test_common.py 47.9 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.func_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
        vllm_runner_kwargs={"enable_mm_embeds": True},
113
        marks=[pytest.mark.core_model, pytest.mark.cpu_model],
114
    ),
Roger Wang's avatar
Roger Wang committed
115
116
    "qwen2_5_vl": VLMTestInfo(
        models=["Qwen/Qwen2.5-VL-3B-Instruct"],
117
118
        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
119
120
        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
121
122
        max_model_len=4096,
        max_num_seqs=2,
123
        auto_cls=AutoModelForImageTextToText,
Roger Wang's avatar
Roger Wang committed
124
125
        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)],
126
        marks=[pytest.mark.core_model, pytest.mark.cpu_model],
Roger Wang's avatar
Roger Wang committed
127
    ),
128
    "qwen2_5_omni": VLMTestInfo(
129
        models=["Qwen/Qwen2.5-Omni-3B"],
130
131
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE, VLMTestType.VIDEO),
        prompt_formatter=lambda img_prompt: f"<|im_start|>User\n{img_prompt}<|im_end|>\n<|im_start|>assistant\n",  # noqa: E501
132
133
        img_idx_to_prompt=lambda idx: "<|vision_bos|><|IMAGE|><|vision_eos|>",
        video_idx_to_prompt=lambda idx: "<|vision_bos|><|VIDEO|><|vision_eos|>",
134
135
        max_model_len=4096,
        max_num_seqs=2,
136
        num_logprobs=6 if current_platform.is_cpu() else 5,
137
        auto_cls=AutoModelForTextToWaveform,
138
        vllm_output_post_proc=model_utils.qwen2_vllm_to_hf_output,
139
        patch_hf_runner=model_utils.qwen2_5_omni_patch_hf_runner,
140
141
142
        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],
    ),
143
    "ultravox": VLMTestInfo(
144
        models=["fixie-ai/ultravox-v0_5-llama-3_2-1b"],
145
        test_type=VLMTestType.AUDIO,
146
        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
147
148
149
150
151
152
153
        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],
    ),
154
155
156
157
158
159
    #### 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,
160
        prompt_formatter=lambda vid_prompt: f"<|im_start|>user\n{vid_prompt}<|im_end|>\n<|im_start|>assistant\n",  # noqa: E501
161
        max_model_len=16384,
162
163
        hf_model_kwargs=model_utils.llava_onevision_hf_model_kwargs(
            "llava-hf/llava-onevision-qwen2-0.5b-ov-hf"
164
        ),
165
166
167
168
169
        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",
170
            "default_torch_num_threads": 1,
171
        },
172
173
174
        # FIXME: Investigate why the test hangs
        # when processing the 3rd prompt in vLLM
        marks=[pytest.mark.core_model, pytest.mark.skip(reason="Test hangs")],
175
    ),
176
177
178
    # Gemma3 has bidirectional mask on images
    "gemma3-transformers": VLMTestInfo(
        models=["google/gemma-3-4b-it"],
179
180
181
182
183
184
185
186
187
188
        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=8192,
189
        auto_cls=AutoModelForImageTextToText,
190
191
        # TODO: Support `do_pan_and_scan` in transformers backend
        # patch_hf_runner=model_utils.gemma3_patch_hf_runner,
192
193
194
195
        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",
196
            # "mm_processor_kwargs": {"do_pan_and_scan": True},
197
198
199
        },
        marks=[pytest.mark.core_model],
    ),
200
201
202
    "idefics3-transformers": VLMTestInfo(
        models=["HuggingFaceTB/SmolVLM-256M-Instruct"],
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
203
        prompt_formatter=lambda img_prompt: f"<|begin_of_text|>User:{img_prompt}<end_of_utterance>\nAssistant:",  # noqa: E501
204
205
206
207
208
209
210
211
212
213
214
        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],
    ),
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
    # PaliGemma has PrefixLM attention
    "paligemma-transformers": 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,
        image_size_factors=[(0.25, 0.5, 1.0)],
        vllm_runner_kwargs={
            "model_impl": "transformers",
