serving.py 84.4 KB
Newer Older
1
2
3
4
5
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project

import asyncio
import time
6
import uuid
7
from collections import deque
8
from collections.abc import AsyncGenerator, AsyncIterator, Callable, Mapping, Sequence
9
from contextlib import AsyncExitStack
10
from copy import copy
11
from http import HTTPStatus
12
from typing import Any, Final
13
14

from fastapi import Request
15
from openai.types.responses import (
16
17
    ResponseContentPartAddedEvent,
    ResponseContentPartDoneEvent,
18
19
    ResponseFunctionCallArgumentsDeltaEvent,
    ResponseFunctionCallArgumentsDoneEvent,
20
    ResponseFunctionToolCall,
21
    ResponseFunctionToolCallItem,
22
23
24
25
26
27
28
29
30
31
32
33
34
35
    ResponseOutputItem,
    ResponseOutputItemAddedEvent,
    ResponseOutputItemDoneEvent,
    ResponseOutputMessage,
    ResponseOutputText,
    ResponseReasoningItem,
    ResponseReasoningTextDeltaEvent,
    ResponseReasoningTextDoneEvent,
    ResponseStatus,
    ResponseTextDeltaEvent,
    ResponseTextDoneEvent,
    response_text_delta_event,
)
from openai.types.responses.response_output_text import Logprob, LogprobTopLogprob
36
from openai.types.responses.response_reasoning_item import (
37
38
    Content as ResponseReasoningTextContent,
)
39
from openai.types.responses.tool import Mcp, Tool
40
from openai_harmony import Message as OpenAIHarmonyMessage
41
from pydantic import TypeAdapter
42

43
from vllm import envs
44
from vllm.config.utils import replace
45
from vllm.engine.protocol import EngineClient
46
47
48
from vllm.entrypoints.chat_utils import (
    ChatCompletionMessageParam,
    ChatTemplateContentFormatOption,
49
    get_tool_call_id_type,
50
)
51
from vllm.entrypoints.logger import RequestLogger
52
from vllm.entrypoints.mcp.tool_server import ToolServer
53
from vllm.entrypoints.openai.engine.protocol import (
54
55
56
57
    DeltaMessage,
    ErrorResponse,
    RequestResponseMetadata,
)
58
from vllm.entrypoints.openai.engine.serving import (
59
60
61
    GenerationError,
    OpenAIServing,
)
62
from vllm.entrypoints.openai.models.serving import OpenAIServingModels
63
64
65
66
67
68
69
70
from vllm.entrypoints.openai.parser.harmony_utils import (
    get_developer_message,
    get_stop_tokens_for_assistant_actions,
    get_system_message,
    get_user_message,
    has_custom_tools,
    render_for_completion,
)
71
72
73
74
75
76
77
from vllm.entrypoints.openai.responses.context import (
    ConversationContext,
    HarmonyContext,
    ParsableContext,
    SimpleContext,
    StreamingHarmonyContext,
)
78
79
80
81
82
83
from vllm.entrypoints.openai.responses.harmony import (
    construct_harmony_previous_input_messages,
    harmony_to_response_output,
    parser_state_to_response_output,
    response_input_to_harmony,
)
84
85
86
87
88
89
from vllm.entrypoints.openai.responses.protocol import (
    InputTokensDetails,
    OutputTokensDetails,
    ResponseCompletedEvent,
    ResponseCreatedEvent,
    ResponseInProgressEvent,
90
    ResponseInputOutputItem,
91
    ResponseInputOutputMessage,
92
93
    ResponseReasoningPartAddedEvent,
    ResponseReasoningPartDoneEvent,
94
95
96
97
98
    ResponsesRequest,
    ResponsesResponse,
    ResponseUsage,
    StreamingResponsesResponse,
)
99
from vllm.entrypoints.openai.responses.streaming_events import (
100
    StreamingState,
101
102
103
104
    emit_content_delta_events,
    emit_previous_item_done_events,
    emit_tool_action_events,
)
105
from vllm.entrypoints.openai.responses.utils import (
106
    construct_input_messages,
107
    construct_tool_dicts,
108
109
    extract_tool_types,
)
110
from vllm.entrypoints.serve.render.serving import OpenAIServingRender
111
from vllm.entrypoints.utils import get_max_tokens
112
from vllm.exceptions import VLLMValidationError
113
from vllm.inputs import EngineInput, tokens_input
114
from vllm.logger import init_logger
115
116
from vllm.logprobs import Logprob as SampleLogprob
from vllm.logprobs import SampleLogprobs
117
from vllm.lora.request import LoRARequest
118
from vllm.outputs import CompletionOutput
119
from vllm.parser import ParserManager
120
from vllm.sampling_params import SamplingParams, StructuredOutputsParams
121
from vllm.tokenizers import TokenizerLike
122
from vllm.tool_parsers import ToolParser
123
from vllm.utils import random_uuid
124
from vllm.utils.collection_utils import as_list
125
126
127
128

logger = init_logger(__name__)


129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
def _extract_allowed_tools_from_mcp_requests(
    tools: list[Tool],
) -> dict[str, list[str] | None]:
    """
    Extract allowed_tools mapping from MCP tool requests.

    Returns a dictionary mapping server_label to allowed_tools list.
    Handles both list format and McpAllowedToolsMcpToolFilter object format.

    Special handling:
    - If allowed_tools is None, returns None (allows all tools)
    - If allowed_tools contains "*", returns None (allows all tools)
    - Otherwise, returns the list of specific tool names

    This function can be reused for both harmony and non-harmony MCP calls.
    """
    allowed_tools_map: dict[str, list[str] | None] = {}
    for tool in tools:
        if not isinstance(tool, Mcp):
            continue

        # allowed_tools can be a list or an object with tool_names
        # Extract the actual list of tool names
        allowed_tools_val = None
        if tool.allowed_tools is not None:
            if isinstance(tool.allowed_tools, list):
                allowed_tools_val = tool.allowed_tools
            elif hasattr(tool.allowed_tools, "tool_names"):
                # It's an McpAllowedToolsMcpToolFilter object
                allowed_tools_val = tool.allowed_tools.tool_names

        # Normalize "*" to None (both mean "allow all tools")
        if allowed_tools_val is not None and "*" in allowed_tools_val:
            allowed_tools_val = None

        allowed_tools_map[tool.server_label] = allowed_tools_val
    return allowed_tools_map


168
169
170
171
172
class OpenAIServingResponses(OpenAIServing):
    def __init__(
        self,
        engine_client: EngineClient,
        models: OpenAIServingModels,
173
        openai_serving_render: OpenAIServingRender,
174
        *,
175
176
        request_logger: RequestLogger | None,
        chat_template: str | None,
177
178
179
180
        chat_template_content_format: ChatTemplateContentFormatOption,
        return_tokens_as_token_ids: bool = False,
        reasoning_parser: str = "",
        enable_auto_tools: bool = False,
181
182
        tool_parser: str | None = None,
        tool_server: ToolServer | None = None,
183
184
        enable_prompt_tokens_details: bool = False,
        enable_force_include_usage: bool = False,
185
        enable_log_outputs: bool = False,
186
187
188
189
190
191
192
193
    ) -> None:
        super().__init__(
            engine_client=engine_client,
            models=models,
            request_logger=request_logger,
            return_tokens_as_token_ids=return_tokens_as_token_ids,
        )

194
        self.openai_serving_render = openai_serving_render
195
196
        self.chat_template = chat_template
        self.chat_template_content_format: Final = chat_template_content_format
197
        self.enable_log_outputs = enable_log_outputs
198

199
200
201
202
203
204
205
        # Set up the unified parser - either a unified parser or fall back to
        # separate parsers accessed through the parser interface
        self.parser = ParserManager.get_parser(
            tool_parser_name=tool_parser,
            reasoning_parser_name=reasoning_parser,
            enable_auto_tools=enable_auto_tools,
            model_name=self.model_config.model,
206
        )
207
208
        self.enable_prompt_tokens_details = enable_prompt_tokens_details
        self.enable_force_include_usage = enable_force_include_usage
209

210
        self.default_sampling_params = self.model_config.get_diff_sampling_param()
211
212
213
214
215
216
        mc = self.model_config
        self.override_max_tokens = (
            self.default_sampling_params.get("max_tokens")
            if mc.generation_config not in ("auto", "vllm")
            else getattr(mc, "override_generation_config", {}).get("max_new_tokens")
        )
217

218
219
220
221
222
        # If False (default), the "store" option is (silently) ignored and the
        # response is not stored. If True, the response is stored in memory.
        # NOTE(woosuk): This may not be intuitive for users, as the default
        # behavior in OpenAI's Responses API is to store the response, but
        # vLLM's default behavior is not.
223
        self.enable_store = envs.VLLM_ENABLE_RESPONSES_API_STORE
224
225
226
227
        if self.enable_store:
            logger.warning_once(
                "`VLLM_ENABLE_RESPONSES_API_STORE` is enabled. This may "
                "cause a memory leak since we never remove responses from "
228
229
                "the store."
            )
230

