serving_responses.py 87.3 KB
Newer Older
1
2
3
4
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project

import asyncio
5
import json
6
import time
7
import uuid
8
from collections import deque
9
from collections.abc import AsyncGenerator, AsyncIterator, Callable, Sequence
10
from contextlib import AsyncExitStack
11
from copy import copy
12
from http import HTTPStatus
13
from typing import Final
14
15
16

import jinja2
from fastapi import Request
17
18
19
20
21
22
from openai.types.responses import (
    ResponseCodeInterpreterCallCodeDeltaEvent,
    ResponseCodeInterpreterCallCodeDoneEvent,
    ResponseCodeInterpreterCallCompletedEvent,
    ResponseCodeInterpreterCallInProgressEvent,
    ResponseCodeInterpreterCallInterpretingEvent,
23
24
25
    ResponseCodeInterpreterToolCallParam,
    ResponseContentPartAddedEvent,
    ResponseContentPartDoneEvent,
26
27
    ResponseFunctionCallArgumentsDeltaEvent,
    ResponseFunctionCallArgumentsDoneEvent,
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
    ResponseFunctionToolCall,
    ResponseFunctionWebSearch,
    ResponseOutputItem,
    ResponseOutputItemAddedEvent,
    ResponseOutputItemDoneEvent,
    ResponseOutputMessage,
    ResponseOutputText,
    ResponseReasoningItem,
    ResponseReasoningTextDeltaEvent,
    ResponseReasoningTextDoneEvent,
    ResponseStatus,
    ResponseTextDeltaEvent,
    ResponseTextDoneEvent,
    ResponseWebSearchCallCompletedEvent,
    ResponseWebSearchCallInProgressEvent,
    ResponseWebSearchCallSearchingEvent,
    response_function_web_search,
    response_text_delta_event,
)
from openai.types.responses.response_output_text import Logprob, LogprobTopLogprob
48
from openai.types.responses.response_reasoning_item import (
49
50
    Content as ResponseReasoningTextContent,
)
51
from openai.types.responses.tool import Mcp, Tool
52
from openai_harmony import Message as OpenAIHarmonyMessage
53
from pydantic import TypeAdapter
54

55
from vllm import envs
56
from vllm.engine.protocol import EngineClient
57
58
59
60
61
62
63
from vllm.entrypoints.chat_utils import (
    ChatCompletionMessageParam,
    ChatTemplateContentFormatOption,
)
from vllm.entrypoints.context import (
    ConversationContext,
    HarmonyContext,
64
    ParsableContext,
65
66
67
    SimpleContext,
    StreamingHarmonyContext,
)
68
69
from vllm.entrypoints.logger import RequestLogger
from vllm.entrypoints.openai.parser.harmony_utils import (
70
    construct_harmony_previous_input_messages,
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
    get_developer_message,
    get_stop_tokens_for_assistant_actions,
    get_system_message,
    get_user_message,
    has_custom_tools,
    parse_output_message,
    parse_remaining_state,
    parse_response_input,
    render_for_completion,
)
from vllm.entrypoints.openai.protocol import (
    DeltaMessage,
    ErrorResponse,
    InputTokensDetails,
    OutputTokensDetails,
    RequestResponseMetadata,
    ResponseCompletedEvent,
    ResponseCreatedEvent,
    ResponseInProgressEvent,
90
    ResponseInputOutputMessage,
91
92
93
94
95
96
97
    ResponseReasoningPartAddedEvent,
    ResponseReasoningPartDoneEvent,
    ResponsesRequest,
    ResponsesResponse,
    ResponseUsage,
    StreamingResponsesResponse,
)
98
99
100
101
from vllm.entrypoints.openai.serving_engine import (
    GenerationError,
    OpenAIServing,
)
102
from vllm.entrypoints.openai.serving_models import OpenAIServingModels
103
from vllm.entrypoints.responses_utils import (
104
    construct_input_messages,
105
    construct_tool_dicts,
106
    extract_tool_types,
107
    make_response_output_items_from_parsable_context,
108
)
109
from vllm.entrypoints.tool_server import ToolServer
110
from vllm.inputs.data import TokensPrompt as EngineTokensPrompt
111
from vllm.logger import init_logger
112
113
from vllm.logprobs import Logprob as SampleLogprob
from vllm.logprobs import SampleLogprobs
114
from vllm.outputs import CompletionOutput
115
from vllm.sampling_params import SamplingParams, StructuredOutputsParams
116
from vllm.tokenizers import TokenizerLike
117
118
119
120
121
from vllm.utils import random_uuid

logger = init_logger(__name__)


122
123
124
125
126
127
128
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
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


161
162
163
164
165
166
class OpenAIServingResponses(OpenAIServing):
    def __init__(
        self,
        engine_client: EngineClient,
        models: OpenAIServingModels,
        *,
167
168
        request_logger: RequestLogger | None,
        chat_template: str | None,
169
170
171
172
        chat_template_content_format: ChatTemplateContentFormatOption,
        return_tokens_as_token_ids: bool = False,
        reasoning_parser: str = "",
        enable_auto_tools: bool = False,
173
174
        tool_parser: str | None = None,
        tool_server: ToolServer | None = None,
175
176
        enable_prompt_tokens_details: bool = False,
        enable_force_include_usage: bool = False,
177
        enable_log_outputs: bool = False,
178
        log_error_stack: bool = False,
179
180
181
182
183
184
    ) -> None:
        super().__init__(
            engine_client=engine_client,
            models=models,
            request_logger=request_logger,
            return_tokens_as_token_ids=return_tokens_as_token_ids,
185
            log_error_stack=log_error_stack,
186
187
188
189
        )

        self.chat_template = chat_template
        self.chat_template_content_format: Final = chat_template_content_format
190
        self.enable_log_outputs = enable_log_outputs
191

192
193
        self.reasoning_parser = self._get_reasoning_parser(
            reasoning_parser_name=reasoning_parser
194
        )
195
196
        self.enable_prompt_tokens_details = enable_prompt_tokens_details
        self.enable_force_include_usage = enable_force_include_usage
197
        self.default_sampling_params = self.model_config.get_diff_sampling_param()
198
199
200
        if self.default_sampling_params:
            source = self.model_config.generation_config
            source = "model" if source == "auto" else source
201
202
203
204
205
            logger.info(
                "Using default chat sampling params from %s: %s",
                source,
                self.default_sampling_params,
            )
206

207
208
209
210
211
        # 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.
212
        self.enable_store = envs.VLLM_ENABLE_RESPONSES_API_STORE
213
214
215
216
        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 "
217
218
                "the store."
            )
219

220
        self.use_harmony = self.model_config.hf_config.model_type == "gpt_oss"
221
        if self.use_harmony:
222
223
224
225
            logger.warning(
                "For gpt-oss, we ignore --enable-auto-tool-choice "
                "and always enable tool use."
            )
226
227
228
229
230
            # 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(
231
232
                get_stop_tokens_for_assistant_actions()
            )
233
        self.enable_auto_tools = enable_auto_tools
234
        # set up tool use
235
236
237
        self.tool_parser = self._get_tool_parser(
            tool_parser_name=tool_parser, enable_auto_tools=enable_auto_tools
        )
238
        # HACK(woosuk): This is a hack. We should use a better store.
239
240
        # FIXME: If enable_store=True, this may cause a memory leak since we
        # never remove responses from the store.
241
242
243
244
        self.response_store: dict[str, ResponsesResponse] = {}
        self.response_store_lock = asyncio.Lock()

        # HACK(woosuk): This is a hack. We should use a better store.
245
246
        # FIXME: If enable_store=True, this may cause a memory leak since we
        # never remove messages from the store.
247
248
        self.msg_store: dict[str, list[ChatCompletionMessageParam]] = {}

249
250
251
        # 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.
252
253
254
        self.event_store: dict[
            str, tuple[deque[StreamingResponsesResponse], asyncio.Event]
        ] = {}
255

