vision_language.py 57.4 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
15
from dataclasses import asdict
from typing import NamedTuple, Optional
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
31
32
33
34

class ModelRequestData(NamedTuple):
    engine_args: EngineArgs
    prompts: list[str]
    stop_token_ids: Optional[list[int]] = None
    lora_requests: Optional[list[LoRARequest]] = None


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
251
        model=model_name,
        max_model_len=2048,
        max_num_seqs=2,
        mm_processor_kwargs={"do_pan_and_scan": True},
252
        limit_mm_per_prompt={modality: 1},
253
    )
254

255
256
257
258
259
260
261
262
    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
263
264
265
266
267
    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )

268

Nicolò Lucchesi's avatar
Nicolò Lucchesi committed
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
# 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
    ]
290
291
292
293
    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
294
295


296
# GLM-4v
297
def run_glm4v(questions: list[str], modality: str) -> ModelRequestData:
298
    assert modality == "image"
299
    model_name = "zai-org/glm-4v-9b"
300

301
302
303
304
305
306
307
    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"]},
308
        limit_mm_per_prompt={modality: 1},
309
    )
310

311
    prompts = [
312
313
314
315
        (
            "<|user|>\n<|begin_of_image|><|endoftext|><|end_of_image|>"
            f"{question}<|assistant|>"
        )
316
        for question in questions
317
    ]
318

319
    stop_token_ids = [151329, 151336, 151338]
320
321
322
323
324
325

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
        stop_token_ids=stop_token_ids,
    )
326
327


328
329
# GLM-4.1V
def run_glm4_1v(questions: list[str], modality: str) -> ModelRequestData:
330
    model_name = "zai-org/GLM-4.1V-9B-Thinking"
331
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

    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,
    )


364
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
# 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,
    )


438
# H2OVL-Mississippi
439
def run_h2ovl(questions: list[str], modality: str) -> ModelRequestData:
440
441
    assert modality == "image"

442
    model_name = "h2oai/h2ovl-mississippi-800m"
443

444
    engine_args = EngineArgs(
445
446
447
        model=model_name,
        trust_remote_code=True,
        max_model_len=8192,
448
        limit_mm_per_prompt={modality: 1},
449
450
    )

451
452
453
454
455
456
457
    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
    )
458
459

    # Stop tokens for H2OVL-Mississippi
460
    # https://huggingface.co/h2oai/h2ovl-mississippi-800m
461
    stop_token_ids = [tokenizer.eos_token_id]
462
463
464
465
466
467

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
        stop_token_ids=stop_token_ids,
    )
468
469


470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
# 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":
            """
488
489
            ocr: List the words in the image in raster order.
                Even if the word order feels unnatural for reading,
490
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
                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,
    )


549
# Idefics3-8B-Llama3
550
def run_idefics3(questions: list[str], modality: str) -> ModelRequestData:
551
552
553
    assert modality == "image"
    model_name = "HuggingFaceM4/Idefics3-8B-Llama3"

554
    engine_args = EngineArgs(
555
556
557
558
559
560
561
        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={
562
            "size": {"longest_edge": 3 * 364},
563
        },
564
        limit_mm_per_prompt={modality: 1},
565
    )
566
567
568
569
    prompts = [
        (f"<|begin_of_text|>User:<image>{question}<end_of_utterance>\nAssistant:")
        for question in questions
    ]
570
571
572
573
574

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
575
576


Lyu Han's avatar
Lyu Han committed
577
578
579
580
581
582
583
584
585
586
587
588
589
# Intern-S1
def run_interns1(questions: list[str], modality: str) -> ModelRequestData:
    model_name = "internlm/Intern-S1"

    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,
    )

590
591
592
593
594
    if modality == "image":
        placeholder = "<IMG_CONTEXT>"
    elif modality == "video":
        placeholder = "<video>"

Lyu Han's avatar
Lyu Han committed
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
    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,
    )


610
# InternVL
611
def run_internvl(questions: list[str], modality: str) -> ModelRequestData:
612
    model_name = "OpenGVLab/InternVL3-2B"
613

