vision_language.py 47.7 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
# Deepseek-VL2
130
def run_deepseek_vl2(questions: list[str], modality: str) -> ModelRequestData:
131
132
    assert modality == "image"

133
    model_name = "deepseek-ai/deepseek-vl2-tiny"
134

135
136
137
138
139
    engine_args = EngineArgs(
        model=model_name,
        max_model_len=4096,
        max_num_seqs=2,
        hf_overrides={"architectures": ["DeepseekVLV2ForCausalLM"]},
140
        limit_mm_per_prompt={modality: 1},
141
    )
142

143
    prompts = [
144
        f"<|User|>: <image>\n{question}\n\n<|Assistant|>:" for question in questions
145
    ]
146
147
148
149
150

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
151
152


153
# Florence2
154
def run_florence2(questions: list[str], modality: str) -> ModelRequestData:
155
156
    assert modality == "image"

157
158
    engine_args = EngineArgs(
        model="microsoft/Florence-2-large",
159
        tokenizer="Isotr0py/Florence-2-tokenizer",
160
161
        max_model_len=4096,
        max_num_seqs=2,
162
163
        trust_remote_code=True,
        dtype="bfloat16",
164
        limit_mm_per_prompt={modality: 1},
165
    )
166

167
168
169
170
171
172
    prompts = ["<MORE_DETAILED_CAPTION>" for _ in questions]

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
173
174


175
# Fuyu
176
def run_fuyu(questions: list[str], modality: str) -> ModelRequestData:
177
178
    assert modality == "image"

179
    prompts = [f"{question}\n" for question in questions]
180
181
182
183
    engine_args = EngineArgs(
        model="adept/fuyu-8b",
        max_model_len=2048,
        max_num_seqs=2,
184
        limit_mm_per_prompt={modality: 1},
185
186
187
188
189
190
    )

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


193
# Gemma 3
194
def run_gemma3(questions: list[str], modality: str) -> ModelRequestData:
195
196
197
    assert modality == "image"
    model_name = "google/gemma-3-4b-it"

198
    engine_args = EngineArgs(
199
200
201
202
        model=model_name,
        max_model_len=2048,
        max_num_seqs=2,
        mm_processor_kwargs={"do_pan_and_scan": True},
203
        limit_mm_per_prompt={modality: 1},
204
    )
205

206
207
208
209
210
211
212
213
    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
214
215
216
217
218
    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )

219

Nicolò Lucchesi's avatar
Nicolò Lucchesi committed
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
# 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
    ]
241
242
243
244
    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
245
246


247
# GLM-4v
248
def run_glm4v(questions: list[str], modality: str) -> ModelRequestData:
249
    assert modality == "image"
250
    model_name = "zai-org/glm-4v-9b"
251

252
253
254
255
256
257
258
    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"]},
259
        limit_mm_per_prompt={modality: 1},
260
    )
261

262
263
    prompts = [
        f"<|user|>\n<|begin_of_image|><|endoftext|><|end_of_image|>\
264
265
        {question}<|assistant|>"
        for question in questions
266
    ]
267

268
    stop_token_ids = [151329, 151336, 151338]
269
270
271
272
273
274

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
        stop_token_ids=stop_token_ids,
    )
275
276


277
278
# GLM-4.1V
def run_glm4_1v(questions: list[str], modality: str) -> ModelRequestData:
279
    model_name = "zai-org/GLM-4.1V-9B-Thinking"
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312

    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,
    )


313
# H2OVL-Mississippi
314
def run_h2ovl(questions: list[str], modality: str) -> ModelRequestData:
315
316
    assert modality == "image"

317
    model_name = "h2oai/h2ovl-mississippi-800m"
318

319
    engine_args = EngineArgs(
320
321
322
        model=model_name,
        trust_remote_code=True,
        max_model_len=8192,
323
        limit_mm_per_prompt={modality: 1},
324
325
    )

326
327
328
329
330
331
332
    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
    )
333
334

    # Stop tokens for H2OVL-Mississippi
335
    # https://huggingface.co/h2oai/h2ovl-mississippi-800m
336
    stop_token_ids = [tokenizer.eos_token_id]
337
338
339
340
341
342

