| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | import sys |
| 4 | from abc import ABC, abstractmethod |
| 5 | |
| 6 | import tokenizers.decoders |
| 7 | from tokenizers import Tokenizer |
| 8 | from transformers import TokenizersBackend |
| 9 | |
| 10 | from vllm.logger import init_logger |
| 11 | from vllm.tokenizers import TokenizerLike |
| 12 | from vllm.tokenizers.detokenizer_utils import ( |
| 13 | convert_prompt_ids_to_tokens, |
| 14 | detokenize_incrementally, |
| 15 | ) |
| 16 | from vllm.utils import length_from_prompt_token_ids_or_embeds |
| 17 | from vllm.v1.engine import EngineCoreRequest |
| 18 | |
| 19 | logger = init_logger(__name__) |
| 20 | |
| 21 | # Error string from https://github.com/huggingface/tokenizers/blob/909fdde2a4ffedd9295206f705eb612be2a91b12/tokenizers/src/tokenizer/mod.rs#L1042 |
| 22 | INVALID_PREFIX_ERR_MSG = "Invalid prefix encountered" |
| 23 | |
| 24 | |
| 25 | def uses_fast_detokenizer(tokenizer: TokenizerLike) -> bool: |
| 26 | """Whether the engine detokenizes with `FastIncrementalDetokenizer`.""" |
| 27 | return isinstance(tokenizer, TokenizersBackend) |
| 28 | |
| 29 | |
| 30 | class IncrementalDetokenizer: |
| 31 | def __init__(self): |
| 32 | self.token_ids: list[int] = [] |
| 33 | |
| 34 | @property |
| 35 | def output_token_ids(self) -> list[int]: |
| 36 | return self.token_ids |
| 37 | |
| 38 | def num_output_tokens(self) -> int: |
| 39 | return len(self.token_ids) |
| 40 | |
| 41 | def update(self, new_token_ids: list[int], stop_terminated: bool) -> str | None: |
| 42 | self.token_ids.extend(new_token_ids) |
| 43 | return None |
| 44 | |
| 45 | def get_next_output_text(self, finished: bool, delta: bool) -> str: |
| 46 | return "" |
| 47 | |
| 48 | @classmethod |
| 49 | def from_new_request( |
| 50 | cls, |
| 51 | tokenizer: TokenizerLike | None, |
| 52 | request: EngineCoreRequest, |
| 53 | ) -> "IncrementalDetokenizer": |
| 54 | assert request.sampling_params is not None |
| 55 | |
| 56 | if tokenizer is None: |
| 57 | # No tokenizer => skipping detokenization. |
| 58 | return IncrementalDetokenizer() |
| 59 | |
| 60 | if uses_fast_detokenizer(tokenizer): |
| 61 | # Fast tokenizer => use tokenizers library DecodeStream. |
| 62 | return FastIncrementalDetokenizer(tokenizer, request) |
| 63 | |
| 64 | # Fall back to slow python-based incremental detokenization. |
| 65 | return SlowIncrementalDetokenizer(tokenizer, request) |
| 66 | |
| 67 | |
| 68 | class BaseIncrementalDetokenizer(IncrementalDetokenizer, ABC): |
| 69 | def __init__(self, request: EngineCoreRequest): |
| 70 | super().__init__() |
| 71 | |
| 72 | # Stop strings |
| 73 | params = request.sampling_params |
| 74 | assert params is not None |
| 75 | if params.stop is None: |
| 76 | self.stop = [] |
| 77 | elif isinstance(params.stop, str): |
| 78 | self.stop = [params.stop] |
| 79 | else: |
| 80 | self.stop = params.stop |
| 81 | self.min_tokens = params.min_tokens |
| 82 | self.include_stop_str_in_output = params.include_stop_str_in_output |
| 83 | |
| 84 | # Number of chars to hold back when stop strings are to be excluded |
| 85 | # from streamed output. |
| 86 | if self.stop and not self.include_stop_str_in_output: |
| 87 | self.stop_buffer_length = max(len(s) for s in self.stop) - 1 |
| 88 | else: |
| 89 | self.stop_buffer_length = 0 |
| 90 | self._last_output_text_offset: int = 0 |
| 91 | |
| 92 | # Generation data |
| 93 | self.output_text = "" |
| 94 | |
| 95 | def update(self, new_token_ids: list[int], stop_terminated: bool) -> str | None: |
| 96 | """Update RequestState for the request_id by: |