        },
        marks=[pytest.mark.core_model],
    ),
236
237
238
239
    # 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,
240
        prompt_formatter=lambda img_prompt: f"<|im_start|>User\n{img_prompt}<|im_end|>\n<|im_start|>assistant\n",  # noqa: E501
241
        img_idx_to_prompt=lambda idx: "<|vision_start|><|image_pad|><|vision_end|>",
242
243
244
245
246
247
248
249
        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",
        },
250
        marks=[large_gpu_mark(min_gb=32)],
251
    ),
252
    #### Extended model tests
253
254
255
    "aria": VLMTestInfo(
        models=["rhymes-ai/Aria"],
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
256
        prompt_formatter=lambda img_prompt: f"<|im_start|>user\n{img_prompt}<|im_end|>\n<|im_start|>assistant\n ",  # noqa: E501
257
258
259
260
        img_idx_to_prompt=lambda idx: "<fim_prefix><|img|><fim_suffix>\n",
        max_model_len=4096,
        max_num_seqs=2,
        auto_cls=AutoModelForImageTextToText,
261
262
263
        single_image_prompts=IMAGE_ASSETS.prompts(
            {
                "stop_sign": "<vlm_image>Please describe the image shortly.",
264
                "cherry_blossom": "<vlm_image>Please infer the season with reason.",
265
266
            }
        ),
267
        multi_image_prompt="<vlm_image><vlm_image>Describe the two images shortly.",
268
269
270
271
272
        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
273
274
    "aya_vision": VLMTestInfo(
        models=["CohereForAI/aya-vision-8b"],
275
        test_type=(VLMTestType.IMAGE),
276
277
278
        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(
            {
279
280
                "stop_sign": "<image>What's the content in the center of the image?",
                "cherry_blossom": "<image>What is the season?",
281
282
            }
        ),
283
        multi_image_prompt="<image><image>Describe the two images in detail.",
284
285
286
287
288
289
290
291
        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),
292
293
294
        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(
            {
295
296
                "stop_sign": "<image>What's the content in the center of the image?",
                "cherry_blossom": "<image>What is the season?",
297
298
            }
        ),
299
        multi_image_prompt="<image><image>Describe the two images in detail.",
300
        max_model_len=4096,
Jennifer Zhao's avatar
Jennifer Zhao committed
301
302
        max_num_seqs=2,
        auto_cls=AutoModelForImageTextToText,
303
304
        vllm_runner_kwargs={"mm_processor_kwargs": {"crop_to_patches": True}},
        marks=[large_gpu_mark(min_gb=32)],
Jennifer Zhao's avatar
Jennifer Zhao committed
305
    ),
306
    "blip2": VLMTestInfo(
307
        models=["Salesforce/blip2-opt-2.7b"],
308
309
310
        test_type=VLMTestType.IMAGE,
        prompt_formatter=lambda img_prompt: f"Question: {img_prompt} Answer:",
        img_idx_to_prompt=lambda idx: "",
311
        auto_cls=AutoModelForImageTextToText,
312
        vllm_output_post_proc=model_utils.blip2_vllm_to_hf_output,
313
314
        # FIXME: https://github.com/huggingface/transformers/pull/38510
        marks=[pytest.mark.skip("Model is broken")],
315
316
317
318
319
320
    ),
    "chameleon": VLMTestInfo(
        models=["facebook/chameleon-7b"],
        test_type=VLMTestType.IMAGE,
        prompt_formatter=lambda img_prompt: f"USER: {img_prompt}\nASSISTANT:",
        max_model_len=4096,
321
        max_num_seqs=2,
322
        auto_cls=AutoModelForImageTextToText,
323
        # For chameleon, we only compare the sequences
324
325
        vllm_output_post_proc=lambda vllm_output, model: vllm_output[:2],
        hf_output_post_proc=lambda hf_output, model: hf_output[:2],
326
327
328
329
        comparator=check_outputs_equal,
        max_tokens=8,
        dtype="bfloat16",
    ),
330
    "deepseek_vl_v2": VLMTestInfo(
331
        models=["Isotr0py/deepseek-vl2-tiny"],  # model repo using dynamic module
332
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
333
