test_common.py 47.8 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
import math
8
import os
9
from collections import defaultdict
10
11
12
from pathlib import PosixPath

import pytest
13
14
15
16
17
from transformers import (
    AutoModel,
    AutoModelForImageTextToText,
    AutoModelForTextToWaveform,
)
18
19

from vllm.platforms import current_platform
20
from vllm.utils import identity
21

22
23
24
25
26
27
28
29
30
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
31
32
33
from ...utils import check_outputs_equal
from .vlm_utils import custom_inputs, model_utils, runners
from .vlm_utils.case_filtering import get_parametrized_options
34
35
36
37
38
39
from .vlm_utils.types import (
    CustomTestOptions,
    ExpandableVLMTestArgs,
    VLMTestInfo,
    VLMTestType,
)
40
41
42
43
44
45
46
47
48
49
50
51
52

# This hack is needed for phi3v & paligemma models
# ROCm Triton FA can run into shared memory issues with these models,
# use other backends in the meantime
# FIXME (mattwong, gshtrasb, hongxiayan)
if current_platform.is_rocm():
    os.environ["VLLM_USE_TRITON_FLASH_ATTN"] = "0"

COMMON_BROADCAST_SETTINGS = {
    "test_type": VLMTestType.IMAGE,
    "dtype": "half",
    "max_tokens": 5,
    "tensor_parallel_size": 2,
53
    "hf_model_kwargs": {"device_map": "auto"},
54
    "image_size_factors": [(0.25, 0.5, 1.0)],
55
56
57
    "distributed_executor_backend": (
        "ray",
        "mp",
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
90
91
92
93
94
}