| 97 | 1) Detokenize the new token ids incrementally. |
| 98 | 2) Evaluate stop criteria. |
| 99 | |
| 100 | Return matched stop string or None. |
| 101 | """ |
| 102 | if not new_token_ids: |
| 103 | # Skip detokenization if no new token ids. |
| 104 | return None |
| 105 | |
| 106 | if stop_terminated and not self.include_stop_str_in_output: |
| 107 | # If stop-terminated, exclude last token from detokenization |
| 108 | # based on include_stop_str_in_output parameter. |
| 109 | skipped_stop_token_id = new_token_ids[-1] |
| 110 | new_token_ids = new_token_ids[:-1] |
| 111 | else: |
| 112 | skipped_stop_token_id = None |
| 113 | |
| 114 | # 1) Detokenize the new token ids incrementally. |
| 115 | stop_check_offset = len(self.output_text) |
| 116 | for new_token_id in new_token_ids: |
| 117 | self.token_ids.append(new_token_id) |
| 118 | self.output_text += self.decode_next(new_token_id) |
| 119 | # Support min_tokens, see https://github.com/vllm-project/vllm/pull/22014 |
| 120 | if self.min_tokens and self.num_output_tokens() <= self.min_tokens: |
| 121 | stop_check_offset = len(self.output_text) |
| 122 | |
| 123 | if skipped_stop_token_id is not None: |
| 124 | # Cleanup after skipping detokenization. |
| 125 | self.token_ids.append(skipped_stop_token_id) |
| 126 | |
| 127 | # 2) Evaluate stop strings. |
| 128 | stop_string = None |
| 129 | if self.stop and self.num_output_tokens() > self.min_tokens: |
| 130 | stop = check_stop_strings( |
| 131 | output_text=self.output_text, |
| 132 | new_char_count=len(self.output_text) - stop_check_offset, |
| 133 | stop=self.stop, |
| 134 | include_in_output=self.include_stop_str_in_output, |
| 135 | ) |
| 136 | if stop is not None: |
| 137 | stop_string, truncate_to = stop |
| 138 | if truncate_to != -1: |
| 139 | self.output_text = self.output_text[:truncate_to] |
| 140 | |
| 141 | return stop_string |
| 142 | |
| 143 | @abstractmethod |
| 144 | def decode_next(self, next_token_id: int) -> str: |
| 145 | raise NotImplementedError |
| 146 | |
| 147 | def get_next_output_text(self, finished: bool, delta: bool) -> str: |
| 148 | """If delta is True, only new text since the last call to |
| 149 | this method is returned""" |
| 150 | # We return the full output text if the sequence is finished. |
| 151 | buffer_length = 0 if finished else self.stop_buffer_length |
| 152 | if not delta: |
| 153 | if not buffer_length: |
| 154 | return self.output_text |
| 155 | return self.output_text[:-buffer_length] |
| 156 | |
| 157 | length = len(self.output_text) - buffer_length |
| 158 | last_offset = self._last_output_text_offset |
| 159 | if last_offset < length: |
| 160 | self._last_output_text_offset = length |
| 161 | return self.output_text[last_offset:length] |
| 162 | return "" |
| 163 | |
| 164 | |
| 165 | class FastIncrementalDetokenizer(BaseIncrementalDetokenizer): |
| 166 | def __init__(self, tokenizer: TokenizersBackend, request: EngineCoreRequest): |
| 167 | super().__init__(request) |
| 168 | |
| 169 | sampling_params = request.sampling_params |
| 170 | assert sampling_params is not None |
| 171 | |
| 172 | self.request_id = request.request_id |
| 173 | self.skip_special_tokens = sampling_params.skip_special_tokens |
| 174 | |
| 175 | self.tokenizer: Tokenizer = tokenizer._tokenizer |
| 176 | |
| 177 | # Use native prefill to prime the decode stream with prompt tokens. |
| 178 | # Look up DecodeStream on the module so backend patches (e.g. the |
| 179 | # fastokens shim that replaces ``tokenizers.decoders.DecodeStream``) |
| 180 | # are honored regardless of import order. |