        prompt_formatter=lambda img_prompt: f"<|User|>: {img_prompt}\n\n<|Assistant|>: ",  # noqa: E501
334
335
        max_model_len=4096,
        max_num_seqs=2,
336
337
        single_image_prompts=IMAGE_ASSETS.prompts(
            {
338
                "stop_sign": "<image>\nWhat's the content in the center of the image?",
339
340
341
342
                "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
343
344
        patch_hf_runner=model_utils.deepseekvl2_patch_hf_runner,
        hf_output_post_proc=model_utils.deepseekvl2_trunc_hf_output,
345
        stop_str=["<|end▁of▁sentence|>", "<|begin▁of▁sentence|>"],
346
        image_size_factors=[(), (1.0,), (1.0, 1.0, 1.0), (0.1, 0.5, 1.0)],
347
    ),
348
349
350
351
352
353
354
    "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,
355
        auto_cls=AutoModelForImageTextToText,
356
357
358
359
        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)],
360
        marks=[large_gpu_mark(min_gb=32)],
361
    ),
362
    "glm4v": VLMTestInfo(
363
        models=["zai-org/glm-4v-9b"],
364
        test_type=VLMTestType.IMAGE,
365
        prompt_formatter=lambda img_prompt: f"<|user|>\n{img_prompt}<|assistant|>",
366
367
368
369
370
371
        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
            }
        ),
372
373
374
        max_model_len=2048,
        max_num_seqs=2,
        get_stop_token_ids=lambda tok: [151329, 151336, 151338],
375
376
377
378
379
380
        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,
381
        marks=[large_gpu_mark(min_gb=32)],
382
    ),
383
    "glm4_1v": VLMTestInfo(
384
        models=["zai-org/GLM-4.1V-9B-Thinking"],
385
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
386
387
388
        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|>",
389
390
391
392
393
394
        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,
395
        marks=[large_gpu_mark(min_gb=32)],
396
397
    ),
    "glm4_1v-video": VLMTestInfo(
398
        models=["zai-org/GLM-4.1V-9B-Thinking"],
399
400
401
402
403
404
        # 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,
405
406
407
408
409
410
        custom_test_opts=[
            CustomTestOptions(
                inputs=custom_inputs.video_with_metadata_glm4_1v(),
                limit_mm_per_prompt={"video": 1},
            )
        ],
411
        marks=[large_gpu_mark(min_gb=32)],
412
    ),
413
    "h2ovl": VLMTestInfo(
414
        models=[
415
            "h2oai/h2ovl-mississippi-800m",
416
            "h2oai/h2ovl-mississippi-2b",
417
418
        ],
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
419
        prompt_formatter=lambda img_prompt: f"<|prompt|>{img_prompt}<|end|><|answer|>",
420
421
        single_image_prompts=IMAGE_ASSETS.prompts(
            {
422
                "stop_sign": "<image>\nWhat's the content in the center of the image?",
423
424
425
                "cherry_blossom": "<image>\nWhat is the season?",
            }
        ),
426
427
428
        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,
429
        num_logprobs=10,
430
        patch_hf_runner=model_utils.h2ovl_patch_hf_runner,
431
    ),
432
    "idefics3": VLMTestInfo(
433
        models=["HuggingFaceTB/SmolVLM-256M-Instruct"],
434
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
435
        prompt_formatter=lambda img_prompt: f"<|begin_of_text|>User:{img_prompt}<end_of_utterance>\nAssistant:",  # noqa: E501
436
437
438
        img_idx_to_prompt=lambda idx: "<image>",
        max_model_len=8192,
        max_num_seqs=2,
439
        auto_cls=AutoModelForImageTextToText,
440
        hf_output_post_proc=model_utils.idefics3_trunc_hf_output,
441
    ),
442
443
444
445
    "intern_vl": VLMTestInfo(
        models=[
            "OpenGVLab/InternVL2-1B",
            "OpenGVLab/InternVL2-2B",
446
447
            # FIXME: Config cannot be loaded in transformers 4.52
            # "OpenGVLab/Mono-InternVL-2B",
448
449
        ],
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
450
451
452
        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(
            {
453
