| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | import random |
| 4 | from typing import Any, NamedTuple, cast |
| 5 | |
| 6 | import numpy as np |
| 7 | import pytest |
| 8 | from transformers import AutoTokenizer, PreTrainedTokenizerBase |
| 9 | |
| 10 | from vllm.benchmarks.datasets import ( |
| 11 | RandomDataset, |
| 12 | RandomMultiModalDataset, |
| 13 | SampleRequest, |
| 14 | ) |
| 15 | |
| 16 | |
| 17 | @pytest.fixture(scope="session") |
| 18 | def hf_tokenizer() -> PreTrainedTokenizerBase: |
| 19 | # Use a small, commonly available tokenizer |
| 20 | return AutoTokenizer.from_pretrained("openai-community/gpt2") |
| 21 | |
| 22 | |
| 23 | class Params(NamedTuple): |
| 24 | num_requests: int |
| 25 | prefix_len: int |
| 26 | range_ratio: float |
| 27 | input_len: int |
| 28 | output_len: int |
| 29 | |
| 30 | |
| 31 | @pytest.fixture(scope="session") |
| 32 | def random_dataset_params() -> Params: |
| 33 | return Params( |
| 34 | num_requests=16, prefix_len=7, range_ratio=0.3, input_len=50, output_len=20 |
| 35 | ) |
| 36 | |
| 37 | |
| 38 | def _fingerprint_sample(req: SampleRequest) -> tuple[str, int, int]: |
| 39 | """Project a SampleRequest into a comparable tuple.""" |
| 40 | assert isinstance(req.prompt, str) |
| 41 | return (req.prompt, req.prompt_len, req.expected_output_len) |
| 42 | |
| 43 | |
| 44 | def _collect_samples( |
| 45 | dataset: RandomDataset, |
| 46 | tokenizer: PreTrainedTokenizerBase, |
| 47 | num_requests: int = 16, |
| 48 | prefix_len: int = 7, |
| 49 | range_ratio: float = 0.3, |
| 50 | input_len: int = 50, |
| 51 | output_len: int = 20, |
| 52 | ) -> list[tuple[str, int, int]]: |
| 53 | samples = dataset.sample( |
| 54 | tokenizer=tokenizer, |
| 55 | num_requests=num_requests, |
| 56 | prefix_len=prefix_len, |
| 57 | range_ratio=range_ratio, |
| 58 | input_len=input_len, |
| 59 | output_len=output_len, |
| 60 | ) |
| 61 | return [_fingerprint_sample(s) for s in samples] |
| 62 | |
| 63 | |
| 64 | @pytest.mark.benchmark |
| 65 | def test_random_dataset_same_seed( |
| 66 | hf_tokenizer: PreTrainedTokenizerBase, random_dataset_params: Params |
| 67 | ) -> None: |
| 68 | """Same seed should yield identical outputs, even if global RNGs change. |
| 69 | |
| 70 | This guards against accidental reliance on Python's random or np.random |
| 71 | in RandomDataset after moving to numpy.default_rng. |
| 72 | """ |
| 73 | p = random_dataset_params |
| 74 | common_seed = 123 |
| 75 | dataset_a = RandomDataset(random_seed=common_seed) |
| 76 | dataset_b = RandomDataset(random_seed=common_seed) |
| 77 | a = _collect_samples( |
| 78 | dataset_a, |
| 79 | hf_tokenizer, |
| 80 | num_requests=p.num_requests, |
| 81 | prefix_len=p.prefix_len, |
| 82 | range_ratio=p.range_ratio, |
| 83 | input_len=p.input_len, |
| 84 | output_len=p.output_len, |
| 85 | ) |
| 86 | |
| 87 | # Perturb global RNG state to ensure isolation |
| 88 | random.seed(999) |
| 89 | _ = [random.random() for _ in range(100)] |
| 90 | np.random.seed(888) |
| 91 | _ = [np.random.random() for _ in range(100)] |
| 92 | |
| 93 | b = _collect_samples( |
| 94 | dataset_b, |
| 95 | hf_tokenizer, |
| 96 | num_requests=p.num_requests, |
| 97 | prefix_len=p.prefix_len, |
| 98 | range_ratio=p.range_ratio, |
| 99 | input_len=p.input_len, |
| 100 | output_len=p.output_len, |
| 101 | ) |
| 102 | assert a == b |
| 103 | |
| 104 | |
| 105 | @pytest.mark.benchmark |
| 106 | def test_random_dataset_different_seeds( |