231
        self.use_harmony = self.model_config.hf_config.model_type == "gpt_oss"
232
        if self.use_harmony:
233
234
235
236
            logger.warning(
                "For gpt-oss, we ignore --enable-auto-tool-choice "
                "and always enable tool use."
            )
237
238
239
240
241
            # OpenAI models have two EOS-like tokens: <|return|> and <|call|>.
            # We need to add them to the stop token ids.
            if "stop_token_ids" not in self.default_sampling_params:
                self.default_sampling_params["stop_token_ids"] = []
            self.default_sampling_params["stop_token_ids"].extend(
242
243
                get_stop_tokens_for_assistant_actions()
            )
244

245
        self.tool_call_id_type = get_tool_call_id_type(self.model_config)
246

247
        self.enable_auto_tools = enable_auto_tools
248
        # HACK(woosuk): This is a hack. We should use a better store.
249
250
        # FIXME: If enable_store=True, this may cause a memory leak since we
        # never remove responses from the store.
251
252
253
254
        self.response_store: dict[str, ResponsesResponse] = {}
        self.response_store_lock = asyncio.Lock()

        # HACK(woosuk): This is a hack. We should use a better store.
255
256
        # FIXME: If enable_store=True, this may cause a memory leak since we
        # never remove messages from the store.
257
258
        self.msg_store: dict[str, list[ChatCompletionMessageParam]] = {}

259
260
261
        # HACK(wuhang): This is a hack. We should use a better store.
        # FIXME: If enable_store=True, this may cause a memory leak since we
        # never remove events from the store.
262
263
264
        self.event_store: dict[
            str, tuple[deque[StreamingResponsesResponse], asyncio.Event]
        ] = {}
265

266
267
        self.background_tasks: dict[str, asyncio.Task] = {}

268
269
        self.tool_server = tool_server

270
    def _validate_generator_input(
271
        self,
272
        engine_input: EngineInput,
273
    ) -> ErrorResponse | None:
274
        """Add validations to the input to the generator here."""
275
        prompt_len = self._extract_prompt_len(engine_input)
276
277
278
        max_model_len = self.model_config.max_model_len

        if prompt_len >= max_model_len:
279
            error_message = (
280
                f"The engine prompt length {prompt_len} "
281
                f"exceeds the max_model_len {max_model_len}. "
282
283
                "Please reduce prompt."
            )
284
285
286
287
            return self.create_error_response(
                err_type="invalid_request_error",
                message=error_message,
                status_code=HTTPStatus.BAD_REQUEST,
288
                param="input",
289
            )
290

291
292
        return None

293
294
295
296
297
298
299
300
    def _validate_create_responses_input(
        self, request: ResponsesRequest
    ) -> ErrorResponse | None:
        if self.use_harmony and request.is_include_output_logprobs():
            return self.create_error_response(
                err_type="invalid_request_error",
                message="logprobs are not supported with gpt-oss models",
                status_code=HTTPStatus.BAD_REQUEST,
301
                param="logprobs",
302
303
304
305
306
307
308
309
310
311
312
313
            )
        if request.store and not self.enable_store and request.background:
            return self.create_error_response(
                err_type="invalid_request_error",
                message=(
                    "This vLLM engine does not support `store=True` and "
                    "therefore does not support the background mode. To "
                    "enable these features, set the environment variable "
                    "`VLLM_ENABLE_RESPONSES_API_STORE=1` when launching "
                    "the vLLM server."
                ),
                status_code=HTTPStatus.BAD_REQUEST,
314
                param="background",
315
            )
316
317
318
319
320
321
        if request.previous_input_messages and request.previous_response_id:
            return self.create_error_response(
                err_type="invalid_request_error",
                message="Only one of `previous_input_messages` and "
                "`previous_response_id` can be set.",
                status_code=HTTPStatus.BAD_REQUEST,
322
                param="previous_response_id",
323
            )
324
325
        return None

326
327
328
    async def create_responses(
        self,
        request: ResponsesRequest,
329
330
331
332
333
334
        raw_request: Request | None = None,
    ) -> (
        AsyncGenerator[StreamingResponsesResponse, None]
        | ResponsesResponse
        | ErrorResponse
    ):
335
336
337
338
        error_check_ret = await self._check_model(request)
        if error_check_ret is not None:
            logger.error("Error with model %s", error_check_ret)
            return error_check_ret
339
340
341
        maybe_validation_error = self._validate_create_responses_input(request)
        if maybe_validation_error is not None:
            return maybe_validation_error
342
343
344
345
346
347
348

        # If the engine is dead, raise the engine's DEAD_ERROR.
        # This is required for the streaming case, where we return a
        # success status before we actually start generating text :).
        if self.engine_client.errored:
            raise self.engine_client.dead_error

349
        if request.store and not self.enable_store:
350
351
352
353
354
355
356
            # Disable the store option.
            # NOTE(woosuk): Although returning an error is possible, we opted
            # to implicitly disable store and process the request anyway, as
            # we assume most users do not intend to actually store the response
            # (i.e., their request's `store=True` just because it's the default
            # value).
            request.store = False
357

358
359
360
361
362
363
364
365
366
367
        # Handle the previous response ID.
        prev_response_id = request.previous_response_id
        if prev_response_id is not None:
            async with self.response_store_lock:
                prev_response = self.response_store.get(prev_response_id)
            if prev_response is None:
                return self._make_not_found_error(prev_response_id)
        else:
            prev_response = None

368
369
        lora_request = self._maybe_get_adapters(request)
        model_name = self.models.model_name(lora_request)
370

371
        if self.use_harmony:
372
            messages, engine_inputs = self._make_request_with_harmony(
373
374
375
                request, prev_response
            )
        else:
376
            messages, engine_inputs = await self._make_request(request, prev_response)
377

378
        request_metadata = RequestResponseMetadata(request_id=request.request_id)
379
380
381
382
        if raw_request:
            raw_request.state.request_metadata = request_metadata

        # Schedule the request and get the result generator.
383
        max_model_len = self.model_config.max_model_len
384
        generators: list[AsyncGenerator[ConversationContext, None]] = []
385

386
387
388
389
390
        # Only include builtin tools that the request actually asked for.
        # Without this filter, tools registered on the server (e.g. via
        # --tool-server demo) would be available for execution even when
        # the request didn't enable them.
        requested_tool_types = extract_tool_types(request.tools)
391
        builtin_tool_list: list[str] = []
392
        if self.tool_server is not None:
393
394
395
396
            if (
                self.tool_server.has_tool("browser")
                and "web_search_preview" in requested_tool_types
            ):
397
                builtin_tool_list.append("browser")
398
399
400
401
            if (
                self.tool_server.has_tool("python")
                and "code_interpreter" in requested_tool_types
            ):
402
                builtin_tool_list.append("python")
403
404
405
406
            if (
                self.tool_server.has_tool("container")
                and "container" in requested_tool_types
            ):
407
                builtin_tool_list.append("container")
408

409
410
411
412
413
        if self.tool_server is not None:
            available_tools = builtin_tool_list
        else:
            assert len(builtin_tool_list) == 0
            available_tools = []
414
415
        tokenizer = self.renderer.get_tokenizer()

416
417
        for engine_input in engine_inputs:
            maybe_error = self._validate_generator_input(engine_input)
418
419
420
421
422
423
            if maybe_error is not None:
                return maybe_error

            default_max_tokens = get_max_tokens(
                max_model_len,
                request.max_output_tokens,
424
                self._extract_prompt_len(engine_input),
425
426
427
                self.default_sampling_params,
                self.override_max_tokens,
            )
428

429
430
431
            sampling_params = request.to_sampling_params(
                default_max_tokens, self.default_sampling_params
            )
432

433
434
435
436
437
            trace_headers = (
                None
                if raw_request is None
                else await self._get_trace_headers(raw_request.headers)
            )
438

439
440
441
442
            context: ConversationContext
            if self.use_harmony:
                if request.stream:
                    context = StreamingHarmonyContext(messages, available_tools)
443
                else:
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
                    context = HarmonyContext(messages, available_tools)
            else:
                if envs.VLLM_USE_EXPERIMENTAL_PARSER_CONTEXT:
                    # This is a feature in development for parsing
                    # tokens during generation instead of at the end
                    context = ParsableContext(
                        response_messages=messages,
                        tokenizer=tokenizer,
                        reasoning_parser_cls=self.parser.reasoning_parser_cls
                        if self.parser
                        else None,
                        request=request,
                        tool_parser_cls=self.parser.tool_parser_cls
                        if self.parser
                        else None,
                        available_tools=available_tools,
                        chat_template=self.chat_template,
                        chat_template_content_format=self.chat_template_content_format,
                    )
                else:
                    context = SimpleContext()

            if self.parser and self.parser.reasoning_parser_cls is not None:
                reasoning_parser = self.parser.reasoning_parser_cls(tokenizer)
                if (
                    isinstance(
                        struct_out := sampling_params.structured_outputs,
                        StructuredOutputsParams,
                    )
                    and struct_out.all_non_structural_tag_constraints_none()
                ):
                    sampling_params.structured_outputs = replace(
                        struct_out,
                        structural_tag=reasoning_parser.prepare_structured_tag(
                            struct_out.structural_tag, self.tool_server
                        ),
                    )
            generator = self._generate_with_builtin_tools(
                request_id=request.request_id,
483
                engine_input=engine_input,
484
485
486
487
488
489
490
                sampling_params=sampling_params,
                context=context,
                lora_request=lora_request,
                priority=request.priority,
                trace_headers=trace_headers,
            )
            generators.append(generator)
491

