| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import json |
| 5 | import time |
| 6 | from http import HTTPStatus |
| 7 | from typing import TYPE_CHECKING, Literal, TypeAlias |
| 8 | |
| 9 | from fastapi import HTTPException, UploadFile |
| 10 | from pydantic import ( |
| 11 | Field, |
| 12 | model_validator, |
| 13 | ) |
| 14 | |
| 15 | from vllm.config.speech_to_text import SpeechToTextParams |
| 16 | from vllm.entrypoints.generate.base.protocol import DeltaMessage |
| 17 | from vllm.entrypoints.serve.engine.protocol import OpenAIBaseModel, UsageInfo |
| 18 | from vllm.exceptions import VLLMValidationError |
| 19 | from vllm.logger import init_logger |
| 20 | from vllm.sampling_params import ( |
| 21 | BeamSearchParams, |
| 22 | RequestOutputKind, |
| 23 | SamplingParams, |
| 24 | ) |
| 25 | from vllm.utils import random_uuid |
| 26 | |
| 27 | from ..base.protocol import _LONG_INFO, TranscriptionResponseFormat |
| 28 | |
| 29 | if TYPE_CHECKING: |
| 30 | import numpy as np |
| 31 | |
| 32 | from vllm.config import ModelConfig, SpeechToTextConfig |
| 33 | |
| 34 | logger = init_logger(__name__) |
| 35 | |
| 36 | |
| 37 | class TranscriptionResponseStreamChoice(OpenAIBaseModel): |
| 38 | delta: DeltaMessage |
| 39 | finish_reason: str | None = None |
| 40 | stop_reason: int | str | None = None |
| 41 | |
| 42 | |
| 43 | class TranscriptionStreamResponse(OpenAIBaseModel): |
| 44 | id: str = Field(default_factory=lambda: f"trsc-{random_uuid()}") |
| 45 | object: Literal["transcription.chunk"] = "transcription.chunk" |
| 46 | created: int = Field(default_factory=lambda: int(time.time())) |
| 47 | model: str |
| 48 | choices: list[TranscriptionResponseStreamChoice] |
| 49 | usage: UsageInfo | None = Field(default=None) |
| 50 | |
| 51 | |
| 52 | class TranscriptionRequest(OpenAIBaseModel): |
| 53 | # Ordered by official OpenAI API documentation |
| 54 | # https://platform.openai.com/docs/api-reference/audio/createTranscription |
| 55 | |
| 56 | file: UploadFile |
| 57 | """ |
| 58 | The audio file object (not file name) to transcribe, in one of these |
| 59 | formats: flac, mp3, mp4, mpeg, mpga, m4a, ogg, wav, or webm. |
| 60 | """ |
| 61 | |
| 62 | model: str | None = None |
| 63 | """ID of the model to use. |
| 64 | """ |
| 65 | |
| 66 | language: str | None = None |
| 67 | """The language of the input audio. |
| 68 | |
| 69 | Supplying the input language in |
| 70 | [ISO-639-1](https://en.wikipedia.org/wiki/List_of_ISO_639-1_codes) format |
| 71 | will improve accuracy and latency. |
| 72 | """ |
| 73 | |
| 74 | hotwords: str | None = None |
| 75 | """ |
| 76 | hotwords refers to a list of important words or phrases that the model |
| 77 | should pay extra attention to during transcription. |
| 78 | """ |
| 79 | |
| 80 | prompt: str = Field(default="") |
| 81 | """An optional text to guide the model's style or continue a previous audio |
| 82 | segment. |
| 83 | |
| 84 | The [prompt](https://platform.openai.com/docs/guides/speech-to-text#prompting) |
| 85 | should match the audio language. |
| 86 | """ |
| 87 | |
| 88 | response_format: TranscriptionResponseFormat = Field(default="json") |
| 89 | """ |
| 90 | The format of the output, in one of these options: `json`, `text`, `srt`, |
| 91 | `verbose_json`, or `vtt`. |
| 92 | """ |
| 93 | |
| 94 | ## TODO (varun) : Support if set to 0, certain thresholds are met !! |
| 95 | |
| 96 | timestamp_granularities: list[Literal["word", "segment"]] = Field( |
| 97 | alias="timestamp_granularities[]", default=[] |
| 98 | ) |
| 99 | """The timestamp granularities to populate for this transcription. |
| 100 | |
| 101 | `response_format` must be set `verbose_json` to use timestamp granularities. |
| 102 | Either or both of these options are supported: `word`, or `segment`. Note: |
| 103 | There is no additional latency for segment timestamps, but generating word |
