModelRequestData:
1211
1212
    assert modality == "image"

1213
1214
1215
1216
    prompts = [
        f"<|user|>\n<|image_1|>\n{question}<|end|>\n<|assistant|>\n"
        for question in questions
    ]
1217

1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
    # num_crops is an override kwarg to the multimodal image processor;
    # For some models, e.g., Phi-3.5-vision-instruct, it is recommended
    # to use 16 for single frame scenarios, and 4 for multi-frame.
    #
    # Generally speaking, a larger value for num_crops results in more
    # tokens per image instance, because it may scale the image more in
    # the image preprocessing. Some references in the model docs and the
    # formula for image tokens after the preprocessing
    # transform can be found below.
    #
    # https://huggingface.co/microsoft/Phi-3.5-vision-instruct#loading-the-model-locally
    # https://huggingface.co/microsoft/Phi-3.5-vision-instruct/blob/main/processing_phi3_v.py#L194
1230
    engine_args = EngineArgs(
1231
1232
        model="microsoft/Phi-3.5-vision-instruct",
        trust_remote_code=True,
1233
        max_model_len=4096,
1234
        max_num_seqs=2,
1235
        # Note - mm_processor_kwargs can also be passed to generate/chat calls
1236
        mm_processor_kwargs={"num_crops": 16},
1237
        limit_mm_per_prompt={modality: 1},
1238
    )
1239
1240
1241
1242
1243

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
1244
1245


1246
# Phi-4-multimodal-instruct
1247
def run_phi4mm(questions: list[str], modality: str) -> ModelRequestData:
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
    """
    Phi-4-multimodal-instruct supports both image and audio inputs. Here, we
    show how to process image inputs.
    """
    assert modality == "image"
    model_path = snapshot_download("microsoft/Phi-4-multimodal-instruct")
    # Since the vision-lora and speech-lora co-exist with the base model,
    # we have to manually specify the path of the lora weights.
    vision_lora_path = os.path.join(model_path, "vision-lora")
    prompts = [
1258
        f"<|user|><|image_1|>{question}<|end|><|assistant|>" for question in questions
1259
    ]
1260
    engine_args = EngineArgs(
1261
1262
        model=model_path,
        trust_remote_code=True,
1263
        max_model_len=5120,
1264
        max_num_seqs=2,
1265
        max_num_batched_tokens=12800,
1266
1267
        enable_lora=True,
        max_lora_rank=320,
1268
1269
        # Note - mm_processor_kwargs can also be passed to generate/chat calls
        mm_processor_kwargs={"dynamic_hd": 16},
1270
        limit_mm_per_prompt={modality: 1},
1271
1272
    )

1273
1274
1275
1276
1277
    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
        lora_requests=[LoRARequest("vision", 1, vision_lora_path)],
    )
1278
1279


1280
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
1314
# HF format Phi-4-multimodal-instruct
def run_phi4_multimodal(questions: list[str], modality: str) -> ModelRequestData:
    """
    Phi-4-multimodal-instruct supports both image and audio inputs. Here, we
    show how to process image inputs.
    """
    assert modality == "image"
    model_path = snapshot_download(
        "microsoft/Phi-4-multimodal-instruct", revision="refs/pr/70"
    )
    # Since the vision-lora and speech-lora co-exist with the base model,
    # we have to manually specify the path of the lora weights.
    vision_lora_path = os.path.join(model_path, "vision-lora")
    prompts = [
        f"<|user|><|image|>{question}<|end|><|assistant|>" for question in questions
    ]
    engine_args = EngineArgs(
        model=model_path,
        max_model_len=5120,
        max_num_seqs=2,
        max_num_batched_tokens=12800,
        enable_lora=True,
        max_lora_rank=320,
        # Note - mm_processor_kwargs can also be passed to generate/chat calls
        mm_processor_kwargs={"dynamic_hd": 16},
        limit_mm_per_prompt={"image": 1},
    )

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
        lora_requests=[LoRARequest("vision", 1, vision_lora_path)],
    )


1315
# Pixtral HF-format
1316
def run_pixtral_hf(questions: list[str], modality: str) -> ModelRequestData:
1317
1318
1319
1320
    assert modality == "image"

    model_name = "mistral-community/pixtral-12b"

