aphrodite-engine/aphrodite/endpoints/openai/serving_chat.py

import codecs
import time
from typing import AsyncGenerator, AsyncIterator, List, Optional, Union

from fastapi import Request
from loguru import logger

from aphrodite.common.outputs import RequestOutput
from aphrodite.common.utils import random_uuid
from aphrodite.endpoints.openai.protocol import (
    ChatCompletionNamedToolChoiceParam, ChatCompletionRequest,
    ChatCompletionResponse, ChatCompletionResponseChoice,
    ChatCompletionResponseStreamChoice, ChatCompletionStreamResponse,
    ChatMessage, DeltaMessage, ErrorResponse, FunctionCall, ToolCall,
    UsageInfo)
from aphrodite.endpoints.openai.serving_engine import LoRA, OpenAIServing
from aphrodite.engine.async_aphrodite import AsyncAphrodite
from aphrodite.modeling.guided_decoding import \
    get_guided_decoding_logits_processor


class OpenAIServingChat(OpenAIServing):

    def __init__(self,
                 engine: AsyncAphrodite,
                 served_model_names: List[str],
                 response_role: str,
                 lora_modules: Optional[List[LoRA]] = None,
                 chat_template=None):
        super().__init__(engine=engine,
                         served_model_names=served_model_names,
                         lora_modules=lora_modules)
        self.response_role = response_role
        self._load_chat_template(chat_template)

    async def create_chat_completion(
        self, request: ChatCompletionRequest, raw_request: Request
    ) -> Union[ErrorResponse, AsyncGenerator[str, None],
               ChatCompletionResponse]:
        """Completion API similar to OpenAI's API.

        See https://platform.openai.com/docs/api-reference/chat/create
        for the API specification. This API mimics the OpenAI
        ChatCompletion API.

        NOTE: Currently we do not support the following feature:
            - function_call (Users should implement this by themselves)
        """
        error_check_ret = await self._check_model(request)
        if error_check_ret is not None:
            return error_check_ret

        # Deal with list in messages.content
        # Just replace the content list with the very first text message
        for message in request.messages:
            if message["role"] == "user" and isinstance(
                    message["content"], list):
                message["content"] = next((content["text"]
                                           for content in message["content"]
                                           if content["type"] == "text"), "")

        try:
            prompt = self.tokenizer.apply_chat_template(
                conversation=request.messages,
                tokenize=False,
                add_generation_prompt=request.add_generation_prompt)
        except Exception as e:
            logger.error(
                f"Error in applying chat template from request: {str(e)}")
            return self.create_error_response(str(e))

        request_id = f"cmpl-{random_uuid()}"
        try:
            # Tokenize/detokenize depending on prompt format (string/token list)
            prompt_ids, prompt_text = self._validate_prompt_and_tokenize(
                request, prompt=prompt)
            sampling_params = request.to_sampling_params(
                self.tokenizer.vocab_size)
            lora_request = self._maybe_get_lora(request)
            guided_decode_logits_processor = (
                await get_guided_decoding_logits_processor(
                    request.guided_decoding_backend, request, await
                    self.engine.get_tokenizer()))
            if guided_decode_logits_processor:
                sampling_params.logits_processors.append(
                    guided_decode_logits_processor)
        except ValueError as e:
            return self.create_error_response(str(e))

        result_generator = self.engine.generate(
            {
                "prompt": prompt_text,
                "prompt_token_ids": prompt_ids
            },
            sampling_params,
            request_id,
            lora_request,
        )
        # Streaming response
        if request.stream:
            return self.chat_completion_stream_generator(
                request, result_generator, request_id)
        else:
            try:
                return await self.chat_completion_full_generator(
                    request, raw_request, result_generator, request_id)
            except ValueError as e:
                # TODO: Use an aphrodite-specific Validation Error
                return self.create_error_response(str(e))

    def get_chat_request_role(self, request: ChatCompletionRequest) -> str:
        if request.add_generation_prompt:
            return self.response_role
        else:
            return request.messages[-1]["role"]

    async def chat_completion_stream_generator(
            self, request: ChatCompletionRequest,
            result_generator: AsyncIterator[RequestOutput], request_id: str
    ) -> Union[ErrorResponse, AsyncGenerator[str, None]]:

        model_name = self.served_model_names[0]
        created_time = int(time.time())
        chunk_object_type = "chat.completion.chunk"
        first_iteration = True

