test_common.py 50.4 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
from transformers import (
    AutoModel,
15
    AutoModelForCausalLM,
16
17
18
    AutoModelForImageTextToText,
    AutoModelForTextToWaveform,
)
19
from transformers import __version__ as TRANSFORMERS_VERSION
20
21

from vllm.platforms import current_platform
22
from vllm.utils.func_utils import identity
23

24
25
26
27
28
29
30
31
32
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
33
34
35
from ...utils import check_outputs_equal
from .vlm_utils import custom_inputs, model_utils, runners
from .vlm_utils.case_filtering import get_parametrized_options
36
37
38
39
40
41
from .vlm_utils.types import (
    CustomTestOptions,
    ExpandableVLMTestArgs,
    VLMTestInfo,
    VLMTestType,
)
42
43
44
45
46
47

COMMON_BROADCAST_SETTINGS = {
    "test_type": VLMTestType.IMAGE,
    "dtype": "half",
    "max_tokens": 5,
    "tensor_parallel_size": 2,
48
    "hf_model_kwargs": {"device_map": "auto"},
49
    "image_size_factors": [(0.25, 0.5, 1.0)],
50
51
52
    "distributed_executor_backend": (
        "ray",
        "mp",
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
89
}

### 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 = {
90
91
92
    #### Core tests to always run in the CI
    "llava": VLMTestInfo(
        models=["llava-hf/llava-1.5-7b-hf"],
93
        test_type=(VLMTestType.EMBEDDING, VLMTestType.IMAGE, VLMTestType.CUSTOM_INPUTS),
94
95
96
        prompt_formatter=lambda img_prompt: f"USER: {img_prompt}\nASSISTANT:",
        convert_assets_to_embeddings=model_utils.get_llava_embeddings,
        max_model_len=4096,
97
        auto_cls=AutoModelForImageTextToText,
98
        vllm_output_post_proc=model_utils.llava_image_vllm_to_hf_output,
99
100
101
102
103
104
105
106
        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},
            )
        ],
107
        vllm_runner_kwargs={"enable_mm_embeds": True},
108
        marks=[pytest.mark.core_model, pytest.mark.cpu_model],
109
    ),
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
    "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
129
130
    "qwen2_5_vl": VLMTestInfo(
        models=["Qwen/Qwen2.5-VL-3B-Instruct"],
131
132
        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
133
134
        img_idx_to_prompt=lambda idx: "<|vision_start|><|image_pad|><|vision_end|>",
        video_idx_to_prompt=lambda idx: "<|vision_start|><|video_pad|><|vision_end|>",
135
        enforce_eager=False,
Roger Wang's avatar
Roger Wang committed
136
137
        max_model_len=4096,
        max_num_seqs=2,
138
        auto_cls=AutoModelForImageTextToText,
Roger Wang's avatar
Roger Wang committed
139
        vllm_output_post_proc=model_utils.qwen2_vllm_to_hf_output,
140
        image_size_factors=[(0.25,), (0.25, 0.25, 0.25), (0.25, 0.2, 0.15)],
141
        marks=[pytest.mark.core_model, pytest.mark.cpu_model],
Roger Wang's avatar
Roger Wang committed
142
    ),
143
    "qwen2_5_omni": VLMTestInfo(
144
        models=["Qwen/Qwen2.5-Omni-3B"],
145
146
        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
147
148
        img_idx_to_prompt=lambda idx: "<|vision_bos|><|IMAGE|><|vision_eos|>",
        video_idx_to_prompt=lambda idx: "<|vision_bos|><|VIDEO|><|vision_eos|>",
149
150
        max_model_len=4096,
        max_num_seqs=2,
151
        num_logprobs=6 if current_platform.is_cpu() else 5,
152
        auto_cls=AutoModelForTextToWaveform,
153
        vllm_output_post_proc=model_utils.qwen2_vllm_to_hf_output,
154
        patch_hf_runner=model_utils.qwen2_5_omni_patch_hf_runner,
155
        image_size_factors=[(0.25,), (0.25, 0.25, 0.25), (0.25, 0.2, 0.15)],
156
157
        marks=[pytest.mark.core_model, pytest.mark.cpu_model],
    ),
158
159
160
161
162
163
164
    "qwen3_vl": VLMTestInfo(
        models=["Qwen/Qwen3-VL-4B-Instruct"],
        test_type=(
            VLMTestType.IMAGE,
            VLMTestType.MULTI_IMAGE,
            VLMTestType.VIDEO,
        ),
165
        enforce_eager=False,
166
167
168
169
170
171
172
173
174
175
        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,
176
        image_size_factors=[(0.25,), (0.25, 0.25, 0.25), (0.25, 0.2, 0.15)],
177
178
179
180
        marks=[
            pytest.mark.core_model,
        ],
    ),
181
    "ultravox": VLMTestInfo(
182
        models=["fixie-ai/ultravox-v0_5-llama-3_2-1b"],
183
        test_type=VLMTestType.AUDIO,
184
        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
185
186
187
188
189
190
191
        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],
    ),
192
193
194
195
196
197
    #### 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,
198
        prompt_formatter=lambda vid_prompt: f"<|im_start|>user\n{vid_prompt}<|im_end|>\n<|im_start|>assistant\n",  # noqa: E501
199
        max_model_len=16384,
200
201
        hf_model_kwargs=model_utils.llava_onevision_hf_model_kwargs(
            "llava-hf/llava-onevision-qwen2-0.5b-ov-hf"
202
        ),
203
204
205
206
207
        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",
208