| 107 | hf_tokenizer: PreTrainedTokenizerBase, random_dataset_params: Params |
| 108 | ) -> None: |
| 109 | """Different seeds should change outputs with overwhelming likelihood.""" |
| 110 | p = random_dataset_params |
| 111 | seed_a = 0 |
| 112 | dataset_a = RandomDataset(random_seed=seed_a) |
| 113 | a = _collect_samples( |
| 114 | dataset_a, |
| 115 | hf_tokenizer, |
| 116 | num_requests=p.num_requests, |
| 117 | prefix_len=p.prefix_len, |
| 118 | range_ratio=p.range_ratio, |
| 119 | input_len=p.input_len, |
| 120 | output_len=p.output_len, |
| 121 | ) |
| 122 | |
| 123 | seed_b = 999 |
| 124 | dataset_b = RandomDataset(random_seed=seed_b) |
| 125 | # Perturb global RNG with same seed as dataset_a to ensure isolation |
| 126 | random.seed(seed_a) |
| 127 | np.random.seed(seed_a) |
| 128 | b = _collect_samples( |
| 129 | dataset_b, |
| 130 | hf_tokenizer, |
| 131 | num_requests=p.num_requests, |
| 132 | prefix_len=p.prefix_len, |
| 133 | range_ratio=p.range_ratio, |
| 134 | input_len=p.input_len, |
| 135 | output_len=p.output_len, |
| 136 | ) |
| 137 | assert a != b |
| 138 | |
| 139 | |
| 140 | # ----------------------------- |
| 141 | # RandomMultiModalDataset tests |
| 142 | # ----------------------------- |
| 143 | |
| 144 | |
| 145 | def _mm_fingerprint_sample( |
| 146 | req: SampleRequest, |
| 147 | ) -> tuple[str, int, int, int, list[str]]: |
| 148 | """Create a compact fingerprint for multimodal samples. |
| 149 | |
| 150 | Includes: |
| 151 | - prompt string |
| 152 | - prompt_len |
| 153 | - expected_output_len |
| 154 | - count of multimodal items |
| 155 | - per-item type and URL prefix (e.g., 'data:image/jpeg;base64,') |
| 156 | """ |
| 157 | assert isinstance(req.prompt, str) |
| 158 | items = req.multi_modal_data or [] |
| 159 | item_prefixes: list[str] = [] |
| 160 | for it in items: |
| 161 | if isinstance(it, dict) and it.get("type") == "image_url": |
| 162 | url = it.get("image_url", {}).get("url", "") |
| 163 | # Only keep a short identifying prefix to avoid huge strings |
| 164 | item_prefixes.append(f"image:{url[:22]}") |
| 165 | elif isinstance(it, dict) and it.get("type") == "video_url": |
| 166 | url = it.get("video_url", {}).get("url", "") |
| 167 | item_prefixes.append(f"video:{url[:22]}") |
| 168 | else: |
| 169 | item_prefixes.append("unknown:") |
| 170 | return ( |
| 171 | req.prompt, |
| 172 | req.prompt_len, |
| 173 | req.expected_output_len, |
| 174 | len(items), |
| 175 | item_prefixes, |
| 176 | ) |
| 177 | |
| 178 | |
| 179 | def _collect_mm_samples( |
| 180 | dataset: RandomMultiModalDataset, |
| 181 | tokenizer: PreTrainedTokenizerBase, |
| 182 | *, |
| 183 | num_requests: int = 8, |
| 184 | prefix_len: int = 3, |
| 185 | range_ratio: float = 0.0, |
| 186 | input_len: int = 20, |
| 187 | output_len: int = 5, |
| 188 | base_items_per_request: int = 2, |
| 189 | num_mm_items_range_ratio: float = 0.0, |
| 190 | limit_mm_per_prompt: dict[str, int] | None = None, |
| 191 | bucket_config: dict[tuple[int, int, int], float] | None = None, |
| 192 | enable_multimodal_chat: bool = False, |
| 193 | ) -> list[SampleRequest]: |
| 194 | if limit_mm_per_prompt is None: |
| 195 | limit_mm_per_prompt = {"image": 5, "video": 0} |
| 196 | if bucket_config is None: |
| 197 | bucket_config = {(32, 32, 1): 0.5, (52, 64, 1): 0.5} |
| 198 | return dataset.sample( |
| 199 | tokenizer=tokenizer, |
| 200 | num_requests=num_requests, |
| 201 | prefix_len=prefix_len, |
| 202 | range_ratio=range_ratio, |
