| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | from dataclasses import dataclass |
| 4 | from typing import Any, Final |
| 5 | |
| 6 | from fastapi import Request |
| 7 | |
| 8 | from vllm.entrypoints.chat_utils import ChatTemplateContentFormatOption |
| 9 | from vllm.entrypoints.openai.models.serving import ( |
| 10 | OpenAIModelRegistry, |
| 11 | OpenAIServingModels, |
| 12 | ) |
| 13 | from vllm.entrypoints.serve.engine.protocol import ErrorResponse |
| 14 | from vllm.entrypoints.serve.engine.serving import BaseServing |
| 15 | from vllm.entrypoints.serve.tokenize.protocol import ( |
| 16 | DetokenizeRequest, |
| 17 | DetokenizeResponse, |
| 18 | TokenizeChatRequest, |
| 19 | TokenizeRequest, |
| 20 | TokenizeResponse, |
| 21 | TokenizerInfoResponse, |
| 22 | ) |
| 23 | from vllm.entrypoints.serve.utils.request_logger import RequestLogger |
| 24 | from vllm.inputs import TokensPrompt, tokens_input |
| 25 | from vllm.logger import init_logger |
| 26 | from vllm.renderers.online_renderer import OnlineRenderer |
| 27 | from vllm.tokenizers import TokenizerLike |
| 28 | |
| 29 | logger = init_logger(__name__) |
| 30 | |
| 31 | |
| 32 | class ServingTokenization(BaseServing): |
| 33 | def __init__( |
| 34 | self, |
| 35 | models: OpenAIServingModels | OpenAIModelRegistry, |
| 36 | online_renderer: OnlineRenderer, |
| 37 | *, |
| 38 | chat_template: str | None, |
| 39 | chat_template_content_format: ChatTemplateContentFormatOption, |
| 40 | default_chat_template_kwargs: dict[str, Any] | None = None, |
| 41 | trust_request_chat_template: bool = False, |
| 42 | request_logger: RequestLogger | None = None, |
| 43 | ) -> None: |
| 44 | super().__init__( |
| 45 | models=models, |
| 46 | model_config=online_renderer.model_config, |
| 47 | request_logger=request_logger, |
| 48 | ) |
| 49 | |
| 50 | self.renderer = online_renderer.renderer |
| 51 | self.online_renderer = online_renderer |
| 52 | self.chat_template = chat_template |
| 53 | self.chat_template_content_format: Final = chat_template_content_format |
| 54 | self.default_chat_template_kwargs = default_chat_template_kwargs or {} |
| 55 | self.trust_request_chat_template = trust_request_chat_template |
| 56 | |
| 57 | async def create_tokenize( |
| 58 | self, |
| 59 | request: TokenizeRequest, |
| 60 | raw_request: Request, |
| 61 | ) -> TokenizeResponse | ErrorResponse: |
| 62 | error_check_ret = await self._check_model(request) |
| 63 | if error_check_ret is not None: |
| 64 | return error_check_ret |
| 65 | |
| 66 | request_id = f"tokenize-{self._base_request_id(raw_request)}" |
| 67 | |
| 68 | lora_request = self._maybe_get_adapters(request) |
| 69 | |
| 70 | if isinstance(request, TokenizeChatRequest): |
| 71 | tool_dicts = ( |
| 72 | None |
| 73 | if request.tools is None |
| 74 | else [tool.model_dump() for tool in request.tools] |
| 75 | ) |
| 76 | error_check_ret = self.online_renderer.validate_chat_template( |
| 77 | request_chat_template=request.chat_template, |
| 78 | chat_template_kwargs=request.chat_template_kwargs, |
| 79 | trust_request_chat_template=self.trust_request_chat_template, |
| 80 | ) |
| 81 | if error_check_ret is not None: |
| 82 | return error_check_ret |
| 83 | |
| 84 | _, engine_inputs = await self.online_renderer.preprocess_chat( |
| 85 | request, |
| 86 | request.messages, |
| 87 | default_template=self.chat_template, |
| 88 | default_template_content_format=self.chat_template_content_format, |
| 89 | default_template_kwargs=self.default_chat_template_kwargs, |
| 90 | tool_dicts=tool_dicts, |
| 91 | skip_mm_cache=True, |
| 92 | ) |
| 93 | else: |
| 94 | engine_inputs = await self.online_renderer.preprocess_completion( |