                "stop_sign": "<image>\nWhat's the content in the center of the image?",
454
455
456
                "cherry_blossom": "<image>\nWhat is the season?",
            }
        ),
457
458
459
460
461
        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,
    ),
462
463
464
465
466
    "intern_vl-video": VLMTestInfo(
        models=[
            "OpenGVLab/InternVL3-1B",
        ],
        test_type=VLMTestType.VIDEO,
467
        prompt_formatter=lambda img_prompt: f"<|im_start|>User\n{img_prompt}<|im_end|>\n<|im_start|>Assistant\n",  # noqa: E501
468
469
470
471
472
        video_idx_to_prompt=lambda idx: "<video>",
        max_model_len=8192,
        use_tokenizer_eos=True,
        patch_hf_runner=model_utils.internvl_patch_hf_runner,
    ),
473
474
475
476
477
478
479
    "intern_vl-hf": VLMTestInfo(
        models=["OpenGVLab/InternVL3-1B-hf"],
        test_type=(
            VLMTestType.IMAGE,
            VLMTestType.MULTI_IMAGE,
            VLMTestType.VIDEO,
        ),
480
        prompt_formatter=lambda img_prompt: f"<|im_start|>User\n{img_prompt}<|im_end|>\n<|im_start|>Assistant\n",  # noqa: E501
481
482
483
484
485
486
        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,
    ),
487
488
489
    "kimi_vl": VLMTestInfo(
        models=["moonshotai/Kimi-VL-A3B-Instruct"],
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
490
        prompt_formatter=lambda img_prompt: f"<|im_user|>user<|im_middle|>{img_prompt}<|im_end|><|im_assistant|>assistant<|im_middle|>",  # noqa: E501
491
492
493
494
495
496
497
498
        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)],
    ),
499
500
    "llama4": VLMTestInfo(
        models=["meta-llama/Llama-4-Scout-17B-16E-Instruct"],
501
        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
502
503
504
        img_idx_to_prompt=lambda _: "<|image|>",
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
        distributed_executor_backend="mp",
505
        image_size_factors=[(0.25, 0.5, 1.0)],
506
507
508
509
510
        hf_model_kwargs={"device_map": "auto"},
        max_model_len=8192,
        max_num_seqs=4,
        dtype="bfloat16",
        auto_cls=AutoModelForImageTextToText,
511
512
        tensor_parallel_size=4,
        marks=multi_gpu_marks(num_gpus=4),
513
    ),
514
515
516
517
518
    "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,
519
        auto_cls=AutoModelForImageTextToText,
520
        vllm_output_post_proc=model_utils.llava_image_vllm_to_hf_output,
521
522
523
524
525
526
527
528
        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},
            )
        ],
529
    ),
530
    "llava_onevision": VLMTestInfo(
531
532
        models=["llava-hf/llava-onevision-qwen2-0.5b-ov-hf"],
        test_type=VLMTestType.CUSTOM_INPUTS,
533
        prompt_formatter=lambda vid_prompt: f"<|im_start|>user\n{vid_prompt}<|im_end|>\n<|im_start|>assistant\n",  # noqa: E501
534
535
        num_video_frames=16,
        max_model_len=16384,
536
537
        hf_model_kwargs=model_utils.llava_onevision_hf_model_kwargs(
            "llava-hf/llava-onevision-qwen2-0.5b-ov-hf"
538
        ),
539
        auto_cls=AutoModelForImageTextToText,
540
        vllm_output_post_proc=model_utils.llava_onevision_vllm_to_hf_output,
541
542
543
544
545
546
547
548
        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},
            )
        ],
549
550
551
552
553
554
555
    ),
    "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,
556
        max_num_seqs=2,
557
        auto_cls=AutoModelForImageTextToText,
558
559
        vllm_output_post_proc=model_utils.llava_video_vllm_to_hf_output,
    ),
560
561
562
563
564
565
    "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],
566
        auto_cls=AutoModelForImageTextToText,
567
568
569
        vllm_output_post_proc=model_utils.mantis_vllm_to_hf_output,
        patch_hf_runner=model_utils.mantis_patch_hf_runner,
    ),
570
    "minicpmv_25": VLMTestInfo(
571
        models=["openbmb/MiniCPM-Llama3-V-2_5"],
572
        test_type=VLMTestType.IMAGE,
573
574
575
576
577
        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],
