| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Unit tests for Qwen2.5-Omni embed_input_ids to verify embeddings are |
| 4 | correctly assigned to audio/image/video token positions. |
| 5 | |
| 6 | Regression test for: https://github.com/vllm-project/vllm/issues/34506 |
| 7 | - Non-interleaved mixed modalities (audio + image + video) should correctly |
| 8 | assign audio embeddings to audio positions, image to image, video to video. |
| 9 | - Interleaved (use_audio_in_video) should also work correctly. |
| 10 | """ |
| 11 | |
| 12 | from unittest.mock import Mock |
| 13 | |
| 14 | import pytest |
| 15 | import torch |
| 16 | |
| 17 | from vllm.model_executor.models.qwen2_5_omni_thinker import ( |
| 18 | check_interleaved_audio_video, |
| 19 | merge_interleaved_embeddings, |
| 20 | ) |
| 21 | from vllm.multimodal.utils import set_mm_embedding_modality |
| 22 | |
| 23 | # Fake token IDs |
| 24 | AUDIO_TOKEN_ID = 1001 |
| 25 | IMAGE_TOKEN_ID = 1002 |
| 26 | VIDEO_TOKEN_ID = 1003 |
| 27 | TEXT_TOKEN_ID = 0 |
| 28 | |
| 29 | |
| 30 | def _mm_embed(shape: tuple[int, ...], value: float, modality: str) -> torch.Tensor: |
| 31 | return set_mm_embedding_modality(torch.full(shape, value), modality) |
| 32 | |
| 33 | |
| 34 | # --------------------------------------------------------------------------- |
| 35 | # Helpers |
| 36 | # --------------------------------------------------------------------------- |
| 37 | |
| 38 | |
| 39 | def make_token_seq( |
| 40 | audio_n: int, image_n: int, video_n: int, text_prefix: int = 3, text_sep: int = 2 |
| 41 | ): |
| 42 | """Build a flat token sequence: |
| 43 | [text_prefix] [AUDIO * audio_n] [text_sep] [IMAGE * image_n] |
| 44 | [text_sep] [VIDEO * video_n] [text_sep] |
| 45 | Returns (input_ids tensor, is_multimodal mask, positions dict). |
| 46 | """ |
| 47 | tokens = ( |
| 48 | [TEXT_TOKEN_ID] * text_prefix |
| 49 | + [AUDIO_TOKEN_ID] * audio_n |
| 50 | + [TEXT_TOKEN_ID] * text_sep |
| 51 | + [IMAGE_TOKEN_ID] * image_n |
| 52 | + [TEXT_TOKEN_ID] * text_sep |
| 53 | + [VIDEO_TOKEN_ID] * video_n |
| 54 | + [TEXT_TOKEN_ID] * text_sep |
| 55 | ) |
| 56 | input_ids = torch.tensor(tokens) |
| 57 | is_multimodal = ( |
| 58 | (input_ids == AUDIO_TOKEN_ID) |
| 59 | | (input_ids == IMAGE_TOKEN_ID) |
| 60 | | (input_ids == VIDEO_TOKEN_ID) |
| 61 | ) |
| 62 | return input_ids, is_multimodal |
| 63 | |
| 64 | |
| 65 | def make_interleaved_seq( |
| 66 | video_chunks: list[int], audio_chunks: list[int], text_prefix: int = 2 |
| 67 | ): |
| 68 | """Build an interleaved sequence like use_audio_in_video: |
| 69 | [text] [V*v0] [A*a0] [V*v1] [A*a1] ... |
| 70 | """ |
| 71 | tokens = [TEXT_TOKEN_ID] * text_prefix |
| 72 | for v, a in zip(video_chunks, audio_chunks): |
| 73 | tokens += [VIDEO_TOKEN_ID] * v + [AUDIO_TOKEN_ID] * a |
| 74 | input_ids = torch.tensor(tokens) |
| 75 | is_multimodal = (input_ids == VIDEO_TOKEN_ID) | (input_ids == AUDIO_TOKEN_ID) |
| 76 | return input_ids, is_multimodal |
| 77 | |
| 78 | |
| 79 | # --------------------------------------------------------------------------- |
| 80 | # Tests for check_interleaved_audio_video |
| 81 | # --------------------------------------------------------------------------- |
| 82 | |
| 83 | |
| 84 | class TestCheckInterleavedAudioVideo: |
| 85 | def test_non_interleaved_audio_then_video(self): |
