test_common.py 49.1 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3
4
5
"""Common tests for testing .generate() functionality for single / multiple
image, embedding, and video support for different VLMs in vLLM.
"""
6

7
8
import math
from collections import defaultdict
9
10
11
from pathlib import PosixPath

import pytest
12
from packaging.version import Version
13
14
15
16
17
from transformers import (
    AutoModel,
    AutoModelForImageTextToText,
    AutoModelForTextToWaveform,
)
18
from transformers import __version__ as TRANSFORMERS_VERSION
19
20

from vllm.platforms import current_platform
21
from vllm.utils.func_utils import identity
22

23
24
25
26
27
28
29
30
31
from ....conftest import (
    IMAGE_ASSETS,
    AudioTestAssets,
    HfRunner,
    ImageTestAssets,
    VideoTestAssets,
    VllmRunner,
)
from ....utils import create_new_process_for_each_test, large_gpu_mark, multi_gpu_marks
32
33
34
from ...utils import check_outputs_equal
from .vlm_utils import custom_inputs, model_utils, runners
from .vlm_utils.case_filtering import get_parametrized_options
35
36
37
38
39
40
from .vlm_utils.types import (
    CustomTestOptions,
    ExpandableVLMTestArgs,
    VLMTestInfo,
    VLMTestType,
)
41
42
43
44
45
46

COMMON_BROADCAST_SETTINGS = {
    "test_type": VLMTestType.IMAGE,
    "dtype": "half",
    "max_tokens": 5,
    "tensor_parallel_size": 2,
47
    "hf_model_kwargs": {"device_map": "auto"},
48
    "image_size_factors": [(0.25, 0.5, 1.0)],
49
50
51
    "distributed_executor_backend": (
        "ray",
        "mp",
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
78
79
80
81
82
83
84
85
86
87
88
}

### Test configuration for specific models
# NOTE: The convention of the test settings below is to lead each test key
# with the name of the model arch used in the test, using underscores in place
# of hyphens; this makes it more convenient to filter tests for a specific kind
# of model. For example....
#
# To run all test types for a specific key:
#     use the k flag to substring match with a leading square bracket; if the
#     model arch happens to be a substring of another one, you can add a
#     trailing hyphen. E.g.,
#                 - pytest $TEST_FILE -k "[llava-"
#     prevents matching on "[llava_next-" & will match just the enabled cases
#     for llava, i.e., single image, image embedding, and custom input tests.
#
# To run a test for a Test Info for just one of multiple models:
#     use the k flag to substring match the model name, e.g.,
#                 - pytest $TEST_FILE -k OpenGVLab/InternVL2-1B
#     prevents matching on nGVLab/InternVL2-2B.
#
# You can also combine substrings to match more granularly.
#     ex 1:
#        pytest $TEST_FILE -k "test_single_image and OpenGVLab/InternVL2-1B"
#     will run only test_single_image* for OpenGVLab/InternVL2-1B; this would
#     match both wrappers for single image tests, since it also matches
#     test_single_image_heavy (which forks if we have a distributed backend)
#     ex 2:
#        pytest $TEST_FILE -k  "[llava- or [intern_vl-"
#     will run all of the tests for only llava & internvl.
#
# NOTE you can add --collect-only to any of the above commands to see
# which cases would be selected and deselected by pytest. In general,
# this is a good idea for checking your command first, since tests are slow.

