vision_language.py 70.7 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3
"""
Cyrus Leung's avatar
Cyrus Leung committed
4
5
This example shows how to use vLLM for running offline inference with
the correct prompt format on vision language models for text generation.
6
7
8
9

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

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

17
from huggingface_hub import snapshot_download
18
from transformers import AutoProcessor, AutoTokenizer
19

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
from vllm.utils.argparse_utils import FlexibleArgumentParser
26

27
28
29
30

class ModelRequestData(NamedTuple):
    engine_args: EngineArgs
    prompts: list[str]
31
32
    stop_token_ids: list[int] | None = None
    lora_requests: list[LoRARequest] | None = None
33
    sampling_params: list[SamplingParams] | None = None
34
35


36
37
38
39
# 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.

40

41
# Aria
42
def run_aria(questions: list[str], modality: str) -> ModelRequestData:
43
44
45
    assert modality == "image"
    model_name = "rhymes-ai/Aria"

46
    # NOTE: Need L40 (or equivalent) to avoid OOM
47
48
49
50
51
    engine_args = EngineArgs(
        model=model_name,
        max_model_len=4096,
        max_num_seqs=2,
        dtype="bfloat16",
52
        limit_mm_per_prompt={modality: 1},
53
    )
54

55
56
57
58
59
60
61
    prompts = [
        (
            f"<|im_start|>user\n<fim_prefix><|img|><fim_suffix>{question}"
            "<|im_end|>\n<|im_start|>assistant\n"
        )
        for question in questions
    ]
62
63

    stop_token_ids = [93532, 93653, 944, 93421, 1019, 93653, 93519]
64
65
66
67
68
69

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
        stop_token_ids=stop_token_ids,
    )
70
71


Jennifer Zhao's avatar
Jennifer Zhao committed
72
73
74
# Aya Vision
def run_aya_vision(questions: list[str], modality: str) -> ModelRequestData:
    assert modality == "image"
75
    model_name = "CohereLabs/aya-vision-8b"
Jennifer Zhao's avatar
Jennifer Zhao committed
76
77
78
79
80
81

    engine_args = EngineArgs(
        model=model_name,
        max_model_len=2048,
        max_num_seqs=2,
        mm_processor_kwargs={"crop_to_patches": True},
82
        limit_mm_per_prompt={modality: 1},
Jennifer Zhao's avatar
Jennifer Zhao committed
83
84
85
86
87
88
89
90
91
92
93
    )
    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,
    )


94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
# Bee-8B
def run_bee(questions: list[str], modality: str) -> ModelRequestData:
    assert modality == "image"
    model_name = "Open-Bee/Bee-8B-RL"

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

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

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


121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
def run_bagel(questions: list[str], modality: str) -> ModelRequestData:
    assert modality == "image"
    model_name = "ByteDance-Seed/BAGEL-7B-MoT"

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

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

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


147
# BLIP-2
148
def run_blip2(questions: list[str], modality: str) -> ModelRequestData:
149
150
151
152
    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
153
    prompts = [f"Question: {question} Answer:" for question in questions]
154
    engine_args = EngineArgs(
155
        model="Salesforce/blip2-opt-2.7b",
156
        limit_mm_per_prompt={modality: 1},
157
158
159
160
161
162
    )

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
163
164
165


# Chameleon
166
def run_chameleon(questions: list[str], modality: str) -> ModelRequestData:
167
168
    assert modality == "image"

169
    prompts = [f"{question}<image>" for question in questions]
170
171
172
173
    engine_args = EngineArgs(
        model="facebook/chameleon-7b",
        max_model_len=4096,
        max_num_seqs=2,
174
        limit_mm_per_prompt={modality: 1},
175
176
177
178
179
180
    )

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
181
182


183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
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,
    )


206
# Deepseek-VL2
207
def run_deepseek_vl2(questions: list[str], modality: str) -> ModelRequestData:
208
209
    assert modality == "image"

210
    model_name = "deepseek-ai/deepseek-vl2-tiny"
211

212
213
214
215
216
    engine_args = EngineArgs(
        model=model_name,
        max_model_len=4096,
        max_num_seqs=2,
        hf_overrides={"architectures": ["DeepseekVLV2ForCausalLM"]},
217
        limit_mm_per_prompt={modality: 1},
218
    )
219

220
    prompts = [
221
        f"<|User|>: <image>\n{question}\n\n<|Assistant|>:" for question in questions
222
    ]
223
224
225
226
227

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
228
229


230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
def run_deepseek_ocr(questions: list[str], modality: str) -> ModelRequestData:
    from vllm.model_executor.models.deepseek_ocr import NGramPerReqLogitsProcessor

    assert modality == "image"

    model_name = "deepseek-ai/DeepSeek-OCR"

    engine_args = EngineArgs(
        model=model_name,
        limit_mm_per_prompt={modality: 1},
        logits_processors=[NGramPerReqLogitsProcessor],
    )

    # deepseek-ocr use plain prompt template
    prompts = [f"<image>\n{question}" for question in questions]

    # The following sampling params config is taken from
    # the official Deepseek-OCR inference example.
    # (IMPORTANT) Use the custom logits processor and avoid skipping
    # special tokens for this model for the optimal OCR performance.
    sampling_params = [
        SamplingParams(
            temperature=0.0,
            max_tokens=8192,
            # ngram logit processor args
            extra_args=dict(
                ngram_size=30,
                window_size=90,
                # whitelist: <td>, </td>
                whitelist_token_ids={128821, 128822},
            ),
            skip_special_tokens=False,
        )
        for _ in questions
    ]

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


RED's avatar
RED committed
273
274
275
276
277
278
279
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
313
314
315
def run_deepseek_ocr2(questions: list[str], modality: str) -> ModelRequestData:
    from vllm.model_executor.models.deepseek_ocr import NGramPerReqLogitsProcessor

    assert modality == "image"

    model_name = "deepseek-ai/DeepSeek-OCR-2"

    engine_args = EngineArgs(
        model=model_name,
        limit_mm_per_prompt={modality: 1},
        logits_processors=[NGramPerReqLogitsProcessor],
    )

    # deepseek-ocr use plain prompt template
    prompts = [f"<image>\n{question}" for question in questions]

    # The following sampling params config is taken from
    # the official Deepseek-OCR inference example.
    # (IMPORTANT) Use the custom logits processor and avoid skipping
    # special tokens for this model for the optimal OCR performance.
    sampling_params = [
        SamplingParams(
            temperature=0.0,
            max_tokens=8192,
            # ngram logit processor args
            extra_args=dict(
                ngram_size=30,
                window_size=90,
                # whitelist: <td>, </td>
                whitelist_token_ids={128821, 128822},
            ),
            skip_special_tokens=False,
        )
        for _ in questions
    ]