| 86 | """Audio entirely before video → not interleaved.""" |
| 87 | input_ids, is_multimodal = make_token_seq(5, 0, 4) |
| 88 | is_video = is_multimodal & (input_ids == VIDEO_TOKEN_ID) |
| 89 | is_audio = is_multimodal & (input_ids == AUDIO_TOKEN_ID) |
| 90 | assert not check_interleaved_audio_video( |
| 91 | is_video, is_audio, is_video.sum().item(), is_audio.sum().item() |
| 92 | ) |
| 93 | |
| 94 | def test_non_interleaved_with_image(self): |
| 95 | """Audio + image + video (the mixed_modalities case) → not interleaved.""" |
| 96 | input_ids, is_multimodal = make_token_seq(5, 4, 6) |
| 97 | is_video = is_multimodal & (input_ids == VIDEO_TOKEN_ID) |
| 98 | is_audio = is_multimodal & (input_ids == AUDIO_TOKEN_ID) |
| 99 | assert not check_interleaved_audio_video( |
| 100 | is_video, is_audio, is_video.sum().item(), is_audio.sum().item() |
| 101 | ) |
| 102 | |
| 103 | def test_no_audio(self): |
| 104 | """Video only → not interleaved.""" |
| 105 | input_ids, is_multimodal = make_token_seq(0, 0, 6) |
| 106 | is_video = is_multimodal & (input_ids == VIDEO_TOKEN_ID) |
| 107 | is_audio = is_multimodal & (input_ids == AUDIO_TOKEN_ID) |
| 108 | assert not check_interleaved_audio_video( |
| 109 | is_video, is_audio, is_video.sum().item(), is_audio.sum().item() |
| 110 | ) |
| 111 | |
| 112 | def test_interleaved(self): |
| 113 | """V A V A interleaved → True.""" |
| 114 | input_ids, is_multimodal = make_interleaved_seq([4, 4], [3, 3]) |
| 115 | is_video = is_multimodal & (input_ids == VIDEO_TOKEN_ID) |
| 116 | is_audio = is_multimodal & (input_ids == AUDIO_TOKEN_ID) |
| 117 | assert check_interleaved_audio_video( |
| 118 | is_video, is_audio, is_video.sum().item(), is_audio.sum().item() |
| 119 | ) |
| 120 | |
| 121 | def test_multi_video_with_boundary_tokens(self): |
| 122 | """Two interleaved videos separated by boundary tokens → still True. |
| 123 | |
| 124 | use_audio_in_video expands each video into a local V/A span bounded by |
| 125 | non-pad tokens. A global density check would fail across those gaps. |
| 126 | """ |
| 127 | # [text][V A V A][text boundary][V A V A] |
| 128 | first_ids, _ = make_interleaved_seq([2, 2], [2, 2], text_prefix=1) |
| 129 | second_ids, _ = make_interleaved_seq([2, 2], [2, 2], text_prefix=0) |
| 130 | boundary = torch.tensor([TEXT_TOKEN_ID, TEXT_TOKEN_ID]) |
| 131 | input_ids = torch.cat([first_ids, boundary, second_ids]) |
| 132 | is_multimodal = (input_ids == VIDEO_TOKEN_ID) | (input_ids == AUDIO_TOKEN_ID) |
| 133 | is_video = is_multimodal & (input_ids == VIDEO_TOKEN_ID) |
| 134 | is_audio = is_multimodal & (input_ids == AUDIO_TOKEN_ID) |
| 135 | assert check_interleaved_audio_video( |
| 136 | is_video, is_audio, is_video.sum().item(), is_audio.sum().item() |
| 137 | ) |
| 138 | |
| 139 | def test_batched_non_interleaved_no_false_positive(self): |
| 140 | """Regression test for https://github.com/vllm-project/vllm/issues/35394. |
| 141 | |
| 142 | 5 identical non-interleaved mixed-modality requests batched together: |
| 143 | each has [audio][image][video] in separate blocks with text between them. |
| 144 | Across the batch, audio from request N falls between video blocks of |
| 145 | request N and request N+1, causing the global ranges to overlap. |
