| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | from abc import ABC, abstractmethod |
| 4 | from collections.abc import AsyncGenerator, Sequence |
| 5 | from typing import Generic, TypeVar |
| 6 | |
| 7 | from vllm.config import VllmConfig |
| 8 | from vllm.inputs import PromptType |
| 9 | from vllm.outputs import PoolingRequestOutput |
| 10 | from vllm.pooling_params import PoolingParams |
| 11 | from vllm.renderers import BaseRenderer |
| 12 | from vllm.sampling_params import SamplingParams |
| 13 | |
| 14 | IOProcessorInput = TypeVar("IOProcessorInput") |
| 15 | IOProcessorOutput = TypeVar("IOProcessorOutput") |
| 16 | |
| 17 | |
| 18 | class IOProcessor(ABC, Generic[IOProcessorInput, IOProcessorOutput]): |
| 19 | """Abstract interface for pre/post-processing of engine I/O.""" |
| 20 | |
| 21 | def __init__(self, vllm_config: VllmConfig, renderer: BaseRenderer): |
| 22 | super().__init__() |
| 23 | |
| 24 | self.vllm_config = vllm_config |
| 25 | |
| 26 | def parse_data(self, data: object) -> IOProcessorInput: |
| 27 | raise NotImplementedError |
| 28 | |
| 29 | def merge_sampling_params( |
| 30 | self, |
| 31 | params: SamplingParams | None = None, |
| 32 | ) -> SamplingParams: |
| 33 | return params or SamplingParams() |
| 34 | |
| 35 | def merge_pooling_params( |
| 36 | self, |
| 37 | params: PoolingParams | None = None, |
| 38 | ) -> PoolingParams: |
| 39 | return params or PoolingParams(task="plugin") |
| 40 | |
| 41 | @abstractmethod |
| 42 | def pre_process( |
| 43 | self, |
| 44 | prompt: IOProcessorInput, |
| 45 | request_id: str | None = None, |
| 46 | **kwargs, |
| 47 | ) -> PromptType | Sequence[PromptType]: |
| 48 | raise NotImplementedError |
| 49 | |
| 50 | async def pre_process_async( |
| 51 | self, |
| 52 | prompt: IOProcessorInput, |
| 53 | request_id: str | None = None, |
| 54 | **kwargs, |
| 55 | ) -> PromptType | Sequence[PromptType]: |
| 56 | return self.pre_process(prompt, request_id, **kwargs) |
| 57 | |
| 58 | @abstractmethod |
| 59 | def post_process( |
| 60 | self, |
| 61 | model_output: Sequence[PoolingRequestOutput], |
| 62 | request_id: str | None = None, |
| 63 | **kwargs, |
| 64 | ) -> IOProcessorOutput: |
| 65 | raise NotImplementedError |
| 66 | |
| 67 | async def post_process_async( |
| 68 | self, |
| 69 | model_output: AsyncGenerator[tuple[int, PoolingRequestOutput]], |
| 70 | request_id: str | None = None, |
| 71 | **kwargs, |
| 72 | ) -> IOProcessorOutput: |
| 73 | # We cannot guarantee outputs are returned in the same order they were |
| 74 | # fed to vLLM. |
| 75 | # Let's sort them by id before post_processing |
| 76 | sorted_output = sorted( |
| 77 | [(i, item) async for i, item in model_output], key=lambda output: output[0] |
| 78 | ) |
| 79 | collected_output = [output[1] for output in sorted_output] |
| 80 | return self.post_process(collected_output, request_id=request_id, **kwargs) |