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
        stop_token_ids=stop_token_ids,
    )
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
412
413
414
415
416
417
418
419
420
421
422
423
# 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":
            """
            ocr: List the words in the image in raster order. 
                Even if the word order feels unnatural for reading, 
                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,
    )


424
# Idefics3-8B-Llama3
425
def run_idefics3(questions: list[str], modality: str) -> ModelRequestData:
426
427
428
    assert modality == "image"
    model_name = "HuggingFaceM4/Idefics3-8B-Llama3"

429
    engine_args = EngineArgs(
430
431
432
433
434
435
436
        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={
437
            "size": {"longest_edge": 3 * 364},
438
        },
439
        limit_mm_per_prompt={modality: 1},
440
    )
441
442
443
444
    prompts = [
        (f"<|begin_of_text|>User:<image>{question}<end_of_utterance>\nAssistant:")
        for question in questions
    ]
445
446
447
448
449

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
450
451


Lyu Han's avatar
Lyu Han committed
452
453
454
455
456
457
458
459
460
461
462
463
464
# 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,
    )

465
466
467
468
469
    if modality == "image":
        placeholder = "<IMG_CONTEXT>"
    elif modality == "video":
        placeholder = "<video>"

Lyu Han's avatar
Lyu Han committed
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
    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,
    )


485
# InternVL
486
def run_internvl(questions: list[str], modality: str) -> ModelRequestData:
487
    model_name = "OpenGVLab/InternVL3-2B"
488

489
    engine_args = EngineArgs(
490
491
        model=model_name,
        trust_remote_code=True,
492
        max_model_len=8192,
493
        limit_mm_per_prompt={modality: 1},
494
495
    )

496
497
498
499
500
    if modality == "image":
        placeholder = "<image>"
    elif modality == "video":
        placeholder = "<video>"

501
502
503
504
505
506
507
508
    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
    )
509
510
511
512
513
514
515

    # 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]
516
    stop_token_ids = [token_id for token_id in stop_token_ids if token_id is not None]
517
518
519
520
521
522

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
        stop_token_ids=stop_token_ids,
    )
523
524


525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
# 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,
    )


556
557
558
559
560
561
562
# 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|>"
563
564
        "<|im_assistant|>assistant<|im_middle|>"
        for question in questions
565
566
567
568
569
    ]

    engine_args = EngineArgs(
        model="moonshotai/Kimi-VL-A3B-Instruct",
        trust_remote_code=True,
Cyrus Leung's avatar
Cyrus Leung committed
570
        max_model_len=4096,
571
        limit_mm_per_prompt={modality: 1},
572
573
574
575
576
577
578
579
    )

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


580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
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,
    )


615
# LLaVA-1.5
616
def run_llava(questions: list[str], modality: str) -> ModelRequestData:
617
    assert modality == "image"
618

619
    prompts = [f"USER: <image>\n{question}\nASSISTANT:" for question in questions]
620

621
622
623
    engine_args = EngineArgs(
        model="llava-hf/llava-1.5-7b-hf",
        max_model_len=4096,
624
        limit_mm_per_prompt={modality: 1},
625
626
627
628
629
630
    )

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
631
632
633


# LLaVA-1.6/LLaVA-NeXT
634
def run_llava_next(questions: list[str], modality: str) -> ModelRequestData:
635
    assert modality == "image"
636

637
    prompts = [f"[INST] <image>\n{question} [/INST]" for question in questions]
638
639
640
    engine_args = EngineArgs(
        model="llava-hf/llava-v1.6-mistral-7b-hf",
        max_model_len=8192,
641
        limit_mm_per_prompt={modality: 1},
642
643
644
645
646
647
    )

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
648
649
650
651


# LlaVA-NeXT-Video
# Currently only support for video input
652
def run_llava_next_video(questions: list[str], modality: str) -> ModelRequestData:
653
654
    assert modality == "video"