578
        hf_output_post_proc=model_utils.minicpmv_trunc_hf_output,
579
        patch_hf_runner=model_utils.minicpmv_25_patch_hf_runner,
580
581
        # FIXME: https://huggingface.co/openbmb/MiniCPM-V-2_6/discussions/55
        marks=[pytest.mark.skip("HF import fails")],
582
    ),
583
584
    "minicpmo_26": VLMTestInfo(
        models=["openbmb/MiniCPM-o-2_6"],
585
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
586
587
588
589
        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,
590
591
        get_stop_token_ids=lambda tok: tok.convert_tokens_to_ids(
            ["<|im_end|>", "<|endoftext|>"]
592
        ),
593
594
        hf_output_post_proc=model_utils.minicpmv_trunc_hf_output,
        patch_hf_runner=model_utils.minicpmo_26_patch_hf_runner,
595
        # FIXME: https://huggingface.co/openbmb/MiniCPM-o-2_6/discussions/49
596
        marks=[pytest.mark.skip("HF import fails")],
597
    ),
598
599
    "minicpmv_26": VLMTestInfo(
        models=["openbmb/MiniCPM-V-2_6"],
600
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
601
602
603
604
        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,
605
606
        get_stop_token_ids=lambda tok: tok.convert_tokens_to_ids(
            ["<|im_end|>", "<|endoftext|>"]
607
        ),
608
609
610
        hf_output_post_proc=model_utils.minicpmv_trunc_hf_output,
        patch_hf_runner=model_utils.minicpmv_26_patch_hf_runner,
    ),
611
612
    "minimax_vl_01": VLMTestInfo(
        models=["MiniMaxAI/MiniMax-VL-01"],
613
        prompt_formatter=lambda img_prompt: f"<beginning_of_sentence>user: {img_prompt} assistant:<end_of_sentence>",  # noqa: E501
614
615
616
617
618
619
620
621
622
623
        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)],
    ),
624
625
    "molmo": VLMTestInfo(
        models=["allenai/Molmo-7B-D-0924"],
626
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
627
        prompt_formatter=identity,
628
629
        max_model_len=4096,
        max_num_seqs=2,
630
        patch_hf_runner=model_utils.molmo_patch_hf_runner,
631
    ),
632
633
634
    "ovis1_6-gemma2": VLMTestInfo(
        models=["AIDC-AI/Ovis1.6-Gemma2-9B"],
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
635
        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
636
        img_idx_to_prompt=lambda idx: "<image>\n",
637
638
639
640
641
642
643
644
        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)],
    ),
645
646
647
    "ovis2": VLMTestInfo(
        models=["AIDC-AI/Ovis2-1B"],
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
648
        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
649
        img_idx_to_prompt=lambda idx: "<image>\n",
650
651
652
653
654
        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"},
655
        patch_hf_runner=model_utils.ovis_patch_hf_runner,
656
    ),
657
658
    "ovis2_5": VLMTestInfo(
        models=["AIDC-AI/Ovis2.5-2B"],
659
660
        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
661
        img_idx_to_prompt=lambda idx: "<image>\n",
662
663
664
665
666
667
        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,
668
        hf_model_kwargs={"revision": "refs/pr/5"},
669
    ),
670
671
672
    "phi3v": VLMTestInfo(
        models=["microsoft/Phi-3.5-vision-instruct"],
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
673
        prompt_formatter=lambda img_prompt: f"<|user|>\n{img_prompt}<|end|>\n<|assistant|>\n",  # noqa: E501
674
675
676
        img_idx_to_prompt=lambda idx: f"<|image_{idx}|>\n",
        max_model_len=4096,
        max_num_seqs=2,
677
        runner="generate",
678
679
        # use sdpa mode for hf runner since phi3v didn't work with flash_attn
        hf_model_kwargs={"_attn_implementation": "sdpa"},
680
681
682
683
        use_tokenizer_eos=True,
        vllm_output_post_proc=model_utils.phi3v_vllm_to_hf_output,
        num_logprobs=10,
    ),
684
685
686
687
688
689
690
    "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,
691
        auto_cls=AutoModelForImageTextToText,
692
        marks=[large_gpu_mark(min_gb=48)],
693
    ),
694
