vision_language.py 52.7 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
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


129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
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,
    )


152
# Deepseek-VL2
153
def run_deepseek_vl2(questions: list[str], modality: str) -> ModelRequestData:
154
155
    assert modality == "image"

156
    model_name = "deepseek-ai/deepseek-vl2-tiny"
157

158
159
160
161
162
    engine_args = EngineArgs(
        model=model_name,
        max_model_len=4096,
        max_num_seqs=2,
        hf_overrides={"architectures": ["DeepseekVLV2ForCausalLM"]},
163
        limit_mm_per_prompt={modality: 1},
164
    )
165

166
    prompts = [
167
        f"<|User|>: <image>\n{question}\n\n<|Assistant|>:" for question in questions
168
    ]
169
170
171
172
173

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
174
175


176
# Florence2
177
def run_florence2(questions: list[str], modality: str) -> ModelRequestData:
178
179
    assert modality == "image"

180
181
    engine_args = EngineArgs(
        model="microsoft/Florence-2-large",
182
        tokenizer="Isotr0py/Florence-2-tokenizer",
183
184
        max_model_len=4096,
        max_num_seqs=2,
185
186
        trust_remote_code=True,
        dtype="bfloat16",
187
        limit_mm_per_prompt={modality: 1},
188
    )
189

190
191
192
193
194
195
    prompts = ["<MORE_DETAILED_CAPTION>" for _ in questions]

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
196
197


198
# Fuyu
199
def run_fuyu(questions: list[str], modality: str) -> ModelRequestData:
200
201
    assert modality == "image"

202
    prompts = [f"{question}\n" for question in questions]
203
204
205
206
    engine_args = EngineArgs(
        model="adept/fuyu-8b",
        max_model_len=2048,
        max_num_seqs=2,
207
        limit_mm_per_prompt={modality: 1},
208
209
210
211
212
213
    )

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
214
215


216
# Gemma 3
217
def run_gemma3(questions: list[str], modality: str) -> ModelRequestData:
218
219
220
    assert modality == "image"
    model_name = "google/gemma-3-4b-it"

221
    engine_args = EngineArgs(
222
223
224
225
        model=model_name,
        max_model_len=2048,
        max_num_seqs=2,
        mm_processor_kwargs={"do_pan_and_scan": True},
226
        limit_mm_per_prompt={modality: 1},
227
    )
228

229
230
231
232
233
234
235
236
    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
237
238
239
240
241
    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )

242

Nicolò Lucchesi's avatar
Nicolò Lucchesi committed
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
# 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
    ]
264
265
266
267
    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
268
269


270
# GLM-4v
271
def run_glm4v(questions: list[str], modality: str) -> ModelRequestData:
272
    assert modality == "image"
273
    model_name = "zai-org/glm-4v-9b"
274

275
276
277
278
279
280
281
    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"]},
282
        limit_mm_per_prompt={modality: 1},
283
    )
284

285
    prompts = [
286
287
288
289
        (
            "<|user|>\n<|begin_of_image|><|endoftext|><|end_of_image|>"
            f"{question}<|assistant|>"
        )
290
        for question in questions
291
    ]
292

293
    stop_token_ids = [151329, 151336, 151338]
294
295
296
297
298
299

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
        stop_token_ids=stop_token_ids,
    )
300
301


302
303
# GLM-4.1V
def run_glm4_1v(questions: list[str], modality: str) -> ModelRequestData:
304
    model_name = "zai-org/GLM-4.1V-9B-Thinking"
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337

    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,
    )


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
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
# 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,
    )


412
# H2OVL-Mississippi
413
def run_h2ovl(questions: list[str], modality: str) -> ModelRequestData:
414
415
    assert modality == "image"

416
    model_name = "h2oai/h2ovl-mississippi-800m"
417

418
    engine_args = EngineArgs(
419
420
421
        model=model_name,
        trust_remote_code=True,
        max_model_len=8192,
422
        limit_mm_per_prompt={modality: 1},
423
424
    )

425
426
427
428
429
430
431
    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
    )
432
433

    # Stop tokens for H2OVL-Mississippi
434
    # https://huggingface.co/h2oai/h2ovl-mississippi-800m
435
    stop_token_ids = [tokenizer.eos_token_id]
436
437
438
439
440
441

