| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Common tests for testing .generate() functionality for single / multiple |
| 4 | image, embedding, and video support for different VLMs in vLLM. |
| 5 | """ |
| 6 | |
| 7 | import math |
| 8 | from collections import defaultdict |
| 9 | from pathlib import PosixPath |
| 10 | |
| 11 | import pytest |
| 12 | from packaging.version import Version |
| 13 | from transformers import ( |
| 14 | AutoModel, |
| 15 | AutoModelForCausalLM, |
| 16 | AutoModelForImageTextToText, |
| 17 | AutoModelForTextToWaveform, |
| 18 | ) |
| 19 | from transformers import __version__ as TRANSFORMERS_VERSION |
| 20 | |
| 21 | from vllm.platforms import current_platform |
| 22 | from vllm.utils.func_utils import identity |
| 23 | |
| 24 | from ....conftest import ( |
| 25 | IMAGE_ASSETS, |
| 26 | AudioTestAssets, |
| 27 | HfRunner, |
| 28 | ImageTestAssets, |
| 29 | VideoTestAssets, |
| 30 | VllmRunner, |
| 31 | ) |
| 32 | from ....utils import create_new_process_for_each_test, large_gpu_mark, multi_gpu_marks |
| 33 | from .vlm_utils import custom_inputs, model_utils, runners |
| 34 | from .vlm_utils.case_filtering import get_parametrized_options |
| 35 | from .vlm_utils.types import ( |
| 36 | CustomTestOptions, |
| 37 | ExpandableVLMTestArgs, |
| 38 | VLMTestInfo, |
| 39 | VLMTestType, |
| 40 | ) |
| 41 | |
| 42 | COMMON_BROADCAST_SETTINGS = { |
| 43 | "test_type": VLMTestType.IMAGE, |
| 44 | "dtype": "half", |
| 45 | "max_tokens": 5, |
| 46 | "tensor_parallel_size": 2, |
| 47 | "hf_model_kwargs": {"device_map": "auto"}, |
| 48 | "image_size_factors": [(0.25, 0.5, 1.0)], |
| 49 | "distributed_executor_backend": ( |
| 50 | "ray", |
| 51 | "mp", |
| 52 | ), |
| 53 | } |
| 54 | |
| 55 | ### Test configuration for specific models |
| 56 | # NOTE: The convention of the test settings below is to lead each test key |
| 57 | # with the name of the model arch used in the test, using underscores in place |
| 58 | # of hyphens; this makes it more convenient to filter tests for a specific kind |
| 59 | # of model. For example.... |
| 60 | # |
| 61 | # To run all test types for a specific key: |
| 62 | # use the k flag to substring match with a leading square bracket; if the |
| 63 | # model arch happens to be a substring of another one, you can add a |
| 64 | # trailing hyphen. E.g., |
| 65 | # - pytest $TEST_FILE -k "[llava-" |
| 66 | # prevents matching on "[llava_next-" & will match just the enabled cases |
| 67 | # for llava, i.e., single image, image embedding, and custom input tests. |
| 68 | # |
| 69 | # To run a test for a Test Info for just one of multiple models: |
| 70 | # use the k flag to substring match the model name, e.g., |
| 71 | # - pytest $TEST_FILE -k OpenGVLab/InternVL2-1B |
| 72 | # prevents matching on nGVLab/InternVL2-2B. |
| 73 | # |
| 74 | # You can also combine substrings to match more granularly. |
| 75 | # ex 1: |
| 76 | # pytest $TEST_FILE -k "test_single_image and OpenGVLab/InternVL2-1B" |
| 77 | # will run only test_single_image* for OpenGVLab/InternVL2-1B; this would |
| 78 | # match both wrappers for single image tests, since it also matches |
| 79 | # test_single_image_heavy (which forks if we have a distributed backend) |
| 80 | # ex 2: |
| 81 | # pytest $TEST_FILE -k "[llava- or [intern_vl-" |
| 82 | # will run all of the tests for only llava & internvl. |
| 83 | # |
| 84 | # NOTE you can add --collect-only to any of the above commands to see |
| 85 | # which cases would be selected and deselected by pytest. In general, |
| 86 | # this is a good idea for checking your command first, since tests are slow. |
| 87 | |
| 88 | |
| 89 | def _granite4_vision_vllm_to_hf_output(vllm_output, model): |
| 90 | """Post-processor for granite4_vision vLLM output. |
| 91 | |
| 92 | Self-contained to avoid calling AutoConfig/AutoTokenizer without |
| 93 | trust_remote_code (needed while the model is not in upstream HF). |
| 94 | """ |
| 95 | output_ids, output_str, out_logprobs = vllm_output |
| 96 | mm_token_id = 100352 |
| 97 | hf_output_ids = [ |
| 98 | token_id |
| 99 | for idx, token_id in enumerate(output_ids) |
| 100 | if token_id != mm_token_id or idx == 0 or output_ids[idx - 1] != mm_token_id |
| 101 | ] |
| 102 | hf_output_str = ( |
| 103 | output_str[1:] if output_str and output_str[0] == " " else output_str |
| 104 | ) |
| 105 | eos_token_id = 100257 |
| 106 | if hf_output_ids and hf_output_ids[-1] == eos_token_id: |
| 107 | hf_output_str = hf_output_str + "<|end_of_text|>" |
| 108 | return hf_output_ids, hf_output_str, out_logprobs |
| 109 | |
| 110 | |
| 111 | VLM_TEST_SETTINGS = { |
| 112 | #### Core tests to always run in the CI |
| 113 | "llava": VLMTestInfo( |
| 114 | models=["llava-hf/llava-1.5-7b-hf"], |
| 115 | test_type=(VLMTestType.EMBEDDING, VLMTestType.IMAGE, VLMTestType.CUSTOM_INPUTS), |
| 116 | prompt_formatter=lambda img_prompt: f"USER: {img_prompt}\nASSISTANT:", |
| 117 | convert_assets_to_embeddings=model_utils.get_llava_embeddings, |
| 118 | max_model_len=4096, |
| 119 | auto_cls=AutoModelForImageTextToText, |
| 120 | vllm_output_post_proc=model_utils.llava_image_vllm_to_hf_output, |
| 121 | custom_test_opts=[ |
| 122 | CustomTestOptions( |
| 123 | inputs=custom_inputs.multi_image_multi_aspect_ratio_inputs( |
| 124 | formatter=lambda img_prompt: f"USER: {img_prompt}\nASSISTANT:" |
| 125 | ), |
| 126 | limit_mm_per_prompt={"image": 4}, |
| 127 | ) |
| 128 | ], |
| 129 | vllm_runner_kwargs={"enable_mm_embeds": True}, |
| 130 | marks=[pytest.mark.core_model, pytest.mark.cpu_model], |
| 131 | ), |
| 132 | "paligemma": VLMTestInfo( |
| 133 | models=["google/paligemma-3b-mix-224"], |
| 134 | test_type=VLMTestType.IMAGE, |
| 135 | prompt_formatter=identity, |
| 136 | img_idx_to_prompt=lambda idx: "", |
| 137 | # Paligemma uses its own sample prompts because the default one fails |
| 138 | single_image_prompts=IMAGE_ASSETS.prompts( |
| 139 | { |
| 140 | "stop_sign": "caption es", |
| 141 | "cherry_blossom": "What is in the picture?", |
| 142 | } |
| 143 | ), |
| 144 | auto_cls=AutoModelForImageTextToText, |
| 145 | vllm_output_post_proc=model_utils.paligemma_vllm_to_hf_output, |
| 146 | ), |
| 147 | "qwen2_5_vl": VLMTestInfo( |
| 148 | models=["Qwen/Qwen2.5-VL-3B-Instruct"], |
| 149 | test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE, VLMTestType.VIDEO), |
| 150 | prompt_formatter=lambda img_prompt: f"<|im_start|>User\n{img_prompt}<|im_end|>\n<|im_start|>assistant\n", # noqa: E501 |
| 151 | img_idx_to_prompt=lambda idx: "<|vision_start|><|image_pad|><|vision_end|>", |
| 152 | video_idx_to_prompt=lambda idx: "<|vision_start|><|video_pad|><|vision_end|>", |
| 153 | enforce_eager=False, |
| 154 | needs_video_metadata=True, |
| 155 | max_model_len=4096, |
| 156 | max_num_seqs=2, |
| 157 | auto_cls=AutoModelForImageTextToText, |
| 158 | vllm_output_post_proc=model_utils.qwen2_vllm_to_hf_output, |
| 159 | patch_hf_runner=model_utils.qwen3_vl_patch_hf_runner, |
| 160 | image_size_factors=[(0.25,), (0.25, 0.25, 0.25), (0.25, 0.2, 0.15)], |
| 161 | marks=[pytest.mark.core_model, pytest.mark.cpu_model], |
| 162 | ), |
| 163 | "qwen2_5_omni": VLMTestInfo( |
| 164 | models=["Qwen/Qwen2.5-Omni-3B"], |
| 165 | test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE, VLMTestType.VIDEO), |