VLM_TEST_SETTINGS = {
89
90
91
    #### Core tests to always run in the CI
    "llava": VLMTestInfo(
        models=["llava-hf/llava-1.5-7b-hf"],
92
        test_type=(VLMTestType.EMBEDDING, VLMTestType.IMAGE, VLMTestType.CUSTOM_INPUTS),
93
94
95
        prompt_formatter=lambda img_prompt: f"USER: {img_prompt}\nASSISTANT:",
        convert_assets_to_embeddings=model_utils.get_llava_embeddings,
        max_model_len=4096,
96
        auto_cls=AutoModelForImageTextToText,
97
        vllm_output_post_proc=model_utils.llava_image_vllm_to_hf_output,
98
99
100
101
102
103
104
105
        custom_test_opts=[
            CustomTestOptions(
                inputs=custom_inputs.multi_image_multi_aspect_ratio_inputs(
                    formatter=lambda img_prompt: f"USER: {img_prompt}\nASSISTANT:"
                ),
                limit_mm_per_prompt={"image": 4},
            )
        ],
106
        vllm_runner_kwargs={"enable_mm_embeds": True},
107
        marks=[pytest.mark.core_model, pytest.mark.cpu_model],
108
    ),
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
    "paligemma": VLMTestInfo(
        models=["google/paligemma-3b-mix-224"],
        test_type=VLMTestType.IMAGE,
        prompt_formatter=identity,
        img_idx_to_prompt=lambda idx: "",
        # Paligemma uses its own sample prompts because the default one fails
        single_image_prompts=IMAGE_ASSETS.prompts(
            {
                "stop_sign": "caption es",
                "cherry_blossom": "What is in the picture?",
            }
        ),
        auto_cls=AutoModelForImageTextToText,
        vllm_output_post_proc=model_utils.paligemma_vllm_to_hf_output,
        dtype="bfloat16",
        marks=[
            pytest.mark.skip(reason="vLLM does not support PrefixLM attention mask")
        ],
    ),
Roger Wang's avatar
Roger Wang committed
128
129
    "qwen2_5_vl": VLMTestInfo(
        models=["Qwen/Qwen2.5-VL-3B-Instruct"],
130
131
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE, VLMTestType.VIDEO),
        prompt_formatter=lambda img_prompt: f"<|im_start|>User\n{img_prompt}<|im_end|>\n<|im_start|>assistant\n",  # noqa: E501
132
133
        img_idx_to_prompt=lambda idx: "<|vision_start|><|image_pad|><|vision_end|>",
        video_idx_to_prompt=lambda idx: "<|vision_start|><|video_pad|><|vision_end|>",
Roger Wang's avatar
Roger Wang committed
134
135
        max_model_len=4096,
        max_num_seqs=2,
136
        auto_cls=AutoModelForImageTextToText,
Roger Wang's avatar
Roger Wang committed
137
138
        vllm_output_post_proc=model_utils.qwen2_vllm_to_hf_output,
        image_size_factors=[(), (0.25,), (0.25, 0.25, 0.25), (0.25, 0.2, 0.15)],
139
        marks=[pytest.mark.core_model, pytest.mark.cpu_model],
Roger Wang's avatar
Roger Wang committed
140
    ),
141
    "qwen2_5_omni": VLMTestInfo(
142
        models=["Qwen/Qwen2.5-Omni-3B"],
143
144
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE, VLMTestType.VIDEO),
        prompt_formatter=lambda img_prompt: f"<|im_start|>User\n{img_prompt}<|im_end|>\n<|im_start|>assistant\n",  # noqa: E501
145
146
        img_idx_to_prompt=lambda idx: "<|vision_bos|><|IMAGE|><|vision_eos|>",
        video_idx_to_prompt=lambda idx: "<|vision_bos|><|VIDEO|><|vision_eos|>",
147
148
        max_model_len=4096,
        max_num_seqs=2,
149
        num_logprobs=6 if current_platform.is_cpu() else 5,
150
        auto_cls=AutoModelForTextToWaveform,
151
        vllm_output_post_proc=model_utils.qwen2_vllm_to_hf_output,
152
        patch_hf_runner=model_utils.qwen2_5_omni_patch_hf_runner,
153
154
155
        image_size_factors=[(), (0.25,), (0.25, 0.25, 0.25), (0.25, 0.2, 0.15)],
        marks=[pytest.mark.core_model, pytest.mark.cpu_model],
    ),
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
    "qwen3_vl": VLMTestInfo(
        models=["Qwen/Qwen3-VL-4B-Instruct"],
        test_type=(
            VLMTestType.IMAGE,
            VLMTestType.MULTI_IMAGE,
            VLMTestType.VIDEO,
        ),
        needs_video_metadata=True,
        prompt_formatter=lambda img_prompt: f"<|im_start|>User\n{img_prompt}<|im_end|>\n<|im_start|>assistant\n",  # noqa: E501
        img_idx_to_prompt=lambda idx: "<|vision_start|><|image_pad|><|vision_end|>",  # noqa: E501
        video_idx_to_prompt=lambda idx: "<|vision_start|><|video_pad|><|vision_end|>",  # noqa: E501
        max_model_len=4096,
        max_num_seqs=2,
        num_logprobs=20,
        auto_cls=AutoModelForImageTextToText,
        vllm_output_post_proc=model_utils.qwen2_vllm_to_hf_output,
        patch_hf_runner=model_utils.qwen3_vl_patch_hf_runner,
        image_size_factors=[(), (0.25,), (0.25, 0.25, 0.25), (0.25, 0.2, 0.15)],
        marks=[
            pytest.mark.core_model,
        ],
    ),
178
    "ultravox": VLMTestInfo(
179
        models=["fixie-ai/ultravox-v0_5-llama-3_2-1b"],
180
        test_type=VLMTestType.AUDIO,
181
        prompt_formatter=lambda audio_prompt: f"<|begin_of_text|><|start_header_id|>user<|end_header_id|>\n\n{audio_prompt}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n",  # noqa: E501
182
183
184
185
186
187
188
        audio_idx_to_prompt=lambda idx: "<|audio|>",
        max_model_len=4096,
        max_num_seqs=2,
        auto_cls=AutoModel,
        hf_output_post_proc=model_utils.ultravox_trunc_hf_output,
        marks=[pytest.mark.core_model, pytest.mark.cpu_model],
    ),
189
190
191
192
193
194
    #### Transformers fallback to test
    ## To reduce test burden, we only test batching arbitrary image size
    # Dynamic image length and number of patches
    "llava-onevision-transformers": VLMTestInfo(
        models=["llava-hf/llava-onevision-qwen2-0.5b-ov-hf"],
        test_type=VLMTestType.IMAGE,
195
        prompt_formatter=lambda vid_prompt: f"<|im_start|>user\n{vid_prompt}<|im_end|>\n<|im_start|>assistant\n",  # noqa: E501
196
        max_model_len=16384,
197
198
        hf_model_kwargs=model_utils.llava_onevision_hf_model_kwargs(
            "llava-hf/llava-onevision-qwen2-0.5b-ov-hf"
199
        ),
200
201
202
203
204
        auto_cls=AutoModelForImageTextToText,
        vllm_output_post_proc=model_utils.llava_onevision_vllm_to_hf_output,
        image_size_factors=[(0.25, 0.5, 1.0)],
        vllm_runner_kwargs={
            "model_impl": "transformers",
205
            "default_torch_num_threads": 1,
206
        },
207
208
209
        # FIXME: Investigate why the test hangs
        # when processing the 3rd prompt in vLLM
        marks=[pytest.mark.core_model, pytest.mark.skip(reason="Test hangs")],
210
    ),
211
212
213
    # Gemma3 has bidirectional mask on images
    "gemma3-transformers": VLMTestInfo(
        models=["google/gemma-3-4b-it"],
214
215
216
        test_type=VLMTestType.IMAGE,
        prompt_formatter=lambda vid_prompt: f"<'<bos><start_of_turn>user\n{vid_prompt}<start_of_image><end_of_turn>\n<start_of_turn>model\n",  # noqa: E501
        max_model_len=4096,
217
218
219
220
221
222
223
224
        auto_cls=AutoModelForImageTextToText,
        vllm_output_post_proc=model_utils.gemma3_vllm_to_hf_output,
        image_size_factors=[(0.25, 0.5, 1.0)],
        vllm_runner_kwargs={
            "model_impl": "transformers",
        },
        marks=[pytest.mark.core_model],
    ),
225
226
227
    "idefics3-transformers": VLMTestInfo(
        models=["HuggingFaceTB/SmolVLM-256M-Instruct"],
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
228
        prompt_formatter=lambda img_prompt: f"<|begin_of_text|>User:{img_prompt}<end_of_utterance>\nAssistant:",  # noqa: E501
229
230
231
232
233
234
235
236
237
238
239
        img_idx_to_prompt=lambda idx: "<image>",
        max_model_len=8192,
        max_num_seqs=2,
        auto_cls=AutoModelForImageTextToText,
        hf_output_post_proc=model_utils.idefics3_trunc_hf_output,
        image_size_factors=[(0.25, 0.5, 1.0)],
        vllm_runner_kwargs={
            "model_impl": "transformers",
        },
        marks=[pytest.mark.core_model],
    ),
240
241
242
243
    # Pixel values from processor are not 4D or 5D arrays
    "qwen2_5_vl-transformers": VLMTestInfo(
        models=["Qwen/Qwen2.5-VL-3B-Instruct"],