### 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 = {
95
96
97
    #### Core tests to always run in the CI
    "llava": VLMTestInfo(
        models=["llava-hf/llava-1.5-7b-hf"],
98
        test_type=(VLMTestType.EMBEDDING, VLMTestType.IMAGE, VLMTestType.CUSTOM_INPUTS),
99
100
101
        prompt_formatter=lambda img_prompt: f"USER: {img_prompt}\nASSISTANT:",
        convert_assets_to_embeddings=model_utils.get_llava_embeddings,
        max_model_len=4096,
102
        auto_cls=AutoModelForImageTextToText,
103
        vllm_output_post_proc=model_utils.llava_image_vllm_to_hf_output,
104
105
106
107
108
109
110
111
        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},
            )
        ],
112
113
        # TODO: Revert to "auto" when CPU backend can use torch > 2.6
        dtype="bfloat16" if current_platform.is_cpu() else "auto",
114
        marks=[pytest.mark.core_model, pytest.mark.cpu_model],
115
116
117
118
119
    ),
    "paligemma": VLMTestInfo(
        models=["google/paligemma-3b-mix-224"],
        test_type=VLMTestType.IMAGE,
        prompt_formatter=identity,
120
        img_idx_to_prompt=lambda idx: "",
121
        # Paligemma uses its own sample prompts because the default one fails
122
123
124
125
126
127
        single_image_prompts=IMAGE_ASSETS.prompts(
            {
                "stop_sign": "caption es",
                "cherry_blossom": "What is in the picture?",
            }
        ),
128
        auto_cls=AutoModelForImageTextToText,
129
        vllm_output_post_proc=model_utils.paligemma_vllm_to_hf_output,
130
        dtype="bfloat16",
131
132
        marks=[
            pytest.mark.skip(reason="vLLM does not support PrefixLM attention mask")
133
        ],
134
    ),
Roger Wang's avatar
Roger Wang committed
135
136
    "qwen2_5_vl": VLMTestInfo(
        models=["Qwen/Qwen2.5-VL-3B-Instruct"],
137
138
        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
139
140
        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
141
142
        max_model_len=4096,
        max_num_seqs=2,
143
        auto_cls=AutoModelForImageTextToText,
Roger Wang's avatar
Roger Wang committed
144
145
        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)],
146
        marks=[pytest.mark.core_model, pytest.mark.cpu_model],
Roger Wang's avatar
Roger Wang committed
147
    ),
148
    "qwen2_5_omni": VLMTestInfo(
149
        models=["Qwen/Qwen2.5-Omni-3B"],
150
151
        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
152
153
        img_idx_to_prompt=lambda idx: "<|vision_bos|><|IMAGE|><|vision_eos|>",
        video_idx_to_prompt=lambda idx: "<|vision_bos|><|VIDEO|><|vision_eos|>",
154
155
        max_model_len=4096,
        max_num_seqs=2,
156
        num_logprobs=6 if current_platform.is_cpu() else 5,
157
        auto_cls=AutoModelForTextToWaveform,
158
        vllm_output_post_proc=model_utils.qwen2_vllm_to_hf_output,
159
        patch_hf_runner=model_utils.qwen2_5_omni_patch_hf_runner,
160
161
162
        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],
    ),
163
    "ultravox": VLMTestInfo(
164
        models=["fixie-ai/ultravox-v0_5-llama-3_2-1b"],
165
        test_type=VLMTestType.AUDIO,
166
        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
167
168
169
170
171
172
173
        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],
    ),
174
175
176
177
178
179
    #### 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,
180
        prompt_formatter=lambda vid_prompt: f"<|im_start|>user\n{vid_prompt}<|im_end|>\n<|im_start|>assistant\n",  # noqa: E501
181
        max_model_len=16384,
182
183
        hf_model_kwargs=model_utils.llava_onevision_hf_model_kwargs(
            "llava-hf/llava-onevision-qwen2-0.5b-ov-hf"
184
        ),
185
186
187
188
189
        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",
190
            "default_torch_num_threads": 1,
191
        },
192
193
194
        # FIXME: Investigate why the test hangs
        # when processing the 3rd prompt in vLLM
        marks=[pytest.mark.core_model, pytest.mark.skip(reason="Test hangs")],
195
    ),
196
197
198
    "idefics3-transformers": VLMTestInfo(
        models=["HuggingFaceTB/SmolVLM-256M-Instruct"],
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
199
        prompt_formatter=lambda img_prompt: f"<|begin_of_text|>User:{img_prompt}<end_of_utterance>\nAssistant:",  # noqa: E501
200
201
202
203
204
205
206
207
208
209
210
        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],
    ),
211
212
213
214
    # 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,
215
        prompt_formatter=lambda img_prompt: f"<|im_start|>User\n{img_prompt}<|im_end|>\n<|im_start|>assistant\n",  # noqa: E501
216
        img_idx_to_prompt=lambda idx: "<|vision_start|><|image_pad|><|vision_end|>",
217
218
219
220
221
222
223
224
        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",
        },
225
        # FIXME: Investigate mrope issue
226
        marks=[large_gpu_mark(min_gb=32), pytest.mark.skip(reason="Mrope issue")],
227
    ),
228
    #### Extended model tests
229
230
231
    "aria": VLMTestInfo(
        models=["rhymes-ai/Aria"],
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
232
        prompt_formatter=lambda img_prompt: f"<|im_start|>user\n{img_prompt}<|im_end|>\n<|im_start|>assistant\n ",  # noqa: E501
233
234
235
236
        img_idx_to_prompt=lambda idx: "<fim_prefix><|img|><fim_suffix>\n",
        max_model_len=4096,