492
        assert len(generators) == 1
493
        (result_generator,) = generators
494
495
496
497

        # Store the input messages.
        if request.store:
            self.msg_store[request.request_id] = messages
498

499
500
501
502
503
504
505
506
507
508
509
510
511
        if request.background:
            created_time = int(time.time())
            response = ResponsesResponse.from_request(
                request,
                sampling_params,
                model_name=model_name,
                created_time=created_time,
                output=[],
                status="queued",
                usage=None,
            )
            async with self.response_store_lock:
                self.response_store[response.id] = response
512

513
            # Run the request in the background.
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
            if request.stream:
                task = asyncio.create_task(
                    self._run_background_request_stream(
                        request,
                        sampling_params,
                        result_generator,
                        context,
                        model_name,
                        tokenizer,
                        request_metadata,
                        created_time,
                    ),
                    name=f"create_{request.request_id}",
                )
            else:
                task = asyncio.create_task(
                    self._run_background_request(
                        request,
                        sampling_params,
                        result_generator,
                        context,
                        model_name,
                        tokenizer,
                        request_metadata,
                        created_time,
                    ),
                    name=f"create_{response.id}",
                )
542

543
544
545
546
            # For cleanup.
            response_id = response.id
            self.background_tasks[response_id] = task
            task.add_done_callback(
547
548
                lambda _: self.background_tasks.pop(response_id, None)
            )
549
550

            if request.stream:
551
                return self.responses_background_stream_generator(request.request_id)
552
553
554
555
556
557
558
559
560
561
562
563
564
            return response

        if request.stream:
            return self.responses_stream_generator(
                request,
                sampling_params,
                result_generator,
                context,
                model_name,
                tokenizer,
                request_metadata,
            )

565
566
567
568
569
570
571
572
573
        return await self.responses_full_generator(
            request,
            sampling_params,
            result_generator,
            context,
            model_name,
            tokenizer,
            request_metadata,
        )
574

575
576
577
    async def _make_request(
        self,
        request: ResponsesRequest,
578
        prev_response: ResponsesResponse | None,
579
    ):
580
        tool_dicts = construct_tool_dicts(request.tools, request.tool_choice)
581
        # Construct the input messages.
582
583
584
585
586
587
        messages = construct_input_messages(
            request_instructions=request.instructions,
            request_input=request.input,
            prev_msg=self.msg_store.get(prev_response.id) if prev_response else None,
            prev_response_output=prev_response.output if prev_response else None,
        )
588

589
        _, engine_inputs = await self.openai_serving_render.preprocess_chat(
590
591
            request,
            messages,
592
593
594
            default_template=self.chat_template,
            default_template_content_format=self.chat_template_content_format,
            default_template_kwargs=None,
595
            tool_dicts=tool_dicts,
596
            tool_parser=self.parser.tool_parser_cls if self.parser else None,
597
            reasoning_parser=self.parser.reasoning_parser_cls if self.parser else None,
598
        )
599
        return messages, engine_inputs
600

601
602
603
604
605
    async def _render_next_turn(
        self,
        request: ResponsesRequest,
        messages: list[ResponseInputOutputItem],
        tool_dicts: list[dict[str, Any]] | None,
606
        tool_parser: type[ToolParser] | None,
607
608
609
610
611
612
613
        chat_template: str | None,
        chat_template_content_format: ChatTemplateContentFormatOption,
    ):
        new_messages = construct_input_messages(
            request_input=messages,
        )

614
        _, engine_inputs = await self.openai_serving_render.preprocess_chat(
615
616
617
618
619
620
621
            request,
            new_messages,
            default_template=chat_template,
            default_template_content_format=chat_template_content_format,
            default_template_kwargs=None,
            tool_dicts=tool_dicts,
            tool_parser=tool_parser,
622
            reasoning_parser=self.parser.reasoning_parser_cls if self.parser else None,
623
        )
624
        return engine_inputs
625
626
627
628

    async def _generate_with_builtin_tools(
        self,
        request_id: str,
629
        engine_input: EngineInput,
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
        sampling_params: SamplingParams,
        context: ConversationContext,
        lora_request: LoRARequest | None = None,
        priority: int = 0,
        trace_headers: Mapping[str, str] | None = None,
    ):
        max_model_len = self.model_config.max_model_len

        orig_priority = priority
        sub_request = 0
        while True:
            # Ensure that each sub-request has a unique request id.
            sub_request_id = f"{request_id}_{sub_request}"

            self._log_inputs(
                sub_request_id,
646
                engine_input,
647
648
649
650
651
                params=sampling_params,
                lora_request=lora_request,
            )

            generator = self.engine_client.generate(
652
                engine_input,
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
                sampling_params,
                sub_request_id,
                lora_request=lora_request,
                trace_headers=trace_headers,
                priority=priority,
            )

            async for res in generator:
                context.append_output(res)
                # NOTE(woosuk): The stop condition is handled by the engine.
                yield context

            if not context.need_builtin_tool_call():
                # The model did not ask for a tool call, so we're done.
                break

            # Call the tool and update the context with the result.
            tool_output = await context.call_tool()
            context.append_tool_output(tool_output)

            # TODO: uncomment this and enable tool output streaming
            # yield context

            # Create inputs for the next turn.
            # Render the next prompt token ids and update sampling_params.
            if isinstance(context, (HarmonyContext, StreamingHarmonyContext)):
                token_ids = context.render_for_completion()
680
                engine_input = tokens_input(token_ids)
681
682
683

                sampling_params.max_tokens = max_model_len - len(token_ids)
            elif isinstance(context, ParsableContext):
684
                (engine_input,) = await self._render_next_turn(
685
686
687
688
689
690
691
692
693
694
695
                    context.request,
                    context.parser.response_messages,
                    context.tool_dicts,
                    context.tool_parser_cls,
                    context.chat_template,
                    context.chat_template_content_format,
                )

                sampling_params.max_tokens = get_max_tokens(
                    max_model_len,
                    context.request.max_output_tokens,
696
                    self._extract_prompt_len(engine_input),
697
698
699
700
701
702
703
704
                    self.default_sampling_params,  # type: ignore
                    self.override_max_tokens,  # type: ignore
                )

            # OPTIMIZATION
            priority = orig_priority - 1
            sub_request += 1

705
706
707
    def _make_request_with_harmony(
        self,
        request: ResponsesRequest,
708
        prev_response: ResponsesResponse | None,
709
710
711
    ):
        if request.tool_choice != "auto":
            raise NotImplementedError(
712
713
                "Only 'auto' tool_choice is supported in response API with Harmony"
            )
714

715
        arrival_time = time.time()
716
        messages = self._construct_input_messages_with_harmony(request, prev_response)
717
        prompt_token_ids = render_for_completion(messages)
718
719
        engine_input = tokens_input(prompt_token_ids, cache_salt=request.cache_salt)
        engine_input["arrival_time"] = arrival_time
720

721
        return messages, [engine_input]
722

723
724
725
726
727
728
    async def _initialize_tool_sessions(
        self,
        request: ResponsesRequest,
        context: ConversationContext,
        exit_stack: AsyncExitStack,
    ):
729
730
731
732
        # we should only initialize the tool session if the request needs tools
        if len(request.tools) == 0:
            return
        mcp_tools = {
733
            tool.server_label: tool for tool in request.tools if tool.type == "mcp"
734
        }
735
736
737
        await context.init_tool_sessions(
            self.tool_server, exit_stack, request.request_id, mcp_tools
        )
738

739
740
741
742
    async def responses_full_generator(
        self,
        request: ResponsesRequest,
        sampling_params: SamplingParams,
743
        result_generator: AsyncIterator[ConversationContext],
744
        context: ConversationContext,
745
        model_name: str,
746
        tokenizer: TokenizerLike,
747
        request_metadata: RequestResponseMetadata,
748
749
        created_time: int | None = None,
    ) -> ErrorResponse | ResponsesResponse:
750
751
752
        if created_time is None:
            created_time = int(time.time())

753
754
        async with AsyncExitStack() as exit_stack:
            try:
755
                await self._initialize_tool_sessions(request, context, exit_stack)
756
757
758
759
                async for _ in result_generator:
                    pass
            except asyncio.CancelledError:
                return self.create_error_response("Client disconnected")
760

761
        # NOTE: Implementation of status is still WIP, but for now
762
763
764
765
        # we guarantee that if the status is not "completed", it is accurate.
        # "completed" is implemented as the "catch-all" for now.
        status: ResponseStatus = "completed"

766
767
        input_messages: ResponseInputOutputMessage | None = None
        output_messages: ResponseInputOutputMessage | None = None
768
769
770
        if self.use_harmony:
            assert isinstance(context, HarmonyContext)
            output = self._make_response_output_items_with_harmony(context)
771
            if request.enable_response_messages:
772
773
                input_messages = context.messages[: context.num_init_messages]
                output_messages = context.messages[context.num_init_messages :]
774
            num_tool_output_tokens = context.num_tool_output_tokens
775
776
777
778
779
            if len(output) > 0:
                if context.finish_reason == "length":
                    status = "incomplete"
                elif context.finish_reason == "abort":
                    status = "cancelled"
780
781
                else:
                    self._raise_if_error(context.finish_reason, request.request_id)
782
783
            else:
                status = "incomplete"
784
        elif isinstance(context, ParsableContext):
785
            output = context.parser.make_response_output_items_from_parsable_context()
786
787

            if request.enable_response_messages:
788
789
                input_messages = context.input_messages
                output_messages = context.output_messages
790
791
792
793