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
        stop_token_ids=stop_token_ids,
    )
442
443


444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
# 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":
            """
462
463
            ocr: List the words in the image in raster order.
                Even if the word order feels unnatural for reading,
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
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
                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,
    )


523
# Idefics3-8B-Llama3
524
def run_idefics3(questions: list[str], modality: str) -> ModelRequestData:
525
526
527
    assert modality == "image"
    model_name = "HuggingFaceM4/Idefics3-8B-Llama3"

528
    engine_args = EngineArgs(
529
530
531
532
533
534
535
        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={
536
            "size": {"longest_edge": 3 * 364},
537
        },
538
        limit_mm_per_prompt={modality: 1},
539
    )
540
541
542
543
    prompts = [
        (f"<|begin_of_text|>User:<image>{question}<end_of_utterance>\nAssistant:")
        for question in questions
    ]
544
545
546
547
548

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
549
550


Lyu Han's avatar
Lyu Han committed
551
552
553
554
555
556
557
558
559
560
561
562
563
# 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,
    )

564
565
566
567
568
    if modality == "image":
        placeholder = "<IMG_CONTEXT>"
    elif modality == "video":
        placeholder = "<video>"

Lyu Han's avatar
Lyu Han committed
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
    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,
    )


584
# InternVL
585
def run_internvl(questions: list[str], modality: str) -> ModelRequestData:
586
    model_name = "OpenGVLab/InternVL3-2B"
587

588
    engine_args = EngineArgs(
589
590
        model=model_name,
        trust_remote_code=True,
591
        max_model_len=8192,
592
        limit_mm_per_prompt={modality: 1},
593
594
    )

595
596
597
598
599
    if modality == "image":
        placeholder = "<image>"
    elif modality == "video":
        placeholder = "<video>"

600
601
602
603
604
605
606
607
    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
    )
608
609
610
611
612
613
614

    # 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]
615
    stop_token_ids = [token_id for token_id in stop_token_ids if token_id is not None]
616
617
618
619
620
621

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
        stop_token_ids=stop_token_ids,
    )
622
623


624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
# 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,
    )


655
656
657
658
659
660
661
# 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|>"
662
663
        "<|im_assistant|>assistant<|im_middle|>"
        for question in questions
664
665
666
667
668
    ]

    engine_args = EngineArgs(
        model="moonshotai/Kimi-VL-A3B-Instruct",
        trust_remote_code=True,
Cyrus Leung's avatar
Cyrus Leung committed
669
        max_model_len=4096,
670
        limit_mm_per_prompt={modality: 1},
671
672
673
674
675
676
677
678
    )

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


679
680
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
712
713
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,
    )


714
# LLaVA-1.5
715
def run_llava(questions: list[str], modality: str) -> ModelRequestData:
716
    assert modality == "image"
717

718
    prompts = [f"USER: <image>\n{question}\nASSISTANT:" for question in questions]
719

720
721
722
    engine_args = EngineArgs(
        model="llava-hf/llava-1.5-7b-hf",
        max_model_len=4096,
723
        limit_mm_per_prompt={modality: 1},
724
725
726
727
728
729
    )

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
730
731
732


# LLaVA-1.6/LLaVA-NeXT
733
def run_llava_next(questions: list[str], modality: str) -> ModelRequestData:
734
    assert modality == "image"
735

736
    prompts = [f"[INST] <image>\n{question} [/INST]" for question in questions]
737
738
739
    engine_args = EngineArgs(
        model="llava-hf/llava-v1.6-mistral-7b-hf",
        max_model_len=8192,
740
        limit_mm_per_prompt={modality: 1},
741
742
743
744
745
746
    )

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
747
748
749
750


# LlaVA-NeXT-Video
# Currently only support for video input
751
def run_llava_next_video(questions: list[str], modality: str) -> ModelRequestData:
752
753
    assert modality == "video"

754
    prompts = [f"USER: <video>\n{question} ASSISTANT:" for question in questions]
755
756
757
    engine_args = EngineArgs(
        model="llava-hf/LLaVA-NeXT-Video-7B-hf",
        max_model_len=8192,
758
        max_num_seqs=2,
759
        limit_mm_per_prompt={modality: 1},
760
761
762
763
764
765
    )