            "default_torch_num_threads": 1,
209
        },
210
211
212
        # FIXME: Investigate why the test hangs
        # when processing the 3rd prompt in vLLM
        marks=[pytest.mark.core_model, pytest.mark.skip(reason="Test hangs")],
213
    ),
214
215
216
    # Gemma3 has bidirectional mask on images
    "gemma3-transformers": VLMTestInfo(
        models=["google/gemma-3-4b-it"],
217
218
219
        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,
220
221
222
223
224
225
226
227
        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],
    ),
228
229
230
    "idefics3-transformers": VLMTestInfo(
        models=["HuggingFaceTB/SmolVLM-256M-Instruct"],
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
231
        prompt_formatter=lambda img_prompt: f"<|begin_of_text|>User:{img_prompt}<end_of_utterance>\nAssistant:",  # noqa: E501
232
233
234
235
236
237
238
239
240
241
242
        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],
    ),
243
244
245
246
    # 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,
247
        prompt_formatter=lambda img_prompt: f"<|im_start|>User\n{img_prompt}<|im_end|>\n<|im_start|>assistant\n",  # noqa: E501
248
        img_idx_to_prompt=lambda idx: "<|vision_start|><|image_pad|><|vision_end|>",
249
250
251
252
253
254
255
256
        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",
        },
257
        marks=[large_gpu_mark(min_gb=32)],
258
    ),
259
    #### Extended model tests
260
261
262
    "aria": VLMTestInfo(
        models=["rhymes-ai/Aria"],
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
263
        prompt_formatter=lambda img_prompt: f"<|im_start|>user\n{img_prompt}<|im_end|>\n<|im_start|>assistant\n ",  # noqa: E501
264
265
266
267
        img_idx_to_prompt=lambda idx: "<fim_prefix><|img|><fim_suffix>\n",
        max_model_len=4096,
        max_num_seqs=2,
        auto_cls=AutoModelForImageTextToText,
268
269
270
        single_image_prompts=IMAGE_ASSETS.prompts(
            {
                "stop_sign": "<vlm_image>Please describe the image shortly.",
271
                "cherry_blossom": "<vlm_image>Please infer the season with reason.",
272
273
            }
        ),
274
        multi_image_prompt="<vlm_image><vlm_image>Describe the two images shortly.",
275
276
277
278
279
        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
280
    "aya_vision": VLMTestInfo(
281
        models=["CohereLabs/aya-vision-8b"],
282
        test_type=(VLMTestType.IMAGE),
283
284
285
        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(
            {
286
287
                "stop_sign": "<image>What's the content in the center of the image?",
                "cherry_blossom": "<image>What is the season?",
288
289
            }
        ),
290
        multi_image_prompt="<image><image>Describe the two images in detail.",
291
292
293
294
295
296
        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(
297
        models=["CohereLabs/aya-vision-8b"],
298
        test_type=(VLMTestType.MULTI_IMAGE),
299
300
301
        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(
            {
302
303
                "stop_sign": "<image>What's the content in the center of the image?",
                "cherry_blossom": "<image>What is the season?",
304
305
            }
        ),
306
        multi_image_prompt="<image><image>Describe the two images in detail.",
307
        max_model_len=4096,
Jennifer Zhao's avatar
Jennifer Zhao committed
308
309
        max_num_seqs=2,
        auto_cls=AutoModelForImageTextToText,
310
311
        vllm_runner_kwargs={"mm_processor_kwargs": {"crop_to_patches": True}},
        marks=[large_gpu_mark(min_gb=32)],
Jennifer Zhao's avatar
Jennifer Zhao committed
312
    ),
313
    "blip2": VLMTestInfo(
314
        models=["Salesforce/blip2-opt-2.7b"],
315
316
317
        test_type=VLMTestType.IMAGE,
        prompt_formatter=lambda img_prompt: f"Question: {img_prompt} Answer:",
        img_idx_to_prompt=lambda idx: "",
318
        auto_cls=AutoModelForImageTextToText,
319
        vllm_output_post_proc=model_utils.blip2_vllm_to_hf_output,
320
321
        # FIXME: https://github.com/huggingface/transformers/pull/38510
        marks=[pytest.mark.skip("Model is broken")],
322
323
324
325
326
327
    ),
    "chameleon": VLMTestInfo(
        models=["facebook/chameleon-7b"],
        test_type=VLMTestType.IMAGE,
        prompt_formatter=lambda img_prompt: f"USER: {img_prompt}\nASSISTANT:",
        max_model_len=4096,
328
        max_num_seqs=2,
329
        auto_cls=AutoModelForImageTextToText,
330
        # For chameleon, we only compare the sequences
331
332
        vllm_output_post_proc=lambda vllm_output, model: vllm_output[:2],
        hf_output_post_proc=lambda hf_output, model: hf_output[:2],
333
334
335
336
        comparator=check_outputs_equal,
        max_tokens=8,
        dtype="bfloat16",
    ),
337
    "deepseek_vl_v2": VLMTestInfo(
338
        models=["Isotr0py/deepseek-vl2-tiny"],  # model repo using dynamic module
339
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