| 203 | input_len=input_len, |
| 204 | output_len=output_len, |
| 205 | base_items_per_request=base_items_per_request, |
| 206 | num_mm_items_range_ratio=num_mm_items_range_ratio, |
| 207 | limit_mm_per_prompt=limit_mm_per_prompt, |
| 208 | bucket_config=bucket_config, |
| 209 | enable_multimodal_chat=enable_multimodal_chat, |
| 210 | ) |
| 211 | |
| 212 | |
| 213 | @pytest.mark.benchmark |
| 214 | def test_random_mm_same_seed(hf_tokenizer: PreTrainedTokenizerBase) -> None: |
| 215 | seed = 42 |
| 216 | ds_a = RandomMultiModalDataset(random_seed=seed) |
| 217 | ds_b = RandomMultiModalDataset(random_seed=seed) |
| 218 | a = _collect_mm_samples(ds_a, hf_tokenizer) |
| 219 | b = _collect_mm_samples(ds_b, hf_tokenizer) |
| 220 | fa = [_mm_fingerprint_sample(s) for s in a] |
| 221 | fb = [_mm_fingerprint_sample(s) for s in b] |
| 222 | assert fa == fb |
| 223 | |
| 224 | |
| 225 | @pytest.mark.benchmark |
| 226 | def test_random_mm_different_seeds( |
| 227 | hf_tokenizer: PreTrainedTokenizerBase, |
| 228 | ) -> None: |
| 229 | ds_a = RandomMultiModalDataset(random_seed=0) |
| 230 | ds_b = RandomMultiModalDataset(random_seed=999) |
| 231 | a = _collect_mm_samples(ds_a, hf_tokenizer) |
| 232 | b = _collect_mm_samples(ds_b, hf_tokenizer) |
| 233 | fa = [_mm_fingerprint_sample(s) for s in a] |
| 234 | fb = [_mm_fingerprint_sample(s) for s in b] |
| 235 | assert fa != fb |
| 236 | |
| 237 | |
| 238 | @pytest.mark.benchmark |
| 239 | def test_random_mm_respects_limits( |
| 240 | hf_tokenizer: PreTrainedTokenizerBase, |
| 241 | ) -> None: |
| 242 | ds = RandomMultiModalDataset(random_seed=0) |
| 243 | # Requesting 3 items with a per-prompt limit of 1 should error per current |
| 244 | # design (dataset refuses to silently clamp below the requested baseline). |
| 245 | with pytest.raises(ValueError): |
| 246 | _collect_mm_samples( |
| 247 | ds, |
| 248 | hf_tokenizer, |
| 249 | num_requests=12, |
| 250 | base_items_per_request=3, |
| 251 | num_mm_items_range_ratio=0.0, |
| 252 | limit_mm_per_prompt={"image": 1, "video": 0}, |
| 253 | bucket_config={(32, 32, 1): 1.0}, |
| 254 | ) |
| 255 | |
| 256 | |
| 257 | @pytest.mark.benchmark |
| 258 | def test_random_mm_zero_prob_entries_are_removed( |
| 259 | hf_tokenizer: PreTrainedTokenizerBase, |
| 260 | ) -> None: |
| 261 | ds = RandomMultiModalDataset(random_seed=0) |
| 262 | # Second bucket has zero probability and should be ignored after |
| 263 | # normalization |
| 264 | samples = _collect_mm_samples( |
| 265 | ds, |
| 266 | hf_tokenizer, |
| 267 | num_requests=6, |
| 268 | base_items_per_request=2, |
| 269 | num_mm_items_range_ratio=0.0, |
| 270 | limit_mm_per_prompt={"image": 10, "video": 0}, |
| 271 | bucket_config={(32, 32, 1): 1.0, (52, 64, 1): 0.0}, |
| 272 | ) |
| 273 | for s in samples: |
| 274 | assert isinstance(s.multi_modal_data, list) |
| 275 | typed_mm = cast(list[dict[str, Any]], s.multi_modal_data) |
| 276 | for it in typed_mm: |
| 277 | assert it.get("type") == "image_url" |
| 278 | |
| 279 | |
| 280 | @pytest.mark.benchmark |
| 281 | def test_random_mm_zero_items(hf_tokenizer: PreTrainedTokenizerBase) -> None: |
| 282 | ds = RandomMultiModalDataset(random_seed=0) |
| 283 | samples = _collect_mm_samples( |
| 284 | ds, |
| 285 | hf_tokenizer, |
| 286 | num_requests=5, |
| 287 | base_items_per_request=0, |
| 288 | num_mm_items_range_ratio=0.0, |
| 289 | limit_mm_per_prompt={"image": 5, "video": 0}, |
| 290 | bucket_config={(32, 32, 1): 1.0}, |
| 291 | ) |