1321
    # NOTE: Need L40 (or equivalent) to avoid OOM
1322
    engine_args = EngineArgs(
1323
        model=model_name,
1324
        max_model_len=6144,
1325
        max_num_seqs=2,
1326
        limit_mm_per_prompt={modality: 1},
1327
1328
    )

1329
    prompts = [f"<s>[INST]{question}\n[IMG][/INST]" for question in questions]
1330
1331
1332
1333
1334

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
1335
1336


1337
# Qwen-VL
1338
def run_qwen_vl(questions: list[str], modality: str) -> ModelRequestData:
1339
1340
    assert modality == "image"

1341
    engine_args = EngineArgs(
1342
        model="Qwen/Qwen-VL",
1343
        trust_remote_code=True,
1344
1345
        max_model_len=1024,
        max_num_seqs=2,
1346
        hf_overrides={"architectures": ["QwenVLForConditionalGeneration"]},
1347
        limit_mm_per_prompt={modality: 1},
1348
1349
    )

1350
    prompts = [f"{question}Picture 1: <img></img>\n" for question in questions]
1351
1352
1353
1354
1355

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
1356
1357


1358
# Qwen2-VL
1359
def run_qwen2_vl(questions: list[str], modality: str) -> ModelRequestData:
1360
    model_name = "Qwen/Qwen2-VL-7B-Instruct"
1361

1362
    engine_args = EngineArgs(
1363
        model=model_name,
1364
1365
1366
        max_model_len=4096,
        max_num_seqs=5,
        # Note - mm_processor_kwargs can also be passed to generate/chat calls
1367
        mm_processor_kwargs={
1368
1369
            "min_pixels": 28 * 28,
            "max_pixels": 1280 * 28 * 28,
1370
        },
1371
        limit_mm_per_prompt={modality: 1},
1372
    )
1373

1374
1375
1376
1377
1378
    if modality == "image":
        placeholder = "<|image_pad|>"
    elif modality == "video":
        placeholder = "<|video_pad|>"

1379
    prompts = [
1380
1381
1382
1383
1384
1385
1386
        (
            "<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n"
            f"<|im_start|>user\n<|vision_start|>{placeholder}<|vision_end|>"
            f"{question}<|im_end|>\n"
            "<|im_start|>assistant\n"
        )
        for question in questions
1387
    ]
1388
1389
1390
1391
1392

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
1393
1394


Roger Wang's avatar
Roger Wang committed
1395
# Qwen2.5-VL
1396
def run_qwen2_5_vl(questions: list[str], modality: str) -> ModelRequestData:
Roger Wang's avatar
Roger Wang committed
1397
1398
    model_name = "Qwen/Qwen2.5-VL-3B-Instruct"

1399
    engine_args = EngineArgs(
Roger Wang's avatar
Roger Wang committed
1400
1401
1402
1403
1404
1405
1406
1407
        model=model_name,
        max_model_len=4096,
        max_num_seqs=5,
        mm_processor_kwargs={
            "min_pixels": 28 * 28,
            "max_pixels": 1280 * 28 * 28,
            "fps": 1,
        },
1408
        limit_mm_per_prompt={modality: 1},
Roger Wang's avatar
Roger Wang committed
1409
1410
1411
1412
1413
1414
1415
    )

    if modality == "image":
        placeholder = "<|image_pad|>"
    elif modality == "video":
        placeholder = "<|video_pad|>"

1416
    prompts = [
1417
1418
1419
1420
1421
1422
1423
        (
            "<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n"
            f"<|im_start|>user\n<|vision_start|>{placeholder}<|vision_end|>"
            f"{question}<|im_end|>\n"
            "<|im_start|>assistant\n"
        )
        for question in questions
1424
    ]
1425
1426
1427
1428
1429

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
Roger Wang's avatar
Roger Wang committed
1430
1431


1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
# Qwen2.5-Omni
def run_qwen2_5_omni(questions: list[str], modality: str):
    model_name = "Qwen/Qwen2.5-Omni-7B"

    engine_args = EngineArgs(
        model=model_name,
        max_model_len=4096,
        max_num_seqs=5,
        mm_processor_kwargs={
            "min_pixels": 28 * 28,
            "max_pixels": 1280 * 28 * 28,
            "fps": [1],
        },
1445
        limit_mm_per_prompt={modality: 1},
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
    )

    if modality == "image":
        placeholder = "<|IMAGE|>"
    elif modality == "video":
        placeholder = "<|VIDEO|>"

    default_system = (
        "You are Qwen, a virtual human developed by the Qwen Team, Alibaba "
        "Group, capable of perceiving auditory and visual inputs, as well as "
1456
1457
        "generating text and speech."
    )
1458