| 146 | check_interleaved_audio_video must return False (not a false positive). |
| 147 | """ |
| 148 | # Build one request: [text][audio*5][text][image*4][text][video*6][text] |
| 149 | single_ids, _ = make_token_seq(5, 4, 6) |
| 150 | # Batch 5 identical requests (separated by text tokens to simulate padding) |
| 151 | sep = torch.tensor([TEXT_TOKEN_ID] * 3) |
| 152 | batched_ids = torch.cat([single_ids, sep] * 5) |
| 153 | is_multimodal = ( |
| 154 | (batched_ids == AUDIO_TOKEN_ID) |
| 155 | | (batched_ids == IMAGE_TOKEN_ID) |
| 156 | | (batched_ids == VIDEO_TOKEN_ID) |
| 157 | ) |
| 158 | is_video = is_multimodal & (batched_ids == VIDEO_TOKEN_ID) |
| 159 | is_audio = is_multimodal & (batched_ids == AUDIO_TOKEN_ID) |
| 160 | assert not check_interleaved_audio_video( |
| 161 | is_video, is_audio, is_video.sum().item(), is_audio.sum().item() |
| 162 | ), "Batched non-interleaved requests should not be detected as interleaved" |
| 163 | |
| 164 | |
| 165 | # --------------------------------------------------------------------------- |
| 166 | # Tests for embed_input_ids via a minimal mock |
| 167 | # --------------------------------------------------------------------------- |
| 168 | |
| 169 | |
| 170 | def make_mock_model(hidden: int = 8): |
| 171 | """Return a minimal mock of Qwen2_5OmniThinkerForConditionalGeneration |
| 172 | that has enough structure to run embed_input_ids. |
| 173 | """ |
| 174 | from vllm.model_executor.models.qwen2_5_omni_thinker import ( |
| 175 | Qwen2_5OmniThinkerForConditionalGeneration, |
| 176 | ) |
| 177 | |
| 178 | model = Mock(spec=Qwen2_5OmniThinkerForConditionalGeneration) |
| 179 | |
| 180 | # Config with token IDs |
| 181 | cfg = Mock() |
| 182 | cfg.video_token_index = VIDEO_TOKEN_ID |
| 183 | cfg.audio_token_index = AUDIO_TOKEN_ID |
| 184 | model.config = cfg |
| 185 | |
| 186 | # embed_input_ids: simply embed each token as a one-hot-like vector |
| 187 | # token_id * ones so we can verify which embedding ends up where. |
| 188 | def fake_lm_embed(ids: torch.Tensor) -> torch.Tensor: |
| 189 | # Use .clone() so the tensor is contiguous (expand() creates a strided |
| 190 | # view with shared memory, which masked_scatter_ cannot handle). |
| 191 | return ids.float().unsqueeze(-1).expand(-1, hidden).clone() |
| 192 | |
| 193 | lang_model = Mock() |
| 194 | lang_model.embed_input_ids = fake_lm_embed |
| 195 | model.get_language_model = Mock(return_value=lang_model) |
| 196 | |
| 197 | # _embed_text_input_ids: delegate to SupportsMultiModal's implementation |
| 198 | from vllm.model_executor.models.interfaces import SupportsMultiModal |
| 199 | |
| 200 | model._embed_text_input_ids = ( |
| 201 | lambda *a, **kw: SupportsMultiModal._embed_text_input_ids(model, *a, **kw) |
| 202 | ) |
| 203 | |
| 204 | # super().embed_input_ids → use SupportsMultiModal.embed_input_ids |
| 205 | def fake_super_embed( |
| 206 | ids, |
| 207 | mm_embs=None, |
| 208 | *, |
| 209 | is_multimodal=None, |
| 210 | ): |
| 211 | return SupportsMultiModal.embed_input_ids( |
| 212 | model, |
| 213 | ids, |
| 214 | mm_embs, |
| 215 | is_multimodal=is_multimodal, |
| 216 | ) |
| 217 | |
| 218 | # Bind embed_input_ids as the real method |
| 219 | model.embed_input_ids = ( |
| 220 | lambda *a, **kw: Qwen2_5OmniThinkerForConditionalGeneration.embed_input_ids( |
| 221 | model, *a, **kw |
