| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import asyncio |
| 5 | import time |
| 6 | from collections import deque |
| 7 | from collections.abc import AsyncGenerator, AsyncIterator, Callable, Mapping, Sequence |
| 8 | from contextlib import AsyncExitStack |
| 9 | from http import HTTPStatus |
| 10 | from typing import Any, Final, cast |
| 11 | |
| 12 | from fastapi import Request |
| 13 | from openai.types.responses import ( |
| 14 | ResponseOutputItem, |
| 15 | ResponseOutputMessage, |
| 16 | ResponseOutputText, |
| 17 | ResponseStatus, |
| 18 | ) |
| 19 | from openai.types.responses.response_output_text import Logprob, LogprobTopLogprob |
| 20 | from pydantic import TypeAdapter |
| 21 | |
| 22 | from vllm import envs |
| 23 | from vllm.config.utils import replace |
| 24 | from vllm.engine.protocol import EngineClient |
| 25 | from vllm.entrypoints.chat_utils import ( |
| 26 | ChatTemplateContentFormatOption, |
| 27 | ) |
| 28 | from vllm.entrypoints.generate.base.protocol import ( |
| 29 | DeltaMessage, |
| 30 | RequestResponseMetadata, |
| 31 | ) |
| 32 | from vllm.entrypoints.generate.base.serving import ( |
| 33 | GenerateBaseServing, |
| 34 | build_per_request_timing_metrics, |
| 35 | ) |
| 36 | from vllm.entrypoints.mcp.tool_server import ToolServer |
| 37 | from vllm.entrypoints.openai.models.serving import OpenAIServingModels |
| 38 | from vllm.entrypoints.openai.responses.context import ( |
| 39 | ConversationContext, |
| 40 | HarmonyContext, |
| 41 | ParsableContext, |
| 42 | SimpleContext, |
| 43 | ) |
| 44 | from vllm.entrypoints.openai.responses.harmony import harmony_to_response_output |
| 45 | from vllm.entrypoints.openai.responses.protocol import ( |
| 46 | InputTokensDetails, |
| 47 | OutputTokensDetails, |
| 48 | ResponseCompletedEvent, |
| 49 | ResponseCreatedEvent, |
| 50 | ResponseInProgressEvent, |
| 51 | ResponseInputOutputItem, |
| 52 | ResponseInputOutputMessage, |
| 53 | ResponsesRequest, |
| 54 | ResponsesResponse, |
| 55 | ResponseUsage, |
| 56 | StreamingResponsesResponse, |
| 57 | ) |
| 58 | from vllm.entrypoints.openai.responses.streaming_events import ( |
| 59 | SimpleStreamingEventProcessor, |
| 60 | StreamingState, |
| 61 | _StateType, |
| 62 | emit_content_delta_events, |
| 63 | emit_previous_item_done_events, |
| 64 | emit_tool_action_events, |
| 65 | split_delta, |
| 66 | ) |
| 67 | from vllm.entrypoints.openai.responses.utils import ( |
| 68 | build_response_output_items, |
| 69 | extract_function_tool_names, |
| 70 | extract_tool_types, |
| 71 | ) |
| 72 | from vllm.entrypoints.serve.engine.protocol import ErrorResponse |
| 73 | from vllm.entrypoints.serve.utils.api_utils import get_max_tokens |
| 74 | from vllm.entrypoints.serve.utils.request_logger import RequestLogger |
| 75 | from vllm.entrypoints.serve.utils.tool_calls_utils import ( |
| 76 | maybe_filter_parallel_tool_calls, |
| 77 | ) |
| 78 | from vllm.exceptions import GenerationError, VLLMValidationError |
| 79 | from vllm.inputs import EngineInput |
| 80 | from vllm.logger import init_logger |
| 81 | from vllm.logprobs import Logprob as SampleLogprob |
| 82 | from vllm.logprobs import SampleLogprobs |
| 83 | from vllm.lora.request import LoRARequest |
| 84 | from vllm.outputs import CompletionOutput |
| 85 | from vllm.parser import Parser, ParserManager |
| 86 | from vllm.renderers import TokenizeParams |
| 87 | from vllm.renderers.online_renderer import ( |
| 88 | OnlineRenderer, |
| 89 | ResponsesPreviousMessages, |
| 90 | ResponsesRenderResult, |
| 91 | ) |
| 92 | from vllm.sampling_params import SamplingParams, StructuredOutputsParams |
| 93 | from vllm.tokenizers import TokenizerLike |
| 94 | from vllm.utils import random_uuid |
| 95 | from vllm.utils.collection_utils import as_list |
| 96 | |
| 97 | logger = init_logger(__name__) |
| 98 | |
| 99 | |
| 100 | class OpenAIServingResponses(GenerateBaseServing): |
| 101 | def __init__( |
| 102 | self, |
| 103 | engine_client: EngineClient, |
| 104 | models: OpenAIServingModels, |
| 105 | online_renderer: OnlineRenderer, |
| 106 | *, |
| 107 | request_logger: RequestLogger | None, |
| 108 | chat_template: str | None, |
| 109 | chat_template_content_format: ChatTemplateContentFormatOption, |
| 110 | return_tokens_as_token_ids: bool = False, |
| 111 | reasoning_parser: str = "", |
| 112 | enable_auto_tools: bool = False, |
| 113 | tool_parser: str | None = None, |
| 114 | tool_strict_level: str = "auto", |
| 115 | tool_server: ToolServer | None = None, |
| 116 | enable_prompt_tokens_details: bool = False, |
| 117 | enable_force_include_usage: bool = False, |
| 118 | enable_per_request_metrics: bool = False, |
| 119 | enable_log_outputs: bool = False, |
| 120 | default_chat_template_kwargs: dict[str, Any] | None = None, |
| 121 | ) -> None: |
| 122 | super().__init__( |
| 123 | engine_client=engine_client, |
| 124 | models=models, |
| 125 | request_logger=request_logger, |
| 126 | return_tokens_as_token_ids=return_tokens_as_token_ids, |
| 127 | ) |
| 128 | |
| 129 | self.online_renderer = online_renderer |
| 130 | self.chat_template = chat_template |
| 131 | self.chat_template_content_format: Final = chat_template_content_format |
| 132 | self.chat_template_kwargs = default_chat_template_kwargs or {} |
| 133 | self.enable_log_outputs = enable_log_outputs |
| 134 | |
| 135 | # Set up the unified parser - either a unified parser or fall back to |
| 136 | # separate parsers accessed through the parser interface |
| 137 | self.parser = ParserManager.get_parser( |
| 138 | tool_parser_name=tool_parser, |
| 139 | reasoning_parser_name=reasoning_parser, |
| 140 | enable_auto_tools=enable_auto_tools, |
| 141 | tool_strict_level=tool_strict_level, |
| 142 | model_name=self.model_config.model, |
| 143 | is_harmony=self.model_config.hf_config.model_type == "gpt_oss", |
| 144 | tokenizer=self.renderer.tokenizer, |
| 145 | ) |
| 146 | self.enable_prompt_tokens_details = enable_prompt_tokens_details |
| 147 | self.enable_force_include_usage = enable_force_include_usage |
| 148 | self.enable_per_request_metrics = enable_per_request_metrics |
| 149 | |
| 150 | self.default_sampling_params = self.model_config.get_diff_sampling_param() |
| 151 | mc = self.model_config |
| 152 | self.override_max_tokens = ( |
| 153 | self.default_sampling_params.get("max_tokens") |
| 154 | if mc.generation_config not in ("auto", "vllm") |
| 155 | else getattr(mc, "override_generation_config", {}).get("max_new_tokens") |
| 156 | ) |
| 157 | |
| 158 | # If False (default), the "store" option is (silently) ignored and the |
| 159 | # response is not stored. If True, the response is stored in memory. |
| 160 | # NOTE(woosuk): This may not be intuitive for users, as the default |