655
    prompts = [f"USER: <video>\n{question} ASSISTANT:" for question in questions]
656
657
658
    engine_args = EngineArgs(
        model="llava-hf/LLaVA-NeXT-Video-7B-hf",
        max_model_len=8192,
659
        max_num_seqs=2,
660
        limit_mm_per_prompt={modality: 1},
661
662
663
664
665
666
    )

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
667
668


669
# LLaVA-OneVision
670
def run_llava_onevision(questions: list[str], modality: str) -> ModelRequestData:
671
    if modality == "video":
672
673
        prompts = [
            f"<|im_start|>user <video>\n{question}<|im_end|> \
674
675
        <|im_start|>assistant\n"
            for question in questions
676
        ]
677
678

    elif modality == "image":
679
680
        prompts = [
            f"<|im_start|>user <image>\n{question}<|im_end|> \
681
682
        <|im_start|>assistant\n"
            for question in questions
683
        ]
684

685
686
687
    engine_args = EngineArgs(
        model="llava-hf/llava-onevision-qwen2-7b-ov-hf",
        max_model_len=16384,
688
        limit_mm_per_prompt={modality: 1},
689
690
691
692
693
694
    )

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
695
696


697
# Mantis
698
def run_mantis(questions: list[str], modality: str) -> ModelRequestData:
699
    assert modality == "image"
700

701
702
    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]
703

704
    engine_args = EngineArgs(
705
        model="TIGER-Lab/Mantis-8B-siglip-llama3",
706
        max_model_len=4096,
707
        hf_overrides={"architectures": ["MantisForConditionalGeneration"]},
708
        limit_mm_per_prompt={modality: 1},
709
    )
710
    stop_token_ids = [128009]
711
712
713
714
715
716

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
        stop_token_ids=stop_token_ids,
    )
717
718
719


# MiniCPM-V
720
def run_minicpmv_base(questions: list[str], modality: str, model_name):
721
722
    assert modality in ["image", "video"]
    # If you want to use `MiniCPM-o-2_6` with audio inputs, check `audio_language.py` # noqa
723
724
725
726
727
728
729

    # 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
730
731
    # model_name = "openbmb/MiniCPM-Llama3-V-2_5"

732
    # 2.6
733
734
735
736
737
738
739
740
741
    # 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"
742
    tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
743
    engine_args = EngineArgs(
744
        model=model_name,
745
746
        max_model_len=4096,
        max_num_seqs=2,
747
        trust_remote_code=True,
748
        limit_mm_per_prompt={modality: 1},
749
    )
750
751
752
753
754
755
756
    # 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]

757
    # 2.6 / o2.6
758
    stop_tokens = ["<|im_end|>", "<|endoftext|>"]
759
    stop_token_ids = [tokenizer.convert_tokens_to_ids(i) for i in stop_tokens]
760

761
762
763
764
765
    modality_placeholder = {
        "image": "(<image>./</image>)",
        "video": "(<video>./</video>)",
    }

766
767
    prompts = [
        tokenizer.apply_chat_template(
768
769
770
771
772
773
            [
                {
                    "role": "user",
                    "content": f"{modality_placeholder[modality]}\n{question}",
                }
            ],
774
            tokenize=False,
775
776
777
            add_generation_prompt=True,
        )
        for question in questions
778
    ]
779
780
781
782
783
784

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
        stop_token_ids=stop_token_ids,
    )
785
786


787
def run_minicpmo(questions: list[str], modality: str) -> ModelRequestData:
788
    return run_minicpmv_base(questions, modality, "openbmb/MiniCPM-o-2_6")
789
790


791
def run_minicpmv(questions: list[str], modality: str) -> ModelRequestData:
792
    return run_minicpmv_base(questions, modality, "openbmb/MiniCPM-V-2_6")
793
794


795
796
797
798
799
800
801
802
803
804
805
806
# 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,
807
        limit_mm_per_prompt={modality: 1},
808
        ignore_patterns=["consolidated.safetensors"],
809
810
811
812
813
814
815
816
817
818
    )

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

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


819
# LLama 3.2
820
def run_mllama(questions: list[str], modality: str) -> ModelRequestData:
821
822
    assert modality == "image"

