| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | from collections.abc import MutableSequence |
| 5 | from collections.abc import Sequence as GenericSequence |
| 6 | from dataclasses import dataclass |
| 7 | from typing import Any, Generic |
| 8 | |
| 9 | import numpy as np |
| 10 | import torch |
| 11 | from typing_extensions import TypeVar |
| 12 | |
| 13 | from vllm.logger import init_logger |
| 14 | from vllm.logprobs import PromptLogprobs, SampleLogprobs |
| 15 | from vllm.lora.request import LoRARequest |
| 16 | from vllm.v1.metrics.stats import RequestSpecDecodeMetrics, RequestStateStats |
| 17 | |
| 18 | logger = init_logger(__name__) |
| 19 | |
| 20 | |
| 21 | @dataclass(frozen=True) |
| 22 | class RequestError: |
| 23 | """A request-level error that does not mark the engine as unhealthy.""" |
| 24 | |
| 25 | code: str |
| 26 | message: str |
| 27 | retryable: bool = False |
| 28 | |
| 29 | |
| 30 | @dataclass |
| 31 | class SamplingMask: |
| 32 | """Per-token sampling support sets aligned with completion token IDs. |
| 33 | |
| 34 | Each inner list contains the vocabulary token IDs that survived |
| 35 | top-k / top-p / min-p filtering for the corresponding generated token. |
| 36 | """ |
| 37 | |
| 38 | token_ids: list[list[int]] |
| 39 | |
| 40 | |
| 41 | @dataclass |
| 42 | class CompletionOutput: |
| 43 | """The output data of one completion output of a request. |
| 44 | |
| 45 | Args: |
| 46 | index: The index of the output in the request. |
| 47 | text: The generated output text. |
| 48 | token_ids: The token IDs of the generated output text. |
| 49 | cumulative_logprob: The cumulative log probability of the generated |
| 50 | output text. |
| 51 | logprobs: The log probabilities of the top probability words at each |
| 52 | position if the logprobs are requested. |
| 53 | sampling_mask: The post-processing token support set for each generated |
| 54 | token, if requested. |
| 55 | finish_reason: The reason why the sequence is finished. |
| 56 | stop_reason: The stop string or token id that caused the completion |
| 57 | to stop, None if the completion finished for some other reason |
| 58 | including encountering the EOS token. |
| 59 | lora_request: The LoRA request that was used to generate the output. |
| 60 | spec_decode_metrics: Per-sequence speculative-decoding acceptance metrics, |
| 61 | populated on finish when speculative decoding ran and |
| 62 | ``--per-request-spec-decode-metrics`` is enabled; None otherwise. |
| 63 | Surfaced in the response as ``metrics.speculative_decoding`` for |
| 64 | single-sequence (``n == 1``) requests. |
| 65 | |
| 66 | """ |
| 67 | |
| 68 | index: int |
| 69 | text: str |
| 70 | token_ids: GenericSequence[int] |
| 71 | cumulative_logprob: float | None |
| 72 | logprobs: SampleLogprobs | None |
| 73 | routed_experts: np.ndarray | None = None # [seq_len,layer_num,topk] |
| 74 | finish_reason: str | None = None |
| 75 | stop_reason: int | str | None = None |
| 76 | lora_request: LoRARequest | None = None |
| 77 | sampling_mask: SamplingMask | None = None |
| 78 | spec_decode_metrics: RequestSpecDecodeMetrics | None = None |
| 79 | |
| 80 | def finished(self) -> bool: |
| 81 | return self.finish_reason is not None |
| 82 | |
| 83 | def __repr__(self) -> str: |
| 84 | return ( |
| 85 | f"CompletionOutput(index={self.index}, " |
| 86 | f"text={self.text!r}, " |
| 87 | f"token_ids={self.token_ids}, " |
| 88 | f"routed_experts={self.routed_experts}, " |