256
257
        self.background_tasks: dict[str, asyncio.Task] = {}

258
259
        self.tool_server = tool_server

260
    def _validate_generator_input(
261
        self, engine_prompt: EngineTokensPrompt
262
    ) -> ErrorResponse | None:
263
264
265
266
267
268
        """Add validations to the input to the generator here."""
        if self.max_model_len <= len(engine_prompt["prompt_token_ids"]):
            error_message = (
                "The engine prompt length"
                f" {len(engine_prompt['prompt_token_ids'])} "
                f"exceeds the max_model_len {self.max_model_len}. "
269
270
                "Please reduce prompt."
            )
271
272
273
274
275
276
277
            return self.create_error_response(
                err_type="invalid_request_error",
                message=error_message,
                status_code=HTTPStatus.BAD_REQUEST,
            )
        return None

278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
    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,
            )
        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,
            )
299
300
301
302
303
304
305
        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,
            )
306
307
        return None

308
309
310
    async def create_responses(
        self,
        request: ResponsesRequest,
311
312
313
314
315
316
        raw_request: Request | None = None,
    ) -> (
        AsyncGenerator[StreamingResponsesResponse, None]
        | ResponsesResponse
        | ErrorResponse
    ):
317
318
319
320
        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
321
322
323
        maybe_validation_error = self._validate_create_responses_input(request)
        if maybe_validation_error is not None:
            return maybe_validation_error
324
325
326
327
328
329
330

        # 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

331
        if request.store and not self.enable_store:
332
333
334
335
336
337
338
            # 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
339

340
341
342
343
344
345
346
347
348
349
350
        # 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

        try:
351
            lora_request = self._maybe_get_adapters(request)
352
            model_name = self.models.model_name(lora_request)
353
            tokenizer = await self.engine_client.get_tokenizer()
354

355
356
            if self.use_harmony:
                messages, request_prompts, engine_prompts = (
357
358
                    self._make_request_with_harmony(request, prev_response)
                )
359
            else:
360
361
362
                messages, request_prompts, engine_prompts = await self._make_request(
                    request, prev_response, tokenizer
                )
363

364
365
366
367
368
369
370
        except (
            ValueError,
            TypeError,
            RuntimeError,
            jinja2.TemplateError,
            NotImplementedError,
        ) as e:
371
372
373
            logger.exception("Error in preprocessing prompt inputs")
            return self.create_error_response(f"{e} {e.__cause__}")

374
        request_metadata = RequestResponseMetadata(request_id=request.request_id)
375
376
377
378
        if raw_request:
            raw_request.state.request_metadata = request_metadata

        # Schedule the request and get the result generator.
379
        generators: list[AsyncGenerator[ConversationContext, None]] = []
380
381

        builtin_tool_list: list[str] = []
382
        if self.tool_server is not None:
383
384
385
386
            if self.tool_server.has_tool("browser"):
                builtin_tool_list.append("browser")
            if self.tool_server.has_tool("python"):
                builtin_tool_list.append("python")
387
388
            if self.tool_server.has_tool("container"):
                builtin_tool_list.append("container")
389

390
391
392
393
394
395
396
        if self.tool_server is not None:
            available_tools = builtin_tool_list
        else:
            assert len(builtin_tool_list) == 0
            available_tools = []
        try:
            for i, engine_prompt in enumerate(engine_prompts):
397
398
399
400
                maybe_error = self._validate_generator_input(engine_prompt)
                if maybe_error is not None:
                    return maybe_error

401
                default_max_tokens = self.max_model_len - len(
402
403
                    engine_prompt["prompt_token_ids"]
                )
404

405
                sampling_params = request.to_sampling_params(
406
407
                    default_max_tokens, self.default_sampling_params
                )
408

409
410
411
412
413
                trace_headers = (
                    None
                    if raw_request is None
                    else await self._get_trace_headers(raw_request.headers)
                )
414
415
416
417

                context: ConversationContext
                if self.use_harmony:
                    if request.stream:
418
                        context = StreamingHarmonyContext(messages, available_tools)
419
420
421
                    else:
                        context = HarmonyContext(messages, available_tools)
                else:
422
423
424
425
426
427
428
429
                    if envs.VLLM_USE_EXPERIMENTAL_PARSER_CONTEXT:
                        # This is an feature in development for parsing
                        # tokens during generation instead of at the end
                        context = ParsableContext(
                            response_messages=messages,
                            tokenizer=tokenizer,
                            reasoning_parser_cls=self.reasoning_parser,
                            request=request,
430
431
432
433
                            tool_parser_cls=self.tool_parser,
                            available_tools=available_tools,
                            chat_template=self.chat_template,
                            chat_template_content_format=self.chat_template_content_format,
434
435
436
                        )
                    else:
                        context = SimpleContext()
437
438
439
440
441
442
443
444
445
446
447
448
449

                if self.reasoning_parser is not None:
                    reasoning_parser = self.reasoning_parser(tokenizer)
                    if sampling_params.structured_outputs is None:
                        sampling_params.structured_outputs = StructuredOutputsParams()
                    struct_out = sampling_params.structured_outputs
                    if struct_out.all_non_structural_tag_constraints_none():
                        sampling_params.structured_outputs.structural_tag = (
                            reasoning_parser.prepare_structured_tag(
                                sampling_params.structured_outputs.structural_tag,
                                self.tool_server,
                            )
                        )
450
451
452
453
454
455
456
457
458
                generator = self._generate_with_builtin_tools(
                    request_id=request.request_id,
                    request_prompt=request_prompts[i],
                    engine_prompt=engine_prompt,
                    sampling_params=sampling_params,
                    context=context,
                    lora_request=lora_request,
                    priority=request.priority,
                    trace_headers=trace_headers,
459
                )
460
461
462
463
                generators.append(generator)
        except ValueError as e:
            # TODO: Use a vllm-specific Validation Error
            return self.create_error_response(str(e))
464

465
        assert len(generators) == 1
466
        (result_generator,) = generators
467
468
469
470

        # Store the input messages.
        if request.store:
            self.msg_store[request.request_id] = messages
471

472
473
474
475
476
477
478
479
480
481
482
483
484
        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
485

486
            # Run the request in the background.
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
            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}",
                )
515

516
517
518
519
            # For cleanup.
            response_id = response.id
            self.background_tasks[response_id] = task
            task.add_done_callback(
520
521
                lambda _: self.background_tasks.pop(response_id, None)
            )
522
523

            if request.stream:
524
                return self.responses_background_stream_generator(request.request_id)
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
            return response

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

        try:
            return await self.responses_full_generator(
                request,
                sampling_params,
                result_generator,
                context,
                model_name,
                tokenizer,
                request_metadata,
            )
548
549
        except GenerationError as e:
            return self._convert_generation_error_to_response(e)
550
551
        except Exception as e:
            return self.create_error_response(str(e))
552

553
554
555
    async def _make_request(
        self,
        request: ResponsesRequest,
556
        prev_response: ResponsesResponse | None,
557
        tokenizer: TokenizerLike,
558
    ):
559
        tool_dicts = construct_tool_dicts(request.tools, request.tool_choice)
560
        # Construct the input messages.
561
562
563
564
565
566
        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,
        )
567
568
569
570
        _, request_prompts, engine_prompts = await self._preprocess_chat(
            request,
            tokenizer,
            messages,
571
572
            tool_dicts=tool_dicts,
            tool_parser=self.tool_parser,
573
574
575
576
577
578
579
580
            chat_template=self.chat_template,
            chat_template_content_format=self.chat_template_content_format,
        )
        return messages, request_prompts, engine_prompts

    def _make_request_with_harmony(
        self,
        request: ResponsesRequest,
581
        prev_response: ResponsesResponse | None,
582
583
584
    ):
        if request.tool_choice != "auto":
            raise NotImplementedError(
585
586
587
                "Only 'auto' tool_choice is supported in response API with Harmony"
            )
        messages = self._construct_input_messages_with_harmony(request, prev_response)
588
589
        prompt_token_ids = render_for_completion(messages)
        engine_prompt = EngineTokensPrompt(prompt_token_ids=prompt_token_ids)
590
591
592
593
594