        test_type=VLMTestType.IMAGE,
244
        prompt_formatter=lambda img_prompt: f"<|im_start|>User\n{img_prompt}<|im_end|>\n<|im_start|>assistant\n",  # noqa: E501
245
        img_idx_to_prompt=lambda idx: "<|vision_start|><|image_pad|><|vision_end|>",
246
247
248
249
250
251
252
253
        max_model_len=4096,
        max_num_seqs=2,
        auto_cls=AutoModelForImageTextToText,
        vllm_output_post_proc=model_utils.qwen2_vllm_to_hf_output,
        image_size_factors=[(0.25, 0.2, 0.15)],
        vllm_runner_kwargs={
            "model_impl": "transformers",
        },
254
        marks=[large_gpu_mark(min_gb=32)],
255
    ),
256
    #### Extended model tests
257
258
259
    "aria": VLMTestInfo(
        models=["rhymes-ai/Aria"],
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
260
        prompt_formatter=lambda img_prompt: f"<|im_start|>user\n{img_prompt}<|im_end|>\n<|im_start|>assistant\n ",  # noqa: E501
261
262
263
264
        img_idx_to_prompt=lambda idx: "<fim_prefix><|img|><fim_suffix>\n",
        max_model_len=4096,
        max_num_seqs=2,
        auto_cls=AutoModelForImageTextToText,
265
266
267
        single_image_prompts=IMAGE_ASSETS.prompts(
            {
                "stop_sign": "<vlm_image>Please describe the image shortly.",
268
                "cherry_blossom": "<vlm_image>Please infer the season with reason.",
269
270
            }
        ),
271
        multi_image_prompt="<vlm_image><vlm_image>Describe the two images shortly.",
272
273
274
275
276
        stop_str=["<|im_end|>"],
        image_size_factors=[(0.10, 0.15)],
        max_tokens=64,
        marks=[large_gpu_mark(min_gb=64)],
    ),
Jennifer Zhao's avatar
Jennifer Zhao committed
277
278
    "aya_vision": VLMTestInfo(
        models=["CohereForAI/aya-vision-8b"],
279
        test_type=(VLMTestType.IMAGE),
280
281
282
        prompt_formatter=lambda img_prompt: f"<|START_OF_TURN_TOKEN|><|USER_TOKEN|>{img_prompt}<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>",  # noqa: E501
        single_image_prompts=IMAGE_ASSETS.prompts(
            {
283
284
                "stop_sign": "<image>What's the content in the center of the image?",
                "cherry_blossom": "<image>What is the season?",
285
286
            }
        ),
287
        multi_image_prompt="<image><image>Describe the two images in detail.",
288
289
290
291
292
293
294
295
        max_model_len=4096,
        max_num_seqs=2,
        auto_cls=AutoModelForImageTextToText,
        vllm_runner_kwargs={"mm_processor_kwargs": {"crop_to_patches": True}},
    ),
    "aya_vision-multi_image": VLMTestInfo(
        models=["CohereForAI/aya-vision-8b"],
        test_type=(VLMTestType.MULTI_IMAGE),
296
297
298
        prompt_formatter=lambda img_prompt: f"<|START_OF_TURN_TOKEN|><|USER_TOKEN|>{img_prompt}<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>",  # noqa: E501
        single_image_prompts=IMAGE_ASSETS.prompts(
            {
299
300
                "stop_sign": "<image>What's the content in the center of the image?",
                "cherry_blossom": "<image>What is the season?",
301
302
            }
        ),
303
        multi_image_prompt="<image><image>Describe the two images in detail.",
304
        max_model_len=4096,
Jennifer Zhao's avatar
Jennifer Zhao committed
305
306
        max_num_seqs=2,
        auto_cls=AutoModelForImageTextToText,
307
308
        vllm_runner_kwargs={"mm_processor_kwargs": {"crop_to_patches": True}},
        marks=[large_gpu_mark(min_gb=32)],
Jennifer Zhao's avatar
Jennifer Zhao committed
309
    ),
310
    "blip2": VLMTestInfo(
311
        models=["Salesforce/blip2-opt-2.7b"],
312
313
314
        test_type=VLMTestType.IMAGE,
        prompt_formatter=lambda img_prompt: f"Question: {img_prompt} Answer:",
        img_idx_to_prompt=lambda idx: "",
315
        auto_cls=AutoModelForImageTextToText,
316
        vllm_output_post_proc=model_utils.blip2_vllm_to_hf_output,
317
318
        # FIXME: https://github.com/huggingface/transformers/pull/38510
        marks=[pytest.mark.skip("Model is broken")],
319
320
321
322
323
324
    ),
    "chameleon": VLMTestInfo(
        models=["facebook/chameleon-7b"],
        test_type=VLMTestType.IMAGE,
        prompt_formatter=lambda img_prompt: f"USER: {img_prompt}\nASSISTANT:",
        max_model_len=4096,
325
        max_num_seqs=2,
326
        auto_cls=AutoModelForImageTextToText,
327
        # For chameleon, we only compare the sequences
328
329
        vllm_output_post_proc=lambda vllm_output, model: vllm_output[:2],
        hf_output_post_proc=lambda hf_output, model: hf_output[:2],
330
331
332
333
        comparator=check_outputs_equal,
        max_tokens=8,
        dtype="bfloat16",
    ),
334
    "deepseek_vl_v2": VLMTestInfo(
335
        models=["Isotr0py/deepseek-vl2-tiny"],  # model repo using dynamic module
336
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
337
        prompt_formatter=lambda img_prompt: f"<|User|>: {img_prompt}\n\n<|Assistant|>: ",  # noqa: E501
338
339
        max_model_len=4096,
        max_num_seqs=2,
340
341
        single_image_prompts=IMAGE_ASSETS.prompts(
            {
342
                "stop_sign": "<image>\nWhat's the content in the center of the image?",
343
344
345
346
                "cherry_blossom": "<image>\nPlease infer the season with reason in details.",  # noqa: E501
            }
        ),
        multi_image_prompt="image_1:<image>\nimage_2:<image>\nWhich image can we see the car and the tower?",  # noqa: E501
347
348
        patch_hf_runner=model_utils.deepseekvl2_patch_hf_runner,
        hf_output_post_proc=model_utils.deepseekvl2_trunc_hf_output,
349
        stop_str=["<|end▁of▁sentence|>", "<|begin▁of▁sentence|>"],
350
        image_size_factors=[(), (1.0,), (1.0, 1.0, 1.0), (0.1, 0.5, 1.0)],
351
    ),
352
353
354
355
356
357
358
    "fuyu": VLMTestInfo(
        models=["adept/fuyu-8b"],
        test_type=VLMTestType.IMAGE,
        prompt_formatter=lambda img_prompt: f"{img_prompt}\n",
        img_idx_to_prompt=lambda idx: "",
        max_model_len=2048,
        max_num_seqs=2,
359
        auto_cls=AutoModelForImageTextToText,
360
361
362
363
        use_tokenizer_eos=True,
        vllm_output_post_proc=model_utils.fuyu_vllm_to_hf_output,
        num_logprobs=10,
        image_size_factors=[(), (0.25,), (0.25, 0.25, 0.25), (0.25, 0.2, 0.15)],
364
        marks=[large_gpu_mark(min_gb=32)],
365
    ),
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
    "gemma3": VLMTestInfo(
        models=["google/gemma-3-4b-it"],
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
        prompt_formatter=lambda img_prompt: f"<bos><start_of_turn>user\n{img_prompt}<end_of_turn>\n<start_of_turn>model\n",  # noqa: E501
        single_image_prompts=IMAGE_ASSETS.prompts(
            {
                "stop_sign": "<start_of_image>What's the content in the center of the image?",  # noqa: E501
                "cherry_blossom": "<start_of_image>What is the season?",
            }
        ),
        multi_image_prompt="<start_of_image><start_of_image>Describe the two images in detail.",  # noqa: E501
        max_model_len=4096,
        max_num_seqs=2,
        auto_cls=AutoModelForImageTextToText,
        vllm_runner_kwargs={"mm_processor_kwargs": {"do_pan_and_scan": True}},
        patch_hf_runner=model_utils.gemma3_patch_hf_runner,
        num_logprobs=10,
    ),
384
    "glm4v": VLMTestInfo(
385
        models=["zai-org/glm-4v-9b"],
386
        test_type=VLMTestType.IMAGE,
387
        prompt_formatter=lambda img_prompt: f"<|user|>\n{img_prompt}<|assistant|>",
388
389
390
391
392
393
        single_image_prompts=IMAGE_ASSETS.prompts(
            {
                "stop_sign": "<|begin_of_image|><|endoftext|><|end_of_image|>What's the content in the center of the image?",  # noqa: E501
                "cherry_blossom": "<|begin_of_image|><|endoftext|><|end_of_image|>What is the season?",  # noqa: E501
            }
        ),
394
395
396
        max_model_len=2048,
        max_num_seqs=2,
        get_stop_token_ids=lambda tok: [151329, 151336, 151338],
397
398
399
400
401
402