614
    engine_args = EngineArgs(
615
616
        model=model_name,
        trust_remote_code=True,
617
        max_model_len=8192,
618
        limit_mm_per_prompt={modality: 1},
619
620
    )

621
622
623
624
625
    if modality == "image":
        placeholder = "<image>"
    elif modality == "video":
        placeholder = "<video>"

626
627
628
629
630
631
632
633
    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
    )
634
635
636
637
638
639
640

    # 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]
641
    stop_token_ids = [token_id for token_id in stop_token_ids if token_id is not None]
642
643
644
645
646
647

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
        stop_token_ids=stop_token_ids,
    )
648
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
# 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,
    )


681
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
# 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,
    )


712
713
714
715
716
717
718
# 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|>"
719
720
        "<|im_assistant|>assistant<|im_middle|>"
        for question in questions
721
722
723
724
725
    ]

    engine_args = EngineArgs(
        model="moonshotai/Kimi-VL-A3B-Instruct",
        trust_remote_code=True,
Cyrus Leung's avatar
Cyrus Leung committed
726
        max_model_len=4096,
727
        limit_mm_per_prompt={modality: 1},
728
729
730
731
732
733
734
735
    )

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


736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
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,
    )


771
# LLaVA-1.5
772
def run_llava(questions: list[str], modality: str) -> ModelRequestData:
773
    assert modality == "image"
774

775
    prompts = [f"USER: <image>\n{question}\nASSISTANT:" for question in questions]
776

777
778
779
    engine_args = EngineArgs(
        model="llava-hf/llava-1.5-7b-hf",
        max_model_len=4096,
780
        limit_mm_per_prompt={modality: 1},
781
782
783
784
785
786
    )

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
787
788
789


# LLaVA-1.6/LLaVA-NeXT
790
def run_llava_next(questions: list[str], modality: str) -> ModelRequestData:
791
    assert modality == "image"
792

793
    prompts = [f"[INST] <image>\n{question} [/INST]" for question in questions]
794
795
796
    engine_args = EngineArgs(
        model="llava-hf/llava-v1.6-mistral-7b-hf",
        max_model_len=8192,
797
        limit_mm_per_prompt={modality: 1},
798
799
800
801
802
803
    )

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
804
805
806
807


# LlaVA-NeXT-Video
# Currently only support for video input
808
def run_llava_next_video(questions: list[str], modality: str) -> ModelRequestData:
809
810
    assert modality == "video"

811
    prompts = [f"USER: <video>\n{question} ASSISTANT:" for question in questions]
812
813
814
    engine_args = EngineArgs(
        model="llava-hf/LLaVA-NeXT-Video-7B-hf",
        max_model_len=8192,
815
        max_num_seqs=2,
816
        limit_mm_per_prompt={modality: 1},
817
818
819
820
821
822
    )

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
823
824


825
# LLaVA-OneVision
826
def run_llava_onevision(questions: list[str], modality: str) -> ModelRequestData:
827
    if modality == "video":
828
        prompts = [
829
            f"<|im_start|>user <video>\n{question}<|im_end|><|im_start|>assistant\n"
830
            for question in questions
831
        ]
832
833

    elif modality == "image":
834
        prompts = [
835
            f"<|im_start|>user <image>\n{question}<|im_end|><|im_start|>assistant\n"
836
            for question in questions
837
        ]
838

839
840
841
    engine_args = EngineArgs(
        model="llava-hf/llava-onevision-qwen2-7b-ov-hf",
        max_model_len=16384,
842
        limit_mm_per_prompt={modality: 1},
843
844
845
846
847
848
    )

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
849
850


851
# Mantis
852
def run_mantis(questions: list[str], modality: str) -> ModelRequestData:
853
    assert modality == "image"
854

855
856
    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]
857

858
    engine_args = EngineArgs(
859
        model="TIGER-Lab/Mantis-8B-siglip-llama3",
860
        max_model_len=4096,
861
        hf_overrides={"architectures": ["MantisForConditionalGeneration"]},
862
        limit_mm_per_prompt={modality: 1},
863
    )
864
    stop_token_ids = [128009]
865
866
867
868
869
870