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


316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
# Dots-OCR
def run_dots_ocr(questions: list[str], modality: str) -> ModelRequestData:
    assert modality == "image"

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

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


333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
# Eagle2.5-VL
def run_eagle2_5(questions: list[str], modality: str) -> ModelRequestData:
    assert modality == "image"

    model_name = "nvidia/Eagle2.5-8B"

    engine_args = EngineArgs(
        model=model_name,
        max_model_len=4096,
        max_num_seqs=2,
        trust_remote_code=True,
        limit_mm_per_prompt={modality: 1},
    )

    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 Eagle2.5 (Qwen2 based)
    stop_tokens = ["<|endoftext|>", "<|im_start|>", "<|im_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]

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


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

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

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

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

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


398
# Fuyu
399
def run_fuyu(questions: list[str], modality: str) -> ModelRequestData:
400
401
    assert modality == "image"

402
    prompts = [f"{question}\n" for question in questions]
403
404
405
406
    engine_args = EngineArgs(
        model="adept/fuyu-8b",
        max_model_len=2048,
        max_num_seqs=2,
407
        limit_mm_per_prompt={modality: 1},
408
409
410
411
412
413
    )

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
414
415


416
# Gemma 3
417
def run_gemma3(questions: list[str], modality: str) -> ModelRequestData:
418
419
420
    assert modality == "image"
    model_name = "google/gemma-3-4b-it"

421
    engine_args = EngineArgs(
422
423
424
        model=model_name,
        max_model_len=2048,
        max_num_seqs=2,
425
        mm_processor_kwargs={"do_pan_and_scan": True},
426
        limit_mm_per_prompt={modality: 1},
427
    )
428

429
430
431
432
433
434
435
436
    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
437
438
439
440
441
    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )

442

Nicolò Lucchesi's avatar
Nicolò Lucchesi committed
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
# 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
    ]
464
465
466
467
    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
468
469


470
# GLM-4v
471
def run_glm4v(questions: list[str], modality: str) -> ModelRequestData:
472
    assert modality == "image"
473
    model_name = "zai-org/glm-4v-9b"
474

475
476
477
478
479
480
481
    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"]},
482
        limit_mm_per_prompt={modality: 1},
483
    )
484

485
    prompts = [
486
487
488
489
        (
            "<|user|>\n<|begin_of_image|><|endoftext|><|end_of_image|>"
            f"{question}<|assistant|>"
        )
490
        for question in questions
491
    ]
492

493
    stop_token_ids = [151329, 151336, 151338]
494
495
496
497
498
499

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
        stop_token_ids=stop_token_ids,
    )
500
501


502
503
# GLM-4.1V
def run_glm4_1v(questions: list[str], modality: str) -> ModelRequestData:
504
    model_name = "zai-org/GLM-4.1V-9B-Thinking"
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537

    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,
    )


538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
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
# 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,
    )


612
613
614
615
616
617
618
619
620
621
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
# GLM-OCR
def run_glm_ocr(questions: list[str], modality: str) -> ModelRequestData:
    model_name = "zai-org/GLM-OCR"

    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,
    )


648
# H2OVL-Mississippi
649
def run_h2ovl(questions: list[str], modality: str) -> ModelRequestData:
650
651
    assert modality == "image"

652
    model_name = "h2oai/h2ovl-mississippi-800m"
653

654
    engine_args = EngineArgs(
655
656
657
        model=model_name,
        trust_remote_code=True,
        max_model_len=8192,
658
        limit_mm_per_prompt={modality: 1},
659
660
    )

661
662
663
664
665
666
667
    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
    )
668
669

    # Stop tokens for H2OVL-Mississippi
670
    # https://huggingface.co/h2oai/h2ovl-mississippi-800m
671
    stop_token_ids = [tokenizer.eos_token_id]
672
673
674
675
676
677

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
        stop_token_ids=stop_token_ids,
    )
678
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
# HunyuanOCR
def run_hunyuan_vl(questions: list[str], modality: str) -> ModelRequestData:
    assert modality == "image"

    model_name = "tencent/HunyuanOCR"

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

    placeholder = "<|hy_place▁holder▁no▁100|><|hy_place▁holder▁no▁102|><|hy_place▁holder▁no▁101|>"  # noqa: E501
    prompts = [
        f"<|hy_begin▁of▁sentence|>{placeholder}{question}<|hy_User|>"
        for question in questions
    ]

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


705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
# 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":
            """
723
724
            ocr: List the words in the image in raster order.
                Even if the word order feels unnatural for reading,
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
                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,
    )


784
# Idefics3-8B-Llama3
785
def run_idefics3(questions: list[str], modality: str) -> ModelRequestData:
786
787
788
    assert modality == "image"
    model_name = "HuggingFaceM4/Idefics3-8B-Llama3"

789
    engine_args = EngineArgs(
790
791
792
793
794
795
796
        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={
797
            "size": {"longest_edge": 3 * 364},
798
        },
799
        limit_mm_per_prompt={modality: 1},
800
    )
801
802
803
804
    prompts = [
        (f"<|begin_of_text|>User:<image>{question}<end_of_utterance>\nAssistant:")
        for question in questions
    ]
805
806
807
808
809

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
810
811


Lyu Han's avatar
Lyu Han committed
812
813
# Intern-S1
def run_interns1(questions: list[str], modality: str) -> ModelRequestData:
814
    model_name = "internlm/Intern-S1-mini"
Lyu Han's avatar
Lyu Han committed
815
816
817
818
819
820
821
822
823
824

    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,
    )

825
826
827
828
829
    if modality == "image":
        placeholder = "<IMG_CONTEXT>"
    elif modality == "video":
        placeholder = "<video>"

Lyu Han's avatar
Lyu Han committed
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
    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,
    )


zxy's avatar
zxy committed
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
# Intern-S1-Pro
def run_interns1_pro(questions: list[str], modality: str) -> ModelRequestData:
    model_name = "internlm/Intern-S1-Pro"

    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,
        tensor_parallel_size=4,
    )

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

    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,
    )


879
# InternVL
880
def run_internvl(questions: list[str], modality: str) -> ModelRequestData:
881
    model_name = "OpenGVLab/InternVL3-2B"
882