| 89 | f"sampling_mask={self.sampling_mask}, " |
| 90 | f"cumulative_logprob={self.cumulative_logprob}, " |
| 91 | f"logprobs={self.logprobs}, " |
| 92 | f"finish_reason={self.finish_reason}, " |
| 93 | f"stop_reason={self.stop_reason})" |
| 94 | ) |
| 95 | |
| 96 | |
| 97 | @dataclass |
| 98 | class PoolingOutput: |
| 99 | """The output data of one pooling output of a request. |
| 100 | |
| 101 | Args: |
| 102 | data: The extracted hidden states. |
| 103 | |
| 104 | """ |
| 105 | |
| 106 | data: torch.Tensor |
| 107 | |
| 108 | def __repr__(self) -> str: |
| 109 | return f"PoolingOutput(data={self.data})" |
| 110 | |
| 111 | def __eq__(self, other: object) -> bool: |
| 112 | return isinstance(other, self.__class__) and bool( |
| 113 | (self.data == other.data).all() |
| 114 | ) |
| 115 | |
| 116 | |
| 117 | class RequestOutput: |
| 118 | """The output data of a completion request to the LLM. |
| 119 | |
| 120 | Args: |
| 121 | request_id: The unique ID of the request. |
| 122 | prompt: The prompt string of the request. |
| 123 | For encoder/decoder models, this is the |
| 124 | decoder input prompt. |
| 125 | prompt_token_ids: The token IDs of the prompt. |
| 126 | For encoder/decoder models, this is the |
| 127 | decoder input prompt token ids. |
| 128 | prompt_logprobs: The log probabilities to return per prompt token. |
| 129 | prompt_token_id_logprobs: Logprobs of prompt_logprob_token_ids, shaped |
| 130 | [max(prompt_len - 1 - prompt_logprob_start, 0), |
| 131 | len(prompt_logprob_token_ids)]; row i scores them as predictions of |
| 132 | prompt token prompt_logprob_start + i + 1. |
| 133 | outputs: The output sequences of the request. |
| 134 | finished: Whether the whole request is finished. |
| 135 | metrics: Metrics associated with the request. |
| 136 | lora_request: The LoRA request that was used to generate the output. |
| 137 | encoder_prompt: The encoder prompt string of the request. |
| 138 | None if decoder-only. |
| 139 | encoder_prompt_token_ids: The token IDs of the encoder prompt. |
| 140 | None if decoder-only. |
| 141 | num_cached_tokens: The number of tokens with prefix cache hit. |
| 142 | num_cache_creation_tokens: Prompt tokens currently counted as local |
| 143 | prefix-cache writes for this request. |
| 144 | kv_transfer_params: The params for remote K/V transfer. |
| 145 | ec_transfer_params: The params for remote encoder-cache transfer. |
| 146 | |
| 147 | """ |
| 148 | |
| 149 | def __init__( |
| 150 | self, |
| 151 | request_id: str, |
| 152 | prompt: str | None, |
| 153 | prompt_token_ids: list[int] | None, |
| 154 | prompt_logprobs: PromptLogprobs | None, |
| 155 | outputs: list[CompletionOutput], |
| 156 | finished: bool, |
| 157 | metrics: RequestStateStats | None = None, |
| 158 | lora_request: LoRARequest | None = None, |
| 159 | encoder_prompt: str | None = None, |
| 160 | encoder_prompt_token_ids: list[int] | None = None, |
| 161 | num_cached_tokens: int | None = None, |
| 162 | num_cache_creation_tokens: int | None = None, |
| 163 | *, |
| 164 | prompt_token_id_logprobs: np.ndarray | None = None, |
| 165 | kv_transfer_params: dict[str, Any] | None = None, |
| 166 | ec_transfer_params: dict[str, Any] | None = None, |
| 167 | # Forward compatibility, code that uses args added in new release can |
| 168 | # still run with older versions of vLLM without breaking. |
| 169 | **kwargs: Any, |
| 170 | ) -> None: |
| 171 | if kwargs: |
| 172 | logger.warning_once( |
| 173 | "RequestOutput: Ignoring extra arguments: %s", str(kwargs) |