1459
1460
1461
1462
1463
1464
1465
1466
1467
    prompts = [
        (
            f"<|im_start|>system\n{default_system}<|im_end|>\n"
            f"<|im_start|>user\n<|vision_bos|>{placeholder}<|vision_eos|>"
            f"{question}<|im_end|>\n"
            "<|im_start|>assistant\n"
        )
        for question in questions
    ]
1468
1469
1470
1471
1472
1473
    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )


1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
# Qwen3-VL-Dense
def run_qwen3_vl(questions: list[str], modality: str) -> ModelRequestData:
    model_name = "Qwen/Qwen3-VL-4B-Instruct"

    engine_args = EngineArgs(
        model=model_name,
        max_model_len=4096,
        max_num_seqs=5,
        mm_processor_kwargs={
            "min_pixels": 28 * 28,
            "max_pixels": 1280 * 28 * 28,
            "fps": 1,
        },
        limit_mm_per_prompt={modality: 1},
    )

    if modality == "image":
        placeholder = "<|image_pad|>"
    elif modality == "video":
        placeholder = "<|video_pad|>"

    prompts = [
        (
            "<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n"
            f"<|im_start|>user\n<|vision_start|>{placeholder}<|vision_end|>"
            f"{question}<|im_end|>\n"
            "<|im_start|>assistant\n"
        )
        for question in questions
    ]

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )


# Qwen3-VL-MOE
def run_qwen3_vl_moe(questions: list[str], modality: str) -> ModelRequestData:
    model_name = "Qwen/Qwen3-VL-30B-A3B-Instruct"

    engine_args = EngineArgs(
        model=model_name,
        max_model_len=4096,
        max_num_seqs=5,
        mm_processor_kwargs={
            "min_pixels": 28 * 28,
            "max_pixels": 1280 * 28 * 28,
            "fps": 1,
        },
        limit_mm_per_prompt={modality: 1},
    )

    if modality == "image":
        placeholder = "<|image_pad|>"
    elif modality == "video":
        placeholder = "<|video_pad|>"

    prompts = [
        (
            "<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n"
            f"<|im_start|>user\n<|vision_start|>{placeholder}<|vision_end|>"
            f"{question}<|im_end|>\n"
            "<|im_start|>assistant\n"
        )
        for question in questions
    ]

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )


1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
# R-4B
def run_r_vl(questions: list[str], modality: str) -> ModelRequestData:
    assert modality == "image"
    model_name = "YannQi/R-4B"

    prompts = [
        f"<|im_start|>user <image>\n{question}<|im_end|><|im_start|>assistant\n"
        for question in questions
    ]

    engine_args = EngineArgs(
        model=model_name,
        max_model_len=16384,
        limit_mm_per_prompt={modality: 1},
    )

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )


1570
1571
# SkyworkR1V
def run_skyworkr1v(questions: list[str], modality: str) -> ModelRequestData:
1572
    assert modality == "image"
1573
1574

    model_name = "Skywork/Skywork-R1V-38B"
1575
1576
1577
1578
1579
1580
1581

    engine_args = EngineArgs(
        model=model_name,
        trust_remote_code=True,
        max_model_len=4096,
        limit_mm_per_prompt={modality: 1},
    )
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594

    tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
    messages = [
        [{"role": "user", "content": f"<image>\n{question}"}] for question in questions
    ]
    prompts = tokenizer.apply_chat_template(
        messages, tokenize=False, add_generation_prompt=True
    )

    # Stop tokens for SkyworkR1V
    # https://huggingface.co/Skywork/Skywork-R1V-38B/blob/main/conversation.py
    stop_tokens = ["<|end▁of▁sentence|>", "<|endoftext|>"]
    stop_token_ids = [tokenizer.convert_tokens_to_ids(i) for i in stop_tokens]
1595
1596
1597
1598

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
1599
        stop_token_ids=stop_token_ids,
1600
1601
1602
    )