883
    engine_args = EngineArgs(
884
885
        model=model_name,
        trust_remote_code=True,
886
        max_model_len=8192,
887
        limit_mm_per_prompt={modality: 1},
888
889
    )

890
891
892
893
894
    if modality == "image":
        placeholder = "<image>"
    elif modality == "video":
        placeholder = "<video>"

895
896
897
898
899
900
901
902
    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
    )
903
904
905
906
907
908
909

    # 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]
910
    stop_token_ids = [token_id for token_id in stop_token_ids if token_id is not None]
911
912
913
914
915
916

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


919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
# Kanana-V
def run_kanana_v(questions: list[str], modality: str) -> ModelRequestData:
    assert modality == "image"

    model_name = "kakaocorp/kanana-1.5-v-3b-instruct"

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

    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
    )

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


946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
# 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,
    )


977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
# Keye-VL-1.5
def run_keye_vl1_5(questions: list[str], modality: str) -> ModelRequestData:
    model_name = "Kwai-Keye/Keye-VL-1.5-8B"

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

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

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

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


1008
1009
1010
1011
1012
1013
1014
# 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|>"
1015
1016
        "<|im_assistant|>assistant<|im_middle|>"
        for question in questions
1017
1018
1019
1020
1021
    ]

    engine_args = EngineArgs(
        model="moonshotai/Kimi-VL-A3B-Instruct",
        trust_remote_code=True,
Cyrus Leung's avatar
Cyrus Leung committed
1022
        max_model_len=4096,
1023
        limit_mm_per_prompt={modality: 1},
1024
1025
1026
1027
1028
1029
1030
1031
    )

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


1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
# Kimi-VL
def run_kimi_k25(questions: list[str], modality: str) -> ModelRequestData:
    assert modality == "vision_chunk"

    prompts = [
        "<|im_user|>user<|media_begin|>image<|media_content|>"
        f"<|media_pad|><|media_end|>{question}<|im_end|>"
        "<|im_assistant|>assistant<|im_middle|>"
        for question in questions
    ]

    engine_args = EngineArgs(
        model="moonshotai/Kimi-K2.5",
        trust_remote_code=True,
        max_model_len=4096,
        limit_mm_per_prompt={modality: 1},
        tensor_parallel_size=4,
    )

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


1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
# LightOnOCR
def run_lightonocr(questions: list[str], modality: str) -> ModelRequestData:
    assert modality == "image"

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

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

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


1077
1078
1079
1080
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
def run_lfm2_vl(questions: list[str], modality: str) -> ModelRequestData:
    assert modality == "image"

    model_name = "LiquidAI/LFM2-VL-450M"

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

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

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


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
1139
1140
1141
1142
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,
    )


1143
# LLaVA-1.5
1144
def run_llava(questions: list[str], modality: str) -> ModelRequestData:
1145
    assert modality == "image"
1146

1147
    prompts = [f"USER: <image>\n{question}\nASSISTANT:" for question in questions]
1148

1149
1150
1151
    engine_args = EngineArgs(
        model="llava-hf/llava-1.5-7b-hf",
        max_model_len=4096,
1152
        limit_mm_per_prompt={modality: 1},
1153
1154
1155
1156
1157
1158
    )

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
1159
1160
1161


# LLaVA-1.6/LLaVA-NeXT
1162
def run_llava_next(questions: list[str], modality: str) -> ModelRequestData:
1163
    assert modality == "image"
1164

1165
    prompts = [f"[INST] <image>\n{question} [/INST]" for question in questions]
1166
1167
1168
    engine_args = EngineArgs(
        model="llava-hf/llava-v1.6-mistral-7b-hf",
        max_model_len=8192,
1169
        limit_mm_per_prompt={modality: 1},
1170
1171
1172
1173
1174
1175
    )

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
1176
1177
1178
1179


# LlaVA-NeXT-Video
# Currently only support for video input
1180
def run_llava_next_video(questions: list[str], modality: str) -> ModelRequestData:
1181
1182
    assert modality == "video"

1183
    prompts = [f"USER: <video>\n{question} ASSISTANT:" for question in questions]
1184
1185
1186
    engine_args = EngineArgs(
        model="llava-hf/LLaVA-NeXT-Video-7B-hf",
        max_model_len=8192,
1187
        max_num_seqs=2,
1188
        limit_mm_per_prompt={modality: 1},
1189
1190
1191
1192
1193
1194
    )

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
1195
1196


1197
# LLaVA-OneVision
1198
def run_llava_onevision(questions: list[str], modality: str) -> ModelRequestData:
1199
    if modality == "video":
1200
        prompts = [
1201
            f"<|im_start|>user <video>\n{question}<|im_end|><|im_start|>assistant\n"
1202
            for question in questions
1203
        ]
1204
1205

    elif modality == "image":
1206
        prompts = [
1207
            f"<|im_start|>user <image>\n{question}<|im_end|><|im_start|>assistant\n"
1208
            for question in questions
1209
        ]
1210

1211
1212
1213
    engine_args = EngineArgs(
        model="llava-hf/llava-onevision-qwen2-7b-ov-hf",
        max_model_len=16384,
1214
        limit_mm_per_prompt={modality: 1},
1215
1216
1217
1218
1219
1220
    )

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
1221
1222


1223
# Mantis
1224
def run_mantis(questions: list[str], modality: str) -> ModelRequestData:
1225
    assert modality == "image"
1226

1227
1228
    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]
1229

1230
    engine_args = EngineArgs(
1231
        model="TIGER-Lab/Mantis-8B-siglip-llama3",
1232
        max_model_len=4096,
1233
        hf_overrides={"architectures": ["MantisForConditionalGeneration"]},
1234
        limit_mm_per_prompt={modality: 1},
1235
    )
1236
    stop_token_ids = [128009]
1237
1238
1239
1240
1241
1242

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
        stop_token_ids=stop_token_ids,
    )
1243
1244
1245


# MiniCPM-V
1246
def run_minicpmv_base(questions: list[str], modality: str, model_name):
1247
1248
    assert modality in ["image", "video"]
    # If you want to use `MiniCPM-o-2_6` with audio inputs, check `audio_language.py` # noqa
1249
1250
1251
1252
1253
1254
1255

    # 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
1256
1257
    # model_name = "openbmb/MiniCPM-Llama3-V-2_5"

1258
    # 2.6
1259
1260
1261
1262
1263
1264
1265
1266
1267
    # 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"
1268
    tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
1269
    engine_args = EngineArgs(
1270
        model=model_name,
1271
1272
        max_model_len=4096,
        max_num_seqs=2,
1273
        trust_remote_code=True,
1274
        limit_mm_per_prompt={modality: 1},
1275
    )
1276
1277
1278
1279
1280
1281
1282
    # 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]

