| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import asyncio |
| 5 | import time |
| 6 | from collections.abc import AsyncGenerator |
| 7 | from http import HTTPStatus |
| 8 | |
| 9 | from fastapi import Request |
| 10 | |
| 11 | from vllm.entrypoints.chat_utils import ConversationMessage |
| 12 | from vllm.entrypoints.generate.base.protocol import RequestResponseMetadata |
| 13 | from vllm.entrypoints.openai.chat_completion.protocol import ( |
| 14 | BatchChatCompletionRequest, |
| 15 | ChatCompletionRequest, |
| 16 | ChatCompletionResponse, |
| 17 | ChatCompletionResponseChoice, |
| 18 | ChatMessage, |
| 19 | ) |
| 20 | from vllm.entrypoints.openai.chat_completion.serving import OpenAIServingChat |
| 21 | from vllm.entrypoints.serve.engine.protocol import ErrorResponse, UsageInfo |
| 22 | from vllm.entrypoints.serve.utils.api_utils import get_max_tokens |
| 23 | from vllm.inputs import EngineInput |
| 24 | from vllm.logger import init_logger |
| 25 | from vllm.outputs import RequestOutput |
| 26 | from vllm.parser.abstract_parser import Parser |
| 27 | from vllm.tokenizers import TokenizerLike |
| 28 | from vllm.utils.async_utils import merge_async_iterators |
| 29 | from vllm.utils.collection_utils import as_list |
| 30 | |
| 31 | logger = init_logger(__name__) |
| 32 | |
| 33 | |
| 34 | class OpenAIServingChatBatch(OpenAIServingChat): |
| 35 | """Extends OpenAIServingChat with the /v1/chat/completions/batch endpoint. |
| 36 | |
| 37 | Processes N conversations from a single request concurrently and returns |
| 38 | one choice per conversation indexed 0, 1, ..., N-1. |
| 39 | """ |
| 40 | |
| 41 | async def render_batch_chat_request( |
| 42 | self, |
| 43 | request: BatchChatCompletionRequest, |
| 44 | ) -> ( |
| 45 | tuple[ |
| 46 | list[ChatCompletionRequest], |
| 47 | list[list[ConversationMessage]], |
| 48 | list[EngineInput], |
| 49 | ] |
| 50 | | ErrorResponse |
| 51 | ): |
| 52 | """Validate the model and preprocess a batched chat completion request. |
| 53 | |
| 54 | Performs engine-aware checks then delegates per-conversation |
| 55 | preprocessing to OnlineRenderer, validating the chat template |
| 56 | once for the whole batch. |
| 57 | |
| 58 | Returns: |
| 59 | A tuple of (single_requests, all_conversations, engine_prompts) on |
| 60 | success — one entry per conversation — or an ErrorResponse on |
| 61 | failure. The renderer adjusts ``single_requests`` in place, so |
| 62 | sampling must read them rather than rebuilding from ``request``. |
| 63 | |
| 64 | """ |
| 65 | error_check_ret = await self._check_model(request) |
| 66 | if error_check_ret is not None: |
| 67 | logger.error("Error with model %s", error_check_ret) |
| 68 | return error_check_ret |
| 69 | |
| 70 | self._preflight() |
| 71 | |
| 72 | renderer = self.online_renderer |
| 73 | |
| 74 | if not renderer.use_harmony: |
| 75 | # Common case: validate the chat template once for the whole batch. |
| 76 | error_check_ret = renderer.validate_chat_template( |
| 77 | request_chat_template=request.chat_template, |
| 78 | chat_template_kwargs=request.chat_template_kwargs, |
| 79 | trust_request_chat_template=renderer.trust_request_chat_template, |
| 80 | ) |
| 81 | if error_check_ret is not None: |
| 82 | return error_check_ret |
| 83 | |
| 84 | parser = renderer.parser |
| 85 | tool_dicts: list[dict] | None = None |
