zai-org/codegeex2-6b

1import os2import torch3from typing import List, Optional, Union, Dict4from sentencepiece import SentencePieceProcessor5from transformers import PreTrainedTokenizer6from transformers.utils import logging, PaddingStrategy7from transformers.tokenization_utils_base import EncodedInput, BatchEncoding8 9 10class SPTokenizer:11    def __init__(self, model_path: str):12        # reload tokenizer13        assert os.path.isfile(model_path), model_path14        self.sp_model = SentencePieceProcessor(model_file=model_path)15 16        # BOS / EOS token IDs17        self.n_words: int = self.sp_model.vocab_size()18        self.bos_id: int = self.sp_model.bos_id()19        self.eos_id: int = self.sp_model.eos_id()20        self.pad_id: int = self.sp_model.unk_id()21        assert self.sp_model.vocab_size() == self.sp_model.get_piece_size()22 23        special_tokens = ["[MASK]", "[gMASK]", "[sMASK]", "sop", "eop"]24        self.special_tokens = {}25        self.index_special_tokens = {}26        for token in special_tokens:27            self.special_tokens[token] = self.n_words28            self.index_special_tokens[self.n_words] = token29            self.n_words += 130 31    def tokenize(self, s: str):32        return self.sp_model.EncodeAsPieces(s)33 34    def encode(self, s: str, bos: bool = False, eos: bool = False) -> List[int]:35        assert type(s) is str36        t = self.sp_model.encode(s)37        if bos:38            t = [self.bos_id] + t39        if eos:40            t = t + [self.eos_id]41        return t42 43    def decode(self, t: List[int]) -> str:44        return self.sp_model.decode(t)45 46    def decode_tokens(self, tokens: List[str]) -> str:47        text = self.sp_model.DecodePieces(tokens)48        return text49 50    def convert_token_to_id(self, token):51        """ Converts a token (str) in an id using the vocab. """52        if token in self.special_tokens:53            return self.special_tokens[token]54        return self.sp_model.PieceToId(token)55 56    def convert_id_to_token(self, index):57        """Converts an index (integer) in a token (str) using the vocab."""58        if index in self.index_special_tokens or index in [self.eos_id, self.bos_id, self.pad_id] or index < 0:59            return ""60        return self.sp_model.IdToPiece(index)61 62 63class ChatGLMTokenizer(PreTrainedTokenizer):64    vocab_files_names = {"vocab_file": "tokenizer.model"}65 66    model_input_names = ["input_ids", "attention_mask", "position_ids"]67 68    def __init__(self, vocab_file, padding_side="left", clean_up_tokenization_spaces=False, **kwargs):69        self.name = "GLMTokenizer"70 71        self.vocab_file = vocab_file72        self.tokenizer = SPTokenizer(vocab_file)73        self.special_tokens = {74            "<bos>": self.tokenizer.bos_id,75            "<eos>": self.tokenizer.eos_id,76            "<pad>": self.tokenizer.pad_id77        }78        super().__init__(padding_side=padding_side, clean_up_tokenization_spaces=clean_up_tokenization_spaces, **kwargs)79    def get_command(self, token):80        if token in self.special_tokens:81            return self.special_tokens[token]82        assert token in self.tokenizer.special_tokens, f"{token} is not a special token for {self.name}"83        return self.tokenizer.special_tokens[token]84 85    @property86    def unk_token(self) -> str:87        return "<unk>"88 89    @property90    def pad_token(self) -> str:91        return "<unk>"92 93    @property94    def pad_token_id(self):95        return self.get_command("<pad>")96 97    @property98    def eos_token(self) -> str:99        return "</s>"100 101    @property102    def eos_token_id(self):103        return self.get_command("<eos>")104 105    @property106    def vocab_size(self):107        return self.tokenizer.n_words108 109    def get_vocab(self):110        """ Returns vocab as a dict """111        vocab = {self._convert_id_to_token(i): i for i in range(self.vocab_size)}112        vocab.update(self.added_tokens_encoder)113        return vocab114 115    def _tokenize(self, text, **kwargs):116        return self.tokenizer.tokenize(text)117 118    def _convert_token_to_id(self, token):119        """ Converts a token (str) in an id using the vocab. """120        return self.tokenizer.convert_token_to_id(token)121 122    def _convert_id_to_token(self, index):123        """Converts an index (integer) in a token (str) using the vocab."""124        return self.tokenizer.convert_id_to_token(index)125 126    def convert_tokens_to_string(self, tokens: List[str]) -> str:127        return self.tokenizer.decode_tokens(tokens)128 129    def save_vocabulary(self, save_directory, filename_prefix=None):130        """131        Save the vocabulary and special tokens file to a directory.132 133        Args:134            save_directory (`str`):135                The directory in which to save the vocabulary.136            filename_prefix (`str`, *optional*):137                An optional prefix to add to the named of the saved files.138 139        Returns:140            `Tuple(str)`: Paths to the files saved.141        """142        if os.path.isdir(save_directory):143            vocab_file = os.path.join(144                save_directory, self.vocab_files_names["vocab_file"]145            )146        else:147            vocab_file = save_directory148 149        with open(self.vocab_file, 'rb') as fin:150            proto_str = fin.read()151 152        with open(vocab_file, "wb") as writer:153            