| 104 | timestamps incurs additional latency. |
| 105 | """ |
| 106 | |
| 107 | stream: bool | None = False |
| 108 | """When set, it will enable output to be streamed in a similar fashion |
| 109 | as the Chat Completion endpoint. |
| 110 | """ |
| 111 | # --8<-- [start:transcription-extra-params] |
| 112 | # Flattened stream option to simplify form data. |
| 113 | stream_include_usage: bool | None = False |
| 114 | stream_continuous_usage_stats: bool | None = False |
| 115 | |
| 116 | vllm_xargs: dict[str, str | int | float | list[str | int | float]] | None = Field( |
| 117 | default=None, |
| 118 | description=( |
| 119 | "Additional request parameters with (list of) string or " |
| 120 | "numeric values, used by custom extensions." |
| 121 | ), |
| 122 | ) |
| 123 | # --8<-- [end:transcription-extra-params] |
| 124 | |
| 125 | to_language: str | None = None |
| 126 | """The language of the output audio we transcribe to. |
| 127 | |
| 128 | Please note that this is not currently used by supported models at this |
| 129 | time, but it is a placeholder for future use, matching translation api. |
| 130 | """ |
| 131 | |
| 132 | # --8<-- [start:transcription-sampling-params] |
| 133 | use_beam_search: bool = False |
| 134 | """Whether or not beam search should be used.""" |
| 135 | |
| 136 | n: int = 1 |
| 137 | """The number of beams to be used in beam search.""" |
| 138 | |
| 139 | length_penalty: float = 1.0 |
| 140 | """Length penalty to be used for beam search.""" |
| 141 | |
| 142 | include_stop_str_in_output: bool = False |
| 143 | """Whether to include the stop strings in output text.""" |
| 144 | |
| 145 | temperature: float = Field(default=0.0) |
| 146 | """The sampling temperature, between 0 and 1. |
| 147 | |
| 148 | Higher values like 0.8 will make the output more random, while lower values |
| 149 | like 0.2 will make it more focused / deterministic. If set to 0, the model |
| 150 | will use [log probability](https://en.wikipedia.org/wiki/Log_probability) |
| 151 | to automatically increase the temperature until certain thresholds are hit. |
| 152 | """ |
| 153 | |
| 154 | top_p: float | None = None |
| 155 | """Enables nucleus (top-p) sampling, where tokens are selected from the |
| 156 | smallest possible set whose cumulative probability exceeds `p`. |
| 157 | """ |
| 158 | |
| 159 | top_k: int | None = None |
| 160 | """Limits sampling to the `k` most probable tokens at each step.""" |
| 161 | |
| 162 | min_p: float | None = None |
| 163 | """Filters out tokens with a probability lower than `min_p`, ensuring a |
| 164 | minimum likelihood threshold during sampling. |
| 165 | """ |
| 166 | |
| 167 | seed: int | None = Field(None, ge=_LONG_INFO.min, le=_LONG_INFO.max) |
| 168 | """The seed to use for sampling.""" |
| 169 | |
| 170 | frequency_penalty: float | None = 0.0 |
| 171 | """The frequency penalty to use for sampling.""" |
| 172 | |
| 173 | repetition_penalty: float | None = None |
| 174 | """The repetition penalty to use for sampling.""" |
| 175 | |
| 176 | presence_penalty: float | None = 0.0 |
| 177 | """The presence penalty to use for sampling.""" |
| 178 | |
| 179 | max_completion_tokens: int | None = None |
| 180 | """The maximum number of tokens to generate.""" |
| 181 | # --8<-- [end:transcription-sampling-params] |
| 182 | |
| 183 | # Default sampling parameters for transcription requests. |
| 184 | _DEFAULT_SAMPLING_PARAMS: dict = { |
| 185 | "repetition_penalty": 1.0, |
| 186 | "temperature": 1.0, |
| 187 | "top_p": 1.0, |
| 188 | "top_k": 0, |
| 189 | "min_p": 0.0, |
| 190 | } |
| 191 | |
| 192 | def build_stt_params( |
| 193 | self, |
| 194 | audio: "np.ndarray", |
| 195 | stt_config: "SpeechToTextConfig", |
| 196 | model_config: "ModelConfig", |
| 197 | task_type: str, |
| 198 | ) -> SpeechToTextParams: |
| 199 | return SpeechToTextParams( |
| 200 | audio=audio, |
| 201 | stt_config=stt_config, |