            # TODO: Calculate usage.
            # assert final_res.prompt_token_ids is not None
            num_tool_output_tokens = 0
794
795
796
797

            # Check finish reason from the parser
            if context.parser.finish_reason == "length":
                status = "incomplete"
798
        else:
799
            assert isinstance(context, SimpleContext)
800
801
            # Use final_output which has accumulated text/token_ids/logprobs
            final_res = context.final_output
802
803
804
805
            assert final_res is not None
            assert len(final_res.outputs) == 1
            final_output = final_res.outputs[0]

806
807
808
            # finish_reason='error' indicates retryable internal error
            self._raise_if_error(final_output.finish_reason, request.request_id)

809
810
811
812
            # Check if generation was stopped due to max_tokens
            if final_output.finish_reason == "length":
                status = "incomplete"

813
            output = self._make_response_output_items(request, final_output, tokenizer)
814

815
            if request.enable_response_messages:
816
817
818
                input_messages = context.input_messages
                output_messages = context.output_messages

819
820
            # Calculate usage.
            assert final_res.prompt_token_ids is not None
821
822
            num_tool_output_tokens = 0

823
        assert isinstance(context, (SimpleContext, HarmonyContext, ParsableContext))
824
825
826
827
        num_prompt_tokens = context.num_prompt_tokens
        num_generated_tokens = context.num_output_tokens
        num_cached_tokens = context.num_cached_tokens
        num_reasoning_tokens = context.num_reasoning_tokens
828
829
830
831
832
833
834
835
836
837
838
839
840
        # For text-based reasoning parsers (e.g., <think>...</think>),
        # HarmonyContext already counts reasoning tokens via channels.
        # For Simple/Parsable contexts, derive reasoning_tokens from
        # accumulated output token IDs using the parser if not already set.
        if (
            num_reasoning_tokens == 0
            and self.parser is not None
            and self.parser.reasoning_parser_cls is not None
            and isinstance(context, (SimpleContext, ParsableContext))
        ):
            reasoning_parser = self.parser.reasoning_parser_cls(tokenizer)
            accumulated = getattr(context, "_accumulated_token_ids", []) or []
            num_reasoning_tokens = reasoning_parser.count_reasoning_tokens(accumulated)
841
842
843
844

        usage = ResponseUsage(
            input_tokens=num_prompt_tokens,
            output_tokens=num_generated_tokens,
845
            total_tokens=num_prompt_tokens + num_generated_tokens,
846
847
848
849
850
851
852
853
854
            input_tokens_details=InputTokensDetails(
                cached_tokens=num_cached_tokens,
                input_tokens_per_turn=[
                    turn.input_tokens for turn in context.all_turn_metrics
                ],
                cached_tokens_per_turn=[
                    turn.cached_input_tokens for turn in context.all_turn_metrics
                ],
            ),
855
            output_tokens_details=OutputTokensDetails(
856
                reasoning_tokens=num_reasoning_tokens,
857
                tool_output_tokens=num_tool_output_tokens,
858
859
860
861
862
863
                output_tokens_per_turn=[
                    turn.output_tokens for turn in context.all_turn_metrics
                ],
                tool_output_tokens_per_turn=[
                    turn.tool_output_tokens for turn in context.all_turn_metrics
                ],
864
            ),
865
866
867
868
        )
        response = ResponsesResponse.from_request(
            request,
            sampling_params,
869
870
            input_messages=input_messages,
            output_messages=output_messages,
871
872
873
            model_name=model_name,
            created_time=created_time,
            output=output,
874
            status=status,
875
            usage=usage,
876
            kv_transfer_params=context.kv_transfer_params,
877
878
879
880
881
882
        )

        if request.store:
            async with self.response_store_lock:
                stored_response = self.response_store.get(response.id)
                # If the response is already cancelled, don't update it.
883
                if stored_response is None or stored_response.status != "cancelled":
884
885
886
                    self.response_store[response.id] = response
        return response

887
888
889
890
    def _topk_logprobs(
        self,
        logprobs: dict[int, SampleLogprob],
        top_logprobs: int,
891
        tokenizer: TokenizerLike,
892
    ) -> list[LogprobTopLogprob]:
893
894
895
896
897
        """Returns the top-k logprobs from the logprobs dictionary."""
        out = []
        for i, (token_id, _logprob) in enumerate(logprobs.items()):
            if i >= top_logprobs:
                break
898
899
900
901
902
            text = self._get_decoded_token(
                logprob=_logprob,
                token_id=token_id,
                tokenizer=tokenizer,
                return_as_token_id=self.return_tokens_as_token_ids,
903
            )
904
905
906
907
908
            out.append(
                LogprobTopLogprob(
                    token=text,
                    logprob=max(_logprob.logprob, -9999.0),
                    bytes=list(text.encode("utf-8", errors="replace")),
909
910
                )
            )
911
912
913
        return out

    def _create_response_logprobs(
914
915
        self,
        token_ids: Sequence[int],
916
        logprobs: SampleLogprobs | None,
917
        tokenizer: TokenizerLike,
918
        top_logprobs: int | None = None,
919
    ) -> list[Logprob]:
920
921
        assert logprobs is not None, "logprobs must be provided"
        assert len(token_ids) == len(logprobs), (
922
923
            "token_ids and logprobs.token_ids must have the same length"
        )
924
925
926
927
        out = []
        for i, token_id in enumerate(token_ids):
            logprob = logprobs[i]
            token_logprob = logprob[token_id]
928
929
930
931
932
            text = self._get_decoded_token(
                logprob=token_logprob,
                token_id=token_id,
                tokenizer=tokenizer,
                return_as_token_id=self.return_tokens_as_token_ids,
933
            )
934
935
936
937
938
            out.append(
                Logprob(
                    token=text,
                    logprob=max(token_logprob.logprob, -9999.0),
                    bytes=list(text.encode("utf-8", errors="replace")),
939
940
941
942
943
944
945
                    top_logprobs=(
                        self._topk_logprobs(
                            logprob, top_logprobs=top_logprobs, tokenizer=tokenizer
                        )
                        if top_logprobs
                        else []
                    ),
946
947
                )
            )
948
949
        return out

950
951
952
    def _create_stream_response_logprobs(
        self,
        token_ids: Sequence[int],
953
        logprobs: SampleLogprobs | None,
954
        tokenizer: TokenizerLike,
955
        top_logprobs: int | None = None,
956
    ) -> list[response_text_delta_event.Logprob]:
957
958
959
960
961
962
        lgs = self._create_response_logprobs(
            token_ids=token_ids,
            logprobs=logprobs,
            tokenizer=tokenizer,
            top_logprobs=top_logprobs,
        )
963
964
965
966
967
968
        return [
            response_text_delta_event.Logprob(
                token=lg.token,
                logprob=lg.logprob,
                top_logprobs=[
                    response_text_delta_event.LogprobTopLogprob(
969
970
                        token=tl.token, logprob=tl.logprob
                    )
971
                    for tl in lg.top_logprobs
972
973
974
                ],
            )
            for lg in lgs
975
976
        ]

977
978
979
980
    def _make_response_output_items(
        self,
        request: ResponsesRequest,
        final_output: CompletionOutput,
981
        tokenizer: TokenizerLike,
982
    ) -> list[ResponseOutputItem]:
983
984
        # Log complete response if output logging is enabled
        if self.enable_log_outputs and self.request_logger:
985
986
987
988
989
990
991
992
            self.request_logger.log_outputs(
                request_id=request.request_id,
                outputs=final_output.text,
                output_token_ids=final_output.token_ids,
                finish_reason=final_output.finish_reason,
                is_streaming=False,
                delta=False,
            )
993

994
995
996
997
998
999
1000
1001
        # Compute logprobs if requested
        logprobs = None
        if request.is_include_output_logprobs() and final_output.logprobs:
            logprobs = self._create_response_logprobs(
                token_ids=final_output.token_ids,
                logprobs=final_output.logprobs,
                tokenizer=tokenizer,
                top_logprobs=request.top_logprobs,
1002
            )
1003

1004
1005
        # Use parser to extract and create response output items
        if self.parser:
1006
            parser = self.parser(tokenizer, request.tools)
1007
1008
            return parser.extract_response_outputs(
                model_output=final_output.text,
1009
                model_output_token_ids=final_output.token_ids,
1010
1011
1012
1013
1014
1015
1016
1017
1018
                request=request,
                enable_auto_tools=self.enable_auto_tools,
                tool_call_id_type=self.tool_call_id_type,
                logprobs=logprobs,
            )

