| 1 | """Schema and utilities for input prompts to the LLM API.""" |
| 2 | |
| 3 | # SPDX-License-Identifier: Apache-2.0 |
| 4 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 5 | from collections.abc import Mapping, Sequence |
| 6 | from typing import TYPE_CHECKING, Any, TypeAlias, TypeVar, final |
| 7 | |
| 8 | from typing_extensions import NotRequired, TypedDict |
| 9 | |
| 10 | if TYPE_CHECKING: |
| 11 | import torch |
| 12 | |
| 13 | from vllm.multimodal.inputs import AudioItem, ImageItem, VideoItem, VisionChunk |
| 14 | |
| 15 | |
| 16 | _T = TypeVar("_T") |
| 17 | |
| 18 | ModalityData: TypeAlias = _T | list[_T | None] | None |
| 19 | """ |
| 20 | Either a single data item, or a list of data items. Can only be None if UUID |
| 21 | is provided. |
| 22 | |
| 23 | The number of data items allowed per modality is restricted by |
| 24 | `--limit-mm-per-prompt`. |
| 25 | """ |
| 26 | |
| 27 | |
| 28 | @final |
| 29 | class MultiModalDataBuiltins(TypedDict, total=False): |
| 30 | """Type annotations for modality types predefined by vLLM.""" |
| 31 | |
| 32 | image: ModalityData["ImageItem"] |
| 33 | """The input image(s).""" |
| 34 | |
| 35 | video: ModalityData["VideoItem"] |
| 36 | """The input video(s).""" |
| 37 | |
| 38 | audio: ModalityData["AudioItem"] |
| 39 | """The input audio(s).""" |
| 40 | |
| 41 | vision_chunk: ModalityData["VisionChunk"] |
| 42 | """The input visual atom(s) - unified modality for images and video chunks.""" |
| 43 | |
| 44 | |
| 45 | MultiModalDataDict: TypeAlias = Mapping[str, ModalityData[Any]] |
| 46 | """ |
| 47 | A dictionary containing an entry for each modality type to input. |
| 48 | |
| 49 | The built-in modalities are defined by |
| 50 | [`MultiModalDataBuiltins`][vllm.inputs.llm.MultiModalDataBuiltins]. |
| 51 | """ |
| 52 | |
| 53 | MultiModalUUIDDict: TypeAlias = Mapping[str, Sequence[str | None] | str] |
| 54 | """ |
| 55 | A dictionary containing user-provided UUIDs for items in each modality. |
| 56 | If a UUID for an item is not provided, its entry will be `None` and |
| 57 | MultiModalHasher will compute a hash for the item. |
| 58 | |
| 59 | The UUID will be used to identify the item for all caching purposes |
| 60 | (input processing caching, embedding caching, prefix caching, etc). |
| 61 | """ |
| 62 | |
| 63 | |
| 64 | class _PromptOptions(TypedDict): |
| 65 | """Additional options available to all |
| 66 | [`SingletonPrompt`][vllm.inputs.llm.SingletonPrompt] types. |
| 67 | """ |
| 68 | |
| 69 | multi_modal_data: NotRequired[MultiModalDataDict | None] |
| 70 | """ |
| 71 | Optional multi-modal data to pass to the model, |
| 72 | if the model supports it. |
| 73 | """ |
| 74 | |
| 75 | media_io_kwargs: NotRequired[dict[str, dict[str, Any]] | None] |
| 76 | """ |
| 77 | Optional per-modality media loading and decoding arguments. |
| 78 | Used only for hash derivation when using multi-modal UUIDs. |
| 79 | """ |
| 80 | |
| 81 | mm_processor_kwargs: NotRequired[dict[str, Any] | None] |
| 82 | """ |
| 83 | Optional multi-modal processor kwargs to be forwarded to the |
| 84 | multimodal input mapper & processor. Note that if multiple modalities |