| 202 | model_config=model_config, |
| 203 | language=self.language, |
| 204 | task_type=task_type, |
| 205 | request_prompt=self.prompt, |
| 206 | to_language=self.to_language, |
| 207 | hotwords=self.hotwords, |
| 208 | ) |
| 209 | |
| 210 | def to_beam_search_params( |
| 211 | self, |
| 212 | default_max_tokens: int, |
| 213 | default_sampling_params: dict | None = None, |
| 214 | ) -> BeamSearchParams: |
| 215 | if default_sampling_params is None: |
| 216 | default_sampling_params = {} |
| 217 | |
| 218 | max_tokens = default_max_tokens |
| 219 | n = self.n if self.n is not None else 1 |
| 220 | |
| 221 | # NOTE: Temp 0 is a different fallback than completions |
| 222 | if (temperature := self.temperature) is None: |
| 223 | temperature = default_sampling_params.get("temperature", 0) |
| 224 | |
| 225 | return BeamSearchParams( |
| 226 | beam_width=n, |
| 227 | max_tokens=max_tokens, |
| 228 | temperature=temperature, |
| 229 | length_penalty=self.length_penalty, |
| 230 | include_stop_str_in_output=self.include_stop_str_in_output, |
| 231 | ) |
| 232 | |
| 233 | def to_sampling_params( |
| 234 | self, default_max_tokens: int, default_sampling_params: dict | None = None |
| 235 | ) -> SamplingParams: |
| 236 | max_tokens = default_max_tokens |
| 237 | |
| 238 | if default_sampling_params is None: |
| 239 | default_sampling_params = {} |
| 240 | |
| 241 | # Default parameters |
| 242 | if (temperature := self.temperature) is None: |
| 243 | temperature = default_sampling_params.get( |
| 244 | "temperature", self._DEFAULT_SAMPLING_PARAMS["temperature"] |
| 245 | ) |
| 246 | if (top_p := self.top_p) is None: |
| 247 | top_p = default_sampling_params.get( |
| 248 | "top_p", self._DEFAULT_SAMPLING_PARAMS["top_p"] |
| 249 | ) |
| 250 | if (top_k := self.top_k) is None: |
| 251 | top_k = default_sampling_params.get( |
| 252 | "top_k", self._DEFAULT_SAMPLING_PARAMS["top_k"] |
| 253 | ) |
| 254 | if (min_p := self.min_p) is None: |
| 255 | min_p = default_sampling_params.get( |
| 256 | "min_p", self._DEFAULT_SAMPLING_PARAMS["min_p"] |
| 257 | ) |
| 258 | |
| 259 | if (repetition_penalty := self.repetition_penalty) is None: |
| 260 | repetition_penalty = default_sampling_params.get( |
| 261 | "repetition_penalty", |
| 262 | self._DEFAULT_SAMPLING_PARAMS["repetition_penalty"], |
| 263 | ) |
| 264 | |
| 265 | return SamplingParams.from_optional( |
| 266 | temperature=temperature, |
| 267 | max_tokens=max_tokens, |
| 268 | seed=self.seed, |
| 269 | top_p=top_p, |
| 270 | top_k=top_k, |
| 271 | min_p=min_p, |
| 272 | frequency_penalty=self.frequency_penalty, |
| 273 | repetition_penalty=repetition_penalty, |
| 274 | presence_penalty=self.presence_penalty, |
| 275 | output_kind=RequestOutputKind.DELTA |
| 276 | if self.stream |
| 277 | else RequestOutputKind.FINAL_ONLY, |
| 278 | extra_args=self.vllm_xargs, |
| 279 | skip_clone=True, # Created fresh per request, safe to skip clone |
| 280 | ) |
| 281 | |
| 282 | @model_validator(mode="before") |
| 283 | @classmethod |
| 284 | def validate_transcription_request(cls, data): |
| 285 | if not isinstance(data, dict): |
| 286 | return data |
| 287 | if isinstance(data.get("file"), str): |
| 288 | raise HTTPException( |
| 289 | status_code=HTTPStatus.UNPROCESSABLE_ENTITY, |
| 290 | detail="Expected 'file' to be a file-like object, not 'str'.", |
| 291 | ) |
| 292 | |
| 293 | stream_opts = ["stream_include_usage", "stream_continuous_usage_stats"] |
| 294 | stream = data.get("stream", False) |
| 295 | if any(bool(data.get(so, False)) for so in stream_opts) and not stream: |
| 296 | # Find which specific stream option was set |
| 297 | invalid_param = next( |
| 298 | (so for so in stream_opts if data.get(so, False)), |
| 299 | "stream_include_usage", |
| 300 | ) |
| 301 | raise VLLMValidationError( |
| 302 | "Stream options can only be defined when `stream=True`.", |
| 303 | parameter=invalid_param, |
