| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import numpy as np |
| 5 | import pytest |
| 6 | import torch |
| 7 | |
| 8 | from tests.v1.sample.utils import create_allowed_token_ids |
| 9 | from vllm.platforms import current_platform |
| 10 | from vllm.utils.platform_utils import is_pin_memory_available |
| 11 | from vllm.utils.torch_utils import make_tensor_with_pad |
| 12 | from vllm.v1.sample.logits_processor import LogitsProcessors |
| 13 | from vllm.v1.sample.metadata import SamplingMetadata |
| 14 | from vllm.v1.sample.sampler import Sampler |
| 15 | |
| 16 | PIN_MEMORY_AVAILABLE = is_pin_memory_available() |
| 17 | MAX_NUM_REQS = 256 |
| 18 | VOCAB_SIZE = 1024 |
| 19 | NUM_OUTPUT_TOKENS = 20 |
| 20 | DEVICE_TYPE = current_platform.device_type |
| 21 | DEVICES = [ |
| 22 | f"{DEVICE_TYPE}:{i}" |
| 23 | for i in range(1 if current_platform.device_count() == 1 else 2) |
| 24 | ] |
| 25 | MAX_NUM_PROMPT_TOKENS = 64 |
| 26 | |
| 27 | |
| 28 | def _create_fake_logits(batch_size: int, vocab_size: int) -> torch.Tensor: |
| 29 | fake_logits = torch.full((batch_size, vocab_size), 1e-2, dtype=torch.float) |
| 30 | return fake_logits |
| 31 | |
| 32 | |
| 33 | def _create_penalty_tensor( |
| 34 | batch_size: int, penalty_value: float, device: torch.device |
| 35 | ) -> torch.Tensor: |
| 36 | return torch.full( |
| 37 | (batch_size,), fill_value=penalty_value, dtype=torch.float, device=device |
| 38 | ) |
| 39 | |
| 40 | |
| 41 | def _create_prompt_tokens_tensor( |
| 42 | prompt_token_ids: list[list[int]], |
| 43 | vocab_size: int, |
| 44 | device: torch.device, |
| 45 | ) -> torch.Tensor: |
| 46 | return make_tensor_with_pad( |
| 47 | prompt_token_ids, |
| 48 | pad=vocab_size, |
| 49 | device=device, |
| 50 | dtype=torch.int64, |
| 51 | pin_memory=False, |
| 52 | ) |
| 53 | |
| 54 | |
| 55 | def _create_bad_words_token_ids( |
| 56 | batch_size: int, |
| 57 | vocab_size: int, |
| 58 | bad_words_lengths: tuple[int, ...], |
| 59 | ) -> dict[int, list[list[int]]]: |
| 60 | bad_words_token_ids = {} |
| 61 | for batch_idx in range(batch_size): |
| 62 | token_ids_single_batch = [] |
| 63 | for bad_words_length in bad_words_lengths: |
| 64 | token_ids = np.random.choice( |
| 65 | vocab_size, size=bad_words_length, replace=True |
| 66 | ).tolist() |
| 67 | token_ids_single_batch.append(token_ids) |
| 68 | bad_words_token_ids[batch_idx] = token_ids_single_batch |
| 69 | if batch_size >= 2: |
| 70 | # Test no bad_words for some batch |
| 71 | no_bad_words_batch_idx = np.random.choice(batch_size) |
| 72 | bad_words_token_ids.pop(no_bad_words_batch_idx, None) |
| 73 | return bad_words_token_ids |
| 74 | |
| 75 | |
| 76 | # Returns all last tokens of bad word sequences that share the same prefix |
| 77 | # as `given_prefix` (excluding the last token). |
| 78 | def _collect_suffixes_with_same_prefix( |
| 79 | given_prefix: list[int], bad_words_token_ids: list[list[int]] |
| 80 | ) -> list[int]: |
| 81 | return [bwt[-1] for bwt in bad_words_token_ids if bwt[:-1] == given_prefix] |
| 82 | |
| 83 | |
| 84 | # generate a valid token id that is not in bad_words_token_ids |
| 85 | def _generate_valid_token_id( |
| 86 | bad_words_token_ids: list[list[int]], vocab_size: int |
| 87 | ) -> int: |
| 88 | forbidden_start_tokens = set() |
| 89 | for bad_word in bad_words_token_ids: |