1283
    # 2.6 / o2.6
1284
    stop_tokens = ["<|im_end|>", "<|endoftext|>"]
1285
    stop_token_ids = [tokenizer.convert_tokens_to_ids(i) for i in stop_tokens]
1286

1287
1288
1289
1290
1291
    modality_placeholder = {
        "image": "(<image>./</image>)",
        "video": "(<video>./</video>)",
    }

1292
1293
    prompts = [
        tokenizer.apply_chat_template(
1294
1295
1296
1297
1298
1299
            [
                {
                    "role": "user",
                    "content": f"{modality_placeholder[modality]}\n{question}",
                }
            ],
1300
            tokenize=False,
1301
1302
1303
            add_generation_prompt=True,
        )
        for question in questions
1304
    ]
1305
1306
1307
1308
1309
1310

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
        stop_token_ids=stop_token_ids,
    )
1311
1312


1313
def run_minicpmo(questions: list[str], modality: str) -> ModelRequestData:
1314
    return run_minicpmv_base(questions, modality, "openbmb/MiniCPM-o-2_6")
1315
1316


1317
def run_minicpmv(questions: list[str], modality: str) -> ModelRequestData:
1318
    return run_minicpmv_base(questions, modality, "openbmb/MiniCPM-V-2_6")
1319
1320


1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
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,
    )


1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
# 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,
1366
        limit_mm_per_prompt={modality: 1},
1367
        ignore_patterns=["consolidated.safetensors"],
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
    )

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

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


1378
1379
# Molmo
def run_molmo(questions: list[str], modality: str) -> ModelRequestData:
1380
1381
    assert modality == "image"

1382
    model_name = "allenai/Molmo-7B-D-0924"
1383
1384
1385

    engine_args = EngineArgs(
        model=model_name,
1386
1387
        trust_remote_code=True,
        dtype="bfloat16",
1388
        limit_mm_per_prompt={modality: 1},
1389
1390
    )

1391
    prompts = [
1392
        f"<|im_start|>user <image>\n{question}<|im_end|><|im_start|>assistant\n"
1393
1394
        for question in questions
    ]
1395

1396
1397
1398
1399
1400
1401
    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )


1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
# Molmo2
def run_molmo2(questions: list[str], modality: str) -> ModelRequestData:
    model_name = "allenai/Molmo2-8B"

    engine_args = EngineArgs(
        model=model_name,
        trust_remote_code=True,
        dtype="bfloat16",
        limit_mm_per_prompt={modality: 1},
        max_num_batched_tokens=36864,
    )

    if modality == "image":
        placeholder = "<|image|>"
    elif modality == "video":
        placeholder = "<|video|>"
    else:
        raise ValueError(f"Unsupported modality for molmo2: {modality}")

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

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


1432
1433
1434
# Nemontron_VL
def run_nemotron_vl(questions: list[str], modality: str) -> ModelRequestData:
    model_name = "nvidia/Llama-3.1-Nemotron-Nano-VL-8B-V1"
1435

1436
    engine_args = EngineArgs(
1437
        model=model_name,
1438
        trust_remote_code=True,
1439
        max_model_len=8192,
1440
        limit_mm_per_prompt={modality: 1},
1441
    )
1442

1443
1444
1445
1446
1447
1448
    assert modality == "image"
    placeholder = "<image>"

    tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
    messages = [
        [{"role": "user", "content": f"{placeholder}\n{question}"}]
1449
        for question in questions
1450
    ]
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
    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]
1462
1463
1464
1465

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
1466
        stop_token_ids=stop_token_ids,
1467
    )
1468
1469


1470
# NVLM-D
1471
def run_nvlm_d(questions: list[str], modality: str) -> ModelRequestData:
1472
1473
1474
1475
1476
    assert modality == "image"

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

    # Adjust this as necessary to fit in GPU
1477
    engine_args = EngineArgs(
1478
1479
1480
1481
        model=model_name,
        trust_remote_code=True,
        max_model_len=4096,
        tensor_parallel_size=4,
1482
        limit_mm_per_prompt={modality: 1},
1483
1484
    )

1485
1486
1487
1488
1489
1490
1491
    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
    )
1492
1493
1494
1495
1496

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
1497
1498


1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
# OpenPangu
def run_openpangu_vl(questions: list[str], modality: str) -> ModelRequestData:
    model_name = "FreedomIntelligence/openPangu-VL-7B"

    engine_args = EngineArgs(
        model=model_name,
        max_model_len=4096,
        max_num_seqs=4,
        trust_remote_code=True,
        enforce_eager=True,
        limit_mm_per_prompt={modality: 1},
    )

    if modality == "image":
        placeholder = "[unused19]"
    elif modality == "video":
        placeholder = "[unused32]"

    prompts = [
        (
            f"<s>[unused9]系统:[unused10][unused9]用户:[unused18]{placeholder}[unused20]{question}[unused10][unused9]助手:"
        )
        for question in questions
    ]

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


1530
1531
# Ovis
def run_ovis(questions: list[str], modality: str) -> ModelRequestData:
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
    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",
1542
        limit_mm_per_prompt={modality: 1},
1543
1544
    )

1545
1546
1547
1548
1549
1550
1551
    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
    )
1552
1553
1554
1555
1556
1557
1558

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


1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
# 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>"

1576
1577
    prompts = [
        f"<|im_start|>user\n\n{placeholder}\n{question}<|im_end|>\n<|im_start|>assistant\n"
1578
1579
1580
1581
1582
1583
1584
1585
1586
        for question in questions
    ]

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


1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
# PaddleOCR-VL
def run_paddleocr_vl(questions: list[str], modality: str) -> ModelRequestData:
    assert modality == "image"

    model_name = "PaddlePaddle/PaddleOCR-VL"

    engine_args = EngineArgs(
        model=model_name,
        max_model_len=4096,
        max_num_seqs=2,
        limit_mm_per_prompt={modality: 1},
        trust_remote_code=True,
    )

    placeholder = "<|IMAGE_START|><|IMAGE_PLACEHOLDER|><|IMAGE_END|>"
    prompts = [
        (f"<|begin_of_sentence|>User: {question}{placeholder}\nAssistant: ")
        for question in questions
    ]

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


1613
# PaliGemma
1614
def run_paligemma(questions: list[str], modality: str) -> ModelRequestData:
1615
    assert modality == "image"
1616