| 166 | prompt_formatter=lambda img_prompt: f"<|im_start|>User\n{img_prompt}<|im_end|>\n<|im_start|>assistant\n", # noqa: E501 |
| 167 | img_idx_to_prompt=lambda idx: "<|vision_bos|><|IMAGE|><|vision_eos|>", |
| 168 | video_idx_to_prompt=lambda idx: "<|vision_bos|><|VIDEO|><|vision_eos|>", |
| 169 | max_model_len=4096, |
| 170 | max_num_seqs=2, |
| 171 | num_logprobs=6 if current_platform.is_cpu() else 5, |
| 172 | auto_cls=AutoModelForTextToWaveform, |
| 173 | vllm_output_post_proc=model_utils.qwen2_vllm_to_hf_output, |
| 174 | patch_hf_runner=model_utils.qwen2_5_omni_patch_hf_runner, |
| 175 | image_size_factors=[(0.25,), (0.25, 0.25, 0.25), (0.25, 0.2, 0.15)], |
| 176 | marks=[pytest.mark.core_model, pytest.mark.cpu_model], |
| 177 | ), |
| 178 | "qwen3_vl": VLMTestInfo( |
| 179 | models=["Qwen/Qwen3-VL-4B-Instruct"], |
| 180 | test_type=( |
| 181 | VLMTestType.IMAGE, |
| 182 | VLMTestType.MULTI_IMAGE, |
| 183 | VLMTestType.VIDEO, |
| 184 | ), |
| 185 | enforce_eager=False, |
| 186 | needs_video_metadata=True, |
| 187 | prompt_formatter=lambda img_prompt: f"<|im_start|>User\n{img_prompt}<|im_end|>\n<|im_start|>assistant\n", # noqa: E501 |
| 188 | img_idx_to_prompt=lambda idx: "<|vision_start|><|image_pad|><|vision_end|>", # noqa: E501 |
| 189 | video_idx_to_prompt=lambda idx: "<|vision_start|><|video_pad|><|vision_end|>", # noqa: E501 |
| 190 | max_model_len=4096, |
| 191 | max_num_seqs=2, |
| 192 | num_logprobs=20, |
| 193 | auto_cls=AutoModelForImageTextToText, |
| 194 | vllm_output_post_proc=model_utils.qwen2_vllm_to_hf_output, |
| 195 | patch_hf_runner=model_utils.qwen3_vl_patch_hf_runner, |
| 196 | image_size_factors=[(0.25,), (0.25, 0.25, 0.25), (0.25, 0.2, 0.15)], |
| 197 | marks=[ |
| 198 | pytest.mark.core_model, |
| 199 | ], |
| 200 | vllm_runner_kwargs={"attention_backend": "TRITON_ATTN"} |
| 201 | if current_platform.is_rocm() |
| 202 | else {}, |
| 203 | ), |
| 204 | "ultravox": VLMTestInfo( |
| 205 | models=["fixie-ai/ultravox-v0_5-llama-3_2-1b"], |
| 206 | test_type=VLMTestType.AUDIO, |
| 207 | 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 |
| 208 | audio_idx_to_prompt=lambda idx: "<|audio|>", |
| 209 | max_model_len=4096, |
| 210 | max_num_seqs=2, |
| 211 | auto_cls=AutoModel, |
| 212 | hf_output_post_proc=model_utils.ultravox_trunc_hf_output, |
| 213 | marks=[ |
| 214 | pytest.mark.core_model, |
| 215 | pytest.mark.cpu_model, |
| 216 | # TODO: Remove skip once model has been upstreamed to Transformers |
| 217 | pytest.mark.skip( |
| 218 | reason="Custom model code is not compatible with Transformers v5" |
| 219 | ), |
| 220 | ], |
| 221 | ), |
| 222 | #### Transformers fallback to test |
| 223 | ## To reduce test burden, we only test batching arbitrary image size |
| 224 | # Dynamic image length and number of patches |
| 225 | "llava-onevision-transformers": VLMTestInfo( |
| 226 | models=["llava-hf/llava-onevision-qwen2-0.5b-ov-hf"], |
| 227 | test_type=VLMTestType.IMAGE, |
| 228 | prompt_formatter=lambda vid_prompt: f"<|im_start|>user\n{vid_prompt}<|im_end|>\n<|im_start|>assistant\n", # noqa: E501 |
| 229 | max_model_len=16384, |
| 230 | hf_model_kwargs=model_utils.llava_onevision_hf_model_kwargs( |
| 231 | "llava-hf/llava-onevision-qwen2-0.5b-ov-hf" |
| 232 | ), |
| 233 | auto_cls=AutoModelForImageTextToText, |
| 234 | vllm_output_post_proc=model_utils.llava_onevision_vllm_to_hf_output, |
| 235 | image_size_factors=[(0.25, 0.5, 1.0)], |
| 236 | vllm_runner_kwargs={ |
| 237 | "model_impl": "transformers", |
| 238 | }, |
| 239 | marks=[pytest.mark.core_model], |
| 240 | ), |
| 241 | # Gemma3 has bidirectional mask on images |
| 242 | "gemma3-transformers": VLMTestInfo( |
| 243 | models=["google/gemma-3-4b-it"], |
| 244 | test_type=VLMTestType.IMAGE, |
| 245 | prompt_formatter=lambda vid_prompt: f"<'<bos><start_of_turn>user\n{vid_prompt}<start_of_image><end_of_turn>\n<start_of_turn>model\n", # noqa: E501 |
| 246 | max_model_len=4096, |
| 247 | auto_cls=AutoModelForImageTextToText, |
| 248 | vllm_output_post_proc=model_utils.gemma3_vllm_to_hf_output, |
| 249 | image_size_factors=[(0.25, 0.5, 1.0)], |
| 250 | vllm_runner_kwargs={ |
| 251 | "model_impl": "transformers", |
| 252 | }, |
| 253 | marks=[pytest.mark.core_model], |
| 254 | ), |
| 255 | "idefics3-transformers": VLMTestInfo( |
| 256 | models=["HuggingFaceTB/SmolVLM-256M-Instruct"], |
| 257 | test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE), |
| 258 | prompt_formatter=lambda img_prompt: f"<|begin_of_text|>User:{img_prompt}<end_of_utterance>\nAssistant:", # noqa: E501 |
| 259 | img_idx_to_prompt=lambda idx: "<image>", |
| 260 | max_model_len=8192, |
| 261 | max_num_seqs=2, |
| 262 | auto_cls=AutoModelForImageTextToText, |
| 263 | hf_output_post_proc=model_utils.idefics3_trunc_hf_output, |
| 264 | image_size_factors=[(0.25, 0.5, 1.0)], |
| 265 | vllm_runner_kwargs={ |
| 266 | "model_impl": "transformers", |
| 267 | }, |
| 268 | marks=[pytest.mark.core_model], |
| 269 | ), |
| 270 | # Pixel values from processor are not 4D or 5D arrays |
| 271 | "qwen2_5_vl-transformers": VLMTestInfo( |
| 272 | models=["Qwen/Qwen2.5-VL-3B-Instruct"], |
| 273 | test_type=VLMTestType.IMAGE, |
| 274 | prompt_formatter=lambda img_prompt: f"<|im_start|>User\n{img_prompt}<|im_end|>\n<|im_start|>assistant\n", # noqa: E501 |
| 275 | img_idx_to_prompt=lambda idx: "<|vision_start|><|image_pad|><|vision_end|>", |
| 276 | max_model_len=4096, |
| 277 | max_num_seqs=2, |
| 278 | auto_cls=AutoModelForImageTextToText, |
| 279 | vllm_output_post_proc=model_utils.qwen2_vllm_to_hf_output, |
| 280 | image_size_factors=[(0.25, 0.2, 0.15)], |
| 281 | vllm_runner_kwargs={ |
| 282 | "model_impl": "transformers", |
| 283 | }, |
| 284 | marks=[large_gpu_mark(min_gb=32)], |
| 285 | ), |
| 286 | #### Extended model tests |
| 287 | "aria": VLMTestInfo( |
| 288 | models=["rhymes-ai/Aria"], |
| 289 | test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE), |
| 290 | prompt_formatter=lambda img_prompt: f"<|im_start|>user\n{img_prompt}<|im_end|>\n<|im_start|>assistant\n ", # noqa: E501 |
| 291 | img_idx_to_prompt=lambda idx: "<fim_prefix><|img|><fim_suffix>\n", |
| 292 | max_model_len=4096, |
| 293 | max_num_seqs=2, |
| 294 | auto_cls=AutoModelForImageTextToText, |
| 295 | single_image_prompts=IMAGE_ASSETS.prompts( |
| 296 | { |
| 297 | "stop_sign": "<vlm_image>Please describe the image shortly.", |
| 298 | "cherry_blossom": "<vlm_image>Please infer the season with reason.", |
| 299 | } |
| 300 | ), |
| 301 | multi_image_prompt="<vlm_image><vlm_image>Describe the two images shortly.", |
| 302 | stop_str=["<|im_end|>"], |
| 303 | image_size_factors=[(0.10, 0.15)], |
| 304 | max_tokens=64, |
| 305 | marks=[ |
| 306 | pytest.mark.skip( |
| 307 | reason="Aria needs to update for latest transformers, " |
| 308 | "must have a vision_processor.py." |
| 309 | "An issue has been filed:" |
| 310 | "https://huggingface.co/rhymes-ai/Aria/discussions/23" |
| 311 | ), |
| 312 | large_gpu_mark(min_gb=64), |
| 313 | ], |
| 314 | ), |
| 315 | "blip2": VLMTestInfo( |
| 316 | models=["Salesforce/blip2-opt-2.7b"], |
| 317 | test_type=VLMTestType.IMAGE, |
| 318 | prompt_formatter=lambda img_prompt: f"Question: {img_prompt} Answer:", |
| 319 | img_idx_to_prompt=lambda idx: "", |
| 320 | auto_cls=AutoModelForImageTextToText, |
