| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | from collections.abc import AsyncGenerator |
| 4 | |
| 5 | from fastapi import Request |
| 6 | |
| 7 | from vllm.engine.protocol import EngineClient |
| 8 | from vllm.entrypoints.generate.base.protocol import RequestResponseMetadata |
| 9 | from vllm.entrypoints.openai.models.serving import OpenAIServingModels |
| 10 | from vllm.entrypoints.serve.utils.request_logger import RequestLogger |
| 11 | from vllm.logger import init_logger |
| 12 | from vllm.outputs import RequestOutput |
| 13 | |
| 14 | from ...serve.engine.protocol import ErrorResponse |
| 15 | from ..base.serving import SpeechToTextBaseServing |
| 16 | from .protocol import ( |
| 17 | TranscriptionRequest, |
| 18 | TranscriptionResponse, |
| 19 | TranscriptionResponseDiarized, |
| 20 | TranscriptionResponseStreamChoice, |
| 21 | TranscriptionResponseVerbose, |
| 22 | TranscriptionStreamResponse, |
| 23 | ) |
| 24 | |
| 25 | logger = init_logger(__name__) |
| 26 | |
| 27 | |
| 28 | class OpenAIServingTranscription(SpeechToTextBaseServing): |
| 29 | """Handles transcription requests.""" |
| 30 | |
| 31 | def __init__( |
| 32 | self, |
| 33 | engine_client: EngineClient, |
| 34 | models: OpenAIServingModels, |
| 35 | *, |
| 36 | request_logger: RequestLogger | None, |
| 37 | return_tokens_as_token_ids: bool = False, |
| 38 | enable_force_include_usage: bool = False, |
| 39 | ): |
| 40 | super().__init__( |
| 41 | engine_client=engine_client, |
| 42 | models=models, |
| 43 | request_logger=request_logger, |
| 44 | return_tokens_as_token_ids=return_tokens_as_token_ids, |
| 45 | task_type="transcribe", |
| 46 | enable_force_include_usage=enable_force_include_usage, |
| 47 | ) |
| 48 | |
| 49 | async def create_transcription( |
| 50 | self, |
| 51 | audio_data: bytes, |
| 52 | request: TranscriptionRequest, |
| 53 | raw_request: Request | None = None, |
| 54 | ) -> ( |
| 55 | TranscriptionResponse |
| 56 | | TranscriptionResponseVerbose |
| 57 | | TranscriptionResponseDiarized |
| 58 | | AsyncGenerator[str, None] |
| 59 | | ErrorResponse |
| 60 | ): |
| 61 | """Transcription API similar to OpenAI's API. |
| 62 | |
| 63 | See https://platform.openai.com/docs/api-reference/audio/createTranscription |
| 64 | for the API specification. This API mimics the OpenAI transcription API. |
| 65 | """ |
| 66 | return await self._create_speech_to_text( |
| 67 | audio_data=audio_data, |
| 68 | request=request, |
| 69 | raw_request=raw_request, |
| 70 | response_class=( |
| 71 | TranscriptionResponseVerbose |
| 72 | if request.response_format == "verbose_json" |
| 73 | else TranscriptionResponseDiarized |
| 74 | if request.response_format == "diarized_json" |
| 75 | else TranscriptionResponse |
| 76 | ), |
| 77 | stream_generator_method=self.transcription_stream_generator, |
| 78 | ) |
| 79 | |
| 80 | async def transcription_stream_generator( |
| 81 | self, |
| 82 | request: TranscriptionRequest, |
| 83 | result_generator: list[AsyncGenerator[RequestOutput, None]], |
| 84 | request_id: str, |
| 85 | request_metadata: RequestResponseMetadata, |
| 86 | audio_duration_s: float, |
| 87 | separator: str, |
| 88 | ) -> AsyncGenerator[str, None]: |
| 89 | generator = self._speech_to_text_stream_generator( |
| 90 | request=request, |
| 91 | list_result_generator=result_generator, |
| 92 | request_id=request_id, |
| 93 | request_metadata=request_metadata, |
| 94 | audio_duration_s=audio_duration_s, |
| 95 | chunk_object_type="transcription.chunk", |
| 96 | response_stream_choice_class=TranscriptionResponseStreamChoice, |
| 97 | stream_response_class=TranscriptionStreamResponse, |
| 98 | separator=separator, |
| 99 | ) |
| 100 | async for chunk in generator: |
| 101 | yield chunk |