823
    model_name = "meta-llama/Llama-3.2-11B-Vision-Instruct"
824

825
826
827
828
829
    # 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.
830
    engine_args = EngineArgs(
831
        model=model_name,
832
        max_model_len=8192,
833
        max_num_seqs=2,
834
        limit_mm_per_prompt={modality: 1},
835
836
    )

837
    tokenizer = AutoTokenizer.from_pretrained(model_name)
838
839
840
841
842
843
844
845
846
847
848
849
    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
    )
850
851
852
853
854

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
855
856


857
858
# Molmo
def run_molmo(questions: list[str], modality: str) -> ModelRequestData:
859
860
    assert modality == "image"

861
    model_name = "allenai/Molmo-7B-D-0924"
862
863
864

    engine_args = EngineArgs(
        model=model_name,
865
866
        trust_remote_code=True,
        dtype="bfloat16",
867
        limit_mm_per_prompt={modality: 1},
868
869
    )

870
871
872
    prompts = [
        f"<|im_start|>user <image>\n{question}<|im_end|> \
        <|im_start|>assistant\n"
873
874
        for question in questions
    ]
875

876
877
878
879
880
881
    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )


882
883
884
# Nemontron_VL
def run_nemotron_vl(questions: list[str], modality: str) -> ModelRequestData:
    model_name = "nvidia/Llama-3.1-Nemotron-Nano-VL-8B-V1"
885

886
    engine_args = EngineArgs(
887
        model=model_name,
888
        trust_remote_code=True,
889
        max_model_len=8192,
890
        limit_mm_per_prompt={modality: 1},
891
    )
892

893
894
895
896
897
898
    assert modality == "image"
    placeholder = "<image>"

    tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
    messages = [
        [{"role": "user", "content": f"{placeholder}\n{question}"}]
899
        for question in questions
900
    ]
901
902
903
904
905
906
907
908
909
910
911
    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]
912
913
914
915

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
916
        stop_token_ids=stop_token_ids,
917
    )
918
919


920
# NVLM-D
921
def run_nvlm_d(questions: list[str], modality: str) -> ModelRequestData:
922
923
924
925
926
    assert modality == "image"

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

    # Adjust this as necessary to fit in GPU
927
    engine_args = EngineArgs(
928
929
930
931
        model=model_name,
        trust_remote_code=True,
        max_model_len=4096,
        tensor_parallel_size=4,
932
        limit_mm_per_prompt={modality: 1},
933
934
    )

935
936
937
938
939
940
941
    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
    )
942
943
944
945
946

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
947
948


949
950
# Ovis
def run_ovis(questions: list[str], modality: str) -> ModelRequestData:
951
952
953
954
955
956
957
958
959
960
    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",
961
        limit_mm_per_prompt={modality: 1},
962
963
    )

964
965
966
967
968
969
970
    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
    )
971
972
973
974
975
976
977

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


978
# PaliGemma
979
def run_paligemma(questions: list[str], modality: str) -> ModelRequestData:
980
    assert modality == "image"
981

982
    # PaliGemma has special prompt format for VQA
983
984
985
    prompts = ["caption en" for _ in questions]
    engine_args = EngineArgs(
        model="google/paligemma-3b-mix-224",
986
        limit_mm_per_prompt={modality: 1},
987
    )
988
989
990
991
992

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
993
994


995
# PaliGemma 2
996
def run_paligemma2(questions: list[str], modality: str) -> ModelRequestData:
997
    assert modality == "image"
998

999
    # PaliGemma 2 has special prompt format for VQA
1000
1001
1002
    prompts = ["caption en" for _ in questions]
    engine_args = EngineArgs(
        model="google/paligemma2-3b-ft-docci-448",
1003
        limit_mm_per_prompt={modality: 1},
1004
    )
1005
1006
1007
1008
1009

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
1010
1011


1012
# Phi-3-Vision
1013
def run_phi3v(questions: list[str], modality: str) -> ModelRequestData:
1014
1015
    assert modality == "image"