| 90 | forbidden_start_tokens.add(bad_word[0]) |
| 91 | # Get a safe token that's not in forbidden starts |
| 92 | safe_token_candidates = list(set(range(vocab_size)) - forbidden_start_tokens) |
| 93 | # Pick a random safe token |
| 94 | return np.random.choice(safe_token_candidates) |
| 95 | |
| 96 | |
| 97 | def _update_output_token_ids_for_bad_words( |
| 98 | metadata: SamplingMetadata, vocab_size: int |
| 99 | ) -> dict[int, list[int]]: |
| 100 | bad_words_last_tokens = {} |
| 101 | for batch_idx, bad_words_token_ids in metadata.bad_words_token_ids.items(): |
| 102 | output_token_ids = metadata.output_token_ids[batch_idx] |
| 103 | bad_words_last_token: list[int] = [] |
| 104 | for i, bad_word_token_ids in enumerate(bad_words_token_ids): |
| 105 | if len(bad_word_token_ids) == 1: |
| 106 | # Single token id always affects logits |
| 107 | bad_words_last_token.append(bad_word_token_ids[0]) |
| 108 | else: |
| 109 | prefix_length = len(bad_word_token_ids) - 1 |
| 110 | has_bad_words = np.random.choice([True, False]) |
| 111 | if has_bad_words: |
| 112 | prefix = bad_word_token_ids[:-1] |
| 113 | output_token_ids[-prefix_length:] = prefix |
| 114 | # Collect all last tokens from other bad words |
| 115 | # that share this prefix |
| 116 | bad_words_last_token.extend( |
| 117 | _collect_suffixes_with_same_prefix(prefix, bad_words_token_ids) |
| 118 | ) |
| 119 | break # Maximum one update to output_token_ids |
| 120 | else: # Make sure no accidental match to bad words |
| 121 | output_token_ids[-1] = _generate_valid_token_id( |
| 122 | bad_words_token_ids, vocab_size |
| 123 | ) |
| 124 | bad_words_last_tokens[batch_idx] = bad_words_last_token |
| 125 | return bad_words_last_tokens |
| 126 | |
| 127 | |
| 128 | def _create_default_sampling_metadata( |
| 129 | num_output_tokens: int, |
| 130 | batch_size: int, |
| 131 | vocab_size: int, |
| 132 | device: torch.device, |
| 133 | ) -> SamplingMetadata: |
| 134 | output_token_ids: list[list[int]] = [] |
| 135 | prompt_token_ids: list[list[int]] = [] |
| 136 | for _ in range(batch_size): |
| 137 | output_token_ids.append( |
| 138 | np.random.randint(0, vocab_size, size=num_output_tokens).tolist() |
| 139 | ) |
| 140 | prompt_token_ids.append( |
| 141 | np.random.randint( |
| 142 | 0, vocab_size, size=np.random.randint(1, MAX_NUM_PROMPT_TOKENS) |
| 143 | ).tolist() |
| 144 | ) |
| 145 | fake_sampling_metadata = SamplingMetadata( |
| 146 | temperature=torch.full((batch_size,), 0.0), |
| 147 | all_greedy=True, |
| 148 | all_random=False, |
| 149 | top_p=None, |
| 150 | top_k=None, |
| 151 | generators={}, |
| 152 | max_num_logprobs=0, |
| 153 | prompt_token_ids=_create_prompt_tokens_tensor( |
| 154 | prompt_token_ids, vocab_size, device |
| 155 | ), |
| 156 | output_token_ids=output_token_ids, |
| 157 | spec_token_ids=[[] for _ in range(batch_size)], |
| 158 | frequency_penalties=_create_penalty_tensor(batch_size, 0.0, device), |
| 159 | presence_penalties=_create_penalty_tensor(batch_size, 0.0, device), |
| 160 | repetition_penalties=_create_penalty_tensor(batch_size, 1.0, device), |
| 161 | no_penalties=True, |
| 162 | allowed_token_ids_mask=None, |
| 163 | bad_words_token_ids={}, |
| 164 | logitsprocs=LogitsProcessors(), |
| 165 | ) |
| 166 | return fake_sampling_metadata |
| 167 | |
| 168 | |
| 169 | def _create_weighted_output_token_list( |
