test_common.py 14.7 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

4
from functools import partial
5
from typing import Optional, Union
6
7
8

import numpy as np
import pytest
9
from mistral_common.protocol.instruct.messages import ImageChunk, TextChunk, UserMessage
10
from mistral_common.protocol.instruct.request import ChatCompletionRequest
11
12
13
from PIL import Image

from vllm.config import ModelConfig
14
15
16
17
18
19
from vllm.config.multimodal import (
    AudioDummyOptions,
    BaseDummyOptions,
    ImageDummyOptions,
    VideoDummyOptions,
)
20
from vllm.multimodal import MULTIMODAL_REGISTRY, MultiModalDataDict
21
from vllm.multimodal.cache import MultiModalProcessorOnlyCache
22
from vllm.multimodal.inputs import MultiModalInputs
23
24
25
26
27
28
29
from vllm.multimodal.processing import BaseMultiModalProcessor, InputProcessingContext
from vllm.transformers_utils.tokenizer import (
    AnyTokenizer,
    MistralTokenizer,
    cached_tokenizer_from_config,
    encode_tokens,
)
30
31

from ....multimodal.utils import random_audio, random_image, random_video
32
from ...registry import HF_EXAMPLE_MODELS
33
34


35
36
37
38
39
40
def glm4_1v_patch_mm_data(mm_data: MultiModalDataDict) -> MultiModalDataDict:
    """
    Patch the multimodal data for GLM4.1V model.
    """
    # Ensure video metadata is included
    if "video" in mm_data:
41
        # GLM4.1V doesn't support multiple videos
42
        video = mm_data["video"]
43
        num_frames = len(video)
44
45
46
47
48
49
50
51
52
53
54
        mm_data["video"] = (
            video,
            {
                "total_num_frames": num_frames,
                "fps": num_frames,
                "duration": 1,
                "frames_indices": [i for i in range(num_frames)],
                "video_backend": "opencv",
                "do_sample_frames": True,
            },
        )
55
56
57
    return mm_data


58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
def qwen3_vl_patch_mm_data(mm_data: MultiModalDataDict) -> MultiModalDataDict:
    """
    Patch the multimodal data for Qwen3-VL model.
    """

    def create_metadata(frames: np.ndarray):
        num_frames = len(frames)
        return {
            "total_num_frames": num_frames,
            "fps": 2.0,
            "duration": num_frames / 2.0,
            "video_backend": "opencv",
            "frames_indices": list(range(num_frames)),
            "do_sample_frames": True,
        }

    # Ensure video metadata is included
    if "video" in mm_data:
        video = mm_data["video"]
        if isinstance(video, list):
            # multiple videos
            mm_data["video"] = [(vid, create_metadata(vid)) for vid in video]
        else:
            # single video
            mm_data["video"] = (video, create_metadata(video))
    return mm_data


86
def _test_processing_correctness(
87
    model_id_or_arch: str,
88
89
90
91
    hit_rate: float,
    num_batches: int,
    simplify_rate: float,
):
92
93
94
95
96
97
98
    if model_id_or_arch in HF_EXAMPLE_MODELS.get_supported_archs():
        # Use model architecture to get the default model id
        model_info = HF_EXAMPLE_MODELS.get_hf_info(model_id_or_arch)
        model_id = model_info.default
    else:
        model_info = HF_EXAMPLE_MODELS.find_hf_info(model_id_or_arch)
        model_id = model_id_or_arch
99
100
    model_info.check_available_online(on_fail="skip")
    model_info.check_transformers_version(on_fail="skip")
101
102
103