340
        prompt_formatter=lambda img_prompt: f"<|User|>: {img_prompt}\n\n<|Assistant|>: ",  # noqa: E501
341
342
        max_model_len=4096,
        max_num_seqs=2,
343
344
        single_image_prompts=IMAGE_ASSETS.prompts(
            {
345
                "stop_sign": "<image>\nWhat's the content in the center of the image?",
346
347
348
349
                "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
350
351
        patch_hf_runner=model_utils.deepseekvl2_patch_hf_runner,
        hf_output_post_proc=model_utils.deepseekvl2_trunc_hf_output,
352
        stop_str=["<|end▁of▁sentence|>", "<|begin▁of▁sentence|>"],
353
        image_size_factors=[(1.0,), (1.0, 1.0, 1.0), (0.1, 0.5, 1.0)],
354
    ),
355
356
357
358
359
360
361
    "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,
362
        auto_cls=AutoModelForImageTextToText,
363
364
365
366
        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)],
367
        marks=[large_gpu_mark(min_gb=32)],
368
    ),
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
    "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,
    ),
386
    "glm4v": VLMTestInfo(
387
        models=["zai-org/glm-4v-9b"],
388
        test_type=VLMTestType.IMAGE,
389
        prompt_formatter=lambda img_prompt: f"<|user|>\n{img_prompt}<|assistant|>",
390
391
392
393
394
395
        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
            }
        ),
396
397
398
        max_model_len=2048,
        max_num_seqs=2,
        get_stop_token_ids=lambda tok: [151329, 151336, 151338],
399
400
401
402
403
404
        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,
405
        auto_cls=AutoModelForCausalLM,
406
        marks=[large_gpu_mark(min_gb=32)],
407
    ),
408
    "glm4_1v": VLMTestInfo(
409
        models=["zai-org/GLM-4.1V-9B-Thinking"],
410
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
411
        prompt_formatter=lambda img_prompt: f"[gMASK]<|user|>\n{img_prompt}<|assistant|>\n",  # noqa: E501
412
413
        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|>",
414
415
416
417
418
419
        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,
420
        marks=[large_gpu_mark(min_gb=32)],
421
422
    ),
    "glm4_1v-video": VLMTestInfo(
423
        models=["zai-org/GLM-4.1V-9B-Thinking"],
424
425
        # GLM4.1V require include video metadata for input
        test_type=VLMTestType.CUSTOM_INPUTS,
426
        prompt_formatter=lambda vid_prompt: f"[gMASK]<|user|>\n{vid_prompt}<|assistant|>\n",  # noqa: E501
427
428
429
430
        max_model_len=4096,
        max_num_seqs=2,
        auto_cls=AutoModelForImageTextToText,
        patch_hf_runner=model_utils.glm4_1v_patch_hf_runner,
431
432
433
434
435
436
        custom_test_opts=[
            CustomTestOptions(
                inputs=custom_inputs.video_with_metadata_glm4_1v(),
                limit_mm_per_prompt={"video": 1},
            )
        ],
437
        marks=[large_gpu_mark(min_gb=32)],
438
    ),
439
    "h2ovl": VLMTestInfo(
440
        models=[
441
            "h2oai/h2ovl-mississippi-800m",
442
            "h2oai/h2ovl-mississippi-2b",
443
444
        ],
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
445
        prompt_formatter=lambda img_prompt: f"<|prompt|>{img_prompt}<|end|><|answer|>",
446
447
        single_image_prompts=IMAGE_ASSETS.prompts(
            {
448
                "stop_sign": "<image>\nWhat's the content in the center of the image?",
449
450
451
                "cherry_blossom": "<image>\nWhat is the season?",
            }
        ),
452
453
454
        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,
455
        num_logprobs=10,
456
        patch_hf_runner=model_utils.h2ovl_patch_hf_runner,
457
    ),
458
    "idefics3": VLMTestInfo(
459
        models=["HuggingFaceTB/SmolVLM-256M-Instruct"],
460
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
461
        prompt_formatter=lambda img_prompt: f"<|begin_of_text|>User:{img_prompt}<end_of_utterance>\nAssistant:",  # noqa: E501
462
463
464
        img_idx_to_prompt=lambda idx: "<image>",
        max_model_len=8192,
        max_num_seqs=2,
465
        auto_cls=AutoModelForImageTextToText,
466
        hf_output_post_proc=model_utils.idefics3_trunc_hf_output,
467
    ),
468
469
470
471
    "intern_vl": VLMTestInfo(
        models=[
            "OpenGVLab/InternVL2-1B",
            "OpenGVLab/InternVL2-2B",
472
473
            # FIXME: Config cannot be loaded in transformers 4.52
            # "OpenGVLab/Mono-InternVL-2B",
474
475
        ],
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
476
477
478
        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(
            {
479
                "stop_sign": "<image>\nWhat's the content in the center of the image?",
480
481
482
                "cherry_blossom": "<image>\nWhat is the season?",
            }
        ),
483
484
485
486
487
        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,
    ),
488
489
490
491
492
    "intern_vl-video": VLMTestInfo(
        models=[
            "OpenGVLab/InternVL3-1B",
