| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | |
| 5 | from typing import Annotated, Any, TypeAlias |
| 6 | |
| 7 | from pydantic import ConfigDict, Field, model_validator |
| 8 | |
| 9 | from vllm.config import ModelConfig |
| 10 | from vllm.entrypoints.chat_utils import ( |
| 11 | ChatCompletionMessageParam, |
| 12 | ChatTemplateContentFormatOption, |
| 13 | ) |
| 14 | from vllm.entrypoints.openai.chat_completion.protocol import ChatCompletionToolsParam |
| 15 | from vllm.entrypoints.serve.engine.protocol import OpenAIBaseModel |
| 16 | from vllm.exceptions import VLLMValidationError |
| 17 | from vllm.renderers import ChatParams, TokenizeParams, merge_kwargs |
| 18 | |
| 19 | |
| 20 | class TokenizeCompletionRequest(OpenAIBaseModel): |
| 21 | model: str | None = None |
| 22 | prompt: str |
| 23 | |
| 24 | add_special_tokens: bool = Field( |
| 25 | default=True, |
| 26 | description=( |
| 27 | "If true (the default), special tokens (e.g. BOS) will be added to " |
| 28 | "the prompt." |
| 29 | ), |
| 30 | ) |
| 31 | return_token_strs: bool | None = Field( |
| 32 | default=False, |
| 33 | description=( |
| 34 | "If true, also return the token strings corresponding to the token ids." |
| 35 | ), |
| 36 | ) |
| 37 | |
| 38 | def build_tok_params(self, model_config: ModelConfig) -> TokenizeParams: |
| 39 | return TokenizeParams( |
| 40 | max_total_tokens=None, |
| 41 | max_output_tokens=0, |
| 42 | add_special_tokens=self.add_special_tokens, |
| 43 | ) |
| 44 | |
| 45 | |
| 46 | class TokenizeChatRequest(OpenAIBaseModel): |
| 47 | model: str | None = None |
| 48 | messages: list[ChatCompletionMessageParam] |
| 49 | |
| 50 | add_generation_prompt: bool = Field( |
| 51 | default=True, |
| 52 | description=( |
| 53 | "If true, the generation prompt will be added to the chat template. " |
| 54 | "This is a parameter used by chat template in tokenizer config of the " |
| 55 | "model." |
| 56 | ), |
| 57 | ) |
| 58 | return_token_strs: bool | None = Field( |
| 59 | default=False, |
| 60 | description=( |
| 61 | "If true, also return the token strings corresponding to the token ids." |
| 62 | ), |
| 63 | ) |
| 64 | continue_final_message: bool = Field( |
| 65 | default=False, |
| 66 | description=( |
| 67 | "If this is set, the chat will be formatted so that the final " |
| 68 | "message in the chat is open-ended, without any EOS tokens. The " |
| 69 | "model will continue this message rather than starting a new one. " |
| 70 | 'This allows you to "prefill" part of the model\'s response for it. ' |
| 71 | "Cannot be used at the same time as `add_generation_prompt`." |
| 72 | ), |
| 73 | ) |
| 74 | add_special_tokens: bool = Field( |
| 75 | default=False, |
| 76 | description=( |
| 77 | "If true, special tokens (e.g. BOS) will be added to the prompt " |
| 78 | "on top of what is added by the chat template. " |
| 79 | "For most models, the chat template takes care of adding the " |
| 80 | "special tokens so this should be set to false (as is the " |
| 81 | "default)." |
| 82 | ), |
| 83 | ) |
| 84 | chat_template: str | None = Field( |
| 85 | default=None, |
| 86 | description=( |
| 87 | "A Jinja template to use for this conversion. " |
| 88 | "As of transformers v4.44, default chat template is no longer " |
| 89 | "allowed, so you must provide a chat template if the tokenizer " |
| 90 | "does not define one." |
| 91 | ), |
| 92 | ) |
| 93 | chat_template_kwargs: dict[str, Any] | None = Field( |
| 94 | default=None, |