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
        stop_token_ids=stop_token_ids,
    )
871
872
873


# MiniCPM-V
874
def run_minicpmv_base(questions: list[str], modality: str, model_name):
875
876
    assert modality in ["image", "video"]
    # If you want to use `MiniCPM-o-2_6` with audio inputs, check `audio_language.py` # noqa
877
878
879
880
881
882
883

    # 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
884
885
    # model_name = "openbmb/MiniCPM-Llama3-V-2_5"

886
    # 2.6
887
888
889
890
891
892
893
894
895
    # 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"
896
    tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
897
    engine_args = EngineArgs(
898
        model=model_name,
899
900
        max_model_len=4096,
        max_num_seqs=2,
901
        trust_remote_code=True,
902
        limit_mm_per_prompt={modality: 1},
903
    )
904
905
906
907
908
909
910
    # 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]

911
    # 2.6 / o2.6
912
    stop_tokens = ["<|im_end|>", "<|endoftext|>"]
913
    stop_token_ids = [tokenizer.convert_tokens_to_ids(i) for i in stop_tokens]
914

915
916
917
918
919
    modality_placeholder = {
        "image": "(<image>./</image>)",
        "video": "(<video>./</video>)",
    }

920
921
    prompts = [
        tokenizer.apply_chat_template(
922
923
924
925
926
927
            [
                {
                    "role": "user",
                    "content": f"{modality_placeholder[modality]}\n{question}",
                }
            ],
928
            tokenize=False,
929
930
931
            add_generation_prompt=True,
        )
        for question in questions
932
    ]
933
934
935
936
937
938

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
        stop_token_ids=stop_token_ids,
    )
939
940


941
def run_minicpmo(questions: list[str], modality: str) -> ModelRequestData:
942
    return run_minicpmv_base(questions, modality, "openbmb/MiniCPM-o-2_6")
943
944


945
def run_minicpmv(questions: list[str], modality: str) -> ModelRequestData:
946
    return run_minicpmv_base(questions, modality, "openbmb/MiniCPM-V-2_6")
947
948


949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
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,
    )


982
983
984
985
986
987
988
989
990
991
992
993
# 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,
994
        limit_mm_per_prompt={modality: 1},
995
        ignore_patterns=["consolidated.safetensors"],
996
997
998
999
1000
1001
1002
1003
1004
1005
    )

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

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


1006
1007
# Molmo
def run_molmo(questions: list[str], modality: str) -> ModelRequestData:
1008
1009
    assert modality == "image"

1010
    model_name = "allenai/Molmo-7B-D-0924"
1011
1012
1013

    engine_args = EngineArgs(
        model=model_name,
1014
1015
        trust_remote_code=True,
        dtype="bfloat16",
1016
        limit_mm_per_prompt={modality: 1},
1017
1018
    )

1019
    prompts = [
1020
        f"<|im_start|>user <image>\n{question}<|im_end|><|im_start|>assistant\n"
1021
1022
        for question in questions
    ]
1023

1024
1025
1026
1027
1028
1029
    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )


1030
1031
1032
# Nemontron_VL
def run_nemotron_vl(questions: list[str], modality: str) -> ModelRequestData:
    model_name = "nvidia/Llama-3.1-Nemotron-Nano-VL-8B-V1"
1033

1034
    engine_args = EngineArgs(
1035
        model=model_name,
1036
        trust_remote_code=True,
1037
        max_model_len=8192,
1038
        limit_mm_per_prompt={modality: 1},
1039
    )
1040

1041
1042
1043
1044
1045
1046
    assert modality == "image"
    placeholder = "<image>"

    tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
    messages = [
        [{"role": "user", "content": f"{placeholder}\n{question}"}]
1047
        for question in questions
1048
    ]
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
    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]
1060
1061
1062
1063

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
1064
        stop_token_ids=stop_token_ids,
1065
    )
1066
1067


1068
# NVLM-D
1069
def run_nvlm_d(questions: list[str], modality: str) -> ModelRequestData:
1070
1071
1072
1073
1074
    assert modality == "image"

