| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | from argparse import Namespace |
| 4 | |
| 5 | from starlette.datastructures import State |
| 6 | |
| 7 | from vllm.config import VllmConfig |
| 8 | from vllm.entrypoints.chat_utils import load_chat_template |
| 9 | from vllm.entrypoints.launchers.cli_args import resolve_default_chat_template_kwargs |
| 10 | from vllm.entrypoints.mcp.tool_server import init_tool_server |
| 11 | from vllm.entrypoints.openai.models.protocol import BaseModelPath |
| 12 | from vllm.entrypoints.openai.models.serving import OpenAIModelRegistry |
| 13 | from vllm.entrypoints.scale_out.factories import init_render_state |
| 14 | from vllm.entrypoints.serve.tokenize.serving import ServingTokenization |
| 15 | from vllm.entrypoints.serve.utils.request_logger import RequestLogger |
| 16 | from vllm.plugins.endpoint_plugins.interface import init_endpoint_plugins_state |
| 17 | from vllm.renderers import renderer_from_config |
| 18 | from vllm.renderers.online_derenderer import OnlineDerenderer |
| 19 | from vllm.renderers.online_renderer import OnlineRenderer |
| 20 | |
| 21 | |
| 22 | async def init_render_app_state( |
| 23 | vllm_config: VllmConfig, |
| 24 | state: State, |
| 25 | args: Namespace, |
| 26 | ) -> None: |
| 27 | """Initialise FastAPI app state for a CPU-only render server. |
| 28 | |
| 29 | Unlike :func:`init_app_state` this function does not require an |
| 30 | :class:`~vllm.engine.protocol.EngineClient`; it bootstraps the |
| 31 | preprocessing pipeline (renderer, input_processor) |
| 32 | directly from the :class:`~vllm.config.VllmConfig`. |
| 33 | """ |
| 34 | served_model_names = args.served_model_name or [args.model] |
| 35 | model_registry = OpenAIModelRegistry( |
| 36 | model_config=vllm_config.model_config, |
| 37 | base_model_paths=[ |
| 38 | BaseModelPath(name=name, model_path=args.model) |
| 39 | for name in served_model_names |
| 40 | ], |
| 41 | ) |
| 42 | |
| 43 | if args.enable_log_requests: |
| 44 | request_logger = RequestLogger(max_log_len=args.max_log_len) |
| 45 | else: |
| 46 | request_logger = None |
| 47 | |
| 48 | renderer = renderer_from_config(vllm_config) |
| 49 | resolved_chat_template = load_chat_template(args.chat_template) |
| 50 | default_chat_template_kwargs = resolve_default_chat_template_kwargs(args) |
| 51 | state.tool_server = await init_tool_server(args) |
| 52 | |
| 53 | state.online_renderer = OnlineRenderer( |
| 54 | model_config=vllm_config.model_config, |
| 55 | renderer=renderer, |
| 56 | request_logger=request_logger, |
| 57 | chat_template=resolved_chat_template, |
| 58 | chat_template_content_format=args.chat_template_content_format, |
| 59 | trust_request_chat_template=args.trust_request_chat_template, |
| 60 | trust_request_mm_kwargs=args.trust_request_mm_kwargs, |
| 61 | enable_auto_tools=args.enable_auto_tool_choice, |
| 62 | exclude_tools_when_tool_choice_none=args.exclude_tools_when_tool_choice_none, |
| 63 | tool_parser=args.tool_call_parser, |
| 64 | tool_strict_level=args.tool_strict_level, |
| 65 | reasoning_parser=args.reasoning_parser, |
| 66 | default_chat_template_kwargs=default_chat_template_kwargs, |
| 67 | log_error_stack=args.log_error_stack, |
| 68 | ) |
| 69 | state.online_renderer.warmup() |
| 70 | |
| 71 | state.online_derenderer = OnlineDerenderer( |
| 72 | model_config=vllm_config.model_config, |
| 73 | renderer=renderer, |
| 74 | request_logger=request_logger, |
| 75 | chat_template=resolved_chat_template, |
| 76 | chat_template_content_format=args.chat_template_content_format, |
| 77 | trust_request_chat_template=args.trust_request_chat_template, |
| 78 | enable_auto_tools=args.enable_auto_tool_choice, |
| 79 | exclude_tools_when_tool_choice_none=args.exclude_tools_when_tool_choice_none, |
| 80 | tool_parser=args.tool_call_parser, |
| 81 | tool_strict_level=args.tool_strict_level, |
| 82 | reasoning_parser=args.reasoning_parser, |
| 83 | default_chat_template_kwargs=default_chat_template_kwargs, |
| 84 | log_error_stack=args.log_error_stack, |
| 85 | ) |
| 86 | |
| 87 | state.openai_serving_models = model_registry |
| 88 | state.serving_tokenization = ServingTokenization( |
| 89 | model_registry, |
| 90 | state.online_renderer, |
| 91 | request_logger=request_logger, |
| 92 | chat_template=resolved_chat_template, |
| 93 | chat_template_content_format=args.chat_template_content_format, |
| 94 | default_chat_template_kwargs=default_chat_template_kwargs, |
| 95 | trust_request_chat_template=args.trust_request_chat_template, |
| 96 | ) |
| 97 | |
| 98 | init_render_state(state, request_logger) |
| 99 | |
| 100 | state.vllm_config = vllm_config |
| 101 | # Disable stats logging — there is no engine to poll. |
| 102 | state.log_stats = False |
| 103 | state.engine_client = None |
| 104 | state.args = args |
| 105 | state.enable_server_load_tracking = False |
| 106 | state.server_load_metrics = 0 |
| 107 | |
| 108 | # No `EngineClient` exists for the render server, so plugins get `None` and |
| 109 | # must handle it themselves (see `EndpointPlugin.init_state`). |
| 110 | await init_endpoint_plugins_state(None, state, args) |