test_common.py 16.4 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

import numpy as np
zhuwenwen's avatar
zhuwenwen committed
8
import os
9
import pytest
10
11
12
from mistral_common.protocol.instruct.messages import (ImageChunk, TextChunk,
                                                       UserMessage)
from mistral_common.protocol.instruct.request import ChatCompletionRequest
13
14
15
from PIL import Image

from vllm.config import ModelConfig
16
from vllm.multimodal import MULTIMODAL_REGISTRY, MultiModalDataDict
17
from vllm.multimodal.cache import MultiModalProcessorOnlyCache
18
from vllm.multimodal.inputs import MultiModalInputs
19
20
from vllm.multimodal.processing import (BaseMultiModalProcessor,
                                        InputProcessingContext)
21
22
23
from vllm.transformers_utils.tokenizer import (AnyTokenizer, MistralTokenizer,
                                               cached_tokenizer_from_config,
                                               encode_tokens)
24
25

from ....multimodal.utils import random_audio, random_image, random_video
26
from ...registry import HF_EXAMPLE_MODELS
zhuwenwen's avatar
zhuwenwen committed
27
from ....utils import models_path_prefix
28
29


30
31
32
33
34
35
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:
36
        # GLM4.1V doesn't support multiple videos
37
        video = mm_data["video"]
38
        num_frames = len(video)
39
        mm_data["video"] = (video, {
40
41
            "total_num_frames": num_frames,
            "fps": num_frames,
42
            "duration": 1,
43
44
45
            "frames_indices": [i for i in range(num_frames)],
            "video_backend": "opencv",
            "do_sample_frames": True,
46
47
48
49
        })
    return mm_data


50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
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


78
def _test_processing_correctness(
79
    model_id_or_arch: str,
80
81
82
83
    hit_rate: float,
    num_batches: int,
    simplify_rate: float,
):
84
85
86
87
88
89
90
    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
91
92
    model_info.check_available_online(on_fail="skip")
    model_info.check_transformers_version(on_fail="skip")
93
94
95

    model_config = ModelConfig(
        model_id,
96
97
        tokenizer=model_info.tokenizer or model_id,
        tokenizer_mode=model_info.tokenizer_mode,
98
        revision=model_info.revision,
99
        trust_remote_code=model_info.trust_remote_code,
100
        hf_overrides=model_info.hf_overrides,
101
102
        # Ensure that the cache can fit all of the data
        mm_processor_cache_gb=2048,
103
104
105
        skip_tokenizer_init=model_info.skip_tokenizer_init,
        enforce_eager=model_info.enforce_eager,
        dtype=model_info.dtype)
106
107
108
109
110

    model_cls = MULTIMODAL_REGISTRY._get_model_cls(model_config)
    factories = MULTIMODAL_REGISTRY._processor_factories[model_cls]
    ctx = InputProcessingContext(
        model_config,
111
        tokenizer=cached_tokenizer_from_config(model_config),
112
    )
113
    cache = MultiModalProcessorOnlyCache(model_config)
114

115
116
117
118
119
120
121
122
123
    processing_info = factories.info(ctx)
    supported_mm_limits = processing_info.get_supported_mm_limits()
    limit_mm_per_prompt = {
        modality: 3 if limit is None else limit
        for modality, limit in supported_mm_limits.items()
    }

    model_config.get_multimodal_config().limit_per_prompt = limit_mm_per_prompt

124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
    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),
        "audio": (np.zeros((512, )), 16000),
    }
    input_factory = {
        "image":
        partial(random_image, rng, min_wh=128, max_wh=256),
        "video":
        partial(random_video,
                rng,
                min_frames=2,
143
                max_frames=16,
144
145
146
147
148
149
150
151
152
153
                min_wh=128,
                max_wh=256),
        "audio":
        partial(random_audio, rng, min_len=512, max_len=1024, sr=16000),
    }

    for batch_idx in range(num_batches):
        mm_data = {
            k:
            [(input_to_hit[k] if rng.rand() < hit_rate else input_factory[k]())
154
             for _ in range(rng.randint(limit + 1))]
155
            for k, limit in limit_mm_per_prompt.items()
156
157
158
        }

        mm_counts = {k: len(vs) for k, vs in mm_data.items()}
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175

        # Mistral chat outputs tokens directly, rather than text prompts
        if isinstance(tokenizer, MistralTokenizer):
            images = mm_data.get("image", [])
            request = ChatCompletionRequest(messages=[
                UserMessage(content=[
                    TextChunk(text=""),
                    *(ImageChunk(image=image) for image in images),
                ]),
            ])
            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
176
177
178
179
180
181
182
183
184

        # 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]

185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
        _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,
202
    "ovis2_5": False,
203
    "paligemma": False,
204
205
206
207
208
209
210
211
212
213
214
    "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"},
}