1603
1604
1605
1606
# SmolVLM2-2.2B-Instruct
def run_smolvlm(questions: list[str], modality: str) -> ModelRequestData:
    assert modality == "image"
    model_name = "HuggingFaceTB/SmolVLM2-2.2B-Instruct"
1607
1608
1609

    engine_args = EngineArgs(
        model=model_name,
1610
1611
1612
1613
1614
1615
        max_model_len=8192,
        max_num_seqs=2,
        enforce_eager=True,
        mm_processor_kwargs={
            "max_image_size": {"longest_edge": 384},
        },
1616
1617
        limit_mm_per_prompt={modality: 1},
    )
1618
1619
1620
1621
    prompts = [
        (f"<|im_start|>User:<image>{question}<end_of_utterance>\nAssistant:")
        for question in questions
    ]
1622

1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
1644
    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )


# Step3
def run_step3(questions: list[str], modality: str) -> ModelRequestData:
    assert modality == "image"

    model_name = "stepfun-ai/step3-fp8"

    # NOTE: Below are verified configurations for step3-fp8
    # on 8xH100 GPUs.
    engine_args = EngineArgs(
        model=model_name,
        max_num_batched_tokens=4096,
        gpu_memory_utilization=0.85,
        tensor_parallel_size=8,
        limit_mm_per_prompt={modality: 1},
        reasoning_parser="step3",
    )
1645
1646

    prompts = [
1647
1648
        "<|begin▁of▁sentence|> You are a helpful assistant. <|BOT|>user\n "
        f"<im_patch>{question} <|EOT|><|BOT|>assistant\n<think>\n"
1649
1650
1651
1652
1653
1654
1655
1656
1657
        for question in questions
    ]

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )


1658
1659
# omni-research/Tarsier-7b
def run_tarsier(questions: list[str], modality: str) -> ModelRequestData:
1660
    assert modality == "image"
1661
    model_name = "omni-research/Tarsier-7b"
1662
1663
1664
1665
1666

    engine_args = EngineArgs(
        model=model_name,
        trust_remote_code=True,
        max_model_len=4096,
1667
        limit_mm_per_prompt={modality: 1},
1668
    )
1669
    prompts = [(f"USER: <image>\n{question} ASSISTANT:") for question in questions]
1670

1671
1672
1673
    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
1674
    )
1675

1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
1690
1691
1692
1693
1694
1695
1696
1697
1698
1699
1700

def run_tarsier2(questions: list[str], modality: str) -> ModelRequestData:
    model_name = "omni-research/Tarsier2-Recap-7b"

    engine_args = EngineArgs(
        model=model_name,
        max_model_len=4096,
        hf_overrides={"architectures": ["Tarsier2ForConditionalGeneration"]},
        limit_mm_per_prompt={modality: 1},
    )

    if modality == "image":
        placeholder = "<|image_pad|>"
    elif modality == "video":
        placeholder = "<|video_pad|>"

    prompts = [
        (
            "<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n"
            f"<|im_start|>user\n<|vision_start|>{placeholder}<|vision_end|>"
            f"{question}<|im_end|>\n"
            "<|im_start|>assistant\n"
        )
        for question in questions
    ]
1701
1702
1703
1704
1705
1706
1707

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )


1708
model_example_map = {
1709
    "aria": run_aria,
Jennifer Zhao's avatar
Jennifer Zhao committed
1710
    "aya_vision": run_aya_vision,
1711
1712
    "blip-2": run_blip2,
    "chameleon": run_chameleon,
Roger Wang's avatar
Roger Wang committed
1713
    "dots_ocr": run_dots_ocr,
1714
    "command_a_vision": run_command_a_vision,
1715
    "deepseek_vl_v2": run_deepseek_vl2,
1716
    "ernie45_vl": run_ernie45_vl,
1717
    "fuyu": run_fuyu,
1718
    "gemma3": run_gemma3,
Nicolò Lucchesi's avatar
Nicolò Lucchesi committed
1719
    "gemma3n": run_gemma3n,
1720
    "glm4v": run_glm4v,
1721
    "glm4_1v": run_glm4_1v,
1722
1723
    "glm4_5v": run_glm4_5v,
    "glm4_5v_fp8": run_glm4_5v_fp8,
1724
    "h2ovl_chat": run_h2ovl,
1725
    "hyperclovax_seed_vision": run_hyperclovax_seed_vision,
1726
    "idefics3": run_idefics3,
Lyu Han's avatar
Lyu Han committed
1727
    "interns1": run_interns1,
1728
    "internvl_chat": run_internvl,
1729
    "keye_vl": run_keye_vl,
1730
    "keye_vl1_5": run_keye_vl1_5,
1731
    "kimi_vl": run_kimi_vl,
1732
    "lightonocr": run_lightonocr,
1733
    "llama4": run_llama4,
1734
1735
    "llava": run_llava,
    "llava-next": run_llava_next,
1736
    "llava-next-video": run_llava_next_video,
1737
    "llava-onevision": run_llava_onevision,
1738
    "mantis": run_mantis,
1739
    "minicpmo": run_minicpmo,
1740
    "minicpmv": run_minicpmv,
1741
    "minimax_vl_01": run_minimax_vl_01,
1742
    "mistral3": run_mistral3,
1743
    "molmo": run_molmo,
1744
    "nemotron_vl": run_nemotron_vl,
1745
    "NVLM_D": run_nvlm_d,
1746
    "ovis": run_ovis,
1747
    "ovis2_5": run_ovis2_5,
1748
1749
1750
    "paligemma": run_paligemma,
    "paligemma2": run_paligemma2,
    "phi3_v": run_phi3v,
1751
    "phi4_mm": run_phi4mm,
1752
    "phi4_multimodal": run_phi4_multimodal,
1753
    "pixtral_hf": run_pixtral_hf,
1754
    "qwen_vl": run_qwen_vl,
1755
    "qwen2_vl": run_qwen2_vl,
Roger Wang's avatar
Roger Wang committed
1756
    "qwen2_5_vl": run_qwen2_5_vl,
1757
    "qwen2_5_omni": run_qwen2_5_omni,
1758
1759
    "qwen3_vl": run_qwen3_vl,
    "qwen3_vl_moe": run_qwen3_vl_moe,
1760
    "rvl": run_r_vl,
1761
    "skywork_chat": run_skyworkr1v,
1762
    "smolvlm": run_smolvlm,
1763
    "step3": run_step3,
汪志鹏's avatar
汪志鹏 committed
1764
    "tarsier": run_tarsier,
1765
    "tarsier2": run_tarsier2,
1766
1767
1768
}


1769
1770
1771
1772
MODELS_NEED_VIDEO_METADATA = [
    "glm4_1v",
    "glm4_5v",
    "glm4_5v_fp8",
1773
1774
    "qwen3_vl",
    "qwen3_vl_moe",
1775
1776
1777
]


1778
1779
1780
1781
1782
1783
1784
1785
1786
def get_multi_modal_input(args):
    """
    return {
        "data": image or video,
        "question": question,
    }
    """
    if args.modality == "image":
        # Input image and question
1787
        image = convert_image_mode(ImageAsset("cherry_blossom").pil_image, "RGB")
1788
1789
1790
1791
1792
1793
        img_questions = [
            "What is the content of this image?",
            "Describe the content of this image in detail.",
            "What's in the image?",
            "Where is this image taken?",
        ]
1794
1795
1796

        return {
            "data": image,
1797
            "questions": img_questions,
1798
1799
1800
1801
        }

    if args.modality == "video":
        # Input video and question
1802
        needs_metadata = args.model_type in MODELS_NEED_VIDEO_METADATA
1803
        video = VideoAsset(name="baby_reading", num_frames=args.num_frames).np_ndarrays
1804
        metadata = VideoAsset(name="baby_reading", num_frames=args.num_frames).metadata
1805
        vid_questions = ["Why is this video funny?"]
1806
1807

        return {
1808
            "data": ([(video, metadata)] if needs_metadata else video),
1809
            "questions": vid_questions,
1810
1811
1812
1813
1814
1815
        }

    msg = f"Modality {args.modality} is not supported."
    raise ValueError(msg)