| 161 | # behavior in OpenAI's Responses API is to store the response, but |
| 162 | # vLLM's default behavior is not. |
| 163 | self.enable_store = envs.VLLM_ENABLE_RESPONSES_API_STORE |
| 164 | if self.enable_store: |
| 165 | logger.warning_once( |
| 166 | "`VLLM_ENABLE_RESPONSES_API_STORE` is enabled. This may " |
| 167 | "cause a memory leak since we never remove responses from " |
| 168 | "the store." |
| 169 | ) |
| 170 | |
| 171 | self.use_harmony = self.model_config.hf_config.model_type == "gpt_oss" |
| 172 | if self.use_harmony: |
| 173 | logger.warning( |
| 174 | "For gpt-oss, we ignore --enable-auto-tool-choice " |
| 175 | "and always enable tool use." |
| 176 | ) |
| 177 | self.enable_auto_tools = enable_auto_tools |
| 178 | # HACK(woosuk): This is a hack. We should use a better store. |
| 179 | # FIXME: If enable_store=True, this may cause a memory leak since we |
| 180 | # never remove responses from the store. |
| 181 | self.response_store: dict[str, ResponsesResponse] = {} |
| 182 | self.response_store_lock = asyncio.Lock() |
| 183 | |
| 184 | # HACK(woosuk): This is a hack. We should use a better store. |
| 185 | # FIXME: If enable_store=True, this may cause a memory leak since we |
| 186 | # never remove messages from the store. |
| 187 | self.msg_store: dict[str, ResponsesPreviousMessages] = {} |
| 188 | |
| 189 | # HACK(wuhang): This is a hack. We should use a better store. |
| 190 | # FIXME: If enable_store=True, this may cause a memory leak since we |
| 191 | # never remove events from the store. |
| 192 | self.event_store: dict[ |
| 193 | str, tuple[deque[StreamingResponsesResponse], asyncio.Event] |
| 194 | ] = {} |
| 195 | |
| 196 | self.background_tasks: dict[str, asyncio.Task] = {} |
| 197 | |
| 198 | self.tool_server = tool_server |
| 199 | |
| 200 | def _effective_chat_template_kwargs( |
| 201 | self, request: ResponsesRequest |
| 202 | ) -> dict[str, Any]: |
| 203 | return ( |
| 204 | request.build_chat_params( |
| 205 | self.chat_template, |
| 206 | self.chat_template_content_format, |
| 207 | ) |
| 208 | .with_defaults(self.chat_template_kwargs) |
| 209 | .chat_template_kwargs |
| 210 | ) |
| 211 | |
| 212 | def _make_response_parser( |
| 213 | self, |
| 214 | request: ResponsesRequest, |
| 215 | tokenizer: TokenizerLike, |
| 216 | chat_template_kwargs: dict[str, Any], |
| 217 | ) -> Parser | None: |
| 218 | if self.parser is None: |
| 219 | return None |
| 220 | return self.parser( |
| 221 | tokenizer, |
| 222 | request.tools, |
| 223 | chat_template_kwargs=chat_template_kwargs, |
| 224 | model_config=self.model_config, |
| 225 | ) |
| 226 | |
| 227 | def _validate_generator_input( |
| 228 | self, |
| 229 | engine_input: EngineInput, |
| 230 | ) -> ErrorResponse | None: |
| 231 | """Add validations to the input to the generator here.""" |
| 232 | prompt_len = self._extract_prompt_len(engine_input) |
| 233 | max_model_len = self.model_config.max_model_len |
| 234 | |
| 235 | if prompt_len >= max_model_len: |
| 236 | error_message = ( |
| 237 | f"The engine prompt length {prompt_len} " |
| 238 | f"exceeds the max_model_len {max_model_len}. " |
| 239 | "Please reduce prompt." |
| 240 | ) |
| 241 | return self.create_error_response( |
| 242 | err_type="invalid_request_error", |
| 243 | message=error_message, |
| 244 | status_code=HTTPStatus.BAD_REQUEST, |
| 245 | param="input", |
| 246 | ) |
| 247 | |
| 248 | return None |
| 249 | |
| 250 | def _validate_create_responses_input( |
| 251 | self, request: ResponsesRequest |
| 252 | ) -> ErrorResponse | None: |
| 253 | if self.use_harmony and request.is_include_output_logprobs(): |
| 254 | return self.create_error_response( |
| 255 | err_type="invalid_request_error", |
| 256 | message="logprobs are not supported with gpt-oss models", |
| 257 | status_code=HTTPStatus.BAD_REQUEST, |
| 258 | param="logprobs", |
| 259 | ) |
| 260 | if request.store and not self.enable_store and request.background: |
| 261 | return self.create_error_response( |
| 262 | err_type="invalid_request_error", |
| 263 | message=( |
| 264 | "This vLLM engine does not support `store=True` and " |
| 265 | "therefore does not support the background mode. To " |
| 266 | "enable these features, set the environment variable " |
| 267 | "`VLLM_ENABLE_RESPONSES_API_STORE=1` when launching " |
| 268 | "the vLLM server." |
| 269 | ), |
| 270 | status_code=HTTPStatus.BAD_REQUEST, |
| 271 | param="background", |
| 272 | ) |
| 273 | if request.previous_input_messages and request.previous_response_id: |
| 274 | return self.create_error_response( |
| 275 | err_type="invalid_request_error", |
| 276 | message="Only one of `previous_input_messages` and " |
| 277 | "`previous_response_id` can be set.", |
| 278 | status_code=HTTPStatus.BAD_REQUEST, |
| 279 | param="previous_response_id", |
| 280 | ) |
| 281 | return None |
| 282 | |
| 283 | async def _render_resolved_response_inputs( |
| 284 | self, |
| 285 | request: ResponsesRequest, |
| 286 | prev_response: ResponsesResponse | None, |
| 287 | *, |
| 288 | skip_mm_cache: bool = False, |
| 289 | ) -> ResponsesRenderResult | ErrorResponse: |
| 290 | previous_messages: ResponsesPreviousMessages | None = None |
| 291 | previous_response_outputs: list[ResponseOutputItem] | None = None |
| 292 | if prev_response is not None: |
| 293 | stored_messages = ( |
| 294 | self.msg_store[prev_response.id] |
| 295 | if self.use_harmony |
| 296 | else self.msg_store.get(prev_response.id) |
| 297 | ) |
| 298 | if stored_messages is not None: |
| 299 | previous_messages = stored_messages.copy() |
| 300 | previous_response_outputs = list(prev_response.output) |
| 301 | |
| 302 | return await self.online_renderer.render_responses( |
| 303 | request, |
| 304 | previous_messages=previous_messages, |
| 305 | previous_response_outputs=previous_response_outputs, |
| 306 | tool_server=self.tool_server, |
| 307 | skip_mm_cache=skip_mm_cache, |
| 308 | ) |
| 309 | |
| 310 | async def create_responses( |
| 311 | self, |
| 312 | request: ResponsesRequest, |
| 313 | raw_request: Request | None = None, |
| 314 | ) -> ( |
| 315 | AsyncGenerator[StreamingResponsesResponse, None] |
| 316 | | ResponsesResponse |
| 317 | | ErrorResponse |
| 318 | ): |
| 319 | return await self._with_kv_transfer_rejection_cleanup( |
| 320 | self._create_responses(request, raw_request), request, raw_request |
| 321 | ) |
| 322 | |
| 323 | async def _create_responses( |
| 324 | self, request: ResponsesRequest, raw_request: Request | None = None |
| 325 | ) -> ( |
| 326 | AsyncGenerator[StreamingResponsesResponse, None] |
| 327 | | ResponsesResponse |
| 328 | | ErrorResponse |
| 329 | ): |