| 222 | ) |
| 223 | ) |
| 224 | |
| 225 | # Store super-embed for use inside the method |
| 226 | model._super_embed_input_ids = fake_super_embed |
| 227 | |
| 228 | return model, hidden |
| 229 | |
| 230 | |
| 231 | def build_mm_embeds( |
| 232 | audio_n, image_n, video_n, hidden, audio_val=10.0, image_val=20.0, video_val=30.0 |
| 233 | ): |
| 234 | """Build multimodal_embeddings list in position order (audio, image, video). |
| 235 | Each embedding is filled with a distinct constant so we can verify placement. |
| 236 | """ |
| 237 | embs = [] |
| 238 | if audio_n: |
| 239 | embs.append(torch.full((audio_n, hidden), audio_val)) |
| 240 | if image_n: |
| 241 | embs.append(torch.full((image_n, hidden), image_val)) |
| 242 | if video_n: |
| 243 | embs.append(torch.full((video_n, hidden), video_val)) |
| 244 | return embs |
| 245 | |
| 246 | |
| 247 | class TestEmbedInputIds: |
| 248 | def _run(self, audio_n, image_n, video_n, hidden=8): |
| 249 | """Run embed_input_ids for a non-interleaved mixed-modality sequence. |
| 250 | Returns (result_embeds, input_ids, is_multimodal). |
| 251 | """ |
| 252 | input_ids, is_multimodal = make_token_seq(audio_n, image_n, video_n) |
| 253 | mm_embeds = build_mm_embeds(audio_n, image_n, video_n, hidden) |
| 254 | |
| 255 | model, _ = make_mock_model(hidden) |
| 256 | result = model.embed_input_ids( |
| 257 | input_ids, mm_embeds, is_multimodal=is_multimodal |
| 258 | ) |
| 259 | return result, input_ids, is_multimodal |
| 260 | |
| 261 | def test_audio_only(self): |
| 262 | """Audio-only: audio positions get audio embeddings.""" |
| 263 | audio_n, hidden = 5, 8 |
| 264 | audio_val = 10.0 |
| 265 | result, input_ids, is_multimodal = self._run(audio_n, 0, 0, hidden) |
| 266 | |
| 267 | audio_pos = (input_ids == AUDIO_TOKEN_ID).nonzero(as_tuple=True)[0] |
| 268 | assert result[audio_pos].allclose(torch.full((audio_n, hidden), audio_val)), ( |
| 269 | "Audio positions should get audio embeddings" |
| 270 | ) |
| 271 | |
| 272 | def test_video_only(self): |
| 273 | """Video-only: video positions get video embeddings.""" |
| 274 | video_n, hidden = 6, 8 |
| 275 | video_val = 30.0 |
| 276 | result, input_ids, is_multimodal = self._run(0, 0, video_n, hidden) |
| 277 | |
| 278 | video_pos = (input_ids == VIDEO_TOKEN_ID).nonzero(as_tuple=True)[0] |
| 279 | assert result[video_pos].allclose(torch.full((video_n, hidden), video_val)), ( |
| 280 | "Video positions should get video embeddings" |
| 281 | ) |
| 282 | |
| 283 | def test_mixed_modalities_audio_goes_to_audio_pos(self): |
| 284 | """Regression test for GitHub issue #34506: |
| 285 | With audio + image + video (non-interleaved), audio positions must |
| 286 | receive audio embeddings (not image or video embeddings). |
| 287 | """ |
| 288 | audio_n, image_n, video_n, hidden = 5, 4, 6, 8 |
| 289 | audio_val, image_val, video_val = 10.0, 20.0, 30.0 |
| 290 | |
| 291 | result, input_ids, is_multimodal = self._run(audio_n, image_n, video_n, hidden) |
| 292 | |
| 293 | audio_pos = (input_ids == AUDIO_TOKEN_ID).nonzero(as_tuple=True)[0] |
| 294 | image_pos = (input_ids == IMAGE_TOKEN_ID).nonzero(as_tuple=True)[0] |
| 295 | video_pos = (input_ids == VIDEO_TOKEN_ID).nonzero(as_tuple=True)[0] |
| 296 | |
| 297 | mean_a = result[audio_pos].mean().item() |