        patch_hf_runner=model_utils.glm4v_patch_hf_runner,
        # The image embeddings match with HF but the outputs of the language
        # decoder are only consistent up to 2 decimal places.
        # So, we need to reduce the number of tokens for the test to pass.
        max_tokens=8,
        num_logprobs=10,
403
        marks=[large_gpu_mark(min_gb=32)],
404
    ),
405
    "glm4_1v": VLMTestInfo(
406
        models=["zai-org/GLM-4.1V-9B-Thinking"],
407
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
408
409
410
        prompt_formatter=lambda img_prompt: f"<|user|>\n{img_prompt}<|assistant|>",
        img_idx_to_prompt=lambda idx: "<|begin_of_image|><|image|><|end_of_image|>",
        video_idx_to_prompt=lambda idx: "<|begin_of_video|><|video|><|end_of_video|>",
411
412
413
414
415
416
        max_model_len=2048,
        max_num_seqs=2,
        get_stop_token_ids=lambda tok: [151329, 151336, 151338],
        num_logprobs=10,
        image_size_factors=[(), (0.25,), (0.25, 0.25, 0.25), (0.25, 0.2, 0.15)],
        auto_cls=AutoModelForImageTextToText,
417
        marks=[large_gpu_mark(min_gb=32)],
418
419
    ),
    "glm4_1v-video": VLMTestInfo(
420
        models=["zai-org/GLM-4.1V-9B-Thinking"],
421
422
423
424
425
426
        # GLM4.1V require include video metadata for input
        test_type=VLMTestType.CUSTOM_INPUTS,
        max_model_len=4096,
        max_num_seqs=2,
        auto_cls=AutoModelForImageTextToText,
        patch_hf_runner=model_utils.glm4_1v_patch_hf_runner,
427
428
429
430
431
432
        custom_test_opts=[
            CustomTestOptions(
                inputs=custom_inputs.video_with_metadata_glm4_1v(),
                limit_mm_per_prompt={"video": 1},
            )
        ],
433
        marks=[large_gpu_mark(min_gb=32)],
434
    ),
435
    "h2ovl": VLMTestInfo(
436
        models=[
437
            "h2oai/h2ovl-mississippi-800m",
438
            "h2oai/h2ovl-mississippi-2b",
439
440
        ],
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
441
        prompt_formatter=lambda img_prompt: f"<|prompt|>{img_prompt}<|end|><|answer|>",
442
443
        single_image_prompts=IMAGE_ASSETS.prompts(
            {
444
                "stop_sign": "<image>\nWhat's the content in the center of the image?",
445
446
447
                "cherry_blossom": "<image>\nWhat is the season?",
            }
        ),
448
449
450
        multi_image_prompt="Image-1: <image>\nImage-2: <image>\nDescribe the two images in short.",  # noqa: E501
        max_model_len=8192,
        use_tokenizer_eos=True,
451
        num_logprobs=10,
452
        patch_hf_runner=model_utils.h2ovl_patch_hf_runner,
453
    ),
454
    "idefics3": VLMTestInfo(
455
        models=["HuggingFaceTB/SmolVLM-256M-Instruct"],
456
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
457
        prompt_formatter=lambda img_prompt: f"<|begin_of_text|>User:{img_prompt}<end_of_utterance>\nAssistant:",  # noqa: E501
458
459
460
        img_idx_to_prompt=lambda idx: "<image>",
        max_model_len=8192,
        max_num_seqs=2,
461
        auto_cls=AutoModelForImageTextToText,
462
        hf_output_post_proc=model_utils.idefics3_trunc_hf_output,
463
    ),
464
465
466
467
    "intern_vl": VLMTestInfo(
        models=[
            "OpenGVLab/InternVL2-1B",
            "OpenGVLab/InternVL2-2B",
468
469
            # FIXME: Config cannot be loaded in transformers 4.52
            # "OpenGVLab/Mono-InternVL-2B",
470
471
        ],
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
472
473
474
        prompt_formatter=lambda img_prompt: f"<|im_start|>User\n{img_prompt}<|im_end|>\n<|im_start|>Assistant\n",  # noqa: E501
        single_image_prompts=IMAGE_ASSETS.prompts(
            {
475
                "stop_sign": "<image>\nWhat's the content in the center of the image?",
476
477
478
                "cherry_blossom": "<image>\nWhat is the season?",
            }
        ),
479
480
481
482
483
        multi_image_prompt="Image-1: <image>\nImage-2: <image>\nDescribe the two images in short.",  # noqa: E501
        max_model_len=4096,
        use_tokenizer_eos=True,
        patch_hf_runner=model_utils.internvl_patch_hf_runner,
    ),
484
485
486
487
488
    "intern_vl-video": VLMTestInfo(
        models=[
            "OpenGVLab/InternVL3-1B",
        ],
        test_type=VLMTestType.VIDEO,
489
        prompt_formatter=lambda img_prompt: f"<|im_start|>User\n{img_prompt}<|im_end|>\n<|im_start|>Assistant\n",  # noqa: E501
490
491
492
493
494
        video_idx_to_prompt=lambda idx: "<video>",
        max_model_len=8192,
        use_tokenizer_eos=True,
        patch_hf_runner=model_utils.internvl_patch_hf_runner,
    ),
495
496
497
498
499
500
501
    "intern_vl-hf": VLMTestInfo(
        models=["OpenGVLab/InternVL3-1B-hf"],
        test_type=(
            VLMTestType.IMAGE,
            VLMTestType.MULTI_IMAGE,
            VLMTestType.VIDEO,
        ),
502
        prompt_formatter=lambda img_prompt: f"<|im_start|>User\n{img_prompt}<|im_end|>\n<|im_start|>Assistant\n",  # noqa: E501
503
504
505
506
507
508
        img_idx_to_prompt=lambda idx: "<IMG_CONTEXT>",
        video_idx_to_prompt=lambda idx: "<video>",
        max_model_len=8192,
        use_tokenizer_eos=True,
        auto_cls=AutoModelForImageTextToText,
    ),
509
510
511
    "kimi_vl": VLMTestInfo(
        models=["moonshotai/Kimi-VL-A3B-Instruct"],
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
512
        prompt_formatter=lambda img_prompt: f"<|im_user|>user<|im_middle|>{img_prompt}<|im_end|><|im_assistant|>assistant<|im_middle|>",  # noqa: E501
513
514
515
516
517
518
519
520
        img_idx_to_prompt=lambda _: "<|media_start|>image<|media_content|><|media_pad|><|media_end|>",  # noqa: E501
        max_model_len=8192,
        max_num_seqs=2,
        dtype="bfloat16",
        tensor_parallel_size=1,
        vllm_output_post_proc=model_utils.kimiv_vl_vllm_to_hf_output,
        marks=[large_gpu_mark(min_gb=48)],
    ),
521
522
    "llama4": VLMTestInfo(
        models=["meta-llama/Llama-4-Scout-17B-16E-Instruct"],
523
        prompt_formatter=lambda img_prompt: f"<|begin_of_text|><|header_start|>user<|header_end|>\n\n{img_prompt}<|eot|><|header_start|>assistant<|header_end|>\n\n",  # noqa: E501
524
525
526
        img_idx_to_prompt=lambda _: "<|image|>",
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
        distributed_executor_backend="mp",
527
        image_size_factors=[(0.25, 0.5, 1.0)],
528
529
530
531
532
        hf_model_kwargs={"device_map": "auto"},
        max_model_len=8192,
        max_num_seqs=4,
        dtype="bfloat16",
        auto_cls=AutoModelForImageTextToText,
533
534
        tensor_parallel_size=4,
        marks=multi_gpu_marks(num_gpus=4),
535
    ),
536
537
538
539
540
    "llava_next": VLMTestInfo(
        models=["llava-hf/llava-v1.6-mistral-7b-hf"],
        test_type=(VLMTestType.IMAGE, VLMTestType.CUSTOM_INPUTS),
        prompt_formatter=lambda img_prompt: f"[INST] {img_prompt} [/INST]",
        max_model_len=10240,
541
        auto_cls=AutoModelForImageTextToText,
542
        vllm_output_post_proc=model_utils.llava_image_vllm_to_hf_output,
543
544
545
546
547
548
549
550
        custom_test_opts=[
            CustomTestOptions(
                inputs=custom_inputs.multi_image_multi_aspect_ratio_inputs(
                    formatter=lambda img_prompt: f"[INST] {img_prompt} [/INST]"
                ),
                limit_mm_per_prompt={"image": 4},
            )
        ],
551
    ),
552
    "llava_onevision": VLMTestInfo(
553
554
        models=["llava-hf/llava-onevision-qwen2-0.5b-ov-hf"],
        test_type=VLMTestType.CUSTOM_INPUTS,
555
        prompt_formatter=lambda vid_prompt: f"<|im_start|>user\n{vid_prompt}<|im_end|>\n<|im_start|>assistant\n",  # noqa: E501
556
557
        num_video_frames=16,
        max_model_len=16384,
558
559
        hf_model_kwargs=model_utils.llava_onevision_hf_model_kwargs(
            "llava-hf/llava-onevision-qwen2-0.5b-ov-hf"
560
        ),
561
        auto_cls=AutoModelForImageTextToText,
562
        vllm_output_post_proc=model_utils.llava_onevision_vllm_to_hf_output,