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
766
767


768
# LLaVA-OneVision
769
def run_llava_onevision(questions: list[str], modality: str) -> ModelRequestData:
770
    if modality == "video":
771
        prompts = [
772
            f"<|im_start|>user <video>\n{question}<|im_end|><|im_start|>assistant\n"
773
            for question in questions
774
        ]
775
776

    elif modality == "image":
777
        prompts = [
778
            f"<|im_start|>user <image>\n{question}<|im_end|><|im_start|>assistant\n"
779
            for question in questions
780
        ]
781

782
783
784
    engine_args = EngineArgs(
        model="llava-hf/llava-onevision-qwen2-7b-ov-hf",
        max_model_len=16384,
785
        limit_mm_per_prompt={modality: 1},
786
787
788
789
790
791
    )

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
792
793


794
# Mantis
795
def run_mantis(questions: list[str], modality: str) -> ModelRequestData:
796
    assert modality == "image"
797

798
799
    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]
800

801
    engine_args = EngineArgs(
802
        model="TIGER-Lab/Mantis-8B-siglip-llama3",
803
        max_model_len=4096,
804
        hf_overrides={"architectures": ["MantisForConditionalGeneration"]},
805
        limit_mm_per_prompt={modality: 1},
806
    )
807
    stop_token_ids = [128009]
808
809
810
811
812
813

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
        stop_token_ids=stop_token_ids,
    )
814
815
816


# MiniCPM-V
817
def run_minicpmv_base(questions: list[str], modality: str, model_name):
818
819
    assert modality in ["image", "video"]
    # If you want to use `MiniCPM-o-2_6` with audio inputs, check `audio_language.py` # noqa
820
821
822
823
824
825
826

    # 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
827
828
    # model_name = "openbmb/MiniCPM-Llama3-V-2_5"

829
    # 2.6
830
831
832
833
834
835
836
837
838
    # 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"
839
    tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
840
    engine_args = EngineArgs(
841
        model=model_name,
842
843
        max_model_len=4096,
        max_num_seqs=2,
844
        trust_remote_code=True,
845
        limit_mm_per_prompt={modality: 1},
846
    )
847
848
849
850
851
852
853
    # 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]

854
    # 2.6 / o2.6
855
    stop_tokens = ["<|im_end|>", "<|endoftext|>"]
856
    stop_token_ids = [tokenizer.convert_tokens_to_ids(i) for i in stop_tokens]
857

858
859
860
861
862
    modality_placeholder = {
        "image": "(<image>./</image>)",
        "video": "(<video>./</video>)",
    }

863
864
    prompts = [
        tokenizer.apply_chat_template(
865
866
867
868
869
870
            [
                {
                    "role": "user",
                    "content": f"{modality_placeholder[modality]}\n{question}",
                }
            ],
871
            tokenize=False,
872
873
874
            add_generation_prompt=True,
        )
        for question in questions
875
    ]
876
877
878
879
880
881

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
        stop_token_ids=stop_token_ids,
    )
882
883


884
def run_minicpmo(questions: list[str], modality: str) -> ModelRequestData:
885
    return run_minicpmv_base(questions, modality, "openbmb/MiniCPM-o-2_6")
886
887


888
def run_minicpmv(questions: list[str], modality: str) -> ModelRequestData:
889
    return run_minicpmv_base(questions, modality, "openbmb/MiniCPM-V-2_6")
890
891


892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
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,
    )


925
926
927
928
929
930
931
932
933
934
935
936
# 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,
937
        limit_mm_per_prompt={modality: 1},
938
        ignore_patterns=["consolidated.safetensors"],
939
940
941
942
943
944
945
946
947
948
    )

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

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


949
# LLama 3.2
950
def run_mllama(questions: list[str], modality: str) -> ModelRequestData:
951
952
    assert modality == "image"

953
    model_name = "meta-llama/Llama-3.2-11B-Vision-Instruct"
954

955
956
957
958
959
    # Note: The default setting of max_num_seqs (256) and
    # max_model_len (131072) for this model may cause OOM.
    # You may lower either to run this example on lower-end GPUs.