| 95 | request, |
| 96 | prompt_input=request.prompt, |
| 97 | prompt_embeds=None, |
| 98 | skip_mm_cache=True, |
| 99 | ) |
| 100 | |
| 101 | input_ids: list[int] = [] |
| 102 | for engine_input in engine_inputs: |
| 103 | self._log_inputs( |
| 104 | request_id, |
| 105 | engine_input, |
| 106 | params=None, |
| 107 | lora_request=lora_request, |
| 108 | ) |
| 109 | |
| 110 | prompt_components = self._extract_prompt_components(engine_input) |
| 111 | if prompt_components.token_ids is not None: |
| 112 | input_ids.extend(prompt_components.token_ids) |
| 113 | |
| 114 | token_strs = None |
| 115 | if request.return_token_strs: |
| 116 | tokenizer = self.renderer.get_tokenizer() |
| 117 | token_strs = tokenizer.convert_ids_to_tokens(input_ids) |
| 118 | |
| 119 | return TokenizeResponse( |
| 120 | tokens=input_ids, |
| 121 | token_strs=token_strs, |
| 122 | count=len(input_ids), |
| 123 | max_model_len=self.model_config.max_model_len, |
| 124 | ) |
| 125 | |
| 126 | async def create_detokenize( |
| 127 | self, |
| 128 | request: DetokenizeRequest, |
| 129 | raw_request: Request, |
| 130 | ) -> DetokenizeResponse | ErrorResponse: |
| 131 | error_check_ret = await self._check_model(request) |
| 132 | if error_check_ret is not None: |
| 133 | return error_check_ret |
| 134 | |
| 135 | request_id = f"tokenize-{self._base_request_id(raw_request)}" |
| 136 | |
| 137 | lora_request = self._maybe_get_adapters(request) |
| 138 | |
| 139 | self._log_inputs( |
| 140 | request_id, |
| 141 | tokens_input(request.tokens), |
| 142 | params=None, |
| 143 | lora_request=lora_request, |
| 144 | ) |
| 145 | |
| 146 | tok_prompt = await self.renderer.tokenize_prompt_async( |
| 147 | TokensPrompt(prompt_token_ids=request.tokens), |
| 148 | request.build_tok_params(self.model_config), |
| 149 | ) |
| 150 | prompt_text = tok_prompt["prompt"] # type: ignore[typeddict-item] |
| 151 | |
| 152 | return DetokenizeResponse(prompt=prompt_text) |
| 153 | |
| 154 | async def get_tokenizer_info( |
| 155 | self, |
| 156 | ) -> TokenizerInfoResponse | ErrorResponse: |
| 157 | """Get comprehensive tokenizer information.""" |
| 158 | tokenizer = self.renderer.get_tokenizer() |
| 159 | info = TokenizerInfo(tokenizer, self.chat_template).to_dict() |
| 160 | return TokenizerInfoResponse(**info) |
| 161 | |
| 162 | |
| 163 | @dataclass |
| 164 | class TokenizerInfo: |
| 165 | tokenizer: TokenizerLike |
| 166 | chat_template: str | None |
| 167 | |
| 168 | def to_dict(self) -> dict[str, Any]: |
| 169 | """Return the tokenizer configuration.""" |
| 170 | return self._get_tokenizer_config() |
| 171 | |
| 172 | def _get_tokenizer_config(self) -> dict[str, Any]: |
| 173 | """Get tokenizer configuration directly from the tokenizer object.""" |
| 174 | config = dict(getattr(self.tokenizer, "init_kwargs", None) or {}) |
| 175 | |
| 176 | # Remove file path fields |
| 177 | config.pop("vocab_file", None) |
| 178 | config.pop("merges_file", None) |
| 179 | |
| 180 | config = self._make_json_serializable(config) |
| 181 | config["tokenizer_class"] = type(self.tokenizer).__name__ |
| 182 | if self.chat_template: |
| 183 | config["chat_template"] = self.chat_template |
| 184 | return config |
| 185 | |
| 186 | def _make_json_serializable(self, obj): |
| 187 | """Convert any non-JSON-serializable objects to serializable format.""" |
| 188 | if hasattr(obj, "content"): |
| 189 | return obj.content |
| 190 | elif isinstance(obj, dict): |
| 191 | return {k: self._make_json_serializable(v) for k, v in obj.items()} |
| 192 | elif isinstance(obj, list): |
| 193 | return [self._make_json_serializable(item) for item in obj] |
| 194 | else: |
| 195 | return obj |