563
564
565
566
567
568
569
570
        custom_test_opts=[
            CustomTestOptions(
                inputs=custom_inputs.multi_video_multi_aspect_ratio_inputs(
                    formatter=lambda vid_prompt: f"<|im_start|>user\n{vid_prompt}<|im_end|>\n<|im_start|>assistant\n",  # noqa: E501
                ),
                limit_mm_per_prompt={"video": 4},
            )
        ],
571
572
573
574
575
576
577
    ),
    "llava_next_video": VLMTestInfo(
        models=["llava-hf/LLaVA-NeXT-Video-7B-hf"],
        test_type=VLMTestType.VIDEO,
        prompt_formatter=lambda vid_prompt: f"USER: {vid_prompt} ASSISTANT:",
        num_video_frames=16,
        max_model_len=4096,
578
        max_num_seqs=2,
579
        auto_cls=AutoModelForImageTextToText,
580
581
        vllm_output_post_proc=model_utils.llava_video_vllm_to_hf_output,
    ),
582
583
584
585
586
587
    "mantis": VLMTestInfo(
        models=["TIGER-Lab/Mantis-8B-siglip-llama3"],
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
        prompt_formatter=lambda img_prompt: f"<|start_header_id|>user<|end_header_id|>\n\n{img_prompt}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n",  # noqa: E501
        max_model_len=4096,
        get_stop_token_ids=lambda tok: [128009],
588
        auto_cls=AutoModelForImageTextToText,
589
590
591
        vllm_output_post_proc=model_utils.mantis_vllm_to_hf_output,
        patch_hf_runner=model_utils.mantis_patch_hf_runner,
    ),
592
    "minicpmv_25": VLMTestInfo(
593
        models=["openbmb/MiniCPM-Llama3-V-2_5"],
594
        test_type=VLMTestType.IMAGE,
595
596
597
598
599
        prompt_formatter=lambda img_prompt: f"<|begin_of_text|><|start_header_id|>user<|end_header_id|>\n\n{img_prompt}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n",  # noqa: E501
        img_idx_to_prompt=lambda idx: "(<image>./</image>)\n",
        max_model_len=4096,
        max_num_seqs=2,
        get_stop_token_ids=lambda tok: [tok.eos_id, tok.eot_id],
600
        hf_output_post_proc=model_utils.minicpmv_trunc_hf_output,
601
        patch_hf_runner=model_utils.minicpmv_25_patch_hf_runner,
602
603
        # FIXME: https://huggingface.co/openbmb/MiniCPM-V-2_6/discussions/55
        marks=[pytest.mark.skip("HF import fails")],
604
    ),
605
606
    "minicpmo_26": VLMTestInfo(
        models=["openbmb/MiniCPM-o-2_6"],
607
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
608
609
610
611
        prompt_formatter=lambda img_prompt: f"<|begin_of_text|><|start_header_id|>user<|end_header_id|>\n\n{img_prompt}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n",  # noqa: E501
        img_idx_to_prompt=lambda idx: "(<image>./</image>)\n",
        max_model_len=4096,
        max_num_seqs=2,
612
613
        get_stop_token_ids=lambda tok: tok.convert_tokens_to_ids(
            ["<|im_end|>", "<|endoftext|>"]
614
        ),
615
616
        hf_output_post_proc=model_utils.minicpmv_trunc_hf_output,
        patch_hf_runner=model_utils.minicpmo_26_patch_hf_runner,
617
        # FIXME: https://huggingface.co/openbmb/MiniCPM-o-2_6/discussions/49
618
        marks=[pytest.mark.skip("HF import fails")],
619
    ),
620
621
    "minicpmv_26": VLMTestInfo(
        models=["openbmb/MiniCPM-V-2_6"],
622
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
623
624
625
626
        prompt_formatter=lambda img_prompt: f"<|begin_of_text|><|start_header_id|>user<|end_header_id|>\n\n{img_prompt}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n",  # noqa: E501
        img_idx_to_prompt=lambda idx: "(<image>./</image>)\n",
        max_model_len=4096,
        max_num_seqs=2,
627
628
        get_stop_token_ids=lambda tok: tok.convert_tokens_to_ids(
            ["<|im_end|>", "<|endoftext|>"]
629
        ),
630
631
632
        hf_output_post_proc=model_utils.minicpmv_trunc_hf_output,
        patch_hf_runner=model_utils.minicpmv_26_patch_hf_runner,
    ),
633
634
    "minimax_vl_01": VLMTestInfo(
        models=["MiniMaxAI/MiniMax-VL-01"],
635
        prompt_formatter=lambda img_prompt: f"<beginning_of_sentence>user: {img_prompt} assistant:<end_of_sentence>",  # noqa: E501
636
637
638
639
640
641
642
643
644
645
        img_idx_to_prompt=lambda _: "<image>",
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
        max_model_len=8192,
        max_num_seqs=4,
        dtype="bfloat16",
        hf_output_post_proc=model_utils.minimax_vl_01_hf_output,
        patch_hf_runner=model_utils.minimax_vl_01_patch_hf_runner,
        auto_cls=AutoModelForImageTextToText,
        marks=[large_gpu_mark(min_gb=80)],
    ),
646
647
    "molmo": VLMTestInfo(
        models=["allenai/Molmo-7B-D-0924"],
648
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
649
        prompt_formatter=identity,
650
651
        max_model_len=4096,
        max_num_seqs=2,
652
        patch_hf_runner=model_utils.molmo_patch_hf_runner,
653
    ),
654
655
656
    "ovis1_6-gemma2": VLMTestInfo(
        models=["AIDC-AI/Ovis1.6-Gemma2-9B"],
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
657
        prompt_formatter=lambda img_prompt: f"<bos><start_of_turn>user\n{img_prompt}<end_of_turn>\n<start_of_turn>model\n",  # noqa: E501
658
        img_idx_to_prompt=lambda idx: "<image>\n",
659
660
661
662
663
664
665
666
        max_model_len=4096,
        max_num_seqs=2,
        dtype="half",
        # use sdpa mode for hf runner since ovis2 didn't work with flash_attn
        hf_model_kwargs={"llm_attn_implementation": "sdpa"},
        patch_hf_runner=model_utils.ovis_patch_hf_runner,
        marks=[large_gpu_mark(min_gb=32)],
    ),
667
668
669
    "ovis2": VLMTestInfo(
        models=["AIDC-AI/Ovis2-1B"],
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
670
        prompt_formatter=lambda img_prompt: f"<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n<|im_start|>user\n{img_prompt}<|im_end|>\n<|im_start|>assistant\n",  # noqa: E501
671
        img_idx_to_prompt=lambda idx: "<image>\n",
672
673
674
675
676
        max_model_len=4096,
        max_num_seqs=2,
        dtype="half",
        # use sdpa mode for hf runner since ovis2 didn't work with flash_attn
        hf_model_kwargs={"llm_attn_implementation": "sdpa"},
677
        patch_hf_runner=model_utils.ovis_patch_hf_runner,
678
    ),
679
680
    "ovis2_5": VLMTestInfo(
        models=["AIDC-AI/Ovis2.5-2B"],
681
682
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE, VLMTestType.VIDEO),
        prompt_formatter=lambda img_prompt: f"<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n<|im_start|>user\n{img_prompt}<|im_end|>\n<|im_start|>assistant\n",  # noqa: E501
683
        img_idx_to_prompt=lambda idx: "<image>\n",
684
685
686
687
688
689
        video_idx_to_prompt=lambda idx: "<video>\n",
        max_model_len=4096,
        max_num_seqs=2,
        dtype="half",
        num_logprobs=10,
        patch_hf_runner=model_utils.ovis2_5_patch_hf_runner,
690
        hf_model_kwargs={"revision": "refs/pr/5"},
691
    ),
692
693
694
    "phi3v": VLMTestInfo(
        models=["microsoft/Phi-3.5-vision-instruct"],
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
695
        prompt_formatter=lambda img_prompt: f"<|user|>\n{img_prompt}<|end|>\n<|assistant|>\n",  # noqa: E501
696
697
698
        img_idx_to_prompt=lambda idx: f"<|image_{idx}|>\n",
        max_model_len=4096,
        max_num_seqs=2,
699
        runner="generate",
700
701
        # use sdpa mode for hf runner since phi3v didn't work with flash_attn
        hf_model_kwargs={"_attn_implementation": "sdpa"},
702
703
704
705
        use_tokenizer_eos=True,
        vllm_output_post_proc=model_utils.phi3v_vllm_to_hf_output,
        num_logprobs=10,
    ),
706
707
708
709
710
711
712
    "pixtral_hf": VLMTestInfo(
        models=["nm-testing/pixtral-12b-FP8-dynamic"],
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
        prompt_formatter=lambda img_prompt: f"<s>[INST]{img_prompt}[/INST]",
        img_idx_to_prompt=lambda idx: "[IMG]",
        max_model_len=8192,
        max_num_seqs=2,
713
        auto_cls=AutoModelForImageTextToText,
714
        marks=[large_gpu_mark(min_gb=48)],
715
    ),
716