        # Add cache_salt if provided in the request
        if request.cache_salt is not None:
            engine_prompt["cache_salt"] = request.cache_salt

595
596
        return messages, [prompt_token_ids], [engine_prompt]

597
598
599
600
601
602
    async def _initialize_tool_sessions(
        self,
        request: ResponsesRequest,
        context: ConversationContext,
        exit_stack: AsyncExitStack,
    ):
603
604
605
606
        # we should only initialize the tool session if the request needs tools
        if len(request.tools) == 0:
            return
        mcp_tools = {
607
            tool.server_label: tool for tool in request.tools if tool.type == "mcp"
608
        }
609
610
611
        await context.init_tool_sessions(
            self.tool_server, exit_stack, request.request_id, mcp_tools
        )
612

613
614
615
616
    async def responses_full_generator(
        self,
        request: ResponsesRequest,
        sampling_params: SamplingParams,
617
        result_generator: AsyncIterator[ConversationContext],
618
        context: ConversationContext,
619
        model_name: str,
620
        tokenizer: TokenizerLike,
621
        request_metadata: RequestResponseMetadata,
622
623
        created_time: int | None = None,
    ) -> ErrorResponse | ResponsesResponse:
624
625
626
        if created_time is None:
            created_time = int(time.time())

627
628
        async with AsyncExitStack() as exit_stack:
            try:
629
                await self._initialize_tool_sessions(request, context, exit_stack)
630
631
632
633
634
635
636
                async for _ in result_generator:
                    pass
            except asyncio.CancelledError:
                return self.create_error_response("Client disconnected")
            except ValueError as e:
                # TODO: Use a vllm-specific Validation Error
                return self.create_error_response(str(e))
637

638
639
640
641
642
        # NOTE: Implementation of stauts is still WIP, but for now
        # we guarantee that if the status is not "completed", it is accurate.
        # "completed" is implemented as the "catch-all" for now.
        status: ResponseStatus = "completed"

643
644
        input_messages: ResponseInputOutputMessage | None = None
        output_messages: ResponseInputOutputMessage | None = None
645
646
647
        if self.use_harmony:
            assert isinstance(context, HarmonyContext)
            output = self._make_response_output_items_with_harmony(context)
648
            if request.enable_response_messages:
649
650
                input_messages = context.messages[: context.num_init_messages]
                output_messages = context.messages[context.num_init_messages :]
651
            num_tool_output_tokens = context.num_tool_output_tokens
652
653
654
655
656
            if len(output) > 0:
                if context.finish_reason == "length":
                    status = "incomplete"
                elif context.finish_reason == "abort":
                    status = "cancelled"
657
658
                else:
                    self._raise_if_error(context.finish_reason, request.request_id)
659
660
            else:
                status = "incomplete"
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
        elif isinstance(context, ParsableContext):
            response_messages = context.parser.response_messages[
                context.parser.num_init_messages :
            ]
            output = make_response_output_items_from_parsable_context(response_messages)

            # TODO: context for non-gptoss models doesn't use messages
            # so we can't get them out yet
            if request.enable_response_messages:
                raise NotImplementedError(
                    "enable_response_messages is currently only supported for gpt-oss"
                )

            # TODO: Calculate usage.
            # assert final_res.prompt_token_ids is not None
            num_tool_output_tokens = 0
677
        else:
678
679
680
681
682
683
            assert isinstance(context, SimpleContext)
            final_res = context.last_output
            assert final_res is not None
            assert len(final_res.outputs) == 1
            final_output = final_res.outputs[0]

684
685
686
            # finish_reason='error' indicates retryable internal error
            self._raise_if_error(final_output.finish_reason, request.request_id)

687
            output = self._make_response_output_items(request, final_output, tokenizer)
688

689
            if request.enable_response_messages:
690
691
692
                input_messages = context.input_messages
                output_messages = context.output_messages

693
694
            # Calculate usage.
            assert final_res.prompt_token_ids is not None
695
696
            num_tool_output_tokens = 0

697
        assert isinstance(context, (SimpleContext, HarmonyContext, ParsableContext))
698
699
700
701
        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
702
703
704
705

        usage = ResponseUsage(
            input_tokens=num_prompt_tokens,
            output_tokens=num_generated_tokens,
706
            total_tokens=num_prompt_tokens + num_generated_tokens,
707
708
709
710
711
712
713
714
715
            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
                ],
            ),
716
            output_tokens_details=OutputTokensDetails(
717
                reasoning_tokens=num_reasoning_tokens,
718
                tool_output_tokens=num_tool_output_tokens,
719
720
721
722
723
724
                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
                ],
725
            ),
726
727
728
729
        )
        response = ResponsesResponse.from_request(
            request,
            sampling_params,
730
731
            input_messages=input_messages,
            output_messages=output_messages,
732
733
734
            model_name=model_name,
            created_time=created_time,
            output=output,
735
            status=status,
736
737
738
739
740
741
742
            usage=usage,
        )

        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.
743
                if stored_response is None or stored_response.status != "cancelled":
744
745
746
                    self.response_store[response.id] = response
        return response

747
748
749
750
    def _topk_logprobs(
        self,
        logprobs: dict[int, SampleLogprob],
        top_logprobs: int,
751
        tokenizer: TokenizerLike,
752
    ) -> list[LogprobTopLogprob]:
753
754
755
756
757
        """Returns the top-k logprobs from the logprobs dictionary."""
        out = []
        for i, (token_id, _logprob) in enumerate(logprobs.items()):
            if i >= top_logprobs:
                break
758
759
760
761
762
            text = (
                _logprob.decoded_token
                if _logprob.decoded_token is not None
                else tokenizer.decode([token_id])
            )
763
764
765
766
767
            out.append(
                LogprobTopLogprob(
                    token=text,
                    logprob=max(_logprob.logprob, -9999.0),
                    bytes=list(text.encode("utf-8", errors="replace")),
768
769
                )
            )
770
771
772
        return out

    def _create_response_logprobs(
773
774
        self,
        token_ids: Sequence[int],
775
        logprobs: SampleLogprobs | None,
776
        tokenizer: TokenizerLike,
777
        top_logprobs: int | None = None,
778
    ) -> list[Logprob]:
779
780
        assert logprobs is not None, "logprobs must be provided"
        assert len(token_ids) == len(logprobs), (
781
782
            "token_ids and logprobs.token_ids must have the same length"
        )
783
784
785
786
        out = []
        for i, token_id in enumerate(token_ids):
            logprob = logprobs[i]
            token_logprob = logprob[token_id]
787
788
789
790
791
            text = (
                token_logprob.decoded_token
                if token_logprob.decoded_token is not None
                else tokenizer.decode([token_id])
            )
792
793
794
795
796
            out.append(
                Logprob(
                    token=text,
                    logprob=max(token_logprob.logprob, -9999.0),
                    bytes=list(text.encode("utf-8", errors="replace")),
797
798
799
800
801
802
803
                    top_logprobs=(
                        self._topk_logprobs(
                            logprob, top_logprobs=top_logprobs, tokenizer=tokenizer
                        )
                        if top_logprobs
                        else []
                    ),
804
805
                )
            )
806
807
        return out

808
809
810
    def _create_stream_response_logprobs(
        self,
        token_ids: Sequence[int],
811
        logprobs: SampleLogprobs | None,
812
        tokenizer: TokenizerLike,
813
        top_logprobs: int | None = None,
814
    ) -> list[response_text_delta_event.Logprob]:
815
816
817
818
819
820
        lgs = self._create_response_logprobs(
            token_ids=token_ids,
            logprobs=logprobs,
            tokenizer=tokenizer,
            top_logprobs=top_logprobs,
        )
821
822
823
824
825
826
        return [
            response_text_delta_event.Logprob(
                token=lg.token,
                logprob=lg.logprob,
                top_logprobs=[
                    response_text_delta_event.LogprobTopLogprob(
827
828
                        token=tl.token, logprob=tl.logprob
                    )
829
                    for tl in lg.top_logprobs
830
831
832
                ],
            )
            for lg in lgs
833
834
        ]