writer.write(proto_str)154 155        return (vocab_file,)156 157    def get_prefix_tokens(self):158        prefix_tokens = [self.get_command("[gMASK]"), self.get_command("sop")]159        return prefix_tokens160 161    def build_prompt(self, query, history=None):162        if history is None:163            history = []164        prompt = ""165        for i, (old_query, response) in enumerate(history):166            prompt += "[Round {}]\n\n问：{}\n\n答：{}\n\n".format(i + 1, old_query, response)167        prompt += "[Round {}]\n\n问：{}\n\n答：".format(len(history) + 1, query)168        return prompt169 170    def build_inputs_with_special_tokens(171            self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None172    ) -> List[int]:173        """174        Build model inputs from a sequence or a pair of sequence for sequence classification tasks by concatenating and175        adding special tokens. A BERT sequence has the following format:176 177        - single sequence: `[CLS] X [SEP]`178        - pair of sequences: `[CLS] A [SEP] B [SEP]`179 180        Args:181            token_ids_0 (`List[int]`):182                List of IDs to which the special tokens will be added.183            token_ids_1 (`List[int]`, *optional*):184                Optional second list of IDs for sequence pairs.185 186        Returns:187            `List[int]`: List of [input IDs](../glossary#input-ids) with the appropriate special tokens.188        """189        prefix_tokens = self.get_prefix_tokens()190        token_ids_0 = prefix_tokens + token_ids_0191        if token_ids_1 is not None:192            token_ids_0 = token_ids_0 + token_ids_1 + [self.get_command("<eos>")]193        return token_ids_0194 195    def _pad(196            self,197            encoded_inputs: Union[Dict[str, EncodedInput], BatchEncoding],198            max_length: Optional[int] = None,199            padding_strategy: PaddingStrategy = PaddingStrategy.DO_NOT_PAD,200            pad_to_multiple_of: Optional[int] = None,201            return_attention_mask: Optional[bool] = None,202    ) -> dict:203        """204        Pad encoded inputs (on left/right and up to predefined length or max length in the batch)205 206        Args:207            encoded_inputs:208                Dictionary of tokenized inputs (`List[int]`) or batch of tokenized inputs (`List[List[int]]`).209            max_length: maximum length of the returned list and optionally padding length (see below).210                Will truncate by taking into account the special tokens.211            padding_strategy: PaddingStrategy to use for padding.212 213                - PaddingStrategy.LONGEST Pad to the longest sequence in the batch214                - PaddingStrategy.MAX_LENGTH: Pad to the max length (default)215                - PaddingStrategy.DO_NOT_PAD: Do not pad216                The tokenizer padding sides are defined in self.padding_side:217 218                    - 'left': pads on the left of the sequences219                    - 'right': pads on the right of the sequences220            pad_to_multiple_of: (optional) Integer if set will pad the sequence to a multiple of the provided value.221                This is especially useful to enable the use of Tensor Core on NVIDIA hardware with compute capability222                `>= 7.5` (Volta).223            return_attention_mask:224                (optional) Set to False to avoid returning attention mask (default: set to model specifics)225        """226        # Load from model defaults227        # assert self.padding_side == "left"228 229        required_input = encoded_inputs[self.model_input_names[0]]230        seq_length = len(required_input)231 232        if padding_strategy == PaddingStrategy.LONGEST:233            max_length = len(required_input)234 235        if max_length is not None and pad_to_multiple_of is not None and (max_length % pad_to_multiple_of != 0):236            max_length = ((max_length // pad_to_multiple_of) + 1) * pad_to_multiple_of237 238        needs_to_be_padded = padding_strategy != PaddingStrategy.DO_NOT_PAD and len(required_input) != max_length239 240        # Initialize attention mask if not present.241        if "attention_mask" not in encoded_inputs:242            encoded_inputs["attention_mask"] = [1] * seq_length243 244        if "position_ids" not in encoded_inputs:245            encoded_inputs["position_ids"] = list(range(seq_length))246 247        if needs_to_be_padded:248            difference = max_length - len(required_input)249 250            if self.padding_side == "left":251                if "attention_mask" in encoded_inputs:252                    encoded_inputs["attention_mask"] = [0] * difference + encoded_inputs["attention_mask"]253                if "position_ids" in encoded_inputs:254                    encoded_inputs["position_ids"] = [0] * difference + encoded_inputs["position_ids"]255                encoded_inputs[self.model_input_names[0]] = [self.pad_token_id] * difference + required_input256            else:257                if "attention_mask" in encoded_inputs:258                    encoded_inputs["attention_mask"] = encoded_inputs["attention_mask"] + [0] * difference259                if "position_ids" in encoded_inputs:260                    encoded_inputs["position_ids"] = encoded_inputs["position_ids"] + [0] * difference261                encoded_inputs[self.model_input_names[0]] = required_input + [self.pad_token_id] * difference262 263        return encoded_inputs264