1617
    # PaliGemma has special prompt format for VQA
1618
1619
1620
    prompts = ["caption en" for _ in questions]
    engine_args = EngineArgs(
        model="google/paligemma-3b-mix-224",
1621
        limit_mm_per_prompt={modality: 1},
1622
    )
1623
1624
1625
1626
1627

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
1628
1629


1630
# PaliGemma 2
1631
def run_paligemma2(questions: list[str], modality: str) -> ModelRequestData:
1632
    assert modality == "image"
1633

1634
    # PaliGemma 2 has special prompt format for VQA
1635
1636
1637
    prompts = ["caption en" for _ in questions]
    engine_args = EngineArgs(
        model="google/paligemma2-3b-ft-docci-448",
1638
        limit_mm_per_prompt={modality: 1},
1639
    )
1640
1641
1642
1643
1644

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
1645
1646


1647
# Phi-3-Vision
1648
def run_phi3v(questions: list[str], modality: str) -> ModelRequestData:
1649
1650
    assert modality == "image"

1651
1652
1653
1654
    prompts = [
        f"<|user|>\n<|image_1|>\n{question}<|end|>\n<|assistant|>\n"
        for question in questions
    ]
1655

1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
    # 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
1668
    engine_args = EngineArgs(
1669
1670
        model="microsoft/Phi-3.5-vision-instruct",
        trust_remote_code=True,
1671
        max_model_len=4096,
1672
        max_num_seqs=2,
1673
        # Note - mm_processor_kwargs can also be passed to generate/chat calls
1674
        mm_processor_kwargs={"num_crops": 16},
1675
        limit_mm_per_prompt={modality: 1},
1676
    )
1677
1678
1679
1680
1681

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
1682
1683


1684
# Phi-4-multimodal-instruct
1685
def run_phi4mm(questions: list[str], modality: str) -> ModelRequestData:
1686
1687
1688
1689
1690
1691
1692
1693
1694
1695
    """
    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 = [
1696
        f"<|user|><|image_1|>{question}<|end|><|assistant|>" for question in questions
1697
    ]
1698
    engine_args = EngineArgs(
1699
1700
        model=model_path,
        trust_remote_code=True,
1701
        max_model_len=5120,
1702
        max_num_seqs=2,
1703
        max_num_batched_tokens=12800,
1704
1705
        enable_lora=True,
        max_lora_rank=320,
1706
1707
        # Note - mm_processor_kwargs can also be passed to generate/chat calls
        mm_processor_kwargs={"dynamic_hd": 16},
1708
        limit_mm_per_prompt={modality: 1},
1709
1710
    )

1711
1712
1713
1714
1715
    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
        lora_requests=[LoRARequest("vision", 1, vision_lora_path)],
    )
1716
1717


1718
# Pixtral HF-format
1719
def run_pixtral_hf(questions: list[str], modality: str) -> ModelRequestData:
1720
1721
1722
1723
    assert modality == "image"

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

1724
    # NOTE: Need L40 (or equivalent) to avoid OOM
1725
    engine_args = EngineArgs(
1726
        model=model_name,
1727
        max_model_len=6144,
1728
        max_num_seqs=2,
1729
        limit_mm_per_prompt={modality: 1},
1730
1731
    )

1732
    prompts = [f"<s>[INST]{question}\n[IMG][/INST]" for question in questions]
1733
1734
1735
1736
1737

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
1738
1739


1740
# Qwen-VL
1741
def run_qwen_vl(questions: list[str], modality: str) -> ModelRequestData:
1742
1743
    assert modality == "image"

1744
    engine_args = EngineArgs(
1745
        model="Qwen/Qwen-VL",
1746
        trust_remote_code=True,
1747
1748
        max_model_len=1024,
        max_num_seqs=2,
1749
        hf_overrides={"architectures": ["QwenVLForConditionalGeneration"]},
1750
        limit_mm_per_prompt={modality: 1},
1751
1752
    )

1753
    prompts = [f"{question}Picture 1: <img></img>\n" for question in questions]
1754
1755
1756
1757
1758

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
1759
1760


1761
# Qwen2-VL
1762
def run_qwen2_vl(questions: list[str], modality: str) -> ModelRequestData:
1763
    model_name = "Qwen/Qwen2-VL-7B-Instruct"
1764

1765
    engine_args = EngineArgs(
1766
        model=model_name,
1767
1768
1769
        max_model_len=4096,
        max_num_seqs=5,
        # Note - mm_processor_kwargs can also be passed to generate/chat calls
1770
        mm_processor_kwargs={
1771
1772
            "min_pixels": 28 * 28,
            "max_pixels": 1280 * 28 * 28,
1773
        },
1774
        limit_mm_per_prompt={modality: 1},
1775
    )
1776

1777
1778
1779
1780
1781
    if modality == "image":
        placeholder = "<|image_pad|>"
    elif modality == "video":
        placeholder = "<|video_pad|>"

1782
    prompts = [
1783
1784
1785
1786
1787
1788
1789
        (
            "<|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
1790
    ]
1791
1792
1793
1794
1795

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
1796
1797


Roger Wang's avatar
Roger Wang committed
1798
# Qwen2.5-VL
1799
def run_qwen2_5_vl(questions: list[str], modality: str) -> ModelRequestData:
Roger Wang's avatar
Roger Wang committed
1800
1801
    model_name = "Qwen/Qwen2.5-VL-3B-Instruct"

1802
    engine_args = EngineArgs(
Roger Wang's avatar
Roger Wang committed
1803
1804
1805
1806
1807
1808
1809
1810
        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,
        },
1811
        limit_mm_per_prompt={modality: 1},
Roger Wang's avatar
Roger Wang committed
1812
1813
1814
1815
1816
1817
1818
    )

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

1819
    prompts = [
1820
1821
1822
1823
1824
1825
1826
        (
            "<|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
1827
    ]
1828
1829
1830
1831
1832

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )
Roger Wang's avatar
Roger Wang committed
1833
1834


1835
1836
1837
1838
1839
1840
1841
1842
1843
1844
1845
# 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,
1846
            "fps": 1,
1847
        },
1848
        limit_mm_per_prompt={modality: 1},
1849
1850
1851
1852
1853
1854
1855
1856
1857
1858
    )

    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 "
1859
1860
        "generating text and speech."
    )
1861

1862
1863
1864
1865
1866
1867
1868
1869
1870
    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
    ]
1871
1872
1873
1874
1875
1876
    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
    )