| 321 | vllm_output_post_proc=model_utils.blip2_vllm_to_hf_output, |
| 322 | ), |
| 323 | "cosmos3": VLMTestInfo( |
| 324 | models=["nvidia/Cosmos3-Nano"], |
| 325 | test_type=( |
| 326 | VLMTestType.IMAGE, |
| 327 | VLMTestType.MULTI_IMAGE, |
| 328 | VLMTestType.VIDEO, |
| 329 | ), |
| 330 | enforce_eager=False, |
| 331 | needs_video_metadata=True, |
| 332 | prompt_formatter=lambda img_prompt: f"<|im_start|>User\n{img_prompt}<|im_end|>\n<|im_start|>assistant\n", # noqa: E501 |
| 333 | img_idx_to_prompt=lambda idx: "<|vision_start|><|image_pad|><|vision_end|>", # noqa: E501 |
| 334 | video_idx_to_prompt=lambda idx: "<|vision_start|><|video_pad|><|vision_end|>", # noqa: E501 |
| 335 | max_model_len=4096, |
| 336 | max_num_seqs=2, |
| 337 | num_logprobs=20, |
| 338 | auto_cls=AutoModelForImageTextToText, |
| 339 | vllm_output_post_proc=model_utils.qwen2_vllm_to_hf_output, |
| 340 | patch_hf_runner=model_utils.qwen3_vl_patch_hf_runner, |
| 341 | image_size_factors=[(0.25,), (0.25, 0.25, 0.25), (0.25, 0.2, 0.15)], |
| 342 | marks=[ |
| 343 | pytest.mark.skip( |
| 344 | reason="Transformers has no cosmos3_omni mapping, so the HF " |
| 345 | "reference runner cannot load the checkpoint." |
| 346 | ) |
| 347 | ], |
| 348 | ), |
| 349 | "deepseek_vl_v2": VLMTestInfo( |
| 350 | models=["Isotr0py/deepseek-vl2-tiny"], # model repo using dynamic module |
| 351 | test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE), |
| 352 | prompt_formatter=lambda img_prompt: f"<|User|>: {img_prompt}\n\n<|Assistant|>: ", # noqa: E501 |
| 353 | max_model_len=4096, |
| 354 | max_num_seqs=2, |
| 355 | single_image_prompts=IMAGE_ASSETS.prompts( |
| 356 | { |
| 357 | "stop_sign": "<image>\nWhat's the content in the center of the image?", |
| 358 | "cherry_blossom": "<image>\nPlease infer the season with reason in details.", # noqa: E501 |
| 359 | } |
| 360 | ), |
| 361 | multi_image_prompt="image_1:<image>\nimage_2:<image>\nWhich image can we see the car and the tower?", # noqa: E501 |
| 362 | patch_hf_runner=model_utils.deepseekvl2_patch_hf_runner, |
| 363 | hf_output_post_proc=model_utils.deepseekvl2_trunc_hf_output, |
| 364 | stop_str=["<｜end▁of▁sentence｜>", "<｜begin▁of▁sentence｜>"], |
| 365 | image_size_factors=[(1.0,), (1.0, 1.0, 1.0), (0.1, 0.5, 1.0)], |
| 366 | ), |
| 367 | "gemma3": VLMTestInfo( |
| 368 | models=["google/gemma-3-4b-it"], |
| 369 | test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE), |
| 370 | 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 |
| 371 | single_image_prompts=IMAGE_ASSETS.prompts( |
| 372 | { |
| 373 | "stop_sign": "<start_of_image>What's the content in the center of the image?", # noqa: E501 |
| 374 | "cherry_blossom": "<start_of_image>What is the season?", |
| 375 | } |
| 376 | ), |
| 377 | multi_image_prompt="<start_of_image><start_of_image>Describe the two images in detail.", # noqa: E501 |
| 378 | max_model_len=4096, |
| 379 | max_num_seqs=2, |
| 380 | auto_cls=AutoModelForImageTextToText, |
| 381 | vllm_runner_kwargs={"mm_processor_kwargs": {"do_pan_and_scan": True}}, |
| 382 | patch_hf_runner=model_utils.gemma3_patch_hf_runner, |
| 383 | ), |
| 384 | "gemma4": VLMTestInfo( |
| 385 | models=["google/gemma-4-E2B-it"], |
| 386 | test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE), |
| 387 | prompt_formatter=lambda img_prompt: f"<bos><|turn>user\n{img_prompt}<turn|>\n<|turn>model\n", # noqa: E501 |
| 388 | single_image_prompts=IMAGE_ASSETS.prompts( |
| 389 | { |
| 390 | "stop_sign": "<|image|>What's the content in the center of the image?", # noqa: E501 |
| 391 | "cherry_blossom": "<|image|>What is the season?", |
| 392 | } |
| 393 | ), |
| 394 | multi_image_prompt="<|image|><|image|>Describe the two images in detail.", # noqa: E501 |
| 395 | max_model_len=4096, |
| 396 | max_num_seqs=2, |
| 397 | auto_cls=AutoModelForImageTextToText, |
| 398 | vllm_runner_kwargs={"limit_mm_per_prompt": {"image": 4}}, |
| 399 | ), |
| 400 | "granite_vision": VLMTestInfo( |
| 401 | models=["ibm-granite/granite-vision-3.3-2b"], |
| 402 | test_type=(VLMTestType.IMAGE), |
| 403 | prompt_formatter=lambda img_prompt: f"<|user|>\n{img_prompt}\n<|assistant|>\n", |
| 404 | max_model_len=8192, |
| 405 | auto_cls=AutoModelForImageTextToText, |
| 406 | vllm_output_post_proc=model_utils.llava_image_vllm_to_hf_output, |
| 407 | ), |
| 408 | "glm4v": VLMTestInfo( |
| 409 | models=["zai-org/glm-4v-9b"], |
| 410 | test_type=VLMTestType.IMAGE, |
| 411 | prompt_formatter=lambda img_prompt: f"<|user|>\n{img_prompt}<|assistant|>", |
| 412 | single_image_prompts=IMAGE_ASSETS.prompts( |
| 413 | { |
| 414 | "stop_sign": "<|begin_of_image|><|endoftext|><|end_of_image|>What's the content in the center of the image?", # noqa: E501 |
| 415 | "cherry_blossom": "<|begin_of_image|><|endoftext|><|end_of_image|>What is the season?", # noqa: E501 |
| 416 | } |
| 417 | ), |
| 418 | max_model_len=2048, |
| 419 | max_num_seqs=2, |
| 420 | get_stop_token_ids=lambda tok: [151329, 151336, 151338], |
| 421 | patch_hf_runner=model_utils.glm4v_patch_hf_runner, |
| 422 | # The image embeddings match with HF but the outputs of the language |
| 423 | # decoder are only consistent up to 2 decimal places. |
| 424 | # So, we need to reduce the number of tokens for the test to pass. |
| 425 | max_tokens=8, |
| 426 | num_logprobs=10, |
| 427 | auto_cls=AutoModelForCausalLM, |
| 428 | marks=[ |
| 429 | pytest.mark.skip( |
| 430 | reason="The code for this model has a bug." |
| 431 | "Please see the issue here:" |
| 432 | "https://huggingface.co/zai-org/glm-4v-9b/discussions/46." |
| 433 | ), |
| 434 | large_gpu_mark(min_gb=32), |
| 435 | ], |
| 436 | ), |
| 437 | "glm4_1v": VLMTestInfo( |
| 438 | models=["zai-org/GLM-4.1V-9B-Thinking"], |
| 439 | test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE), |
| 440 | prompt_formatter=lambda img_prompt: f"[gMASK]<|user|>\n{img_prompt}<|assistant|>\n", # noqa: E501 |
| 441 | img_idx_to_prompt=lambda idx: "<|begin_of_image|><|image|><|end_of_image|>", |
| 442 | video_idx_to_prompt=lambda idx: "<|begin_of_video|><|video|><|end_of_video|>", |
| 443 | max_model_len=2048, |
| 444 | max_num_seqs=2, |
| 445 | get_stop_token_ids=lambda tok: [151329, 151336, 151338], |
| 446 | num_logprobs=10, |
| 447 | image_size_factors=[(0.25,), (0.25, 0.25, 0.25), (0.25, 0.2, 0.15)], |
| 448 | auto_cls=AutoModelForImageTextToText, |
| 449 | marks=[large_gpu_mark(min_gb=32)], |
| 450 | ), |
| 451 | "glm4_1v-video": VLMTestInfo( |
| 452 | models=["zai-org/GLM-4.1V-9B-Thinking"], |
| 453 | # GLM4.1V require include video metadata for input |
| 454 | test_type=VLMTestType.CUSTOM_INPUTS, |
| 455 | prompt_formatter=lambda vid_prompt: f"[gMASK]<|user|>\n{vid_prompt}<|assistant|>\n", # noqa: E501 |
| 456 | max_model_len=4096, |
| 457 | max_num_seqs=2, |
| 458 | auto_cls=AutoModelForImageTextToText, |
| 459 | patch_hf_runner=model_utils.glm4_1v_patch_hf_runner, |
| 460 | custom_test_opts=[ |
| 461 | CustomTestOptions( |
| 462 | inputs=custom_inputs.video_with_metadata_glm4_1v(), |
| 463 | limit_mm_per_prompt={"video": 1}, |
| 464 | ) |
| 465 | ], |
| 466 | marks=[large_gpu_mark(min_gb=32)], |
| 467 | ), |
| 468 | "glm_ocr": VLMTestInfo( |
| 469 | models=["zai-org/GLM-OCR"], |
| 470 | test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE), |