    model_config = ModelConfig(
        model_id,
104
105
        tokenizer=model_info.tokenizer or model_id,
        tokenizer_mode=model_info.tokenizer_mode,
106
        revision=model_info.revision,
107
        trust_remote_code=model_info.trust_remote_code,
108
        hf_overrides=model_info.hf_overrides,
109
110
        # Ensure that the cache can fit all of the data
        mm_processor_cache_gb=2048,
111
112
        skip_tokenizer_init=model_info.skip_tokenizer_init,
        enforce_eager=model_info.enforce_eager,
113
114
        dtype=model_info.dtype,
    )
115
116
117
118
119

    model_cls = MULTIMODAL_REGISTRY._get_model_cls(model_config)
    factories = MULTIMODAL_REGISTRY._processor_factories[model_cls]
    ctx = InputProcessingContext(
        model_config,
120
        tokenizer=cached_tokenizer_from_config(model_config),
121
    )
122
    cache = MultiModalProcessorOnlyCache(model_config)
123

124
125
    processing_info = factories.info(ctx)
    supported_mm_limits = processing_info.get_supported_mm_limits()
126
127
    # Keep integer limits for local data generation
    limit_mm_per_prompt_ints = {
128
129
130
131
        modality: 3 if limit is None else limit
        for modality, limit in supported_mm_limits.items()
    }

132
133
134
135
136
137
138
139
140
141
142
143
144
145
    def _to_dummy_options(modality: str, count: int) -> BaseDummyOptions:
        if modality == "video":
            return VideoDummyOptions(count=count)
        if modality == "image":
            return ImageDummyOptions(count=count)
        if modality == "audio":
            return AudioDummyOptions(count=count)
        return BaseDummyOptions(count=count)

    # Assign normalized DummyOptions to the model config
    model_config.get_multimodal_config().limit_per_prompt = {
        modality: _to_dummy_options(modality, count)
        for modality, count in limit_mm_per_prompt_ints.items()
    }
146

147
148
149
150
151
152
153
154
155
156
    baseline_processor = factories.build_processor(ctx, cache=None)
    cached_processor = factories.build_processor(ctx, cache=cache)
    dummy_inputs = baseline_processor.dummy_inputs
    tokenizer = baseline_processor.info.get_tokenizer()

    rng = np.random.RandomState(0)

    input_to_hit = {
        "image": Image.new("RGB", size=(128, 128)),
        "video": np.zeros((4, 128, 128, 3), dtype=np.uint8),
157
        "audio": (np.zeros((512,)), 16000),
158
159
    }
    input_factory = {
160
161
162
163
164
        "image": partial(random_image, rng, min_wh=128, max_wh=256),
        "video": partial(
            random_video, rng, min_frames=2, max_frames=16, min_wh=128, max_wh=256
        ),
        "audio": partial(random_audio, rng, min_len=512, max_len=1024, sr=16000),
165
166
167
168
    }

    for batch_idx in range(num_batches):
        mm_data = {
169
170
171
172
            k: [
                (input_to_hit[k] if rng.rand() < hit_rate else input_factory[k]())
                for _ in range(rng.randint(limit + 1))
            ]
173
            for k, limit in limit_mm_per_prompt_ints.items()
174
175
176
        }

        mm_counts = {k: len(vs) for k, vs in mm_data.items()}
177
178
179
180

        # Mistral chat outputs tokens directly, rather than text prompts
        if isinstance(tokenizer, MistralTokenizer):
            images = mm_data.get("image", [])
181
182
183
184
185
186
187
188
189
190
            request = ChatCompletionRequest(
                messages=[
                    UserMessage(
                        content=[
                            TextChunk(text=""),
                            *(ImageChunk(image=image) for image in images),
                        ]
                    ),
                ]
            )
191
192
193
194
195
196
197
            res = tokenizer.mistral.encode_chat_completion(request)
            prompt = res.tokens
        else:
            prompt = dummy_inputs.get_dummy_processor_inputs(
                model_config.max_model_len,
                mm_counts,
            ).prompt
198
199
200
201
202
203
204
205
206