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

    # Adjust this as necessary to fit in GPU
1075
    engine_args = EngineArgs(
1076
1077
1078
1079
        model=model_name,
        trust_remote_code=True,
        max_model_len=4096,
        tensor_parallel_size=4,
1080
        limit_mm_per_prompt={modality: 1},
1081
1082
    )

1083
1084
1085
1086
1087
1088
1089
    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
    )
1090
1091
1092
1093
1094

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
1095
1096


1097
1098
# Ovis
def run_ovis(questions: list[str], modality: str) -> ModelRequestData:
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
    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",
1109
        limit_mm_per_prompt={modality: 1},
1110
1111
    )

1112
1113
1114
1115
1116
1117
1118
    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
    )
1119
1120
1121
1122
1123
1124
1125

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


1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
# 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>"

    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,
    )


1158
# PaliGemma
1159
def run_paligemma(questions: list[str], modality: str) -> ModelRequestData:
1160
    assert modality == "image"
1161

1162
    # PaliGemma has special prompt format for VQA
1163
1164
1165
    prompts = ["caption en" for _ in questions]
    engine_args = EngineArgs(
        model="google/paligemma-3b-mix-224",
1166
        limit_mm_per_prompt={modality: 1},
1167
    )
1168
1169
1170
1171
1172

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
1173
1174


1175
# PaliGemma 2
1176
def run_paligemma2(questions: list[str], modality: str) -> ModelRequestData:
1177
    assert modality == "image"
1178

1179
    # PaliGemma 2 has special prompt format for VQA
1180
1181
1182
    prompts = ["caption en" for _ in questions]
    engine_args = EngineArgs(
        model="google/paligemma2-3b-ft-docci-448",
1183
        limit_mm_per_prompt={modality: 1},
1184
    )
1185
1186
1187
1188
1189

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


1192
# Phi-3-Vision
1193
def run_phi3v(questions: list[str], modality: str) -> ModelRequestData:
1194
1195
    assert modality == "image"

1196
1197
1198
1199
    prompts = [
        f"<|user|>\n<|image_1|>\n{question}<|end|>\n<|assistant|>\n"
        for question in questions
    ]
1200

1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
    # 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
1213
    engine_args = EngineArgs(
1214
1215
        model="microsoft/Phi-3.5-vision-instruct",
        trust_remote_code=True,
1216
        max_model_len=4096,
1217
        max_num_seqs=2,
1218
        # Note - mm_processor_kwargs can also be passed to generate/chat calls
1219
        mm_processor_kwargs={"num_crops": 16},
1220
        limit_mm_per_prompt={modality: 1},
1221
    )
1222
1223
1224
1225
1226

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
1227
1228


1229
# Phi-4-multimodal-instruct
1230
def run_phi4mm(questions: list[str], modality: str) -> ModelRequestData:
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
    """
    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 = [
1241
        f"<|user|><|image_1|>{question}<|end|><|assistant|>" for question in questions
1242
    ]
1243
    engine_args = EngineArgs(
1244
1245
        model=model_path,
        trust_remote_code=True,
1246
        max_model_len=5120,
1247
        max_num_seqs=2,
1248
        max_num_batched_tokens=12800,
1249
1250
        enable_lora=True,
        max_lora_rank=320,
1251
1252
        # Note - mm_processor_kwargs can also be passed to generate/chat calls
        mm_processor_kwargs={"dynamic_hd": 16},
1253
        limit_mm_per_prompt={modality: 1},
1254
1255
    )

1256
1257
1258
1259
1260
    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
        lora_requests=[LoRARequest("vision", 1, vision_lora_path)],
    )
1261
1262


1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
# 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)],
    )


1298
# Pixtral HF-format
1299
def run_pixtral_hf(questions: list[str], modality: str) -> ModelRequestData:
1300
1301
1302
1303
    assert modality == "image"

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

1304
    # NOTE: Need L40 (or equivalent) to avoid OOM
1305
    engine_args = EngineArgs(
1306
        model=model_name,
1307
        max_model_len=6144,
1308
        max_num_seqs=2,
1309
        limit_mm_per_prompt={modality: 1},
1310
1311
    )