        max_num_seqs=2,
        auto_cls=AutoModelForImageTextToText,
237
238
239
        single_image_prompts=IMAGE_ASSETS.prompts(
            {
                "stop_sign": "<vlm_image>Please describe the image shortly.",
240
                "cherry_blossom": "<vlm_image>Please infer the season with reason.",
241
242
            }
        ),
243
        multi_image_prompt="<vlm_image><vlm_image>Describe the two images shortly.",
244
245
246
247
248
        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
249
250
    "aya_vision": VLMTestInfo(
        models=["CohereForAI/aya-vision-8b"],
251
        test_type=(VLMTestType.IMAGE),
252
253
254
        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(
            {
255
256
                "stop_sign": "<image>What's the content in the center of the image?",
                "cherry_blossom": "<image>What is the season?",
257
258
            }
        ),
259
        multi_image_prompt="<image><image>Describe the two images in detail.",
260
261
262
263
264
265
266
267
        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),
268
269
270
        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(
            {
271
272
                "stop_sign": "<image>What's the content in the center of the image?",
                "cherry_blossom": "<image>What is the season?",
273
274
            }
        ),
275
        multi_image_prompt="<image><image>Describe the two images in detail.",
276
        max_model_len=4096,
Jennifer Zhao's avatar
Jennifer Zhao committed
277
278
        max_num_seqs=2,
        auto_cls=AutoModelForImageTextToText,
279
280
        vllm_runner_kwargs={"mm_processor_kwargs": {"crop_to_patches": True}},
        marks=[large_gpu_mark(min_gb=32)],
Jennifer Zhao's avatar
Jennifer Zhao committed
281
    ),
282
    "blip2": VLMTestInfo(
283
        models=["Salesforce/blip2-opt-2.7b"],
284
285
286
        test_type=VLMTestType.IMAGE,
        prompt_formatter=lambda img_prompt: f"Question: {img_prompt} Answer:",
        img_idx_to_prompt=lambda idx: "",
287
        auto_cls=AutoModelForImageTextToText,
288
        vllm_output_post_proc=model_utils.blip2_vllm_to_hf_output,
289
290
        # FIXME: https://github.com/huggingface/transformers/pull/38510
        marks=[pytest.mark.skip("Model is broken")],
291
292
293
294
295
296
    ),
    "chameleon": VLMTestInfo(
        models=["facebook/chameleon-7b"],
        test_type=VLMTestType.IMAGE,
        prompt_formatter=lambda img_prompt: f"USER: {img_prompt}\nASSISTANT:",
        max_model_len=4096,
297
        max_num_seqs=2,
298
        auto_cls=AutoModelForImageTextToText,
299
        # For chameleon, we only compare the sequences
300
301
        vllm_output_post_proc=lambda vllm_output, model: vllm_output[:2],
        hf_output_post_proc=lambda hf_output, model: hf_output[:2],
302
303
304
305
        comparator=check_outputs_equal,
        max_tokens=8,
        dtype="bfloat16",
    ),
306
    "deepseek_vl_v2": VLMTestInfo(
307
        models=["Isotr0py/deepseek-vl2-tiny"],  # model repo using dynamic module
308
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
309
        prompt_formatter=lambda img_prompt: f"<|User|>: {img_prompt}\n\n<|Assistant|>: ",  # noqa: E501
310
311
        max_model_len=4096,
        max_num_seqs=2,
312
313
        single_image_prompts=IMAGE_ASSETS.prompts(
            {
314
                "stop_sign": "<image>\nWhat's the content in the center of the image?",
315
316
317
318
                "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
319
320
        patch_hf_runner=model_utils.deepseekvl2_patch_hf_runner,
        hf_output_post_proc=model_utils.deepseekvl2_trunc_hf_output,
321
        stop_str=["<|end▁of▁sentence|>", "<|begin▁of▁sentence|>"],
322
        image_size_factors=[(), (1.0,), (1.0, 1.0, 1.0), (0.1, 0.5, 1.0)],
323
    ),
324
325
326
327
328
329
330
    "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,
331
        auto_cls=AutoModelForImageTextToText,
332
333
334
335
        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)],
336
        marks=[large_gpu_mark(min_gb=32)],
337
    ),
338
339
340
    "gemma3": VLMTestInfo(
        models=["google/gemma-3-4b-it"],
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
341
342
343
344
        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
345
                "cherry_blossom": "<start_of_image>What is the season?",
346
347
            }
        ),
348
349
350
        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,
351
        auto_cls=AutoModelForImageTextToText,
352
353
        vllm_runner_kwargs={"mm_processor_kwargs": {"do_pan_and_scan": True}},
        patch_hf_runner=model_utils.gemma3_patch_hf_runner,
354
        num_logprobs=10,
355
    ),
356
    "glm4v": VLMTestInfo(
357
        models=["zai-org/glm-4v-9b"],
358
        test_type=VLMTestType.IMAGE,
359
        prompt_formatter=lambda img_prompt: f"<|user|>\n{img_prompt}<|assistant|>",
360
361
362
363
364
365
        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
            }
        ),
366
367
368
        max_model_len=2048,
        max_num_seqs=2,
        get_stop_token_ids=lambda tok: [151329, 151336, 151338],
369
370
371
372
373
374
        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,
375
        marks=[large_gpu_mark(min_gb=32)],
376
    ),
377
    "glm4_1v": VLMTestInfo(
378
        models=["zai-org/GLM-4.1V-9B-Thinking"],
379