    "qwen_vl": VLMTestInfo(
695
696
697
698
699
700
701
702
703
        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,
    ),
704
705
    "qwen2_vl": VLMTestInfo(
        models=["Qwen/Qwen2-VL-2B-Instruct"],
706
707
        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
708
709
        img_idx_to_prompt=lambda idx: "<|vision_start|><|image_pad|><|vision_end|>",
        video_idx_to_prompt=lambda idx: "<|vision_start|><|video_pad|><|vision_end|>",
710
        multi_image_prompt="Picture 1: <vlm_image>\nPicture 2: <vlm_image>\nDescribe these two images with one paragraph respectively.",  # noqa: E501
711
712
        max_model_len=4096,
        max_num_seqs=2,
713
        auto_cls=AutoModelForImageTextToText,
714
715
716
717
        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],
    ),
718
719
720
    "skywork_r1v": VLMTestInfo(
        models=["Skywork/Skywork-R1V-38B"],
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
721
722
723
        prompt_formatter=lambda img_prompt: f"<|begin▁of▁sentence|><|User|>\n{img_prompt}<|Assistant|><think>\n",  # noqa: E501
        single_image_prompts=IMAGE_ASSETS.prompts(
            {
724
                "stop_sign": "<image>\nWhat's the content in the center of the image?",
725
726
727
                "cherry_blossom": "<image>\nWhat is the season?",
            }
        ),
728
        multi_image_prompt="<image>\n<image>\nDescribe the two images in short.",
729
730
731
732
733
        max_model_len=4096,
        use_tokenizer_eos=True,
        patch_hf_runner=model_utils.skyworkr1v_patch_hf_runner,
        marks=[large_gpu_mark(min_gb=80)],
    ),
734
735
736
    "smolvlm": VLMTestInfo(
        models=["HuggingFaceTB/SmolVLM2-2.2B-Instruct"],
        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
737
        prompt_formatter=lambda img_prompt: f"<|im_start|>User:{img_prompt}<end_of_utterance>\nAssistant:",  # noqa: E501
738
739
740
741
742
        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,
743
        num_logprobs=10,
744
    ),
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
    "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,
        ),
761
        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
762
763
        img_idx_to_prompt=lambda idx: "<|vision_start|><|image_pad|><|vision_end|>",
        video_idx_to_prompt=lambda idx: "<|vision_start|><|video_pad|><|vision_end|>",
764
765
766
767
768
769
        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")],
    ),
770
    ### Tensor parallel / multi-gpu broadcast tests
771
    "chameleon-broadcast": VLMTestInfo(
772
773
774
        models=["facebook/chameleon-7b"],
        prompt_formatter=lambda img_prompt: f"USER: {img_prompt}\nASSISTANT:",
        max_model_len=4096,
775
        auto_cls=AutoModelForImageTextToText,
776
777
        vllm_output_post_proc=lambda vllm_output, model: vllm_output[:2],
        hf_output_post_proc=lambda hf_output, model: hf_output[:2],
778
        comparator=check_outputs_equal,
779
        marks=multi_gpu_marks(num_gpus=2),
780
        **COMMON_BROADCAST_SETTINGS,  # type: ignore
781
    ),
782
    "llava-broadcast": VLMTestInfo(
783
784
785
        models=["llava-hf/llava-1.5-7b-hf"],
        prompt_formatter=lambda img_prompt: f"USER: {img_prompt}\nASSISTANT:",
        max_model_len=4096,
786
        auto_cls=AutoModelForImageTextToText,
787
        vllm_output_post_proc=model_utils.llava_image_vllm_to_hf_output,
788
        marks=multi_gpu_marks(num_gpus=2),
789
        **COMMON_BROADCAST_SETTINGS,  # type: ignore
790
    ),
791
    "llava_next-broadcast": VLMTestInfo(
792
793
794
        models=["llava-hf/llava-v1.6-mistral-7b-hf"],
        prompt_formatter=lambda img_prompt: f"[INST] {img_prompt} [/INST]",
        max_model_len=10240,
795
        auto_cls=AutoModelForImageTextToText,
796
        vllm_output_post_proc=model_utils.llava_image_vllm_to_hf_output,
797
        marks=multi_gpu_marks(num_gpus=2),
798
        **COMMON_BROADCAST_SETTINGS,  # type: ignore
799
800
801
802
    ),
    ### Custom input edge-cases for specific models
    "intern_vl-diff-patches": VLMTestInfo(
