| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import asyncio |
| 5 | from collections.abc import AsyncGenerator |
| 6 | from functools import cached_property |
| 7 | from typing import Literal, cast |
| 8 | |
| 9 | import numpy as np |
| 10 | |
| 11 | from vllm.engine.protocol import EngineClient, StreamingInput |
| 12 | from vllm.entrypoints.generate.base.serving import GenerateBaseServing |
| 13 | from vllm.entrypoints.openai.models.serving import OpenAIServingModels |
| 14 | from vllm.entrypoints.serve.utils.request_logger import RequestLogger |
| 15 | from vllm.inputs import PromptType |
| 16 | from vllm.logger import init_logger |
| 17 | from vllm.model_executor.models.interfaces import SupportsRealtime |
| 18 | from vllm.renderers.inputs.preprocess import parse_model_prompt |
| 19 | |
| 20 | logger = init_logger(__name__) |
| 21 | |
| 22 | |
| 23 | class OpenAIServingRealtime(GenerateBaseServing): |
| 24 | """Realtime audio transcription service via WebSocket streaming. |
| 25 | |
| 26 | Provides streaming audio-to-text transcription by transforming audio chunks |
| 27 | into StreamingInput objects that can be consumed by the engine. |
| 28 | """ |
| 29 | |
| 30 | def __init__( |
| 31 | self, |
| 32 | engine_client: EngineClient, |
| 33 | models: OpenAIServingModels, |
| 34 | *, |
| 35 | request_logger: RequestLogger | None, |
| 36 | ): |
| 37 | super().__init__( |
| 38 | engine_client=engine_client, |
| 39 | models=models, |
| 40 | request_logger=request_logger, |
| 41 | ) |
| 42 | |
| 43 | self.task_type: Literal["realtime"] = "realtime" |
| 44 | |
| 45 | logger.info("OpenAIServingRealtime initialized for task: %s", self.task_type) |
| 46 | |
| 47 | @cached_property |
| 48 | def model_cls(self) -> type[SupportsRealtime]: |
| 49 | """Get the model class that supports transcription.""" |
| 50 | from vllm.model_executor.model_loader import get_model_cls |
| 51 | |
| 52 | model_cls = get_model_cls(self.model_config) |
| 53 | return cast(type[SupportsRealtime], model_cls) |
| 54 | |
| 55 | async def transcribe_realtime( |
| 56 | self, |
| 57 | audio_stream: AsyncGenerator[np.ndarray, None], |
| 58 | input_stream: asyncio.Queue[list[int]], |
| 59 | ) -> AsyncGenerator[StreamingInput, None]: |
| 60 | """Transform audio stream into StreamingInput for engine.generate(). |
| 61 | |
| 62 | Args: |
| 63 | audio_stream: Async generator yielding float32 numpy audio arrays |
| 64 | input_stream: Queue containing context token IDs from previous |
| 65 | generation outputs. Used for autoregressive multi-turn |
| 66 | processing where each generation's output becomes the context |
| 67 | for the next iteration. |
| 68 | |
| 69 | Yields: |
| 70 | StreamingInput objects containing audio prompts for the engine |
| 71 | |
| 72 | """ |
| 73 | model_config = self.model_config |
| 74 | renderer = self.renderer |
| 75 | |
| 76 | # mypy is being stupid |
| 77 | # TODO(Patrick) - fix this |
| 78 | stream_input_iter = cast( |
| 79 | AsyncGenerator[PromptType, None], |
| 80 | self.model_cls.buffer_realtime_audio( |
| 81 | audio_stream, input_stream, model_config |
| 82 | ), |
| 83 | ) |
| 84 | |
| 85 | async for prompt in stream_input_iter: |
| 86 | parsed_prompt = parse_model_prompt(model_config, prompt) |
| 87 | (engine_input,) = await renderer.render_cmpl_async([parsed_prompt]) |
| 88 | |
| 89 | yield StreamingInput(prompt=engine_input) |