1016
1017
1018
1019
    prompts = [
        f"<|user|>\n<|image_1|>\n{question}<|end|>\n<|assistant|>\n"
        for question in questions
    ]
1020

1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
    # 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
1033
    engine_args = EngineArgs(
1034
1035
        model="microsoft/Phi-3.5-vision-instruct",
        trust_remote_code=True,
1036
        max_model_len=4096,
1037
        max_num_seqs=2,
1038
        # Note - mm_processor_kwargs can also be passed to generate/chat calls
1039
        mm_processor_kwargs={"num_crops": 16},
1040
        limit_mm_per_prompt={modality: 1},
1041
    )
1042
1043
1044
1045
1046

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
1047
1048


1049
# Phi-4-multimodal-instruct
1050
def run_phi4mm(questions: list[str], modality: str) -> ModelRequestData:
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
    """
    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 = [
1061
        f"<|user|><|image_1|>{question}<|end|><|assistant|>" for question in questions
1062
    ]
1063
    engine_args = EngineArgs(
1064
1065
        model=model_path,
        trust_remote_code=True,
1066
        max_model_len=5120,
1067
        max_num_seqs=2,
1068
        max_num_batched_tokens=12800,
1069
1070
        enable_lora=True,
        max_lora_rank=320,
1071
1072
        # Note - mm_processor_kwargs can also be passed to generate/chat calls
        mm_processor_kwargs={"dynamic_hd": 16},
1073
        limit_mm_per_prompt={modality: 1},
1074
1075
    )

1076
1077
1078
1079
1080
    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
        lora_requests=[LoRARequest("vision", 1, vision_lora_path)],
    )
1081
1082


1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
# 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)],
    )


1118
# Pixtral HF-format
1119
def run_pixtral_hf(questions: list[str], modality: str) -> ModelRequestData:
1120
1121
1122
1123
    assert modality == "image"

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

1124
    # NOTE: Need L40 (or equivalent) to avoid OOM
1125
    engine_args = EngineArgs(
1126
        model=model_name,
1127
        max_model_len=6144,
1128
        max_num_seqs=2,
1129
        limit_mm_per_prompt={modality: 1},
1130
1131
    )

1132
    prompts = [f"<s>[INST]{question}\n[IMG][/INST]" for question in questions]
1133
1134
1135
1136
1137

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
1138
1139


1140
# Qwen-VL
1141
def run_qwen_vl(questions: list[str], modality: str) -> ModelRequestData:
1142
1143
    assert modality == "image"

1144
    engine_args = EngineArgs(
1145
        model="Qwen/Qwen-VL",
1146
        trust_remote_code=True,
1147
1148
        max_model_len=1024,
        max_num_seqs=2,
1149
        hf_overrides={"architectures": ["QwenVLForConditionalGeneration"]},
1150
        limit_mm_per_prompt={modality: 1},
1151
1152
    )

1153
    prompts = [f"{question}Picture 1: <img></img>\n" for question in questions]
1154
1155
1156
1157
1158

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
1159
1160


1161
# Qwen2-VL
1162
def run_qwen2_vl(questions: list[str], modality: str) -> ModelRequestData:
1163
    model_name = "Qwen/Qwen2-VL-7B-Instruct"
1164

1165
    engine_args = EngineArgs(
1166
        model=model_name,
1167
1168
1169
        max_model_len=4096,
        max_num_seqs=5,
        # Note - mm_processor_kwargs can also be passed to generate/chat calls
1170
        mm_processor_kwargs={
1171
1172
            "min_pixels": 28 * 28,
            "max_pixels": 1280 * 28 * 28,
1173
        },
1174
        limit_mm_per_prompt={modality: 1},
1175
    )
1176

1177
1178
1179
1180
1181
    if modality == "image":
        placeholder = "<|image_pad|>"
    elif modality == "video":
        placeholder = "<|video_pad|>"

1182
    prompts = [
1183
1184
1185
1186
1187
1188
1189
        (
            "<|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
1190
    ]
1191
1192
1193
1194
1195

