Skip to content

GitLab

Switch branch/tag
Find file Normal viewHistoryPermalink
deepseek_v32.py 3.42 KB
Newer Older
Cyrus Leung's avatar
[Frontend] Introduce Renderer for processing chat messages (using `ModelConfig`) (#30200)  
Cyrus Leung committed
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119 
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from typing import Any

from vllm.config import ModelConfig
from vllm.entrypoints.chat_utils import (
    ChatCompletionMessageParam,
    ConversationMessage,
    parse_chat_messages,
    parse_chat_messages_async,
)
from vllm.inputs import TextPrompt, TokensPrompt
from vllm.logger import init_logger
from vllm.tokenizers import cached_get_tokenizer
from vllm.tokenizers.deepseek_v32 import DeepseekV32Tokenizer

from .protocol import RendererLike

logger = init_logger(__name__)


class DeepseekV32Renderer(RendererLike):
    @classmethod
    def from_config(
        cls,
        config: ModelConfig,
        tokenizer_kwargs: dict[str, Any],
    ) -> "RendererLike":
        return cls(config, tokenizer_kwargs)

    def __init__(
        self,
        config: ModelConfig,
        tokenizer_kwargs: dict[str, Any],
    ) -> None:
        super().__init__()

        self.config = config

        if config.skip_tokenizer_init:
            tokenizer = None
        else:
            tokenizer = cached_get_tokenizer(
                tokenizer_cls=DeepseekV32Tokenizer,
                **tokenizer_kwargs,
            )

        self._tokenizer = tokenizer

    @property
    def tokenizer(self) -> DeepseekV32Tokenizer | None:
        return self._tokenizer

    def get_tokenizer(self) -> DeepseekV32Tokenizer:
        tokenizer = self.tokenizer
        if tokenizer is None:
            raise ValueError("Tokenizer not available when `skip_tokenizer_init=True`")

        return tokenizer

    def render_messages(
        self,
        messages: list[ChatCompletionMessageParam],
        **kwargs,
    ) -> tuple[list[ConversationMessage], TextPrompt | TokensPrompt]:
        tokenizer = self.get_tokenizer()
        conversation, mm_data, mm_uuids = parse_chat_messages(
            messages,
            self.config,
            content_format="string",
        )

        prompt_raw = tokenizer.apply_chat_template(
            conversation=conversation,
            messages=messages,
            **kwargs,
        )

        prompt = (
            TextPrompt(prompt=prompt_raw)
            if isinstance(prompt_raw, str)
            else TokensPrompt(prompt_token_ids=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  # type: ignore[return-value]

    async def render_messages_async(
        self,
        messages: list[ChatCompletionMessageParam],
        **kwargs,
    ) -> tuple[list[ConversationMessage], TextPrompt | TokensPrompt]:
        tokenizer = self.get_tokenizer()
        conversation, mm_data, mm_uuids = await parse_chat_messages_async(
            messages,
            self.config,
            content_format="string",
        )

        prompt_raw = tokenizer.apply_chat_template(
            conversation=conversation,
            messages=messages,
            **kwargs,
        )

        prompt = (
            TextPrompt(prompt=prompt_raw)
            if isinstance(prompt_raw, str)
            else TokensPrompt(prompt_token_ids=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  # type: ignore[return-value]