        # Send response for each token for each request.n (index)
        previous_texts = [""] * request.n
        previous_num_tokens = [0] * request.n
        finish_reason_sent = [False] * request.n
        try:
            async for res in result_generator:
                res: RequestOutput
                # We need to do it here, because if there are exceptions in
                # the result_generator, it needs to be sent as the FIRST
                # response (by the try...catch).
                if first_iteration:
                    # Send first response for each request.n (index) with
                    # the role
                    role = self.get_chat_request_role(request)
                    for i in range(request.n):
                        choice_data = ChatCompletionResponseStreamChoice(
                            index=i,
                            delta=DeltaMessage(role=role),
                            logprobs=None,
                            finish_reason=None)
                        chunk = ChatCompletionStreamResponse(
                            id=request_id,
                            object=chunk_object_type,
                            created=created_time,
                            choices=[choice_data],
                            model=model_name)
                        if (request.stream_options
                                and request.stream_options.include_usage):
                            chunk.usage = None
                        data = chunk.model_dump_json(exclude_unset=True)
                        yield f"data: {data}\n\n"

                    # Send response to echo the input portion of the
                    # last message
                    if request.echo:
                        last_msg_content = ""
                        if request.messages and isinstance(
                                request.messages,
                                list) and request.messages[-1].get(
                                    "content") and request.messages[-1].get(
                                        "role") == role:
                            last_msg_content = request.messages[-1]["content"]

                        if last_msg_content:
                            for i in range(request.n):
                                choice_data = (
                                    ChatCompletionResponseStreamChoice(
                                        index=i,
                                        delta=DeltaMessage(
                                            content=last_msg_content),
                                        finish_reason=None))
                                chunk = ChatCompletionStreamResponse(
                                    id=request_id,
                                    object=chunk_object_type,
                                    created=created_time,
                                    choices=[choice_data],
                                    logprobs=None,
                                    model=model_name)
                                if (request.stream_options and
                                        request.stream_options.include_usage):
                                    chunk.usage = None
                                data = chunk.model_dump_json(
                                    exclude_unset=True)
                                yield f"data: {data}\n\n"
                    first_iteration = False

                for output in res.outputs:
                    i = output.index

                    if finish_reason_sent[i]:
                        continue

                    delta_token_ids = output.token_ids[previous_num_tokens[i]:]
                    top_logprobs = output.logprobs[
                        previous_num_tokens[i]:] if output.logprobs else None

                    if request.logprobs:
                        logprobs = self._create_logprobs(
                            token_ids=delta_token_ids,
                            top_logprobs=top_logprobs,
                            num_output_top_logprobs=request.logprobs,
                            initial_text_offset=len(previous_texts[i]),
                        )
                    else:
                        logprobs = None

                    delta_text = output.text[len(previous_texts[i]):]
                    previous_texts[i] = output.text
                    previous_num_tokens[i] = len(output.token_ids)
                    if request.tool_choice and type(
                            request.tool_choice
                    ) is ChatCompletionNamedToolChoiceParam:
                        delta_message = DeltaMessage(tool_calls=[
                            ToolCall(function=FunctionCall(
                                name=request.tool_choice.function.name,
                                arguments=delta_text))
                        ])
                    else:
                        delta_message = DeltaMessage(content=delta_text)
                    if output.finish_reason is None:
                        # Send token-by-token response for each request.n
                        choice_data = ChatCompletionResponseStreamChoice(
                            index=i,
                            delta=delta_message,
                            logprobs=logprobs,
                            finish_reason=None)
                        chunk = ChatCompletionStreamResponse(
                            id=request_id,
                            object=chunk_object_type,
                            created=created_time,
                            choices=[choice_data],
                            model=model_name)
                        if (request.stream_options
                                and request.stream_options.include_usage):
                            chunk.usage = None
                        data = chunk.model_dump_json(exclude_unset=True)
                        yield f"data: {data}\n\n"
                    else:
                        # Send the finish response for each request.n only once
                        prompt_tokens = len(res.prompt_token_ids)
                        choice_data = ChatCompletionResponseStreamChoice(
                            index=i,
                            delta=delta_message,
                            logprobs=logprobs,
                            finish_reason=output.finish_reason,
                            stop_reason=output.stop_reason)
                        chunk = ChatCompletionStreamResponse(
                            id=request_id,
                            object=chunk_object_type,
                            created=created_time,
                            choices=[choice_data],
                            model=model_name)
                        if (request.stream_options
                                and request.stream_options.include_usage):
                            chunk.usage = None
                        data = chunk.model_dump_json(exclude_unset=True)
                        yield f"data: {data}\n\n"
                        finish_reason_sent[i] = True

                    if (request.stream_options
                            and request.stream_options.include_usage):
                        final_usage = UsageInfo(
                            prompt_tokens=prompt_tokens,
                            completion_tokens=previous_num_tokens[i],
                            total_tokens=prompt_tokens +
                            previous_num_tokens[i],
                        )