        models=["OpenGVLab/InternVL2-2B"],
803
        prompt_formatter=lambda img_prompt: f"<|im_start|>User\n{img_prompt}<|im_end|>\n<|im_start|>Assistant\n",  # noqa: E501
804
805
806
807
808
809
810
811
        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},
812
813
            )
            for inp in custom_inputs.different_patch_input_cases_internvl()
814
815
        ],
    ),
816
    "llava_onevision-multiple-images": VLMTestInfo(
817
818
819
820
        models=["llava-hf/llava-onevision-qwen2-0.5b-ov-hf"],
        test_type=VLMTestType.CUSTOM_INPUTS,
        max_model_len=16384,
        max_num_seqs=2,
821
        auto_cls=AutoModelForImageTextToText,
822
823
        hf_model_kwargs=model_utils.llava_onevision_hf_model_kwargs(
            "llava-hf/llava-onevision-qwen2-0.5b-ov-hf"
824
        ),
825
        vllm_output_post_proc=model_utils.llava_onevision_vllm_to_hf_output,
826
827
828
829
830
831
832
833
        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},
            )
        ],
834
    ),
835
836
837
838
839
840
    # 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,
841
        auto_cls=AutoModelForImageTextToText,
842
        vllm_output_post_proc=model_utils.qwen2_vllm_to_hf_output,
843
844
845
846
847
848
        custom_test_opts=[
            CustomTestOptions(
                inputs=custom_inputs.windows_attention_image_qwen2_5_vl(),
                limit_mm_per_prompt={"image": 1},
            )
        ],
849
    ),
850
851
852
}


853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
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):
871
        models_in_group = models[i * split_size : (i + 1) * split_size]
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887

        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)


888
889
890
891
892
893
### Test wrappers
# Wrappers around the core test running func for:
# - single image
# - multi-image
# - image embeddings
# - video
894
# - audio
895
# - custom inputs
896
897
898
899
900
@pytest.mark.parametrize(
    "model_type,test_case",
    get_parametrized_options(
        VLM_TEST_SETTINGS,
        test_type=VLMTestType.IMAGE,
901
        create_new_process_for_each_test=False,
902
903
    ),
)
904
905
906
907
908
909
910
911
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,
):
912
913
914
915
916
917
918
919
920
921
922
    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,
    )


923
924
925
926
927
@pytest.mark.parametrize(
    "model_type,test_case",
    get_parametrized_options(
        VLM_TEST_SETTINGS,
        test_type=VLMTestType.MULTI_IMAGE,
928
        create_new_process_for_each_test=False,
929
930
    ),
)
931
932
933
934
935
936
937
938
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,
):
939
940
941
942
943
944
945
946
947
948
949
    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,
    )


950
951
952
953
954
@pytest.mark.parametrize(
    "model_type,test_case",
    get_parametrized_options(
        VLM_TEST_SETTINGS,
        test_type=VLMTestType.EMBEDDING,
955
        create_new_process_for_each_test=False,
956
957
    ),
)
958
959
960
961
962
963
964
def test_image_embedding_models(
    model_type: str,
    test_case: ExpandableVLMTestArgs,
    hf_runner: type[HfRunner],
    vllm_runner: type[VllmRunner],
    image_assets: ImageTestAssets,
):
965
966
967
968
969
970
971
972
973
974
    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,
    )


975
976
977
978
979
@pytest.mark.parametrize(
    "model_type,test_case",
    get_parametrized_options(
        VLM_TEST_SETTINGS,
        test_type=VLMTestType.VIDEO,
980
        create_new_process_for_each_test=False,
981
982
    ),
)
983
984
985
986
987
988
989
def test_video_models(
    model_type: str,
    test_case: ExpandableVLMTestArgs,
    hf_runner: type[HfRunner],
    vllm_runner: type[VllmRunner],
    video_assets: VideoTestAssets,
):
990
991
992
993
994
995
996
997
998
999
    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,
    )


1000
1001
1002
1003
1004
1005
@pytest.mark.parametrize(
    "model_type,test_case",