| 170 | batch_size: int, vocab_size: int |
| 171 | ) -> tuple[list[list[int]], list[list[int]]]: |
| 172 | """Creates an output token list where each token occurs a distinct |
| 173 | number of times. |
| 174 | |
| 175 | For each batch, a random subset of token IDs is selected from the |
| 176 | vocabulary. The selected tokens are then added to the output token |
| 177 | list, each with a different frequency. |
| 178 | |
| 179 | Returns: |
| 180 | tuple[list[list[int]], list[list[int]]]: |
| 181 | - The first element is the output token list, where each sublist |
| 182 | corresponds to a batch and contains tokens with weighted |
| 183 | frequencies. |
| 184 | - The second element is a list of distinct token IDs for each |
| 185 | batch, ordered by their frequency in the corresponding output |
| 186 | list. |
| 187 | |
| 188 | """ |
| 189 | output_token_ids: list[list[int]] = [] |
| 190 | sorted_token_ids_in_output: list[list[int]] = [] |
| 191 | for _ in range(batch_size): |
| 192 | distinct_token_ids = np.random.choice( |
| 193 | vocab_size, size=np.random.randint(1, 10), replace=False |
| 194 | ).tolist() |
| 195 | sorted_token_ids_in_output.append(distinct_token_ids) |
| 196 | output_token_ids_for_batch = [] |
| 197 | for index, token_id in enumerate(distinct_token_ids): |
| 198 | output_token_ids_for_batch.extend([token_id for _ in range(index + 1)]) |
| 199 | output_token_ids.append(output_token_ids_for_batch) |
| 200 | return output_token_ids, sorted_token_ids_in_output |
| 201 | |
| 202 | |
| 203 | @pytest.mark.parametrize("device", DEVICES) |
| 204 | @pytest.mark.parametrize("batch_size", [1, 2, 32]) |
| 205 | @pytest.mark.parametrize("presence_penalty", [-2.0, 2.0]) |
| 206 | def test_sampler_presence_penalty( |
| 207 | device: str, batch_size: int, presence_penalty: float |
| 208 | ): |
| 209 | """Test to verify that if presence penalty is enabled then tokens |
| 210 | are penalized as per their presence in the existing output. |
| 211 | """ |
| 212 | torch.set_default_device(device) |
| 213 | # Create fake logits where each token is assigned the same |
| 214 | # logit value. |
| 215 | fake_logits = _create_fake_logits(batch_size, VOCAB_SIZE) |
| 216 | sampling_metadata = _create_default_sampling_metadata( |
| 217 | NUM_OUTPUT_TOKENS, batch_size, VOCAB_SIZE, torch.device(device) |
| 218 | ) |
| 219 | output_token_ids = sampling_metadata.output_token_ids |
| 220 | sampling_metadata.presence_penalties = _create_penalty_tensor( |
| 221 | batch_size, presence_penalty, torch.device(device) |
| 222 | ) |
| 223 | sampling_metadata.no_penalties = False |
| 224 | sampler = Sampler() |
| 225 | logits = sampler.apply_penalties( |
| 226 | fake_logits, sampling_metadata, sampling_metadata.output_token_ids |
| 227 | ) |
| 228 | logits = logits.cpu() |
| 229 | for batch_idx in range(batch_size): |
| 230 | # Since all tokens initially have the same logits, the non-penalized |
| 231 | # token ID will be the one with the highest logit value, while the |
| 232 | # penalized token ID will be the one with the lowest logit value. |
| 233 | non_penalized_token_id = logits[batch_idx].argmax().item() |
| 234 | penalized_token_id = logits[batch_idx].argmin().item() |
| 235 | if presence_penalty > 0: |
| 236 | # If `presence_penalty` is set to a value greater than 0, it |