| 330 | error_check_ret = await self._check_model(request) |
| 331 | if error_check_ret is not None: |
| 332 | logger.error("Error with model %s", error_check_ret) |
| 333 | return error_check_ret |
| 334 | maybe_validation_error = self._validate_create_responses_input(request) |
| 335 | if maybe_validation_error is not None: |
| 336 | return maybe_validation_error |
| 337 | |
| 338 | maybe_template_error = self.online_renderer.validate_chat_template( |
| 339 | request_chat_template=None, |
| 340 | chat_template_kwargs=request.chat_template_kwargs, |
| 341 | trust_request_chat_template=( |
| 342 | self.online_renderer.trust_request_chat_template |
| 343 | ), |
| 344 | ) |
| 345 | if maybe_template_error is not None: |
| 346 | return maybe_template_error |
| 347 | |
| 348 | self._preflight() |
| 349 | |
| 350 | if request.store and not self.enable_store: |
| 351 | # Disable the store option. |
| 352 | # NOTE(woosuk): Although returning an error is possible, we opted |
| 353 | # to implicitly disable store and process the request anyway, as |
| 354 | # we assume most users do not intend to actually store the response |
| 355 | # (i.e., their request's `store=True` just because it's the default |
| 356 | # value). |
| 357 | request.store = False |
| 358 | |
| 359 | # Handle the previous response ID. |
| 360 | prev_response_id = request.previous_response_id |
| 361 | if prev_response_id is not None: |
| 362 | async with self.response_store_lock: |
| 363 | prev_response = self.response_store.get(prev_response_id) |
| 364 | if prev_response is None: |
| 365 | return self._make_not_found_error(prev_response_id) |
| 366 | else: |
| 367 | prev_response = None |
| 368 | |
| 369 | lora_request = self._maybe_get_adapters(request) |
| 370 | model_name = self.models.model_name(lora_request) |
| 371 | |
| 372 | render_result = await self._render_resolved_response_inputs( |
| 373 | request, |
| 374 | prev_response, |
| 375 | ) |
| 376 | if isinstance(render_result, ErrorResponse): |
| 377 | return render_result |
| 378 | messages = render_result.messages |
| 379 | engine_inputs = [render_result.engine_input] |
| 380 | |
| 381 | request_metadata = RequestResponseMetadata(request_id=request.request_id) |
| 382 | if raw_request: |
| 383 | raw_request.state.request_metadata = request_metadata |
| 384 | |
| 385 | # Schedule the request and get the result generator. |
| 386 | max_model_len = self.model_config.max_model_len |
| 387 | generators: list[AsyncGenerator[ConversationContext, None]] = [] |
| 388 | |
| 389 | # Only include builtin tools that the request actually asked for. |
| 390 | # Without this filter, tools registered on the server (e.g. via |
| 391 | # --tool-server demo) would be available for execution even when |
| 392 | # the request didn't enable them. |
| 393 | requested_tool_types = extract_tool_types(request.tools) |
| 394 | builtin_tool_list: list[str] = [] |
| 395 | if self.tool_server is not None: |
| 396 | if ( |
| 397 | self.tool_server.has_tool("browser") |
| 398 | and "web_search_preview" in requested_tool_types |
| 399 | ): |
| 400 | builtin_tool_list.append("browser") |
| 401 | if ( |
| 402 | self.tool_server.has_tool("python") |
| 403 | and "code_interpreter" in requested_tool_types |
| 404 | ): |
| 405 | builtin_tool_list.append("python") |
| 406 | if ( |
| 407 | self.tool_server.has_tool("container") |
| 408 | and "container" in requested_tool_types |
| 409 | ): |
| 410 | builtin_tool_list.append("container") |
| 411 | |
| 412 | if self.tool_server is not None: |
| 413 | available_tools = builtin_tool_list |
| 414 | else: |
| 415 | assert len(builtin_tool_list) == 0 |
| 416 | available_tools = [] |
| 417 | tokenizer = self.renderer.get_tokenizer() |
| 418 | |
| 419 | for engine_input in engine_inputs: |
| 420 | maybe_error = self._validate_generator_input(engine_input) |
| 421 | if maybe_error is not None: |
| 422 | return maybe_error |
| 423 | |
| 424 | default_max_tokens = get_max_tokens( |
| 425 | max_model_len, |
| 426 | request.max_output_tokens, |
| 427 | self._extract_prompt_len(engine_input), |
| 428 | self.default_sampling_params, |
| 429 | self.override_max_tokens, |
| 430 | truncate_prompt_tokens=( |
| 431 | -1 if request.truncation != "disabled" else None |
| 432 | ), |
| 433 | ) |
| 434 | |
| 435 | sampling_params = request.to_sampling_params( |
| 436 | default_max_tokens, self.default_sampling_params |
| 437 | ) |
| 438 | |
| 439 | trace_headers = ( |
| 440 | None |
| 441 | if raw_request is None |
| 442 | else await self._get_trace_headers(raw_request.headers) |
| 443 | ) |
| 444 | session_id = self._get_session_id(request, raw_request) |
| 445 | |
| 446 | chat_template_kwargs = self._effective_chat_template_kwargs(request) |
| 447 | response_parser = self._make_response_parser( |
| 448 | request, tokenizer, chat_template_kwargs |
| 449 | ) |
| 450 | if response_parser is not None: |
| 451 | prompt_token_ids = self._extract_prompt_components( |
| 452 | engine_input |
| 453 | ).token_ids |
| 454 | if prompt_token_ids is not None: |
| 455 | response_parser.set_prompt_token_ids(prompt_token_ids) |
| 456 | |
| 457 | context: ConversationContext |
| 458 | function_tool_names = extract_function_tool_names(request.tools) |
| 459 | if self.use_harmony: |
| 460 | context = HarmonyContext( |
| 461 | messages, |
| 462 | available_tools, |
| 463 | function_tool_names, |
| 464 | response_parser=response_parser, |
| 465 | request=request, |
| 466 | ) |
| 467 | else: |
| 468 | if envs.VLLM_USE_EXPERIMENTAL_PARSER_CONTEXT: |
| 469 | # This is a feature in development for parsing |
| 470 | # tokens during generation instead of at the end |
| 471 | context = ParsableContext( |
| 472 | response_messages=messages, |
| 473 | tokenizer=tokenizer, |
| 474 | parser_cls=self.parser, |
| 475 | request=request, |
| 476 | response_parser=response_parser, |
| 477 | available_tools=available_tools, |
| 478 | chat_template=self.chat_template, |
| 479 | chat_template_content_format=self.chat_template_content_format, |
| 480 | enable_auto_tools=self.enable_auto_tools, |
| 481 | ) |
| 482 | else: |
| 483 | context = SimpleContext( |
| 484 | response_parser=response_parser, |
| 485 | ) |
| 486 | |
| 487 | reasoning_parser_kwargs = None |
| 488 | if ( |
| 489 | context.response_parser is not None |
| 490 | and context.response_parser.reasoning_parser is not None |
| 491 | ): |
| 492 | reasoning_parser_kwargs = { |
| 493 | "chat_template_kwargs": chat_template_kwargs, |
| 494 | } |
| 495 | if ( |
| 496 | isinstance( |
| 497 | struct_out := sampling_params.structured_outputs, |
| 498 | StructuredOutputsParams, |
| 499 | ) |
| 500 | and struct_out.all_non_structural_tag_constraints_none() |
| 501 | ): |