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
380
381
382
        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|>",
383
384
385
386
387
388
        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,
389
        marks=[large_gpu_mark(min_gb=32)],
390
391
    ),
    "glm4_1v-video": VLMTestInfo(
392
        models=["zai-org/GLM-4.1V-9B-Thinking"],
393
394
395
396
397
398
        # 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,
399
400
401
402
403
404
        custom_test_opts=[
            CustomTestOptions(
                inputs=custom_inputs.video_with_metadata_glm4_1v(),
                limit_mm_per_prompt={"video": 1},
            )
        ],
405
        marks=[large_gpu_mark(min_gb=32)],
406
    ),
407
    "h2ovl": VLMTestInfo(
408
        models=[
409
            "h2oai/h2ovl-mississippi-800m",
410
            "h2oai/h2ovl-mississippi-2b",
411
412
        ],
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
413
        prompt_formatter=lambda img_prompt: f"<|prompt|>{img_prompt}<|end|><|answer|>",
414
415
        single_image_prompts=IMAGE_ASSETS.prompts(
            {
416
                "stop_sign": "<image>\nWhat's the content in the center of the image?",
417
418
419
                "cherry_blossom": "<image>\nWhat is the season?",
            }
        ),
420
421
422
        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,
423
        num_logprobs=10,
424
        patch_hf_runner=model_utils.h2ovl_patch_hf_runner,
425
    ),
426
    "idefics3": VLMTestInfo(
427
        models=["HuggingFaceTB/SmolVLM-256M-Instruct"],
428
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
429
        prompt_formatter=lambda img_prompt: f"<|begin_of_text|>User:{img_prompt}<end_of_utterance>\nAssistant:",  # noqa: E501
430
431
432
        img_idx_to_prompt=lambda idx: "<image>",
        max_model_len=8192,
        max_num_seqs=2,
433
        auto_cls=AutoModelForImageTextToText,
434
        hf_output_post_proc=model_utils.idefics3_trunc_hf_output,
435
    ),
436
437
438
439
    "intern_vl": VLMTestInfo(
        models=[
            "OpenGVLab/InternVL2-1B",
            "OpenGVLab/InternVL2-2B",
440
441
            # FIXME: Config cannot be loaded in transformers 4.52
            # "OpenGVLab/Mono-InternVL-2B",
442
443
        ],
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
444
445
446
        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(
            {
447
                "stop_sign": "<image>\nWhat's the content in the center of the image?",
448
449
450
                "cherry_blossom": "<image>\nWhat is the season?",
            }
        ),
451
452
453
454
455
        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,
    ),
456
457
458
459
460
    "intern_vl-video": VLMTestInfo(
        models=[
            "OpenGVLab/InternVL3-1B",
        ],
        test_type=VLMTestType.VIDEO,
461
        prompt_formatter=lambda img_prompt: f"<|im_start|>User\n{img_prompt}<|im_end|>\n<|im_start|>Assistant\n",  # noqa: E501
462
463
464
465
466
        video_idx_to_prompt=lambda idx: "<video>",
        max_model_len=8192,
        use_tokenizer_eos=True,
        patch_hf_runner=model_utils.internvl_patch_hf_runner,
    ),
467
468
469
470
471
472
473
    "intern_vl-hf": VLMTestInfo(
        models=["OpenGVLab/InternVL3-1B-hf"],
        test_type=(
            VLMTestType.IMAGE,
            VLMTestType.MULTI_IMAGE,
            VLMTestType.VIDEO,
        ),
474
        prompt_formatter=lambda img_prompt: f"<|im_start|>User\n{img_prompt}<|im_end|>\n<|im_start|>Assistant\n",  # noqa: E501
475
476
477
478
479
480
        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,
    ),
481
482
483
    "kimi_vl": VLMTestInfo(
        models=["moonshotai/Kimi-VL-A3B-Instruct"],
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
484
        prompt_formatter=lambda img_prompt: f"<|im_user|>user<|im_middle|>{img_prompt}<|im_end|><|im_assistant|>assistant<|im_middle|>",  # noqa: E501
485
486
487
488
489
490
491
492
        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)],
    ),
493
494
    "llama4": VLMTestInfo(
        models=["meta-llama/Llama-4-Scout-17B-16E-Instruct"],
495
        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
496
497
498
        img_idx_to_prompt=lambda _: "<|image|>",
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
        distributed_executor_backend="mp",
499
        image_size_factors=[(0.25, 0.5, 1.0)],
500
501
502
503
504
        hf_model_kwargs={"device_map": "auto"},
        max_model_len=8192,
        max_num_seqs=4,
        dtype="bfloat16",
        auto_cls=AutoModelForImageTextToText,
505
506
        tensor_parallel_size=4,
        marks=multi_gpu_marks(num_gpus=4),
507
    ),
508
509
510
511
512
    "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,
513
        auto_cls=AutoModelForImageTextToText,
514
        vllm_output_post_proc=model_utils.llava_image_vllm_to_hf_output,
515
516
517
518
519
520
521
522
        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},
            )
        ],
523
    ),
524
    "llava_onevision": VLMTestInfo(
525
526
        models=["llava-hf/llava-onevision-qwen2-0.5b-ov-hf"],
        test_type=VLMTestType.CUSTOM_INPUTS,
527
        prompt_formatter=lambda vid_prompt: f"<|im_start|>user\n{vid_prompt}<|im_end|>\n<|im_start|>assistant\n",  # noqa: E501
528
529
        num_video_frames=16,
        max_model_len=16384,
530
531
        hf_model_kwargs=model_utils.llava_onevision_hf_model_kwargs(