    "qwen_vl": VLMTestInfo(
717
718
719
720
721
722
723
724
725
        models=["Qwen/Qwen-VL"],
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
        prompt_formatter=identity,
        img_idx_to_prompt=lambda idx: f"Picture {idx}: <img></img>\n",
        max_model_len=1024,
        max_num_seqs=2,
        vllm_output_post_proc=model_utils.qwen_vllm_to_hf_output,
        prompt_path_encoder=model_utils.qwen_prompt_path_encoder,
    ),
726
727
    "qwen2_vl": VLMTestInfo(
        models=["Qwen/Qwen2-VL-2B-Instruct"],
728
729
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE, VLMTestType.VIDEO),
        prompt_formatter=lambda img_prompt: f"<|im_start|>User\n{img_prompt}<|im_end|>\n<|im_start|>assistant\n",  # noqa: E501
730
731
        img_idx_to_prompt=lambda idx: "<|vision_start|><|image_pad|><|vision_end|>",
        video_idx_to_prompt=lambda idx: "<|vision_start|><|video_pad|><|vision_end|>",
732
        multi_image_prompt="Picture 1: <vlm_image>\nPicture 2: <vlm_image>\nDescribe these two images with one paragraph respectively.",  # noqa: E501
733
734
        max_model_len=4096,
        max_num_seqs=2,
735
        auto_cls=AutoModelForImageTextToText,
736
737
738
739
        vllm_output_post_proc=model_utils.qwen2_vllm_to_hf_output,
        image_size_factors=[(), (0.25,), (0.25, 0.25, 0.25), (0.25, 0.2, 0.15)],
        marks=[pytest.mark.cpu_model],
    ),
740
741
742
    "skywork_r1v": VLMTestInfo(
        models=["Skywork/Skywork-R1V-38B"],
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
743
744
745
        prompt_formatter=lambda img_prompt: f"<|begin▁of▁sentence|><|User|>\n{img_prompt}<|Assistant|><think>\n",  # noqa: E501
        single_image_prompts=IMAGE_ASSETS.prompts(
            {
746
                "stop_sign": "<image>\nWhat's the content in the center of the image?",
747
748
749
                "cherry_blossom": "<image>\nWhat is the season?",
            }
        ),
750
        multi_image_prompt="<image>\n<image>\nDescribe the two images in short.",
751
752
753
754
755
        max_model_len=4096,
        use_tokenizer_eos=True,
        patch_hf_runner=model_utils.skyworkr1v_patch_hf_runner,
        marks=[large_gpu_mark(min_gb=80)],
    ),
756
757
758
    "smolvlm": VLMTestInfo(
        models=["HuggingFaceTB/SmolVLM2-2.2B-Instruct"],
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
759
        prompt_formatter=lambda img_prompt: f"<|im_start|>User:{img_prompt}<end_of_utterance>\nAssistant:",  # noqa: E501
760
761
762
763
764
        img_idx_to_prompt=lambda idx: "<image>",
        max_model_len=8192,
        max_num_seqs=2,
        auto_cls=AutoModelForImageTextToText,
        hf_output_post_proc=model_utils.smolvlm_trunc_hf_output,
765
        num_logprobs=10,
766
    ),
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
    "tarsier": VLMTestInfo(
        models=["omni-research/Tarsier-7b"],
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
        prompt_formatter=lambda img_prompt: f"USER: {img_prompt} ASSISTANT:",
        max_model_len=4096,
        max_num_seqs=2,
        auto_cls=AutoModelForImageTextToText,
        patch_hf_runner=model_utils.tarsier_patch_hf_runner,
    ),
    "tarsier2": VLMTestInfo(
        models=["omni-research/Tarsier2-Recap-7b"],
        test_type=(
            VLMTestType.IMAGE,
            VLMTestType.MULTI_IMAGE,
            VLMTestType.VIDEO,
        ),
783
        prompt_formatter=lambda img_prompt: f"<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n<|im_start|>user\n{img_prompt}<|im_end|>\n<|im_start|>assistant\n",  # noqa: E501
784
785
        img_idx_to_prompt=lambda idx: "<|vision_start|><|image_pad|><|vision_end|>",
        video_idx_to_prompt=lambda idx: "<|vision_start|><|video_pad|><|vision_end|>",
786
787
788
789
790
791
        max_model_len=4096,
        max_num_seqs=2,
        auto_cls=AutoModelForImageTextToText,
        image_size_factors=[(), (0.25,), (0.25, 0.25, 0.25), (0.25, 0.2, 0.15)],
        marks=[pytest.mark.skip("Model initialization hangs")],
    ),
792
    ### Tensor parallel / multi-gpu broadcast tests
793
    "chameleon-broadcast": VLMTestInfo(
794
795
796
        models=["facebook/chameleon-7b"],
        prompt_formatter=lambda img_prompt: f"USER: {img_prompt}\nASSISTANT:",
        max_model_len=4096,
797
        auto_cls=AutoModelForImageTextToText,
798
799
        vllm_output_post_proc=lambda vllm_output, model: vllm_output[:2],
        hf_output_post_proc=lambda hf_output, model: hf_output[:2],
800
        comparator=check_outputs_equal,
801
        marks=multi_gpu_marks(num_gpus=2),
802
        **COMMON_BROADCAST_SETTINGS,  # type: ignore
803
    ),
804
    "llava-broadcast": VLMTestInfo(
805
806
807
        models=["llava-hf/llava-1.5-7b-hf"],
        prompt_formatter=lambda img_prompt: f"USER: {img_prompt}\nASSISTANT:",
        max_model_len=4096,
808
        auto_cls=AutoModelForImageTextToText,
809
        vllm_output_post_proc=model_utils.llava_image_vllm_to_hf_output,
810
        marks=multi_gpu_marks(num_gpus=2),
811
        **COMMON_BROADCAST_SETTINGS,  # type: ignore
812
    ),
813
    "llava_next-broadcast": VLMTestInfo(
814
815
816
        models=["llava-hf/llava-v1.6-mistral-7b-hf"],
        prompt_formatter=lambda img_prompt: f"[INST] {img_prompt} [/INST]",
        max_model_len=10240,
817
        auto_cls=AutoModelForImageTextToText,
818
        vllm_output_post_proc=model_utils.llava_image_vllm_to_hf_output,
819
        marks=multi_gpu_marks(num_gpus=2),
820
        **COMMON_BROADCAST_SETTINGS,  # type: ignore
821
822
823
824
    ),
    ### Custom input edge-cases for specific models
    "intern_vl-diff-patches": VLMTestInfo(
        models=["OpenGVLab/InternVL2-2B"],
825
        prompt_formatter=lambda img_prompt: f"<|im_start|>User\n{img_prompt}<|im_end|>\n<|im_start|>Assistant\n",  # noqa: E501
826
827
828
829
830
831
832
833
        test_type=VLMTestType.CUSTOM_INPUTS,
        max_model_len=4096,
        use_tokenizer_eos=True,
        patch_hf_runner=model_utils.internvl_patch_hf_runner,
        custom_test_opts=[
            CustomTestOptions(
                inputs=inp,
                limit_mm_per_prompt={"image": 2},
834
835
            )
            for inp in custom_inputs.different_patch_input_cases_internvl()
836
837
        ],
    ),
838
    "llava_onevision-multiple-images": VLMTestInfo(
839
840
841
842
        models=["llava-hf/llava-onevision-qwen2-0.5b-ov-hf"],
        test_type=VLMTestType.CUSTOM_INPUTS,
        max_model_len=16384,
        max_num_seqs=2,
843
        auto_cls=AutoModelForImageTextToText,
844
845
        hf_model_kwargs=model_utils.llava_onevision_hf_model_kwargs(
            "llava-hf/llava-onevision-qwen2-0.5b-ov-hf"
846
        ),
847
        vllm_output_post_proc=model_utils.llava_onevision_vllm_to_hf_output,
848
849
850
851
852
853
854
855
        custom_test_opts=[
            CustomTestOptions(
                inputs=custom_inputs.multi_image_multi_aspect_ratio_inputs(
                    formatter=lambda vid_prompt: f"<|im_start|>user\n{vid_prompt}<|im_end|>\n<|im_start|>assistant\n",  # noqa: E501
                ),
                limit_mm_per_prompt={"image": 4},
            )
        ],
856
857
858
859
860
861
        marks=[
            pytest.mark.skipif(
                Version(TRANSFORMERS_VERSION) == Version("4.57.1"),
                reason="This model is broken in Transformers v4.57.1",
            )
        ],
862
    ),
863
864
865
866
867
868
    # regression test for https://github.com/vllm-project/vllm/issues/15122
    "qwen2_5_vl-windows-attention": VLMTestInfo(
        models=["Qwen/Qwen2.5-VL-3B-Instruct"],
        test_type=VLMTestType.CUSTOM_INPUTS,
        max_model_len=4096,
        max_num_seqs=2,
869
        auto_cls=AutoModelForImageTextToText,
870
        vllm_output_post_proc=model_utils.qwen2_vllm_to_hf_output,
871
872
873
874
875
876
        custom_test_opts=[
            CustomTestOptions(
                inputs=custom_inputs.windows_attention_image_qwen2_5_vl(),
                limit_mm_per_prompt={"image": 1},
            )
        ],
877
    ),
878
879
880
}