        ],
        test_type=VLMTestType.VIDEO,
493
        prompt_formatter=lambda img_prompt: f"<|im_start|>User\n{img_prompt}<|im_end|>\n<|im_start|>Assistant\n",  # noqa: E501
494
495
496
497
498
        video_idx_to_prompt=lambda idx: "<video>",
        max_model_len=8192,
        use_tokenizer_eos=True,
        patch_hf_runner=model_utils.internvl_patch_hf_runner,
    ),
499
500
501
502
503
504
505
    "intern_vl-hf": VLMTestInfo(
        models=["OpenGVLab/InternVL3-1B-hf"],
        test_type=(
            VLMTestType.IMAGE,
            VLMTestType.MULTI_IMAGE,
            VLMTestType.VIDEO,
        ),
506
        prompt_formatter=lambda img_prompt: f"<|im_start|>User\n{img_prompt}<|im_end|>\n<|im_start|>Assistant\n",  # noqa: E501
507
508
509
510
511
512
        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,
    ),
513
514
515
    "kimi_vl": VLMTestInfo(
        models=["moonshotai/Kimi-VL-A3B-Instruct"],
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
516
        prompt_formatter=lambda img_prompt: f"<|im_user|>user<|im_middle|>{img_prompt}<|im_end|><|im_assistant|>assistant<|im_middle|>",  # noqa: E501
517
518
519
520
521
522
523
524
        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)],
    ),
525
526
    "llama4": VLMTestInfo(
        models=["meta-llama/Llama-4-Scout-17B-16E-Instruct"],
527
        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
528
529
530
        img_idx_to_prompt=lambda _: "<|image|>",
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
        distributed_executor_backend="mp",
531
        image_size_factors=[(0.25, 0.5, 1.0)],
532
533
534
535
536
        hf_model_kwargs={"device_map": "auto"},
        max_model_len=8192,
        max_num_seqs=4,
        dtype="bfloat16",
        auto_cls=AutoModelForImageTextToText,
537
538
        tensor_parallel_size=4,
        marks=multi_gpu_marks(num_gpus=4),
539
    ),
540
541
542
543
544
    "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,
545
        auto_cls=AutoModelForImageTextToText,
546
        vllm_output_post_proc=model_utils.llava_image_vllm_to_hf_output,
547
548
549
550
551
552
553
554
        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},
            )
        ],
555
    ),
556
    "llava_onevision": VLMTestInfo(
557
558
        models=["llava-hf/llava-onevision-qwen2-0.5b-ov-hf"],
        test_type=VLMTestType.CUSTOM_INPUTS,
559
        prompt_formatter=lambda vid_prompt: f"<|im_start|>user\n{vid_prompt}<|im_end|>\n<|im_start|>assistant\n",  # noqa: E501
560
561
        num_video_frames=16,
        max_model_len=16384,
562
563
        hf_model_kwargs=model_utils.llava_onevision_hf_model_kwargs(
            "llava-hf/llava-onevision-qwen2-0.5b-ov-hf"
564
        ),
565
        auto_cls=AutoModelForImageTextToText,
566
        vllm_output_post_proc=model_utils.llava_onevision_vllm_to_hf_output,
567
568
569
570
571
572
573
574
        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},
            )
        ],
575
576
577
578
579
580
581
    ),
    "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,
582
        max_num_seqs=2,
583
        auto_cls=AutoModelForImageTextToText,
584
585
        vllm_output_post_proc=model_utils.llava_video_vllm_to_hf_output,
    ),
586
587
588
589
590
591
    "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],
592
        auto_cls=AutoModelForImageTextToText,
593
594
595
        vllm_output_post_proc=model_utils.mantis_vllm_to_hf_output,
        patch_hf_runner=model_utils.mantis_patch_hf_runner,
    ),
596
    "minicpmv_25": VLMTestInfo(
597
        models=["openbmb/MiniCPM-Llama3-V-2_5"],
598
        test_type=VLMTestType.IMAGE,
599
600
601
602
603
        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],
604
        hf_output_post_proc=model_utils.minicpmv_trunc_hf_output,
605
        patch_hf_runner=model_utils.minicpmv_25_patch_hf_runner,
606
607
        # FIXME: https://huggingface.co/openbmb/MiniCPM-V-2_6/discussions/55
        marks=[pytest.mark.skip("HF import fails")],
608
    ),
609
610
    "minicpmo_26": VLMTestInfo(
        models=["openbmb/MiniCPM-o-2_6"],
611
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
612
613
614
615
        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,
616
617
        get_stop_token_ids=lambda tok: tok.convert_tokens_to_ids(
            ["<|im_end|>", "<|endoftext|>"]
618
        ),
619
620
        hf_output_post_proc=model_utils.minicpmv_trunc_hf_output,
        patch_hf_runner=model_utils.minicpmo_26_patch_hf_runner,
621
        # FIXME: https://huggingface.co/openbmb/MiniCPM-o-2_6/discussions/49
622
        marks=[pytest.mark.skip("HF import fails")],
623
    ),
624
625
    "minicpmv_26": VLMTestInfo(
        models=["openbmb/MiniCPM-V-2_6"],
626
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
627
628
629
630
        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,
631
632
        get_stop_token_ids=lambda tok: tok.convert_tokens_to_ids(
            ["<|im_end|>", "<|endoftext|>"]
633
        ),
634
635
636
        hf_output_post_proc=model_utils.minicpmv_trunc_hf_output,