| 237 | # indicates a preference for new tokens over those already |
| 238 | # present in the output. |
| 239 | # Verify that the penalized token ID exists in the output, while the |
| 240 | # non-penalized token ID does not. |
| 241 | assert penalized_token_id in output_token_ids[batch_idx] |
| 242 | assert non_penalized_token_id not in output_token_ids[batch_idx] |
| 243 | elif presence_penalty < 0: |
| 244 | # If `presence_penalty` is set to a value less than 0, it indicates |
| 245 | # a preference for existing tokens over new ones. Verify that the |
| 246 | # non-penalized token ID exists in the output, while the penalized |
| 247 | # token ID does not. |
| 248 | assert non_penalized_token_id in output_token_ids[batch_idx] |
| 249 | assert penalized_token_id not in output_token_ids[batch_idx] |
| 250 | |
| 251 | |
| 252 | @pytest.mark.parametrize("device", DEVICES) |
| 253 | @pytest.mark.parametrize("batch_size", [1, 2, 32]) |
| 254 | @pytest.mark.parametrize("frequency_penalty", [-2.0, 2.0]) |
| 255 | def test_sampler_frequency_penalty( |
| 256 | device: str, batch_size: int, frequency_penalty: float |
| 257 | ): |
| 258 | """Test to verify that if frequency penalty is enabled then tokens are |
| 259 | penalized as per their frequency of occurrence. |
| 260 | """ |
| 261 | torch.set_default_device(device) |
| 262 | # Create fake logits where each token is assigned the same |
| 263 | # logit value. |
| 264 | fake_logits = _create_fake_logits(batch_size, VOCAB_SIZE) |
| 265 | sampling_metadata = _create_default_sampling_metadata( |
| 266 | NUM_OUTPUT_TOKENS, batch_size, VOCAB_SIZE, torch.device(device) |
| 267 | ) |
| 268 | sampling_metadata.frequency_penalties = _create_penalty_tensor( |
| 269 | batch_size, frequency_penalty, torch.device(device) |
| 270 | ) |
| 271 | output_token_ids, sorted_token_ids_in_output = _create_weighted_output_token_list( |
| 272 | batch_size, |
| 273 | VOCAB_SIZE, |
| 274 | ) |
| 275 | sampling_metadata.output_token_ids = output_token_ids |
| 276 | sampling_metadata.no_penalties = False |
| 277 | sampler = Sampler() |
| 278 | logits = sampler.apply_penalties( |
| 279 | fake_logits, sampling_metadata, sampling_metadata.output_token_ids |
| 280 | ) |
| 281 | logits = logits.cpu() |
| 282 | for batch_idx in range(batch_size): |
| 283 | non_penalized_token_id = logits[batch_idx].argmax().item() |
| 284 | penalized_token_id = logits[batch_idx].argmin().item() |
| 285 | distinct_sorted_token_ids_in_output = sorted_token_ids_in_output[batch_idx] |
| 286 | most_frequent_token_id = distinct_sorted_token_ids_in_output[ |
| 287 | len(distinct_sorted_token_ids_in_output) - 1 |
| 288 | ] |
| 289 | if frequency_penalty > 0: |
| 290 | # If `frequency_penalty` is set to > 0, it indicates |
| 291 | # a preference for new tokens over existing ones. Verify that the |
| 292 | # non-penalized token ID is not present in the output, while the |
| 293 | # most penalized token is the one that occurs most frequently in |
| 294 | # the output. |
| 295 | assert non_penalized_token_id not in distinct_sorted_token_ids_in_output |
| 296 | assert penalized_token_id == most_frequent_token_id |
| 297 | elif frequency_penalty < 0: |
| 298 | # If `frequency_penalty` is set to < 0, it indicates |
| 299 | # a preference for existing tokens over new ones. Verify that the |