835
836
837
838
    def _make_response_output_items(
        self,
        request: ResponsesRequest,
        final_output: CompletionOutput,
839
        tokenizer: TokenizerLike,
840
841
842
843
844
845
846
847
    ) -> list[ResponseOutputItem]:
        if self.reasoning_parser:
            try:
                reasoning_parser = self.reasoning_parser(tokenizer)
            except RuntimeError as e:
                logger.exception("Error in reasoning parser creation.")
                raise e

848
            reasoning, content = reasoning_parser.extract_reasoning(
849
850
                final_output.text, request=request
            )
851
        else:
852
            reasoning = None
853
854
            content = final_output.text

855
856
857
858
859
        # Log complete response if output logging is enabled
        if self.enable_log_outputs and self.request_logger:
            output_text = ""
            if content:
                output_text = content
860
861
            elif reasoning:
                output_text = f"[reasoning: {reasoning}]"
862
863
864
865
866
867
868
869
870
871
872

            if output_text:
                self.request_logger.log_outputs(
                    request_id=request.request_id,
                    outputs=output_text,
                    output_token_ids=final_output.token_ids,
                    finish_reason=final_output.finish_reason,
                    is_streaming=False,
                    delta=False,
                )

873
874
        reasoning_item = None
        message_item = None
875
        if reasoning:
876
877
878
879
880
            reasoning_item = ResponseReasoningItem(
                id=f"rs_{random_uuid()}",
                summary=[],
                type="reasoning",
                content=[
881
                    ResponseReasoningTextContent(text=reasoning, type="reasoning_text")
882
883
884
                ],
                status=None,  # NOTE: Only the last output item has status.
            )
885
886
887
888
889
890
891
        tool_calls, content = self._parse_tool_calls_from_content(
            request=request,
            tokenizer=tokenizer,
            content=content,
            enable_auto_tools=self.enable_auto_tools,
            tool_parser_cls=self.tool_parser,
        )
892
893
894
895
896
        if content:
            output_text = ResponseOutputText(
                text=content,
                annotations=[],  # TODO
                type="output_text",
897
898
899
900
901
902
903
904
905
906
                logprobs=(
                    self._create_response_logprobs(
                        token_ids=final_output.token_ids,
                        logprobs=final_output.logprobs,
                        tokenizer=tokenizer,
                        top_logprobs=request.top_logprobs,
                    )
                    if request.is_include_output_logprobs()
                    else None
                ),
907
            )
908
            message_item = ResponseOutputMessage(
909
910
911
912
913
914
                id=f"msg_{random_uuid()}",
                content=[output_text],
                role="assistant",
                status="completed",
                type="message",
            )
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
        outputs = []

        if reasoning_item:
            outputs.append(reasoning_item)
        if message_item:
            outputs.append(message_item)
        if tool_calls:
            tool_call_items = [
                ResponseFunctionToolCall(
                    id=f"fc_{random_uuid()}",
                    call_id=f"call_{random_uuid()}",
                    type="function_call",
                    status="completed",
                    name=tool_call.name,
                    arguments=tool_call.arguments,
                )
                for tool_call in tool_calls
            ]
            outputs.extend(tool_call_items)
        return outputs
935
936
937
938
939

    def _make_response_output_items_with_harmony(
        self,
        context: HarmonyContext,
    ) -> list[ResponseOutputItem]:
940
        output_items: list[ResponseOutputItem] = []
941
942
943
944
945
946
947
948
949
        num_init_messages = context.num_init_messages
        for msg in context.messages[num_init_messages:]:
            output_items.extend(parse_output_message(msg))
        # Handle the generation stopped in the middle (if any).
        last_items = parse_remaining_state(context.parser)
        if last_items:
            output_items.extend(last_items)
        return output_items

950
    def _construct_harmony_system_input_message(
951
        self, request: ResponsesRequest, with_custom_tools: bool, tool_types: set[str]
952
953
    ) -> OpenAIHarmonyMessage:
        reasoning_effort = request.reasoning.effort if request.reasoning else None
954
955
956
957
958
959
960
961
962
963
964

        # 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
965
966
            and self.tool_server is not None
            and self.tool_server.has_tool("browser")
967
            else None
968
        )
969
970
971
972
973
        python_description = (
            self.tool_server.get_tool_description(
                "python", allowed_tools_map.get("code_interpreter")
            )
            if "code_interpreter" in tool_types
974
975
            and self.tool_server is not None
            and self.tool_server.has_tool("python")
976
            else None
977
        )
978
979
980
981
982
        container_description = (
            self.tool_server.get_tool_description(
                "container", allowed_tools_map.get("container")
            )
            if "container" in tool_types
983
984
            and self.tool_server is not None
            and self.tool_server.has_tool("container")
985
            else None
986
        )
987

988
989
        sys_msg = get_system_message(
            reasoning_effort=reasoning_effort,
990
991
992
            browser_description=browser_description,
            python_description=python_description,
            container_description=container_description,
993
994
995
996
997
            instructions=request.instructions,
            with_custom_tools=with_custom_tools,
        )
        return sys_msg

998
999
1000
    def _construct_input_messages_with_harmony(
        self,
        request: ResponsesRequest,
1001
        prev_response: ResponsesResponse | None,
1002
1003
1004
1005
    ) -> list[OpenAIHarmonyMessage]:
        messages: list[OpenAIHarmonyMessage] = []
        if prev_response is None:
            # New conversation.
1006
            tool_types = extract_tool_types(request.tools)
1007
            with_custom_tools = has_custom_tools(tool_types)
1008
1009
1010

            sys_msg = self._construct_harmony_system_input_message(
                request, with_custom_tools, tool_types
1011
1012
            )
            messages.append(sys_msg)
1013
1014
            if with_custom_tools:
                dev_msg = get_developer_message(
1015
1016
                    instructions=request.instructions, tools=request.tools
                )
1017
                messages.append(dev_msg)
1018
1019
            messages += construct_harmony_previous_input_messages(request)

1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
        else:
            # Continue the previous conversation.
            # FIXME(woosuk): Currently, request params like reasoning and
            # instructions are ignored.
            prev_msgs = self.msg_store[prev_response.id]
            # Remove the previous chain-of-thoughts if there is a new "final"
            # message. Note that this also removes these messages from the
            # msg_store.
            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
1039
1040
                    recent_turn_msgs = prev_msgs[prev_final_msg_idx + 1 :]
                    del prev_msgs[prev_final_msg_idx + 1 :]
1041
1042
1043
1044
1045
1046
                    for msg in recent_turn_msgs:
                        assert isinstance(msg, OpenAIHarmonyMessage)
                        if msg.channel != "analysis":
                            prev_msgs.append(msg)
            messages.extend(prev_msgs)
        # Append the new input.
co63oc's avatar
co63oc committed
1047
        # Responses API supports simple text inputs without chat format.
1048
1049
1050
1051
1052
1053
1054
1055
        if isinstance(request.input, str):
            messages.append(get_user_message(request.input))
        else:
            if prev_response is not None:
                prev_outputs = copy(prev_response.output)
            else:
                prev_outputs = []
            for response_msg in request.input:
1056
                messages.append(parse_response_input(response_msg, prev_outputs))
1057
                # User passes in a tool call request and its output. We need
1058
1059
1060
1061
1062
1063
1064
                # to add the tool call request to prev_outputs so that the
                # parse_response_input can find the tool call request when
                # parsing the tool call output.
                if isinstance(response_msg, ResponseFunctionToolCall):
                    prev_outputs.append(response_msg)
        return messages