1816
1817
1818
1819
def apply_image_repeat(
    image_repeat_prob, num_prompts, data, prompts: list[str], modality
):
    """Repeats images with provided probability of "image_repeat_prob".
1820
1821
    Used to simulate hit/miss for the MM preprocessor cache.
    """
1822
    assert image_repeat_prob <= 1.0 and image_repeat_prob >= 0
1823
1824
1825
1826
    no_yes = [0, 1]
    probs = [1.0 - image_repeat_prob, image_repeat_prob]

    inputs = []
1827
    inputs_with_empty_media = []
1828
1829
1830
1831
1832
1833
1834
1835
1836
1837
    cur_image = data
    for i in range(num_prompts):
        if image_repeat_prob is not None:
            res = random.choices(no_yes, probs)[0]
            if res == 0:
                # No repeat => Modify one pixel
                cur_image = cur_image.copy()
                new_val = (i // 256 // 256, i // 256, i % 256)
                cur_image.putpixel((0, 0), new_val)

1838
1839
        uuid = "uuid_{}".format(i)

1840
1841
1842
1843
        inputs.append(
            {
                "prompt": prompts[i % len(prompts)],
                "multi_modal_data": {modality: cur_image},
1844
1845
1846
1847
1848
1849
1850
1851
1852
                "multi_modal_uuids": {modality: uuid},
            }
        )

        inputs_with_empty_media.append(
            {
                "prompt": prompts[i % len(prompts)],
                "multi_modal_data": {modality: None},
                "multi_modal_uuids": {modality: uuid},
1853
            }
1854
        )
1855

1856
    return inputs, inputs_with_empty_media
1857
1858


1859
1860
1861
1862
@contextmanager
def time_counter(enable: bool):
    if enable:
        import time
1863

1864
1865
1866
1867
1868
1869
1870
1871
1872
1873
        start_time = time.time()
        yield
        elapsed_time = time.time() - start_time
        print("-" * 50)
        print("-- generate time = {}".format(elapsed_time))
        print("-" * 50)
    else:
        yield


1874
1875
def parse_args():
    parser = FlexibleArgumentParser(
1876
1877
1878
1879
1880
1881
1882
1883
1884
1885
1886
1887
1888
1889
1890
1891
1892
1893
1894
1895
1896
1897
1898
1899
1900
1901
1902
1903
1904
1905
1906
1907
1908
        description="Demo on using vLLM for offline inference with "
        "vision language models for text generation"
    )
    parser.add_argument(
        "--model-type",
        "-m",
        type=str,
        default="llava",
        choices=model_example_map.keys(),
        help='Huggingface "model_type".',
    )
    parser.add_argument(
        "--num-prompts", type=int, default=4, help="Number of prompts to run."
    )
    parser.add_argument(
        "--modality",
        type=str,
        default="image",
        choices=["image", "video"],
        help="Modality of the input.",
    )
    parser.add_argument(
        "--num-frames",
        type=int,
        default=16,
        help="Number of frames to extract from the video.",
    )
    parser.add_argument(
        "--seed",
        type=int,
        default=None,
        help="Set the seed when initializing `vllm.LLM`.",
    )
1909
1910

    parser.add_argument(
1911
        "--image-repeat-prob",
1912
1913
        type=float,
        default=None,
1914
1915
        help="Simulates the hit-ratio for multi-modal preprocessor cache (if enabled)",
    )
1916
1917

    parser.add_argument(
1918
        "--disable-mm-processor-cache",
1919
        action="store_true",
1920
        help="If True, disables caching of multi-modal processor.",
1921
    )
1922
1923

    parser.add_argument(
1924
1925
1926
1927
        "--time-generate",
        action="store_true",
        help="If True, then print the total generate() call time",
    )
1928
1929

    parser.add_argument(
1930
1931
1932
1933
1934
        "--use-different-prompt-per-request",
        action="store_true",
        help="If True, then use different prompt (with the same multi-modal "
        "data) for each request.",
    )
1935
1936
1937
1938
1939
1940
1941

    parser.add_argument(
        "--verify-mm-cache-hit-with-uuids",
        action="store_true",
        help="If True, will send all requests in a second batch with empty mm "
        "data to verify cache hits with UUIDs.",
    )
1942
1943
1944
    return parser.parse_args()


1945
1946
1947
1948
1949
def main(args):
    model = args.model_type
    if model not in model_example_map:
        raise ValueError(f"Model type {model} is not supported.")