| 471 | prompt_formatter=lambda img_prompt: f"[gMASK]<|user|>\n{img_prompt}<|assistant|>\n", # noqa: E501 |
| 472 | img_idx_to_prompt=lambda idx: "<|begin_of_image|><|image|><|end_of_image|>", |
| 473 | video_idx_to_prompt=lambda idx: "<|begin_of_video|><|video|><|end_of_video|>", |
| 474 | max_model_len=2048, |
| 475 | max_num_seqs=2, |
| 476 | get_stop_token_ids=lambda tok: [151329, 151336, 151338], |
| 477 | num_logprobs=10, |
| 478 | image_size_factors=[(0.25,), (0.25, 0.25, 0.25), (0.25, 0.2, 0.15)], |
| 479 | auto_cls=AutoModelForImageTextToText, |
| 480 | marks=[ |
| 481 | pytest.mark.skip( |
| 482 | reason="This test fails on both AMD and NV" |
| 483 | "hardware. please see the issue:" |
| 484 | "https://github.com/vllm-project/vllm/issues/42016" |
| 485 | ), |
| 486 | large_gpu_mark(min_gb=32), |
| 487 | ], |
| 488 | ), |
| 489 | "granite4_vision": VLMTestInfo( |
| 490 | models=["ibm-granite/granite-vision-4.1-4b"], |
| 491 | test_type=(VLMTestType.IMAGE), |
| 492 | prompt_formatter=lambda img_prompt: f"<|user|>\n{img_prompt}\n<|assistant|>\n", |
| 493 | max_model_len=8192, |
| 494 | auto_cls=AutoModelForImageTextToText, |
| 495 | vllm_output_post_proc=_granite4_vision_vllm_to_hf_output, |
| 496 | image_size_factors=[(1.0,)], |
| 497 | vllm_runner_kwargs={ |
| 498 | "enable_lora": True, |
| 499 | "max_lora_rank": 256, |
| 500 | "default_mm_loras": {"image": "ibm-granite/granite-vision-4.1-4b"}, |
| 501 | }, |
| 502 | ), |
| 503 | "h2ovl": VLMTestInfo( |
| 504 | models=[ |
| 505 | "h2oai/h2ovl-mississippi-800m", |
| 506 | "h2oai/h2ovl-mississippi-2b", |
| 507 | ], |
| 508 | test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE), |
| 509 | prompt_formatter=lambda img_prompt: f"<|prompt|>{img_prompt}<|end|><|answer|>", |
| 510 | single_image_prompts=IMAGE_ASSETS.prompts( |
| 511 | { |
| 512 | "stop_sign": "<image>\nWhat's the content in the center of the image?", |
| 513 | "cherry_blossom": "<image>\nWhat is the season?", |
| 514 | } |
| 515 | ), |
| 516 | multi_image_prompt="Image-1: <image>\nImage-2: <image>\nDescribe the two images in short.", # noqa: E501 |
| 517 | max_model_len=8192, |
| 518 | use_tokenizer_eos=True, |
| 519 | num_logprobs=10, |
| 520 | patch_hf_runner=model_utils.h2ovl_patch_hf_runner, |
| 521 | ), |
| 522 | "idefics3": VLMTestInfo( |
| 523 | models=["HuggingFaceTB/SmolVLM-256M-Instruct"], |
| 524 | test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE), |
| 525 | prompt_formatter=lambda img_prompt: f"<|begin_of_text|>User:{img_prompt}<end_of_utterance>\nAssistant:", # noqa: E501 |
| 526 | img_idx_to_prompt=lambda idx: "<image>", |
| 527 | max_model_len=8192, |
| 528 | max_num_seqs=2, |
| 529 | auto_cls=AutoModelForImageTextToText, |
| 530 | hf_output_post_proc=model_utils.idefics3_trunc_hf_output, |
| 531 | ), |
| 532 | "intern_vl": VLMTestInfo( |
| 533 | models=[ |
| 534 | "OpenGVLab/InternVL2-1B", |
| 535 | "OpenGVLab/InternVL2-2B", |
| 536 | ], |
| 537 | test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE), |
| 538 | prompt_formatter=lambda img_prompt: f"<|im_start|>User\n{img_prompt}<|im_end|>\n<|im_start|>Assistant\n", # noqa: E501 |
| 539 | single_image_prompts=IMAGE_ASSETS.prompts( |
| 540 | { |
| 541 | "stop_sign": "<image>\nWhat's the content in the center of the image?", |
| 542 | "cherry_blossom": "<image>\nWhat is the season?", |
| 543 | } |
| 544 | ), |
| 545 | multi_image_prompt="Image-1: <image>\nImage-2: <image>\nDescribe the two images in short.", # noqa: E501 |
| 546 | max_model_len=4096, |
| 547 | use_tokenizer_eos=True, |
| 548 | patch_hf_runner=model_utils.internvl_patch_hf_runner, |
| 549 | # TODO: Remove skip once model has been upstreamed to Transformers |
| 550 | marks=[ |
| 551 | pytest.mark.skip( |
| 552 | reason="Custom model code tries to access data from meta-tensor" |
| 553 | ) |
| 554 | ], |
| 555 | ), |
| 556 | "intern_vl-video": VLMTestInfo( |
| 557 | models=[ |
| 558 | "OpenGVLab/InternVL3-1B", |
| 559 | ], |
| 560 | test_type=VLMTestType.VIDEO, |
| 561 | prompt_formatter=lambda img_prompt: f"<|im_start|>User\n{img_prompt}<|im_end|>\n<|im_start|>Assistant\n", # noqa: E501 |
| 562 | video_idx_to_prompt=lambda idx: "<video>", |
| 563 | max_model_len=8192, |
| 564 | use_tokenizer_eos=True, |
| 565 | patch_hf_runner=model_utils.internvl_patch_hf_runner, |
| 566 | num_logprobs=10 if current_platform.is_rocm() else 5, |
| 567 | # TODO: Remove skip once model has been upstreamed to Transformers |
| 568 | marks=[ |
| 569 | pytest.mark.skip( |
| 570 | reason="Custom model code tries to access data from meta-tensor" |
| 571 | ) |
| 572 | ], |
| 573 | ), |
| 574 | "intern_vl-hf": VLMTestInfo( |
| 575 | models=["OpenGVLab/InternVL3-1B-hf"], |
| 576 | test_type=( |
| 577 | VLMTestType.IMAGE, |
| 578 | VLMTestType.MULTI_IMAGE, |
| 579 | VLMTestType.VIDEO, |
| 580 | ), |
| 581 | prompt_formatter=lambda img_prompt: f"<|im_start|>User\n{img_prompt}<|im_end|>\n<|im_start|>Assistant\n", # noqa: E501 |
| 582 | img_idx_to_prompt=lambda idx: "<IMG_CONTEXT>", |
| 583 | video_idx_to_prompt=lambda idx: "<video>", |
| 584 | max_model_len=8192, |
| 585 | use_tokenizer_eos=True, |
| 586 | auto_cls=AutoModelForImageTextToText, |
| 587 | ), |
| 588 | "isaac": VLMTestInfo( |
| 589 | # NOTE: PerceptronAI/Isaac-0.1 removed because the upstream HF |
| 590 | # repo has a stale model.safetensors.index.json that references |
| 591 | # shard files which no longer exist (consolidated into a single |
| 592 | # model.safetensors on 2026-03-20). Re-add once upstream fixes |
| 593 | # the index file. |
| 594 | models=[ |
| 595 | "PerceptronAI/Isaac-0.2-2B-Preview", |
| 596 | ], |
| 597 | test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE), |
| 598 | prompt_formatter=lambda img_prompt: ( |
| 599 | f"<|im_start|>User\n{img_prompt}<|im_end|>\n<|im_start|>assistant\n" |
| 600 | ), |
| 601 | img_idx_to_prompt=lambda idx: "<image>", |
| 602 | single_image_prompts=IMAGE_ASSETS.prompts( |
| 603 | { |
| 604 | "stop_sign": "<vlm_image>Please describe the image shortly.", |
| 605 | "cherry_blossom": "<vlm_image>Please infer the season with reason.", |
| 606 | } |
| 607 | ), |
| 608 | multi_image_prompt=( |
| 609 | "Picture 1: <vlm_image>\n" |
| 610 | "Picture 2: <vlm_image>\n" |
| 611 | "Describe these two images with one paragraph respectively." |
| 612 | ), |
| 613 | enforce_eager=False, |
| 614 | max_model_len=4096, |
| 615 | max_num_seqs=2, |
| 616 | hf_model_kwargs={"device_map": "auto"}, |
| 617 | patch_hf_runner=model_utils.isaac_patch_hf_runner, |
| 618 | image_size_factors=[(0.25,), (0.25, 0.25, 0.25), (0.25, 0.2, 0.15)], |
| 619 | # TODO: Remove skip once model has been upstreamed to Transformers |
| 620 | marks=[pytest.mark.skip(reason="Custom model imports deleted object")], # noqa: E501 |
| 621 | ), |
| 622 | "kimi_vl": VLMTestInfo( |
| 623 | models=["moonshotai/Kimi-VL-A3B-Instruct"], |
| 624 | test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE), |
| 625 | prompt_formatter=lambda img_prompt: f"<|im_user|>user<|im_middle|>{img_prompt}<|im_end|><|im_assistant|>assistant<|im_middle|>", # noqa: E501 |
| 626 | img_idx_to_prompt=lambda _: "<|media_start|>image<|media_content|><|media_pad|><|media_end|>", # noqa: E501 |
| 627 | max_model_len=8192, |
| 628 | max_num_seqs=2, |
| 629 | dtype="bfloat16", |
| 630 | tensor_parallel_size=1, |