| 298 | assert result[audio_pos].allclose(torch.full((audio_n, hidden), audio_val)), ( |
| 299 | f"Audio emb wrong: expected {audio_val}, got mean={mean_a:.1f}" |
| 300 | ) |
| 301 | |
| 302 | mean_i = result[image_pos].mean().item() |
| 303 | assert result[image_pos].allclose(torch.full((image_n, hidden), image_val)), ( |
| 304 | f"Image emb wrong: expected {image_val}, got mean={mean_i:.1f}" |
| 305 | ) |
| 306 | |
| 307 | mean_v = result[video_pos].mean().item() |
| 308 | assert result[video_pos].allclose(torch.full((video_n, hidden), video_val)), ( |
| 309 | f"Video emb wrong: expected {video_val}, got mean={mean_v:.1f}" |
| 310 | ) |
| 311 | |
| 312 | def test_text_positions_unchanged(self): |
| 313 | """Text positions should keep their text embeddings.""" |
| 314 | audio_n, image_n, video_n, hidden = 3, 2, 4, 8 |
| 315 | result, input_ids, is_multimodal = self._run(audio_n, image_n, video_n, hidden) |
| 316 | |
| 317 | text_pos = (~is_multimodal).nonzero(as_tuple=True)[0] |
| 318 | # Text tokens have value TEXT_TOKEN_ID=0, so embed → 0.0 |
| 319 | assert result[text_pos].allclose(torch.zeros(len(text_pos), hidden)), ( |
| 320 | "Text positions should keep text embeddings" |
| 321 | ) |
| 322 | |
| 323 | def test_interleaved_use_audio_in_video(self): |
| 324 | """Interleaved (use_audio_in_video): video chunks interleaved with audio. |
| 325 | Video embeddings must go to video positions, audio to audio positions. |
| 326 | """ |
| 327 | hidden = 8 |
| 328 | audio_val, video_val = 10.0, 30.0 |
| 329 | # Two video chunks of 4, two audio chunks of 3 |
| 330 | video_chunks = [4, 4] |
| 331 | audio_chunks = [3, 3] |
| 332 | input_ids, is_multimodal = make_interleaved_seq(video_chunks, audio_chunks) |
| 333 | |
| 334 | video_n = sum(video_chunks) # 8 |
| 335 | audio_n = sum(audio_chunks) # 6 |
| 336 | |
| 337 | # mm_embeds come in [video, audio] order (video feature first in |
| 338 | # mm_features when positions are the same for use_audio_in_video) |
| 339 | mm_embeds = [ |
| 340 | _mm_embed((video_n, hidden), video_val, "video"), |
| 341 | _mm_embed((audio_n, hidden), audio_val, "audio"), |
| 342 | ] |
| 343 | |
| 344 | model, _ = make_mock_model(hidden) |
| 345 | # Modalities are attached on the embedding tensors (as in encoder gather). |
| 346 | result = model.embed_input_ids( |
| 347 | input_ids, |
| 348 | mm_embeds, |
| 349 | is_multimodal=is_multimodal, |
| 350 | ) |
| 351 | |
| 352 | video_pos = (input_ids == VIDEO_TOKEN_ID).nonzero(as_tuple=True)[0] |
| 353 | audio_pos = (input_ids == AUDIO_TOKEN_ID).nonzero(as_tuple=True)[0] |
| 354 | |
| 355 | assert result[video_pos].allclose(torch.full((video_n, hidden), video_val)), ( |
| 356 | "Interleaved: video positions should get video embeddings" |
| 357 | ) |
| 358 | |
| 359 | assert result[audio_pos].allclose(torch.full((audio_n, hidden), audio_val)), ( |
| 360 | "Interleaved: audio positions should get audio embeddings" |
| 361 | ) |
| 362 | |
| 363 | |
| 364 | # --------------------------------------------------------------------------- |
| 365 | # Tests for merge_interleaved_embeddings helper |
| 366 | # --------------------------------------------------------------------------- |
| 367 | |
| 368 | |
| 369 | class TestMergeInterleavedEmbeddings: |
| 370 | def test_basic_interleaved(self): |