1065
1066
1067
1068
1069
1070
    async def _run_background_request_stream(
        self,
        request: ResponsesRequest,
        *args,
        **kwargs,
    ):
1071
        event_deque: deque[StreamingResponsesResponse] = deque()
1072
1073
1074
1075
        new_event_signal = asyncio.Event()
        self.event_store[request.request_id] = (event_deque, new_event_signal)
        response = None
        try:
1076
            generator = self.responses_stream_generator(request, *args, **kwargs)
1077
1078
1079
            async for event in generator:
                event_deque.append(event)
                new_event_signal.set()  # Signal new event available
1080
1081
        except GenerationError as e:
            response = self._convert_generation_error_to_response(e)
1082
        except Exception as e:
1083
            logger.exception("Background request failed for %s", request.request_id)
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
            response = self.create_error_response(str(e))
        finally:
            new_event_signal.set()

        if response is not None and 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"

1097
1098
1099
1100
1101
1102
1103
    async def _run_background_request(
        self,
        request: ResponsesRequest,
        *args,
        **kwargs,
    ):
        try:
1104
            response = await self.responses_full_generator(request, *args, **kwargs)
1105
1106
        except GenerationError as e:
            response = self._convert_generation_error_to_response(e)
1107
        except Exception as e:
1108
            logger.exception("Background request failed for %s", request.request_id)
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
            response = self.create_error_response(str(e))

        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"

1120
1121
1122
    async def responses_background_stream_generator(
        self,
        response_id: str,
1123
        starting_after: int | None = None,
1124
    ) -> AsyncGenerator[StreamingResponsesResponse, None]:
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
        if response_id not in self.event_store:
            raise ValueError(f"Unknown response_id: {response_id}")

        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
1139
                if getattr(event, "type", "unknown") == "response.completed":
1140
                    return
1141
1142
1143
1144
                current_index += 1

            await new_event_signal.wait()

1145
1146
1147
    async def retrieve_responses(
        self,
        response_id: str,
1148
1149
1150
1151
1152
1153
1154
        starting_after: int | None,
        stream: bool | None,
    ) -> (
        ErrorResponse
        | ResponsesResponse
        | AsyncGenerator[StreamingResponsesResponse, None]
    ):
1155
1156
1157
1158
1159
        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)
1160
1161
1162
1163
1164
1165

        if stream:
            return self.responses_background_stream_generator(
                response_id,
                starting_after,
            )
1166
1167
1168
1169
1170
        return response

    async def cancel_responses(
        self,
        response_id: str,
1171
    ) -> ErrorResponse | ResponsesResponse:
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
        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.",
                )

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

        # Abort the request.
1188
        if task := self.background_tasks.get(response_id):
1189
1190
1191
1192
            task.cancel()
            try:
                await task
            except asyncio.CancelledError:
1193
                logger.exception("Background task for %s was cancelled", response_id)
1194
1195
1196
1197
1198
1199
1200
1201
        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,
        )
1202
1203
1204
1205

    def _make_store_not_supported_error(self) -> ErrorResponse:
        return self.create_error_response(
            err_type="invalid_request_error",
1206
1207
1208
1209
1210
1211
            message=(
                "`store=True` (default) is not supported. Please set "
                "`store=False` in Responses API or set "
                "`VLLM_ENABLE_RESPONSES_API_STORE=1` in the env var when "
                "starting the vLLM server."
            ),
1212
1213
            status_code=HTTPStatus.BAD_REQUEST,
        )
1214

1215
    async def _process_simple_streaming_events(
1216
1217
1218
        self,
        request: ResponsesRequest,
        sampling_params: SamplingParams,
1219
        result_generator: AsyncIterator[ConversationContext | None],
1220
1221
        context: ConversationContext,
        model_name: str,
1222
        tokenizer: TokenizerLike,
1223
        request_metadata: RequestResponseMetadata,
1224
        created_time: int,
1225
        _increment_sequence_number_and_return: Callable[
1226
1227
            [StreamingResponsesResponse], StreamingResponsesResponse
        ],
1228
    ) -> AsyncGenerator[StreamingResponsesResponse, None]:
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
        current_content_index = 0
        current_output_index = 0
        current_item_id = ""
        reasoning_parser = None
        if self.reasoning_parser:
            reasoning_parser = self.reasoning_parser(tokenizer)
        previous_text = ""
        previous_token_ids: list[int] = []
        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]
1245
1246
                # finish_reason='error' indicates a retryable error
                self._raise_if_error(output.finish_reason, request.request_id)
1247
                if reasoning_parser:
1248
1249
1250
1251
1252
1253
1254
                    delta_message = reasoning_parser.extract_reasoning_streaming(
                        previous_text=previous_text,
                        current_text=previous_text + output.text,
                        delta_text=output.text,
                        previous_token_ids=previous_token_ids,
                        current_token_ids=previous_token_ids + output.token_ids,
                        delta_token_ids=output.token_ids,
1255
1256
                    )
                else:
1257
1258
1259
                    delta_message = DeltaMessage(
                        content=output.text,
                    )
1260
1261
1262
1263
1264
1265
                previous_text += output.text
                previous_token_ids += output.token_ids
                if not delta_message:
                    continue
                if not first_delta_sent:
                    current_item_id = str(uuid.uuid4())
1266
                    if delta_message.reasoning:
1267
                        yield _increment_sequence_number_and_return(
1268
1269
1270
1271
                            ResponseOutputItemAddedEvent(
                                type="response.output_item.added",
                                sequence_number=-1,
                                output_index=current_output_index,
1272
                                item=ResponseReasoningItem(
1273
1274
1275
1276
1277
                                    type="reasoning",
                                    id=current_item_id,
                                    summary=[],
                                    status="in_progress",
                                ),
1278
1279
                            )
                        )
1280
                    else:
1281
                        yield _increment_sequence_number_and_return(
1282
1283
1284
1285
                            ResponseOutputItemAddedEvent(
                                type="response.output_item.added",
                                sequence_number=-1,
                                output_index=current_output_index,
1286
                                item=ResponseOutputMessage(
1287
1288
1289
1290
1291
1292
                                    id=current_item_id,
                                    type="message",
                                    role="assistant",
                                    content=[],
                                    status="in_progress",
                                ),
1293
1294
                            )
                        )
1295
                    yield _increment_sequence_number_and_return(
1296
                        ResponseContentPartAddedEvent(
1297
1298
1299
1300
1301
                            type="response.content_part.added",
                            sequence_number=-1,
                            output_index=current_output_index,
                            item_id=current_item_id,
                            content_index=current_content_index,
1302
                            part=ResponseOutputText(
1303
1304
1305
1306
1307
                                type="output_text",
                                text="",
                                annotations=[],
                                logprobs=[],
                            ),
1308
1309
                        )
                    )
1310
1311
1312
1313
1314
1315
                    current_content_index += 1
                    first_delta_sent = True
                # todo(kebe7jun) tool call support

                # check delta message and previous delta message are
                # same as content or reasoning content
1316
1317
                if (
                    previous_delta_messages
1318
                    and previous_delta_messages[-1].reasoning is not None
1319
1320
                    and delta_message.content is not None
                ):
1321
1322
                    # from reasoning to normal content, send done
                    # event for reasoning
1323
                    reason_content = "".join(
1324
                        pm.reasoning
1325
                        for pm in previous_delta_messages
1326
                        if pm.reasoning is not None
1327
                    )
1328
                    yield _increment_sequence_number_and_return(
1329
1330
1331
1332
1333
1334
1335
                        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,
1336
1337
                        )
                    )
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
                    current_content_index = 0
                    reasoning_item = ResponseReasoningItem(
                        type="reasoning",
                        content=[
                            ResponseReasoningTextContent(
                                text=reason_content,
                                type="reasoning_text",
                            ),
                        ],
                        status="completed",
                        id=current_item_id,
                        summary=[],
                    )
1351
                    yield _increment_sequence_number_and_return(
1352
1353
1354
1355
1356
                        ResponseOutputItemDoneEvent(
                            type="response.output_item.done",
                            sequence_number=-1,
                            output_index=current_output_index,
                            item=reasoning_item,
1357
1358
                        )
                    )
1359
                    yield _increment_sequence_number_and_return(
1360
                        ResponseOutputItemAddedEvent(
1361
1362
1363
                            type="response.output_item.added",
                            sequence_number=-1,
                            output_index=current_output_index,
1364
                            item=ResponseOutputMessage(
1365
1366
1367
1368
1369
1370
                                id=current_item_id,
                                type="message",
                                role="assistant",
                                content=[],
                                status="in_progress",
                            ),
1371
1372
                        )
                    )
1373
1374
                    current_output_index += 1
                    current_item_id = str(uuid.uuid4())
1375
                    yield _increment_sequence_number_and_return(
1376
                        ResponseContentPartAddedEvent(
1377
1378
1379
1380
1381
                            type="response.content_part.added",
                            sequence_number=-1,
                            output_index=current_output_index,
                            item_id=current_item_id,
                            content_index=current_content_index,
1382
                            part=ResponseOutputText(
1383
1384
1385
1386
1387
                                type="output_text",
                                text="",
                                annotations=[],
                                logprobs=[],
                            ),
1388
1389
                        )
                    )
1390
1391
1392
                    current_content_index += 1
                    # reset previous delta messages
                    previous_delta_messages = []
1393