        # Fallback when no parser is configured
        return [
            ResponseOutputMessage(
1019
                id=f"msg_{random_uuid()}",
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
                content=[
                    ResponseOutputText(
                        text=final_output.text,
                        annotations=[],
                        type="output_text",
                        logprobs=logprobs,
                    )
                ]
                if final_output.text
                else [],
1030
1031
1032
1033
                role="assistant",
                status="completed",
                type="message",
            )
1034
        ]
1035
1036
1037
1038
1039

    def _make_response_output_items_with_harmony(
        self,
        context: HarmonyContext,
    ) -> list[ResponseOutputItem]:
1040
        output_items: list[ResponseOutputItem] = []
1041
1042
        num_init_messages = context.num_init_messages
        for msg in context.messages[num_init_messages:]:
1043
            output_items.extend(harmony_to_response_output(msg))
1044
        # Handle the generation stopped in the middle (if any).
1045
        last_items = parser_state_to_response_output(context.parser)
1046
1047
1048
1049
        if last_items:
            output_items.extend(last_items)
        return output_items

1050
1051
1052
    def _extract_system_message_from_request(
        self, request: ResponsesRequest
    ) -> str | None:
1053
1054
1055
1056
1057
1058
1059
        system_msg = None
        if not isinstance(request.input, str):
            for response_msg in request.input:
                if (
                    isinstance(response_msg, dict)
                    and response_msg.get("role") == "system"
                ):
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
                    content = response_msg.get("content")
                    if isinstance(content, str):
                        system_msg = content
                    elif isinstance(content, list):
                        for param in content:
                            if (
                                isinstance(param, dict)
                                and param.get("type") == "input_text"
                            ):
                                system_msg = param.get("text")
                                break
1071
1072
1073
                    break
        return system_msg

1074
    def _construct_harmony_system_input_message(
1075
        self, request: ResponsesRequest, with_custom_tools: bool, tool_types: set[str]
1076
    ) -> OpenAIHarmonyMessage:
1077
1078
        model_identity = self._extract_system_message_from_request(request)

1079
        reasoning_effort = request.reasoning.effort if request.reasoning else None
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090

        # Extract allowed_tools from MCP tool requests
        allowed_tools_map = _extract_allowed_tools_from_mcp_requests(request.tools)

        # Get filtered tool descriptions first.
        # If get_tool_description returns None (due to filtering), the tool is disabled.
        browser_description = (
            self.tool_server.get_tool_description(
                "browser", allowed_tools_map.get("web_search_preview")
            )
            if "web_search_preview" in tool_types
1091
1092
            and self.tool_server is not None
            and self.tool_server.has_tool("browser")
1093
            else None
1094
        )
1095
1096
1097
1098
1099
        python_description = (
            self.tool_server.get_tool_description(
                "python", allowed_tools_map.get("code_interpreter")
            )
            if "code_interpreter" in tool_types
1100
1101
            and self.tool_server is not None
            and self.tool_server.has_tool("python")
1102
            else None
1103
        )
1104
1105
1106
1107
1108
        container_description = (
            self.tool_server.get_tool_description(
                "container", allowed_tools_map.get("container")
            )
            if "container" in tool_types
1109
1110
            and self.tool_server is not None
            and self.tool_server.has_tool("container")
1111
            else None
1112
        )
1113

1114
        sys_msg = get_system_message(
1115
            model_identity=model_identity,
1116
            reasoning_effort=reasoning_effort,
1117
1118
1119
            browser_description=browser_description,
            python_description=python_description,
            container_description=container_description,
1120
1121
1122
1123
1124
            instructions=request.instructions,
            with_custom_tools=with_custom_tools,
        )
        return sys_msg

1125
1126
1127
    def _construct_input_messages_with_harmony(
        self,
        request: ResponsesRequest,
1128
        prev_response: ResponsesResponse | None,
1129
1130
1131
1132
    ) -> list[OpenAIHarmonyMessage]:
        messages: list[OpenAIHarmonyMessage] = []
        if prev_response is None:
            # New conversation.
1133
            tool_types = extract_tool_types(request.tools)
1134
            with_custom_tools = has_custom_tools(tool_types)
1135
1136
1137

            sys_msg = self._construct_harmony_system_input_message(
                request, with_custom_tools, tool_types
1138
1139
            )
            messages.append(sys_msg)
1140
1141
            if with_custom_tools:
                dev_msg = get_developer_message(
1142
1143
                    instructions=request.instructions, tools=request.tools
                )
1144
                messages.append(dev_msg)
1145
1146
            messages += construct_harmony_previous_input_messages(request)

1147
1148
1149
1150
1151
        else:
            # Continue the previous conversation.
            # FIXME(woosuk): Currently, request params like reasoning and
            # instructions are ignored.
            prev_msgs = self.msg_store[prev_response.id]
1152
1153
1154
1155
1156
1157
1158
1159
1160

            # FIXME(woosuk): The slice-delete-reappend cycle below is
            # currently a no-op --- it removes messages then puts them all
            # back unfiltered. It may be intentionally deferred (see FIXME
            # above) or redundant if the Harmony encoder already strips
            # analysis messages at render time. If analysis messages need
            # to be dropped here, add a channel != "analysis" filter when
            # re-appending, similar to auto_drop_analysis_messages in
            # harmony_utils.py.
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
            if len(prev_msgs) > 0:
                last_msg = prev_msgs[-1]
                assert isinstance(last_msg, OpenAIHarmonyMessage)
                if last_msg.channel == "final":
                    prev_final_msg_idx = -1
                    for i in range(len(prev_msgs) - 2, -1, -1):
                        prev_msg_i = prev_msgs[i]
                        assert isinstance(prev_msg_i, OpenAIHarmonyMessage)
                        if prev_msg_i.channel == "final":
                            prev_final_msg_idx = i
                            break
1172
1173
                    recent_turn_msgs = prev_msgs[prev_final_msg_idx + 1 :]
                    del prev_msgs[prev_final_msg_idx + 1 :]
1174
1175
                    for msg in recent_turn_msgs:
                        assert isinstance(msg, OpenAIHarmonyMessage)
1176
                        prev_msgs.append(msg)
1177
1178
            messages.extend(prev_msgs)
        # Append the new input.
co63oc's avatar
co63oc committed
1179
        # Responses API supports simple text inputs without chat format.
1180
        if isinstance(request.input, str):
1181
1182
1183
1184
1185
            # Skip empty string input when previous_input_messages supplies
            # the full conversation history --- an empty trailing user message
            # confuses the model into thinking nothing was sent.
            if request.input or not request.previous_input_messages:
                messages.append(get_user_message(request.input))
1186
1187
1188
1189
1190
1191
        else:
            if prev_response is not None:
                prev_outputs = copy(prev_response.output)
            else:
                prev_outputs = []
            for response_msg in request.input:
1192
                new_msg = response_input_to_harmony(response_msg, prev_outputs)
1193
                if new_msg is not None and new_msg.author.role != "system":
1194
1195
                    messages.append(new_msg)

1196
                # User passes in a tool call request and its output. We need
1197
1198
                # to add the tool call request to prev_outputs so that
                # response_input_to_harmony can find the tool call request when
1199
1200
1201
1202
1203
                # parsing the tool call output.
                if isinstance(response_msg, ResponseFunctionToolCall):
                    prev_outputs.append(response_msg)
        return messages

1204
1205
1206
1207
1208
1209
    async def _run_background_request_stream(
        self,
        request: ResponsesRequest,
        *args,
        **kwargs,
    ):
1210
        event_deque: deque[StreamingResponsesResponse] = deque()
1211
1212
        new_event_signal = asyncio.Event()
        self.event_store[request.request_id] = (event_deque, new_event_signal)
1213
        generator = self.responses_stream_generator(request, *args, **kwargs)
1214
1215
1216
1217
1218
1219
1220
        try:
            async for event in generator:
                event_deque.append(event)
                new_event_signal.set()  # Signal new event available
        finally:
            new_event_signal.set()

1221
1222
1223
1224
1225
1226
    async def _run_background_request(
        self,
        request: ResponsesRequest,
        *args,
        **kwargs,
    ):
1227
        response = await self.responses_full_generator(request, *args, **kwargs)
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237

        if isinstance(response, ErrorResponse):
            # If the request has failed, update the status to "failed".
            response_id = request.request_id
            async with self.response_store_lock:
                stored_response = self.response_store.get(response_id)
                assert stored_response is not None
                if stored_response.status not in ("completed", "cancelled"):
                    stored_response.status = "failed"

1238
1239
1240
    async def responses_background_stream_generator(
        self,
        response_id: str,
1241
        starting_after: int | None = None,
1242
    ) -> AsyncGenerator[StreamingResponsesResponse, None]:
1243
        if response_id not in self.event_store:
1244
1245
1246
1247
1248
            raise VLLMValidationError(
                f"Unknown response_id: {response_id}",
                parameter="response_id",
                value=response_id,
            )
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260

        event_deque, new_event_signal = self.event_store[response_id]
        start_index = 0 if starting_after is None else starting_after + 1
        current_index = start_index

        while True:
            new_event_signal.clear()

            # Yield existing events from start_index
            while current_index < len(event_deque):
                event = event_deque[current_index]
                yield event
1261
                if getattr(event, "type", "unknown") == "response.completed":
1262
                    return
1263
1264
1265
1266
                current_index += 1

            await new_event_signal.wait()

