grok2.py

# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project

from vllm.entrypoints.chat_utils import (
    ChatCompletionMessageParam,
    ConversationMessage,
    parse_chat_messages,
    parse_chat_messages_async,
)
from vllm.logger import init_logger
from vllm.tokenizers.grok2 import Grok2Tokenizer

from .base import BaseRenderer
from .inputs import DictPrompt
from .inputs.preprocess import parse_dec_only_prompt
from .params import ChatParams

logger = init_logger(__name__)


class Grok2Renderer(BaseRenderer[Grok2Tokenizer]):
    def render_messages(
        self,
        messages: list[ChatCompletionMessageParam],
        params: ChatParams,
    ) -> tuple[list[ConversationMessage], DictPrompt]:
        tokenizer = self.get_tokenizer()
        conversation, mm_data, mm_uuids = parse_chat_messages(
            messages,
            self.model_config,
            content_format="string",
            media_io_kwargs=params.media_io_kwargs,
            mm_processor_kwargs=params.mm_processor_kwargs,
        )

        prompt_raw = tokenizer.apply_chat_template(
            conversation=conversation,
            messages=messages,
            **params.get_apply_chat_template_kwargs(),
        )

        prompt = parse_dec_only_prompt(prompt_raw)
        if mm_data is not None:
            prompt["multi_modal_data"] = mm_data
        if mm_uuids is not None:
            prompt["multi_modal_uuids"] = mm_uuids

        return conversation, prompt

    async def render_messages_async(
        self,
        messages: list[ChatCompletionMessageParam],
        params: ChatParams,
    ) -> tuple[list[ConversationMessage], DictPrompt]:
        tokenizer = self.get_tokenizer()
        conversation, mm_data, mm_uuids = await parse_chat_messages_async(
            messages,
            self.model_config,
            content_format="string",
            media_io_kwargs=params.media_io_kwargs,
            mm_processor_kwargs=params.mm_processor_kwargs,
        )

        prompt_raw = tokenizer.apply_chat_template(
            conversation=conversation,
            messages=messages,
            **params.get_apply_chat_template_kwargs(),
        )

        prompt = parse_dec_only_prompt(prompt_raw)
        if mm_data is not None:
            prompt["multi_modal_data"] = mm_data
        if mm_uuids is not None:
            prompt["multi_modal_uuids"] = mm_uuids

        return conversation, prompt