1394
                if delta_message.reasoning is not None:
1395
                    yield _increment_sequence_number_and_return(
1396
1397
1398
1399
1400
1401
                        ResponseReasoningTextDeltaEvent(
                            type="response.reasoning_text.delta",
                            sequence_number=-1,
                            content_index=current_content_index,
                            output_index=current_output_index,
                            item_id=current_item_id,
1402
                            delta=delta_message.reasoning,
1403
1404
                        )
                    )
1405
                elif delta_message.content is not None:
1406
                    yield _increment_sequence_number_and_return(
1407
                        ResponseTextDeltaEvent(
1408
1409
1410
1411
1412
1413
                            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,
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
                            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 []
                            ),
1424
1425
                        )
                    )
1426
1427
1428
1429
                current_content_index += 1

                previous_delta_messages.append(delta_message)
        if previous_delta_messages:
1430
            if previous_delta_messages[-1].reasoning is not None:
1431
                reason_content = "".join(
1432
                    pm.reasoning
1433
                    for pm in previous_delta_messages
1434
                    if pm.reasoning is not None
1435
                )
1436
                yield _increment_sequence_number_and_return(
1437
1438
1439
1440
1441
1442
1443
                    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,
1444
1445
                    )
                )
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
                current_content_index += 1
                reasoning_item = ResponseReasoningItem(
                    type="reasoning",
                    content=[
                        ResponseReasoningTextContent(
                            text=reason_content,
                            type="reasoning_text",
                        ),
                    ],
                    status="completed",
                    id=current_item_id,
                    summary=[],
                )
1459
                yield _increment_sequence_number_and_return(
1460
1461
1462
1463
1464
                    ResponseOutputItemDoneEvent(
                        type="response.output_item.done",
                        sequence_number=-1,
                        output_index=current_output_index,
                        item=reasoning_item,
1465
1466
                    )
                )
1467
            elif previous_delta_messages[-1].content is not None:
1468
1469
1470
1471
1472
                final_content = "".join(
                    pm.content
                    for pm in previous_delta_messages
                    if pm.content is not None
                )
1473
                yield _increment_sequence_number_and_return(
1474
                    ResponseTextDoneEvent(
1475
1476
1477
1478
1479
1480
1481
                        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,
1482
1483
                    )
                )
1484
1485
1486
1487
1488
1489
                current_content_index += 1
                part = ResponseOutputText(
                    text=final_content,
                    type="output_text",
                    annotations=[],
                )
1490
                yield _increment_sequence_number_and_return(
1491
                    ResponseContentPartDoneEvent(
1492
1493
1494
1495
1496
1497
                        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,
1498
1499
                    )
                )
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
                current_content_index += 1
                item = ResponseOutputMessage(
                    type="message",
                    role="assistant",
                    content=[
                        part,
                    ],
                    status="completed",
                    id=current_item_id,
                    summary=[],
                )
1511
                yield _increment_sequence_number_and_return(
1512
1513
1514
1515
1516
                    ResponseOutputItemDoneEvent(
                        type="response.output_item.done",
                        sequence_number=-1,
                        output_index=current_output_index,
                        item=item,
1517
1518
                    )
                )
1519
1520
1521
1522
1523

    async def _process_harmony_streaming_events(
        self,
        request: ResponsesRequest,
        sampling_params: SamplingParams,
1524
        result_generator: AsyncIterator[ConversationContext | None],
1525
1526
        context: ConversationContext,
        model_name: str,
1527
        tokenizer: TokenizerLike,
1528
1529
        request_metadata: RequestResponseMetadata,
        created_time: int,
1530
        _increment_sequence_number_and_return: Callable[
1531
1532
            [StreamingResponsesResponse], StreamingResponsesResponse
        ],
1533
    ) -> AsyncGenerator[StreamingResponsesResponse, None]:
1534
        current_content_index = -1
1535
        current_output_index = 0
1536
        current_item_id: str = ""
1537
        sent_output_item_added = False
1538
        is_first_function_call_delta = False
1539
1540
1541
        async for ctx in result_generator:
            assert isinstance(ctx, StreamingHarmonyContext)

1542
1543
1544
            # finish_reason='error' indicates a retryable error
            self._raise_if_error(ctx.finish_reason, request.request_id)

1545
1546
1547
            if ctx.is_expecting_start():
                current_output_index += 1
                sent_output_item_added = False
1548
                is_first_function_call_delta = False
1549
1550
1551
                if len(ctx.parser.messages) > 0:
                    previous_item = ctx.parser.messages[-1]
                    if previous_item.recipient is not None:
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
                        # Deal with tool call
                        if previous_item.recipient.startswith("functions."):
                            function_name = previous_item.recipient[len("functions.") :]
                            yield _increment_sequence_number_and_return(
                                ResponseFunctionCallArgumentsDoneEvent(
                                    type="response.function_call_arguments.done",
                                    arguments=previous_item.content[0].text,
                                    name=function_name,
                                    item_id=current_item_id,
                                    output_index=current_output_index,
                                    sequence_number=-1,
                                )
                            )
                            function_call_item = ResponseFunctionToolCall(
                                type="function_call",
                                arguments=previous_item.content[0].text,
                                name=function_name,
                                item_id=current_item_id,
                                output_index=current_output_index,
                                sequence_number=-1,
                                call_id=f"fc_{random_uuid()}",
                                status="completed",
                            )
                            yield _increment_sequence_number_and_return(
                                ResponseOutputItemDoneEvent(
                                    type="response.output_item.done",
                                    sequence_number=-1,
                                    output_index=current_output_index,
                                    item=function_call_item,
                                )
                            )
1583
                    elif previous_item.channel == "analysis":
1584
1585
1586
1587
                        content = ResponseReasoningTextContent(
                            text=previous_item.content[0].text,
                            type="reasoning_text",
                        )
1588
1589
                        reasoning_item = ResponseReasoningItem(
                            type="reasoning",
1590
                            content=[content],
1591
                            status="completed",
1592
1593
                            id=current_item_id,
                            summary=[],
1594
                        )
1595
                        yield _increment_sequence_number_and_return(
1596
1597
1598
1599
1600
1601
1602
                            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=previous_item.content[0].text,
1603
1604
                            )
                        )
1605
1606
1607
1608
1609
1610
1611
1612
                        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=content,
1613
1614
                            )
                        )
1615
                        yield _increment_sequence_number_and_return(
1616
1617
1618
1619
1620
                            ResponseOutputItemDoneEvent(
                                type="response.output_item.done",
                                sequence_number=-1,
                                output_index=current_output_index,
                                item=reasoning_item,
1621
1622
                            )
                        )
1623
1624
1625
1626
1627
1628
                    elif previous_item.channel == "final":
                        text_content = ResponseOutputText(
                            type="output_text",
                            text=previous_item.content[0].text,
                            annotations=[],
                        )
1629
                        yield _increment_sequence_number_and_return(
1630
                            ResponseTextDoneEvent(
1631
1632
1633
1634
1635
1636
1637
                                type="response.output_text.done",
                                sequence_number=-1,
                                output_index=current_output_index,
                                content_index=current_content_index,
                                text=previous_item.content[0].text,
                                logprobs=[],
                                item_id=current_item_id,
1638
1639
                            )
                        )
1640
                        yield _increment_sequence_number_and_return(
1641
1642
1643
1644
1645
1646
1647
                            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=text_content,
1648
1649
                            )
                        )
1650
                        yield _increment_sequence_number_and_return(
1651
                            ResponseOutputItemDoneEvent(
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
                                type="response.output_item.done",
                                sequence_number=-1,
                                output_index=current_output_index,
                                item=ResponseOutputMessage(
                                    id=current_item_id,
                                    type="message",
                                    role="assistant",
                                    content=[text_content],
                                    status="completed",
                                ),
1662
1663
                            )
                        )
1664

