test_common.py 15.2 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

4
5
6
7
from functools import partial

import numpy as np
import pytest
8
9
from mistral_common.protocol.instruct.chunk import ImageChunk, TextChunk
from mistral_common.protocol.instruct.messages import 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
113
        skip_tokenizer_init=model_info.require_embed_inputs,
        enable_prompt_embeds=model_info.require_embed_inputs,
        enable_mm_embeds=model_info.require_embed_inputs,
114
        enforce_eager=model_info.enforce_eager,
115
116
        dtype=model_info.dtype,
    )
117
118
119
120
121

    model_cls = MULTIMODAL_REGISTRY._get_model_cls(model_config)
    factories = MULTIMODAL_REGISTRY._processor_factories[model_cls]
    ctx = InputProcessingContext(
        model_config,
122
        tokenizer=cached_tokenizer_from_config(model_config),
123
    )
124
    cache = MultiModalProcessorOnlyCache(model_config)
125

126
127
    processing_info = factories.info(ctx)
    supported_mm_limits = processing_info.get_supported_mm_limits()
128
129
    # Keep integer limits for local data generation
    limit_mm_per_prompt_ints = {
130
131
132
133
        modality: 3 if limit is None else limit
        for modality, limit in supported_mm_limits.items()
    }

134
135
136
137
138
139
140
141
142
143
144
145
146
147
    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()
    }
148

149
150
151
152
153
154
155
156
157
158
    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),
159
        "audio": (np.zeros((512,)), 16000),
160
161
    }
    input_factory = {
162
163
164
165
166
        "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),
167
168
169
170
    }

    for batch_idx in range(num_batches):
        mm_data = {
171
172
173
174
            k: [
                (input_to_hit[k] if rng.rand() < hit_rate else input_factory[k]())
                for _ in range(rng.randint(limit + 1))
            ]
175
            for k, limit in limit_mm_per_prompt_ints.items()
176
177
178
        }

        mm_counts = {k: len(vs) for k, vs in mm_data.items()}
179
180
181
182

        # Mistral chat outputs tokens directly, rather than text prompts
        if isinstance(tokenizer, MistralTokenizer):
            images = mm_data.get("image", [])
183
184
185
186
187
188
189
190
191
192
            request = ChatCompletionRequest(
                messages=[
                    UserMessage(
                        content=[
                            TextChunk(text=""),
                            *(ImageChunk(image=image) for image in images),
                        ]
                    ),
                ]
            )
193
194
195
196
197
198
199
            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
200
201
202
203
204
205
206
207
208

        # 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]

209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
        _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,
226
    "ovis2_5": False,
227
    "paligemma": False,
228
229
230
231
232
233
234
235
236
237
238
    "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"},
}

239
MM_DATA_PATCHES = {
240
    # GLM4.1V and Qwen3-VL requires video metadata to be included in the input
241
    "glm4v": glm4_1v_patch_mm_data,
242
    "glm4v_moe": glm4_1v_patch_mm_data,
243
244
    "qwen3_vl": qwen3_vl_patch_mm_data,
    "qwen3_vl_moe": qwen3_vl_patch_mm_data,
245
246
}

247
248

def _test_processing_correctness_one(
249
    model_config: ModelConfig,
250
    tokenizer: AnyTokenizer,
251
    prompt: str | list[int],
252
253
254
255
256
    mm_data: MultiModalDataDict,
    baseline_processor: BaseMultiModalProcessor,
    cached_processor: BaseMultiModalProcessor,
    batch_idx: int,
):
257
258
    model_type = model_config.hf_config.model_type
    ignore_mm_keys = _IGNORE_MM_KEYS.get(model_type, set[str]())
259
260
    if model_type in MM_DATA_PATCHES:
        mm_data = MM_DATA_PATCHES[model_type](mm_data)
261
262
263
264
265
266
267
268

    if isinstance(prompt, str):
        text_prompt = prompt
        token_prompt = encode_tokens(
            tokenizer,
            prompt,
            add_special_tokens=_ADD_SPECIAL_TOKENS_OVERRIDES.get(model_type),
        )
269
    else:
270
271
272
        # Mistral does not support decode_tokens with skip_special_tokens=False
        text_prompt = None
        token_prompt = prompt
273
274
275
276
277
278
279
280
281
282
283
284
285

    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={},
    )

286
    _assert_inputs_equal(
287
288
        baseline_tokenized_result,
        cached_tokenized_result,
289
        ignore_mm_keys=ignore_mm_keys,
290
        msg=f"Failed ({batch_idx=}, {token_prompt=}, {mm_data=})",
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
    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,
316
            msg=f"Failed ({batch_idx=}, {text_prompt=}, {token_prompt=}, {mm_data=})",
317
318
319
320
321
322
        )

        _assert_inputs_equal(
            cached_text_result,
            cached_tokenized_result,
            ignore_mm_keys=ignore_mm_keys,
323
            msg=f"Failed ({batch_idx=}, {text_prompt=}, {token_prompt=}, {mm_data=})",
324
325
        )