| 292 | for s in samples: |
| 293 | assert s.multi_modal_data == [] |
| 294 | |
| 295 | |
| 296 | @pytest.mark.benchmark |
| 297 | def test_random_mm_num_items_per_prompt(hf_tokenizer: PreTrainedTokenizerBase) -> None: |
| 298 | ds = RandomMultiModalDataset(random_seed=0) |
| 299 | # Fixed number of images per prompt |
| 300 | # set num_mm_items_range_ratio to 0.0 |
| 301 | # TODO: modify video values when video sampling is implemented |
| 302 | samples_fixed_items = _collect_mm_samples( |
| 303 | ds, |
| 304 | hf_tokenizer, |
| 305 | num_requests=5, |
| 306 | base_items_per_request=3, |
| 307 | num_mm_items_range_ratio=0.0, |
| 308 | limit_mm_per_prompt={"image": 3, "video": 0}, |
| 309 | bucket_config={(32, 32, 1): 1.0}, |
| 310 | ) |
| 311 | # Must have 5 requests each with 3 mm items per prompt |
| 312 | assert len(samples_fixed_items) == 5 |
| 313 | for s in samples_fixed_items: |
| 314 | mm_data = cast(list[dict[str, Any]], s.multi_modal_data) |
| 315 | assert len(mm_data) == 3 |
| 316 | for it in mm_data: |
| 317 | assert it.get("type") == "image_url" |
| 318 | |
| 319 | |
| 320 | @pytest.mark.benchmark |
| 321 | def test_random_mm_bucket_config_not_mutated( |
| 322 | hf_tokenizer: PreTrainedTokenizerBase, |
| 323 | ) -> None: |
| 324 | ds = RandomMultiModalDataset(random_seed=0) |
| 325 | # This bucket config is not normalized to sum to 1 |
| 326 | # and has more buckets than requested images |
| 327 | original = {(32, 32, 1): 0.2, (52, 64, 1): 6, (25, 64, 1): 3} |
| 328 | # Keep a snapshot to compare after sampling |
| 329 | snapshot = dict(original) |
| 330 | |
| 331 | _ = _collect_mm_samples( |
| 332 | ds, |
| 333 | hf_tokenizer, |
| 334 | num_requests=4, |
| 335 | base_items_per_request=1, |
| 336 | num_mm_items_range_ratio=0.0, |
| 337 | limit_mm_per_prompt={"image": 1, "video": 0}, |
| 338 | bucket_config=original, |
| 339 | ) |
| 340 | |
| 341 | # Ensure the original dict content is unchanged |
| 342 | assert original == snapshot |
| 343 | |
| 344 | # Vary number of mm items per prompt |
| 345 | # set num_mm_items_range_ratio to 0.5 |
| 346 | samples_varying_items = _collect_mm_samples( |
| 347 | ds, |
| 348 | hf_tokenizer, |
| 349 | num_requests=5, |
| 350 | base_items_per_request=2, |
| 351 | num_mm_items_range_ratio=0.5, |
| 352 | limit_mm_per_prompt={"image": 4, "video": 0}, |
| 353 | bucket_config={(32, 32, 1): 1.0}, |
| 354 | ) |
| 355 | # Must have 5 requests each with less than 4 mm items per prompt |
| 356 | # but at least 1 mm item per prompt |
| 357 | assert len(samples_varying_items) == 5 |
| 358 | for s in samples_varying_items: |
| 359 | mm_data = cast(list[dict[str, Any]], s.multi_modal_data) |
| 360 | assert len(mm_data) <= 4 |
| 361 | assert len(mm_data) >= 1 |
| 362 | for it in mm_data: |
| 363 | assert it.get("type") == "image_url" |
| 364 | |
| 365 | |
| 366 | @pytest.mark.benchmark |
| 367 | def test_random_mm_video_sampling(hf_tokenizer: PreTrainedTokenizerBase) -> None: |
| 368 | """Test video sampling functionality in RandomMultiModalDataset.""" |
| 369 | ds = RandomMultiModalDataset(random_seed=42) |
| 370 | |
| 371 | # Test with video bucket configuration |
| 372 | bucket_config = { |
| 373 | (64, 64, 1): 0.3, # Images |
| 374 | (64, 64, 8): 0.7, # Videos |
| 375 | } |
| 376 | |
| 377 | limit_mm_per_prompt = {"image": 2, "video": 2} |
| 378 | |
| 379 | samples = _collect_mm_samples( |
| 380 | ds, |
| 381 | hf_tokenizer, |
| 382 | num_requests=5, |