1312
    prompts = [f"<s>[INST]{question}\n[IMG][/INST]" for question in questions]
1313
1314
1315
1316
1317

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
1318
1319


1320
# Qwen-VL
1321
def run_qwen_vl(questions: list[str], modality: str) -> ModelRequestData:
1322
1323
    assert modality == "image"

1324
    engine_args = EngineArgs(
1325
        model="Qwen/Qwen-VL",
1326
        trust_remote_code=True,
1327
1328
        max_model_len=1024,
        max_num_seqs=2,
1329
        hf_overrides={"architectures": ["QwenVLForConditionalGeneration"]},
1330
        limit_mm_per_prompt={modality: 1},
1331
1332
    )

1333
    prompts = [f"{question}Picture 1: <img></img>\n" for question in questions]
1334
1335
1336
1337
1338

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
1339
1340


1341
# Qwen2-VL
1342
def run_qwen2_vl(questions: list[str], modality: str) -> ModelRequestData:
1343
    model_name = "Qwen/Qwen2-VL-7B-Instruct"
1344

1345
    engine_args = EngineArgs(
1346
        model=model_name,
1347
1348
1349
        max_model_len=4096,
        max_num_seqs=5,
        # Note - mm_processor_kwargs can also be passed to generate/chat calls
1350
        mm_processor_kwargs={
1351
1352
            "min_pixels": 28 * 28,
            "max_pixels": 1280 * 28 * 28,
1353
        },
1354
        limit_mm_per_prompt={modality: 1},
1355
    )
1356

1357
1358
1359
1360
1361
    if modality == "image":
        placeholder = "<|image_pad|>"
    elif modality == "video":
        placeholder = "<|video_pad|>"

1362
    prompts = [
1363
1364
1365
1366
1367
1368
1369
        (
            "<|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
1370
    ]
1371
1372
1373
1374
1375

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
1376
1377


Roger Wang's avatar
Roger Wang committed
1378
# Qwen2.5-VL
1379
def run_qwen2_5_vl(questions: list[str], modality: str) -> ModelRequestData:
Roger Wang's avatar
Roger Wang committed
1380
1381
    model_name = "Qwen/Qwen2.5-VL-3B-Instruct"

1382
    engine_args = EngineArgs(
Roger Wang's avatar
Roger Wang committed
1383
1384
1385
1386
1387
1388
1389
1390
        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,
        },
1391
        limit_mm_per_prompt={modality: 1},
Roger Wang's avatar
Roger Wang committed
1392
1393
1394
1395
1396
1397
1398
    )

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

1399
    prompts = [
1400
1401
1402
1403
1404
1405
1406
        (
            "<|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
1407
    ]
1408
1409
1410
1411
1412

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
Roger Wang's avatar
Roger Wang committed
1413
1414


1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
# 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],
        },
1428
        limit_mm_per_prompt={modality: 1},
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
    )

    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 "
1439
1440
        "generating text and speech."
    )
1441

1442
1443
1444
1445
1446
1447
1448
1449
1450
    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
    ]
1451
1452
1453
1454
1455
1456
    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )


1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
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
# 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,
    )


1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
# 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,
    )


1553
1554
# SkyworkR1V
def run_skyworkr1v(questions: list[str], modality: str) -> ModelRequestData:
1555
    assert modality == "image"
1556
1557

    model_name = "Skywork/Skywork-R1V-38B"
1558
1559
1560
1561
1562
1563
1564

    engine_args = EngineArgs(
        model=model_name,
        trust_remote_code=True,
        max_model_len=4096,
        limit_mm_per_prompt={modality: 1},
    )
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577

    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]
1578
1579
1580
1581

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
1582
        stop_token_ids=stop_token_ids,
1583
1584
1585
    )