        patch_hf_runner=model_utils.minicpmv_26_patch_hf_runner,
    ),
637
638
    "minimax_vl_01": VLMTestInfo(
        models=["MiniMaxAI/MiniMax-VL-01"],
639
        prompt_formatter=lambda img_prompt: f"<beginning_of_sentence>user: {img_prompt} assistant:<end_of_sentence>",  # noqa: E501
640
641
642
643
644
645
646
647
648
649
        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)],
    ),
650
651
    "molmo": VLMTestInfo(
        models=["allenai/Molmo-7B-D-0924"],
652
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
653
        prompt_formatter=identity,
654
655
        max_model_len=4096,
        max_num_seqs=2,
656
        patch_hf_runner=model_utils.molmo_patch_hf_runner,
657
    ),
658
659
660
    "ovis1_6-gemma2": VLMTestInfo(
        models=["AIDC-AI/Ovis1.6-Gemma2-9B"],
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
661
        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
662
        img_idx_to_prompt=lambda idx: "<image>\n",
663
664
665
666
667
668
669
670
        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)],
    ),
671
672
673
    "ovis2": VLMTestInfo(
        models=["AIDC-AI/Ovis2-1B"],
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
674
        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
675
        img_idx_to_prompt=lambda idx: "<image>\n",
676
677
678
679
680
        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"},
681
        patch_hf_runner=model_utils.ovis_patch_hf_runner,
682
    ),
683
684
    "ovis2_5": VLMTestInfo(
        models=["AIDC-AI/Ovis2.5-2B"],
685
686
        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
687
        img_idx_to_prompt=lambda idx: "<image>\n",
688
689
690
691
692
693
        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,
694
        hf_model_kwargs={"revision": "refs/pr/5"},
695
    ),
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
    "paddleocr_vl": VLMTestInfo(
        models=["PaddlePaddle/PaddleOCR-VL"],
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
        prompt_formatter=lambda img_prompt: f"USER: {img_prompt}\nASSISTANT:",
        img_idx_to_prompt=lambda idx: (
            "<|IMAGE_START|><|IMAGE_PLACEHOLDER|><|IMAGE_END|>"
        ),
        multi_image_prompt=(
            "Image-1: <|IMAGE_START|><|IMAGE_PLACEHOLDER|><|IMAGE_END|>\n"
            "Image-2: <|IMAGE_START|><|IMAGE_PLACEHOLDER|><|IMAGE_END|>\n"
            "Describe these two images separately."
        ),
        max_model_len=8192,
        max_num_seqs=2,
        auto_cls=AutoModelForCausalLM,
711
        image_size_factors=[(0.25,)],
712
713
714
715
716
717
        marks=[
            pytest.mark.skipif(
                Version(TRANSFORMERS_VERSION) == Version("4.57.3"),
                reason="This model is broken in Transformers v4.57.3",
            )
        ],
718
    ),
719
720
721
    "phi3v": VLMTestInfo(
        models=["microsoft/Phi-3.5-vision-instruct"],
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
722
        prompt_formatter=lambda img_prompt: f"<|user|>\n{img_prompt}<|end|>\n<|assistant|>\n",  # noqa: E501
723
724
725
        img_idx_to_prompt=lambda idx: f"<|image_{idx}|>\n",
        max_model_len=4096,
        max_num_seqs=2,
726
        runner="generate",
727
728
        # use sdpa mode for hf runner since phi3v didn't work with flash_attn
        hf_model_kwargs={"_attn_implementation": "sdpa"},
729
730
731
732
        use_tokenizer_eos=True,
        vllm_output_post_proc=model_utils.phi3v_vllm_to_hf_output,
        num_logprobs=10,
    ),
733
734
735
736
737
738
739
    "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,
740
        auto_cls=AutoModelForImageTextToText,
741
742
743
744
745
746
747
        marks=[
            large_gpu_mark(min_gb=48),
            pytest.mark.skipif(
                current_platform.is_rocm(),
                reason="Model produces a vector of <UNK> output in HF on ROCm",
            ),
        ],
748
    ),
749
    "qwen_vl": VLMTestInfo(
750
751
752
753
754
755
756
757
758
        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,
    ),
759
760
    "qwen2_vl": VLMTestInfo(
        models=["Qwen/Qwen2-VL-2B-Instruct"],
761
762
        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
763
764
        img_idx_to_prompt=lambda idx: "<|vision_start|><|image_pad|><|vision_end|>",
        video_idx_to_prompt=lambda idx: "<|vision_start|><|video_pad|><|vision_end|>",
765
        multi_image_prompt="Picture 1: <vlm_image>\nPicture 2: <vlm_image>\nDescribe these two images with one paragraph respectively.",  # noqa: E501
766
767
        max_model_len=4096,
        max_num_seqs=2,
768
        auto_cls=AutoModelForImageTextToText,
769
        vllm_output_post_proc=model_utils.qwen2_vllm_to_hf_output,
770
        image_size_factors=[(0.25,), (0.25, 0.25, 0.25), (0.25, 0.2, 0.15)],
771
772
        marks=[pytest.mark.cpu_model],
    ),
773
774
775
    "skywork_r1v": VLMTestInfo(
        models=["Skywork/Skywork-R1V-38B"],
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
776