| 300 | # non-penalized token ID is the one that occurs most frequently |
| 301 | # in the output, while the penalized token ID is one that has not |
| 302 | # yet appeared. |
| 303 | assert non_penalized_token_id == most_frequent_token_id |
| 304 | assert penalized_token_id not in distinct_sorted_token_ids_in_output |
| 305 | |
| 306 | |
| 307 | @pytest.mark.parametrize("device", DEVICES) |
| 308 | @pytest.mark.parametrize("batch_size", [1, 2, 32]) |
| 309 | @pytest.mark.parametrize("repetition_penalty", [0.1, 1.9]) |
| 310 | def test_sampler_repetition_penalty( |
| 311 | device: str, batch_size: int, repetition_penalty: float |
| 312 | ): |
| 313 | """Test to verify that when the repetition penalty is enabled, tokens |
| 314 | are penalized based on their presence in the prompt or the existing |
| 315 | output. |
| 316 | """ |
| 317 | torch.set_default_device(device) |
| 318 | # Create fake logits where each token is assigned the same |
| 319 | # logit value. |
| 320 | fake_logits = _create_fake_logits(batch_size, VOCAB_SIZE) |
| 321 | sampling_metadata = _create_default_sampling_metadata( |
| 322 | NUM_OUTPUT_TOKENS, batch_size, VOCAB_SIZE, torch.device(device) |
| 323 | ) |
| 324 | sampling_metadata.repetition_penalties = _create_penalty_tensor( |
| 325 | batch_size, repetition_penalty, torch.device(device) |
| 326 | ) |
| 327 | sampling_metadata.no_penalties = False |
| 328 | sampler = Sampler() |
| 329 | logits = sampler.apply_penalties( |
| 330 | fake_logits, sampling_metadata, sampling_metadata.output_token_ids |
| 331 | ) |
| 332 | logits = logits.cpu() |
| 333 | for batch_idx in range(batch_size): |
| 334 | non_penalized_token_id = logits[batch_idx].argmax().item() |
| 335 | penalized_token_id = logits[batch_idx].argmin().item() |
| 336 | assert sampling_metadata.prompt_token_ids is not None |
| 337 | prompt_tokens = sampling_metadata.prompt_token_ids[batch_idx][:].tolist() |
| 338 | output_tokens = sampling_metadata.output_token_ids[batch_idx] |
| 339 | if repetition_penalty > 1.0: |
| 340 | # If `repetition_penalty` > 1.0, verify that the non-penalized |
| 341 | # token ID has not been seen before, while the penalized token ID |
| 342 | # exists either in the prompt or the output. |
| 343 | assert ( |
| 344 | non_penalized_token_id not in prompt_tokens |
| 345 | and non_penalized_token_id not in output_tokens |
| 346 | ) |
| 347 | assert ( |
| 348 | penalized_token_id in prompt_tokens |
| 349 | or penalized_token_id in output_tokens |
| 350 | ) |
| 351 | elif repetition_penalty < 1.0: |
| 352 | # If `repetition_penalty` < 1.0, verify that the penalized |
| 353 | # token ID has not been seen before, while the non-penalized |
| 354 | # token ID exists either in the prompt or the output. |
| 355 | assert ( |
| 356 | penalized_token_id not in prompt_tokens |
| 357 | and penalized_token_id not in output_tokens |
| 358 | ) |
| 359 | assert ( |
| 360 | non_penalized_token_id in prompt_tokens |
| 361 | or non_penalized_token_id in output_tokens |
| 362 | ) |
| 363 | |
| 364 | |
| 365 | @pytest.mark.parametrize("device", DEVICES) |
| 366 | @pytest.mark.parametrize("batch_size", [1, 2, 32]) |
| 367 | @pytest.mark.parametrize("num_allowed_token_ids", [0, 1, 2]) |
| 368 | def test_sampler_allowed_token_ids( |
| 369 | device: str, batch_size: int, num_allowed_token_ids: int |
| 370 | ): |