| 85 | have registered mappers etc for the model being considered, we attempt |
| 86 | to pass the mm_processor_kwargs to each of them. |
| 87 | """ |
| 88 | |
| 89 | multi_modal_uuids: NotRequired[MultiModalUUIDDict] |
| 90 | """ |
| 91 | Optional user-specified UUIDs for multimodal items, mapped by modality. |
| 92 | Lists must match the number of items per modality and may contain `None`. |
| 93 | For `None` entries, the hasher will compute IDs automatically; non-None |
| 94 | entries override the default hashes for caching, and MUST be unique per |
| 95 | multimodal item. |
| 96 | """ |
| 97 | |
| 98 | cache_salt: NotRequired[str] |
| 99 | """ |
| 100 | Optional cache salt to be used for prefix caching. |
| 101 | """ |
| 102 | |
| 103 | |
| 104 | class TextPrompt(_PromptOptions): |
| 105 | """Schema for a text prompt.""" |
| 106 | |
| 107 | prompt: str |
| 108 | """The input text to be tokenized before passing to the model.""" |
| 109 | |
| 110 | |
| 111 | class TokensPrompt(_PromptOptions): |
| 112 | """Schema for a tokenized prompt.""" |
| 113 | |
| 114 | prompt_token_ids: list[int] |
| 115 | """A list of token IDs to pass to the model.""" |
| 116 | |
| 117 | prompt: NotRequired[str] |
| 118 | """The prompt text corresponding to the token IDs, if available.""" |
| 119 | |
| 120 | token_type_ids: NotRequired[list[int]] |
| 121 | """A list of token type IDs to pass to the cross encoder model.""" |
| 122 | |
| 123 | prompt_token_offsets: NotRequired[list[tuple[int, int]] | None] |
| 124 | """Char-level (start, end) offsets per token, relative to the |
| 125 | tokenized source string. Present only when offsets were requested |
| 126 | AND a Fast (Rust-backed) tokenizer was used AND no multimodal data |
| 127 | was present. The list length equals the length of `prompt_token_ids`.""" |
| 128 | |
| 129 | |
| 130 | class EmbedsPrompt(_PromptOptions): |
| 131 | """Schema for a prompt provided via token embeddings.""" |
| 132 | |
| 133 | prompt_embeds: "torch.Tensor" |
| 134 | """The embeddings of the prompt.""" |
| 135 | |
| 136 | prompt: NotRequired[str] |
| 137 | """The prompt text corresponding to the token embeddings, if available.""" |
| 138 | |
| 139 | prompt_token_ids: NotRequired[list[int]] |
| 140 | """Token IDs for mixed-mode inputs (chat completion with |
| 141 | `prompt_embeds` content parts). The tokens at positions where |
| 142 | `prompt_is_token_ids` is `False` are placeholder tokens that |
| 143 | get replaced by entries from `prompt_embeds` in the forward pass.""" |
| 144 | |
| 145 | prompt_is_token_ids: NotRequired[list[bool]] |
| 146 | """Per-position mask, `True` uses the real token ID, `False` uses |
| 147 | the corresponding entry from `prompt_embeds`. |
| 148 | Must be the same length as `prompt_token_ids` when both are set.""" |
| 149 | |
| 150 | |
| 151 | DecoderOnlyPrompt: TypeAlias = ( |
| 152 | str | TextPrompt | list[int] | TokensPrompt | EmbedsPrompt |
| 153 | ) |
| 154 | """ |
| 155 | Schema of a prompt for a decoder-only model: |
| 156 | |
| 157 | - A text prompt (string or [`TextPrompt`][vllm.inputs.llm.TextPrompt]) |
| 158 | - A tokenized prompt (list of token IDs, or |
| 159 | [`TokensPrompt`][vllm.inputs.llm.TokensPrompt]) |