777
778
        prompt_formatter=lambda img_prompt: f"<|begin▁of▁sentence|><|User|>\n{img_prompt}<|Assistant|><think>\n",  # noqa: E501
        single_image_prompts=IMAGE_ASSETS.prompts(
            {
779
                "stop_sign": "<image>\nWhat's the content in the center of the image?",
780
781
782
                "cherry_blossom": "<image>\nWhat is the season?",
            }
        ),
783
        multi_image_prompt="<image>\n<image>\nDescribe the two images in short.",
784
785
786
787
788
        max_model_len=4096,
        use_tokenizer_eos=True,
        patch_hf_runner=model_utils.skyworkr1v_patch_hf_runner,
        marks=[large_gpu_mark(min_gb=80)],
    ),
789
790
791
    "smolvlm": VLMTestInfo(
        models=["HuggingFaceTB/SmolVLM2-2.2B-Instruct"],
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
792
        prompt_formatter=lambda img_prompt: f"<|im_start|>User:{img_prompt}<end_of_utterance>\nAssistant:",  # noqa: E501
793
794
795
796
797
        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,
798
        num_logprobs=10,
799
    ),
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
    "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,
        ),
816
        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
817
818
        img_idx_to_prompt=lambda idx: "<|vision_start|><|image_pad|><|vision_end|>",
        video_idx_to_prompt=lambda idx: "<|vision_start|><|video_pad|><|vision_end|>",
819
820
821
        max_model_len=4096,
        max_num_seqs=2,
        auto_cls=AutoModelForImageTextToText,
822
        image_size_factors=[(0.25,), (0.25, 0.25, 0.25), (0.25, 0.2, 0.15)],
823
824
        marks=[pytest.mark.skip("Model initialization hangs")],
    ),
825
    ### Tensor parallel / multi-gpu broadcast tests
826
    "chameleon-broadcast": VLMTestInfo(
827
828
829
        models=["facebook/chameleon-7b"],
        prompt_formatter=lambda img_prompt: f"USER: {img_prompt}\nASSISTANT:",
        max_model_len=4096,
830
        auto_cls=AutoModelForImageTextToText,
831
832
        vllm_output_post_proc=lambda vllm_output, model: vllm_output[:2],
        hf_output_post_proc=lambda hf_output, model: hf_output[:2],
833
        comparator=check_outputs_equal,
834
        marks=multi_gpu_marks(num_gpus=2),
835
        **COMMON_BROADCAST_SETTINGS,  # type: ignore
836
    ),
837
    "llava-broadcast": VLMTestInfo(
838
839
840
        models=["llava-hf/llava-1.5-7b-hf"],
        prompt_formatter=lambda img_prompt: f"USER: {img_prompt}\nASSISTANT:",
        max_model_len=4096,
841
        auto_cls=AutoModelForImageTextToText,
842
        vllm_output_post_proc=model_utils.llava_image_vllm_to_hf_output,
843
        marks=multi_gpu_marks(num_gpus=2),
844
        **COMMON_BROADCAST_SETTINGS,  # type: ignore
845
    ),
846
    "llava_next-broadcast": VLMTestInfo(
847
848
849
        models=["llava-hf/llava-v1.6-mistral-7b-hf"],
        prompt_formatter=lambda img_prompt: f"[INST] {img_prompt} [/INST]",
        max_model_len=10240,
850
        auto_cls=AutoModelForImageTextToText,
851
        vllm_output_post_proc=model_utils.llava_image_vllm_to_hf_output,
852
        marks=multi_gpu_marks(num_gpus=2),
853
        **COMMON_BROADCAST_SETTINGS,  # type: ignore
854
855
856
857
    ),
    ### Custom input edge-cases for specific models
    "intern_vl-diff-patches": VLMTestInfo(
        models=["OpenGVLab/InternVL2-2B"],
858
        prompt_formatter=lambda img_prompt: f"<|im_start|>User\n{img_prompt}<|im_end|>\n<|im_start|>Assistant\n",  # noqa: E501
859
860
861
862
863
864
865
866
        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},
867
868
            )
            for inp in custom_inputs.different_patch_input_cases_internvl()
869
870
        ],
    ),
871
    "llava_onevision-multiple-images": VLMTestInfo(
872
873
874
875
        models=["llava-hf/llava-onevision-qwen2-0.5b-ov-hf"],
        test_type=VLMTestType.CUSTOM_INPUTS,
        max_model_len=16384,
        max_num_seqs=2,
876
        auto_cls=AutoModelForImageTextToText,
877
878
        hf_model_kwargs=model_utils.llava_onevision_hf_model_kwargs(
            "llava-hf/llava-onevision-qwen2-0.5b-ov-hf"
879
        ),
880
        vllm_output_post_proc=model_utils.llava_onevision_vllm_to_hf_output,
881
882
883
884
885
886
887
888
        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},
            )
        ],
889
890
891
892
893
894
        marks=[
            pytest.mark.skipif(
                Version(TRANSFORMERS_VERSION) == Version("4.57.1"),
                reason="This model is broken in Transformers v4.57.1",
            )
        ],
895
    ),
896
897
898
899
900
901
    # 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,
902
        auto_cls=AutoModelForImageTextToText,
903
        vllm_output_post_proc=model_utils.qwen2_vllm_to_hf_output,
904
905
906
907
908
909
        custom_test_opts=[
            CustomTestOptions(
                inputs=custom_inputs.windows_attention_image_qwen2_5_vl(),
                limit_mm_per_prompt={"image": 1},
            )
        ],
910
    ),
911
912
913
}