1267
1268
1269
    async def retrieve_responses(
        self,
        response_id: str,
1270
1271
1272
1273
1274
1275
1276
        starting_after: int | None,
        stream: bool | None,
    ) -> (
        ErrorResponse
        | ResponsesResponse
        | AsyncGenerator[StreamingResponsesResponse, None]
    ):
1277
1278
1279
1280
1281
        async with self.response_store_lock:
            response = self.response_store.get(response_id)

        if response is None:
            return self._make_not_found_error(response_id)
1282
1283
1284
1285
1286
1287

        if stream:
            return self.responses_background_stream_generator(
                response_id,
                starting_after,
            )
1288
1289
1290
1291
1292
        return response

    async def cancel_responses(
        self,
        response_id: str,
1293
    ) -> ErrorResponse | ResponsesResponse:
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
        async with self.response_store_lock:
            response = self.response_store.get(response_id)
            if response is None:
                return self._make_not_found_error(response_id)

            prev_status = response.status
            if prev_status not in ("queued", "in_progress"):
                return self.create_error_response(
                    err_type="invalid_request_error",
                    message="Cannot cancel a synchronous response.",
1304
                    param="response_id",
1305
1306
1307
1308
1309
1310
                )

            # Update the status to "cancelled".
            response.status = "cancelled"

        # Abort the request.
1311
        if task := self.background_tasks.get(response_id):
1312
1313
1314
1315
            task.cancel()
            try:
                await task
            except asyncio.CancelledError:
1316
                logger.exception("Background task for %s was cancelled", response_id)
1317
1318
1319
1320
1321
1322
1323
        return response

    def _make_not_found_error(self, response_id: str) -> ErrorResponse:
        return self.create_error_response(
            err_type="invalid_request_error",
            message=f"Response with id '{response_id}' not found.",
            status_code=HTTPStatus.NOT_FOUND,
1324
            param="response_id",
1325
        )
1326

1327
    async def _process_simple_streaming_events(
1328
1329
1330
        self,
        request: ResponsesRequest,
        sampling_params: SamplingParams,
1331
        result_generator: AsyncIterator[ConversationContext | None],
1332
1333
        context: ConversationContext,
        model_name: str,
1334
        tokenizer: TokenizerLike,
1335
        request_metadata: RequestResponseMetadata,
1336
        created_time: int,
1337
        _increment_sequence_number_and_return: Callable[
1338
1339
            [StreamingResponsesResponse], StreamingResponsesResponse
        ],
1340
    ) -> AsyncGenerator[StreamingResponsesResponse, None]:
1341
1342
1343
        current_content_index = 0
        current_output_index = 0
        current_item_id = ""
1344
        current_tool_call_index: int | None = None
1345
        parser = self.parser(tokenizer, request.tools) if self.parser else None
1346
1347
1348
1349
1350
1351
1352
1353
        first_delta_sent = False
        previous_delta_messages: list[DeltaMessage] = []
        async for ctx in result_generator:
            assert isinstance(ctx, SimpleContext)
            if ctx.last_output is None:
                continue
            if ctx.last_output.outputs:
                output = ctx.last_output.outputs[0]
1354
1355
                # finish_reason='error' indicates a retryable error
                self._raise_if_error(output.finish_reason, request.request_id)
1356
1357
                delta_text = output.text
                delta_token_ids = as_list(output.token_ids)
1358
1359
1360

                if parser:
                    delta_message = parser.parse_delta(
1361
1362
                        delta_text=delta_text,
                        delta_token_ids=delta_token_ids,
1363
1364
                        request=request,
                        prompt_token_ids=ctx.last_output.prompt_token_ids,
1365
1366
                    )
                else:
1367
1368
1369
                    delta_message = DeltaMessage(
                        content=output.text,
                    )
1370
1371
                if not delta_message:
                    continue
1372
                tool_call_item_started = False
1373
                if not first_delta_sent:
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
                    current_item_id = random_uuid()
                    if delta_message.tool_calls:
                        current_tool_call_id = f"call_{random_uuid()}"
                        assert len(delta_message.tool_calls) == 1, (
                            "Multiple tool calls in one delta is not supported"
                        )
                        assert delta_message.tool_calls[0].function is not None, (
                            "Tool call without function is not supported"
                        )
                        assert delta_message.tool_calls[0].function.name is not None, (
                            "Tool call without function name is not supported"
                        )
                        current_tool_call_name = delta_message.tool_calls[
                            0
                        ].function.name
1389
                        current_tool_call_index = delta_message.tool_calls[0].index
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
                        yield _increment_sequence_number_and_return(
                            ResponseOutputItemAddedEvent(
                                type="response.output_item.added",
                                sequence_number=-1,
                                output_index=current_output_index,
                                item=ResponseFunctionToolCallItem(
                                    type="function_call",
                                    id=current_item_id,
                                    call_id=current_tool_call_id,
                                    name=current_tool_call_name,
1400
                                    arguments="",
1401
1402
1403
1404
                                    status="in_progress",
                                ),
                            )
                        )
1405
                        tool_call_item_started = True
1406
                    elif delta_message.reasoning:
1407
                        yield _increment_sequence_number_and_return(
1408
1409
1410
1411
                            ResponseOutputItemAddedEvent(
                                type="response.output_item.added",
                                sequence_number=-1,
                                output_index=current_output_index,
1412
                                item=ResponseReasoningItem(
1413
1414
1415
1416
1417
                                    type="reasoning",
                                    id=current_item_id,
                                    summary=[],
                                    status="in_progress",
                                ),
1418
1419
                            )
                        )
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
                        yield _increment_sequence_number_and_return(
                            ResponseReasoningPartAddedEvent(
                                type="response.reasoning_part.added",
                                sequence_number=-1,
                                output_index=current_output_index,
                                item_id=current_item_id,
                                content_index=current_content_index,
                                part=ResponseReasoningTextContent(
                                    text="",
                                    type="reasoning_text",
                                ),
                            )
                        )
1433
                    elif not delta_message.tool_calls:
1434
                        yield _increment_sequence_number_and_return(
1435
1436
1437
1438
                            ResponseOutputItemAddedEvent(
                                type="response.output_item.added",
                                sequence_number=-1,
                                output_index=current_output_index,
1439
                                item=ResponseOutputMessage(
1440
1441
1442
1443
1444
1445
                                    id=current_item_id,
                                    type="message",
                                    role="assistant",
                                    content=[],
                                    status="in_progress",
                                ),
1446
1447
                            )
                        )
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
                        yield _increment_sequence_number_and_return(
                            ResponseContentPartAddedEvent(
                                type="response.content_part.added",
                                sequence_number=-1,
                                output_index=current_output_index,
                                item_id=current_item_id,
                                content_index=current_content_index,
                                part=ResponseOutputText(
                                    type="output_text",
                                    text="",
                                    annotations=[],
                                    logprobs=[],
                                ),
                            )
1462
                        )
1463
1464
1465
1466
                    first_delta_sent = True

                # check delta message and previous delta message are
                # same as content or reasoning content
1467
1468
                if (
                    previous_delta_messages
1469
                    and previous_delta_messages[-1].reasoning is not None
1470
1471
                    and delta_message.content is not None
                ):
1472
1473
                    # from reasoning to normal content, send done
                    # event for reasoning
1474
                    reason_content = "".join(
1475
                        pm.reasoning
1476
                        for pm in previous_delta_messages
1477
                        if pm.reasoning is not None
1478
                    )
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498

                    # delta message could have both reasoning and
                    # content. Include current delta's reasoning in the
                    # finalization since it may carry the tail end of
                    # reasoning text (e.g. when reasoning end and
                    # content start arrive in the same delta).
                    if delta_message.reasoning is not None:
                        yield _increment_sequence_number_and_return(
                            ResponseReasoningTextDeltaEvent(
                                type="response.reasoning_text.delta",
                                sequence_number=-1,
                                content_index=current_content_index,
                                output_index=current_output_index,
                                item_id=current_item_id,
                                delta=delta_message.reasoning,
                            )
                        )
                        reason_content += delta_message.reasoning
                        delta_message = DeltaMessage(content=delta_message.content)