| 174 | ) |
| 175 | self.request_id = request_id |
| 176 | self.prompt = prompt |
| 177 | self.prompt_token_ids = prompt_token_ids |
| 178 | self.prompt_logprobs = prompt_logprobs |
| 179 | self.prompt_token_id_logprobs = prompt_token_id_logprobs |
| 180 | self.outputs = outputs |
| 181 | self.finished = finished |
| 182 | self.metrics = metrics |
| 183 | self.lora_request = lora_request |
| 184 | self.encoder_prompt = encoder_prompt |
| 185 | self.encoder_prompt_token_ids = encoder_prompt_token_ids |
| 186 | self.num_cached_tokens = num_cached_tokens |
| 187 | self.num_cache_creation_tokens = num_cache_creation_tokens |
| 188 | self.kv_transfer_params = kv_transfer_params |
| 189 | self.ec_transfer_params = ec_transfer_params |
| 190 | |
| 191 | def add(self, next_output: "RequestOutput", aggregate: bool) -> None: |
| 192 | """Merge subsequent RequestOutput into this one""" |
| 193 | self.finished |= next_output.finished |
| 194 | self.kv_transfer_params = next_output.kv_transfer_params |
| 195 | self.ec_transfer_params = next_output.ec_transfer_params |
| 196 | |
| 197 | for next_completion in next_output.outputs: |
| 198 | for i, completion in enumerate(self.outputs): |
| 199 | if completion.index == next_completion.index: |
| 200 | if aggregate: |
| 201 | # Merge outputs with same index |
| 202 | completion.text += next_completion.text |
| 203 | if not isinstance(completion.token_ids, MutableSequence): |
| 204 | completion.token_ids = list(completion.token_ids) |
| 205 | completion.token_ids.extend(next_completion.token_ids) |
| 206 | if next_completion.logprobs: |
| 207 | assert completion.logprobs is not None |
| 208 | completion.logprobs.extend(next_completion.logprobs) # type: ignore[arg-type] |
| 209 | if next_completion.sampling_mask is not None: |
| 210 | if completion.sampling_mask is None: |
| 211 | completion.sampling_mask = SamplingMask([]) |
| 212 | completion.sampling_mask.token_ids.extend( |
| 213 | next_completion.sampling_mask.token_ids |
| 214 | ) |
| 215 | completion.cumulative_logprob = ( |
| 216 | next_completion.cumulative_logprob |
| 217 | ) |
| 218 | # R3 is returned on the terminal output and must survive |
| 219 | # aggregation with earlier chunks that have no R3. |
| 220 | if next_completion.routed_experts is not None: |
| 221 | completion.routed_experts = next_completion.routed_experts |
| 222 | completion.finish_reason = next_completion.finish_reason |
| 223 | completion.stop_reason = next_completion.stop_reason |
| 224 | else: |
| 225 | # Replace the output with the new one |
| 226 | self.outputs[i] = next_completion |
| 227 | break |
| 228 | else: |
| 229 | self.outputs.append(next_completion) |
| 230 | |
| 231 | def __repr__(self) -> str: |
| 232 | return ( |
| 233 | f"RequestOutput(request_id={self.request_id}, " |
| 234 | f"prompt={self.prompt!r}, " |
| 235 | f"prompt_token_ids={self.prompt_token_ids}, " |
| 236 | f"encoder_prompt={self.encoder_prompt!r}, " |
| 237 | f"encoder_prompt_token_ids={self.encoder_prompt_token_ids}, " |
| 238 | f"prompt_logprobs={self.prompt_logprobs}, " |
| 239 | f"prompt_token_id_logprobs={self.prompt_token_id_logprobs}, " |
| 240 | f"outputs={self.outputs}, " |
| 241 | f"finished={self.finished}, " |
| 242 | f"metrics={self.metrics}, " |
| 243 | f"lora_request={self.lora_request}, " |
| 244 | f"num_cached_tokens={self.num_cached_tokens}, " |
| 245 | f"num_cache_creation_tokens={self.num_cache_creation_tokens})" |
| 246 | ) |