                        final_usage_chunk = ChatCompletionStreamResponse(
                            id=request_id,
                            object=chunk_object_type,
                            created=created_time,
                            choices=[],
                            model=model_name,
                            usage=final_usage)
                        final_usage_data = (final_usage_chunk.model_dump_json(
                            exclude_unset=True, exclude_none=True))
                        yield f"data: {final_usage_data}\n\n"
        except ValueError as e:
            # TODO: Use an aphrodite-specific Validation Error
            data = self.create_streaming_error_response(str(e))
            yield f"data: {data}\n\n"
        # Send the final done message after all response.n are finished
        yield "data: [DONE]\n\n"

    async def chat_completion_full_generator(
            self, request: ChatCompletionRequest, raw_request: Request,
            result_generator: AsyncIterator[RequestOutput],
            request_id: str) -> Union[ErrorResponse, ChatCompletionResponse]:

        model_name = self.served_model_names[0]
        created_time = int(time.time())
        final_res: RequestOutput = None

        async for res in result_generator:
            if await raw_request.is_disconnected():
                # Abort the request if the client disconnects.
                await self.engine.abort(request_id)
                return self.create_error_response("Client disconnected")
            final_res = res
        assert final_res is not None

        choices = []

        role = self.get_chat_request_role(request)
        for output in final_res.outputs:
            token_ids = output.token_ids
            top_logprobs = output.logprobs

            if request.logprobs:
                logprobs = self._create_logprobs(
                    token_ids=token_ids,
                    top_logprobs=top_logprobs,
                    num_output_top_logprobs=request.logprobs,
                )
            else:
                logprobs = None

            if request.tool_choice and type(
                    request.tool_choice) is ChatCompletionNamedToolChoiceParam:
                message = ChatMessage(
                    role=role,
                    content="",
                    tool_calls=[
                        ToolCall(function=FunctionCall(
                            name=request.tool_choice.function.name,
                            arguments=output.text))
                    ])
            elif not request.tool_choice or request.tool_choice == "none":
                message = ChatMessage(role=role, content=output.text)

            choice_data = ChatCompletionResponseChoice(
                index=output.index,
                message=message,
                logprobs=logprobs,
                finish_reason=output.finish_reason,
                stop_reason=output.stop_reason,
            )
            choices.append(choice_data)

        if request.echo:
            last_msg_content = ""
            if request.messages and isinstance(
                    request.messages, list) and request.messages[-1].get(
                        "content") and request.messages[-1].get(
                            "role") == role:
                last_msg_content = request.messages[-1]["content"]

            for choice in choices:
                full_message = last_msg_content + choice.message.content
                choice.message.content = full_message

        num_prompt_tokens = len(final_res.prompt_token_ids)
        num_generated_tokens = sum(
            len(output.token_ids) for output in final_res.outputs)
        usage = UsageInfo(
            prompt_tokens=num_prompt_tokens,
            completion_tokens=num_generated_tokens,
            total_tokens=num_prompt_tokens + num_generated_tokens,
        )
        response = ChatCompletionResponse(
            id=request_id,
            created=created_time,
            model=model_name,
            choices=choices,
            usage=usage,
        )

        return response

    def _load_chat_template(self, chat_template):
        if chat_template is not None:
            try:
                with open(chat_template, "r") as f:
                    self.tokenizer.chat_template = f.read()
            except OSError:
                # If opening a file fails, set chat template to be args to
                # ensure we decode so our escape are interpreted correctly
                self.tokenizer.chat_template = codecs.decode(
                    chat_template, "unicode_escape")

            logger.info("Using the supplied chat template.")
        elif self.tokenizer.chat_template is not None:
            logger.info("Using the default chat template")
        else:
            logger.warning(
                "No chat template provided. Chat API will not work.")