    # The configuration below has been confirmed to launch on a single L40 GPU.
960
    engine_args = EngineArgs(
961
        model=model_name,
962
        max_model_len=8192,
963
        max_num_seqs=2,
964
        limit_mm_per_prompt={modality: 1},
965
966
    )

967
    tokenizer = AutoTokenizer.from_pretrained(model_name)
968
969
970
971
972
973
974
975
976
977
978
979
    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
    )
980
981
982
983
984

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
985
986


987
988
# Molmo
def run_molmo(questions: list[str], modality: str) -> ModelRequestData:
989
990
    assert modality == "image"

991
    model_name = "allenai/Molmo-7B-D-0924"
992
993
994

    engine_args = EngineArgs(
        model=model_name,
995
996
        trust_remote_code=True,
        dtype="bfloat16",
997
        limit_mm_per_prompt={modality: 1},
998
999
    )

1000
    prompts = [
1001
        f"<|im_start|>user <image>\n{question}<|im_end|><|im_start|>assistant\n"
1002
1003
        for question in questions
    ]
1004

1005
1006
1007
1008
1009
1010
    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )


1011
1012
1013
# Nemontron_VL
def run_nemotron_vl(questions: list[str], modality: str) -> ModelRequestData:
    model_name = "nvidia/Llama-3.1-Nemotron-Nano-VL-8B-V1"
1014

1015
    engine_args = EngineArgs(
1016
        model=model_name,
1017
        trust_remote_code=True,
1018
        max_model_len=8192,
1019
        limit_mm_per_prompt={modality: 1},
1020
    )
1021

1022
1023
1024
1025
1026
1027
    assert modality == "image"
    placeholder = "<image>"

    tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
    messages = [
        [{"role": "user", "content": f"{placeholder}\n{question}"}]
1028
        for question in questions
1029
    ]
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
    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]
1041
1042
1043
1044

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


1049
# NVLM-D
1050
def run_nvlm_d(questions: list[str], modality: str) -> ModelRequestData:
1051
1052
1053
1054
1055
    assert modality == "image"

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

    # Adjust this as necessary to fit in GPU
1056
    engine_args = EngineArgs(
1057
1058
1059
1060
        model=model_name,
        trust_remote_code=True,
        max_model_len=4096,
        tensor_parallel_size=4,
1061
        limit_mm_per_prompt={modality: 1},
1062
1063
    )

1064
1065
1066
1067
1068
1069
1070
    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
    )
1071
1072
1073
1074
1075

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
1076
1077


1078
1079
# Ovis
def run_ovis(questions: list[str], modality: str) -> ModelRequestData:
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
    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",
1090
        limit_mm_per_prompt={modality: 1},
1091
1092
    )

1093
1094
1095
1096
1097
1098
1099
    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
    )
1100
1101
1102
1103
1104
1105
1106

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


1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
# 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,
    )


1139
# PaliGemma
1140
def run_paligemma(questions: list[str], modality: str) -> ModelRequestData:
1141
    assert modality == "image"
1142

1143
    # PaliGemma has special prompt format for VQA
1144
1145
1146
    prompts = ["caption en" for _ in questions]
    engine_args = EngineArgs(
        model="google/paligemma-3b-mix-224",
1147
        limit_mm_per_prompt={modality: 1},
1148
    )
1149
1150
1151
1152
1153

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
1154
1155


1156
# PaliGemma 2
1157
def run_paligemma2(questions: list[str], modality: str) -> ModelRequestData:
1158
    assert modality == "image"
1159

1160
    # PaliGemma 2 has special prompt format for VQA
1161
1162
1163
    prompts = ["caption en" for _ in questions]
    engine_args = EngineArgs(
        model="google/paligemma2-3b-ft-docci-448",
1164
        limit_mm_per_prompt={modality: 1},
1165
    )
1166
1167
1168
1169
1170

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
1171
1172


1173
# Phi-3-Vision
1174
def run_phi3v(questions: list[str], modality: str) -> ModelRequestData:
1175
1176
    assert modality == "image"

1177
1178
1179
1180
    prompts = [
        f"<|user|>\n<|image_1|>\n{question}<|end|>\n<|assistant|>\n"
        for question in questions
    ]
1181