| 247 | |
| 248 | |
| 249 | # Sentinel to indicate request is finished, used with streaming inputs. |
| 250 | STREAM_FINISHED = RequestOutput( |
| 251 | request_id="", |
| 252 | prompt=None, |
| 253 | prompt_token_ids=None, |
| 254 | prompt_logprobs=None, |
| 255 | outputs=[], |
| 256 | finished=True, |
| 257 | ) |
| 258 | |
| 259 | _O = TypeVar("_O", default=PoolingOutput) |
| 260 | |
| 261 | |
| 262 | class PoolingRequestOutput(Generic[_O]): |
| 263 | """The output data of a pooling request to the LLM. |
| 264 | |
| 265 | Args: |
| 266 | request_id (str): A unique identifier for the pooling request. |
| 267 | outputs (PoolingOutput): The pooling results for the given input. |
| 268 | prompt_token_ids (list[int]): A list of token IDs used in the prompt. |
| 269 | num_cached_tokens: The number of tokens with prefix cache hit. |
| 270 | finished (bool): A flag indicating whether the pooling is completed. |
| 271 | error: Structured request-level error information, if the request failed. |
| 272 | |
| 273 | """ |
| 274 | |
| 275 | def __init__( |
| 276 | self, |
| 277 | request_id: str, |
| 278 | outputs: _O, |
| 279 | prompt_token_ids: list[int], |
| 280 | num_cached_tokens: int, |
| 281 | finished: bool, |
| 282 | *, |
| 283 | error: RequestError | None = None, |
| 284 | ): |
| 285 | self.request_id = request_id |
| 286 | self.prompt_token_ids = prompt_token_ids |
| 287 | self.num_cached_tokens = num_cached_tokens |
| 288 | self.finished = finished |
| 289 | self.outputs = outputs |
| 290 | self.error = error |
| 291 | |
| 292 | def __repr__(self) -> str: |
| 293 | error = f", error={self.error!r}" if self.error is not None else "" |
| 294 | return ( |
| 295 | f"{type(self).__name__}(request_id={self.request_id!r}, " |
| 296 | f"outputs={self.outputs!r}, " |
| 297 | f"prompt_token_ids={self.prompt_token_ids}, " |
| 298 | f"num_cached_tokens={self.num_cached_tokens}, " |
| 299 | f"finished={self.finished}" |
| 300 | f"{error})" |
| 301 | ) |
| 302 | |
| 303 | |
| 304 | @dataclass |
| 305 | class EmbeddingOutput: |
| 306 | """The output data of one embedding output of a request. |
| 307 | |
| 308 | Args: |
| 309 | embedding: The embedding vector, which is a list of floats. |
| 310 | Its length depends on the hidden dimension of the model. |
| 311 | |
| 312 | """ |
| 313 | |
| 314 | embedding: list[float] |
| 315 | |
| 316 | @staticmethod |
| 317 | def from_base(pooling_output: PoolingOutput) -> "EmbeddingOutput": |
| 318 | pooled_data = pooling_output.data |
| 319 | if pooled_data.ndim != 1: |
| 320 | raise ValueError("pooled_data should be a 1-D embedding vector") |
| 321 | |
| 322 | return EmbeddingOutput(pooled_data.tolist()) |
| 323 | |
| 324 | @property |
| 325 | def hidden_size(self) -> int: |
| 326 | return len(self.embedding) |
| 327 | |
| 328 | def __repr__(self) -> str: |
| 329 | return f"EmbeddingOutput(hidden_size={self.hidden_size})" |
| 330 | |
| 331 | |
| 332 | class EmbeddingRequestOutput(PoolingRequestOutput[EmbeddingOutput]): |
| 333 | @staticmethod |
| 334 | def from_base( |
| 335 | request_output: PoolingRequestOutput, |
| 336 | ) -> "EmbeddingRequestOutput": |
| 337 | outputs = ( |
| 338 | EmbeddingOutput([]) |
| 339 | if request_output.error is not None |
| 340 | else EmbeddingOutput.from_base(request_output.outputs) |
| 341 | ) |
| 342 | return EmbeddingRequestOutput( |
| 343 | request_id=request_output.request_id, |
| 344 | outputs=outputs, |
| 345 | prompt_token_ids=request_output.prompt_token_ids, |