1950
1951
1952
    modality = args.modality
    mm_input = get_multi_modal_input(args)
    data = mm_input["data"]
1953
    questions = mm_input["questions"]
1954

1955
1956
    req_data = model_example_map[model](questions, modality)

1957
1958
1959
    # Disable other modalities to save memory
    default_limits = {"image": 0, "video": 0, "audio": 0}
    req_data.engine_args.limit_mm_per_prompt = default_limits | dict(
1960
1961
        req_data.engine_args.limit_mm_per_prompt or {}
    )
1962
1963
1964

    engine_args = asdict(req_data.engine_args) | {
        "seed": args.seed,
1965
        "mm_processor_cache_gb": 0 if args.disable_mm_processor_cache else 4,
1966
    }
1967
1968
    llm = LLM(**engine_args)

1969
    # Don't want to check the flag multiple times, so just hijack `prompts`.
1970
1971
1972
1973
1974
    prompts = (
        req_data.prompts
        if args.use_different_prompt_per_request
        else [req_data.prompts[0]]
    )
1975
1976
1977

    # We set temperature to 0.2 so that outputs can be different
    # even when all prompts are identical when running batch inference.
1978
1979
1980
    sampling_params = SamplingParams(
        temperature=0.2, max_tokens=64, stop_token_ids=req_data.stop_token_ids
    )
1981
1982
1983
1984

    assert args.num_prompts > 0
    if args.num_prompts == 1:
        # Single inference
1985
        uuid = "uuid_0"
1986
        inputs = {
1987
            "prompt": prompts[0],
1988
            "multi_modal_data": {modality: data},
1989
1990
1991
1992
1993
1994
            "multi_modal_uuids": {modality: uuid},
        }
        inputs_with_empty_media = {
            "prompt": prompts[0],
            "multi_modal_data": {modality: None},
            "multi_modal_uuids": {modality: uuid},
1995
1996
1997
        }
    else:
        # Batch inference
1998
1999
        if args.image_repeat_prob is not None:
            # Repeat images with specified probability of "image_repeat_prob"
2000
2001
2002
2003
2004
2005
            inputs, inputs_with_empty_media = apply_image_repeat(
                args.image_repeat_prob,
                args.num_prompts,
                data,
                prompts,
                modality,
2006
            )
2007
2008
        else:
            # Use the same image for all prompts
2009
2010
2011
2012
2013
2014
2015
2016
2017
2018
2019
2020
2021
2022
2023
2024
2025
2026
            inputs = []
            inputs_with_empty_media = []
            for i in range(args.num_prompts):
                uuid = "uuid_{}".format(i)
                inputs.append(
                    {
                        "prompt": prompts[i % len(prompts)],
                        "multi_modal_data": {modality: data},
                        "multi_modal_uuids": {modality: uuid},
                    }
                )
                inputs_with_empty_media.append(
                    {
                        "prompt": prompts[i % len(prompts)],
                        "multi_modal_data": {modality: None},
                        "multi_modal_uuids": {modality: uuid},
                    }
                )
2027

2028
    # Add LoRA request if applicable
2029
2030
2031
    lora_request = (
        req_data.lora_requests * args.num_prompts if req_data.lora_requests else None
    )
2032

2033
2034
2035
2036
2037
2038
    with time_counter(args.time_generate):
        outputs = llm.generate(
            inputs,
            sampling_params=sampling_params,
            lora_request=lora_request,
        )
2039

2040
    print("-" * 50)
2041
2042
2043
    for o in outputs:
        generated_text = o.outputs[0].text
        print(generated_text)
2044
        print("-" * 50)
2045

2046
2047
2048
2049
2050
2051
2052
2053
2054
2055
2056
2057
2058
2059
2060
2061
2062
2063
2064
2065
    if args.verify_mm_cache_hit_with_uuids:
        try:
            # Verify cache hits with UUIDs
            print(
                "Sending a second batch of requests with empty media"
                " and matching UUIDs."
            )
            outputs = llm.generate(
                inputs_with_empty_media,
                sampling_params=sampling_params,
                lora_request=lora_request,
            )
            print("-" * 50)
            for o in outputs:
                generated_text = o.outputs[0].text
                print(generated_text)
                print("-" * 50)
        except Exception as e:
            print(f"Failed to verify cache hits with UUIDs. Error: {e}")

2066
2067

if __name__ == "__main__":
2068
    args = parse_args()
2069
    main(args)