| 631 | vllm_output_post_proc=model_utils.kimiv_vl_vllm_to_hf_output, |
| 632 | marks=[large_gpu_mark(min_gb=48)], |
| 633 | ), |
| 634 | "llama4": VLMTestInfo( |
| 635 | models=["meta-llama/Llama-4-Scout-17B-16E-Instruct"], |
| 636 | 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 |
| 637 | img_idx_to_prompt=lambda _: "<|image|>", |
| 638 | test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE), |
| 639 | distributed_executor_backend="mp", |
| 640 | image_size_factors=[(0.25, 0.5, 1.0)], |
| 641 | hf_model_kwargs={"device_map": "auto"}, |
| 642 | max_model_len=8192, |
| 643 | max_num_seqs=4, |
| 644 | dtype="bfloat16", |
| 645 | auto_cls=AutoModelForImageTextToText, |
| 646 | tensor_parallel_size=4, |
| 647 | marks=multi_gpu_marks(num_gpus=4), |
| 648 | ), |
| 649 | "llava_next": VLMTestInfo( |
| 650 | models=["llava-hf/llava-v1.6-mistral-7b-hf"], |
| 651 | test_type=(VLMTestType.IMAGE, VLMTestType.CUSTOM_INPUTS), |
| 652 | prompt_formatter=lambda img_prompt: f"[INST] {img_prompt} [/INST]", |
| 653 | max_model_len=10240, |
| 654 | auto_cls=AutoModelForImageTextToText, |
| 655 | vllm_output_post_proc=model_utils.llava_image_vllm_to_hf_output, |
| 656 | custom_test_opts=[ |
| 657 | CustomTestOptions( |
| 658 | inputs=custom_inputs.multi_image_multi_aspect_ratio_inputs( |
| 659 | formatter=lambda img_prompt: f"[INST] {img_prompt} [/INST]" |
| 660 | ), |
| 661 | limit_mm_per_prompt={"image": 4}, |
| 662 | ) |
| 663 | ], |
| 664 | ), |
| 665 | "llava_onevision": VLMTestInfo( |
| 666 | models=["llava-hf/llava-onevision-qwen2-0.5b-ov-hf"], |
| 667 | test_type=VLMTestType.CUSTOM_INPUTS, |
| 668 | prompt_formatter=lambda vid_prompt: f"<|im_start|>user\n{vid_prompt}<|im_end|>\n<|im_start|>assistant\n", # noqa: E501 |
| 669 | num_video_frames=16, |
| 670 | max_model_len=16384, |
| 671 | hf_model_kwargs=model_utils.llava_onevision_hf_model_kwargs( |
| 672 | "llava-hf/llava-onevision-qwen2-0.5b-ov-hf" |
| 673 | ), |
| 674 | auto_cls=AutoModelForImageTextToText, |
| 675 | vllm_output_post_proc=model_utils.llava_onevision_vllm_to_hf_output, |
| 676 | custom_test_opts=[ |
| 677 | CustomTestOptions( |
| 678 | inputs=custom_inputs.multi_video_multi_aspect_ratio_inputs( |
| 679 | formatter=lambda vid_prompt: f"<|im_start|>user\n{vid_prompt}<|im_end|>\n<|im_start|>assistant\n", # noqa: E501 |
| 680 | ), |
| 681 | limit_mm_per_prompt={"video": 4}, |
| 682 | ) |
| 683 | ], |
| 684 | ), |
| 685 | "llava_next_video": VLMTestInfo( |
| 686 | models=["llava-hf/LLaVA-NeXT-Video-7B-hf"], |
| 687 | test_type=VLMTestType.VIDEO, |
| 688 | prompt_formatter=lambda vid_prompt: f"USER: {vid_prompt} ASSISTANT:", |
| 689 | num_video_frames=16, |
| 690 | max_model_len=4096, |
| 691 | max_num_seqs=2, |
| 692 | auto_cls=AutoModelForImageTextToText, |
| 693 | vllm_output_post_proc=model_utils.llava_video_vllm_to_hf_output, |
| 694 | ), |
| 695 | "minicpmv_25": VLMTestInfo( |
| 696 | models=["openbmb/MiniCPM-Llama3-V-2_5"], |
| 697 | test_type=VLMTestType.IMAGE, |
| 698 | 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 |
| 699 | img_idx_to_prompt=lambda idx: "(<image>./</image>)\n", |
| 700 | max_model_len=4096, |
| 701 | max_num_seqs=2, |
| 702 | get_stop_token_ids=lambda tok: [tok.eos_id, tok.eot_id], |
| 703 | hf_output_post_proc=model_utils.minicpmv_trunc_hf_output, |
| 704 | patch_hf_runner=model_utils.minicpmv_25_patch_hf_runner, |
| 705 | ), |
| 706 | "minicpmo_26": VLMTestInfo( |
| 707 | models=["openbmb/MiniCPM-o-2_6"], |
| 708 | test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE), |
| 709 | 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 |
| 710 | img_idx_to_prompt=lambda idx: "(<image>./</image>)\n", |
| 711 | max_model_len=4096, |
| 712 | max_num_seqs=2, |
| 713 | get_stop_token_ids=lambda tok: tok.convert_tokens_to_ids( |
| 714 | ["<|im_end|>", "<|endoftext|>"] |
| 715 | ), |
| 716 | hf_output_post_proc=model_utils.minicpmv_trunc_hf_output, |
| 717 | patch_hf_runner=model_utils.minicpmo_26_patch_hf_runner, |
| 718 | ), |
| 719 | "minicpmv_26": VLMTestInfo( |
| 720 | models=["openbmb/MiniCPM-V-2_6"], |
| 721 | test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE), |
| 722 | 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 |
| 723 | img_idx_to_prompt=lambda idx: "(<image>./</image>)\n", |
| 724 | max_model_len=4096, |
| 725 | max_num_seqs=2, |
| 726 | get_stop_token_ids=lambda tok: tok.convert_tokens_to_ids( |
| 727 | ["<|im_end|>", "<|endoftext|>"] |
| 728 | ), |
| 729 | hf_output_post_proc=model_utils.minicpmv_trunc_hf_output, |
| 730 | patch_hf_runner=model_utils.minicpmv_26_patch_hf_runner, |
| 731 | ), |
| 732 | "molmo": VLMTestInfo( |
| 733 | models=["allenai/Molmo-7B-D-0924"], |
| 734 | test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE), |
| 735 | prompt_formatter=identity, |
| 736 | max_model_len=4096, |
| 737 | max_num_seqs=2, |
| 738 | patch_hf_runner=model_utils.molmo_patch_hf_runner, |
| 739 | ), |
| 740 | "moondream3": VLMTestInfo( |
| 741 | models=["moondream/moondream3-preview"], |
| 742 | test_type=VLMTestType.IMAGE, |
| 743 | prompt_formatter=identity, |
| 744 | img_idx_to_prompt=lambda idx: "<|endoftext|><image>", |
| 745 | # Common-image coverage here targets query/caption. The native |
| 746 | # detect/point skills are not exposed by vLLM. |
| 747 | single_image_prompts=IMAGE_ASSETS.prompts( |
| 748 | { |
| 749 | "stop_sign": "<vlm_image><|md_reserved_0|>query<|md_reserved_1|>What is this sign?<|md_reserved_2|>", # noqa: E501 |
| 750 | "cherry_blossom": ( |
| 751 | "<vlm_image><|md_reserved_0|>query<|md_reserved_1|>What season is shown?<|md_reserved_2|>" # noqa: E501 |
| 752 | ), |
| 753 | } |
| 754 | ), |
| 755 | max_model_len=4096, |
| 756 | max_num_seqs=2, |
| 757 | dtype="bfloat16", |
| 758 | hf_processor=model_utils.moondream3_processor, |
| 759 | patch_hf_runner=model_utils.moondream3_patch_hf_runner, |
| 760 | # Single size factor to avoid GPU OOM when running multiple test |
| 761 | # cases sequentially (9B MoE model uses ~18 GiB per instance). |
| 762 | image_size_factors=[(1.0,)], |
| 763 | # Moondream3 is 9B params with MoE, needs significant GPU memory |
| 764 | marks=[large_gpu_mark(min_gb=48)], |
| 765 | ), |
| 766 | "ovis1_6-gemma2": VLMTestInfo( |
| 767 | models=["AIDC-AI/Ovis1.6-Gemma2-9B"], |
| 768 | test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE), |
| 769 | 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 |
| 770 | img_idx_to_prompt=lambda idx: "<image>\n", |
| 771 | max_model_len=4096, |
| 772 | max_num_seqs=2, |
| 773 | dtype="half", |
| 774 | # use sdpa mode for hf runner since ovis2 didn't work with flash_attn |
| 775 | hf_model_kwargs={"llm_attn_implementation": "sdpa"}, |
| 776 | patch_hf_runner=model_utils.ovis_patch_hf_runner, |
| 777 | marks=[large_gpu_mark(min_gb=32)], |
| 778 | ), |
| 779 | "ovis2": VLMTestInfo( |
| 780 | models=["AIDC-AI/Ovis2-1B"], |
| 781 | test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE), |
| 782 | 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 |
| 783 | img_idx_to_prompt=lambda idx: "<image>\n", |
| 784 | max_model_len=4096, |
| 785 | max_num_seqs=2, |
| 786 | dtype="half", |
| 787 | # use sdpa mode for hf runner since ovis2 didn't work with flash_attn |
| 788 | hf_model_kwargs={"llm_attn_implementation": "sdpa"}, |