| 86 | |
| 87 | single_requests: list[ChatCompletionRequest] = [] |
| 88 | all_conversations: list[list[ConversationMessage]] = [] |
| 89 | all_engine_prompts: list[EngineInput] = [] |
| 90 | |
| 91 | for messages in request.messages: |
| 92 | single_request = request.to_chat_completion_request(messages) |
| 93 | if renderer.use_harmony: |
| 94 | renderer.adjust_harmony_request(single_request) |
| 95 | conversation, engine_prompts = renderer._make_request_with_harmony( |
| 96 | single_request, should_include_tools=tool_dicts is not None |
| 97 | ) |
| 98 | else: |
| 99 | conversation, engine_prompts = await renderer.preprocess_chat( |
| 100 | single_request, |
| 101 | messages, |
| 102 | default_template=renderer.chat_template, |
| 103 | default_template_content_format=renderer.chat_template_content_format, |
| 104 | default_template_kwargs=renderer.default_chat_template_kwargs, |
| 105 | tool_dicts=tool_dicts, |
| 106 | parser=parser, |
| 107 | ) |
| 108 | single_requests.append(single_request) |
| 109 | all_conversations.append(conversation) |
| 110 | all_engine_prompts.append(engine_prompts[0]) |
| 111 | |
| 112 | return single_requests, all_conversations, all_engine_prompts |
| 113 | |
| 114 | async def create_batch_chat_completion( |
| 115 | self, |
| 116 | request: BatchChatCompletionRequest, |
| 117 | raw_request: Request | None = None, |
| 118 | ) -> ChatCompletionResponse | ErrorResponse: |
| 119 | """Batch Chat Completion endpoint (/v1/chat/completions/batch). |
| 120 | |
| 121 | Processes N conversations from a single request concurrently and |
| 122 | returns one choice per conversation indexed 0, 1, ..., N-1. |
| 123 | Streaming, tool use, and beam search are not supported. |
| 124 | """ |
| 125 | tokenizer = self.renderer.tokenizer |
| 126 | assert tokenizer is not None |
| 127 | render_result = await self.render_batch_chat_request(request) |
| 128 | if isinstance(render_result, ErrorResponse): |
| 129 | return render_result |
| 130 | single_requests, all_conversations, engine_prompts = render_result |
| 131 | |
| 132 | request_id = ( |
| 133 | f"chatcmpl-{self._base_request_id(raw_request, request.request_id)}" |
| 134 | ) |
| 135 | request_metadata = RequestResponseMetadata(request_id=request_id) |
| 136 | if raw_request: |
| 137 | raw_request.state.request_metadata = request_metadata |
| 138 | |
| 139 | lora_request = self._maybe_get_adapters(request, supports_default_mm_loras=True) |
| 140 | model_name = self.models.model_name(lora_request) |
| 141 | data_parallel_rank = self._get_data_parallel_rank(raw_request) |
| 142 | max_model_len = self.model_config.max_model_len |
| 143 | |
| 144 | generators: list[AsyncGenerator[RequestOutput, None]] = [] |
| 145 | for i, engine_prompt in enumerate(engine_prompts): |
| 146 | sub_request_id = f"{request_id}_{i}" |
| 147 | max_tokens = get_max_tokens( |
| 148 | max_model_len, |
| 149 | request.max_completion_tokens |
| 150 | if request.max_completion_tokens is not None |
| 151 | else request.max_tokens, |
| 152 | self._extract_prompt_len(engine_prompt), |
| 153 | self.default_sampling_params, |
| 154 | self.override_max_tokens, |
| 155 | ) |
| 156 | single_request = single_requests[i] |
| 157 | sampling_params = single_request.to_sampling_params( |
| 158 | max_tokens, self.default_sampling_params |