1665
            # stream the output of a harmony message
1666
            if ctx.parser.last_content_delta:
1667
1668
1669
1670
                if (
                    ctx.parser.current_channel == "final"
                    and ctx.parser.current_recipient is None
                ):
1671
1672
                    if not sent_output_item_added:
                        sent_output_item_added = True
1673
                        current_item_id = f"msg_{random_uuid()}"
1674
                        yield _increment_sequence_number_and_return(
1675
1676
1677
1678
                            ResponseOutputItemAddedEvent(
                                type="response.output_item.added",
                                sequence_number=-1,
                                output_index=current_output_index,
1679
                                item=ResponseOutputMessage(
1680
1681
1682
1683
1684
1685
                                    id=current_item_id,
                                    type="message",
                                    role="assistant",
                                    content=[],
                                    status="in_progress",
                                ),
1686
1687
                            )
                        )
1688
                        current_content_index += 1
1689
                        yield _increment_sequence_number_and_return(
1690
1691
1692
1693
1694
1695
                            ResponseContentPartAddedEvent(
                                type="response.content_part.added",
                                sequence_number=-1,
                                output_index=current_output_index,
                                item_id=current_item_id,
                                content_index=current_content_index,
1696
                                part=ResponseOutputText(
1697
1698
1699
1700
1701
                                    type="output_text",
                                    text="",
                                    annotations=[],
                                    logprobs=[],
                                ),
1702
1703
                            )
                        )
1704
                    yield _increment_sequence_number_and_return(
1705
                        ResponseTextDeltaEvent(
1706
1707
1708
1709
1710
1711
1712
1713
                            type="response.output_text.delta",
                            sequence_number=-1,
                            content_index=current_content_index,
                            output_index=current_output_index,
                            item_id=current_item_id,
                            delta=ctx.parser.last_content_delta,
                            # TODO, use logprobs from ctx.last_request_output
                            logprobs=[],
1714
1715
1716
1717
1718
1719
                        )
                    )
                elif (
                    ctx.parser.current_channel == "analysis"
                    and ctx.parser.current_recipient is None
                ):
1720
1721
                    if not sent_output_item_added:
                        sent_output_item_added = True
1722
                        current_item_id = f"msg_{random_uuid()}"
1723
                        yield _increment_sequence_number_and_return(
1724
1725
1726
1727
                            ResponseOutputItemAddedEvent(
                                type="response.output_item.added",
                                sequence_number=-1,
                                output_index=current_output_index,
1728
                                item=ResponseReasoningItem(
1729
1730
1731
1732
1733
                                    type="reasoning",
                                    id=current_item_id,
                                    summary=[],
                                    status="in_progress",
                                ),
1734
1735
                            )
                        )
1736
                        current_content_index += 1
1737
                        yield _increment_sequence_number_and_return(
1738
1739
                            ResponseReasoningPartAddedEvent(
                                type="response.reasoning_part.added",
1740
1741
1742
1743
                                sequence_number=-1,
                                output_index=current_output_index,
                                item_id=current_item_id,
                                content_index=current_content_index,
1744
                                part=ResponseReasoningTextContent(
1745
                                    text="",
1746
                                    type="reasoning_text",
1747
                                ),
1748
1749
                            )
                        )
1750
                    yield _increment_sequence_number_and_return(
1751
1752
1753
1754
1755
1756
1757
                        ResponseReasoningTextDeltaEvent(
                            type="response.reasoning_text.delta",
                            item_id=current_item_id,
                            output_index=current_output_index,
                            content_index=current_content_index,
                            delta=ctx.parser.last_content_delta,
                            sequence_number=-1,
1758
1759
                        )
                    )
1760
1761
1762
                # built-in tools will be triggered on the analysis channel
                # However, occasionally built-in tools will
                # still be output to commentary.
1763
1764
1765
1766
                elif (
                    ctx.parser.current_channel == "commentary"
                    or ctx.parser.current_channel == "analysis"
                ) and ctx.parser.current_recipient == "python":
1767
1768
                    if not sent_output_item_added:
                        sent_output_item_added = True
1769
                        current_item_id = f"tool_{random_uuid()}"
1770
                        yield _increment_sequence_number_and_return(
1771
1772
1773
1774
                            ResponseOutputItemAddedEvent(
                                type="response.output_item.added",
                                sequence_number=-1,
                                output_index=current_output_index,
1775
                                item=ResponseCodeInterpreterToolCallParam(
1776
1777
1778
1779
1780
1781
1782
                                    type="code_interpreter_call",
                                    id=current_item_id,
                                    code=None,
                                    container_id="auto",
                                    outputs=None,
                                    status="in_progress",
                                ),
1783
1784
                            )
                        )
1785
                        yield _increment_sequence_number_and_return(
1786
                            ResponseCodeInterpreterCallInProgressEvent(
1787
                                type="response.code_interpreter_call.in_progress",
1788
1789
1790
                                sequence_number=-1,
                                output_index=current_output_index,
                                item_id=current_item_id,
1791
1792
                            )
                        )
1793
                    yield _increment_sequence_number_and_return(
1794
1795
1796
1797
1798
1799
                        ResponseCodeInterpreterCallCodeDeltaEvent(
                            type="response.code_interpreter_call_code.delta",
                            sequence_number=-1,
                            output_index=current_output_index,
                            item_id=current_item_id,
                            delta=ctx.parser.last_content_delta,
1800
1801
                        )
                    )
1802
1803

            # stream tool call outputs
1804
1805
            if ctx.is_assistant_action_turn() and len(ctx.parser.messages) > 0:
                previous_item = ctx.parser.messages[-1]
1806
1807
1808
1809
1810
1811
1812
                if (
                    self.tool_server is not None
                    and self.tool_server.has_tool("browser")
                    and previous_item.recipient is not None
                    and previous_item.recipient.startswith("browser.")
                ):
                    function_name = previous_item.recipient[len("browser.") :]
1813
1814
1815
                    action = None
                    parsed_args = json.loads(previous_item.content[0].text)
                    if function_name == "search":
1816
                        action = response_function_web_search.ActionSearch(
1817
1818
                            type="search",
                            query=parsed_args["query"],
1819
                        )
1820
                    elif function_name == "open":
1821
1822
1823
1824
1825
                        action = response_function_web_search.ActionOpenPage(
                            type="open_page",
                            # TODO: translate to url
                            url=f"cursor:{parsed_args.get('cursor', '')}",
                        )
1826
                    elif function_name == "find":
1827
1828
1829
1830
1831
1832
                        action = response_function_web_search.ActionFind(
                            type="find",
                            pattern=parsed_args["pattern"],
                            # TODO: translate to url
                            url=f"cursor:{parsed_args.get('cursor', '')}",
                        )
1833
                    else:
1834
                        raise ValueError(f"Unknown function name: {function_name}")
1835