914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
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):
932
        models_in_group = models[i * split_size : (i + 1) * split_size]
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948

        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)


949
950
951
952
953
954
### Test wrappers
# Wrappers around the core test running func for:
# - single image
# - multi-image
# - image embeddings
# - video
955
# - audio
956
# - custom inputs
957
958
959
960
961
@pytest.mark.parametrize(
    "model_type,test_case",
    get_parametrized_options(
        VLM_TEST_SETTINGS,
        test_type=VLMTestType.IMAGE,
962
        create_new_process_for_each_test=False,
963
964
    ),
)
965
966
967
968
969
970
971
972
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,
):
973
974
975
976
977
978
979
980
981
982
983
    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,
    )


984
985
986
987
988
@pytest.mark.parametrize(
    "model_type,test_case",
    get_parametrized_options(
        VLM_TEST_SETTINGS,
        test_type=VLMTestType.MULTI_IMAGE,
989
        create_new_process_for_each_test=False,
990
991
    ),
)
992
993
994
995
996
997
998
999
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,
):
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
    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,
    )


1011
1012
1013
1014
1015
@pytest.mark.parametrize(
    "model_type,test_case",
    get_parametrized_options(
        VLM_TEST_SETTINGS,
        test_type=VLMTestType.EMBEDDING,
1016
        create_new_process_for_each_test=False,
1017
1018
    ),
)
1019
1020
1021
1022
1023
1024
1025
def test_image_embedding_models(
    model_type: str,
    test_case: ExpandableVLMTestArgs,
    hf_runner: type[HfRunner],
    vllm_runner: type[VllmRunner],
    image_assets: ImageTestAssets,
):
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
    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,
    )


1036
1037
1038
1039
1040
@pytest.mark.parametrize(
    "model_type,test_case",
    get_parametrized_options(
        VLM_TEST_SETTINGS,
        test_type=VLMTestType.VIDEO,
1041
        create_new_process_for_each_test=False,
1042
1043
    ),
)
1044
1045
1046
1047
1048
1049
1050
def test_video_models(
    model_type: str,
    test_case: ExpandableVLMTestArgs,
    hf_runner: type[HfRunner],
    vllm_runner: type[VllmRunner],
    video_assets: VideoTestAssets,
):
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
    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,
    )