| 159 | ) |
| 160 | self._log_inputs( |
| 161 | sub_request_id, |
| 162 | engine_prompt, |
| 163 | params=sampling_params, |
| 164 | lora_request=lora_request, |
| 165 | ) |
| 166 | trace_headers = ( |
| 167 | None |
| 168 | if raw_request is None |
| 169 | else await self._get_trace_headers(raw_request.headers) |
| 170 | ) |
| 171 | session_id = self._get_session_id(single_request, raw_request) |
| 172 | generators.append( |
| 173 | self.engine_client.generate( |
| 174 | engine_prompt, |
| 175 | sampling_params, |
| 176 | sub_request_id, |
| 177 | lora_request=lora_request, |
| 178 | trace_headers=trace_headers, |
| 179 | priority=request.priority, |
| 180 | data_parallel_rank=data_parallel_rank, |
| 181 | session_id=session_id, |
| 182 | reasoning_ended=None, |
| 183 | ) |
| 184 | ) |
| 185 | |
| 186 | return await self.chat_completion_full_generator_batch( |
| 187 | request, # type: ignore[arg-type] |
| 188 | generators, |
| 189 | request_id, |
| 190 | model_name, |
| 191 | all_conversations, |
| 192 | tokenizer, |
| 193 | request_metadata, |
| 194 | self.parser_cls, |
| 195 | ) |
| 196 | |
| 197 | async def chat_completion_full_generator_batch( |
| 198 | self, |
| 199 | request: BatchChatCompletionRequest, # type: ignore[override] |
| 200 | generators: list[AsyncGenerator[RequestOutput, None]], |
| 201 | request_id: str, |
| 202 | model_name: str, |
| 203 | all_conversations: list[list[ConversationMessage]], |
| 204 | tokenizer: TokenizerLike, |
| 205 | request_metadata: RequestResponseMetadata, |
| 206 | parser_cls: type[Parser] | None = None, |
| 207 | ) -> ErrorResponse | ChatCompletionResponse: |
| 208 | """Handle batched (non-streaming) chat completions. |
| 209 | |
| 210 | Fans out N generators (one per conversation in the batch), collects |
| 211 | the final output for each, and assembles a single |
| 212 | ``ChatCompletionResponse`` whose ``choices`` are indexed 0,...,N-1. |
| 213 | |
| 214 | Tool-use and streaming are rejected upstream by the |
| 215 | ``check_batch_mode`` validator, so neither needs to be handled here. |
| 216 | """ |
| 217 | created_time = int(time.time()) |
| 218 | final_results: dict[int, RequestOutput] = {} |
| 219 | try: |
| 220 | async for prompt_idx, res in merge_async_iterators(*generators): |
| 221 | final_results[prompt_idx] = res |
| 222 | except asyncio.CancelledError: |
| 223 | return self.create_error_response("Client disconnected") |
| 224 | |
| 225 | choices: list[ChatCompletionResponseChoice] = [] |
| 226 | total_prompt_tokens = 0 |
| 227 | total_completion_tokens = 0 |
| 228 | |
| 229 | for prompt_idx in range(len(generators)): |
| 230 | final_res = final_results.get(prompt_idx) |
| 231 | if final_res is None: |
| 232 | return self.create_error_response( |
| 233 | f"No output received from the engine for prompt {prompt_idx}.", |
| 234 | err_type="InternalServerError", |
| 235 | status_code=HTTPStatus.INTERNAL_SERVER_ERROR, |
| 236 | ) |
| 237 | |
| 238 | assert final_res.prompt_token_ids is not None |
| 239 | num_prompt_tokens = len(final_res.prompt_token_ids) |
| 240 | if final_res.encoder_prompt_token_ids is not None: |
| 241 | num_prompt_tokens += len(final_res.encoder_prompt_token_ids) |
| 242 | total_prompt_tokens += num_prompt_tokens |
| 243 | total_completion_tokens += sum( |
| 244 | len(output.token_ids) for output in final_res.outputs |