| 502 | sampling_params.structured_outputs = replace( |
| 503 | struct_out, |
| 504 | structural_tag=( |
| 505 | context.response_parser.reasoning_parser.prepare_structured_tag( |
| 506 | struct_out.structural_tag, self.tool_server |
| 507 | ) |
| 508 | ), |
| 509 | ) |
| 510 | generator = self._generate_with_builtin_tools( |
| 511 | request_id=request.request_id, |
| 512 | engine_input=engine_input, |
| 513 | sampling_params=sampling_params, |
| 514 | context=context, |
| 515 | tok_params=( |
| 516 | request.build_tok_params(self.model_config) |
| 517 | if isinstance(context, HarmonyContext) |
| 518 | else None |
| 519 | ), |
| 520 | lora_request=lora_request, |
| 521 | priority=self._get_priority(request, raw_request), |
| 522 | trace_headers=trace_headers, |
| 523 | session_id=session_id, |
| 524 | reasoning_parser_kwargs=reasoning_parser_kwargs, |
| 525 | ) |
| 526 | generators.append(generator) |
| 527 | |
| 528 | assert len(generators) == 1 |
| 529 | (result_generator,) = generators |
| 530 | |
| 531 | # Store the input messages. |
| 532 | if request.store: |
| 533 | self.msg_store[request.request_id] = messages |
| 534 | |
| 535 | if request.background: |
| 536 | created_time = int(time.time()) |
| 537 | response = ResponsesResponse.from_request( |
| 538 | request, |
| 539 | sampling_params, |
| 540 | model_name=model_name, |
| 541 | created_time=created_time, |
| 542 | output=[], |
| 543 | status="queued", |
| 544 | usage=None, |
| 545 | ) |
| 546 | async with self.response_store_lock: |
| 547 | self.response_store[response.id] = response |
| 548 | |
| 549 | # Run the request in the background. |
| 550 | if request.stream: |
| 551 | task = asyncio.create_task( |
| 552 | self._run_background_request_stream( |
| 553 | request, |
| 554 | sampling_params, |
| 555 | result_generator, |
| 556 | context, |
| 557 | model_name, |
| 558 | tokenizer, |
| 559 | request_metadata, |
| 560 | created_time, |
| 561 | ), |
| 562 | name=f"create_{request.request_id}", |
| 563 | ) |
| 564 | else: |
| 565 | task = asyncio.create_task( |
| 566 | self._run_background_request( |
| 567 | request, |
| 568 | sampling_params, |
| 569 | result_generator, |
| 570 | context, |
| 571 | model_name, |
| 572 | tokenizer, |
| 573 | request_metadata, |
| 574 | created_time, |
| 575 | ), |
| 576 | name=f"create_{response.id}", |
| 577 | ) |
| 578 | |
| 579 | # For cleanup. |
| 580 | response_id = response.id |
| 581 | self.background_tasks[response_id] = task |
| 582 | task.add_done_callback( |
| 583 | lambda _: self.background_tasks.pop(response_id, None) |
| 584 | ) |
| 585 | |
| 586 | if request.stream: |
| 587 | return self.responses_background_stream_generator(request.request_id) |
| 588 | return response |
| 589 | |
| 590 | if request.stream: |
| 591 | return self.responses_stream_generator( |
| 592 | request, |
| 593 | sampling_params, |
| 594 | result_generator, |
| 595 | context, |
| 596 | model_name, |
| 597 | tokenizer, |
| 598 | request_metadata, |
| 599 | ) |
| 600 | |
| 601 | return await self.responses_full_generator( |
| 602 | request, |
| 603 | sampling_params, |
| 604 | result_generator, |
| 605 | context, |
| 606 | model_name, |
| 607 | tokenizer, |
| 608 | request_metadata, |
| 609 | ) |
| 610 | |
| 611 | async def _render_next_turn( |
| 612 | self, |
| 613 | request: ResponsesRequest, |
| 614 | messages: list[ResponseInputOutputItem], |
| 615 | ) -> list[EngineInput]: |
| 616 | render_request = request.model_copy( |
| 617 | update={ |
| 618 | "input": list(messages), |
| 619 | "instructions": None, |
| 620 | "previous_input_messages": None, |
| 621 | "previous_response_id": None, |
| 622 | } |
| 623 | ) |
| 624 | render_result = await self.online_renderer.render_responses( |
| 625 | render_request, |
| 626 | previous_messages=None, |
| 627 | previous_response_outputs=None, |
| 628 | tool_server=self.tool_server, |
| 629 | ) |
| 630 | if isinstance(render_result, ErrorResponse): |
| 631 | raise VLLMValidationError(render_result.error.message) |
| 632 | return [render_result.engine_input] |
| 633 | |
| 634 | async def _generate_with_builtin_tools( |
| 635 | self, |
| 636 | request_id: str, |
| 637 | engine_input: EngineInput, |
| 638 | sampling_params: SamplingParams, |
| 639 | context: ConversationContext, |
| 640 | tok_params: TokenizeParams | None = None, |
| 641 | lora_request: LoRARequest | None = None, |
| 642 | priority: int = 0, |
| 643 | trace_headers: Mapping[str, str] | None = None, |
| 644 | session_id: str | None = None, |
| 645 | reasoning_parser_kwargs: dict[str, Any] | None = None, |
| 646 | ): |
| 647 | max_model_len = self.model_config.max_model_len |
| 648 | cache_salt = cast(str | None, engine_input.get("cache_salt")) |
| 649 | |
| 650 | orig_priority = priority |
| 651 | sub_request = 0 |
| 652 | while True: |
| 653 | # Ensure that each sub-request has a unique request id. |
| 654 | sub_request_id = f"{request_id}_{sub_request}" |
| 655 | |
| 656 | self._log_inputs( |
| 657 | sub_request_id, |
| 658 | engine_input, |
| 659 | params=sampling_params, |
| 660 | lora_request=lora_request, |
| 661 | ) |
| 662 | |
| 663 | generator = self.engine_client.generate( |
| 664 | engine_input, |
| 665 | sampling_params, |
| 666 | sub_request_id, |
| 667 | lora_request=lora_request, |
| 668 | trace_headers=trace_headers, |
| 669 | priority=priority, |
| 670 | session_id=session_id, |
| 671 | reasoning_parser_kwargs=reasoning_parser_kwargs, |
| 672 | ) |
| 673 | |
| 674 | async for res in generator: |
| 675 | context.request_metrics = res.metrics |
| 676 | context.append_output(res) |
| 677 | # NOTE(woosuk): The stop condition is handled by the engine. |
| 678 | yield context |
| 679 | |
| 680 | if not context.need_builtin_tool_call(): |
| 681 | # The model did not ask for a tool call, so we're done. |
| 682 | break |
| 683 | |
| 684 | # Call the tool and update the context with the result. |
| 685 | context.request_metrics_cover_all_generation_turns = False |
| 686 | tool_output = await context.call_tool() |
| 687 | context.append_tool_output(tool_output) |
| 688 | |
| 689 | # TODO: uncomment this and enable tool output streaming |
| 690 | # yield context |
| 691 | |
| 692 | # Create inputs for the next turn. |
| 693 | # Render the next prompt token ids and update sampling_params. |
| 694 | if isinstance(context, HarmonyContext): |
| 695 | engine_input = self.online_renderer.render_responses_harmony_messages( |
| 696 | context.messages, |
| 697 | cache_salt=cache_salt, |
| 698 | tok_params=tok_params, |
| 699 | ) |
| 700 | |
| 701 | sampling_params.max_tokens = get_max_tokens( |
| 702 | max_model_len, |
| 703 | context.request.max_output_tokens if context.request else None, |