| 789 | patch_hf_runner=model_utils.ovis_patch_hf_runner, |
| 790 | ), |
| 791 | "ovis2_5": VLMTestInfo( |
| 792 | models=["AIDC-AI/Ovis2.5-2B"], |
| 793 | test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE, VLMTestType.VIDEO), |
| 794 | 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 |
| 795 | img_idx_to_prompt=lambda idx: "<image>\n", |
| 796 | video_idx_to_prompt=lambda idx: "<video>\n", |
| 797 | max_model_len=4096, |
| 798 | max_num_seqs=2, |
| 799 | dtype="half", |
| 800 | num_logprobs=10, |
| 801 | patch_hf_runner=model_utils.ovis2_5_patch_hf_runner, |
| 802 | ), |
| 803 | "paddleocr_vl": VLMTestInfo( |
| 804 | models=["PaddlePaddle/PaddleOCR-VL"], |
| 805 | test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE), |
| 806 | prompt_formatter=lambda img_prompt: f"USER: {img_prompt}\nASSISTANT:", |
| 807 | img_idx_to_prompt=lambda idx: ( |
| 808 | "<|IMAGE_START|><|IMAGE_PLACEHOLDER|><|IMAGE_END|>" |
| 809 | ), |
| 810 | multi_image_prompt=( |
| 811 | "Image-1: <|IMAGE_START|><|IMAGE_PLACEHOLDER|><|IMAGE_END|>\n" |
| 812 | "Image-2: <|IMAGE_START|><|IMAGE_PLACEHOLDER|><|IMAGE_END|>\n" |
| 813 | "Describe these two images separately." |
| 814 | ), |
| 815 | max_model_len=8192, |
| 816 | max_num_seqs=2, |
| 817 | auto_cls=AutoModelForCausalLM, |
| 818 | patch_hf_runner=model_utils.paddleocr_vl_patch_hf_runner, |
| 819 | image_size_factors=[(0.25,)], |
| 820 | marks=[ |
| 821 | pytest.mark.skipif( |
| 822 | Version(TRANSFORMERS_VERSION) >= Version("5.0.0"), |
| 823 | reason="Model's custom code uses ROPE_INIT_FUNCTIONS" |
| 824 | "['default'] which was removed in transformers v5", |
| 825 | ), |
| 826 | ], |
| 827 | ), |
| 828 | "phi3v": VLMTestInfo( |
| 829 | models=["microsoft/Phi-3.5-vision-instruct"], |
| 830 | test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE), |
| 831 | prompt_formatter=lambda img_prompt: f"<|user|>\n{img_prompt}<|end|>\n<|assistant|>\n", # noqa: E501 |
| 832 | img_idx_to_prompt=lambda idx: f"<|image_{idx}|>\n", |
| 833 | max_model_len=4096, |
| 834 | max_num_seqs=2, |
| 835 | runner="generate", |
| 836 | # use sdpa mode for hf runner since phi3v didn't work with flash_attn |
| 837 | hf_model_kwargs={"_attn_implementation": "sdpa"}, |
| 838 | use_tokenizer_eos=True, |
| 839 | vllm_output_post_proc=model_utils.phi3v_vllm_to_hf_output, |
| 840 | num_logprobs=10, |
| 841 | ), |
| 842 | "pixtral_hf": VLMTestInfo( |
| 843 | models=["nm-testing/pixtral-12b-FP8-dynamic"], |
| 844 | test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE), |
| 845 | prompt_formatter=lambda img_prompt: f"<s>[INST]{img_prompt}[/INST]", |
| 846 | img_idx_to_prompt=lambda idx: "[IMG]", |
| 847 | max_model_len=8192, |
| 848 | max_num_seqs=2, |
| 849 | auto_cls=AutoModelForImageTextToText, |
| 850 | marks=[ |
| 851 | large_gpu_mark(min_gb=48), |
| 852 | pytest.mark.skipif( |
| 853 | current_platform.is_rocm(), |
| 854 | reason="Model produces a vector of <UNK> output in HF on ROCm", |
| 855 | ), |
| 856 | ], |
| 857 | ), |
| 858 | "qianfan_ocr": VLMTestInfo( |
| 859 | models=["baidu/Qianfan-OCR"], |
| 860 | test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE), |
| 861 | prompt_formatter=lambda img_prompt: f"<|im_start|>user\n{img_prompt}<|im_end|>\n<|im_start|>assistant\n", # noqa: E501 |
| 862 | img_idx_to_prompt=lambda idx: "<image>", |
| 863 | max_model_len=4096, |
| 864 | use_tokenizer_eos=True, |
| 865 | auto_cls=AutoModelForImageTextToText, |
| 866 | ), |
| 867 | "qwen2_vl": VLMTestInfo( |
| 868 | models=["Qwen/Qwen2-VL-2B-Instruct"], |
| 869 | test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE, VLMTestType.VIDEO), |
| 870 | prompt_formatter=lambda img_prompt: f"<|im_start|>User\n{img_prompt}<|im_end|>\n<|im_start|>assistant\n", # noqa: E501 |
| 871 | img_idx_to_prompt=lambda idx: "<|vision_start|><|image_pad|><|vision_end|>", |
| 872 | video_idx_to_prompt=lambda idx: "<|vision_start|><|video_pad|><|vision_end|>", |
| 873 | multi_image_prompt="Picture 1: <vlm_image>\nPicture 2: <vlm_image>\nDescribe these two images with one paragraph respectively.", # noqa: E501 |
| 874 | max_model_len=4096, |
| 875 | max_num_seqs=2, |
| 876 | # Qwen2-VL multi-image generation can accumulate backend-specific |
| 877 | # numerical drift during long autoregressive decoding. |
| 878 | # |
| 879 | # On CPU with torch 2.13, HF and vLLM agree for a long prefix before |
| 880 | # diverging near the end of generation. See pytorch/pytorch#187735. |
| 881 | # |
| 882 | # A similar late-generation divergence is observed on XPU with BF16 |
| 883 | # ViT FlashAttention. The issue is not resolved by increasing |
| 884 | # num_logprobs from 10 to 20, while FP32 and BF16 with Triton ViT |
| 885 | # attention pass in local reproduction. |
| 886 | # |
| 887 | # Cap max_tokens on CPU and XPU to keep the correctness comparison |
| 888 | # before the numerically sensitive late-generation region. |
| 889 | max_tokens=64 |
| 890 | if current_platform.is_cpu() or current_platform.is_xpu() |
| 891 | else 128, |
| 892 | num_logprobs=20 if current_platform.is_cpu() else 10, |
| 893 | auto_cls=AutoModelForImageTextToText, |
| 894 | vllm_output_post_proc=model_utils.qwen2_vllm_to_hf_output, |
| 895 | image_size_factors=[(0.25,), (0.25, 0.25, 0.25), (0.25, 0.2, 0.15)], |
| 896 | marks=[pytest.mark.cpu_model], |
| 897 | ), |
| 898 | "skywork_r1v": VLMTestInfo( |
| 899 | models=["Skywork/Skywork-R1V-38B"], |
| 900 | test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE), |
| 901 | prompt_formatter=lambda img_prompt: f"<｜begin▁of▁sentence｜><｜User｜>\n{img_prompt}<｜Assistant｜><think>\n", # noqa: E501 |
| 902 | single_image_prompts=IMAGE_ASSETS.prompts( |
| 903 | { |
| 904 | "stop_sign": "<image>\nWhat's the content in the center of the image?", |
| 905 | "cherry_blossom": "<image>\nWhat is the season?", |
| 906 | } |
| 907 | ), |
| 908 | multi_image_prompt="<image>\n<image>\nDescribe the two images in short.", |
| 909 | max_model_len=4096, |
| 910 | use_tokenizer_eos=True, |
| 911 | patch_hf_runner=model_utils.skyworkr1v_patch_hf_runner, |
| 912 | marks=[large_gpu_mark(min_gb=80)], |
| 913 | ), |
| 914 | "smolvlm": VLMTestInfo( |
| 915 | models=["HuggingFaceTB/SmolVLM2-2.2B-Instruct"], |
| 916 | test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE), |
| 917 | prompt_formatter=lambda img_prompt: f"<|im_start|>User:{img_prompt}<end_of_utterance>\nAssistant:", # noqa: E501 |
| 918 | img_idx_to_prompt=lambda idx: "<image>", |
| 919 | max_model_len=8192, |
| 920 | max_num_seqs=2, |
| 921 | auto_cls=AutoModelForImageTextToText, |
| 922 | hf_output_post_proc=model_utils.smolvlm_trunc_hf_output, |
| 923 | num_logprobs=10, |
| 924 | ), |
| 925 | ### Tensor parallel / multi-gpu broadcast tests |
| 926 | "llava-broadcast": VLMTestInfo( |
| 927 | models=["llava-hf/llava-1.5-7b-hf"], |
| 928 | prompt_formatter=lambda img_prompt: f"USER: {img_prompt}\nASSISTANT:", |
| 929 | max_model_len=4096, |
| 930 | auto_cls=AutoModelForImageTextToText, |
| 931 | vllm_output_post_proc=model_utils.llava_image_vllm_to_hf_output, |
| 932 | marks=multi_gpu_marks(num_gpus=2), |
| 933 | **COMMON_BROADCAST_SETTINGS, # type: ignore |
| 934 | ), |
| 935 | "llava_next-broadcast": VLMTestInfo( |
| 936 | models=["llava-hf/llava-v1.6-mistral-7b-hf"], |
| 937 | prompt_formatter=lambda img_prompt: f"[INST] {img_prompt} [/INST]", |