1586
1587
1588
1589
# SmolVLM2-2.2B-Instruct
def run_smolvlm(questions: list[str], modality: str) -> ModelRequestData:
    assert modality == "image"
    model_name = "HuggingFaceTB/SmolVLM2-2.2B-Instruct"
1590
1591
1592

    engine_args = EngineArgs(
        model=model_name,
1593
1594
1595
1596
1597
1598
        max_model_len=8192,
        max_num_seqs=2,
        enforce_eager=True,
        mm_processor_kwargs={
            "max_image_size": {"longest_edge": 384},
        },
1599
1600
        limit_mm_per_prompt={modality: 1},
    )
1601
1602
1603
1604
    prompts = [
        (f"<|im_start|>User:<image>{question}<end_of_utterance>\nAssistant:")
        for question in questions
    ]
1605

1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
    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",
    )
1628
1629

    prompts = [
1630
1631
        "<|begin▁of▁sentence|> You are a helpful assistant. <|BOT|>user\n "
        f"<im_patch>{question} <|EOT|><|BOT|>assistant\n<think>\n"
1632
1633
1634
1635
1636
1637
1638
1639
1640
        for question in questions
    ]

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


1641
1642
# omni-research/Tarsier-7b
def run_tarsier(questions: list[str], modality: str) -> ModelRequestData:
1643
    assert modality == "image"
1644
    model_name = "omni-research/Tarsier-7b"
1645
1646
1647
1648
1649

    engine_args = EngineArgs(
        model=model_name,
        trust_remote_code=True,
        max_model_len=4096,
1650
        limit_mm_per_prompt={modality: 1},
1651
    )
1652
    prompts = [(f"USER: <image>\n{question} ASSISTANT:") for question in questions]
1653

1654
1655
1656
    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
1657
    )
1658

1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
1671
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681
1682
1683

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
    ]
1684
1685
1686
1687
1688
1689
1690

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


1691
model_example_map = {
1692
    "aria": run_aria,
Jennifer Zhao's avatar
Jennifer Zhao committed
1693
    "aya_vision": run_aya_vision,
1694
1695
    "blip-2": run_blip2,
    "chameleon": run_chameleon,
Roger Wang's avatar
Roger Wang committed
1696
    "dots_ocr": run_dots_ocr,
1697
    "command_a_vision": run_command_a_vision,
1698
    "deepseek_vl_v2": run_deepseek_vl2,
1699
    "ernie45_vl": run_ernie45_vl,
1700
    "fuyu": run_fuyu,
1701
    "gemma3": run_gemma3,
Nicolò Lucchesi's avatar
Nicolò Lucchesi committed
1702
    "gemma3n": run_gemma3n,
1703
    "glm4v": run_glm4v,
1704
    "glm4_1v": run_glm4_1v,
1705
1706
    "glm4_5v": run_glm4_5v,
    "glm4_5v_fp8": run_glm4_5v_fp8,
1707
    "h2ovl_chat": run_h2ovl,
1708
    "hyperclovax_seed_vision": run_hyperclovax_seed_vision,
1709
    "idefics3": run_idefics3,
Lyu Han's avatar
Lyu Han committed
1710
    "interns1": run_interns1,
1711
    "internvl_chat": run_internvl,
1712
    "keye_vl": run_keye_vl,
1713
    "keye_vl1_5": run_keye_vl1_5,
1714
    "kimi_vl": run_kimi_vl,
1715
    "llama4": run_llama4,
1716
1717
    "llava": run_llava,
    "llava-next": run_llava_next,
1718
    "llava-next-video": run_llava_next_video,
1719
    "llava-onevision": run_llava_onevision,
1720
    "mantis": run_mantis,
1721
    "minicpmo": run_minicpmo,
1722
    "minicpmv": run_minicpmv,
1723
    "minimax_vl_01": run_minimax_vl_01,
1724
    "mistral3": run_mistral3,
1725
    "molmo": run_molmo,
1726
    "nemotron_vl": run_nemotron_vl,
1727
    "NVLM_D": run_nvlm_d,
1728
    "ovis": run_ovis,
1729
    "ovis2_5": run_ovis2_5,
1730
1731
1732
    "paligemma": run_paligemma,
    "paligemma2": run_paligemma2,
    "phi3_v": run_phi3v,
1733
    "phi4_mm": run_phi4mm,
1734
    "phi4_multimodal": run_phi4_multimodal,
1735
    "pixtral_hf": run_pixtral_hf,
1736
    "qwen_vl": run_qwen_vl,
1737
    "qwen2_vl": run_qwen2_vl,
Roger Wang's avatar
Roger Wang committed
1738
    "qwen2_5_vl": run_qwen2_5_vl,
1739
    "qwen2_5_omni": run_qwen2_5_omni,
1740
1741
    "qwen3_vl": run_qwen3_vl,
    "qwen3_vl_moe": run_qwen3_vl_moe,
1742
    "rvl": run_r_vl,
1743
    "skywork_chat": run_skyworkr1v,
1744
    "smolvlm": run_smolvlm,
1745
    "step3": run_step3,
汪志鹏's avatar
汪志鹏 committed
1746
    "tarsier": run_tarsier,
1747
    "tarsier2": run_tarsier2,
1748
1749
1750
}