| 160 | - An embeddings prompt ([`EmbedsPrompt`][vllm.inputs.llm.EmbedsPrompt]) |
| 161 | |
| 162 | For encoder-decoder models, passing a singleton prompt is shorthand for passing |
| 163 | `ExplicitEncoderDecoderPrompt(encoder_prompt=prompt, decoder_prompt=None)`. |
| 164 | """ |
| 165 | |
| 166 | |
| 167 | EncoderPrompt: TypeAlias = str | TextPrompt | list[int] | TokensPrompt |
| 168 | """ |
| 169 | Schema of a prompt for the encoder part of a encoder-decoder model: |
| 170 | |
| 171 | - A text prompt (string or [`TextPrompt`][vllm.inputs.llm.TextPrompt]) |
| 172 | - A tokenized prompt (list of token IDs, or |
| 173 | [`TokensPrompt`][vllm.inputs.llm.TokensPrompt]) |
| 174 | """ |
| 175 | |
| 176 | |
| 177 | DecoderPrompt: TypeAlias = str | TextPrompt | list[int] | TokensPrompt |
| 178 | """ |
| 179 | Schema of a prompt for the decoder part of an encoder-decoder model: |
| 180 | |
| 181 | - A text prompt (string or [`TextPrompt`][vllm.inputs.llm.TextPrompt]) |
| 182 | - A tokenized prompt (list of token IDs, or |
| 183 | [`TokensPrompt`][vllm.inputs.llm.TokensPrompt]) |
| 184 | |
| 185 | Note: |
| 186 | Multi-modal inputs are not supported for decoder prompts. |
| 187 | """ |
| 188 | |
| 189 | |
| 190 | class ExplicitEncoderDecoderPrompt(TypedDict): |
| 191 | """Schema for a pair of encoder and decoder singleton prompts. |
| 192 | |
| 193 | Note: |
| 194 | This schema is not valid for decoder-only models. |
| 195 | |
| 196 | """ |
| 197 | |
| 198 | encoder_prompt: EncoderPrompt |
| 199 | """The prompt for the encoder part of the model.""" |
| 200 | |
| 201 | decoder_prompt: DecoderPrompt | None |
| 202 | """ |
| 203 | The prompt for the decoder part of the model. |
| 204 | |
| 205 | Passing `None` will cause the prompt to be inferred automatically. |
| 206 | """ |
| 207 | |
| 208 | |
| 209 | EncoderDecoderPrompt: TypeAlias = EncoderPrompt | ExplicitEncoderDecoderPrompt |
| 210 | """ |
| 211 | Schema for a prompt for an encoder-decoder model. |
| 212 | |
| 213 | You can pass a singleton encoder prompt, in which case the decoder prompt is |
| 214 | considered to be `None` (i.e., infer automatically). |
| 215 | """ |
| 216 | |
| 217 | |
| 218 | SingletonPrompt: TypeAlias = DecoderOnlyPrompt | EncoderPrompt | DecoderPrompt |
| 219 | """ |
| 220 | Schema for a single prompt. This is as opposed to a data structure |
| 221 | which encapsulates multiple prompts, such as |
| 222 | [`ExplicitEncoderDecoderPrompt`][vllm.inputs.llm.ExplicitEncoderDecoderPrompt]. |
| 223 | """ |
| 224 | |
| 225 | |
| 226 | PromptType: TypeAlias = DecoderOnlyPrompt | EncoderDecoderPrompt |
| 227 | """ |
| 228 | Schema for any prompt, regardless of model type. |
| 229 | |
| 230 | This is the input format accepted by most [`LLM`][vllm.entrypoints.llm.LLM] APIs. |
| 231 | """ |
| 232 | |
| 233 | |
| 234 | class DataPrompt(_PromptOptions): |
| 235 | """Represents generic inputs that are converted to |
| 236 | [`PromptType`][vllm.inputs.llm.PromptType] by IO processor plugins. |
| 237 | """ |
| 238 | |
| 239 | data: Any |
| 240 | """The input data.""" |
| 241 | |
| 242 | data_format: str |
| 243 | """The input data format.""" |