| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | from typing import TYPE_CHECKING, Any |
| 4 | |
| 5 | from fastapi import FastAPI |
| 6 | |
| 7 | import vllm.envs as envs |
| 8 | |
| 9 | if TYPE_CHECKING: |
| 10 | from argparse import Namespace |
| 11 | |
| 12 | from starlette.datastructures import State |
| 13 | |
| 14 | from vllm.engine.protocol import EngineClient |
| 15 | from vllm.entrypoints.serve.utils.request_logger import RequestLogger |
| 16 | from vllm.tasks import SupportedTask |
| 17 | else: |
| 18 | RequestLogger = object |
| 19 | |
| 20 | |
| 21 | def register_generate_api_routers(app: FastAPI): |
| 22 | from vllm.entrypoints.openai.chat_completion.api_router import ( |
| 23 | attach_router as register_chat_api_router, |
| 24 | ) |
| 25 | |
| 26 | register_chat_api_router(app) |
| 27 | |
| 28 | from vllm.entrypoints.openai.responses.api_router import ( |
| 29 | attach_router as register_responses_api_router, |
| 30 | ) |
| 31 | |
| 32 | register_responses_api_router(app) |
| 33 | |
| 34 | from vllm.entrypoints.openai.completion.api_router import ( |
| 35 | attach_router as register_completion_api_router, |
| 36 | ) |
| 37 | |
| 38 | register_completion_api_router(app) |
| 39 | |
| 40 | from vllm.entrypoints.anthropic.api_router import ( |
| 41 | attach_router as register_anthropic_api_router, |
| 42 | ) |
| 43 | |
| 44 | register_anthropic_api_router(app) |
| 45 | |
| 46 | from vllm.entrypoints.cohere.api_router import ( |
| 47 | attach_router as register_cohere_api_router, |
| 48 | ) |
| 49 | |
| 50 | register_cohere_api_router(app) |
| 51 | |
| 52 | from .generative_scoring.api_router import register_generative_scoring_api_router |
| 53 | |
| 54 | register_generative_scoring_api_router(app) |
| 55 | |
| 56 | |
| 57 | async def init_generate_state( |
| 58 | engine_client: "EngineClient", |
| 59 | state: "State", |
| 60 | args: "Namespace", |
| 61 | request_logger: RequestLogger | None, |
| 62 | supported_tasks: tuple["SupportedTask", ...], |
| 63 | default_chat_template_kwargs: dict[str, Any], |
| 64 | ): |
| 65 | from vllm.entrypoints.anthropic.serving import AnthropicServingMessages |
| 66 | from vllm.entrypoints.chat_utils import load_chat_template |
| 67 | |
| 68 | # The Cohere serving handler depends on the optional `cohere` SDK for |
| 69 | # its wire-format protocol models, and is additionally gated on the |
| 70 | # `VLLM_ENABLE_COHERE_API` env flag (see |
| 71 | # `vllm.entrypoints.cohere.api_router.attach_router`). Skip the import |
| 72 | # entirely when the endpoint isn't going to be exposed, both because |
| 73 | # the SDK may not be installed and because the serving object holds |
| 74 | # nontrivial state (chat handler, warmup) that would otherwise be |
| 75 | # unused. |
| 76 | if envs.VLLM_ENABLE_COHERE_API: |
| 77 | try: |
| 78 | from vllm.entrypoints.cohere.serving import CohereServingChatV2 |
| 79 | except ImportError: |
| 80 | CohereServingChatV2 = None # type: ignore[assignment,misc] |
| 81 | else: |
| 82 | CohereServingChatV2 = None # type: ignore[assignment,misc] |
| 83 | from vllm.entrypoints.openai.chat_completion.batch_serving import ( |
| 84 | OpenAIServingChatBatch, |
| 85 | ) |
| 86 | from vllm.entrypoints.openai.chat_completion.serving import OpenAIServingChat |
| 87 | from vllm.entrypoints.openai.completion.serving import OpenAIServingCompletion |
| 88 | from vllm.entrypoints.openai.responses.serving import OpenAIServingResponses |
| 89 | from vllm.entrypoints.serve.utils.fingerprint import set_default_fingerprint_mode |
| 90 | |
| 91 | # Applied before any serving class is constructed so that each one picks |
| 92 | # up the chosen mode on its first cache miss. |
| 93 | set_default_fingerprint_mode( |
| 94 | getattr(args, "fingerprint_mode", "full"), |
| 95 | getattr(args, "fingerprint_value", None), |
| 96 | ) |
| 97 | |
| 98 | resolved_chat_template = load_chat_template(args.chat_template) |
| 99 | |
| 100 | # Render endpoints are always backed by OnlineRenderer so that chat, |
| 101 | # completion, and Responses rendering work on both generate-mode and |
| 102 | # render-only servers. Created in init_app_state. |
| 103 | |
| 104 | state.openai_serving_responses = ( |
| 105 | OpenAIServingResponses( |
| 106 | engine_client, |
| 107 | state.openai_serving_models, |
| 108 | state.online_renderer, |
| 109 | request_logger=request_logger, |
| 110 | chat_template=resolved_chat_template, |
| 111 | chat_template_content_format=args.chat_template_content_format, |
| 112 | return_tokens_as_token_ids=args.return_tokens_as_token_ids, |
| 113 | enable_auto_tools=args.enable_auto_tool_choice, |
| 114 | tool_parser=args.tool_call_parser, |
| 115 | tool_strict_level=args.tool_strict_level, |
| 116 | tool_server=state.tool_server, |
| 117 | reasoning_parser=args.structured_outputs_config.reasoning_parser, |
| 118 | enable_prompt_tokens_details=args.enable_prompt_tokens_details, |