1751
1752
1753
1754
MODELS_NEED_VIDEO_METADATA = [
    "glm4_1v",
    "glm4_5v",
    "glm4_5v_fp8",
1755
1756
    "qwen3_vl",
    "qwen3_vl_moe",
1757
1758
1759
]


1760
1761
1762
1763
1764
1765
1766
1767
1768
def get_multi_modal_input(args):
    """
    return {
        "data": image or video,
        "question": question,
    }
    """
    if args.modality == "image":
        # Input image and question
1769
        image = convert_image_mode(ImageAsset("cherry_blossom").pil_image, "RGB")
1770
1771
1772
1773
1774
1775
        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?",
        ]
1776
1777
1778

        return {
            "data": image,
1779
            "questions": img_questions,
1780
1781
1782
1783
        }

    if args.modality == "video":
        # Input video and question
1784
        needs_metadata = args.model_type in MODELS_NEED_VIDEO_METADATA
1785
        video = VideoAsset(name="baby_reading", num_frames=args.num_frames).np_ndarrays
1786
        metadata = VideoAsset(name="baby_reading", num_frames=args.num_frames).metadata
1787
        vid_questions = ["Why is this video funny?"]
1788
1789

        return {
1790
            "data": ([(video, metadata)] if needs_metadata else video),
1791
            "questions": vid_questions,
1792
1793
1794
1795
1796
1797
        }

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


1798
1799
1800
1801
def apply_image_repeat(
    image_repeat_prob, num_prompts, data, prompts: list[str], modality
):
    """Repeats images with provided probability of "image_repeat_prob".
1802
1803
    Used to simulate hit/miss for the MM preprocessor cache.
    """
1804
    assert image_repeat_prob <= 1.0 and image_repeat_prob >= 0
1805
1806
1807
1808
    no_yes = [0, 1]
    probs = [1.0 - image_repeat_prob, image_repeat_prob]

    inputs = []
1809
    inputs_with_empty_media = []
1810
1811
1812
1813
1814
1815
1816
1817
1818
1819
    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)

1820
1821
        uuid = "uuid_{}".format(i)

1822
1823
1824
1825
        inputs.append(
            {
                "prompt": prompts[i % len(prompts)],
                "multi_modal_data": {modality: cur_image},
1826
1827
1828
1829
1830
1831
1832
1833
1834
                "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},
1835
            }
1836
        )
1837

1838
    return inputs, inputs_with_empty_media
1839
1840


1841
1842
1843
1844
@contextmanager
def time_counter(enable: bool):
    if enable:
        import time
1845

1846
1847
1848
1849
1850
1851
1852
1853
1854
1855
        start_time = time.time()
        yield
        elapsed_time = time.time() - start_time
        print("-" * 50)
        print("-- generate time = {}".format(elapsed_time))
        print("-" * 50)
    else:
        yield