215
MM_DATA_PATCHES = {
216
    # GLM4.1V and Qwen3-VL requires video metadata to be included in the input
217
    "glm4v": glm4_1v_patch_mm_data,
218
    "glm4v_moe": glm4_1v_patch_mm_data,
219
220
    "qwen3_vl": qwen3_vl_patch_mm_data,
    "qwen3_vl_moe": qwen3_vl_patch_mm_data,
221
222
}

223
224

def _test_processing_correctness_one(
225
    model_config: ModelConfig,
226
227
    tokenizer: AnyTokenizer,
    prompt: Union[str, list[int]],
228
229
230
231
232
    mm_data: MultiModalDataDict,
    baseline_processor: BaseMultiModalProcessor,
    cached_processor: BaseMultiModalProcessor,
    batch_idx: int,
):
233
234
    model_type = model_config.hf_config.model_type
    ignore_mm_keys = _IGNORE_MM_KEYS.get(model_type, set[str]())
235
236
    if model_type in MM_DATA_PATCHES:
        mm_data = MM_DATA_PATCHES[model_type](mm_data)
237
238
239
240
241
242
243
244

    if isinstance(prompt, str):
        text_prompt = prompt
        token_prompt = encode_tokens(
            tokenizer,
            prompt,
            add_special_tokens=_ADD_SPECIAL_TOKENS_OVERRIDES.get(model_type),
        )
245
    else:
246
247
248
        # Mistral does not support decode_tokens with skip_special_tokens=False
        text_prompt = None
        token_prompt = prompt
249
250
251
252
253
254
255
256
257
258
259
260
261

    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={},
    )

262
    _assert_inputs_equal(
263
264
        baseline_tokenized_result,
        cached_tokenized_result,
265
        ignore_mm_keys=ignore_mm_keys,
266
        msg=f"Failed ({batch_idx=}, {token_prompt=}, {mm_data=})",
267
    )
268

269
270
271
272
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
    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,
            msg=f"Failed ({batch_idx=}, {text_prompt=}, "
            f"{token_prompt=}, {mm_data=})",
        )