1877
1878
1879
1880
1881
1882
1883
1884
1885
1886
1887
1888
1889
1890
1891
1892
1893
1894
1895
1896
1897
1898
1899
1900
1901
1902
1903
1904
1905
1906
1907
1908
1909
1910
1911
1912
1913
1914
1915
1916
1917
1918
1919
1920
1921
1922
1923
1924
1925
1926
1927
1928
1929
1930
1931
1932
1933
1934
1935
1936
1937
1938
1939
1940
1941
1942
1943
1944
1945
1946
1947
1948
1949
1950
# Qwen3-VL-Dense
def run_qwen3_vl(questions: list[str], modality: str) -> ModelRequestData:
    model_name = "Qwen/Qwen3-VL-4B-Instruct"

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

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

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

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


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

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

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

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

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


1951
1952
1953
1954
1955
1956
1957
1958
1959
1960
1961
1962
1963
1964
1965
1966
1967
1968
1969
1970
1971
1972
# 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,
    )


1973
1974
# SkyworkR1V
def run_skyworkr1v(questions: list[str], modality: str) -> ModelRequestData:
1975
    assert modality == "image"
1976
1977

    model_name = "Skywork/Skywork-R1V-38B"
1978
1979
1980
1981
1982
1983
1984

    engine_args = EngineArgs(
        model=model_name,
        trust_remote_code=True,
        max_model_len=4096,
        limit_mm_per_prompt={modality: 1},
    )
1985
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
1996
1997

    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]
1998
1999
2000
2001

    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
2002
        stop_token_ids=stop_token_ids,
2003
2004
2005
    )


2006
2007
2008
2009
# SmolVLM2-2.2B-Instruct
def run_smolvlm(questions: list[str], modality: str) -> ModelRequestData:
    assert modality == "image"
    model_name = "HuggingFaceTB/SmolVLM2-2.2B-Instruct"
2010
2011
2012

    engine_args = EngineArgs(
        model=model_name,
2013
2014
2015
2016
2017
2018
        max_model_len=8192,
        max_num_seqs=2,
        enforce_eager=True,
        mm_processor_kwargs={
            "max_image_size": {"longest_edge": 384},
        },
2019
2020
        limit_mm_per_prompt={modality: 1},
    )
2021
2022
2023
2024
    prompts = [
        (f"<|im_start|>User:<image>{question}<end_of_utterance>\nAssistant:")
        for question in questions
    ]
2025

2026
2027
2028
2029
2030
2031
2032
2033
2034
2035
2036
2037
2038
2039
2040
2041
2042
2043
2044
2045
2046
2047
    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",
    )
2048
2049

    prompts = [
2050
2051
        "<|begin▁of▁sentence|> You are a helpful assistant. <|BOT|>user\n "
        f"<im_patch>{question} <|EOT|><|BOT|>assistant\n<think>\n"
2052
2053
2054
2055
2056
2057
2058
2059
2060
        for question in questions
    ]

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


2061
2062
2063
2064
2065
2066
2067
2068
2069
2070
2071
2072
2073
2074
2075
2076
2077
2078
2079
2080
2081
2082
2083
2084
2085
2086
# StepVL10B
def run_step_vl(questions: list[str], modality: str) -> ModelRequestData:
    assert modality == "image"

    model_name = "stepfun-ai/Step3-VL-10B"
    engine_args = EngineArgs(
        model=model_name,
        max_num_batched_tokens=4096,
        tensor_parallel_size=1,
        trust_remote_code=True,
        limit_mm_per_prompt={modality: 1},
        reasoning_parser="deepseek_r1",
    )

    prompts = [
        "<|begin▁of▁sentence|> You are a helpful assistant.<|BOT|>user\n "
        f"<im_patch>{question} <|EOT|><|BOT|>assistant\n<think>\n"
        for question in questions
    ]

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


2087
2088
# omni-research/Tarsier-7b
def run_tarsier(questions: list[str], modality: str) -> ModelRequestData:
2089
    assert modality == "image"
2090
    model_name = "omni-research/Tarsier-7b"
2091
2092
2093
2094
2095

    engine_args = EngineArgs(
        model=model_name,
        trust_remote_code=True,
        max_model_len=4096,
2096
        limit_mm_per_prompt={modality: 1},
2097
    )
2098
    prompts = [(f"USER: <image>\n{question} ASSISTANT:") for question in questions]
2099

2100
2101
2102
    return ModelRequestData(
        engine_args=engine_args,
        prompts=prompts,
2103
    )
2104

2105
2106
2107
2108
2109
2110
2111

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,
2112
2113
2114
2115
        hf_overrides={
            "architectures": ["Tarsier2ForConditionalGeneration"],
            "model_type": "tarsier2",
        },
2116
2117
2118
2119
2120
2121
2122
2123
2124
2125
2126
2127
2128
2129
2130
2131
2132
        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
    ]
2133
2134
2135
2136
2137
2138
2139