| 383 | base_items_per_request=1, |
| 384 | num_mm_items_range_ratio=0.0, |
| 385 | limit_mm_per_prompt=limit_mm_per_prompt, |
| 386 | bucket_config=bucket_config, |
| 387 | ) |
| 388 | |
| 389 | assert len(samples) == 5 |
| 390 | |
| 391 | # Check that we have both images and videos |
| 392 | video_count = 0 |
| 393 | image_count = 0 |
| 394 | |
| 395 | for s in samples: |
| 396 | mm_data = cast(list[dict[str, Any]], s.multi_modal_data) |
| 397 | assert len(mm_data) == 1 |
| 398 | |
| 399 | item = mm_data[0] |
| 400 | if item.get("type") == "video_url": |
| 401 | video_count += 1 |
| 402 | # Verify video URL format |
| 403 | url = item.get("video_url", {}).get("url", "") |
| 404 | assert url.startswith("data:video/mp4;base64,") |
| 405 | elif item.get("type") == "image_url": |
| 406 | image_count += 1 |
| 407 | # Verify image URL format |
| 408 | url = item.get("image_url", {}).get("url", "") |
| 409 | assert url.startswith("data:image/jpeg;base64,") |
| 410 | |
| 411 | # Should have some videos due to 0.7 probability |
| 412 | assert video_count > 0 |
| 413 | assert image_count > 0 |
| 414 | |
| 415 | |
| 416 | @pytest.mark.benchmark |
| 417 | def test_random_mm_video_only_sampling(hf_tokenizer: PreTrainedTokenizerBase) -> None: |
| 418 | """Test sampling with only video buckets.""" |
| 419 | ds = RandomMultiModalDataset(random_seed=42) |
| 420 | |
| 421 | bucket_config = { |
| 422 | (64, 64, 8): 1.0, # Only videos |
| 423 | } |
| 424 | |
| 425 | limit_mm_per_prompt = {"image": 0, "video": 1} |
| 426 | |
| 427 | samples = _collect_mm_samples( |
| 428 | ds, |
| 429 | hf_tokenizer, |
| 430 | num_requests=3, |
| 431 | base_items_per_request=1, |
| 432 | num_mm_items_range_ratio=0.0, |
| 433 | limit_mm_per_prompt=limit_mm_per_prompt, |
| 434 | bucket_config=bucket_config, |
| 435 | ) |
| 436 | |
| 437 | assert len(samples) == 3 |
| 438 | |
| 439 | for s in samples: |
| 440 | mm_data = cast(list[dict[str, Any]], s.multi_modal_data) |
| 441 | assert len(mm_data) == 1 |
| 442 | |
| 443 | item = mm_data[0] |
| 444 | assert item.get("type") == "video_url" |
| 445 | url = item.get("video_url", {}).get("url", "") |
| 446 | assert url.startswith("data:video/mp4;base64,") |
| 447 | |
| 448 | |
| 449 | @pytest.mark.benchmark |
| 450 | def test_random_mm_video_deterministic_sampling( |
| 451 | hf_tokenizer: PreTrainedTokenizerBase, |
| 452 | ) -> None: |
| 453 | """Test that video sampling is deterministic with same seed.""" |
| 454 | seed = 123 |
| 455 | ds_a = RandomMultiModalDataset(random_seed=seed) |
| 456 | ds_b = RandomMultiModalDataset(random_seed=seed) |
| 457 | |
| 458 | bucket_config = { |
| 459 | (64, 64, 8): 1.0, # Only videos |
| 460 | } |
| 461 | |
| 462 | limit_mm_per_prompt = {"image": 0, "video": 1} |
| 463 | |
| 464 | a = _collect_mm_samples( |
| 465 | ds_a, |
| 466 | hf_tokenizer, |
| 467 | num_requests=3, |
| 468 | base_items_per_request=1, |
| 469 | num_mm_items_range_ratio=0.0, |
| 470 | limit_mm_per_prompt=limit_mm_per_prompt, |
| 471 | bucket_config=bucket_config, |
| 472 | ) |
| 473 | |
| 474 | b = _collect_mm_samples( |
| 475 | ds_b, |
| 476 | hf_tokenizer, |
| 477 | num_requests=3, |
| 478 | base_items_per_request=1, |
| 479 | num_mm_items_range_ratio=0.0, |
| 480 | limit_mm_per_prompt=limit_mm_per_prompt, |
| 481 | bucket_config=bucket_config, |
| 482 | ) |
| 483 | |
| 484 | fa = [_mm_fingerprint_sample(s) for s in a] |
| 485 | fb = [_mm_fingerprint_sample(s) for s in b] |
| 486 | assert fa == fb |