| 119 | enable_force_include_usage=args.enable_force_include_usage, |
| 120 | enable_per_request_metrics=args.enable_per_request_metrics, |
| 121 | enable_log_outputs=args.enable_log_outputs, |
| 122 | default_chat_template_kwargs=default_chat_template_kwargs, |
| 123 | ) |
| 124 | if "generate" in supported_tasks |
| 125 | else None |
| 126 | ) |
| 127 | _chat_kwargs = dict( |
| 128 | engine_client=engine_client, |
| 129 | models=state.openai_serving_models, |
| 130 | response_role=args.response_role, |
| 131 | online_renderer=state.online_renderer, |
| 132 | request_logger=request_logger, |
| 133 | chat_template=resolved_chat_template, |
| 134 | chat_template_content_format=args.chat_template_content_format, |
| 135 | default_chat_template_kwargs=default_chat_template_kwargs, |
| 136 | trust_request_chat_template=args.trust_request_chat_template, |
| 137 | return_tokens_as_token_ids=args.return_tokens_as_token_ids, |
| 138 | enable_auto_tools=args.enable_auto_tool_choice, |
| 139 | exclude_tools_when_tool_choice_none=args.exclude_tools_when_tool_choice_none, |
| 140 | tool_parser=args.tool_call_parser, |
| 141 | tool_strict_level=args.tool_strict_level, |
| 142 | reasoning_parser=args.structured_outputs_config.reasoning_parser, |
| 143 | enable_prompt_tokens_details=args.enable_prompt_tokens_details, |
| 144 | enable_force_include_usage=args.enable_force_include_usage, |
| 145 | enable_log_outputs=args.enable_log_outputs, |
| 146 | enable_log_deltas=args.enable_log_deltas, |
| 147 | enable_per_request_metrics=args.enable_per_request_metrics, |
| 148 | ) |
| 149 | state.openai_serving_chat = ( |
| 150 | OpenAIServingChat(**_chat_kwargs) if "generate" in supported_tasks else None |
| 151 | ) |
| 152 | state.openai_serving_chat_batch = ( |
| 153 | OpenAIServingChatBatch(**_chat_kwargs) |
| 154 | if "generate" in supported_tasks |
| 155 | else None |
| 156 | ) |
| 157 | state.openai_serving_completion = ( |
| 158 | OpenAIServingCompletion( |
| 159 | engine_client, |
| 160 | state.openai_serving_models, |
| 161 | online_renderer=state.online_renderer, |
| 162 | request_logger=request_logger, |
| 163 | return_tokens_as_token_ids=args.return_tokens_as_token_ids, |
| 164 | enable_prompt_tokens_details=args.enable_prompt_tokens_details, |
| 165 | enable_force_include_usage=args.enable_force_include_usage, |
| 166 | enable_per_request_metrics=args.enable_per_request_metrics, |
| 167 | ) |
| 168 | if "generate" in supported_tasks |
| 169 | else None |
| 170 | ) |
| 171 | state.anthropic_serving_messages = ( |
| 172 | AnthropicServingMessages( |
| 173 | engine_client, |
| 174 | state.openai_serving_models, |
| 175 | args.response_role, |
| 176 | online_renderer=state.online_renderer, |
| 177 | request_logger=request_logger, |
| 178 | chat_template=resolved_chat_template, |
| 179 | chat_template_content_format=args.chat_template_content_format, |
| 180 | return_tokens_as_token_ids=args.return_tokens_as_token_ids, |
| 181 | enable_auto_tools=args.enable_auto_tool_choice, |
| 182 | tool_parser=args.tool_call_parser, |
| 183 | tool_strict_level=args.tool_strict_level, |
| 184 | reasoning_parser=args.structured_outputs_config.reasoning_parser, |
| 185 | enable_prompt_tokens_details=args.enable_prompt_tokens_details, |
| 186 | enable_force_include_usage=args.enable_force_include_usage, |
| 187 | default_chat_template_kwargs=default_chat_template_kwargs, |
| 188 | disabled_thinking_effort=args.anthropic_disabled_thinking_effort, |
| 189 | ) |
| 190 | if "generate" in supported_tasks |
| 191 | else None |
| 192 | ) |
| 193 | state.cohere_serving_chat_v2 = ( |
| 194 | CohereServingChatV2( |
| 195 | engine_client, |
| 196 | state.openai_serving_models, |
| 197 | args.response_role, |
| 198 | online_renderer=state.online_renderer, |
| 199 | request_logger=request_logger, |
| 200 | chat_template=resolved_chat_template, |
| 201 | chat_template_content_format=args.chat_template_content_format, |
| 202 | return_tokens_as_token_ids=args.return_tokens_as_token_ids, |
| 203 | enable_auto_tools=args.enable_auto_tool_choice, |
| 204 | tool_parser=args.tool_call_parser, |
| 205 | tool_strict_level=args.tool_strict_level, |
| 206 | reasoning_parser=args.structured_outputs_config.reasoning_parser, |
| 207 | enable_prompt_tokens_details=args.enable_prompt_tokens_details, |
| 208 | enable_force_include_usage=args.enable_force_include_usage, |
| 209 | default_chat_template_kwargs=default_chat_template_kwargs, |
| 210 | is_reasoning_model=args.cohere_is_reasoning_model, |
| 211 | ) |
| 212 | if CohereServingChatV2 is not None and "generate" in supported_tasks |
| 213 | else None |
| 214 | ) |
| 215 | |
| 216 | from .generative_scoring.serving import ServingGenerativeScoring |
| 217 | |
| 218 | state.serving_generative_scoring = ServingGenerativeScoring( |
| 219 | engine_client, |
| 220 | state.openai_serving_models, |
| 221 | request_logger=request_logger, |
| 222 | ) |