1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
    # 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
1194
    engine_args = EngineArgs(
1195
1196
        model="microsoft/Phi-3.5-vision-instruct",
        trust_remote_code=True,
1197
        max_model_len=4096,
1198
        max_num_seqs=2,
1199
        # Note - mm_processor_kwargs can also be passed to generate/chat calls
1200
        mm_processor_kwargs={"num_crops": 16},
1201
        limit_mm_per_prompt={modality: 1},
1202
    )
1203
1204
1205
1206
1207

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
1208
1209


1210
# Phi-4-multimodal-instruct
1211
def run_phi4mm(questions: list[str], modality: str) -> ModelRequestData:
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
    """
    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 = [
1222
        f"<|user|><|image_1|>{question}<|end|><|assistant|>" for question in questions
1223
    ]
1224
    engine_args = EngineArgs(
1225
1226
        model=model_path,
        trust_remote_code=True,
1227
        max_model_len=5120,
1228
        max_num_seqs=2,
1229
        max_num_batched_tokens=12800,
1230
1231
        enable_lora=True,
        max_lora_rank=320,
1232
1233
        # Note - mm_processor_kwargs can also be passed to generate/chat calls
        mm_processor_kwargs={"dynamic_hd": 16},
1234
        limit_mm_per_prompt={modality: 1},
1235
1236
    )

1237
1238
1239
1240
1241
    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
        lora_requests=[LoRARequest("vision", 1, vision_lora_path)],
    )
1242
1243


1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
# 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)],
    )


1279
# Pixtral HF-format
1280
def run_pixtral_hf(questions: list[str], modality: str) -> ModelRequestData:
1281
1282
1283
1284
    assert modality == "image"

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

1285
    # NOTE: Need L40 (or equivalent) to avoid OOM
1286
    engine_args = EngineArgs(
1287
        model=model_name,
1288
        max_model_len=6144,
1289
        max_num_seqs=2,
1290
        limit_mm_per_prompt={modality: 1},
1291
1292
    )

1293
    prompts = [f"<s>[INST]{question}\n[IMG][/INST]" for question in questions]
1294
1295
1296
1297
1298

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
1299
1300


1301
# Qwen-VL
1302
def run_qwen_vl(questions: list[str], modality: str) -> ModelRequestData:
1303
1304
    assert modality == "image"

1305
    engine_args = EngineArgs(
1306
        model="Qwen/Qwen-VL",
1307
        trust_remote_code=True,
1308
1309
        max_model_len=1024,
        max_num_seqs=2,
1310
        hf_overrides={"architectures": ["QwenVLForConditionalGeneration"]},
1311
        limit_mm_per_prompt={modality: 1},
1312
1313
    )

1314
    prompts = [f"{question}Picture 1: <img></img>\n" for question in questions]
1315
1316
1317
1318
1319

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
1320
1321


1322
# Qwen2-VL
1323
def run_qwen2_vl(questions: list[str], modality: str) -> ModelRequestData:
1324
    model_name = "Qwen/Qwen2-VL-7B-Instruct"
1325

1326
    engine_args = EngineArgs(
1327
        model=model_name,
1328
1329
1330
        max_model_len=4096,
        max_num_seqs=5,
        # Note - mm_processor_kwargs can also be passed to generate/chat calls
1331
        mm_processor_kwargs={
1332
1333
            "min_pixels": 28 * 28,
            "max_pixels": 1280 * 28 * 28,
1334
        },
1335
        limit_mm_per_prompt={modality: 1},
1336
    )
1337

1338
1339
1340
1341
1342
    if modality == "image":
        placeholder = "<|image_pad|>"
    elif modality == "video":
        placeholder = "<|video_pad|>"

1343
    prompts = [
1344
1345
1346
1347
1348
1349
1350
        (
            "<|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
1351
    ]
1352
1353
1354
1355
1356

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
1357
1358


Roger Wang's avatar
Roger Wang committed
1359
# Qwen2.5-VL
1360
def run_qwen2_5_vl(questions: list[str], modality: str) -> ModelRequestData:
Roger Wang's avatar
Roger Wang committed
1361
1362
    model_name = "Qwen/Qwen2.5-VL-3B-Instruct"