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
1196
1197


Roger Wang's avatar
Roger Wang committed
1198
# Qwen2.5-VL
1199
def run_qwen2_5_vl(questions: list[str], modality: str) -> ModelRequestData:
Roger Wang's avatar
Roger Wang committed
1200
1201
    model_name = "Qwen/Qwen2.5-VL-3B-Instruct"

1202
    engine_args = EngineArgs(
Roger Wang's avatar
Roger Wang committed
1203
1204
1205
1206
1207
1208
1209
1210
        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,
        },
1211
        limit_mm_per_prompt={modality: 1},
Roger Wang's avatar
Roger Wang committed
1212
1213
1214
1215
1216
1217
1218
    )

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

1219
    prompts = [
1220
1221
1222
1223
1224
1225
1226
        (
            "<|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
1227
    ]
1228
1229
1230
1231
1232

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
Roger Wang's avatar
Roger Wang committed
1233
1234


1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
# 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],
        },
1248
        limit_mm_per_prompt={modality: 1},
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
    )

    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 "
1259
1260
        "generating text and speech."
    )
1261

1262
1263
1264
1265
1266
1267
1268
1269
1270
    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
    ]
1271
1272
1273
1274
1275
1276
    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )


1277
1278
# SkyworkR1V
def run_skyworkr1v(questions: list[str], modality: str) -> ModelRequestData:
1279
    assert modality == "image"
1280
1281

    model_name = "Skywork/Skywork-R1V-38B"
1282
1283
1284
1285
1286
1287
1288

    engine_args = EngineArgs(
        model=model_name,
        trust_remote_code=True,
        max_model_len=4096,
        limit_mm_per_prompt={modality: 1},
    )
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301

    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]
1302
1303
1304
1305

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
1306
        stop_token_ids=stop_token_ids,
1307
1308
1309
    )


1310
1311
1312
1313
# SmolVLM2-2.2B-Instruct
def run_smolvlm(questions: list[str], modality: str) -> ModelRequestData:
    assert modality == "image"
    model_name = "HuggingFaceTB/SmolVLM2-2.2B-Instruct"
1314
1315
1316

    engine_args = EngineArgs(
        model=model_name,
1317
1318
1319
1320
1321
1322
        max_model_len=8192,
        max_num_seqs=2,
        enforce_eager=True,
        mm_processor_kwargs={
            "max_image_size": {"longest_edge": 384},
        },
1323
1324
        limit_mm_per_prompt={modality: 1},
    )
1325
1326
1327
1328
    prompts = [
        (f"<|im_start|>User:<image>{question}<end_of_utterance>\nAssistant:")
        for question in questions
    ]
1329

1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
    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",
    )
1352
1353

    prompts = [
1354
1355
        "<|begin▁of▁sentence|> You are a helpful assistant. <|BOT|>user\n "
        f"<im_patch>{question} <|EOT|><|BOT|>assistant\n<think>\n"
1356
1357
1358
1359
1360
1361
1362
1363
1364
        for question in questions
    ]

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


1365
1366
# omni-research/Tarsier-7b
def run_tarsier(questions: list[str], modality: str) -> ModelRequestData:
1367
    assert modality == "image"
1368
    model_name = "omni-research/Tarsier-7b"
1369
1370
1371
1372
1373

    engine_args = EngineArgs(
        model=model_name,
        trust_remote_code=True,
        max_model_len=4096,
1374
        limit_mm_per_prompt={modality: 1},
1375
    )
1376
    prompts = [(f"USER: <image>\n{question} ASSISTANT:") for question in questions]
1377

1378
1379
1380
    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
1381
    )
1382

1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407

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
    ]
1408
1409
1410
1411
1412
1413
1414