| 371 | """Video chunks + audio chunks scattered to correct positions.""" |
| 372 | hidden = 4 |
| 373 | input_ids, is_multimodal = make_interleaved_seq([3, 3], [2, 2]) |
| 374 | |
| 375 | is_video = is_multimodal & (input_ids == VIDEO_TOKEN_ID) |
| 376 | is_audio = is_multimodal & (input_ids == AUDIO_TOKEN_ID) |
| 377 | num_video = is_video.sum().item() # 6 |
| 378 | num_audio = is_audio.sum().item() # 4 |
| 379 | |
| 380 | inputs_embeds = torch.zeros(len(input_ids), hidden) |
| 381 | mm_embeds = [ |
| 382 | _mm_embed((num_video, hidden), 30.0, "video"), |
| 383 | _mm_embed((num_audio, hidden), 10.0, "audio"), |
| 384 | ] |
| 385 | |
| 386 | result = merge_interleaved_embeddings( |
| 387 | inputs_embeds, |
| 388 | mm_embeds, |
| 389 | is_video, |
| 390 | is_audio, |
| 391 | is_multimodal, |
| 392 | ) |
| 393 | |
| 394 | video_pos = is_video.nonzero(as_tuple=True)[0] |
| 395 | audio_pos = is_audio.nonzero(as_tuple=True)[0] |
| 396 | assert result[video_pos].allclose(torch.full((num_video, hidden), 30.0)) |
| 397 | assert result[audio_pos].allclose(torch.full((num_audio, hidden), 10.0)) |
| 398 | |
| 399 | def test_image_and_video_mixed(self): |
| 400 | """Image embeddings must not be misclassified as video.""" |
| 401 | hidden = 4 |
| 402 | # [text][I I][V A V A] |
| 403 | tokens = ( |
| 404 | [TEXT_TOKEN_ID] * 2 |
| 405 | + [IMAGE_TOKEN_ID] * 2 |
| 406 | + [VIDEO_TOKEN_ID, AUDIO_TOKEN_ID, VIDEO_TOKEN_ID, AUDIO_TOKEN_ID] |
| 407 | ) |
| 408 | input_ids = torch.tensor(tokens) |
| 409 | is_multimodal = ( |
| 410 | (input_ids == IMAGE_TOKEN_ID) |
| 411 | | (input_ids == VIDEO_TOKEN_ID) |
| 412 | | (input_ids == AUDIO_TOKEN_ID) |
| 413 | ) |
| 414 | is_video = is_multimodal & (input_ids == VIDEO_TOKEN_ID) |
| 415 | is_audio = is_multimodal & (input_ids == AUDIO_TOKEN_ID) |
| 416 | is_image = is_multimodal & (input_ids == IMAGE_TOKEN_ID) |
| 417 | |
| 418 | inputs_embeds = torch.zeros(len(input_ids), hidden) |
| 419 | mm_embeds = [ |
| 420 | _mm_embed((2, hidden), 20.0, "image"), |
| 421 | _mm_embed((2, hidden), 30.0, "video"), |
| 422 | _mm_embed((2, hidden), 10.0, "audio"), |
| 423 | ] |
| 424 | result = merge_interleaved_embeddings( |
| 425 | inputs_embeds, |
| 426 | mm_embeds, |
| 427 | is_video, |
| 428 | is_audio, |
| 429 | is_multimodal, |
| 430 | ) |
| 431 | assert result[is_image.nonzero(as_tuple=True)[0]].allclose( |
| 432 | torch.full((2, hidden), 20.0) |
| 433 | ) |
| 434 | assert result[is_video.nonzero(as_tuple=True)[0]].allclose( |
| 435 | torch.full((2, hidden), 30.0) |
| 436 | ) |
| 437 | assert result[is_audio.nonzero(as_tuple=True)[0]].allclose( |
| 438 | torch.full((2, hidden), 10.0) |
| 439 | ) |
| 440 | |
| 441 | def test_missing_modality_raises(self): |
| 442 | hidden = 2 |
| 443 | input_ids, is_multimodal = make_interleaved_seq([2], [2]) |
| 444 | is_video = is_multimodal & (input_ids == VIDEO_TOKEN_ID) |
| 445 | is_audio = is_multimodal & (input_ids == AUDIO_TOKEN_ID) |
| 446 | with pytest.raises(ValueError, match="Missing modality"): |
| 447 | merge_interleaved_embeddings( |
| 448 | torch.zeros(len(input_ids), hidden), |
| 449 | [torch.zeros(2, hidden), torch.zeros(2, hidden)], |
| 450 | is_video, |
| 451 | is_audio, |
| 452 | is_multimodal, |
| 453 | ) |
| 454 | |
| 455 | |
| 456 | if __name__ == "__main__": |
| 457 | pytest.main([__file__, "-v"]) |