        # Drop unnecessary keys and test single -> multi conversion
        if rng.rand() < simplify_rate:
            for k in list(mm_data.keys()):
                if not mm_data[k]:
                    del mm_data[k]
                elif len(mm_data[k]) == 1:
                    mm_data[k] = mm_data[k][0]

207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
        _test_processing_correctness_one(
            model_config,
            tokenizer,
            prompt,
            mm_data,
            baseline_processor,
            cached_processor,
            batch_idx,
        )


# For some multimodal models, tokenizer will always add bos_token
# at the beginning of prompt by default, causing hf_processor outputs
# incorrect token ids. So we need use `add_special_tokens=False` here
# to leave bos_token to be added by the processor.
_ADD_SPECIAL_TOKENS_OVERRIDES = {
    "ovis": False,
224
    "ovis2_5": False,
225
    "paligemma": False,
226
227
228
229
230
231
232
233
234
235
236
    "ultravox": False,
    "whisper": False,
}

_IGNORE_MM_KEYS = {
    # In Ultravox, the audio_features can be different depending on padding
    # The slight difference should not be a problem though, since
    # attention_mask lets us ignore the difference.
    "ultravox": {"audio_features"},
}

237
MM_DATA_PATCHES = {
238
    # GLM4.1V and Qwen3-VL requires video metadata to be included in the input
239
    "glm4v": glm4_1v_patch_mm_data,
240
    "glm4v_moe": glm4_1v_patch_mm_data,
241
242
    "qwen3_vl": qwen3_vl_patch_mm_data,
    "qwen3_vl_moe": qwen3_vl_patch_mm_data,
243
244
}

245
246

def _test_processing_correctness_one(
247
    model_config: ModelConfig,
248
249
    tokenizer: AnyTokenizer,
    prompt: Union[str, list[int]],
250
251
252
253
254
    mm_data: MultiModalDataDict,
    baseline_processor: BaseMultiModalProcessor,
    cached_processor: BaseMultiModalProcessor,
    batch_idx: int,
):
255
256
    model_type = model_config.hf_config.model_type
    ignore_mm_keys = _IGNORE_MM_KEYS.get(model_type, set[str]())
257
258
    if model_type in MM_DATA_PATCHES:
        mm_data = MM_DATA_PATCHES[model_type](mm_data)
259
260
261
262
263
264
265
266

    if isinstance(prompt, str):
        text_prompt = prompt
        token_prompt = encode_tokens(
            tokenizer,
            prompt,
            add_special_tokens=_ADD_SPECIAL_TOKENS_OVERRIDES.get(model_type),
        )
267
    else:
268
269
270
        # Mistral does not support decode_tokens with skip_special_tokens=False
        text_prompt = None
        token_prompt = prompt
271
272
273
274
275
276
277
278
279
280
281
282
283

    baseline_tokenized_result = baseline_processor.apply(
        token_prompt,
        mm_data=mm_data,
        hf_processor_mm_kwargs={},
    )

    cached_tokenized_result = cached_processor.apply(
        token_prompt,
        mm_data=mm_data,
        hf_processor_mm_kwargs={},
    )

284
    _assert_inputs_equal(
285
286
        baseline_tokenized_result,
        cached_tokenized_result,
287
        ignore_mm_keys=ignore_mm_keys,
288
        msg=f"Failed ({batch_idx=}, {token_prompt=}, {mm_data=})",
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
    if text_prompt is not None:
        baseline_text_result = baseline_processor.apply(
            text_prompt,
            mm_data=mm_data,
            hf_processor_mm_kwargs={},
        )
        cached_text_result = cached_processor.apply(
            text_prompt,
            mm_data=mm_data,
            hf_processor_mm_kwargs={},
        )

        _assert_inputs_equal(
            baseline_text_result,
            cached_text_result,
            ignore_mm_keys=ignore_mm_keys,
            msg=f"Failed ({batch_idx=}, {text_prompt=}, {mm_data=})",
        )