            "llava-hf/llava-onevision-qwen2-0.5b-ov-hf"
532
        ),
533
        auto_cls=AutoModelForImageTextToText,
534
        vllm_output_post_proc=model_utils.llava_onevision_vllm_to_hf_output,
535
536
537
538
539
540
541
542
        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},
            )
        ],
543
544
545
546
547
548
549
    ),
    "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,
550
        max_num_seqs=2,
551
        auto_cls=AutoModelForImageTextToText,
552
553
        vllm_output_post_proc=model_utils.llava_video_vllm_to_hf_output,
    ),
554
555
556
557
558
559
    "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],
560
        auto_cls=AutoModelForImageTextToText,
561
562
563
        vllm_output_post_proc=model_utils.mantis_vllm_to_hf_output,
        patch_hf_runner=model_utils.mantis_patch_hf_runner,
    ),
564
    "minicpmv_25": VLMTestInfo(
565
        models=["openbmb/MiniCPM-Llama3-V-2_5"],
566
        test_type=VLMTestType.IMAGE,
567
568
569
570
571
        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],
572
        hf_output_post_proc=model_utils.minicpmv_trunc_hf_output,
573
        patch_hf_runner=model_utils.minicpmv_25_patch_hf_runner,
574
575
        # FIXME: https://huggingface.co/openbmb/MiniCPM-V-2_6/discussions/55
        marks=[pytest.mark.skip("HF import fails")],
576
    ),
577
578
    "minicpmo_26": VLMTestInfo(
        models=["openbmb/MiniCPM-o-2_6"],
579
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
580
581
582
583
        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,
584
585
        get_stop_token_ids=lambda tok: tok.convert_tokens_to_ids(
            ["<|im_end|>", "<|endoftext|>"]
586
        ),
587
588
        hf_output_post_proc=model_utils.minicpmv_trunc_hf_output,
        patch_hf_runner=model_utils.minicpmo_26_patch_hf_runner,
589
        # FIXME: https://huggingface.co/openbmb/MiniCPM-o-2_6/discussions/49
590
        marks=[pytest.mark.skip("HF import fails")],
591
    ),
592
593
    "minicpmv_26": VLMTestInfo(
        models=["openbmb/MiniCPM-V-2_6"],
594
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
595
596
597
598
        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,
599
600
        get_stop_token_ids=lambda tok: tok.convert_tokens_to_ids(
            ["<|im_end|>", "<|endoftext|>"]
601
        ),
602
603
604
        hf_output_post_proc=model_utils.minicpmv_trunc_hf_output,
        patch_hf_runner=model_utils.minicpmv_26_patch_hf_runner,
    ),
605
606
    "minimax_vl_01": VLMTestInfo(
        models=["MiniMaxAI/MiniMax-VL-01"],
607
        prompt_formatter=lambda img_prompt: f"<beginning_of_sentence>user: {img_prompt} assistant:<end_of_sentence>",  # noqa: E501
608
609
610
611
612
613
614
615
616
617
        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)],
    ),
618
619
    "molmo": VLMTestInfo(
        models=["allenai/Molmo-7B-D-0924"],
620
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
621
        prompt_formatter=identity,
622
623
        max_model_len=4096,
        max_num_seqs=2,
624
        patch_hf_runner=model_utils.molmo_patch_hf_runner,
625
    ),
626
627
628
    "ovis1_6-gemma2": VLMTestInfo(
        models=["AIDC-AI/Ovis1.6-Gemma2-9B"],
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
629
        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
630
        img_idx_to_prompt=lambda idx: "<image>\n",
631
632
633
634
635
636
637
638
        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)],
    ),
639
640
641
    "ovis2": VLMTestInfo(
        models=["AIDC-AI/Ovis2-1B"],
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
642
        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
643
        img_idx_to_prompt=lambda idx: "<image>\n",
644
645
646
647
648
        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"},
649
        patch_hf_runner=model_utils.ovis_patch_hf_runner,
650
    ),
651
652
    "ovis2_5": VLMTestInfo(
        models=["AIDC-AI/Ovis2.5-2B"],
653
654
        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
655
        img_idx_to_prompt=lambda idx: "<image>\n",
656
657
658
659
660
661
        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,
662
        hf_model_kwargs={"revision": "refs/pr/5"},
663
    ),
664
665
666
    "phi3v": VLMTestInfo(
        models=["microsoft/Phi-3.5-vision-instruct"],
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
667
        prompt_formatter=lambda img_prompt: f"<|user|>\n{img_prompt}<|end|>\n<|assistant|>\n",  # noqa: E501
668
669
670
        img_idx_to_prompt=lambda idx: f"<|image_{idx}|>\n",
        max_model_len=4096,
        max_num_seqs=2,
671
        runner="generate",
672
673
        # use sdpa mode for hf runner since phi3v didn't work with flash_attn
        hf_model_kwargs={"_attn_implementation": "sdpa"},
674
675
676
677
        use_tokenizer_eos=True,
        vllm_output_post_proc=model_utils.phi3v_vllm_to_hf_output,
        num_logprobs=10,
    ),
678
679
680
681
682
683
684
    "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,
685
        auto_cls=AutoModelForImageTextToText,
686
        marks=[large_gpu_mark(min_gb=48)],
687
    ),
688
    "qwen_vl": VLMTestInfo(
689
690
691
692
693
694
695
696
        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,
697
698
        # FIXME: https://github.com/huggingface/transformers/issues/38358
        marks=[pytest.mark.skip("Model initialization fails")],
699
    ),