| 346 | num_cached_tokens=request_output.num_cached_tokens, |
| 347 | finished=request_output.finished, |
| 348 | error=request_output.error, |
| 349 | ) |
| 350 | |
| 351 | |
| 352 | @dataclass |
| 353 | class ClassificationOutput: |
| 354 | """The output data of one classification output of a request. |
| 355 | |
| 356 | Args: |
| 357 | probs: The probability vector, which is a list of floats. |
| 358 | Its length depends on the number of classes. |
| 359 | |
| 360 | """ |
| 361 | |
| 362 | probs: list[float] |
| 363 | |
| 364 | @staticmethod |
| 365 | def from_base(pooling_output: PoolingOutput) -> "ClassificationOutput": |
| 366 | # pooling_output shape: (num_classes) |
| 367 | pooled_data = pooling_output.data |
| 368 | if pooled_data.ndim != 1: |
| 369 | raise ValueError("pooled_data should be a 1-D probability vector") |
| 370 | |
| 371 | return ClassificationOutput(pooled_data.tolist()) |
| 372 | |
| 373 | @property |
| 374 | def num_classes(self) -> int: |
| 375 | return len(self.probs) |
| 376 | |
| 377 | def __repr__(self) -> str: |
| 378 | return f"ClassificationOutput(num_classes={self.num_classes})" |
| 379 | |
| 380 | |
| 381 | class ClassificationRequestOutput(PoolingRequestOutput[ClassificationOutput]): |
| 382 | @staticmethod |
| 383 | def from_base( |
| 384 | request_output: PoolingRequestOutput, |
| 385 | ) -> "ClassificationRequestOutput": |
| 386 | outputs = ( |
| 387 | ClassificationOutput([]) |
| 388 | if request_output.error is not None |
| 389 | else ClassificationOutput.from_base(request_output.outputs) |
| 390 | ) |
| 391 | return ClassificationRequestOutput( |
| 392 | request_id=request_output.request_id, |
| 393 | outputs=outputs, |
| 394 | prompt_token_ids=request_output.prompt_token_ids, |
| 395 | num_cached_tokens=request_output.num_cached_tokens, |
| 396 | finished=request_output.finished, |
| 397 | error=request_output.error, |
| 398 | ) |
| 399 | |
| 400 | |
| 401 | @dataclass |
| 402 | class ScoringOutput: |
| 403 | """The output data of one scoring output of a request. |
| 404 | |
| 405 | Args: |
| 406 | score: The similarity score, which is a scalar value. |
| 407 | |
| 408 | """ |
| 409 | |
| 410 | score: float |
| 411 | |
| 412 | @staticmethod |
| 413 | def from_base(pooling_output: PoolingOutput) -> "ScoringOutput": |
| 414 | # pooling_output shape: |
| 415 | # classify task: (num_classes) num_classes == 1 |
| 416 | # embed task: a scalar value |
| 417 | pooled_data = pooling_output.data.squeeze() |
| 418 | if pooled_data.ndim != 0: |
| 419 | raise ValueError("pooled_data should be a scalar score") |
| 420 | |
| 421 | return ScoringOutput(pooled_data.item()) |
| 422 | |
| 423 | def __repr__(self) -> str: |
| 424 | return f"ScoringOutput(score={self.score})" |
| 425 | |
| 426 | |
| 427 | class ScoringRequestOutput(PoolingRequestOutput[ScoringOutput]): |
| 428 | @staticmethod |
| 429 | def from_base( |
| 430 | request_output: PoolingRequestOutput, |
| 431 | ) -> "ScoringRequestOutput": |
| 432 | outputs = ( |
| 433 | ScoringOutput(0.0) |
| 434 | if request_output.error is not None |
| 435 | else ScoringOutput.from_base(request_output.outputs) |
| 436 | ) |
| 437 | return ScoringRequestOutput( |
| 438 | request_id=request_output.request_id, |
| 439 | outputs=outputs, |
| 440 | prompt_token_ids=request_output.prompt_token_ids, |
| 441 | num_cached_tokens=request_output.num_cached_tokens, |
| 442 | finished=request_output.finished, |
| 443 | error=request_output.error, |
| 444 | ) |