881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
def _mark_splits(
    test_settings: dict[str, VLMTestInfo],
    *,
    num_groups: int,
) -> dict[str, VLMTestInfo]:
    name_by_test_info_id = {id(v): k for k, v in test_settings.items()}
    test_infos_by_model = defaultdict[str, list[VLMTestInfo]](list)

    for info in test_settings.values():
        for model in info.models:
            test_infos_by_model[model].append(info)

    models = sorted(test_infos_by_model.keys())
    split_size = math.ceil(len(models) / num_groups)

    new_test_settings = dict[str, VLMTestInfo]()

    for i in range(num_groups):
899
        models_in_group = models[i * split_size : (i + 1) * split_size]
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915

        for model in models_in_group:
            for info in test_infos_by_model[model]:
                new_marks = (info.marks or []) + [pytest.mark.split(group=i)]
                new_info = info._replace(marks=new_marks)
                new_test_settings[name_by_test_info_id[id(info)]] = new_info

    missing_keys = test_settings.keys() - new_test_settings.keys()
    assert not missing_keys, f"Missing keys: {missing_keys}"

    return new_test_settings


VLM_TEST_SETTINGS = _mark_splits(VLM_TEST_SETTINGS, num_groups=2)


916
917
918
919
920
921
### Test wrappers
# Wrappers around the core test running func for:
# - single image
# - multi-image
# - image embeddings
# - video
922
# - audio
923
# - custom inputs
924
925
926
927
928
@pytest.mark.parametrize(
    "model_type,test_case",
    get_parametrized_options(
        VLM_TEST_SETTINGS,
        test_type=VLMTestType.IMAGE,
929
        create_new_process_for_each_test=False,
930
931
    ),
)
932
933
934
935
936
937
938
939
def test_single_image_models(
    tmp_path: PosixPath,
    model_type: str,
    test_case: ExpandableVLMTestArgs,
    hf_runner: type[HfRunner],
    vllm_runner: type[VllmRunner],
    image_assets: ImageTestAssets,
):
940
941
942
943
944
945
946
947
948
949
950
    model_test_info = VLM_TEST_SETTINGS[model_type]
    runners.run_single_image_test(
        tmp_path=tmp_path,
        model_test_info=model_test_info,
        test_case=test_case,
        hf_runner=hf_runner,
        vllm_runner=vllm_runner,
        image_assets=image_assets,
    )


951
952
953
954
955
@pytest.mark.parametrize(
    "model_type,test_case",
    get_parametrized_options(
        VLM_TEST_SETTINGS,
        test_type=VLMTestType.MULTI_IMAGE,
956
        create_new_process_for_each_test=False,
957
958
    ),
)
959
960
961
962
963
964
965
966
def test_multi_image_models(
    tmp_path: PosixPath,
    model_type: str,
    test_case: ExpandableVLMTestArgs,
    hf_runner: type[HfRunner],
    vllm_runner: type[VllmRunner],
    image_assets: ImageTestAssets,
):
967
968
969
970
971
972
973
974
975
976
977
    model_test_info = VLM_TEST_SETTINGS[model_type]
    runners.run_multi_image_test(
        tmp_path=tmp_path,
        model_test_info=model_test_info,
        test_case=test_case,
        hf_runner=hf_runner,
        vllm_runner=vllm_runner,
        image_assets=image_assets,
    )


978
979
980
981
982
@pytest.mark.parametrize(
    "model_type,test_case",
    get_parametrized_options(
        VLM_TEST_SETTINGS,
        test_type=VLMTestType.EMBEDDING,
983
        create_new_process_for_each_test=False,
984
985
    ),
)
986
987
988
989
990
991
992
def test_image_embedding_models(
    model_type: str,
    test_case: ExpandableVLMTestArgs,
    hf_runner: type[HfRunner],
    vllm_runner: type[VllmRunner],
    image_assets: ImageTestAssets,
):
993
994
995
996
997
998
999
1000
1001
1002
    model_test_info = VLM_TEST_SETTINGS[model_type]
    runners.run_embedding_test(
        model_test_info=model_test_info,
        test_case=test_case,
        hf_runner=hf_runner,
        vllm_runner=vllm_runner,
        image_assets=image_assets,
    )