1836
                    current_item_id = f"tool_{random_uuid()}"
1837
                    yield _increment_sequence_number_and_return(
1838
                        ResponseOutputItemAddedEvent(
1839
1840
1841
                            type="response.output_item.added",
                            sequence_number=-1,
                            output_index=current_output_index,
1842
                            item=response_function_web_search.ResponseFunctionWebSearch(
1843
1844
1845
1846
1847
1848
                                # TODO: generate a unique id for web search call
                                type="web_search_call",
                                id=current_item_id,
                                action=action,
                                status="in_progress",
                            ),
1849
1850
                        )
                    )
1851
                    yield _increment_sequence_number_and_return(
1852
1853
1854
1855
1856
                        ResponseWebSearchCallInProgressEvent(
                            type="response.web_search_call.in_progress",
                            sequence_number=-1,
                            output_index=current_output_index,
                            item_id=current_item_id,
1857
1858
                        )
                    )
1859
                    yield _increment_sequence_number_and_return(
1860
1861
1862
1863
1864
                        ResponseWebSearchCallSearchingEvent(
                            type="response.web_search_call.searching",
                            sequence_number=-1,
                            output_index=current_output_index,
                            item_id=current_item_id,
1865
1866
                        )
                    )
1867
1868

                    # enqueue
1869
                    yield _increment_sequence_number_and_return(
1870
1871
1872
1873
1874
                        ResponseWebSearchCallCompletedEvent(
                            type="response.web_search_call.completed",
                            sequence_number=-1,
                            output_index=current_output_index,
                            item_id=current_item_id,
1875
1876
                        )
                    )
1877
                    yield _increment_sequence_number_and_return(
1878
                        ResponseOutputItemDoneEvent(
1879
1880
1881
                            type="response.output_item.done",
                            sequence_number=-1,
                            output_index=current_output_index,
1882
                            item=ResponseFunctionWebSearch(
1883
1884
1885
1886
1887
                                type="web_search_call",
                                id=current_item_id,
                                action=action,
                                status="completed",
                            ),
1888
1889
                        )
                    )
1890

1891
1892
1893
1894
1895
1896
                if (
                    self.tool_server is not None
                    and self.tool_server.has_tool("python")
                    and previous_item.recipient is not None
                    and previous_item.recipient.startswith("python")
                ):
1897
                    yield _increment_sequence_number_and_return(
1898
1899
1900
1901
1902
                        ResponseCodeInterpreterCallCodeDoneEvent(
                            type="response.code_interpreter_call_code.done",
                            sequence_number=-1,
                            output_index=current_output_index,
                            item_id=current_item_id,
1903
                            code=previous_item.content[0].text,
1904
1905
                        )
                    )
1906
                    yield _increment_sequence_number_and_return(
1907
1908
1909
1910
1911
                        ResponseCodeInterpreterCallInterpretingEvent(
                            type="response.code_interpreter_call.interpreting",
                            sequence_number=-1,
                            output_index=current_output_index,
                            item_id=current_item_id,
1912
1913
                        )
                    )
1914
                    yield _increment_sequence_number_and_return(
1915
1916
1917
1918
1919
                        ResponseCodeInterpreterCallCompletedEvent(
                            type="response.code_interpreter_call.completed",
                            sequence_number=-1,
                            output_index=current_output_index,
                            item_id=current_item_id,
1920
1921
                        )
                    )
1922
                    yield _increment_sequence_number_and_return(
1923
                        ResponseOutputItemDoneEvent(
1924
1925
1926
                            type="response.output_item.done",
                            sequence_number=-1,
                            output_index=current_output_index,
1927
                            item=ResponseCodeInterpreterToolCallParam(
1928
1929
1930
1931
1932
1933
1934
1935
                                type="code_interpreter_call",
                                id=current_item_id,
                                code=previous_item.content[0].text,
                                container_id="auto",
                                # TODO: add outputs here
                                outputs=[],
                                status="completed",
                            ),
1936
1937
                        )
                    )
1938
1939
1940
1941
1942
1943
1944
1945
1946
1947
1948
1949
1950
1951
1952
1953
1954
1955
1956
1957
1958
1959
1960
1961
1962
1963
1964
1965
1966
1967
1968
1969
1970
1971
1972
1973
1974
            # developer tools will be triggered on the commentary channel
            # and recipient starts with "functions.TOOL_NAME"
            if (
                ctx.parser.current_channel == "commentary"
                and ctx.parser.current_recipient
                and ctx.parser.current_recipient.startswith("functions.")
            ):
                if is_first_function_call_delta is False:
                    is_first_function_call_delta = True
                    fc_name = ctx.parser.current_recipient[len("functions.") :]
                    tool_call_item = ResponseFunctionToolCall(
                        name=fc_name,
                        type="function_call",
                        id=current_item_id,
                        call_id=f"call_{random_uuid()}",
                        arguments="",
                        status="in_progress",
                    )
                    current_item_id = f"fc_{random_uuid()}"
                    yield _increment_sequence_number_and_return(
                        ResponseOutputItemAddedEvent(
                            type="response.output_item.added",
                            sequence_number=-1,
                            output_index=current_output_index,
                            item=tool_call_item,
                        )
                    )
                else:
                    yield _increment_sequence_number_and_return(
                        ResponseFunctionCallArgumentsDeltaEvent(
                            item_id=current_item_id,
                            delta=ctx.parser.last_content_delta,
                            output_index=current_output_index,
                            sequence_number=-1,
                            type="response.function_call_arguments.delta",
                        )
                    )
1975

1976
1977
1978
1979
    async def responses_stream_generator(
        self,
        request: ResponsesRequest,
        sampling_params: SamplingParams,
1980
        result_generator: AsyncIterator[ConversationContext | None],
1981
1982
        context: ConversationContext,
        model_name: str,
1983
        tokenizer: TokenizerLike,
1984
        request_metadata: RequestResponseMetadata,
1985
        created_time: int | None = None,
1986
    ) -> AsyncGenerator[StreamingResponsesResponse, None]:
1987
1988
1989
1990
1991
        # TODO:
        # 1. Handle disconnect

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

1992
1993
        sequence_number = 0

1994
        def _increment_sequence_number_and_return(
1995
            event: StreamingResponsesResponse,
1996
        ) -> StreamingResponsesResponse:
1997
1998
            nonlocal sequence_number
            # Set sequence_number if the event has this attribute
1999
            if hasattr(event, "sequence_number"):
2000
2001
                event.sequence_number = sequence_number
            sequence_number += 1
2002
            return event
2003

2004
        async with AsyncExitStack() as exit_stack:
2005
2006
            processer = None
            if self.use_harmony:
2007
2008
                # TODO: in streaming, we noticed this bug:
                # https://github.com/vllm-project/vllm/issues/25697
2009
                await self._initialize_tool_sessions(request, context, exit_stack)
2010
2011
2012
                processer = self._process_harmony_streaming_events
            else:
                processer = self._process_simple_streaming_events
2013
            # TODO Hanchen make sampling params to include the structural tag
2014
2015
2016
2017
2018
2019
2020
2021
2022
2023

            initial_response = ResponsesResponse.from_request(
                request,
                sampling_params,
                model_name=model_name,
                created_time=created_time,
                output=[],
                status="in_progress",
                usage=None,
            ).model_dump()
2024
            yield _increment_sequence_number_and_return(
2025
2026
2027
2028
                ResponseCreatedEvent(
                    type="response.created",
                    sequence_number=-1,
                    response=initial_response,
2029
2030
                )
            )
2031
            yield _increment_sequence_number_and_return(
2032
2033
2034
2035
                ResponseInProgressEvent(
                    type="response.in_progress",
                    sequence_number=-1,
                    response=initial_response,
2036
2037
                )
            )
2038

2039
2040
2041
2042
2043
2044
2045
2046
2047
2048
2049
2050
2051
2052
2053
2054
2055
2056
2057
            try:
                async for event_data in processer(
                    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
2058
2059
2060
2061
2062
2063
2064
2065
2066
2067
2068
2069
2070
2071
2072
2073
2074

            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,
            )
2075
            yield _increment_sequence_number_and_return(
2076
                ResponseCompletedEvent(
2077
2078
                    type="response.completed",
                    sequence_number=-1,
2079
                    response=final_response,
2080
2081
                )
            )