Skip to content

vllm.tokenizers.hf

HfTokenizer

Bases: TokenizerLike

Source code in vllm/tokenizers/hf.py 
 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
120
121
122
class HfTokenizer(TokenizerLike):
    @classmethod
    def from_pretrained(
        cls,
        path_or_repo_id: str | Path,
        *args,
        trust_remote_code: bool = False,
        revision: str | None = None,
        download_dir: str | None = None,
        **kwargs,
    ) -> "TokenizerLike":
        try:
            tokenizer = AutoTokenizer.from_pretrained(
                path_or_repo_id,
                *args,
                trust_remote_code=trust_remote_code,
                revision=revision,
                cache_dir=download_dir,
                **kwargs,
            )
        except ValueError as e:
            # If the error pertains to the tokenizer class not existing or not
            # currently being imported,
            # suggest using the --trust-remote-code flag.
            if not trust_remote_code and (
                "does not exist or is not currently imported." in str(e)
                or "requires you to execute the tokenizer file" in str(e)
            ):
                err_msg = (
                    "Failed to load the tokenizer. If the tokenizer "
                    "is a custom tokenizer not yet available in the "
                    "HuggingFace transformers library, consider "
                    "setting `trust_remote_code=True` in LLM or using "
                    "the `--trust-remote-code` flag in the CLI."
                )
                raise RuntimeError(err_msg) from e
            else:
                raise e

        # The special_tokens in tokenizer should also be
        # controlled by do_lower_case in encoder_config
        encoder_config = get_sentence_transformer_tokenizer_config(
            path_or_repo_id, revision
        )
        if isinstance(encoder_config, dict) and encoder_config.get(
            "do_lower_case", False
        ):
            special_tokens_map = {
                k: v.lower() for k, v in tokenizer.special_tokens_map.items()
            }
            tokenizer.add_special_tokens(special_tokens_map)

        return get_cached_tokenizer(tokenizer)

from_pretrained classmethod 

from_pretrained(
    path_or_repo_id: str | Path,
    *args,
    trust_remote_code: bool = False,
    revision: str | None = None,
    download_dir: str | None = None,
    **kwargs,
) -> TokenizerLike
Source code in vllm/tokenizers/hf.py 
 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
120
121
122
@classmethod
def from_pretrained(
    cls,
    path_or_repo_id: str | Path,
    *args,
    trust_remote_code: bool = False,
    revision: str | None = None,
    download_dir: str | None = None,
    **kwargs,
) -> "TokenizerLike":
    try:
        tokenizer = AutoTokenizer.from_pretrained(
            path_or_repo_id,
            *args,
            trust_remote_code=trust_remote_code,
            revision=revision,
            cache_dir=download_dir,
            **kwargs,
        )
    except ValueError as e:
        # If the error pertains to the tokenizer class not existing or not
        # currently being imported,
        # suggest using the --trust-remote-code flag.
        if not trust_remote_code and (
            "does not exist or is not currently imported." in str(e)
            or "requires you to execute the tokenizer file" in str(e)
        ):
            err_msg = (
                "Failed to load the tokenizer. If the tokenizer "
                "is a custom tokenizer not yet available in the "
                "HuggingFace transformers library, consider "
                "setting `trust_remote_code=True` in LLM or using "
                "the `--trust-remote-code` flag in the CLI."
            )
            raise RuntimeError(err_msg) from e
        else:
            raise e

    # The special_tokens in tokenizer should also be
    # controlled by do_lower_case in encoder_config
    encoder_config = get_sentence_transformer_tokenizer_config(
        path_or_repo_id, revision
    )
    if isinstance(encoder_config, dict) and encoder_config.get(
        "do_lower_case", False
    ):
        special_tokens_map = {
            k: v.lower() for k, v in tokenizer.special_tokens_map.items()
        }
        tokenizer.add_special_tokens(special_tokens_map)

    return get_cached_tokenizer(tokenizer)

get_cached_tokenizer 

get_cached_tokenizer(
    tokenizer: PreTrainedTokenizer
    | PreTrainedTokenizerFast,
) -> TokenizerLike

By default, transformers will recompute multiple tokenizer properties each time they are called, leading to a significant slowdown. This proxy caches these properties for faster access.

Source code in vllm/tokenizers/hf.py 
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
def get_cached_tokenizer(
    tokenizer: "PreTrainedTokenizer | PreTrainedTokenizerFast",
) -> TokenizerLike:
    """
    By default, transformers will recompute multiple tokenizer properties
    each time they are called, leading to a significant slowdown.
    This proxy caches these properties for faster access.
    """
    cached_tokenizer = copy.copy(tokenizer)

    tokenizer_all_special_ids = tokenizer.all_special_ids
    tokenizer_all_special_tokens = tokenizer.all_special_tokens
    tokenizer_vocab = tokenizer.get_vocab()
    tokenizer_len = len(tokenizer)

    max_token_id = max(tokenizer_vocab.values())
    # Some tokenizers (e.g., QwenTokenizer) have special tokens that
    # are added and included in the implementation of the vocab_size
    # property, but not in get_vocab(); if there is an implementation
    # of vocab size, we should take the greater value.
    if hasattr(tokenizer, "vocab_size"):
        with contextlib.suppress(NotImplementedError):
            max_token_id = max(max_token_id, tokenizer.vocab_size)

    class CachedTokenizer(tokenizer.__class__):  # type: ignore
        @property
        def all_special_ids(self) -> list[int]:
            return tokenizer_all_special_ids

        @property
        def all_special_tokens(self) -> list[str]:
            return tokenizer_all_special_tokens

        @property
        def max_token_id(self) -> int:
            return max_token_id

        def get_vocab(self) -> dict[str, int]:
            return tokenizer_vocab

        def __len__(self) -> int:
            return tokenizer_len

        def __reduce__(self):
            return get_cached_tokenizer, (tokenizer,)

    CachedTokenizer.__name__ = f"Cached{tokenizer.__class__.__name__}"

    cached_tokenizer.__class__ = CachedTokenizer
    return cached_tokenizer  # type: ignore