test_common.py 12 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
10
11
from mistral_common.protocol.instruct.messages import (ImageChunk, TextChunk,
                                                       UserMessage)
from mistral_common.protocol.instruct.request import ChatCompletionRequest
12
13
14
15
from PIL import Image

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

from ....multimodal.utils import random_audio, random_image, random_video
24
from ...registry import HF_EXAMPLE_MODELS
25
26


27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
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:
        video = mm_data["video"]
        mm_data["video"] = (video, {
            "total_num_frames": len(video),
            "fps": len(video),
            "duration": 1,
            "video_backend": "opencv"
        })
    return mm_data


43
def _test_processing_correctness(
44
    model_id_or_arch: str,
45
46
47
48
    hit_rate: float,
    num_batches: int,
    simplify_rate: float,
):
49
50
51
52
53
54
55
    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
56
57
    model_info.check_available_online(on_fail="skip")
    model_info.check_transformers_version(on_fail="skip")
58
59
60
61

    model_config = ModelConfig(
        model_id,
        task="auto",
62
63
        tokenizer=model_info.tokenizer or model_id,
        tokenizer_mode=model_info.tokenizer_mode,
64
        trust_remote_code=model_info.trust_remote_code,
65
        seed=0,
66
        dtype="auto",
67
        revision=model_info.revision,
68
        hf_overrides=model_info.hf_overrides,
69
70
71
72
73
74
    )

    model_cls = MULTIMODAL_REGISTRY._get_model_cls(model_config)
    factories = MULTIMODAL_REGISTRY._processor_factories[model_cls]
    ctx = InputProcessingContext(
        model_config,
75
        tokenizer=cached_tokenizer_from_config(model_config),
76
77
    )
    # Ensure that it can fit all of the data
78
    cache = ProcessingCache(capacity_gb=2048)
79

80
81
82
83
84
85
86
87
88
    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

89
90
91
92
93
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
    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,
                max_frames=8,
                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]())
119
             for _ in range(rng.randint(limit + 1))]
120
            for k, limit in limit_mm_per_prompt.items()
121
122
123
        }

        mm_counts = {k: len(vs) for k, vs in mm_data.items()}
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140

        # 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
141
142
143
144
145
146
147
148
149

        # 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]

150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
        _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 = {
    "mllama": False,
    "ovis": False,
168
    "paligemma": False,
169
170
171
172
173
174
175
176
177
178
179
    "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"},
}

180
181
182
183
184
MM_DATA_PATCHES = {
    # GLM4.1V requires video metadata to be included in the input
    "glm4v": glm4_1v_patch_mm_data,
}

185
186

def _test_processing_correctness_one(
187
    model_config: ModelConfig,
188
189
    tokenizer: AnyTokenizer,
    prompt: Union[str, list[int]],
190
191
192
193
194
    mm_data: MultiModalDataDict,
    baseline_processor: BaseMultiModalProcessor,
    cached_processor: BaseMultiModalProcessor,
    batch_idx: int,
):
195
196
    model_type = model_config.hf_config.model_type
    ignore_mm_keys = _IGNORE_MM_KEYS.get(model_type, set[str]())
197
198
    if model_type in MM_DATA_PATCHES:
        mm_data = MM_DATA_PATCHES[model_type](mm_data)
199
200
201
202
203
204
205
206

    if isinstance(prompt, str):
        text_prompt = prompt
        token_prompt = encode_tokens(
            tokenizer,
            prompt,
            add_special_tokens=_ADD_SPECIAL_TOKENS_OVERRIDES.get(model_type),
        )
207
    else:
208
209
210
        # Mistral does not support decode_tokens with skip_special_tokens=False
        text_prompt = None
        token_prompt = prompt
211
212
213
214
215
216
217
218
219
220
221
222
223

    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={},
    )

224
    _assert_inputs_equal(
225
226
        baseline_tokenized_result,
        cached_tokenized_result,
227
        ignore_mm_keys=ignore_mm_keys,
228
        msg=f"Failed ({batch_idx=}, {token_prompt=}, {mm_data=})",
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
    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=})",
        )