700
701
    "qwen2_vl": VLMTestInfo(
        models=["Qwen/Qwen2-VL-2B-Instruct"],
702
703
        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
704
705
        img_idx_to_prompt=lambda idx: "<|vision_start|><|image_pad|><|vision_end|>",
        video_idx_to_prompt=lambda idx: "<|vision_start|><|video_pad|><|vision_end|>",
706
        multi_image_prompt="Picture 1: <vlm_image>\nPicture 2: <vlm_image>\nDescribe these two images with one paragraph respectively.",  # noqa: E501
707
708
        max_model_len=4096,
        max_num_seqs=2,
709
        auto_cls=AutoModelForImageTextToText,
710
711
712
713
        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],
    ),
714
715
716
    "skywork_r1v": VLMTestInfo(
        models=["Skywork/Skywork-R1V-38B"],
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
717
718
719
        prompt_formatter=lambda img_prompt: f"<|begin▁of▁sentence|><|User|>\n{img_prompt}<|Assistant|><think>\n",  # noqa: E501
        single_image_prompts=IMAGE_ASSETS.prompts(
            {
720
                "stop_sign": "<image>\nWhat's the content in the center of the image?",
721
722
723
                "cherry_blossom": "<image>\nWhat is the season?",
            }
        ),
724
        multi_image_prompt="<image>\n<image>\nDescribe the two images in short.",
725
726
727
728
729
        max_model_len=4096,
        use_tokenizer_eos=True,
        patch_hf_runner=model_utils.skyworkr1v_patch_hf_runner,
        marks=[large_gpu_mark(min_gb=80)],
    ),
730
731
732
    "smolvlm": VLMTestInfo(
        models=["HuggingFaceTB/SmolVLM2-2.2B-Instruct"],
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
733
        prompt_formatter=lambda img_prompt: f"<|im_start|>User:{img_prompt}<end_of_utterance>\nAssistant:",  # noqa: E501
734
735
736
737
738
739
        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,
    ),
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
    "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,
        ),
756
        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
757
758
        img_idx_to_prompt=lambda idx: "<|vision_start|><|image_pad|><|vision_end|>",
        video_idx_to_prompt=lambda idx: "<|vision_start|><|video_pad|><|vision_end|>",
759
760
761
762
763
764
        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")],
    ),
765
    ### Tensor parallel / multi-gpu broadcast tests
766
    "chameleon-broadcast": VLMTestInfo(
767
768
769
        models=["facebook/chameleon-7b"],
        prompt_formatter=lambda img_prompt: f"USER: {img_prompt}\nASSISTANT:",
        max_model_len=4096,
770
        auto_cls=AutoModelForImageTextToText,
771
772
        vllm_output_post_proc=lambda vllm_output, model: vllm_output[:2],
        hf_output_post_proc=lambda hf_output, model: hf_output[:2],
773
        comparator=check_outputs_equal,
774
        marks=multi_gpu_marks(num_gpus=2),
775
        **COMMON_BROADCAST_SETTINGS,  # type: ignore
776
    ),
777
    "llava-broadcast": VLMTestInfo(
778
779
780
        models=["llava-hf/llava-1.5-7b-hf"],
        prompt_formatter=lambda img_prompt: f"USER: {img_prompt}\nASSISTANT:",
        max_model_len=4096,
781
        auto_cls=AutoModelForImageTextToText,
782
        vllm_output_post_proc=model_utils.llava_image_vllm_to_hf_output,
783
        marks=multi_gpu_marks(num_gpus=2),
784
        **COMMON_BROADCAST_SETTINGS,  # type: ignore
785
    ),
786
    "llava_next-broadcast": VLMTestInfo(
787
788
789
        models=["llava-hf/llava-v1.6-mistral-7b-hf"],
        prompt_formatter=lambda img_prompt: f"[INST] {img_prompt} [/INST]",
        max_model_len=10240,
790
        auto_cls=AutoModelForImageTextToText,
791
        vllm_output_post_proc=model_utils.llava_image_vllm_to_hf_output,
792
        marks=multi_gpu_marks(num_gpus=2),
793
        **COMMON_BROADCAST_SETTINGS,  # type: ignore
794
795
796
797
    ),
    ### Custom input edge-cases for specific models
    "intern_vl-diff-patches": VLMTestInfo(
        models=["OpenGVLab/InternVL2-2B"],
798
        prompt_formatter=lambda img_prompt: f"<|im_start|>User\n{img_prompt}<|im_end|>\n<|im_start|>Assistant\n",  # noqa: E501
799
800
801
802
803
804
805
806
        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},
807
808
            )
            for inp in custom_inputs.different_patch_input_cases_internvl()
809
810
        ],
    ),
811
    "llava_onevision-multiple-images": VLMTestInfo(
812
813
814
815
        models=["llava-hf/llava-onevision-qwen2-0.5b-ov-hf"],
        test_type=VLMTestType.CUSTOM_INPUTS,
        max_model_len=16384,
        max_num_seqs=2,
816
        auto_cls=AutoModelForImageTextToText,
817
818
        hf_model_kwargs=model_utils.llava_onevision_hf_model_kwargs(
            "llava-hf/llava-onevision-qwen2-0.5b-ov-hf"
819
        ),
820
        vllm_output_post_proc=model_utils.llava_onevision_vllm_to_hf_output,
821
822
823
824
825
826
827
828
        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},
            )
        ],
829
    ),
830
831
832
833
834
835
    # 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,
836
        auto_cls=AutoModelForImageTextToText,
837
        vllm_output_post_proc=model_utils.qwen2_vllm_to_hf_output,
838
839
840
841
842
843
        custom_test_opts=[
            CustomTestOptions(
                inputs=custom_inputs.windows_attention_image_qwen2_5_vl(),
                limit_mm_per_prompt={"image": 1},
            )
        ],
844
    ),
845
846
847
}