| 181 | self.stream = tokenizers.decoders.DecodeStream( |
| 182 | ids=request.prompt_token_ids, |
| 183 | skip_special_tokens=self.skip_special_tokens, |
| 184 | ) |
| 185 | |
| 186 | self.spaces_between_special_tokens = ( |
| 187 | sampling_params.skip_special_tokens |
| 188 | or sampling_params.spaces_between_special_tokens |
| 189 | ) |
| 190 | |
| 191 | if not self.spaces_between_special_tokens: |
| 192 | # Store dict of added token ids so that we can suppress |
| 193 | # the spaces between them. |
| 194 | added_token_ids = getattr(self.tokenizer, "added_token_ids", None) |
| 195 | if added_token_ids is None: |
| 196 | self.tokenizer.added_token_ids = added_token_ids = { |
| 197 | tid: tok.content |
| 198 | for tid, tok in self.tokenizer.get_added_tokens_decoder().items() |
| 199 | } |
| 200 | |
| 201 | if added_token_ids: |
| 202 | self.last_special = False |
| 203 | self.added_token_ids = added_token_ids |
| 204 | else: |
| 205 | # No added tokens. |
| 206 | self.spaces_between_special_tokens = True |
| 207 | |
| 208 | def decode_next(self, next_token_id: int) -> str: |
| 209 | token = self._protected_step(next_token_id) |
| 210 | |
| 211 | if not self.spaces_between_special_tokens: |
| 212 | special_token = self.added_token_ids.get(next_token_id) |
| 213 | is_special = special_token is not None |
| 214 | if is_special and self.last_special: |
| 215 | # Return raw token string without any prefixed spaces. |
| 216 | token = special_token |
| 217 | self.last_special = is_special |
| 218 | |
| 219 | return token or "" |
| 220 | |
| 221 | def _protected_step(self, next_token_id: int) -> str | None: |
| 222 | try: |
| 223 | token = self.stream.step(self.tokenizer, next_token_id) |
| 224 | except (OverflowError, TypeError): |
| 225 | # Handle rare observed overflow, still to be diagnosed. |
| 226 | # See https://github.com/vllm-project/vllm/issues/21951. |
| 227 | logger.exception("Encountered invalid token id: %r", next_token_id) |
| 228 | token = None |
| 229 | except Exception as e: |
| 230 | if not str(e).startswith(INVALID_PREFIX_ERR_MSG): |
| 231 | raise e |
| 232 | # Recover from edge case where tokenizer can produce non-monotonic, |
| 233 | # invalid UTF-8 output, which breaks the internal state of |
| 234 | # tokenizers' DecodeStream. |
| 235 | # See https://github.com/vllm-project/vllm/issues/17448. |
| 236 | logger.warning( |
| 237 | "Encountered invalid prefix detokenization error" |
| 238 | " for request %s, resetting decode stream.", |
| 239 | self.request_id, |
| 240 | ) |
| 241 | self.stream = tokenizers.decoders.DecodeStream( |
| 242 | skip_special_tokens=self.skip_special_tokens |
| 243 | ) |
| 244 | token = self.stream.step(self.tokenizer, next_token_id) |
| 245 | return token |
| 246 | |
| 247 | |
| 248 | class SlowIncrementalDetokenizer(BaseIncrementalDetokenizer): |
| 249 | def __init__(self, tokenizer: TokenizerLike, request: EngineCoreRequest): |
| 250 | super().__init__(request) |
| 251 | |
| 252 | self.tokenizer = tokenizer |
| 253 | params = request.sampling_params |
| 254 | assert params is not None |
| 255 | |
| 256 | self.prompt_len = length_from_prompt_token_ids_or_embeds( |
| 257 | request.prompt_token_ids, request.prompt_embeds |
| 258 | ) |
| 259 | |
| 260 | # Metadata for incremental detokenization. |
| 261 | if request.prompt_token_ids is not None: |
| 262 | self.tokens, self.prefix_offset, self.read_offset = ( |
| 263 | convert_prompt_ids_to_tokens( |
| 264 | tokenizer=tokenizer, |
| 265 | prompt_ids=request.prompt_token_ids, |
| 266 | skip_special_tokens=params.skip_special_tokens, |