| 704 | self._extract_prompt_len(engine_input), |
| 705 | self.default_sampling_params, |
| 706 | self.override_max_tokens, |
| 707 | ) |
| 708 | elif isinstance(context, ParsableContext): |
| 709 | (engine_input,) = await self._render_next_turn( |
| 710 | context.request, |
| 711 | context.response_messages, |
| 712 | ) |
| 713 | if context.response_parser is not None: |
| 714 | prompt_token_ids = self._extract_prompt_components( |
| 715 | engine_input |
| 716 | ).token_ids |
| 717 | if prompt_token_ids is not None: |
| 718 | context.response_parser.set_prompt_token_ids(prompt_token_ids) |
| 719 | |
| 720 | sampling_params.max_tokens = get_max_tokens( |
| 721 | max_model_len, |
| 722 | context.request.max_output_tokens, |
| 723 | self._extract_prompt_len(engine_input), |
| 724 | self.default_sampling_params, # type: ignore |
| 725 | self.override_max_tokens, # type: ignore |
| 726 | truncate_prompt_tokens=( |
| 727 | -1 if context.request.truncation != "disabled" else None |
| 728 | ), |
| 729 | ) |
| 730 | |
| 731 | # OPTIMIZATION |
| 732 | priority = orig_priority - 1 |
| 733 | sub_request += 1 |
| 734 | |
| 735 | async def _initialize_tool_sessions( |
| 736 | self, |
| 737 | request: ResponsesRequest, |
| 738 | context: ConversationContext, |
| 739 | exit_stack: AsyncExitStack, |
| 740 | ): |
| 741 | # we should only initialize the tool session if the request needs tools |
| 742 | if len(request.tools) == 0: |
| 743 | return |
| 744 | mcp_tools = { |
| 745 | tool.server_label: tool for tool in request.tools if tool.type == "mcp" |
| 746 | } |
| 747 | await context.init_tool_sessions( |
| 748 | self.tool_server, exit_stack, request.request_id, mcp_tools |
| 749 | ) |
| 750 | |
| 751 | async def responses_full_generator( |
| 752 | self, |
| 753 | request: ResponsesRequest, |
| 754 | sampling_params: SamplingParams, |
| 755 | result_generator: AsyncIterator[ConversationContext], |
| 756 | context: ConversationContext, |
| 757 | model_name: str, |
| 758 | tokenizer: TokenizerLike, |
| 759 | request_metadata: RequestResponseMetadata, |
| 760 | created_time: int | None = None, |
| 761 | ) -> ErrorResponse | ResponsesResponse: |
| 762 | if created_time is None: |
| 763 | created_time = int(time.time()) |
| 764 | |
| 765 | async with AsyncExitStack() as exit_stack: |
| 766 | try: |
| 767 | await self._initialize_tool_sessions(request, context, exit_stack) |
| 768 | async for _ in result_generator: |
| 769 | pass |
| 770 | except asyncio.CancelledError: |
| 771 | return self.create_error_response("Client disconnected") |
| 772 | |
| 773 | # NOTE: Implementation of status is still WIP, but for now |
| 774 | # we guarantee that if the status is not "completed", it is accurate. |
| 775 | # "completed" is implemented as the "catch-all" for now. |
| 776 | status: ResponseStatus = "completed" |
| 777 | |
| 778 | input_messages: ResponseInputOutputMessage | None = None |
| 779 | output_messages: ResponseInputOutputMessage | None = None |
| 780 | if self.use_harmony: |
| 781 | assert isinstance(context, HarmonyContext) |
| 782 | output = [] |
| 783 | harmony_msgs = context.messages[context.num_init_messages :] |
| 784 | if harmony_msgs: |
| 785 | fn_names = context.function_tool_names |
| 786 | for msg in harmony_msgs[:-1]: |
| 787 | output.extend(harmony_to_response_output(msg, fn_names)) |
| 788 | output.extend( |
| 789 | harmony_to_response_output( |
| 790 | harmony_msgs[-1], |
| 791 | fn_names, |
| 792 | incomplete=context.last_append_flush_status, |
| 793 | ) |
| 794 | ) |
| 795 | |
| 796 | if request.enable_response_messages: |
| 797 | input_messages = context.messages[: context.num_init_messages] |
| 798 | output_messages = context.messages[context.num_init_messages :] |
| 799 | num_tool_output_tokens = context.num_tool_output_tokens |
| 800 | if len(output) > 0: |
| 801 | if context.finish_reason == "length": |
| 802 | status = "incomplete" |
| 803 | elif context.finish_reason == "abort": |
| 804 | status = "cancelled" |
| 805 | else: |
| 806 | self._raise_if_error(context.finish_reason, request.request_id) |
| 807 | else: |
| 808 | status = "incomplete" |
| 809 | elif isinstance(context, ParsableContext): |
| 810 | output = context.make_response_output_items() |
| 811 | |
| 812 | if request.enable_response_messages: |
| 813 | input_messages = context.input_messages |
| 814 | output_messages = context.output_messages |
| 815 | |
| 816 | # TODO: Calculate usage. |
| 817 | # assert final_res.prompt_token_ids is not None |
| 818 | num_tool_output_tokens = 0 |
| 819 | |
| 820 | # Check finish reason from the parser |
| 821 | if context.finish_reason == "length": |
| 822 | status = "incomplete" |
| 823 | else: |
| 824 | assert isinstance(context, SimpleContext) |
| 825 | # Use final_output which has accumulated text/token_ids/logprobs |
| 826 | final_res = context.final_output |
| 827 | assert final_res is not None |
| 828 | assert len(final_res.outputs) == 1 |
| 829 | final_output = final_res.outputs[0] |
| 830 | |
| 831 | # finish_reason='error' indicates retryable internal error |
| 832 | self._raise_if_error(final_output.finish_reason, request.request_id) |
| 833 | |
| 834 | # Check if generation was stopped due to max_tokens |
| 835 | if final_output.finish_reason == "length": |
| 836 | status = "incomplete" |
| 837 | |
| 838 | streamed_items = context.streamed_output_items |
| 839 | self._log_final_output( |
| 840 | request, final_output, is_streaming=streamed_items is not None |
| 841 | ) |
| 842 | if streamed_items is not None: |
| 843 | output = streamed_items |
| 844 | else: |
| 845 | output = self._make_response_output_items( |
| 846 | request, |
| 847 | final_output, |
| 848 | tokenizer, |
| 849 | parser=context.response_parser, |
| 850 | ) |
| 851 | |
| 852 | if request.enable_response_messages: |
| 853 | input_messages = context.input_messages |
| 854 | output_messages = context.output_messages |
| 855 | |
| 856 | # Calculate usage. |
| 857 | assert final_res.prompt_token_ids is not None |
| 858 | num_tool_output_tokens = 0 |
| 859 | |
| 860 | assert isinstance(context, (SimpleContext, HarmonyContext, ParsableContext)) |
| 861 | num_prompt_tokens = context.num_prompt_tokens |
| 862 | num_generated_tokens = context.num_output_tokens |
| 863 | num_cached_tokens = context.num_cached_tokens |
| 864 | num_reasoning_tokens = context.num_reasoning_tokens |
| 865 | # HarmonyContext and ParsableContext count reasoning tokens as each |
| 866 | # round is appended. SimpleContext is single-round but accumulates |
| 867 | # streaming deltas, so count its full output once here. |
| 868 | if isinstance(context, SimpleContext) and context.response_parser is not None: |