848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
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):
866
        models_in_group = models[i * split_size : (i + 1) * split_size]
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882

        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)


883
884
885
886
887
888
### Test wrappers
# Wrappers around the core test running func for:
# - single image
# - multi-image
# - image embeddings
# - video
889
# - audio
890
# - custom inputs
891
892
893
894
895
@pytest.mark.parametrize(
    "model_type,test_case",
    get_parametrized_options(
        VLM_TEST_SETTINGS,
        test_type=VLMTestType.IMAGE,
896
        create_new_process_for_each_test=False,
897
898
    ),
)
899
900
901
902
903
904
905
906
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,
):
907
908
909
910
911
912
913
914
915
916
917
    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,
    )


918
919
920
921
922
@pytest.mark.parametrize(
    "model_type,test_case",
    get_parametrized_options(
        VLM_TEST_SETTINGS,
        test_type=VLMTestType.MULTI_IMAGE,
923
        create_new_process_for_each_test=False,
924
925
    ),
)
926
927
928
929
930
931
932
933
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,
):
934
935
936
937
938
939
940
941
942
943
944
    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,
    )


945
946
947
948
949
@pytest.mark.parametrize(
    "model_type,test_case",
    get_parametrized_options(
        VLM_TEST_SETTINGS,
        test_type=VLMTestType.EMBEDDING,
950
        create_new_process_for_each_test=False,
951
952
    ),
)
953
954
955
956
957
958
959
def test_image_embedding_models(
    model_type: str,
    test_case: ExpandableVLMTestArgs,
    hf_runner: type[HfRunner],
    vllm_runner: type[VllmRunner],
    image_assets: ImageTestAssets,
):
960
961
962
963
964
965
966
967
968
969
    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,
    )


970
971
972
973
974
@pytest.mark.parametrize(
    "model_type,test_case",
    get_parametrized_options(
        VLM_TEST_SETTINGS,
        test_type=VLMTestType.VIDEO,
975
        create_new_process_for_each_test=False,
976
977
    ),
)
978
979
980
981
982
983
984
def test_video_models(
    model_type: str,
    test_case: ExpandableVLMTestArgs,
    hf_runner: type[HfRunner],
    vllm_runner: type[VllmRunner],
    video_assets: VideoTestAssets,
):
985
986
987
988
989
990
991
992
993
994
    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,
    )