| 95 | description=( |
| 96 | "Additional keyword args to pass to the template renderer. " |
| 97 | "Will be accessible by the chat template." |
| 98 | ), |
| 99 | ) |
| 100 | media_io_kwargs: dict[str, dict[str, Any]] | None = Field( |
| 101 | default=None, |
| 102 | description=( |
| 103 | "Additional kwargs to pass to the media IO connectors, " |
| 104 | "keyed by modality. Merged with engine-level media_io_kwargs." |
| 105 | ), |
| 106 | ) |
| 107 | mm_processor_kwargs: dict[str, Any] | None = Field( |
| 108 | default=None, |
| 109 | description="Additional kwargs to pass to the HF processor.", |
| 110 | ) |
| 111 | tools: list[ChatCompletionToolsParam] | None = Field( |
| 112 | default=None, |
| 113 | description="A list of tools the model may call.", |
| 114 | ) |
| 115 | |
| 116 | @model_validator(mode="before") |
| 117 | @classmethod |
| 118 | def check_generation_prompt(cls, data): |
| 119 | if not isinstance(data, dict): |
| 120 | return data |
| 121 | if data.get("continue_final_message") and data.get("add_generation_prompt"): |
| 122 | raise VLLMValidationError( |
| 123 | "Cannot set both `continue_final_message` and " |
| 124 | "`add_generation_prompt` to True.", |
| 125 | ) |
| 126 | return data |
| 127 | |
| 128 | def build_chat_params( |
| 129 | self, |
| 130 | default_template: str | None, |
| 131 | default_template_content_format: ChatTemplateContentFormatOption, |
| 132 | ) -> ChatParams: |
| 133 | return ChatParams( |
| 134 | chat_template=self.chat_template or default_template, |
| 135 | chat_template_content_format=default_template_content_format, |
| 136 | chat_template_kwargs=merge_kwargs( |
| 137 | self.chat_template_kwargs, |
| 138 | dict( |
| 139 | add_generation_prompt=self.add_generation_prompt, |
| 140 | continue_final_message=self.continue_final_message, |
| 141 | ), |
| 142 | ), |
| 143 | media_io_kwargs=self.media_io_kwargs, |
| 144 | ) |
| 145 | |
| 146 | def build_tok_params(self, model_config: ModelConfig) -> TokenizeParams: |
| 147 | return TokenizeParams( |
| 148 | max_total_tokens=None, |
| 149 | max_output_tokens=0, |
| 150 | add_special_tokens=self.add_special_tokens, |
| 151 | ) |
| 152 | |
| 153 | |
| 154 | TokenizeRequest: TypeAlias = TokenizeCompletionRequest | TokenizeChatRequest |
| 155 | |
| 156 | |
| 157 | class TokenizeResponse(OpenAIBaseModel): |
| 158 | count: int |
| 159 | max_model_len: int |
| 160 | tokens: list[int] |
| 161 | token_strs: list[str] | None = None |
| 162 | |
| 163 | |
| 164 | class DetokenizeRequest(OpenAIBaseModel): |
| 165 | model: str | None = None |
| 166 | # TODO: Factor `torch.iinfo` out. `torch.iinfo` pulls torch into a |
| 167 | # Pydantic protocol file that currently has no torch dependency. |
| 168 | # See: https://github.com/vllm-project/vllm/pull/34468#discussion_r2801173630 |
| 169 | tokens: list[Annotated[int, Field(ge=0, le=2**63 - 1)]] |
| 170 | |
| 171 | def build_tok_params(self, model_config: ModelConfig) -> TokenizeParams: |
| 172 | return TokenizeParams( |
| 173 | max_total_tokens=None, |
| 174 | max_output_tokens=0, |
| 175 | needs_detokenization=True, |
| 176 | ) |
| 177 | |
| 178 | |
| 179 | class DetokenizeResponse(OpenAIBaseModel): |
| 180 | prompt: str |
| 181 | |
| 182 | |
| 183 | class TokenizerInfoResponse(OpenAIBaseModel): |
| 184 | """Response containing tokenizer configuration |
| 185 | equivalent to tokenizer_config.json |
| 186 | """ |
| 187 | |
| 188 | model_config = ConfigDict(extra="allow") |
| 189 | tokenizer_class: str |