| 938 | max_model_len=10240, |
| 939 | auto_cls=AutoModelForImageTextToText, |
| 940 | vllm_output_post_proc=model_utils.llava_image_vllm_to_hf_output, |
| 941 | marks=multi_gpu_marks(num_gpus=2), |
| 942 | **COMMON_BROADCAST_SETTINGS, # type: ignore |
| 943 | ), |
| 944 | ### Custom input edge-cases for specific models |
| 945 | "intern_vl-diff-patches": VLMTestInfo( |
| 946 | models=["OpenGVLab/InternVL2-2B"], |
| 947 | prompt_formatter=lambda img_prompt: f"<|im_start|>User\n{img_prompt}<|im_end|>\n<|im_start|>Assistant\n", # noqa: E501 |
| 948 | test_type=VLMTestType.CUSTOM_INPUTS, |
| 949 | max_model_len=4096, |
| 950 | use_tokenizer_eos=True, |
| 951 | patch_hf_runner=model_utils.internvl_patch_hf_runner, |
| 952 | custom_test_opts=[ |
| 953 | CustomTestOptions( |
| 954 | inputs=inp, |
| 955 | limit_mm_per_prompt={"image": 2}, |
| 956 | ) |
| 957 | for inp in custom_inputs.different_patch_input_cases_internvl() |
| 958 | ], |
| 959 | # TODO: Remove skip once model has been upstreamed to Transformers |
| 960 | marks=[ |
| 961 | pytest.mark.skip( |
| 962 | reason="Custom model code tries to access data from meta-tensor" |
| 963 | ) |
| 964 | ], |
| 965 | ), |
| 966 | "llava_onevision-multiple-images": VLMTestInfo( |
| 967 | models=["llava-hf/llava-onevision-qwen2-0.5b-ov-hf"], |
| 968 | test_type=VLMTestType.CUSTOM_INPUTS, |
| 969 | max_model_len=16384, |
| 970 | max_num_seqs=2, |
| 971 | auto_cls=AutoModelForImageTextToText, |
| 972 | hf_model_kwargs=model_utils.llava_onevision_hf_model_kwargs( |
| 973 | "llava-hf/llava-onevision-qwen2-0.5b-ov-hf" |
| 974 | ), |
| 975 | vllm_output_post_proc=model_utils.llava_onevision_vllm_to_hf_output, |
| 976 | custom_test_opts=[ |
| 977 | CustomTestOptions( |
| 978 | inputs=custom_inputs.multi_image_multi_aspect_ratio_inputs( |
| 979 | formatter=lambda vid_prompt: f"<|im_start|>user\n{vid_prompt}<|im_end|>\n<|im_start|>assistant\n", # noqa: E501 |
| 980 | ), |
| 981 | limit_mm_per_prompt={"image": 4}, |
| 982 | ) |
| 983 | ], |
| 984 | ), |
| 985 | # regression test for https://github.com/vllm-project/vllm/issues/15122 |
| 986 | "qwen2_5_vl-windows-attention": VLMTestInfo( |
| 987 | models=["Qwen/Qwen2.5-VL-3B-Instruct"], |
| 988 | test_type=VLMTestType.CUSTOM_INPUTS, |
| 989 | max_model_len=4096, |
| 990 | max_num_seqs=2, |
| 991 | auto_cls=AutoModelForImageTextToText, |
| 992 | vllm_output_post_proc=model_utils.qwen2_vllm_to_hf_output, |
| 993 | custom_test_opts=[ |
| 994 | CustomTestOptions( |
| 995 | inputs=custom_inputs.windows_attention_image_qwen2_5_vl(), |
| 996 | limit_mm_per_prompt={"image": 1}, |
| 997 | ) |
| 998 | ], |
| 999 | ), |
| 1000 | } |
| 1001 | |
| 1002 | |
| 1003 | def _mark_splits( |
| 1004 | test_settings: dict[str, VLMTestInfo], |
| 1005 | *, |
| 1006 | num_groups: int, |
| 1007 | ) -> dict[str, VLMTestInfo]: |
| 1008 | name_by_test_info_id = {id(v): k for k, v in test_settings.items()} |
| 1009 | test_infos_by_model = defaultdict[str, list[VLMTestInfo]](list) |
| 1010 | |
| 1011 | for info in test_settings.values(): |
| 1012 | for model in info.models: |
| 1013 | test_infos_by_model[model].append(info) |
| 1014 | |
| 1015 | models = sorted(test_infos_by_model.keys()) |
| 1016 | split_size = math.ceil(len(models) / num_groups) |
| 1017 | |
| 1018 | new_test_settings = dict[str, VLMTestInfo]() |
| 1019 | |
| 1020 | for i in range(num_groups): |
| 1021 | models_in_group = models[i * split_size : (i + 1) * split_size] |
| 1022 | |
| 1023 | for model in models_in_group: |
| 1024 | for info in test_infos_by_model[model]: |
| 1025 | new_marks = (info.marks or []) + [pytest.mark.split(group=i)] |
| 1026 | new_info = info._replace(marks=new_marks) |
| 1027 | new_test_settings[name_by_test_info_id[id(info)]] = new_info |
| 1028 | |
| 1029 | missing_keys = test_settings.keys() - new_test_settings.keys() |
| 1030 | assert not missing_keys, f"Missing keys: {missing_keys}" |
| 1031 | |
| 1032 | return new_test_settings |
| 1033 | |
| 1034 | |
| 1035 | VLM_TEST_SETTINGS = _mark_splits(VLM_TEST_SETTINGS, num_groups=2) |
| 1036 | |
| 1037 | |
| 1038 | ### Test wrappers |
| 1039 | # Wrappers around the core test running func for: |
| 1040 | # - single image |
| 1041 | # - multi-image |
| 1042 | # - image embeddings |
| 1043 | # - video |
| 1044 | # - audio |
| 1045 | # - custom inputs |
| 1046 | @pytest.mark.parametrize( |
| 1047 | "model_type,test_case", |
| 1048 | get_parametrized_options( |
| 1049 | VLM_TEST_SETTINGS, |
| 1050 | test_type=VLMTestType.IMAGE, |
| 1051 | create_new_process_for_each_test=False, |
| 1052 | ), |
| 1053 | ) |
| 1054 | def test_single_image_models( |
| 1055 | tmp_path: PosixPath, |
| 1056 | model_type: str, |
| 1057 | test_case: ExpandableVLMTestArgs, |
| 1058 | hf_runner: type[HfRunner], |
| 1059 | vllm_runner: type[VllmRunner], |
| 1060 | image_assets: ImageTestAssets, |
| 1061 | ): |
| 1062 | model_test_info = VLM_TEST_SETTINGS[model_type] |
| 1063 | runners.run_single_image_test( |
| 1064 | tmp_path=tmp_path, |
| 1065 | model_test_info=model_test_info, |
| 1066 | test_case=test_case, |
| 1067 | hf_runner=hf_runner, |
| 1068 | vllm_runner=vllm_runner, |
| 1069 | image_assets=image_assets, |
| 1070 | ) |
| 1071 | |
| 1072 | |
| 1073 | @pytest.mark.parametrize( |
| 1074 | "model_type,test_case", |
| 1075 | get_parametrized_options( |
| 1076 | VLM_TEST_SETTINGS, |
| 1077 | test_type=VLMTestType.MULTI_IMAGE, |
| 1078 | create_new_process_for_each_test=False, |
| 1079 | ), |
| 1080 | ) |
| 1081 | def test_multi_image_models( |
| 1082 | tmp_path: PosixPath, |
| 1083 | model_type: str, |
| 1084 | test_case: ExpandableVLMTestArgs, |
| 1085 | hf_runner: type[HfRunner], |
| 1086 | vllm_runner: type[VllmRunner], |
| 1087 | image_assets: ImageTestAssets, |
| 1088 | ): |
| 1089 | model_test_info = VLM_TEST_SETTINGS[model_type] |
| 1090 | runners.run_multi_image_test( |
| 1091 | tmp_path=tmp_path, |
| 1092 | model_test_info=model_test_info, |
| 1093 | test_case=test_case, |
| 1094 | hf_runner=hf_runner, |
| 1095 | vllm_runner=vllm_runner, |
| 1096 | image_assets=image_assets, |
| 1097 | ) |
| 1098 | |
| 1099 | |
| 1100 | @pytest.mark.parametrize( |
| 1101 | "model_type,test_case", |
| 1102 | get_parametrized_options( |
| 1103 | VLM_TEST_SETTINGS, |
| 1104 | test_type=VLMTestType.EMBEDDING, |
| 1105 | create_new_process_for_each_test=False, |
| 1106 | ), |
| 1107 | ) |
| 1108 | def test_image_embedding_models( |
| 1109 | model_type: str, |
| 1110 | test_case: ExpandableVLMTestArgs, |
| 1111 | hf_runner: type[HfRunner], |
| 1112 | vllm_runner: type[VllmRunner], |
| 1113 | image_assets: ImageTestAssets, |
| 1114 | ): |
| 1115 | model_test_info = VLM_TEST_SETTINGS[model_type] |
| 1116 | runners.run_embedding_test( |
| 1117 | model_test_info=model_test_info, |
| 1118 | test_case=test_case, |
| 1119 | hf_runner=hf_runner, |
| 1120 | vllm_runner=vllm_runner, |
| 1121 | image_assets=image_assets, |
| 1122 | ) |
| 1123 | |
| 1124 | |
| 1125 | @pytest.mark.parametrize( |
| 1126 | "model_type,test_case", |
| 1127 | get_parametrized_options( |
| 1128 | VLM_TEST_SETTINGS, |
| 1129 | test_type=VLMTestType.VIDEO, |
| 1130 | create_new_process_for_each_test=False, |
| 1131 | ), |
| 1132 | ) |
| 1133 | def test_video_models( |
| 1134 | model_type: str, |
| 1135 | test_case: ExpandableVLMTestArgs, |