| 267 | ) |
| 268 | ) |
| 269 | else: |
| 270 | # Prompt embedding requests cannot be detokenized, in general. |
| 271 | self.tokens = [""] * self.prompt_len |
| 272 | self.prefix_offset = 0 |
| 273 | self.read_offset = 0 |
| 274 | |
| 275 | self.token_ids.extend(request.prompt_token_ids or [0] * self.prompt_len) |
| 276 | |
| 277 | self.skip_special_tokens = params.skip_special_tokens |
| 278 | self.spaces_between_special_tokens = params.spaces_between_special_tokens |
| 279 | |
| 280 | @property |
| 281 | def output_token_ids(self) -> list[int]: |
| 282 | if self.prompt_len: |
| 283 | return self.token_ids[self.prompt_len :] |
| 284 | return self.token_ids |
| 285 | |
| 286 | def num_output_tokens(self) -> int: |
| 287 | return len(self.token_ids) - self.prompt_len |
| 288 | |
| 289 | def decode_next(self, next_token_id: int) -> str: |
| 290 | new_tokens, decoded_text, prefix_offset, read_offset = detokenize_incrementally( |
| 291 | tokenizer=self.tokenizer, |
| 292 | all_input_ids=self.token_ids, |
| 293 | prev_tokens=self.tokens, |
| 294 | prefix_offset=self.prefix_offset, |
| 295 | read_offset=self.read_offset, |
| 296 | skip_special_tokens=self.skip_special_tokens, |
| 297 | spaces_between_special_tokens=self.spaces_between_special_tokens, |
| 298 | ) |
| 299 | |
| 300 | self.tokens.extend(new_tokens) |
| 301 | self.prefix_offset = prefix_offset |
| 302 | self.read_offset = read_offset |
| 303 | |
| 304 | return decoded_text |
| 305 | |
| 306 | |
| 307 | def check_stop_strings( |
| 308 | output_text: str, |
| 309 | new_char_count: int, |
| 310 | stop: list[str], |
| 311 | include_in_output: bool, |
| 312 | ) -> tuple[str, int] | None: |
| 313 | """Check if any stop strings are matched and truncate sequence |
| 314 | output text accordingly. |
| 315 | |
| 316 | Returns tuple (stop_string, offset) if matched or else None. |
| 317 | |
| 318 | Where stop_string is the matched stop string and offset is the |
| 319 | length to which output_text should be truncated, or -1 for no |
| 320 | truncation. |
| 321 | |
| 322 | When several stop strings match within the newly generated text (for |
| 323 | example when speculative decoding appends multiple tokens in a single |
| 324 | step), the stop string that completes earliest in the text is selected, |
| 325 | so the result matches appending one token at a time. Ties are broken by |
| 326 | stop-list order. |
| 327 | """ |
| 328 | if not new_char_count or not stop: |
| 329 | return None |
| 330 | |
| 331 | best_stop_str: str | None = None |
| 332 | best_stop_index = 0 |
| 333 | best_end = sys.maxsize |
| 334 | for stop_str in stop: |
| 335 | stop_string_len = len(stop_str) |
| 336 | # Avoid searching already-searched text. |
| 337 | stop_index = output_text.find(stop_str, 1 - new_char_count - stop_string_len) |
| 338 | if stop_index == -1: |
| 339 | continue |
| 340 | |
| 341 | # Prefer the stop string that completes earliest in the text. |
| 342 | end = stop_index + stop_string_len |
| 343 | if end < best_end: |
| 344 | best_stop_str = stop_str |
| 345 | best_stop_index = stop_index |
| 346 | best_end = end |
| 347 | |
| 348 | if best_stop_str is None: |
| 349 | return None |
| 350 | |
| 351 | if include_in_output: |
| 352 | # Truncate to end of stop string. |
| 353 | if best_end >= len(output_text): |
| 354 | # No truncation required. |
| 355 | return best_stop_str, -1 |
| 356 | return best_stop_str, best_end |
| 357 | |
| 358 | # Truncate the output text to the beginning of the stop string. |
| 359 | return best_stop_str, best_stop_index |