93 lines
2.8 KiB
Python
93 lines
2.8 KiB
Python
# SPDX-License-Identifier: Apache-2.0
|
|
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
|
from typing import Any
|
|
|
|
from vllm.config import VllmConfig
|
|
from vllm.entrypoints.chat_utils import (
|
|
ChatCompletionMessageParam,
|
|
ConversationMessage,
|
|
parse_chat_messages,
|
|
parse_chat_messages_async,
|
|
)
|
|
from vllm.logger import init_logger
|
|
from vllm.tokenizers import cached_get_tokenizer
|
|
from vllm.tokenizers.deepseek_v32 import DeepseekV32Tokenizer
|
|
|
|
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 DeepseekV32Renderer(BaseRenderer[DeepseekV32Tokenizer]):
|
|
@classmethod
|
|
def from_config( # type: ignore[override]
|
|
cls,
|
|
config: VllmConfig,
|
|
tokenizer_kwargs: dict[str, Any],
|
|
) -> "DeepseekV32Renderer":
|
|
model_config = config.model_config
|
|
if model_config.skip_tokenizer_init:
|
|
tokenizer = None
|
|
else:
|
|
tokenizer = cached_get_tokenizer(
|
|
tokenizer_cls=DeepseekV32Tokenizer,
|
|
**tokenizer_kwargs,
|
|
)
|
|
|
|
return cls(config, tokenizer)
|
|
|
|
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",
|
|
)
|
|
|
|
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",
|
|
)
|
|
|
|
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
|