1003
1004
1005
1006
1007
@pytest.mark.parametrize(
    "model_type,test_case",
    get_parametrized_options(
        VLM_TEST_SETTINGS,
        test_type=VLMTestType.VIDEO,
1008
        create_new_process_for_each_test=False,
1009
1010
    ),
)
1011
1012
1013
1014
1015
1016
1017
def test_video_models(
    model_type: str,
    test_case: ExpandableVLMTestArgs,
    hf_runner: type[HfRunner],
    vllm_runner: type[VllmRunner],
    video_assets: VideoTestAssets,
):
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
    model_test_info = VLM_TEST_SETTINGS[model_type]
    runners.run_video_test(
        model_test_info=model_test_info,
        test_case=test_case,
        hf_runner=hf_runner,
        vllm_runner=vllm_runner,
        video_assets=video_assets,
    )


1028
1029
1030
1031
1032
1033
@pytest.mark.parametrize(
    "model_type,test_case",
    get_parametrized_options(
        VLM_TEST_SETTINGS,
        test_type=VLMTestType.AUDIO,
        create_new_process_for_each_test=False,
1034
1035
    ),
)
1036
1037
1038
1039
1040
1041
1042
def test_audio_models(
    model_type: str,
    test_case: ExpandableVLMTestArgs,
    hf_runner: type[HfRunner],
    vllm_runner: type[VllmRunner],
    audio_assets: AudioTestAssets,
):
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
    model_test_info = VLM_TEST_SETTINGS[model_type]
    runners.run_audio_test(
        model_test_info=model_test_info,
        test_case=test_case,
        hf_runner=hf_runner,
        vllm_runner=vllm_runner,
        audio_assets=audio_assets,
    )


1053
1054
1055
1056
1057
@pytest.mark.parametrize(
    "model_type,test_case",
    get_parametrized_options(
        VLM_TEST_SETTINGS,
        test_type=VLMTestType.CUSTOM_INPUTS,
1058
        create_new_process_for_each_test=False,
1059
1060
    ),
)
1061
1062
1063
def test_custom_inputs_models(
    model_type: str,
    test_case: ExpandableVLMTestArgs,
1064
1065
    hf_runner: type[HfRunner],
    vllm_runner: type[VllmRunner],
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
):
    model_test_info = VLM_TEST_SETTINGS[model_type]
    runners.run_custom_inputs_test(
        model_test_info=model_test_info,
        test_case=test_case,
        hf_runner=hf_runner,
        vllm_runner=vllm_runner,
    )


#### Tests filtering for things running each test as a new process
1077
1078
1079
1080
1081
@pytest.mark.parametrize(
    "model_type,test_case",
    get_parametrized_options(
        VLM_TEST_SETTINGS,
        test_type=VLMTestType.IMAGE,
1082
        create_new_process_for_each_test=True,
1083
1084
    ),
)
1085
@create_new_process_for_each_test()
1086
1087
1088
1089
1090
1091
1092
1093
def test_single_image_models_heavy(
    tmp_path: PosixPath,
    model_type: str,
    test_case: ExpandableVLMTestArgs,
    hf_runner: type[HfRunner],
    vllm_runner: type[VllmRunner],
    image_assets: ImageTestAssets,
):
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
    model_test_info = VLM_TEST_SETTINGS[model_type]
    runners.run_single_image_test(
        tmp_path=tmp_path,
        model_test_info=model_test_info,
        test_case=test_case,
        hf_runner=hf_runner,
        vllm_runner=vllm_runner,
        image_assets=image_assets,
    )


1105
1106
1107
1108
1109
@pytest.mark.parametrize(
    "model_type,test_case",
    get_parametrized_options(
        VLM_TEST_SETTINGS,
        test_type=VLMTestType.MULTI_IMAGE,
1110
        create_new_process_for_each_test=True,
1111
1112
    ),
)
1113
@create_new_process_for_each_test()
1114
1115
1116
1117
1118
1119
1120
1121
def test_multi_image_models_heavy(
    tmp_path: PosixPath,
    model_type: str,
    test_case: ExpandableVLMTestArgs,
    hf_runner: type[HfRunner],
    vllm_runner: type[VllmRunner],
    image_assets: ImageTestAssets,
):
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
    model_test_info = VLM_TEST_SETTINGS[model_type]
    runners.run_multi_image_test(
        tmp_path=tmp_path,
        model_test_info=model_test_info,
        test_case=test_case,
        hf_runner=hf_runner,
        vllm_runner=vllm_runner,
        image_assets=image_assets,
    )


1133
1134
1135
1136
1137
@pytest.mark.parametrize(
    "model_type,test_case",
    get_parametrized_options(
        VLM_TEST_SETTINGS,
        test_type=VLMTestType.EMBEDDING,
1138
        create_new_process_for_each_test=True,
1139
1140
    ),
)
1141
@create_new_process_for_each_test()
1142
1143
1144
1145
1146
1147
1148
def test_image_embedding_models_heavy(
    model_type: str,
    test_case: ExpandableVLMTestArgs,
    hf_runner: type[HfRunner],
    vllm_runner: type[VllmRunner],
    image_assets: ImageTestAssets,
):
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
    model_test_info = VLM_TEST_SETTINGS[model_type]
    runners.run_embedding_test(
        model_test_info=model_test_info,
        test_case=test_case,
        hf_runner=hf_runner,
        vllm_runner=vllm_runner,
        image_assets=image_assets,
    )


1159
1160
1161
1162
1163
@pytest.mark.parametrize(
    "model_type,test_case",
    get_parametrized_options(
        VLM_TEST_SETTINGS,
        test_type=VLMTestType.VIDEO,
1164
        create_new_process_for_each_test=True,
1165
1166
    ),
)
1167
1168
1169
1170
1171
1172
1173
def test_video_models_heavy(
    model_type: str,
    test_case: ExpandableVLMTestArgs,
    hf_runner: type[HfRunner],
    vllm_runner: type[VllmRunner],
    video_assets: VideoTestAssets,
):
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
    model_test_info = VLM_TEST_SETTINGS[model_type]
    runners.run_video_test(
        model_test_info=model_test_info,
        test_case=test_case,
        hf_runner=hf_runner,
        vllm_runner=vllm_runner,
        video_assets=video_assets,
    )


1184
1185
1186
1187
1188
1189
@pytest.mark.parametrize(
    "model_type,test_case",
    get_parametrized_options(
        VLM_TEST_SETTINGS,
        test_type=VLMTestType.AUDIO,
        create_new_process_for_each_test=True,
1190
1191
    ),
)
1192
1193
1194
1195
1196
1197
1198
def test_audio_models_heavy(
    model_type: str,
    test_case: ExpandableVLMTestArgs,
    hf_runner: type[HfRunner],
    vllm_runner: type[VllmRunner],
    audio_assets: AudioTestAssets,
):
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
    model_test_info = VLM_TEST_SETTINGS[model_type]
    runners.run_audio_test(
        model_test_info=model_test_info,
        test_case=test_case,
        hf_runner=hf_runner,
        vllm_runner=vllm_runner,
        audio_assets=audio_assets,
    )


1209
1210
1211
1212
1213
@pytest.mark.parametrize(
    "model_type,test_case",
    get_parametrized_options(
        VLM_TEST_SETTINGS,
        test_type=VLMTestType.CUSTOM_INPUTS,
1214
        create_new_process_for_each_test=True,
1215
1216
    ),
)
1217
@create_new_process_for_each_test()
1218
1219
1220
def test_custom_inputs_models_heavy(
    model_type: str,
    test_case: ExpandableVLMTestArgs,
1221
1222
    hf_runner: type[HfRunner],
    vllm_runner: type[VllmRunner],
1223
1224
1225
1226
1227
1228
1229
1230
):
    model_test_info = VLM_TEST_SETTINGS[model_type]
    runners.run_custom_inputs_test(
        model_test_info=model_test_info,
        test_case=test_case,
        hf_runner=hf_runner,
        vllm_runner=vllm_runner,
    )