995
996
997
998
999
1000
@pytest.mark.parametrize(
    "model_type,test_case",
    get_parametrized_options(
        VLM_TEST_SETTINGS,
        test_type=VLMTestType.AUDIO,
        create_new_process_for_each_test=False,
1001
1002
    ),
)
1003
1004
1005
1006
1007
1008
1009
def test_audio_models(
    model_type: str,
    test_case: ExpandableVLMTestArgs,
    hf_runner: type[HfRunner],
    vllm_runner: type[VllmRunner],
    audio_assets: AudioTestAssets,
):
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
    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,
    )


1020
1021
1022
1023
1024
@pytest.mark.parametrize(
    "model_type,test_case",
    get_parametrized_options(
        VLM_TEST_SETTINGS,
        test_type=VLMTestType.CUSTOM_INPUTS,
1025
        create_new_process_for_each_test=False,
1026
1027
    ),
)
1028
1029
1030
def test_custom_inputs_models(
    model_type: str,
    test_case: ExpandableVLMTestArgs,
1031
1032
    hf_runner: type[HfRunner],
    vllm_runner: type[VllmRunner],
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
):
    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
1044
1045
1046
1047
1048
@pytest.mark.parametrize(
    "model_type,test_case",
    get_parametrized_options(
        VLM_TEST_SETTINGS,
        test_type=VLMTestType.IMAGE,
1049
        create_new_process_for_each_test=True,
1050
1051
    ),
)
1052
@create_new_process_for_each_test()
1053
1054
1055
1056
1057
1058
1059
1060
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,
):
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
    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,
    )


1072
1073
1074
1075
1076
@pytest.mark.parametrize(
    "model_type,test_case",
    get_parametrized_options(
        VLM_TEST_SETTINGS,
        test_type=VLMTestType.MULTI_IMAGE,
1077
        create_new_process_for_each_test=True,
1078
1079
    ),
)
1080
@create_new_process_for_each_test()
1081
1082
1083
1084
1085
1086
1087
1088
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,
):
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
    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,
    )


1100
1101
1102
1103
1104
@pytest.mark.parametrize(
    "model_type,test_case",
    get_parametrized_options(
        VLM_TEST_SETTINGS,
        test_type=VLMTestType.EMBEDDING,
1105
        create_new_process_for_each_test=True,
1106
1107
    ),
)
1108
@create_new_process_for_each_test()
1109
1110
1111
1112
1113
1114
1115
def test_image_embedding_models_heavy(
    model_type: str,
    test_case: ExpandableVLMTestArgs,
    hf_runner: type[HfRunner],
    vllm_runner: type[VllmRunner],
    image_assets: ImageTestAssets,
):
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
    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,
    )


1126
1127
1128
1129
1130
@pytest.mark.parametrize(
    "model_type,test_case",
    get_parametrized_options(
        VLM_TEST_SETTINGS,
        test_type=VLMTestType.VIDEO,
1131
        create_new_process_for_each_test=True,
1132
1133
    ),
)
1134
1135
1136
1137
1138
1139
1140
def test_video_models_heavy(
    model_type: str,
    test_case: ExpandableVLMTestArgs,
    hf_runner: type[HfRunner],
    vllm_runner: type[VllmRunner],
    video_assets: VideoTestAssets,
):
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
    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,
    )


1151
1152
1153
1154
1155
1156
@pytest.mark.parametrize(
    "model_type,test_case",
    get_parametrized_options(
        VLM_TEST_SETTINGS,
        test_type=VLMTestType.AUDIO,
        create_new_process_for_each_test=True,
1157
1158
    ),
)
1159
1160
1161
1162
1163
1164
1165
def test_audio_models_heavy(
    model_type: str,
    test_case: ExpandableVLMTestArgs,
    hf_runner: type[HfRunner],
    vllm_runner: type[VllmRunner],
    audio_assets: AudioTestAssets,
):
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
    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,
    )


1176
1177
1178
1179
1180
@pytest.mark.parametrize(
    "model_type,test_case",
    get_parametrized_options(
        VLM_TEST_SETTINGS,
        test_type=VLMTestType.CUSTOM_INPUTS,
1181
        create_new_process_for_each_test=True,
1182
1183
    ),
)
1184
@create_new_process_for_each_test()
1185
1186
1187
def test_custom_inputs_models_heavy(
    model_type: str,
    test_case: ExpandableVLMTestArgs,
1188
1189
    hf_runner: type[HfRunner],
    vllm_runner: type[VllmRunner],
1190
1191
1192
1193
1194
1195
1196
1197
):
    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,
    )