1061
1062
1063
1064
1065
1066
@pytest.mark.parametrize(
    "model_type,test_case",
    get_parametrized_options(
        VLM_TEST_SETTINGS,
        test_type=VLMTestType.AUDIO,
        create_new_process_for_each_test=False,
1067
1068
    ),
)
1069
1070
1071
1072
1073
1074
1075
def test_audio_models(
    model_type: str,
    test_case: ExpandableVLMTestArgs,
    hf_runner: type[HfRunner],
    vllm_runner: type[VllmRunner],
    audio_assets: AudioTestAssets,
):
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
    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,
    )


1086
1087
1088
1089
1090
@pytest.mark.parametrize(
    "model_type,test_case",
    get_parametrized_options(
        VLM_TEST_SETTINGS,
        test_type=VLMTestType.CUSTOM_INPUTS,
1091
        create_new_process_for_each_test=False,
1092
1093
    ),
)
1094
1095
1096
def test_custom_inputs_models(
    model_type: str,
    test_case: ExpandableVLMTestArgs,
1097
1098
    hf_runner: type[HfRunner],
    vllm_runner: type[VllmRunner],
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
):
    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
1110
1111
1112
1113
1114
@pytest.mark.parametrize(
    "model_type,test_case",
    get_parametrized_options(
        VLM_TEST_SETTINGS,
        test_type=VLMTestType.IMAGE,
1115
        create_new_process_for_each_test=True,
1116
1117
    ),
)
1118
@create_new_process_for_each_test()
1119
1120
1121
1122
1123
1124
1125
1126
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,
):
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
    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,
    )


1138
1139
1140
1141
1142
@pytest.mark.parametrize(
    "model_type,test_case",
    get_parametrized_options(
        VLM_TEST_SETTINGS,
        test_type=VLMTestType.MULTI_IMAGE,
1143
        create_new_process_for_each_test=True,
1144
1145
    ),
)
1146
@create_new_process_for_each_test()
1147
1148
1149
1150
1151
1152
1153
1154
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,
):
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
    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,
    )


1166
1167
1168
1169
1170
@pytest.mark.parametrize(
    "model_type,test_case",
    get_parametrized_options(
        VLM_TEST_SETTINGS,
        test_type=VLMTestType.EMBEDDING,
1171
        create_new_process_for_each_test=True,
1172
1173
    ),
)
1174
@create_new_process_for_each_test()
1175
1176
1177
1178
1179
1180
1181
def test_image_embedding_models_heavy(
    model_type: str,
    test_case: ExpandableVLMTestArgs,
    hf_runner: type[HfRunner],
    vllm_runner: type[VllmRunner],
    image_assets: ImageTestAssets,
):
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
    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,
    )


1192
1193
1194
1195
1196
@pytest.mark.parametrize(
    "model_type,test_case",
    get_parametrized_options(
        VLM_TEST_SETTINGS,
        test_type=VLMTestType.VIDEO,
1197
        create_new_process_for_each_test=True,
1198
1199
    ),
)
1200
1201
1202
1203
1204
1205
1206
def test_video_models_heavy(
    model_type: str,
    test_case: ExpandableVLMTestArgs,
    hf_runner: type[HfRunner],
    vllm_runner: type[VllmRunner],
    video_assets: VideoTestAssets,
):
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
    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,
    )


1217
1218
1219
1220
1221
1222
@pytest.mark.parametrize(
    "model_type,test_case",
    get_parametrized_options(
        VLM_TEST_SETTINGS,
        test_type=VLMTestType.AUDIO,
        create_new_process_for_each_test=True,
1223
1224
    ),
)
1225
1226
1227
1228
1229
1230
1231
def test_audio_models_heavy(
    model_type: str,
    test_case: ExpandableVLMTestArgs,
    hf_runner: type[HfRunner],
    vllm_runner: type[VllmRunner],
    audio_assets: AudioTestAssets,
):
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
    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,
    )


1242
1243
1244
1245
1246
@pytest.mark.parametrize(
    "model_type,test_case",
    get_parametrized_options(
        VLM_TEST_SETTINGS,
        test_type=VLMTestType.CUSTOM_INPUTS,
1247
        create_new_process_for_each_test=True,
1248
1249
    ),
)
1250
@create_new_process_for_each_test()
1251
1252
1253
def test_custom_inputs_models_heavy(
    model_type: str,
    test_case: ExpandableVLMTestArgs,
1254
1255
    hf_runner: type[HfRunner],
    vllm_runner: type[VllmRunner],
1256
1257
1258
1259
1260
1261
1262
1263
):
    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,
    )