| 869 | num_reasoning_tokens = context.response_parser.count_reasoning_tokens( |
| 870 | context._accumulated_token_ids |
| 871 | ) |
| 872 | |
| 873 | usage = ResponseUsage( |
| 874 | input_tokens=num_prompt_tokens, |
| 875 | output_tokens=num_generated_tokens, |
| 876 | total_tokens=num_prompt_tokens + num_generated_tokens, |
| 877 | input_tokens_details=InputTokensDetails( |
| 878 | cache_write_tokens=getattr(context, "num_cache_creation_tokens", 0), |
| 879 | cached_tokens=num_cached_tokens, |
| 880 | input_tokens_per_turn=[ |
| 881 | turn.input_tokens for turn in context.all_turn_metrics |
| 882 | ], |
| 883 | cached_tokens_per_turn=[ |
| 884 | turn.cached_input_tokens for turn in context.all_turn_metrics |
| 885 | ], |
| 886 | ), |
| 887 | output_tokens_details=OutputTokensDetails( |
| 888 | reasoning_tokens=num_reasoning_tokens, |
| 889 | tool_output_tokens=num_tool_output_tokens, |
| 890 | output_tokens_per_turn=[ |
| 891 | turn.output_tokens for turn in context.all_turn_metrics |
| 892 | ], |
| 893 | tool_output_tokens_per_turn=[ |
| 894 | turn.tool_output_tokens for turn in context.all_turn_metrics |
| 895 | ], |
| 896 | ), |
| 897 | ) |
| 898 | per_request_metrics = None |
| 899 | if ( |
| 900 | self.enable_per_request_metrics |
| 901 | and context.request_metrics_cover_all_generation_turns |
| 902 | ): |
| 903 | per_request_metrics = build_per_request_timing_metrics( |
| 904 | context.request_metrics, num_generated_tokens |
| 905 | ) |
| 906 | response = ResponsesResponse.from_request( |
| 907 | request, |
| 908 | sampling_params, |
| 909 | input_messages=input_messages, |
| 910 | output_messages=output_messages, |
| 911 | model_name=model_name, |
| 912 | created_time=created_time, |
| 913 | output=output, |
| 914 | status=status, |
| 915 | usage=usage, |
| 916 | metrics=per_request_metrics, |
| 917 | kv_transfer_params=context.kv_transfer_params, |
| 918 | ec_transfer_params=context.ec_transfer_params, |
| 919 | ) |
| 920 | |
| 921 | if request.store: |
| 922 | async with self.response_store_lock: |
| 923 | stored_response = self.response_store.get(response.id) |
| 924 | # If the response is already cancelled, don't update it. |
| 925 | if stored_response is None or stored_response.status != "cancelled": |
| 926 | self.response_store[response.id] = response |
| 927 | return response |
| 928 | |
| 929 | def _topk_logprobs( |
| 930 | self, |
| 931 | logprobs: dict[int, SampleLogprob], |
| 932 | top_logprobs: int, |
| 933 | tokenizer: TokenizerLike, |
| 934 | ) -> list[LogprobTopLogprob]: |
| 935 | """Returns the top-k logprobs from the logprobs dictionary.""" |
| 936 | out = [] |
| 937 | for i, (token_id, _logprob) in enumerate(logprobs.items()): |
| 938 | if i >= top_logprobs: |
| 939 | break |
| 940 | text = self._get_decoded_token( |
| 941 | logprob=_logprob, |
| 942 | token_id=token_id, |
| 943 | tokenizer=tokenizer, |
| 944 | return_as_token_id=self.return_tokens_as_token_ids, |
| 945 | ) |
| 946 | out.append( |
| 947 | LogprobTopLogprob( |
| 948 | token=text, |
| 949 | logprob=max(_logprob.logprob, -9999.0), |
| 950 | bytes=list(text.encode("utf-8", errors="replace")), |
| 951 | ) |
| 952 | ) |
| 953 | return out |
| 954 | |
| 955 | def _create_response_logprobs( |
| 956 | self, |
| 957 | token_ids: Sequence[int], |
| 958 | logprobs: SampleLogprobs | None, |
| 959 | tokenizer: TokenizerLike, |
| 960 | top_logprobs: int | None = None, |
| 961 | ) -> list[Logprob]: |
| 962 | assert logprobs is not None, "logprobs must be provided" |
| 963 | assert len(token_ids) == len(logprobs), ( |
| 964 | "token_ids and logprobs.token_ids must have the same length" |
| 965 | ) |
| 966 | out = [] |
| 967 | for i, token_id in enumerate(token_ids): |
| 968 | logprob = logprobs[i] |
| 969 | token_logprob = logprob[token_id] |
| 970 | text = self._get_decoded_token( |
| 971 | logprob=token_logprob, |
| 972 | token_id=token_id, |
| 973 | tokenizer=tokenizer, |
| 974 | return_as_token_id=self.return_tokens_as_token_ids, |
| 975 | ) |
| 976 | out.append( |
| 977 | Logprob( |
| 978 | token=text, |
| 979 | logprob=max(token_logprob.logprob, -9999.0), |
| 980 | bytes=list(text.encode("utf-8", errors="replace")), |
| 981 | top_logprobs=( |
| 982 | self._topk_logprobs( |
| 983 | logprob, top_logprobs=top_logprobs, tokenizer=tokenizer |
| 984 | ) |
| 985 | if top_logprobs |
| 986 | else [] |
| 987 | ), |
| 988 | ) |
| 989 | ) |
| 990 | return out |
| 991 | |
| 992 | def _log_final_output( |
| 993 | self, |
| 994 | request: ResponsesRequest, |
| 995 | final_output: CompletionOutput, |
| 996 | is_streaming: bool, |
| 997 | ) -> None: |
| 998 | if self.enable_log_outputs and self.request_logger: |
| 999 | self.request_logger.log_outputs( |
| 1000 | request_id=request.request_id, |
| 1001 | outputs=final_output.text, |
| 1002 | output_token_ids=final_output.token_ids, |
| 1003 | finish_reason=final_output.finish_reason, |
| 1004 | is_streaming=is_streaming, |
| 1005 | delta=False, |
| 1006 | ) |
| 1007 | |
| 1008 | def _make_response_output_items( |
| 1009 | self, |
| 1010 | request: ResponsesRequest, |
| 1011 | final_output: CompletionOutput, |
| 1012 | tokenizer: TokenizerLike, |
| 1013 | parser: Parser | None = None, |
| 1014 | ) -> list[ResponseOutputItem]: |
| 1015 | # Compute logprobs if requested |
| 1016 | logprobs = None |
| 1017 | if request.is_include_output_logprobs() and final_output.logprobs: |
| 1018 | logprobs = self._create_response_logprobs( |
| 1019 | token_ids=final_output.token_ids, |
| 1020 | logprobs=final_output.logprobs, |
| 1021 | tokenizer=tokenizer, |
| 1022 | top_logprobs=request.top_logprobs, |
| 1023 | ) |
| 1024 | |
| 1025 | # Use parser to extract reasoning, content, and tool calls |
| 1026 | if parser: |
| 1027 | reasoning, content, tool_calls = parser.parse( |
| 1028 | final_output.text, |
| 1029 | request, |
| 1030 | enable_auto_tools=self.enable_auto_tools, |
| 1031 | model_output_token_ids=final_output.token_ids, |
| 1032 | ) |
| 1033 | if not request.include_reasoning: |
| 1034 | reasoning = None |
| 1035 | logprobs = None |
| 1036 | return build_response_output_items( |
| 1037 | reasoning=reasoning, |
| 1038 | content=content, |
| 1039 | tool_calls=maybe_filter_parallel_tool_calls(tool_calls or [], request), |
| 1040 | logprobs=logprobs, |
| 1041 | tools=request.tools, |
| 1042 | ) |
| 1043 | |
| 1044 | # Fallback when no parser is configured |
| 1045 | return [ |
| 1046 | ResponseOutputMessage( |
| 1047 | id=f"msg_{random_uuid()}", |
| 1048 | content=[ |
| 1049 | ResponseOutputText( |
| 1050 | text=final_output.text, |
| 1051 | annotations=[], |
| 1052 | type="output_text", |
| 1053 | logprobs=logprobs, |
| 1054 | ) |
| 1055 | ] |