1499
                    yield _increment_sequence_number_and_return(
1500
1501
1502
1503
1504
1505
1506
                        ResponseReasoningTextDoneEvent(
                            type="response.reasoning_text.done",
                            item_id=current_item_id,
                            sequence_number=-1,
                            output_index=current_output_index,
                            content_index=current_content_index,
                            text=reason_content,
1507
1508
                        )
                    )
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
                    yield _increment_sequence_number_and_return(
                        ResponseReasoningPartDoneEvent(
                            type="response.reasoning_part.done",
                            sequence_number=-1,
                            item_id=current_item_id,
                            output_index=current_output_index,
                            content_index=current_content_index,
                            part=ResponseReasoningTextContent(
                                text=reason_content,
                                type="reasoning_text",
                            ),
                        )
                    )
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
                    current_content_index = 0
                    reasoning_item = ResponseReasoningItem(
                        type="reasoning",
                        content=[
                            ResponseReasoningTextContent(
                                text=reason_content,
                                type="reasoning_text",
                            ),
                        ],
                        status="completed",
                        id=current_item_id,
                        summary=[],
                    )
1535
                    yield _increment_sequence_number_and_return(
1536
1537
1538
1539
1540
                        ResponseOutputItemDoneEvent(
                            type="response.output_item.done",
                            sequence_number=-1,
                            output_index=current_output_index,
                            item=reasoning_item,
1541
1542
                        )
                    )
1543
1544
                    current_output_index += 1
                    current_item_id = str(uuid.uuid4())
1545
                    yield _increment_sequence_number_and_return(
1546
                        ResponseOutputItemAddedEvent(
1547
1548
1549
                            type="response.output_item.added",
                            sequence_number=-1,
                            output_index=current_output_index,
1550
                            item=ResponseOutputMessage(
1551
1552
1553
1554
1555
1556
                                id=current_item_id,
                                type="message",
                                role="assistant",
                                content=[],
                                status="in_progress",
                            ),
1557
1558
                        )
                    )
1559
                    yield _increment_sequence_number_and_return(
1560
                        ResponseContentPartAddedEvent(
1561
1562
1563
1564
1565
                            type="response.content_part.added",
                            sequence_number=-1,
                            output_index=current_output_index,
                            item_id=current_item_id,
                            content_index=current_content_index,
1566
                            part=ResponseOutputText(
1567
1568
1569
1570
1571
                                type="output_text",
                                text="",
                                annotations=[],
                                logprobs=[],
                            ),
1572
1573
                        )
                    )
1574
1575
                    # reset previous delta messages
                    previous_delta_messages = []
1576
                if delta_message.tool_calls and delta_message.tool_calls[0].function:
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
1644
1645
1646
1647
1648
1649
                    tool_call = delta_message.tool_calls[0]
                    tool_call_function = tool_call.function
                    if (
                        current_tool_call_index is not None
                        and tool_call.index is not None
                        and tool_call.index != current_tool_call_index
                        and tool_call_function is not None
                        and tool_call_function.name is not None
                    ):
                        # From one tool call to another, finalize the previous
                        # function-call item before opening the next one.
                        parts = []
                        for pm in previous_delta_messages:
                            if pm.tool_calls:
                                previous_tool_call = pm.tool_calls[0]
                                if previous_tool_call.function is not None:
                                    parts.append(
                                        previous_tool_call.function.arguments or ""
                                    )

                        tool_call_arguments = "".join(parts)
                        yield _increment_sequence_number_and_return(
                            ResponseFunctionCallArgumentsDoneEvent(
                                type="response.function_call_arguments.done",
                                sequence_number=-1,
                                output_index=current_output_index,
                                item_id=current_item_id,
                                arguments=tool_call_arguments,
                                name=current_tool_call_name,
                            )
                        )
                        function_call_item = ResponseFunctionToolCall(
                            type="function_call",
                            name=current_tool_call_name,
                            arguments=tool_call_arguments,
                            status="completed",
                            id=current_item_id,
                            call_id=current_tool_call_id,
                        )
                        yield _increment_sequence_number_and_return(
                            ResponseOutputItemDoneEvent(
                                type="response.output_item.done",
                                sequence_number=-1,
                                output_index=current_output_index,
                                item=function_call_item,
                            )
                        )
                        # Reset previous delta messages so the next tool call
                        # does not reuse arguments from the completed item.
                        previous_delta_messages = []
                        current_output_index += 1
                        current_item_id = random_uuid()
                        current_tool_call_name = tool_call_function.name
                        current_tool_call_id = f"call_{random_uuid()}"
                        current_tool_call_index = tool_call.index
                        yield _increment_sequence_number_and_return(
                            ResponseOutputItemAddedEvent(
                                type="response.output_item.added",
                                sequence_number=-1,
                                output_index=current_output_index,
                                item=ResponseFunctionToolCallItem(
                                    type="function_call",
                                    id=current_item_id,
                                    call_id=current_tool_call_id,
                                    name=current_tool_call_name,
                                    arguments="",
                                    status="in_progress",
                                ),
                            )
                        )
                        current_content_index = 0
                        tool_call_item_started = True

1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
                    if delta_message.tool_calls[0].function.arguments:
                        yield _increment_sequence_number_and_return(
                            ResponseFunctionCallArgumentsDeltaEvent(
                                type="response.function_call_arguments.delta",
                                sequence_number=-1,
                                output_index=current_output_index,
                                item_id=current_item_id,
                                delta=delta_message.tool_calls[0].function.arguments,
                            )
                        )
                    # tool call initiated with no arguments
1661
1662
1663
1664
                    elif (
                        delta_message.tool_calls[0].function.name
                        and not tool_call_item_started
                    ):
1665
1666
1667
1668
1669
1670
1671
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
1690
1691
1692
1693
1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
1709
1710
1711
1712
                        # send done with current content part
                        # and add new function call item
                        yield _increment_sequence_number_and_return(
                            ResponseTextDoneEvent(
                                type="response.output_text.done",
                                sequence_number=-1,
                                output_index=current_output_index,
                                content_index=current_content_index,
                                text="",
                                logprobs=[],
                                item_id=current_item_id,
                            )
                        )
                        yield _increment_sequence_number_and_return(
                            ResponseContentPartDoneEvent(
                                type="response.content_part.done",
                                sequence_number=-1,
                                item_id=current_item_id,
                                output_index=current_output_index,
                                content_index=current_content_index,
                                part=ResponseOutputText(
                                    type="output_text",
                                    text="",
                                    annotations=[],
                                    logprobs=[],
                                ),
                            )
                        )
                        yield _increment_sequence_number_and_return(
                            ResponseOutputItemDoneEvent(
                                type="response.output_item.done",
                                sequence_number=-1,
                                output_index=current_output_index,
                                item=ResponseOutputMessage(
                                    id=current_item_id,
                                    type="message",
                                    role="assistant",
                                    content=[],
                                    status="completed",
                                ),
                            )
                        )
                        current_output_index += 1
                        current_item_id = random_uuid()
                        current_tool_call_name = delta_message.tool_calls[
                            0
                        ].function.name
                        current_tool_call_id = f"call_{random_uuid()}"
1713
                        current_tool_call_index = delta_message.tool_calls[0].index
1714
1715
1716
1717
1718
1719
1720
1721
1722
1723
1724
1725
1726
1727
1728
1729
1730
1731
1732
                        yield _increment_sequence_number_and_return(
                            ResponseOutputItemAddedEvent(
                                type="response.output_item.added",
                                sequence_number=-1,
                                output_index=current_output_index,
                                item=ResponseFunctionToolCallItem(
                                    type="function_call",
                                    id=current_item_id,
                                    call_id=current_tool_call_id,
                                    name=current_tool_call_name,
                                    arguments="",
                                    status="in_progress",
                                ),
                            )
                        )
                        # skip content part for tool call
                        current_content_index = 1
                        continue
                elif delta_message.reasoning is not None:
1733
                    yield _increment_sequence_number_and_return(
1734
1735
1736
1737
1738
1739
                        ResponseReasoningTextDeltaEvent(
                            type="response.reasoning_text.delta",
                            sequence_number=-1,
                            content_index=current_content_index,
                            output_index=current_output_index,
                            item_id=current_item_id,
1740
                            delta=delta_message.reasoning,
1741
1742
                        )
                    )
1743
                elif delta_message.content:
1744
                    yield _increment_sequence_number_and_return(
1745
                        ResponseTextDeltaEvent(
1746
1747
1748
1749
1750
1751
                            type="response.output_text.delta",
                            sequence_number=-1,
                            content_index=current_content_index,
                            output_index=current_output_index,
                            item_id=current_item_id,
                            delta=delta_message.content,
1752
1753
1754
1755
1756
1757
1758
1759
1760
1761
                            logprobs=(
                                self._create_stream_response_logprobs(
                                    token_ids=output.token_ids,
                                    logprobs=output.logprobs,
                                    tokenizer=tokenizer,
                                    top_logprobs=request.top_logprobs,
                                )
                                if request.is_include_output_logprobs()
                                else []
                            ),
1762
1763
                        )
                    )
1764
1765

                previous_delta_messages.append(delta_message)
1766

1767
        if previous_delta_messages:
1768
1769
1770
1771
1772
1773
1774
1775
1776
1777
1778
1779
1780
1781
1782
1783
1784
1785
1786
1787
1788
1789
1790
1791
1792
1793
1794
1795
1796
1797
1798
1799
1800
1801
1802
1803
1804
1805
1806
1807
1808
1809
            parts = []
            for pm in previous_delta_messages:
                if pm.tool_calls:
                    assert len(pm.tool_calls) == 1, (
                        "Multiple tool calls in one delta is not supported"
                    )
                    assert pm.tool_calls[0].function is not None, (
                        "Tool call without function is not supported"
                    )
                    parts.append(pm.tool_calls[0].function.arguments or "")