1363
    engine_args = EngineArgs(
Roger Wang's avatar
Roger Wang committed
1364
1365
1366
1367
1368
1369
1370
1371
        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,
        },
1372
        limit_mm_per_prompt={modality: 1},
Roger Wang's avatar
Roger Wang committed
1373
1374
1375
1376
1377
1378
1379
    )

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

1380
    prompts = [
1381
1382
1383
1384
1385
1386
1387
        (
            "<|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
1388
    ]
1389
1390
1391
1392
1393

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
Roger Wang's avatar
Roger Wang committed
1394
1395


1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
# 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],
        },
1409
        limit_mm_per_prompt={modality: 1},
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
    )

    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 "
1420
1421
        "generating text and speech."
    )
1422

1423
1424
1425
1426
1427
1428
1429
1430
1431
    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
    ]
1432
1433
1434
1435
1436
1437
    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )


1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
# 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,
    )


1460
1461
# SkyworkR1V
def run_skyworkr1v(questions: list[str], modality: str) -> ModelRequestData:
1462
    assert modality == "image"
1463
1464

    model_name = "Skywork/Skywork-R1V-38B"
1465
1466
1467
1468
1469
1470
1471

    engine_args = EngineArgs(
        model=model_name,
        trust_remote_code=True,
        max_model_len=4096,
        limit_mm_per_prompt={modality: 1},
    )
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484

    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]
1485
1486
1487
1488

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
1489
        stop_token_ids=stop_token_ids,
1490
1491
1492
    )


1493
1494
1495
1496
# SmolVLM2-2.2B-Instruct
def run_smolvlm(questions: list[str], modality: str) -> ModelRequestData:
    assert modality == "image"
    model_name = "HuggingFaceTB/SmolVLM2-2.2B-Instruct"
1497
1498
1499

    engine_args = EngineArgs(
        model=model_name,
1500
1501
1502
1503
1504
1505
        max_model_len=8192,
        max_num_seqs=2,
        enforce_eager=True,
        mm_processor_kwargs={
            "max_image_size": {"longest_edge": 384},
        },
1506
1507
        limit_mm_per_prompt={modality: 1},
    )
1508
1509
1510
1511
    prompts = [
        (f"<|im_start|>User:<image>{question}<end_of_utterance>\nAssistant:")
        for question in questions
    ]
1512

1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
    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",
    )
1535
1536

    prompts = [
1537
1538
        "<|begin▁of▁sentence|> You are a helpful assistant. <|BOT|>user\n "
        f"<im_patch>{question} <|EOT|><|BOT|>assistant\n<think>\n"
1539
1540
1541
1542
1543
1544
1545
1546
1547
        for question in questions
    ]

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


1548
1549
# omni-research/Tarsier-7b
def run_tarsier(questions: list[str], modality: str) -> ModelRequestData:
1550
    assert modality == "image"
1551
    model_name = "omni-research/Tarsier-7b"
1552
1553
1554
1555
1556

    engine_args = EngineArgs(
        model=model_name,
        trust_remote_code=True,
        max_model_len=4096,
1557
        limit_mm_per_prompt={modality: 1},
1558
    )
1559
    prompts = [(f"USER: <image>\n{question} ASSISTANT:") for question in questions]
1560

1561
1562
1563
    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
1564
    )
1565

1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590

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
    ]
1591
1592
1593
1594
1595
1596
1597