| 1056 | if final_output.text |
| 1057 | else [], |
| 1058 | role="assistant", |
| 1059 | status="completed", |
| 1060 | type="message", |
| 1061 | ) |
| 1062 | ] |
| 1063 | |
| 1064 | async def _run_background_request_stream( |
| 1065 | self, |
| 1066 | request: ResponsesRequest, |
| 1067 | *args, |
| 1068 | **kwargs, |
| 1069 | ): |
| 1070 | event_deque: deque[StreamingResponsesResponse] = deque() |
| 1071 | new_event_signal = asyncio.Event() |
| 1072 | self.event_store[request.request_id] = (event_deque, new_event_signal) |
| 1073 | generator = self.responses_stream_generator(request, *args, **kwargs) |
| 1074 | try: |
| 1075 | async for event in generator: |
| 1076 | event_deque.append(event) |
| 1077 | new_event_signal.set() # Signal new event available |
| 1078 | finally: |
| 1079 | new_event_signal.set() |
| 1080 | |
| 1081 | async def _run_background_request( |
| 1082 | self, |
| 1083 | request: ResponsesRequest, |
| 1084 | *args, |
| 1085 | **kwargs, |
| 1086 | ): |
| 1087 | response = await self.responses_full_generator(request, *args, **kwargs) |
| 1088 | |
| 1089 | if isinstance(response, ErrorResponse): |
| 1090 | # If the request has failed, update the status to "failed". |
| 1091 | response_id = request.request_id |
| 1092 | async with self.response_store_lock: |
| 1093 | stored_response = self.response_store.get(response_id) |
| 1094 | assert stored_response is not None |
| 1095 | if stored_response.status not in ("completed", "cancelled"): |
| 1096 | stored_response.status = "failed" |
| 1097 | |
| 1098 | async def responses_background_stream_generator( |
| 1099 | self, |
| 1100 | response_id: str, |
| 1101 | starting_after: int | None = None, |
| 1102 | ) -> AsyncGenerator[StreamingResponsesResponse, None]: |
| 1103 | if response_id not in self.event_store: |
| 1104 | raise VLLMValidationError( |
| 1105 | f"Unknown response_id: {response_id}", |
| 1106 | parameter="response_id", |
| 1107 | value=response_id, |
| 1108 | ) |
| 1109 | |
| 1110 | event_deque, new_event_signal = self.event_store[response_id] |
| 1111 | start_index = 0 if starting_after is None else starting_after + 1 |
| 1112 | current_index = start_index |
| 1113 | |
| 1114 | while True: |
| 1115 | new_event_signal.clear() |
| 1116 | |
| 1117 | # Yield existing events from start_index |
| 1118 | while current_index < len(event_deque): |
| 1119 | event = event_deque[current_index] |
| 1120 | yield event |
| 1121 | if getattr(event, "type", "unknown") == "response.completed": |
| 1122 | return |
| 1123 | current_index += 1 |
| 1124 | |
| 1125 | await new_event_signal.wait() |
| 1126 | |
| 1127 | async def retrieve_responses( |
| 1128 | self, |
| 1129 | response_id: str, |
| 1130 | starting_after: int | None, |
| 1131 | stream: bool | None, |
| 1132 | ) -> ( |
| 1133 | ErrorResponse |
| 1134 | | ResponsesResponse |
| 1135 | | AsyncGenerator[StreamingResponsesResponse, None] |
| 1136 | ): |
| 1137 | async with self.response_store_lock: |
| 1138 | response = self.response_store.get(response_id) |
| 1139 | |
| 1140 | if response is None: |
| 1141 | return self._make_not_found_error(response_id) |
| 1142 | |
| 1143 | if stream: |
| 1144 | return self.responses_background_stream_generator( |
| 1145 | response_id, |
| 1146 | starting_after, |
| 1147 | ) |
| 1148 | return response |
| 1149 | |
| 1150 | async def cancel_responses( |
| 1151 | self, |
| 1152 | response_id: str, |
| 1153 | ) -> ErrorResponse | ResponsesResponse: |
| 1154 | async with self.response_store_lock: |
| 1155 | response = self.response_store.get(response_id) |
| 1156 | if response is None: |
| 1157 | return self._make_not_found_error(response_id) |
| 1158 | |
| 1159 | prev_status = response.status |
| 1160 | if prev_status not in ("queued", "in_progress"): |
| 1161 | return self.create_error_response( |
| 1162 | err_type="invalid_request_error", |
| 1163 | message="Cannot cancel a synchronous response.", |
| 1164 | param="response_id", |
| 1165 | ) |
| 1166 | |
| 1167 | # Update the status to "cancelled". |
| 1168 | response.status = "cancelled" |
| 1169 | |
| 1170 | # Abort the request. |
| 1171 | if task := self.background_tasks.get(response_id): |
| 1172 | task.cancel() |
| 1173 | try: |
| 1174 | await task |
| 1175 | except asyncio.CancelledError: |
| 1176 | logger.exception("Background task for %s was cancelled", response_id) |
| 1177 | return response |
| 1178 | |
| 1179 | def _make_not_found_error(self, response_id: str) -> ErrorResponse: |
| 1180 | return self.create_error_response( |
| 1181 | err_type="invalid_request_error", |
| 1182 | message=f"Response with id '{response_id}' not found.", |
| 1183 | status_code=HTTPStatus.NOT_FOUND, |
| 1184 | param="response_id", |
| 1185 | ) |
| 1186 | |
| 1187 | async def _process_simple_streaming_events( |
| 1188 | self, |
| 1189 | request: ResponsesRequest, |
| 1190 | sampling_params: SamplingParams, |
| 1191 | result_generator: AsyncIterator[ConversationContext | None], |
| 1192 | context: ConversationContext, |
| 1193 | model_name: str, |
| 1194 | tokenizer: TokenizerLike, |
| 1195 | request_metadata: RequestResponseMetadata, |
| 1196 | created_time: int, |
| 1197 | _increment_sequence_number_and_return: Callable[ |
| 1198 | [StreamingResponsesResponse], StreamingResponsesResponse |
| 1199 | ], |
| 1200 | ) -> AsyncGenerator[StreamingResponsesResponse, None]: |
| 1201 | processor = SimpleStreamingEventProcessor(tools=request.tools) |
| 1202 | |
| 1203 | hide_stream_metadata = ( |
| 1204 | not request.include_reasoning and context.response_parser is not None |
| 1205 | ) |
| 1206 | |
| 1207 | def _get_logprobs(output: CompletionOutput) -> list[Logprob]: |
| 1208 | if not request.is_include_output_logprobs(): |
| 1209 | return [] |
| 1210 | if hide_stream_metadata: |
| 1211 | return [] |
| 1212 | return self._create_response_logprobs( |
| 1213 | token_ids=output.token_ids, |
| 1214 | logprobs=output.logprobs, |
| 1215 | tokenizer=tokenizer, |
| 1216 | top_logprobs=request.top_logprobs, |
| 1217 | ) |
| 1218 | |
| 1219 | async for ctx in result_generator: |
| 1220 | assert isinstance(ctx, SimpleContext) |
| 1221 | if ctx.last_output is None or not ctx.last_output.outputs: |
| 1222 | continue |
| 1223 | |
| 1224 | output = ctx.last_output.outputs[0] |
| 1225 | self._raise_if_error(output.finish_reason, request.request_id) |
| 1226 | delta_text = output.text |
| 1227 | delta_token_ids = as_list(output.token_ids) |
| 1228 | |
| 1229 | if ctx.response_parser: |
| 1230 | delta_message = ctx.response_parser.parse_delta( |
| 1231 | delta_text=delta_text, |
| 1232 | delta_token_ids=delta_token_ids, |
| 1233 | request=request, |
| 1234 | prompt_token_ids=ctx.last_output.prompt_token_ids, |