| 304 | ) |
| 305 | |
| 306 | # Parse vllm_xargs from JSON string (form data sends it as a string) |
| 307 | xargs = data.get("vllm_xargs") |
| 308 | if isinstance(xargs, str): |
| 309 | try: |
| 310 | data["vllm_xargs"] = json.loads(xargs) |
| 311 | except json.JSONDecodeError as e: |
| 312 | raise VLLMValidationError( |
| 313 | f"Failed to parse vllm_xargs. Must be valid JSON: {e}", |
| 314 | parameter="vllm_xargs", |
| 315 | ) from e |
| 316 | |
| 317 | return data |
| 318 | |
| 319 | |
| 320 | # Transcription response objects |
| 321 | class TranscriptionUsageAudio(OpenAIBaseModel): |
| 322 | type: Literal["duration"] = "duration" |
| 323 | seconds: int |
| 324 | |
| 325 | |
| 326 | class TranscriptionResponse(OpenAIBaseModel): |
| 327 | text: str |
| 328 | """The transcribed text.""" |
| 329 | usage: TranscriptionUsageAudio |
| 330 | |
| 331 | |
| 332 | class TranscriptionWord(OpenAIBaseModel): |
| 333 | end: float |
| 334 | """End time of the word in seconds.""" |
| 335 | |
| 336 | start: float |
| 337 | """Start time of the word in seconds.""" |
| 338 | |
| 339 | word: str |
| 340 | """The text content of the word.""" |
| 341 | |
| 342 | |
| 343 | class TranscriptionSegment(OpenAIBaseModel): |
| 344 | id: int |
| 345 | """Unique identifier of the segment.""" |
| 346 | |
| 347 | avg_logprob: float |
| 348 | """Average logprob of the segment. |
| 349 | |
| 350 | If the value is lower than -1, consider the logprobs failed. |
| 351 | """ |
| 352 | |
| 353 | compression_ratio: float |
| 354 | """Compression ratio of the segment. |
| 355 | |
| 356 | If the value is greater than 2.4, consider the compression failed. |
| 357 | """ |
| 358 | |
| 359 | end: float |
| 360 | """End time of the segment in seconds.""" |
| 361 | |
| 362 | no_speech_prob: float | None = None |
| 363 | """Probability of no speech in the segment. |
| 364 | |
| 365 | If the value is higher than 1.0 and the `avg_logprob` is below -1, consider |
| 366 | this segment silent. |
| 367 | """ |
| 368 | |
| 369 | seek: int |
| 370 | """Seek offset of the segment.""" |
| 371 | |
| 372 | start: float |
| 373 | """Start time of the segment in seconds.""" |
| 374 | |
| 375 | temperature: float |
| 376 | """Temperature parameter used for generating the segment.""" |
| 377 | |
| 378 | text: str |
| 379 | """Text content of the segment.""" |
| 380 | |
| 381 | tokens: list[int] |
| 382 | """Array of token IDs for the text content.""" |
| 383 | |
| 384 | |
| 385 | class TranscriptionResponseVerbose(OpenAIBaseModel): |
| 386 | duration: float |
| 387 | """The duration of the input audio.""" |
| 388 | |
| 389 | language: str |
| 390 | """The language of the input audio.""" |
| 391 | |
| 392 | text: str |
| 393 | """The transcribed text.""" |
| 394 | |
| 395 | segments: list[TranscriptionSegment] | None = None |
| 396 | """Segments of the transcribed text and their corresponding details.""" |
| 397 | |
| 398 | words: list[TranscriptionWord] | None = None |
| 399 | """Extracted words and their corresponding timestamps.""" |
| 400 | |
| 401 | |
| 402 | class TranscriptionDiarizedSegment(OpenAIBaseModel): |
| 403 | """A speaker-attributed transcription segment.""" |
| 404 | |
| 405 | type: Literal["transcript.text.segment"] = "transcript.text.segment" |
| 406 | id: str |
| 407 | start: float |
| 408 | end: float |
| 409 | text: str |
| 410 | speaker: str |
| 411 | |
| 412 | |
| 413 | class TranscriptionResponseDiarized(OpenAIBaseModel): |
| 414 | """OpenAI-compatible diarized transcription response.""" |
| 415 | |
| 416 | task: Literal["transcribe"] = "transcribe" |
| 417 | duration: float |
| 418 | text: str |
| 419 | segments: list[TranscriptionDiarizedSegment] |
| 420 | usage: TranscriptionUsageAudio |
| 421 | |
| 422 | |
| 423 | TranscriptionResponseVariant: TypeAlias = ( |
| 424 | TranscriptionResponse | TranscriptionResponseVerbose | TranscriptionResponseDiarized |
| 425 | ) |