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


2140
model_example_map = {
2141
    "aria": run_aria,
Jennifer Zhao's avatar
Jennifer Zhao committed
2142
    "aya_vision": run_aya_vision,
2143
    "bagel": run_bagel,
2144
    "bee": run_bee,
2145
2146
    "blip-2": run_blip2,
    "chameleon": run_chameleon,
2147
    "command_a_vision": run_command_a_vision,
2148
    "deepseek_vl_v2": run_deepseek_vl2,
2149
    "deepseek_ocr": run_deepseek_ocr,
RED's avatar
RED committed
2150
    "deepseek_ocr2": run_deepseek_ocr2,
2151
    "dots_ocr": run_dots_ocr,
2152
    "eagle2_5": run_eagle2_5,
2153
    "ernie45_vl": run_ernie45_vl,
2154
    "fuyu": run_fuyu,
2155
    "gemma3": run_gemma3,
Nicolò Lucchesi's avatar
Nicolò Lucchesi committed
2156
    "gemma3n": run_gemma3n,
2157
    "glm4v": run_glm4v,
2158
    "glm4_1v": run_glm4_1v,
2159
2160
    "glm4_5v": run_glm4_5v,
    "glm4_5v_fp8": run_glm4_5v_fp8,
2161
    "glm_ocr": run_glm_ocr,
2162
    "h2ovl_chat": run_h2ovl,
2163
    "hunyuan_vl": run_hunyuan_vl,
2164
    "hyperclovax_seed_vision": run_hyperclovax_seed_vision,
2165
    "idefics3": run_idefics3,
Lyu Han's avatar
Lyu Han committed
2166
    "interns1": run_interns1,
zxy's avatar
zxy committed
2167
    "interns1_pro": run_interns1_pro,
2168
    "internvl_chat": run_internvl,
2169
    "kanana_v": run_kanana_v,
2170
    "keye_vl": run_keye_vl,
2171
    "keye_vl1_5": run_keye_vl1_5,
2172
    "kimi_vl": run_kimi_vl,
2173
    "kimi_k25": run_kimi_k25,
2174
    "lightonocr": run_lightonocr,
2175
    "lfm2_vl": run_lfm2_vl,
2176
    "llama4": run_llama4,
2177
2178
    "llava": run_llava,
    "llava-next": run_llava_next,
2179
    "llava-next-video": run_llava_next_video,
2180
    "llava-onevision": run_llava_onevision,
2181
    "mantis": run_mantis,
2182
    "minicpmo": run_minicpmo,
2183
    "minicpmv": run_minicpmv,
2184
    "minimax_vl_01": run_minimax_vl_01,
2185
    "mistral3": run_mistral3,
2186
    "molmo": run_molmo,
2187
    "molmo2": run_molmo2,
2188
    "nemotron_vl": run_nemotron_vl,
2189
    "NVLM_D": run_nvlm_d,
2190
    "openpangu_vl": run_openpangu_vl,
2191
    "ovis": run_ovis,
2192
    "ovis2_5": run_ovis2_5,
2193
    "paddleocr_vl": run_paddleocr_vl,
2194
2195
2196
    "paligemma": run_paligemma,
    "paligemma2": run_paligemma2,
    "phi3_v": run_phi3v,
2197
    "phi4_mm": run_phi4mm,
2198
    "pixtral_hf": run_pixtral_hf,
2199
    "qwen_vl": run_qwen_vl,
2200
    "qwen2_vl": run_qwen2_vl,
Roger Wang's avatar
Roger Wang committed
2201
    "qwen2_5_vl": run_qwen2_5_vl,
2202
    "qwen2_5_omni": run_qwen2_5_omni,
2203
2204
    "qwen3_vl": run_qwen3_vl,
    "qwen3_vl_moe": run_qwen3_vl_moe,
2205
    "rvl": run_r_vl,
2206
    "skywork_chat": run_skyworkr1v,
2207
    "smolvlm": run_smolvlm,
2208
    "step3": run_step3,
2209
    "stepvl": run_step_vl,
汪志鹏's avatar
汪志鹏 committed
2210
    "tarsier": run_tarsier,
2211
    "tarsier2": run_tarsier2,
2212
2213
2214
}


2215
2216
MODELS_NEED_VIDEO_METADATA = [
    "glm4_1v",
2217
    "glm_ocr",
2218
2219
    "glm4_5v",
    "glm4_5v_fp8",
2220
    "molmo2",
2221
2222
    "qwen3_vl",
    "qwen3_vl_moe",
2223
2224
2225
]


2226
2227
2228
2229
2230
2231
2232
2233
2234
def get_multi_modal_input(args):
    """
    return {
        "data": image or video,
        "question": question,
    }
    """
    if args.modality == "image":
        # Input image and question
2235
        image = convert_image_mode(ImageAsset("cherry_blossom").pil_image, "RGB")
2236
2237
2238
2239
2240
2241
        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?",
        ]
2242
2243
2244

        return {
            "data": image,
2245
            "questions": img_questions,
2246
2247
2248
2249
        }

    if args.modality == "video":
        # Input video and question
2250
        needs_metadata = args.model_type in MODELS_NEED_VIDEO_METADATA
2251
        video = VideoAsset(name="baby_reading", num_frames=args.num_frames).np_ndarrays
2252
        metadata = VideoAsset(name="baby_reading", num_frames=args.num_frames).metadata
2253
        vid_questions = ["Why is this video funny?"]
2254
2255

        return {
2256
            "data": ([(video, metadata)] if needs_metadata else video),
2257
            "questions": vid_questions,
2258
2259
        }

2260
2261
2262
2263
2264
2265
2266
2267
2268
2269
2270
2271
2272
    if args.modality == "vision_chunk":
        # Input vision chunks and question
        image = convert_image_mode(ImageAsset("cherry_blossom").pil_image, "RGB")
        vision_chunk_questions = [
            "What is the content of this image chunk?",
            "Describe the content of this image chunk in detail.",
        ]

        return {
            "data": {"type": "image", "image": image},
            "questions": vision_chunk_questions,
        }

2273
2274
2275
2276
    msg = f"Modality {args.modality} is not supported."
    raise ValueError(msg)


2277
2278
2279
2280
def apply_image_repeat(
    image_repeat_prob, num_prompts, data, prompts: list[str], modality
):
    """Repeats images with provided probability of "image_repeat_prob".
2281
2282
    Used to simulate hit/miss for the MM preprocessor cache.
    """
2283
    assert image_repeat_prob <= 1.0 and image_repeat_prob >= 0
2284
2285
2286
2287
    no_yes = [0, 1]
    probs = [1.0 - image_repeat_prob, image_repeat_prob]

    inputs = []
2288
    inputs_with_empty_media = []
2289
2290
2291
2292
2293
2294
2295
2296
2297
2298
    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)

2299
2300
        uuid = "uuid_{}".format(i)

2301
2302
2303
2304
        inputs.append(
            {
                "prompt": prompts[i % len(prompts)],
                "multi_modal_data": {modality: cur_image},
2305
2306
2307
2308
2309
2310
2311
2312
2313
                "multi_modal_uuids": {modality: uuid},
            }
        )

        inputs_with_empty_media.append(
            {
                "prompt": prompts[i % len(prompts)],
                "multi_modal_data": {modality: None},
                "multi_modal_uuids": {modality: uuid},
2314
            }
2315
        )
2316

2317
    return inputs, inputs_with_empty_media
2318
2319


2320
2321
2322
2323
@contextmanager
def time_counter(enable: bool):
    if enable:
        import time
2324

2325
2326
2327
2328
2329
2330
2331
2332
2333
2334
        start_time = time.time()
        yield
        elapsed_time = time.time() - start_time
        print("-" * 50)
        print("-- generate time = {}".format(elapsed_time))
        print("-" * 50)
    else:
        yield