        _assert_inputs_equal(
            baseline_text_result,
            baseline_tokenized_result,
            ignore_mm_keys=ignore_mm_keys,
314
            msg=f"Failed ({batch_idx=}, {text_prompt=}, {token_prompt=}, {mm_data=})",
315
316
317
318
319
320
        )

        _assert_inputs_equal(
            cached_text_result,
            cached_tokenized_result,
            ignore_mm_keys=ignore_mm_keys,
321
            msg=f"Failed ({batch_idx=}, {text_prompt=}, {token_prompt=}, {mm_data=})",
322
323
        )

324
325

# yapf: disable
326
327
@pytest.mark.parametrize("model_id", [
    "rhymes-ai/Aria",
Jennifer Zhao's avatar
Jennifer Zhao committed
328
    "CohereForAI/aya-vision-8b",
329
330
    "Salesforce/blip2-opt-2.7b",
    "facebook/chameleon-7b",
331
    "CohereLabs/command-a-vision-07-2025",
332
    "deepseek-ai/deepseek-vl2-tiny",
333
    "baidu/ERNIE-4.5-VL-28B-A3B-PT",
334
    "adept/fuyu-8b",
335
    "google/gemma-3-4b-it",
Nicolò Lucchesi's avatar
Nicolò Lucchesi committed
336
    "google/gemma-3n-E2B-it",
337
338
    "zai-org/glm-4v-9b",
    "zai-org/GLM-4.1V-9B-Thinking",
339
    "zai-org/GLM-4.5V",
340
    "ibm-granite/granite-speech-3.3-2b",
341
    "h2oai/h2ovl-mississippi-800m",
342
343
    "naver-hyperclovax/HyperCLOVAX-SEED-Vision-Instruct-3B",
    "HuggingFaceM4/Idefics3-8B-Llama3",
344
    "internlm/Intern-S1",
345
    "OpenGVLab/InternVL2-1B",
346
    "OpenGVLab/InternVL3-1B",
347
348
349
    "OpenGVLab/InternVL3_5-1B",
    "OpenGVLab/InternVL3_5-GPT-OSS-20B-A4B-Preview",
    "OpenGVLab/InternVL3_5-30B-A3B",
350
    "Kwai-Keye/Keye-VL-8B-Preview",
351
    "Kwai-Keye/Keye-VL-1_5-8B",
352
    "moonshotai/Kimi-VL-A3B-Instruct",
353
    "meta-llama/Llama-4-Scout-17B-16E-Instruct",
354
355
356
357
358
    "llava-hf/llava-1.5-7b-hf",
    "llava-hf/llava-v1.6-mistral-7b-hf",
    "llava-hf/LLaVA-NeXT-Video-7B-hf",
    "llava-hf/llava-onevision-qwen2-0.5b-ov-hf",
    "TIGER-Lab/Mantis-8B-siglip-llama3",
359
    "mispeech/midashenglm-7b",
360
    "openbmb/MiniCPM-Llama3-V-2_5",
361
362
    "openbmb/MiniCPM-o-2_6",
    "openbmb/MiniCPM-V-2_6",
363
    "MiniMaxAI/MiniMax-VL-01",
364
365
    "allenai/Molmo-7B-D-0924",
    "allenai/Molmo-7B-O-0924",
366
    "nvidia/NVLM-D-72B",
367
    "nvidia/Llama-3.1-Nemotron-Nano-VL-8B-V1",
368
369
    "AIDC-AI/Ovis1.6-Gemma2-9B",
    "AIDC-AI/Ovis1.6-Llama3.2-3B",
370
    "AIDC-AI/Ovis2-1B",
371
    "AIDC-AI/Ovis2.5-2B",
372
373
    "google/paligemma-3b-mix-224",
    "google/paligemma2-3b-ft-docci-448",
374
    "microsoft/Phi-3.5-vision-instruct",
375
    "microsoft/Phi-4-multimodal-instruct",
376
377
    "mistralai/Pixtral-12B-2409",
    "mistral-community/pixtral-12b",
378
379
    "Qwen/Qwen-VL-Chat",
    "Qwen/Qwen2-VL-2B-Instruct",
Roger Wang's avatar
Roger Wang committed
380
    "Qwen/Qwen2.5-VL-3B-Instruct",
381
    "Qwen/Qwen2-Audio-7B-Instruct",
382
    "Qwen/Qwen2.5-Omni-3B",
383
384
    "Qwen/Qwen3-VL-4B-Instruct",
    "Qwen/Qwen3-VL-30B-A3B-Instruct",
385
    "YannQi/R-4B",
386
    "Skywork/Skywork-R1V-38B",
387
388
    "HuggingFaceTB/SmolVLM2-2.2B-Instruct",
    "stepfun-ai/step3",
389
    "fixie-ai/ultravox-v0_5-llama-3_2-1b",
390
    "openai/whisper-large-v3",
汪志鹏's avatar
汪志鹏 committed
391
    "omni-research/Tarsier-7b",
Nicolò Lucchesi's avatar
Nicolò Lucchesi committed
392
    "omni-research/Tarsier2-Recap-7b",
393
    "mistralai/Voxtral-Mini-3B-2507",
394
395
396
397
398
399
400
401
402
403
404
])
@pytest.mark.parametrize("hit_rate", [0.3, 0.5, 1.0])
@pytest.mark.parametrize("num_batches", [32])
@pytest.mark.parametrize("simplify_rate", [1.0])
# yapf: enable
def test_processing_correctness(
    model_id: str,
    hit_rate: float,
    num_batches: int,
    simplify_rate: float,
):
Nicolò Lucchesi's avatar
Nicolò Lucchesi committed
405
406
    if model_id == "google/gemma-3n-E2B-it":
        pytest.skip("Skipping gemma-3n-E2B-it due to transformers #39911 bug.")
407
408
409
410
411
412
    _test_processing_correctness(
        model_id,
        hit_rate=hit_rate,
        num_batches=num_batches,
        simplify_rate=simplify_rate,
    )
413
414