| 245 | ) |
| 246 | |
| 247 | for output in final_res.outputs: |
| 248 | self._raise_if_error(output.finish_reason, request_id) |
| 249 | |
| 250 | if request.logprobs and ( |
| 251 | request.top_logprobs is not None or request.logprob_token_ids |
| 252 | ): |
| 253 | assert output.logprobs is not None, "Did not output logprobs" |
| 254 | logprobs = self._create_chat_logprobs( |
| 255 | token_ids=output.token_ids, |
| 256 | top_logprobs=output.logprobs, |
| 257 | num_output_top_logprobs=request.top_logprobs, |
| 258 | logprob_token_ids=request.logprob_token_ids, |
| 259 | tokenizer=tokenizer, |
| 260 | return_as_token_id=request.return_tokens_as_token_ids, |
| 261 | ) |
| 262 | else: |
| 263 | logprobs = None |
| 264 | |
| 265 | if parser_cls is not None: |
| 266 | single_request = request.to_chat_completion_request( |
| 267 | request.messages[prompt_idx] |
| 268 | ) |
| 269 | parser = parser_cls( |
| 270 | tokenizer, |
| 271 | None, |
| 272 | chat_template_kwargs=self._effective_chat_template_kwargs( |
| 273 | single_request |
| 274 | ), |
| 275 | ) |
| 276 | parser.set_prompt_token_ids(final_res.prompt_token_ids) |
| 277 | reasoning, content, _ = parser.parse( |
| 278 | output.text, |
| 279 | request=single_request, |
| 280 | model_output_token_ids=output.token_ids, |
| 281 | ) |
| 282 | if not request.include_reasoning: |
| 283 | reasoning = None |
| 284 | else: |
| 285 | reasoning = None |
| 286 | content = output.text |
| 287 | |
| 288 | role = ( |
| 289 | self.response_role |
| 290 | if request.add_generation_prompt |
| 291 | else request.messages[prompt_idx][-1]["role"] |
| 292 | ) |
| 293 | |
| 294 | message = ChatMessage(role=role, reasoning=reasoning, content=content) |
| 295 | |
| 296 | if request.echo: |
| 297 | conversation = all_conversations[prompt_idx] |
| 298 | last_msg_content: str | list[dict[str, str]] = "" |
| 299 | if ( |
| 300 | conversation |
| 301 | and "content" in conversation[-1] |
| 302 | and conversation[-1].get("role") == role |
| 303 | ): |
| 304 | last_msg_content = conversation[-1]["content"] or "" |
| 305 | if isinstance(last_msg_content, list): |
| 306 | last_msg_content = "\n".join( |
| 307 | msg["text"] for msg in last_msg_content |
| 308 | ) |
| 309 | message.content = last_msg_content + (message.content or "") |
| 310 | |
| 311 | choice_data = ChatCompletionResponseChoice( |
| 312 | index=prompt_idx, |
| 313 | message=message, |
| 314 | logprobs=logprobs, |
| 315 | finish_reason=output.finish_reason |
| 316 | if output.finish_reason |
| 317 | else "stop", |
| 318 | stop_reason=output.stop_reason, |
| 319 | token_ids=( |
| 320 | as_list(output.token_ids) if request.return_token_ids else None |
| 321 | ), |
| 322 | ) |
| 323 | choices.append(choice_data) |
| 324 | |
| 325 | usage = UsageInfo( |
| 326 | prompt_tokens=total_prompt_tokens, |
| 327 | completion_tokens=total_completion_tokens, |
| 328 | total_tokens=total_prompt_tokens + total_completion_tokens, |
| 329 | ) |
| 330 | request_metadata.final_usage_info = usage |
| 331 | |
| 332 | choices.sort(key=lambda c: c.index) |
| 333 | |
| 334 | return ChatCompletionResponse( |
| 335 | id=request_id, |
| 336 | created=created_time, |
| 337 | model=model_name, |
| 338 | choices=choices, |
| 339 | usage=usage, |
| 340 | system_fingerprint=self.system_fingerprint, |
| 341 | ) |