| 371 | """Test to verify that when the repetition penalty is enabled, tokens |
| 372 | are penalized based on their presence in the prompt or the existing |
| 373 | output. |
| 374 | """ |
| 375 | torch.set_default_device(device) |
| 376 | # Create fake logits where each token is assigned the same |
| 377 | # logit value. |
| 378 | fake_logits = _create_fake_logits(batch_size, VOCAB_SIZE) |
| 379 | sampling_metadata = _create_default_sampling_metadata( |
| 380 | NUM_OUTPUT_TOKENS, batch_size, VOCAB_SIZE, torch.device(device) |
| 381 | ) |
| 382 | mask = create_allowed_token_ids( |
| 383 | batch_size=batch_size, |
| 384 | vocab_size=VOCAB_SIZE, |
| 385 | num_allowed_token_ids=num_allowed_token_ids, |
| 386 | device=device, |
| 387 | ) |
| 388 | sampling_metadata.allowed_token_ids_mask = mask |
| 389 | sampler = Sampler() |
| 390 | logits = sampler.apply_logits_processors( |
| 391 | fake_logits, sampling_metadata, predict_bonus_token=False |
| 392 | ) |
| 393 | logits = logits.cpu() |
| 394 | for batch_idx in range(batch_size): |
| 395 | logits_for_req = logits[batch_idx] |
| 396 | if batch_idx % 2 == 1: |
| 397 | assert torch.all(logits_for_req != -float("inf")) |
| 398 | continue |
| 399 | for token_id in range(VOCAB_SIZE): |
| 400 | start = min(batch_idx, VOCAB_SIZE - 1) |
| 401 | end = min(batch_idx + num_allowed_token_ids, VOCAB_SIZE - 1) |
| 402 | if token_id >= start and token_id < end: |
| 403 | assert logits_for_req[token_id] == -float("inf"), ( |
| 404 | f"{batch_idx}, {token_id}" |
| 405 | ) |
| 406 | else: |
| 407 | assert logits_for_req[token_id] != -float("inf") |
| 408 | |
| 409 | |
| 410 | @pytest.mark.parametrize("device", DEVICES) |
| 411 | @pytest.mark.parametrize("batch_size", [1, 2, 32]) |
| 412 | @pytest.mark.parametrize("bad_words_lengths", [(1,), (1, 3), (2, 2)]) |
| 413 | def test_sampler_bad_words( |
| 414 | device: str, batch_size: int, bad_words_lengths: tuple[int, ...] |
| 415 | ): |
| 416 | """Test to verify that when the bad words restriction is present, tokens |
| 417 | are penalized based on their match with the bad words. |
| 418 | """ |
| 419 | torch.set_default_device(device) |
| 420 | # Create fake logits where each token is assigned the same |
| 421 | # logit value. |
| 422 | fake_logits = _create_fake_logits(batch_size, VOCAB_SIZE) |
| 423 | sampling_metadata = _create_default_sampling_metadata( |
| 424 | NUM_OUTPUT_TOKENS, batch_size, VOCAB_SIZE, torch.device(device) |
| 425 | ) |
| 426 | sampling_metadata.bad_words_token_ids = _create_bad_words_token_ids( |
| 427 | batch_size, VOCAB_SIZE, bad_words_lengths |
| 428 | ) |
| 429 | bad_words_last_tokens = _update_output_token_ids_for_bad_words( |
| 430 | sampling_metadata, VOCAB_SIZE |
| 431 | ) |
| 432 | sampler = Sampler() |
| 433 | logits = sampler.apply_logits_processors( |
| 434 | fake_logits, sampling_metadata, predict_bonus_token=False |
| 435 | ) |
| 436 | logits = logits.cpu() |
| 437 | for batch_idx in range(batch_size): |
| 438 | logits_for_req = logits[batch_idx] |
| 439 | for token_id in range(VOCAB_SIZE): |
| 440 | if ( |
| 441 | batch_idx in bad_words_last_tokens |
| 442 | and token_id in bad_words_last_tokens[batch_idx] |
| 443 | ): |
| 444 | assert logits_for_req[token_id] == -float("inf") |
| 445 | else: |
| 446 | assert logits_for_req[token_id] != -float("inf") |