415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
# Phi4MultimodalForCausalLM share same model repo with original format
# Phi4MMForCausalLM, so we add it as a separate test case
# Remove this test after conversion PR merged:
# https://huggingface.co/microsoft/Phi-4-multimodal-instruct/discussions/70
@pytest.mark.parametrize("model_arch", ["Phi4MultimodalForCausalLM"])
@pytest.mark.parametrize("hit_rate", [0.3, 0.5, 1.0])
@pytest.mark.parametrize("num_batches", [32])
@pytest.mark.parametrize("simplify_rate", [1.0])
def test_processing_correctness_phi4_multimodal(
    model_arch: str,
    hit_rate: float,
    num_batches: int,
    simplify_rate: float,
):
    _test_processing_correctness(
        model_arch,
        hit_rate=hit_rate,
        num_batches=num_batches,
        simplify_rate=simplify_rate,
    )


437
def _assert_inputs_equal(
438
439
    a: MultiModalInputs,
    b: MultiModalInputs,
440
441
442
    *,
    ignore_mm_keys: Optional[set[str]] = None,
    msg: str = "",
443
):
444
445
446
    if ignore_mm_keys is None:
        ignore_mm_keys = set()

447
448
449
450
451
452
453
    a_rest = {k: v for k, v in a.items() if k != "mm_kwargs"}
    b_rest = {k: v for k, v in b.items() if k != "mm_kwargs"}

    assert a_rest == b_rest, msg

    a_data = a["mm_kwargs"].get_data()
    b_data = b["mm_kwargs"].get_data()
454
455

    for key in ignore_mm_keys:
456
457
        a_data.pop(key, None)
        b_data.pop(key, None)
458

459
    assert a_data == b_data, msg