| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | from transformers import BaseImageProcessor, BatchFeature, TensorType |
| 4 | from transformers.processing_utils import ProcessorMixin |
| 5 | |
| 6 | from vllm.multimodal.inputs import VisionChunk |
| 7 | from vllm.tokenizers.hf import HfTokenizer |
| 8 | |
| 9 | |
| 10 | class KimiK25Processor(ProcessorMixin): |
| 11 | attributes = ["image_processor", "tokenizer"] |
| 12 | |
| 13 | def __init__( |
| 14 | self, |
| 15 | image_processor: BaseImageProcessor, |
| 16 | tokenizer: HfTokenizer, |
| 17 | media_token_id: int, |
| 18 | ) -> None: |
| 19 | self.image_processor = image_processor |
| 20 | self.tokenizer = tokenizer |
| 21 | |
| 22 | self.media_token_id = media_token_id |
| 23 | |
| 24 | def __call__( |
| 25 | self, |
| 26 | text: str | list[str] | None = None, |
| 27 | vision_chunks: list[VisionChunk] | None = None, |
| 28 | return_tensors: str | TensorType | None = None, |
| 29 | **kwargs, |
| 30 | ) -> BatchFeature: |
| 31 | """Args: |
| 32 | text: The text to be field to the model. |
| 33 | vision_chunks: List of `VisionChunk` items to be processed. |
| 34 | For image: `VisionChunkImage` with |
| 35 | `type='image', image=PIL.Image` |
| 36 | For video_chunk: `VisionChunkVideo` with |
| 37 | `type='video_chunk', video_chunk=list[PIL.Image]` |
| 38 | Returns: |
| 39 | [`BatchFeature`]: A [`BatchFeature`] with the following fields: |
| 40 | |
| 41 | - **input_ids** -- list of token ids to be fed to a model. |
| 42 | - **pixel_values** -- Pixel values to be fed to a model. |
| 43 | Returned when `vision_chunks` is not `None`. |
| 44 | - **grid_thws** -- list of image 3D grid in LLM. |
| 45 | Returned when `vision_chunks` is not `None`. |
| 46 | |
| 47 | """ |
| 48 | if vision_chunks is not None: |
| 49 | mm_inputs = self.image_processor.preprocess( |
| 50 | vision_chunks, |
| 51 | return_tensors=return_tensors, |
| 52 | ) |
| 53 | else: |
| 54 | mm_inputs = {} |
| 55 | |
| 56 | if text is not None: |
| 57 | if not isinstance(text, list): |
| 58 | text = [text] |
| 59 | |
| 60 | text_inputs = self.tokenizer(text) |
| 61 | |
| 62 | # Note: Modify in-place |
| 63 | input_ids: list[list[int]] = text_inputs["input_ids"] # type: ignore |
| 64 | |
| 65 | if vision_chunks is not None: |
| 66 | num_tokens_per_chunk = [ |
| 67 | self.image_processor.media_tokens_calculator(chunk) |
| 68 | for chunk in vision_chunks |
| 69 | ] |
| 70 | |
| 71 | for i in range(len(input_ids)): |
| 72 | new_input_ids = [] |
| 73 | for token in input_ids[i]: |
| 74 | if token == self.media_token_id: |
| 75 | new_input_ids.extend( |
| 76 | [self.media_token_id] * num_tokens_per_chunk.pop(0) |
| 77 | ) |
| 78 | else: |
| 79 | new_input_ids.append(token) |
| 80 | |
| 81 | input_ids[i] = new_input_ids |
| 82 | else: |
| 83 | text_inputs = {} |
| 84 | |
| 85 | return BatchFeature( |
| 86 | data={**text_inputs, **mm_inputs}, |
| 87 | tensor_type=return_tensors, |
| 88 | ) |