326

327
328
329
330
331
@pytest.mark.parametrize(
    "model_id",
    [
        "rhymes-ai/Aria",
        "CohereForAI/aya-vision-8b",
332
        "Open-Bee/Bee-8B-RL",
333
334
335
336
        "Salesforce/blip2-opt-2.7b",
        "facebook/chameleon-7b",
        "CohereLabs/command-a-vision-07-2025",
        "deepseek-ai/deepseek-vl2-tiny",
337
        "deepseek-ai/DeepSeek-OCR",
338
339
        "baidu/ERNIE-4.5-VL-28B-A3B-PT",
        "adept/fuyu-8b",
340
        "google/gemma-3-4b-it",
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
        "google/gemma-3n-E2B-it",
        "zai-org/glm-4v-9b",
        "zai-org/GLM-4.1V-9B-Thinking",
        "zai-org/GLM-4.5V",
        "ibm-granite/granite-speech-3.3-2b",
        "h2oai/h2ovl-mississippi-800m",
        "naver-hyperclovax/HyperCLOVAX-SEED-Vision-Instruct-3B",
        "HuggingFaceM4/Idefics3-8B-Llama3",
        "internlm/Intern-S1",
        "OpenGVLab/InternVL2-1B",
        "OpenGVLab/InternVL3-1B",
        "OpenGVLab/InternVL3_5-1B",
        "OpenGVLab/InternVL3_5-GPT-OSS-20B-A4B-Preview",
        "OpenGVLab/InternVL3_5-30B-A3B",
        "Kwai-Keye/Keye-VL-8B-Preview",
        "Kwai-Keye/Keye-VL-1_5-8B",
        "moonshotai/Kimi-VL-A3B-Instruct",
        "meta-llama/Llama-4-Scout-17B-16E-Instruct",
        "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",
        "mispeech/midashenglm-7b",
        "openbmb/MiniCPM-Llama3-V-2_5",
        "openbmb/MiniCPM-o-2_6",
        "openbmb/MiniCPM-V-2_6",
        "MiniMaxAI/MiniMax-VL-01",
        "allenai/Molmo-7B-D-0924",
        "allenai/Molmo-7B-O-0924",
        "nvidia/NVLM-D-72B",
        "nvidia/Llama-3.1-Nemotron-Nano-VL-8B-V1",
        "AIDC-AI/Ovis1.6-Gemma2-9B",
        "AIDC-AI/Ovis1.6-Llama3.2-3B",
        "AIDC-AI/Ovis2-1B",
        "AIDC-AI/Ovis2.5-2B",
377
378
        "google/paligemma-3b-mix-224",
        "google/paligemma2-3b-ft-docci-448",
379
380
381
382
383
384
385
386
387
388
389
        "microsoft/Phi-3.5-vision-instruct",
        "microsoft/Phi-4-multimodal-instruct",
        "mistralai/Pixtral-12B-2409",
        "mistral-community/pixtral-12b",
        "Qwen/Qwen-VL-Chat",
        "Qwen/Qwen2-VL-2B-Instruct",
        "Qwen/Qwen2.5-VL-3B-Instruct",
        "Qwen/Qwen2-Audio-7B-Instruct",
        "Qwen/Qwen2.5-Omni-3B",
        "Qwen/Qwen3-VL-4B-Instruct",
        "Qwen/Qwen3-VL-30B-A3B-Instruct",
390
        "Qwen/Qwen3-Omni-30B-A3B-Instruct",
391
392
393
394
395
396
397
398
399
400
401
        "YannQi/R-4B",
        "Skywork/Skywork-R1V-38B",
        "HuggingFaceTB/SmolVLM2-2.2B-Instruct",
        "stepfun-ai/step3",
        "fixie-ai/ultravox-v0_5-llama-3_2-1b",
        "openai/whisper-large-v3",
        "omni-research/Tarsier-7b",
        "omni-research/Tarsier2-Recap-7b",
        "mistralai/Voxtral-Mini-3B-2507",
    ],
)
402
403
404
405
406
407
408
409
410
@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(
    model_id: str,
    hit_rate: float,
    num_batches: int,
    simplify_rate: float,
):
Nicolò Lucchesi's avatar
Nicolò Lucchesi committed
411
412
    if model_id == "google/gemma-3n-E2B-it":
        pytest.skip("Skipping gemma-3n-E2B-it due to transformers #39911 bug.")
413
414
415
416
417
418
    _test_processing_correctness(
        model_id,
        hit_rate=hit_rate,
        num_batches=num_batches,
        simplify_rate=simplify_rate,
    )
419
420


421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
# 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,
    )


443
def _assert_inputs_equal(
444
445
    a: MultiModalInputs,
    b: MultiModalInputs,
446
    *,
447
    ignore_mm_keys: set[str] | None = None,
448
    msg: str = "",
449
):
450
451
452
    if ignore_mm_keys is None:
        ignore_mm_keys = set()

453
454
455
456
457
458
459
    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()
460
461

    for key in ignore_mm_keys:
462
463
        a_data.pop(key, None)
        b_data.pop(key, None)
464

465
    assert a_data == b_data, msg