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


1415
model_example_map = {
1416
    "aria": run_aria,
Jennifer Zhao's avatar
Jennifer Zhao committed
1417
    "aya_vision": run_aya_vision,
1418
1419
    "blip-2": run_blip2,
    "chameleon": run_chameleon,
1420
    "deepseek_vl_v2": run_deepseek_vl2,
1421
    "florence2": run_florence2,
1422
    "fuyu": run_fuyu,
1423
    "gemma3": run_gemma3,
Nicolò Lucchesi's avatar
Nicolò Lucchesi committed
1424
    "gemma3n": run_gemma3n,
1425
    "glm4v": run_glm4v,
1426
    "glm4_1v": run_glm4_1v,
1427
    "h2ovl_chat": run_h2ovl,
1428
    "hyperclovax_seed_vision": run_hyperclovax_seed_vision,
1429
    "idefics3": run_idefics3,
Lyu Han's avatar
Lyu Han committed
1430
    "interns1": run_interns1,
1431
    "internvl_chat": run_internvl,
1432
    "keye_vl": run_keye_vl,
1433
    "kimi_vl": run_kimi_vl,
1434
    "llama4": run_llama4,
1435
1436
    "llava": run_llava,
    "llava-next": run_llava_next,
1437
    "llava-next-video": run_llava_next_video,
1438
    "llava-onevision": run_llava_onevision,
1439
    "mantis": run_mantis,
1440
    "minicpmo": run_minicpmo,
1441
    "minicpmv": run_minicpmv,
1442
    "mistral3": run_mistral3,
1443
1444
    "mllama": run_mllama,
    "molmo": run_molmo,
1445
    "nemotron_vl": run_nemotron_vl,
1446
    "NVLM_D": run_nvlm_d,
1447
    "ovis": run_ovis,
1448
1449
1450
    "paligemma": run_paligemma,
    "paligemma2": run_paligemma2,
    "phi3_v": run_phi3v,
1451
    "phi4_mm": run_phi4mm,
1452
    "phi4_multimodal": run_phi4_multimodal,
1453
    "pixtral_hf": run_pixtral_hf,
1454
    "qwen_vl": run_qwen_vl,
1455
    "qwen2_vl": run_qwen2_vl,
Roger Wang's avatar
Roger Wang committed
1456
    "qwen2_5_vl": run_qwen2_5_vl,
1457
    "qwen2_5_omni": run_qwen2_5_omni,
1458
    "skywork_chat": run_skyworkr1v,
1459
    "smolvlm": run_smolvlm,
1460
    "step3": run_step3,
汪志鹏's avatar
汪志鹏 committed
1461
    "tarsier": run_tarsier,
1462
    "tarsier2": run_tarsier2,
1463
1464
1465
}


1466
1467
1468
1469
1470
1471
1472
1473
1474
def get_multi_modal_input(args):
    """
    return {
        "data": image or video,
        "question": question,
    }
    """
    if args.modality == "image":
        # Input image and question
1475
        image = convert_image_mode(ImageAsset("cherry_blossom").pil_image, "RGB")
1476
1477
1478
1479
1480
1481
        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?",
        ]
1482
1483
1484

        return {
            "data": image,
1485
            "questions": img_questions,
1486
1487
1488
1489
        }

    if args.modality == "video":
        # Input video and question
1490
        video = VideoAsset(name="baby_reading", num_frames=args.num_frames).np_ndarrays
1491
        metadata = VideoAsset(name="baby_reading", num_frames=args.num_frames).metadata
1492
        vid_questions = ["Why is this video funny?"]
1493
1494

        return {
1495
            "data": [(video, metadata)] if args.model_type == "glm4_1v" else video,
1496
            "questions": vid_questions,
1497
1498
1499
1500
1501
1502
        }

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


1503
1504
1505
1506
def apply_image_repeat(
    image_repeat_prob, num_prompts, data, prompts: list[str], modality
):
    """Repeats images with provided probability of "image_repeat_prob".
1507
1508
    Used to simulate hit/miss for the MM preprocessor cache.
    """
1509
    assert image_repeat_prob <= 1.0 and image_repeat_prob >= 0
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
    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)

1524
1525
1526
1527
        inputs.append(
            {
                "prompt": prompts[i % len(prompts)],
                "multi_modal_data": {modality: cur_image},
1528
            }
1529
        )
1530
1531
1532
1533

    return inputs


1534
1535
1536
1537
@contextmanager
def time_counter(enable: bool):
    if enable:
        import time
1538