    get_parametrized_options(
        VLM_TEST_SETTINGS,
        test_type=VLMTestType.AUDIO,
        create_new_process_for_each_test=False,
1006
1007
    ),
)
1008
1009
1010
1011
1012
1013
1014
def test_audio_models(
    model_type: str,
    test_case: ExpandableVLMTestArgs,
    hf_runner: type[HfRunner],
    vllm_runner: type[VllmRunner],
    audio_assets: AudioTestAssets,
):
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
    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,
    )


1025
1026
1027
1028
1029
@pytest.mark.parametrize(
    "model_type,test_case",
    get_parametrized_options(
        VLM_TEST_SETTINGS,
        test_type=VLMTestType.CUSTOM_INPUTS,
1030
        create_new_process_for_each_test=False,
1031
1032
    ),
)
1033
1034
1035
def test_custom_inputs_models(
    model_type: str,
    test_case: ExpandableVLMTestArgs,
1036
1037
    hf_runner: type[HfRunner],
    vllm_runner: type[VllmRunner],
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
):
    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
1049
1050
1051
1052
1053
@pytest.mark.parametrize(
    "model_type,test_case",
    get_parametrized_options(
        VLM_TEST_SETTINGS,
        test_type=VLMTestType.IMAGE,
1054
        create_new_process_for_each_test=True,
1055
1056
    ),
)
1057
@create_new_process_for_each_test()
1058
1059
1060
1061
1062
1063
1064
1065
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,
):
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
    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,
    )


1077
1078
1079
1080
1081
@pytest.mark.parametrize(
    "model_type,test_case",
    get_parametrized_options(
        VLM_TEST_SETTINGS,
        test_type=VLMTestType.MULTI_IMAGE,
1082
        create_new_process_for_each_test=True,
1083
1084
    ),
)
1085
@create_new_process_for_each_test()
1086
1087
1088
1089
1090
1091
1092
1093
def test_multi_image_models_heavy(
    tmp_path: PosixPath,
    model_type: str,
    test_case: ExpandableVLMTestArgs,
    hf_runner: type[HfRunner],
    vllm_runner: type[VllmRunner],
    image_assets: ImageTestAssets,
):
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
    model_test_info = VLM_TEST_SETTINGS[model_type]
    runners.run_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,
    )


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


1131
1132
1133
1134
1135
@pytest.mark.parametrize(
    "model_type,test_case",
    get_parametrized_options(
        VLM_TEST_SETTINGS,
        test_type=VLMTestType.VIDEO,
1136
        create_new_process_for_each_test=True,
1137
1138
    ),
)
1139
1140
1141
1142
1143
1144
1145
def test_video_models_heavy(
    model_type: str,
    test_case: ExpandableVLMTestArgs,
    hf_runner: type[HfRunner],
    vllm_runner: type[VllmRunner],
    video_assets: VideoTestAssets,
):
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
    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,
    )


1156
1157
1158
1159
1160
1161
@pytest.mark.parametrize(
    "model_type,test_case",
    get_parametrized_options(
        VLM_TEST_SETTINGS,
        test_type=VLMTestType.AUDIO,
        create_new_process_for_each_test=True,
1162
1163
    ),
)
1164
1165
1166
1167
1168
1169
1170
def test_audio_models_heavy(
    model_type: str,
    test_case: ExpandableVLMTestArgs,
    hf_runner: type[HfRunner],
    vllm_runner: type[VllmRunner],
    audio_assets: AudioTestAssets,
):
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
    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,
    )


1181
1182
1183
1184
1185
@pytest.mark.parametrize(
    "model_type,test_case",
    get_parametrized_options(
        VLM_TEST_SETTINGS,
        test_type=VLMTestType.CUSTOM_INPUTS,
1186
        create_new_process_for_each_test=True,
1187
1188
    ),
)
1189
@create_new_process_for_each_test()
1190
1191
1192
def test_custom_inputs_models_heavy(
    model_type: str,
    test_case: ExpandableVLMTestArgs,
1193
1194
    hf_runner: type[HfRunner],
    vllm_runner: type[VllmRunner],
1195
1196
1197
1198
1199
1200
1201
1202
):
    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,
    )