| 1235 | finished=output.finish_reason is not None, |
| 1236 | ) |
| 1237 | else: |
| 1238 | delta_message = DeltaMessage(content=output.text) |
| 1239 | |
| 1240 | if not delta_message: |
| 1241 | continue |
| 1242 | delta_message = maybe_filter_parallel_tool_calls(delta_message, request) |
| 1243 | |
| 1244 | for dm in split_delta(delta_message): |
| 1245 | target_state, tool_call = processor.resolve_target_state(dm) |
| 1246 | if target_state == _StateType.NONE: |
| 1247 | continue |
| 1248 | |
| 1249 | if processor.needs_transition(target_state, tool_call): |
| 1250 | for event in processor.close_current(): |
| 1251 | yield _increment_sequence_number_and_return(event) |
| 1252 | for event in processor.open(target_state, tool_call): |
| 1253 | yield _increment_sequence_number_and_return(event) |
| 1254 | |
| 1255 | for event in processor.emit_delta(dm, output, _get_logprobs): |
| 1256 | yield _increment_sequence_number_and_return(event) |
| 1257 | |
| 1258 | for event in processor.close_current(): |
| 1259 | yield _increment_sequence_number_and_return(event) |
| 1260 | |
| 1261 | assert isinstance(context, SimpleContext) |
| 1262 | context.streamed_output_items = processor.output_items |
| 1263 | |
| 1264 | async def _process_harmony_streaming_events( |
| 1265 | self, |
| 1266 | request: ResponsesRequest, |
| 1267 | sampling_params: SamplingParams, |
| 1268 | result_generator: AsyncIterator[ConversationContext | None], |
| 1269 | context: ConversationContext, |
| 1270 | model_name: str, |
| 1271 | tokenizer: TokenizerLike, |
| 1272 | request_metadata: RequestResponseMetadata, |
| 1273 | created_time: int, |
| 1274 | _increment_sequence_number_and_return: Callable[ |
| 1275 | [StreamingResponsesResponse], StreamingResponsesResponse |
| 1276 | ], |
| 1277 | ) -> AsyncGenerator[StreamingResponsesResponse, None]: |
| 1278 | state = StreamingState() |
| 1279 | |
| 1280 | async for ctx in result_generator: |
| 1281 | assert isinstance(ctx, HarmonyContext) |
| 1282 | |
| 1283 | # finish_reason='error' indicates a retryable error |
| 1284 | self._raise_if_error(ctx.finish_reason, request.request_id) |
| 1285 | |
| 1286 | for segment in ctx.last_append_segments: |
| 1287 | if segment.delta: |
| 1288 | for event in emit_content_delta_events( |
| 1289 | segment, state, ctx.function_tool_names |
| 1290 | ): |
| 1291 | yield _increment_sequence_number_and_return(event) |
| 1292 | |
| 1293 | elif completed_message := segment.completed_message: |
| 1294 | # TODO: Fix browser emitted as MCP calls |
| 1295 | for event in emit_previous_item_done_events( |
| 1296 | completed_message, state, ctx.function_tool_names |
| 1297 | ): |
| 1298 | yield _increment_sequence_number_and_return(event) |
| 1299 | |
| 1300 | for event in emit_tool_action_events( |
| 1301 | completed_message, state, self.tool_server |
| 1302 | ): |
| 1303 | yield _increment_sequence_number_and_return(event) |
| 1304 | state.reset_for_new_item() |
| 1305 | |
| 1306 | async def responses_stream_generator( |
| 1307 | self, |
| 1308 | request: ResponsesRequest, |
| 1309 | sampling_params: SamplingParams, |
| 1310 | result_generator: AsyncIterator[ConversationContext | None], |
| 1311 | context: ConversationContext, |
| 1312 | model_name: str, |
| 1313 | tokenizer: TokenizerLike, |
| 1314 | request_metadata: RequestResponseMetadata, |
| 1315 | created_time: int | None = None, |
| 1316 | ) -> AsyncGenerator[StreamingResponsesResponse, None]: |
| 1317 | # TODO: |
| 1318 | # 1. Handle disconnect |
| 1319 | |
| 1320 | created_time = created_time or int(time.time()) |
| 1321 | |
| 1322 | sequence_number = 0 |
| 1323 | |
| 1324 | def _increment_sequence_number_and_return( |
| 1325 | event: StreamingResponsesResponse, |
| 1326 | ) -> StreamingResponsesResponse: |
| 1327 | nonlocal sequence_number |
| 1328 | # Set sequence_number if the event has this attribute |
| 1329 | if hasattr(event, "sequence_number"): |
| 1330 | event.sequence_number = sequence_number |
| 1331 | sequence_number += 1 |
| 1332 | return event |
| 1333 | |
| 1334 | async with AsyncExitStack() as exit_stack: |
| 1335 | if self.use_harmony: |
| 1336 | # TODO: in streaming, we noticed this bug: |
| 1337 | # https://github.com/vllm-project/vllm/issues/25697 |
| 1338 | await self._initialize_tool_sessions(request, context, exit_stack) |
| 1339 | processor = self._process_harmony_streaming_events |
| 1340 | else: |
| 1341 | processor = self._process_simple_streaming_events |
| 1342 | # TODO Hanchen make sampling params to include the structural tag |
| 1343 | |
| 1344 | initial_response = ResponsesResponse.from_request( |
| 1345 | request, |
| 1346 | sampling_params, |
| 1347 | model_name=model_name, |
| 1348 | created_time=created_time, |
| 1349 | output=[], |
| 1350 | status="in_progress", |
| 1351 | usage=None, |
| 1352 | ).model_dump(mode="json", by_alias=True) |
| 1353 | yield _increment_sequence_number_and_return( |
| 1354 | ResponseCreatedEvent( |
| 1355 | type="response.created", |
| 1356 | sequence_number=-1, |
| 1357 | response=initial_response, |
| 1358 | ) |
| 1359 | ) |
| 1360 | yield _increment_sequence_number_and_return( |
| 1361 | ResponseInProgressEvent( |
| 1362 | type="response.in_progress", |
| 1363 | sequence_number=-1, |
| 1364 | response=initial_response, |
| 1365 | ) |
| 1366 | ) |
| 1367 | |
| 1368 | try: |
| 1369 | async for event_data in processor( |
| 1370 | request, |
| 1371 | sampling_params, |
| 1372 | result_generator, |
| 1373 | context, |
| 1374 | model_name, |
| 1375 | tokenizer, |
| 1376 | request_metadata, |
| 1377 | created_time, |
| 1378 | _increment_sequence_number_and_return, |
| 1379 | ): |
| 1380 | yield event_data |
| 1381 | except GenerationError as e: |
| 1382 | error_json = self._convert_generation_error_to_streaming_response(e) |
| 1383 | yield _increment_sequence_number_and_return( |
| 1384 | TypeAdapter(StreamingResponsesResponse).validate_json(error_json) |
| 1385 | ) |
| 1386 | return |
| 1387 | |
| 1388 | async def empty_async_generator(): |
| 1389 | # A hack to trick Python to think this is a generator but |
| 1390 | # in fact it immediately returns. |
| 1391 | if False: |
| 1392 | yield |
| 1393 | |
| 1394 | final_response = await self.responses_full_generator( |
| 1395 | request, |
| 1396 | sampling_params, |
| 1397 | empty_async_generator(), |
| 1398 | context, |
| 1399 | model_name, |
| 1400 | tokenizer, |
| 1401 | request_metadata, |
| 1402 | created_time=created_time, |
| 1403 | ) |
| 1404 | yield _increment_sequence_number_and_return( |
| 1405 | ResponseCompletedEvent( |
| 1406 | type="response.completed", |
| 1407 | sequence_number=-1, |
| 1408 | response=final_response, |
| 1409 | ) |
| 1410 | ) |