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

304
305

# yapf: disable
306
@pytest.mark.parametrize("model_id", [
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
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
    os.path.join(models_path_prefix, "rhymes-ai/Aria"),
    os.path.join(models_path_prefix, "CohereForAI/aya-vision-8b"),
    os.path.join(models_path_prefix, "Salesforce/blip2-opt-2.7b"),
    os.path.join(models_path_prefix, "facebook/chameleon-7b"),
    os.path.join(models_path_prefix, "CohereLabs/command-a-vision-07-2025"),
    os.path.join(models_path_prefix, "deepseek-ai/deepseek-vl2-tiny"),
    os.path.join(models_path_prefix, "baidu/ERNIE-4.5-VL-28B-A3B-PT"),
    os.path.join(models_path_prefix, "adept/fuyu-8b"),
    os.path.join(models_path_prefix, "google/gemma-3-4b-it"),
    os.path.join(models_path_prefix, "google/gemma-3n-E2B-it"),
    os.path.join(models_path_prefix, "zai-org/glm-4v-9b"),
    os.path.join(models_path_prefix, "zai-org/GLM-4.1V-9B-Thinking"),
    os.path.join(models_path_prefix, "zai-org/GLM-4.5V"),
    os.path.join(models_path_prefix, "ibm-granite/granite-speech-3.3-2b"),
    os.path.join(models_path_prefix, "h2oai/h2ovl-mississippi-800m"),
    os.path.join(models_path_prefix, "naver-hyperclovax/HyperCLOVAX-SEED-Vision-Instruct-3B"),
    os.path.join(models_path_prefix, "HuggingFaceM4/Idefics3-8B-Llama3"),
    os.path.join(models_path_prefix, "internlm/Intern-S1"),
    os.path.join(models_path_prefix, "OpenGVLab/InternVL2-1B"),
    os.path.join(models_path_prefix, "OpenGVLab/InternVL3-1B"),
    os.path.join(models_path_prefix, "OpenGVLab/InternVL3_5-1B"),
    os.path.join(models_path_prefix, "OpenGVLab/InternVL3_5-GPT-OSS-20B-A4B-Preview"),
    os.path.join(models_path_prefix, "OpenGVLab/InternVL3_5-30B-A3B"),
    os.path.join(models_path_prefix, "Kwai-Keye/Keye-VL-8B-Preview"),
    os.path.join(models_path_prefix, "Kwai-Keye/Keye-VL-1_5-8B"),
    os.path.join(models_path_prefix, "moonshotai/Kimi-VL-A3B-Instruct"),
    os.path.join(models_path_prefix, "meta-llama/Llama-4-Scout-17B-16E-Instruct"),
    os.path.join(models_path_prefix, "llava-hf/llava-1.5-7b-hf"),
    os.path.join(models_path_prefix, "llava-hf/llava-v1.6-mistral-7b-hf"),
    os.path.join(models_path_prefix, "llava-hf/LLaVA-NeXT-Video-7B-hf"),
    os.path.join(models_path_prefix, "llava-hf/llava-onevision-qwen2-0.5b-ov-hf"),
    os.path.join(models_path_prefix, "TIGER-Lab/Mantis-8B-siglip-llama3"),
    os.path.join(models_path_prefix, "mispeech/midashenglm-7b"),
    os.path.join(models_path_prefix, "openbmb/MiniCPM-Llama3-V-2_5"),
    os.path.join(models_path_prefix, "openbmb/MiniCPM-o-2_6"),
    os.path.join(models_path_prefix, "openbmb/MiniCPM-V-2_6"),
    os.path.join(models_path_prefix, "MiniMaxAI/MiniMax-VL-01"),
    os.path.join(models_path_prefix, "allenai/Molmo-7B-D-0924"),
    os.path.join(models_path_prefix, "allenai/Molmo-7B-O-0924"),
    os.path.join(models_path_prefix, "nvidia/NVLM-D-72B"),
    os.path.join(models_path_prefix, "nvidia/Llama-3.1-Nemotron-Nano-VL-8B-V1"),
    os.path.join(models_path_prefix, "AIDC-AI/Ovis1.6-Gemma2-9B"),
    os.path.join(models_path_prefix, "AIDC-AI/Ovis1.6-Llama3.2-3B"),
    os.path.join(models_path_prefix, "AIDC-AI/Ovis2-1B"),
    os.path.join(models_path_prefix, "AIDC-AI/Ovis2.5-2B"),
    os.path.join(models_path_prefix, "google/paligemma-3b-mix-224"),
    os.path.join(models_path_prefix, "google/paligemma2-3b-ft-docci-448"),
    os.path.join(models_path_prefix, "microsoft/Phi-3.5-vision-instruct"),
    os.path.join(models_path_prefix, "microsoft/Phi-4-multimodal-instruct"),
    os.path.join(models_path_prefix, "mistralai/Pixtral-12B-2409"),
    os.path.join(models_path_prefix, "mistral-community/pixtral-12b"),
    os.path.join(models_path_prefix, "Qwen/Qwen-VL-Chat"),
    os.path.join(models_path_prefix, "Qwen/Qwen2-VL-2B-Instruct"),
    os.path.join(models_path_prefix, "Qwen/Qwen2.5-VL-3B-Instruct"),
    os.path.join(models_path_prefix, "Qwen/Qwen2-Audio-7B-Instruct"),
    os.path.join(models_path_prefix, "Qwen/Qwen2.5-Omni-3B"),
363
364
    os.path.join(models_path_prefix, "Qwen/Qwen3-VL-4B-Instruct"),
    os.path.join(models_path_prefix, "Qwen/Qwen3-VL-30B-A3B-Instruct"),
365
366
367
368
369
370
371
372
373
    os.path.join(models_path_prefix, "YannQi/R-4B"),
    os.path.join(models_path_prefix, "Skywork/Skywork-R1V-38B"),
    os.path.join(models_path_prefix, "HuggingFaceTB/SmolVLM2-2.2B-Instruct"),
    os.path.join(models_path_prefix, "stepfun-ai/step3"),
    os.path.join(models_path_prefix, "fixie-ai/ultravox-v0_5-llama-3_2-1b"),
    os.path.join(models_path_prefix, "openai/whisper-large-v3"),
    os.path.join(models_path_prefix, "omni-research/Tarsier-7b"),
    os.path.join(models_path_prefix, "omni-research/Tarsier2-Recap-7b"),
    os.path.join(models_path_prefix, "mistralai/Voxtral-Mini-3B-2507"),
374
375
376
377
378
379
380
381
382
383
384
])
@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
385
386
    if model_id == "google/gemma-3n-E2B-it":
        pytest.skip("Skipping gemma-3n-E2B-it due to transformers #39911 bug.")
387
388
389
390
391
392
    _test_processing_correctness(
        model_id,
        hit_rate=hit_rate,
        num_batches=num_batches,
        simplify_rate=simplify_rate,
    )
393
394


395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
# 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,
    )


417
def _assert_inputs_equal(
418
419
    a: MultiModalInputs,
    b: MultiModalInputs,
420
421
422
    *,
    ignore_mm_keys: Optional[set[str]] = None,
    msg: str = "",
423
):
424
425
426
    if ignore_mm_keys is None:
        ignore_mm_keys = set()

427
428
429
430
431
432
433
    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()
434
435

    for key in ignore_mm_keys:
436
437
        a_data.pop(key, None)
        b_data.pop(key, None)
438

439
    assert a_data == b_data, msg