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


1598
model_example_map = {
1599
    "aria": run_aria,
Jennifer Zhao's avatar
Jennifer Zhao committed
1600
    "aya_vision": run_aya_vision,
1601
1602
    "blip-2": run_blip2,
    "chameleon": run_chameleon,
1603
    "command_a_vision": run_command_a_vision,
1604
    "deepseek_vl_v2": run_deepseek_vl2,
1605
    "florence2": run_florence2,
1606
    "fuyu": run_fuyu,
1607
    "gemma3": run_gemma3,
Nicolò Lucchesi's avatar
Nicolò Lucchesi committed
1608
    "gemma3n": run_gemma3n,
1609
    "glm4v": run_glm4v,
1610
    "glm4_1v": run_glm4_1v,
1611
1612
    "glm4_5v": run_glm4_5v,
    "glm4_5v_fp8": run_glm4_5v_fp8,
1613
    "h2ovl_chat": run_h2ovl,
1614
    "hyperclovax_seed_vision": run_hyperclovax_seed_vision,
1615
    "idefics3": run_idefics3,
Lyu Han's avatar
Lyu Han committed
1616
    "interns1": run_interns1,
1617
    "internvl_chat": run_internvl,
1618
    "keye_vl": run_keye_vl,
1619
    "kimi_vl": run_kimi_vl,
1620
    "llama4": run_llama4,
1621
1622
    "llava": run_llava,
    "llava-next": run_llava_next,
1623
    "llava-next-video": run_llava_next_video,
1624
    "llava-onevision": run_llava_onevision,
1625
    "mantis": run_mantis,
1626
    "minicpmo": run_minicpmo,
1627
    "minicpmv": run_minicpmv,
1628
    "minimax_vl_01": run_minimax_vl_01,
1629
    "mistral3": run_mistral3,
1630
1631
    "mllama": run_mllama,
    "molmo": run_molmo,
1632
    "nemotron_vl": run_nemotron_vl,
1633
    "NVLM_D": run_nvlm_d,
1634
    "ovis": run_ovis,
1635
    "ovis2_5": run_ovis2_5,
1636
1637
1638
    "paligemma": run_paligemma,
    "paligemma2": run_paligemma2,
    "phi3_v": run_phi3v,
1639
    "phi4_mm": run_phi4mm,
1640
    "phi4_multimodal": run_phi4_multimodal,
1641
    "pixtral_hf": run_pixtral_hf,
1642
    "qwen_vl": run_qwen_vl,
1643
    "qwen2_vl": run_qwen2_vl,
Roger Wang's avatar
Roger Wang committed
1644
    "qwen2_5_vl": run_qwen2_5_vl,
1645
    "qwen2_5_omni": run_qwen2_5_omni,
1646
    "rvl": run_r_vl,
1647
    "skywork_chat": run_skyworkr1v,
1648
    "smolvlm": run_smolvlm,
1649
    "step3": run_step3,
汪志鹏's avatar
汪志鹏 committed
1650
    "tarsier": run_tarsier,
1651
    "tarsier2": run_tarsier2,
1652
1653
1654
}


1655
1656
1657
1658
1659
1660
1661
1662
1663
def get_multi_modal_input(args):
    """
    return {
        "data": image or video,
        "question": question,
    }
    """
    if args.modality == "image":
        # Input image and question
1664
        image = convert_image_mode(ImageAsset("cherry_blossom").pil_image, "RGB")
1665
1666
1667
1668
1669
1670
        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?",
        ]
1671
1672
1673

        return {
            "data": image,
1674
            "questions": img_questions,
1675
1676
1677
1678
        }

    if args.modality == "video":
        # Input video and question
1679
        video = VideoAsset(name="baby_reading", num_frames=args.num_frames).np_ndarrays
1680
        metadata = VideoAsset(name="baby_reading", num_frames=args.num_frames).metadata
1681
        vid_questions = ["Why is this video funny?"]
1682
1683

        return {
1684
            "data": [(video, metadata)] if args.model_type == "glm4_1v" else video,
1685
            "questions": vid_questions,
1686
1687
1688
1689
1690
1691
        }

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


1692
1693
1694
1695
def apply_image_repeat(
    image_repeat_prob, num_prompts, data, prompts: list[str], modality
):
    """Repeats images with provided probability of "image_repeat_prob".
1696
1697
    Used to simulate hit/miss for the MM preprocessor cache.
    """
1698
    assert image_repeat_prob <= 1.0 and image_repeat_prob >= 0
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
1709
1710
1711
1712
    no_yes = [0, 1]
    probs = [1.0 - image_repeat_prob, image_repeat_prob]

    inputs = []
    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)

1713
1714
1715
1716
        inputs.append(
            {
                "prompt": prompts[i % len(prompts)],
                "multi_modal_data": {modality: cur_image},
1717
            }
1718
        )
1719
1720
1721
1722

    return inputs


1723
1724
1725
1726
@contextmanager
def time_counter(enable: bool):
    if enable:
        import time
1727