| 1136 | hf_runner: type[HfRunner], |
| 1137 | vllm_runner: type[VllmRunner], |
| 1138 | video_assets: VideoTestAssets, |
| 1139 | ): |
| 1140 | model_test_info = VLM_TEST_SETTINGS[model_type] |
| 1141 | runners.run_video_test( |
| 1142 | model_test_info=model_test_info, |
| 1143 | test_case=test_case, |
| 1144 | hf_runner=hf_runner, |
| 1145 | vllm_runner=vllm_runner, |
| 1146 | video_assets=video_assets, |
| 1147 | ) |
| 1148 | |
| 1149 | |
| 1150 | @pytest.mark.parametrize( |
| 1151 | "model_type,test_case", |
| 1152 | get_parametrized_options( |
| 1153 | VLM_TEST_SETTINGS, |
| 1154 | test_type=VLMTestType.AUDIO, |
| 1155 | create_new_process_for_each_test=False, |
| 1156 | ), |
| 1157 | ) |
| 1158 | def test_audio_models( |
| 1159 | model_type: str, |
| 1160 | test_case: ExpandableVLMTestArgs, |
| 1161 | hf_runner: type[HfRunner], |
| 1162 | vllm_runner: type[VllmRunner], |
| 1163 | audio_assets: AudioTestAssets, |
| 1164 | ): |
| 1165 | model_test_info = VLM_TEST_SETTINGS[model_type] |
| 1166 | runners.run_audio_test( |
| 1167 | model_test_info=model_test_info, |
| 1168 | test_case=test_case, |
| 1169 | hf_runner=hf_runner, |
| 1170 | vllm_runner=vllm_runner, |
| 1171 | audio_assets=audio_assets, |
| 1172 | ) |
| 1173 | |
| 1174 | |
| 1175 | @pytest.mark.parametrize( |
| 1176 | "model_type,test_case", |
| 1177 | get_parametrized_options( |
| 1178 | VLM_TEST_SETTINGS, |
| 1179 | test_type=VLMTestType.CUSTOM_INPUTS, |
| 1180 | create_new_process_for_each_test=False, |
| 1181 | ), |
| 1182 | ) |
| 1183 | def test_custom_inputs_models( |
| 1184 | model_type: str, |
| 1185 | test_case: ExpandableVLMTestArgs, |
| 1186 | hf_runner: type[HfRunner], |
| 1187 | vllm_runner: type[VllmRunner], |
| 1188 | ): |
| 1189 | model_test_info = VLM_TEST_SETTINGS[model_type] |
| 1190 | runners.run_custom_inputs_test( |
| 1191 | model_test_info=model_test_info, |
| 1192 | test_case=test_case, |
| 1193 | hf_runner=hf_runner, |
| 1194 | vllm_runner=vllm_runner, |
| 1195 | ) |
| 1196 | |
| 1197 | |
| 1198 | #### Tests filtering for things running each test as a new process |
| 1199 | @pytest.mark.parametrize( |
| 1200 | "model_type,test_case", |
| 1201 | get_parametrized_options( |
| 1202 | VLM_TEST_SETTINGS, |
| 1203 | test_type=VLMTestType.IMAGE, |
| 1204 | create_new_process_for_each_test=True, |
| 1205 | ), |
| 1206 | ) |
| 1207 | @create_new_process_for_each_test("spawn") |
| 1208 | def test_single_image_models_heavy( |
| 1209 | tmp_path: PosixPath, |
| 1210 | model_type: str, |
| 1211 | test_case: ExpandableVLMTestArgs, |
| 1212 | hf_runner: type[HfRunner], |
| 1213 | vllm_runner: type[VllmRunner], |
| 1214 | image_assets: ImageTestAssets, |
| 1215 | ): |
| 1216 | model_test_info = VLM_TEST_SETTINGS[model_type] |
| 1217 | runners.run_single_image_test( |
| 1218 | tmp_path=tmp_path, |
| 1219 | model_test_info=model_test_info, |
| 1220 | test_case=test_case, |
| 1221 | hf_runner=hf_runner, |
| 1222 | vllm_runner=vllm_runner, |
| 1223 | image_assets=image_assets, |
| 1224 | ) |
| 1225 | |
| 1226 | |
| 1227 | @pytest.mark.parametrize( |
| 1228 | "model_type,test_case", |
| 1229 | get_parametrized_options( |
| 1230 | VLM_TEST_SETTINGS, |
| 1231 | test_type=VLMTestType.MULTI_IMAGE, |
| 1232 | create_new_process_for_each_test=True, |
| 1233 | ), |
| 1234 | ) |
| 1235 | @create_new_process_for_each_test() |
| 1236 | def test_multi_image_models_heavy( |
| 1237 | tmp_path: PosixPath, |
| 1238 | model_type: str, |
| 1239 | test_case: ExpandableVLMTestArgs, |
| 1240 | hf_runner: type[HfRunner], |
| 1241 | vllm_runner: type[VllmRunner], |
| 1242 | image_assets: ImageTestAssets, |
| 1243 | ): |
| 1244 | model_test_info = VLM_TEST_SETTINGS[model_type] |
| 1245 | runners.run_multi_image_test( |
| 1246 | tmp_path=tmp_path, |
| 1247 | model_test_info=model_test_info, |
| 1248 | test_case=test_case, |
| 1249 | hf_runner=hf_runner, |
| 1250 | vllm_runner=vllm_runner, |
| 1251 | image_assets=image_assets, |
| 1252 | ) |
| 1253 | |
| 1254 | |
| 1255 | @pytest.mark.parametrize( |
| 1256 | "model_type,test_case", |
| 1257 | get_parametrized_options( |
| 1258 | VLM_TEST_SETTINGS, |
| 1259 | test_type=VLMTestType.EMBEDDING, |
| 1260 | create_new_process_for_each_test=True, |
| 1261 | ), |
| 1262 | ) |
| 1263 | @create_new_process_for_each_test() |
| 1264 | def test_image_embedding_models_heavy( |
| 1265 | model_type: str, |
| 1266 | test_case: ExpandableVLMTestArgs, |
| 1267 | hf_runner: type[HfRunner], |
| 1268 | vllm_runner: type[VllmRunner], |
| 1269 | image_assets: ImageTestAssets, |
| 1270 | ): |
| 1271 | model_test_info = VLM_TEST_SETTINGS[model_type] |
| 1272 | runners.run_embedding_test( |
| 1273 | model_test_info=model_test_info, |
| 1274 | test_case=test_case, |
| 1275 | hf_runner=hf_runner, |
| 1276 | vllm_runner=vllm_runner, |
| 1277 | image_assets=image_assets, |
| 1278 | ) |
| 1279 | |
| 1280 | |
| 1281 | @pytest.mark.parametrize( |
| 1282 | "model_type,test_case", |
| 1283 | get_parametrized_options( |
| 1284 | VLM_TEST_SETTINGS, |
| 1285 | test_type=VLMTestType.VIDEO, |
| 1286 | create_new_process_for_each_test=True, |
| 1287 | ), |
| 1288 | ) |
| 1289 | def test_video_models_heavy( |
| 1290 | model_type: str, |
| 1291 | test_case: ExpandableVLMTestArgs, |
| 1292 | hf_runner: type[HfRunner], |
| 1293 | vllm_runner: type[VllmRunner], |
| 1294 | video_assets: VideoTestAssets, |
| 1295 | ): |
| 1296 | model_test_info = VLM_TEST_SETTINGS[model_type] |
| 1297 | runners.run_video_test( |
| 1298 | model_test_info=model_test_info, |
| 1299 | test_case=test_case, |
| 1300 | hf_runner=hf_runner, |
| 1301 | vllm_runner=vllm_runner, |
| 1302 | video_assets=video_assets, |
| 1303 | ) |
| 1304 | |
| 1305 | |
| 1306 | @pytest.mark.parametrize( |
| 1307 | "model_type,test_case", |
| 1308 | get_parametrized_options( |
| 1309 | VLM_TEST_SETTINGS, |
| 1310 | test_type=VLMTestType.AUDIO, |
| 1311 | create_new_process_for_each_test=True, |
| 1312 | ), |
| 1313 | ) |
| 1314 | def test_audio_models_heavy( |
| 1315 | model_type: str, |
| 1316 | test_case: ExpandableVLMTestArgs, |
| 1317 | hf_runner: type[HfRunner], |
| 1318 | vllm_runner: type[VllmRunner], |
| 1319 | audio_assets: AudioTestAssets, |
| 1320 | ): |
| 1321 | model_test_info = VLM_TEST_SETTINGS[model_type] |
| 1322 | runners.run_audio_test( |
| 1323 | model_test_info=model_test_info, |
| 1324 | test_case=test_case, |
| 1325 | hf_runner=hf_runner, |
| 1326 | vllm_runner=vllm_runner, |
| 1327 | audio_assets=audio_assets, |
| 1328 | ) |
| 1329 | |
| 1330 | |
| 1331 | @pytest.mark.parametrize( |
| 1332 | "model_type,test_case", |
| 1333 | get_parametrized_options( |
| 1334 | VLM_TEST_SETTINGS, |
| 1335 | test_type=VLMTestType.CUSTOM_INPUTS, |
| 1336 | create_new_process_for_each_test=True, |
| 1337 | ), |
| 1338 | ) |
| 1339 | @create_new_process_for_each_test() |
| 1340 | def test_custom_inputs_models_heavy( |
| 1341 | model_type: str, |
| 1342 | test_case: ExpandableVLMTestArgs, |
| 1343 | hf_runner: type[HfRunner], |
| 1344 | vllm_runner: type[VllmRunner], |
| 1345 | ): |
| 1346 | model_test_info = VLM_TEST_SETTINGS[model_type] |
| 1347 | runners.run_custom_inputs_test( |
| 1348 | model_test_info=model_test_info, |
| 1349 | test_case=test_case, |
| 1350 | hf_runner=hf_runner, |
| 1351 | vllm_runner=vllm_runner, |
| 1352 | ) |