2335
2336
def parse_args():
    parser = FlexibleArgumentParser(
2337
2338
2339
2340
2341
2342
2343
2344
2345
2346
2347
2348
2349
2350
2351
2352
2353
2354
        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",
2355
        choices=["image", "video", "vision_chunk"],
2356
2357
2358
2359
2360
2361
2362
2363
2364
2365
2366
        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,
2367
        default=0,
2368
2369
        help="Set the seed when initializing `vllm.LLM`.",
    )
2370
2371

    parser.add_argument(
2372
        "--image-repeat-prob",
2373
2374
        type=float,
        default=None,
2375
2376
        help="Simulates the hit-ratio for multi-modal preprocessor cache (if enabled)",
    )
2377
2378

    parser.add_argument(
2379
        "--disable-mm-processor-cache",
2380
        action="store_true",
2381
        help="If True, disables caching of multi-modal processor.",
2382
    )
2383
2384

    parser.add_argument(
2385
2386
2387
2388
        "--time-generate",
        action="store_true",
        help="If True, then print the total generate() call time",
    )
2389
2390

    parser.add_argument(
2391
2392
2393
2394
2395
        "--use-different-prompt-per-request",
        action="store_true",
        help="If True, then use different prompt (with the same multi-modal "
        "data) for each request.",
    )
2396
2397
2398
2399
2400
2401
2402

    parser.add_argument(
        "--verify-mm-cache-hit-with-uuids",
        action="store_true",
        help="If True, will send all requests in a second batch with empty mm "
        "data to verify cache hits with UUIDs.",
    )
2403
2404
2405
2406
2407
2408
2409
    parser.add_argument(
        "--tensor-parallel-size",
        "-tp",
        type=int,
        default=None,
        help="Tensor parallel size to override the model's default setting. ",
    )
2410
2411
2412
    return parser.parse_args()


2413
2414
2415
2416
2417
def main(args):
    model = args.model_type
    if model not in model_example_map:
        raise ValueError(f"Model type {model} is not supported.")

2418
2419
2420
2421
2422
2423
    if args.tensor_parallel_size is not None and args.tensor_parallel_size < 1:
        raise ValueError(
            f"tensor_parallel_size must be a positive integer, "
            f"got {args.tensor_parallel_size}"
        )

2424
2425
2426
    modality = args.modality
    mm_input = get_multi_modal_input(args)
    data = mm_input["data"]
2427
    questions = mm_input["questions"]
2428

2429
2430
    req_data = model_example_map[model](questions, modality)

2431
    # Disable other modalities to save memory
2432
    default_limits = {"image": 0, "video": 0, "audio": 0, "vision_chunk": 0}
2433
    req_data.engine_args.limit_mm_per_prompt = default_limits | dict(
2434
2435
        req_data.engine_args.limit_mm_per_prompt or {}
    )
2436
2437
2438

    engine_args = asdict(req_data.engine_args) | {
        "seed": args.seed,
2439
        "mm_processor_cache_gb": 0 if args.disable_mm_processor_cache else 4,
2440
    }
2441
2442
    if args.tensor_parallel_size is not None:
        engine_args["tensor_parallel_size"] = args.tensor_parallel_size
2443
2444
    llm = LLM(**engine_args)

2445
    # Don't want to check the flag multiple times, so just hijack `prompts`.
2446
2447
2448
2449
2450
    prompts = (
        req_data.prompts
        if args.use_different_prompt_per_request
        else [req_data.prompts[0]]
    )
2451
2452
2453

    # We set temperature to 0.2 so that outputs can be different
    # even when all prompts are identical when running batch inference.
2454
2455
2456
2457
2458
2459
    sampling_params = (
        SamplingParams(
            temperature=0.2, max_tokens=64, stop_token_ids=req_data.stop_token_ids
        )
        if req_data.sampling_params is None
        else req_data.sampling_params
2460
    )
2461
2462
2463
2464

    assert args.num_prompts > 0
    if args.num_prompts == 1:
        # Single inference
2465
        uuid = "uuid_0"
2466
        inputs = {
2467
            "prompt": prompts[0],
2468
            "multi_modal_data": {modality: data},
2469
2470
2471
2472
2473
2474
            "multi_modal_uuids": {modality: uuid},
        }
        inputs_with_empty_media = {
            "prompt": prompts[0],
            "multi_modal_data": {modality: None},
            "multi_modal_uuids": {modality: uuid},
2475
2476
2477
        }
    else:
        # Batch inference
2478
2479
        if args.image_repeat_prob is not None:
            # Repeat images with specified probability of "image_repeat_prob"
2480
2481
2482
2483
2484
2485
            inputs, inputs_with_empty_media = apply_image_repeat(
                args.image_repeat_prob,
                args.num_prompts,
                data,
                prompts,
                modality,
2486
            )
2487
2488
        else:
            # Use the same image for all prompts
2489
2490
2491
2492
2493
2494
2495
2496
2497
2498
2499
2500
2501
2502
2503
2504
2505
2506
            inputs = []
            inputs_with_empty_media = []
            for i in range(args.num_prompts):
                uuid = "uuid_{}".format(i)
                inputs.append(
                    {
                        "prompt": prompts[i % len(prompts)],
                        "multi_modal_data": {modality: data},
                        "multi_modal_uuids": {modality: uuid},
                    }
                )
                inputs_with_empty_media.append(
                    {
                        "prompt": prompts[i % len(prompts)],
                        "multi_modal_data": {modality: None},
                        "multi_modal_uuids": {modality: uuid},
                    }
                )
2507

2508
    # Add LoRA request if applicable
2509
2510
2511
    lora_request = (
        req_data.lora_requests * args.num_prompts if req_data.lora_requests else None
    )
2512

2513
2514
2515
2516
2517
2518
    with time_counter(args.time_generate):
        outputs = llm.generate(
            inputs,
            sampling_params=sampling_params,
            lora_request=lora_request,
        )
2519

2520
    print("-" * 50)
2521
2522
2523
    for o in outputs:
        generated_text = o.outputs[0].text
        print(generated_text)
2524
        print("-" * 50)
2525

2526
2527
2528
2529
2530
2531
2532
2533
2534
2535
2536
2537
2538
2539
2540
2541
2542
2543
2544
2545
    if args.verify_mm_cache_hit_with_uuids:
        try:
            # Verify cache hits with UUIDs
            print(
                "Sending a second batch of requests with empty media"
                " and matching UUIDs."
            )
            outputs = llm.generate(
                inputs_with_empty_media,
                sampling_params=sampling_params,
                lora_request=lora_request,
            )
            print("-" * 50)
            for o in outputs:
                generated_text = o.outputs[0].text
                print(generated_text)
                print("-" * 50)
        except Exception as e:
            print(f"Failed to verify cache hits with UUIDs. Error: {e}")

2546
2547

if __name__ == "__main__":
2548
    args = parse_args()
2549
    main(args)