            tool_call_arguments = "".join(parts)
            if tool_call_arguments:
                yield _increment_sequence_number_and_return(
                    ResponseFunctionCallArgumentsDoneEvent(
                        type="response.function_call_arguments.done",
                        sequence_number=-1,
                        output_index=current_output_index,
                        item_id=current_item_id,
                        arguments=tool_call_arguments,
                        name=current_tool_call_name,
                    )
                )
                current_content_index = 0
                function_call_item = ResponseFunctionToolCall(
                    type="function_call",
                    name=current_tool_call_name,
                    arguments=tool_call_arguments,
                    status="completed",
                    id=current_item_id,
                    call_id=current_tool_call_id,
                )
                yield _increment_sequence_number_and_return(
                    ResponseOutputItemDoneEvent(
                        type="response.output_item.done",
                        sequence_number=-1,
                        output_index=current_output_index,
                        item=function_call_item,
                    )
                )

            elif previous_delta_messages[-1].reasoning is not None:
1810
                reason_content = "".join(
1811
                    pm.reasoning
1812
                    for pm in previous_delta_messages
1813
                    if pm.reasoning is not None
1814
                )
1815
                yield _increment_sequence_number_and_return(
1816
1817
1818
1819
1820
1821
1822
                    ResponseReasoningTextDoneEvent(
                        type="response.reasoning_text.done",
                        item_id=current_item_id,
                        sequence_number=-1,
                        output_index=current_output_index,
                        content_index=current_content_index,
                        text=reason_content,
1823
1824
                    )
                )
1825
1826
1827
1828
1829
1830
1831
1832
1833
1834
1835
1836
1837
                yield _increment_sequence_number_and_return(
                    ResponseReasoningPartDoneEvent(
                        type="response.reasoning_part.done",
                        sequence_number=-1,
                        item_id=current_item_id,
                        output_index=current_output_index,
                        content_index=current_content_index,
                        part=ResponseReasoningTextContent(
                            text=reason_content,
                            type="reasoning_text",
                        ),
                    )
                )
1838
1839
1840
1841
1842
1843
1844
1845
1846
1847
1848
1849
                reasoning_item = ResponseReasoningItem(
                    type="reasoning",
                    content=[
                        ResponseReasoningTextContent(
                            text=reason_content,
                            type="reasoning_text",
                        ),
                    ],
                    status="completed",
                    id=current_item_id,
                    summary=[],
                )
1850
                yield _increment_sequence_number_and_return(
1851
1852
1853
1854
1855
                    ResponseOutputItemDoneEvent(
                        type="response.output_item.done",
                        sequence_number=-1,
                        output_index=current_output_index,
                        item=reasoning_item,
1856
1857
                    )
                )
1858
            elif previous_delta_messages[-1].content:
1859
                final_content = "".join(
1860
                    pm.content for pm in previous_delta_messages if pm.content
1861
                )
1862
                yield _increment_sequence_number_and_return(
1863
                    ResponseTextDoneEvent(
1864
1865
1866
1867
1868
1869
1870
                        type="response.output_text.done",
                        sequence_number=-1,
                        output_index=current_output_index,
                        content_index=current_content_index,
                        text=final_content,
                        logprobs=[],
                        item_id=current_item_id,
1871
1872
                    )
                )
1873
1874
1875
1876
1877
                part = ResponseOutputText(
                    text=final_content,
                    type="output_text",
                    annotations=[],
                )
1878
                yield _increment_sequence_number_and_return(
1879
                    ResponseContentPartDoneEvent(
1880
1881
1882
1883
1884
1885
                        type="response.content_part.done",
                        sequence_number=-1,
                        item_id=current_item_id,
                        output_index=current_output_index,
                        content_index=current_content_index,
                        part=part,
1886
1887
                    )
                )
1888
1889
1890
1891
1892
1893
1894
1895
1896
1897
                item = ResponseOutputMessage(
                    type="message",
                    role="assistant",
                    content=[
                        part,
                    ],
                    status="completed",
                    id=current_item_id,
                    summary=[],
                )
1898
                yield _increment_sequence_number_and_return(
1899
1900
1901
1902
1903
                    ResponseOutputItemDoneEvent(
                        type="response.output_item.done",
                        sequence_number=-1,
                        output_index=current_output_index,
                        item=item,
1904
1905
                    )
                )
1906
1907
1908
1909
1910

    async def _process_harmony_streaming_events(
        self,
        request: ResponsesRequest,
        sampling_params: SamplingParams,
1911
        result_generator: AsyncIterator[ConversationContext | None],
1912
1913
        context: ConversationContext,
        model_name: str,
1914
        tokenizer: TokenizerLike,
1915
1916
        request_metadata: RequestResponseMetadata,
        created_time: int,
1917
        _increment_sequence_number_and_return: Callable[
1918
1919
            [StreamingResponsesResponse], StreamingResponsesResponse
        ],
1920
    ) -> AsyncGenerator[StreamingResponsesResponse, None]:
1921
        state = StreamingState()
1922

1923
1924
1925
        async for ctx in result_generator:
            assert isinstance(ctx, StreamingHarmonyContext)

1926
1927
1928
            # finish_reason='error' indicates a retryable error
            self._raise_if_error(ctx.finish_reason, request.request_id)

1929
1930
1931
            if ctx.is_expecting_start():
                if len(ctx.parser.messages) > 0:
                    previous_item = ctx.parser.messages[-1]
1932
                    for event in emit_previous_item_done_events(previous_item, state):
1933
1934
1935
1936
                        yield _increment_sequence_number_and_return(event)
                state.reset_for_new_item()

            # Stream the output of a harmony message
1937
            for event in emit_content_delta_events(ctx, state):
1938
1939
1940
                yield _increment_sequence_number_and_return(event)

            # Stream tool call outputs
1941
            for event in emit_tool_action_events(ctx, state, self.tool_server):
1942
                yield _increment_sequence_number_and_return(event)
1943

1944
1945
1946
1947
    async def responses_stream_generator(
        self,
        request: ResponsesRequest,
        sampling_params: SamplingParams,
1948
        result_generator: AsyncIterator[ConversationContext | None],
1949
1950
        context: ConversationContext,
        model_name: str,
1951
        tokenizer: TokenizerLike,
1952
        request_metadata: RequestResponseMetadata,
1953
        created_time: int | None = None,
1954
    ) -> AsyncGenerator[StreamingResponsesResponse, None]:
1955
1956
1957
1958
1959
        # TODO:
        # 1. Handle disconnect

        created_time = created_time or int(time.time())

1960
1961
        sequence_number = 0

1962
        def _increment_sequence_number_and_return(
1963
            event: StreamingResponsesResponse,
1964
        ) -> StreamingResponsesResponse:
1965
1966
            nonlocal sequence_number
            # Set sequence_number if the event has this attribute
1967
            if hasattr(event, "sequence_number"):
1968
1969
                event.sequence_number = sequence_number
            sequence_number += 1
1970
            return event
1971

1972
        async with AsyncExitStack() as exit_stack:
1973
            if self.use_harmony:
1974
1975
                # TODO: in streaming, we noticed this bug:
                # https://github.com/vllm-project/vllm/issues/25697
1976
                await self._initialize_tool_sessions(request, context, exit_stack)
Jiayi Yan's avatar
Jiayi Yan committed
1977
                processor = self._process_harmony_streaming_events
1978
            else:
Jiayi Yan's avatar
Jiayi Yan committed
1979
                processor = self._process_simple_streaming_events
1980
            # TODO Hanchen make sampling params to include the structural tag
1981
1982
1983
1984
1985
1986
1987
1988
1989

            initial_response = ResponsesResponse.from_request(
                request,
                sampling_params,
                model_name=model_name,
                created_time=created_time,
                output=[],
                status="in_progress",
                usage=None,
1990
            ).model_dump(mode="json", by_alias=True)
1991
            yield _increment_sequence_number_and_return(
1992
1993
1994
1995
                ResponseCreatedEvent(
                    type="response.created",
                    sequence_number=-1,
                    response=initial_response,
1996
1997
                )
            )
1998
            yield _increment_sequence_number_and_return(
1999
2000
2001
2002
                ResponseInProgressEvent(
                    type="response.in_progress",
                    sequence_number=-1,
                    response=initial_response,
2003
2004
                )
            )
2005

2006
            try:
Jiayi Yan's avatar
Jiayi Yan committed
2007
                async for event_data in processor(
2008
2009
2010
2011
2012
2013
2014
2015
2016
2017
2018
2019
2020
2021
2022
2023
2024
                    request,
                    sampling_params,
                    result_generator,
                    context,
                    model_name,
                    tokenizer,
                    request_metadata,
                    created_time,
                    _increment_sequence_number_and_return,
                ):
                    yield event_data
            except GenerationError as e:
                error_json = self._convert_generation_error_to_streaming_response(e)
                yield _increment_sequence_number_and_return(
                    TypeAdapter(StreamingResponsesResponse).validate_json(error_json)
                )
                return
2025
2026
2027
2028
2029
2030
2031
2032
2033
2034
2035
2036
2037
2038
2039
2040
2041

            async def empty_async_generator():
                # A hack to trick Python to think this is a generator but
                # in fact it immediately returns.
                if False:
                    yield

            final_response = await self.responses_full_generator(
                request,
                sampling_params,
                empty_async_generator(),
                context,
                model_name,
                tokenizer,
                request_metadata,
                created_time=created_time,
            )
2042
            yield _increment_sequence_number_and_return(
2043
                ResponseCompletedEvent(
2044
2045
                    type="response.completed",
                    sequence_number=-1,
2046
                    response=final_response,
2047
2048
                )
            )