1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
        start_time = time.time()
        yield
        elapsed_time = time.time() - start_time
        print("-" * 50)
        print("-- generate time = {}".format(elapsed_time))
        print("-" * 50)
    else:
        yield


1549
1550
def parse_args():
    parser = FlexibleArgumentParser(
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
        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`.",
    )
1584
1585

    parser.add_argument(
1586
        "--image-repeat-prob",
1587
1588
        type=float,
        default=None,
1589
1590
        help="Simulates the hit-ratio for multi-modal preprocessor cache (if enabled)",
    )
1591
1592

    parser.add_argument(
1593
        "--disable-mm-processor-cache",
1594
        action="store_true",
1595
        help="If True, disables caching of multi-modal processor.",
1596
    )
1597
1598

    parser.add_argument(
1599
1600
1601
1602
        "--time-generate",
        action="store_true",
        help="If True, then print the total generate() call time",
    )
1603
1604

    parser.add_argument(
1605
1606
1607
1608
1609
        "--use-different-prompt-per-request",
        action="store_true",
        help="If True, then use different prompt (with the same multi-modal "
        "data) for each request.",
    )
1610
1611
1612
    return parser.parse_args()


1613
1614
1615
1616
1617
def main(args):
    model = args.model_type
    if model not in model_example_map:
        raise ValueError(f"Model type {model} is not supported.")

1618
1619
1620
    modality = args.modality
    mm_input = get_multi_modal_input(args)
    data = mm_input["data"]
1621
    questions = mm_input["questions"]
1622

1623
1624
    req_data = model_example_map[model](questions, modality)

1625
1626
1627
    # 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(
1628
1629
        req_data.engine_args.limit_mm_per_prompt or {}
    )
1630
1631
1632

    engine_args = asdict(req_data.engine_args) | {
        "seed": args.seed,
1633
        "mm_processor_cache_gb": 0 if args.disable_mm_processor_cache else 4,
1634
    }
1635
1636
    llm = LLM(**engine_args)

1637
    # Don't want to check the flag multiple times, so just hijack `prompts`.
1638
1639
1640
1641
1642
    prompts = (
        req_data.prompts
        if args.use_different_prompt_per_request
        else [req_data.prompts[0]]
    )
1643
1644
1645

    # We set temperature to 0.2 so that outputs can be different
    # even when all prompts are identical when running batch inference.
1646
1647
1648
    sampling_params = SamplingParams(
        temperature=0.2, max_tokens=64, stop_token_ids=req_data.stop_token_ids
    )
1649
1650
1651
1652
1653

    assert args.num_prompts > 0
    if args.num_prompts == 1:
        # Single inference
        inputs = {
1654
            "prompt": prompts[0],
1655
            "multi_modal_data": {modality: data},
1656
1657
1658
        }
    else:
        # Batch inference
1659
1660
        if args.image_repeat_prob is not None:
            # Repeat images with specified probability of "image_repeat_prob"
1661
1662
1663
            inputs = apply_image_repeat(
                args.image_repeat_prob, args.num_prompts, data, prompts, modality
            )
1664
1665
        else:
            # Use the same image for all prompts
1666
1667
1668
1669
1670
1671
1672
            inputs = [
                {
                    "prompt": prompts[i % len(prompts)],
                    "multi_modal_data": {modality: data},
                }
                for i in range(args.num_prompts)
            ]
1673

1674
    # Add LoRA request if applicable
1675
1676
1677
    lora_request = (
        req_data.lora_requests * args.num_prompts if req_data.lora_requests else None
    )
1678

1679
1680
1681
1682
1683
1684
    with time_counter(args.time_generate):
        outputs = llm.generate(
            inputs,
            sampling_params=sampling_params,
            lora_request=lora_request,
        )
1685

1686
    print("-" * 50)
1687
1688
1689
    for o in outputs:
        generated_text = o.outputs[0].text
        print(generated_text)
1690
        print("-" * 50)
1691
1692
1693


if __name__ == "__main__":
1694
    args = parse_args()
1695
    main(args)