1856
1857
def parse_args():
    parser = FlexibleArgumentParser(
1858
1859
1860
1861
1862
1863
1864
1865
1866
1867
1868
1869
1870
1871
1872
1873
1874
1875
1876
1877
1878
1879
1880
1881
1882
1883
1884
1885
1886
1887
1888
1889
1890
        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`.",
    )
1891
1892

    parser.add_argument(
1893
        "--image-repeat-prob",
1894
1895
        type=float,
        default=None,
1896
1897
        help="Simulates the hit-ratio for multi-modal preprocessor cache (if enabled)",
    )
1898
1899

    parser.add_argument(
1900
        "--disable-mm-processor-cache",
1901
        action="store_true",
1902
        help="If True, disables caching of multi-modal processor.",
1903
    )
1904
1905

    parser.add_argument(
1906
1907
1908
1909
        "--time-generate",
        action="store_true",
        help="If True, then print the total generate() call time",
    )
1910
1911

    parser.add_argument(
1912
1913
1914
1915
1916
        "--use-different-prompt-per-request",
        action="store_true",
        help="If True, then use different prompt (with the same multi-modal "
        "data) for each request.",
    )
1917
1918
1919
1920
1921
1922
1923

    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.",
    )
1924
1925
1926
    return parser.parse_args()


1927
1928
1929
1930
1931
def main(args):
    model = args.model_type
    if model not in model_example_map:
        raise ValueError(f"Model type {model} is not supported.")

1932
1933
1934
    modality = args.modality
    mm_input = get_multi_modal_input(args)
    data = mm_input["data"]
1935
    questions = mm_input["questions"]
1936

1937
1938
    req_data = model_example_map[model](questions, modality)

1939
1940
1941
    # 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(
1942
1943
        req_data.engine_args.limit_mm_per_prompt or {}
    )
1944
1945
1946

    engine_args = asdict(req_data.engine_args) | {
        "seed": args.seed,
1947
        "mm_processor_cache_gb": 0 if args.disable_mm_processor_cache else 4,
1948
    }
1949
1950
    llm = LLM(**engine_args)

1951
    # Don't want to check the flag multiple times, so just hijack `prompts`.
1952
1953
1954
1955
1956
    prompts = (
        req_data.prompts
        if args.use_different_prompt_per_request
        else [req_data.prompts[0]]
    )
1957
1958
1959

    # We set temperature to 0.2 so that outputs can be different
    # even when all prompts are identical when running batch inference.
1960
1961
1962
    sampling_params = SamplingParams(
        temperature=0.2, max_tokens=64, stop_token_ids=req_data.stop_token_ids
    )
1963
1964
1965
1966

    assert args.num_prompts > 0
    if args.num_prompts == 1:
        # Single inference
1967
        uuid = "uuid_0"
1968
        inputs = {
1969
            "prompt": prompts[0],
1970
            "multi_modal_data": {modality: data},
1971
1972
1973
1974
1975
1976
            "multi_modal_uuids": {modality: uuid},
        }
        inputs_with_empty_media = {
            "prompt": prompts[0],
            "multi_modal_data": {modality: None},
            "multi_modal_uuids": {modality: uuid},
1977
1978
1979
        }
    else:
        # Batch inference
1980
1981
        if args.image_repeat_prob is not None:
            # Repeat images with specified probability of "image_repeat_prob"
1982
1983
1984
1985
1986
1987
            inputs, inputs_with_empty_media = apply_image_repeat(
                args.image_repeat_prob,
                args.num_prompts,
                data,
                prompts,
                modality,
1988
            )
1989
1990
        else:
            # Use the same image for all prompts
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
            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},
                    }
                )
2009

2010
    # Add LoRA request if applicable
2011
2012
2013
    lora_request = (
        req_data.lora_requests * args.num_prompts if req_data.lora_requests else None
    )
2014

2015
2016
2017
2018
2019
2020
    with time_counter(args.time_generate):
        outputs = llm.generate(
            inputs,
            sampling_params=sampling_params,
            lora_request=lora_request,
        )
2021

2022
    print("-" * 50)
2023
2024
2025
    for o in outputs:
        generated_text = o.outputs[0].text
        print(generated_text)
2026
        print("-" * 50)
2027

2028
2029
2030
2031
2032
2033
2034
2035
2036
2037
2038
2039
2040
2041
2042
2043
2044
2045
2046
2047
    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}")

2048
2049

if __name__ == "__main__":
2050
    args = parse_args()
2051
    main(args)