266
267

# yapf: disable
268
269
@pytest.mark.parametrize("model_id", [
    "rhymes-ai/Aria",
Jennifer Zhao's avatar
Jennifer Zhao committed
270
    "CohereForAI/aya-vision-8b",
271
272
273
    "Salesforce/blip2-opt-2.7b",
    "facebook/chameleon-7b",
    "deepseek-ai/deepseek-vl2-tiny",
274
    "microsoft/Florence-2-base",
275
    "adept/fuyu-8b",
276
    "google/gemma-3-4b-it",
277
    "THUDM/glm-4v-9b",
278
    "THUDM/GLM-4.1V-9B-Thinking",
279
    "ibm-granite/granite-speech-3.3-2b",
280
281
    "h2oai/h2ovl-mississippi-800m",
    "OpenGVLab/InternVL2-1B",
282
    "OpenGVLab/InternVL3-1B",
283
    "HuggingFaceM4/Idefics3-8B-Llama3",
284
    "HuggingFaceTB/SmolVLM2-2.2B-Instruct",
285
    "moonshotai/Kimi-VL-A3B-Instruct",
286
    "meta-llama/Llama-4-Scout-17B-16E-Instruct",
287
    "naver-hyperclovax/HyperCLOVAX-SEED-Vision-Instruct-3B",
288
289
290
291
    "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",
292
    "meta-llama/Llama-3.2-11B-Vision-Instruct",
293
    "TIGER-Lab/Mantis-8B-siglip-llama3",
294
    "openbmb/MiniCPM-Llama3-V-2_5",
295
296
    "openbmb/MiniCPM-o-2_6",
    "openbmb/MiniCPM-V-2_6",
297
    "MiniMaxAI/MiniMax-VL-01",
298
299
    "allenai/Molmo-7B-D-0924",
    "allenai/Molmo-7B-O-0924",
300
    "nvidia/NVLM-D-72B",
301
    "nvidia/Llama-3.1-Nemotron-Nano-VL-8B-V1",
302
303
    "AIDC-AI/Ovis1.6-Gemma2-9B",
    "AIDC-AI/Ovis1.6-Llama3.2-3B",
304
    "AIDC-AI/Ovis2-1B",
305
306
    "google/paligemma-3b-mix-224",
    "google/paligemma2-3b-ft-docci-448",
307
    "microsoft/Phi-3.5-vision-instruct",
308
    "microsoft/Phi-4-multimodal-instruct",
309
310
    "mistralai/Pixtral-12B-2409",
    "mistral-community/pixtral-12b",
311
312
    "Qwen/Qwen-VL-Chat",
    "Qwen/Qwen2-VL-2B-Instruct",
Roger Wang's avatar
Roger Wang committed
313
    "Qwen/Qwen2.5-VL-3B-Instruct",
314
    "Qwen/Qwen2-Audio-7B-Instruct",
315
    "Qwen/Qwen2.5-Omni-3B",
316
    "Skywork/Skywork-R1V-38B",
317
    "fixie-ai/ultravox-v0_5-llama-3_2-1b",
318
    "openai/whisper-large-v3",
汪志鹏's avatar
汪志鹏 committed
319
    "omni-research/Tarsier-7b",
320
    "omni-research/Tarsier2-Recap-7b"
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
])
@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,
):
    _test_processing_correctness(
        model_id,
        hit_rate=hit_rate,
        num_batches=num_batches,
        simplify_rate=simplify_rate,
    )
338
339


340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
# 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,
    )


362
def _assert_inputs_equal(
363
364
    a: MultiModalInputs,
    b: MultiModalInputs,
365
366
367
    *,
    ignore_mm_keys: Optional[set[str]] = None,
    msg: str = "",
368
):
369
370
371
    if ignore_mm_keys is None:
        ignore_mm_keys = set()

372
    assert "mm_kwargs" in a and "mm_kwargs" in b, msg
373
374
375
376
377

    for key in ignore_mm_keys:
        a["mm_kwargs"].pop(key, None)
        b["mm_kwargs"].pop(key, None)

378
    assert a == b, msg