1728
1729
1730
1731
1732
1733
1734
1735
1736
1737
        start_time = time.time()
        yield
        elapsed_time = time.time() - start_time
        print("-" * 50)
        print("-- generate time = {}".format(elapsed_time))
        print("-" * 50)
    else:
        yield


1738
1739
def parse_args():
    parser = FlexibleArgumentParser(
1740
1741
1742
1743
1744
1745
1746
1747
1748
1749
1750
1751
1752
1753
1754
1755
1756
1757
1758
1759
1760
1761
1762
1763
1764
1765
1766
1767
1768
1769
1770
1771
1772
        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`.",
    )
1773
1774

    parser.add_argument(
1775
        "--image-repeat-prob",
1776
1777
        type=float,
        default=None,
1778
1779
        help="Simulates the hit-ratio for multi-modal preprocessor cache (if enabled)",
    )
1780
1781

    parser.add_argument(
1782
        "--disable-mm-processor-cache",
1783
        action="store_true",
1784
        help="If True, disables caching of multi-modal processor.",
1785
    )
1786
1787

    parser.add_argument(
1788
1789
1790
1791
        "--time-generate",
        action="store_true",
        help="If True, then print the total generate() call time",
    )
1792
1793

    parser.add_argument(
1794
1795
1796
1797
1798
        "--use-different-prompt-per-request",
        action="store_true",
        help="If True, then use different prompt (with the same multi-modal "
        "data) for each request.",
    )
1799
1800
1801
    return parser.parse_args()


1802
1803
1804
1805
1806
def main(args):
    model = args.model_type
    if model not in model_example_map:
        raise ValueError(f"Model type {model} is not supported.")

1807
1808
1809
    modality = args.modality
    mm_input = get_multi_modal_input(args)
    data = mm_input["data"]
1810
    questions = mm_input["questions"]
1811

1812
1813
    req_data = model_example_map[model](questions, modality)

1814
1815
1816
    # 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(
1817
1818
        req_data.engine_args.limit_mm_per_prompt or {}
    )
1819
1820
1821

    engine_args = asdict(req_data.engine_args) | {
        "seed": args.seed,
1822
        "mm_processor_cache_gb": 0 if args.disable_mm_processor_cache else 4,
1823
    }
1824
1825
    llm = LLM(**engine_args)

1826
    # Don't want to check the flag multiple times, so just hijack `prompts`.
1827
1828
1829
1830
1831
    prompts = (
        req_data.prompts
        if args.use_different_prompt_per_request
        else [req_data.prompts[0]]
    )
1832
1833
1834

    # We set temperature to 0.2 so that outputs can be different
    # even when all prompts are identical when running batch inference.
1835
1836
1837
    sampling_params = SamplingParams(
        temperature=0.2, max_tokens=64, stop_token_ids=req_data.stop_token_ids
    )
1838
1839
1840
1841
1842

    assert args.num_prompts > 0
    if args.num_prompts == 1:
        # Single inference
        inputs = {
1843
            "prompt": prompts[0],
1844
            "multi_modal_data": {modality: data},
1845
1846
1847
        }
    else:
        # Batch inference
1848
1849
        if args.image_repeat_prob is not None:
            # Repeat images with specified probability of "image_repeat_prob"
1850
1851
1852
            inputs = apply_image_repeat(
                args.image_repeat_prob, args.num_prompts, data, prompts, modality
            )
1853
1854
        else:
            # Use the same image for all prompts
1855
1856
1857
1858
1859
1860
1861
            inputs = [
                {
                    "prompt": prompts[i % len(prompts)],
                    "multi_modal_data": {modality: data},
                }
                for i in range(args.num_prompts)
            ]
1862

1863
    # Add LoRA request if applicable
1864
1865
1866
    lora_request = (
        req_data.lora_requests * args.num_prompts if req_data.lora_requests else None
    )
1867

1868
1869
1870
1871
1872
1873
    with time_counter(args.time_generate):
        outputs = llm.generate(
            inputs,
            sampling_params=sampling_params,
            lora_request=lora_request,
        )
1874

1875
    print("-" * 50)
1876
1877
1878
    for o in outputs:
        generated_text = o.outputs[0].text
        print(generated_text)
1879
        print("-" * 50)
1880
1881
1882


if __name__ == "__main__":
1883
    args = parse_args()
1884
    main(args)