zai-org/glm-2b

1import os2from typing import Optional, Tuple, List, Union3from shutil import copyfile4import torch5 6from transformers import PreTrainedTokenizer, RobertaTokenizer, GPT2Tokenizer, BertTokenizer7from transformers.utils import logging8from transformers.tokenization_utils_base import BatchEncoding9from transformers.models.auto.tokenization_auto import get_tokenizer_config10from transformers.utils.generic import _is_torch_device11import sentencepiece as spm12 13logger = logging.get_logger(__name__)14 15 16class GLMBatchEncoding(BatchEncoding):17    def to(self, device: Union[str, "torch.device"]) -> "BatchEncoding":18        """19        Send all values to device by calling `v.to(device)` (PyTorch only).20 21        Args:22            device (`str` or `torch.device`): The device to put the tensors on.23 24        Returns:25            [`BatchEncoding`]: The same instance after modification.26        """27 28        # This check catches things like APEX blindly calling "to" on all inputs to a module29        # Otherwise it passes the casts down and casts the LongTensor containing the token idxs30        # into a HalfTensor31        if isinstance(device, str) or _is_torch_device(device) or isinstance(device, int):32            self.data = {k: v.to(device=device) if torch.is_tensor(v) else v for k, v in self.data.items()}33        else:34            logger.warning(f"Attempting to cast a BatchEncoding to type {str(device)}. This is not supported.")35        return self36 37 38class GLMTokenizerMixin:39    @property40    def sop_token(self) -> Optional[str]:41        return "<|startofpiece|>"42 43    @property44    def sop_token_id(self) -> Optional[int]:45        """46        `Optional[int]`: Id of the start token in the vocabulary, used when training a model with autoregressive blank filling.47        """48        return self.convert_tokens_to_ids(self.sop_token)49 50    @property51    def eop_token(self) -> Optional[str]:52        return "<|endofpiece|>"53 54    @property55    def eop_token_id(self) -> Optional[int]:56        """57        `Optional[int]`: Id of the end token in the vocabulary, used when training a model with autoregressive blank filling.58        """59        return self.convert_tokens_to_ids(self.eop_token)60 61    @property62    def gmask_token_id(self) -> int:63        return self.convert_tokens_to_ids("[gMASK]")64 65    @property66    def smask_token_id(self) -> int:67        return self.convert_tokens_to_ids("[sMASK]")68 69    @property70    def mask_token_ids(self):71        return [self.mask_token_id, self.smask_token_id, self.gmask_token_id]72 73    def _build_input_for_multiple_choice(self, context, choices):74        context_id = context["input_ids"]75        if torch.is_tensor(context_id):76            context_id = context_id.tolist()77 78        division = len(context_id)79        mask_position = context_id.index(self.mask_token_id)80 81        token = torch.tensor(context_id, dtype=torch.long)82        attention_mask = [context["attention_mask"].expand(division, -1)]83        position_id = torch.arange(division, dtype=torch.long)84        block_position_id = torch.zeros(division, dtype=torch.long)85 86        choice_ids, choice_indices = [], []87 88        for choice_str in choices:89            choice = torch.tensor(self(choice_str, add_special_tokens=False, padding=False)['input_ids'],90                                  dtype=torch.long)91            choice_ids.append(choice)92            choice_indices.append(torch.arange(len(token), len(token) + len(choice), dtype=torch.long))93            attention_mask.append(torch.tril(torch.ones((len(choice), len(choice)), dtype=torch.long)))94 95            token = torch.cat((token, torch.tensor([self.sop_token_id], dtype=torch.long), choice[:-1]))96            position_id = torch.cat((position_id, torch.tensor([mask_position] * len(choice), dtype=torch.long)))97            block_position_id = torch.cat((block_position_id, torch.arange(1, 1 + len(choice), dtype=torch.long)))98 99        attention_mask = torch.block_diag(*attention_mask)100        attention_mask[division:, :division] = context["attention_mask"].unsqueeze(0)101 102        return {103            "input_ids": token,104            "position_ids": torch.stack((position_id, block_position_id)),105            "attention_mask": attention_mask,106            "choice_ids": choice_ids,107            "choice_indices": choice_indices108        }109 110    def _pad_batch(self, tokens, position_ids, attention_mask, max_seq_length):111        pad_length = max_seq_length - len(tokens)112        attention_mask = torch.nn.functional.pad(113            attention_mask,114            (0, pad_length, 0, pad_length),115            mode="constant",116            value=0,117        )118        tokens = torch.cat((tokens, torch.zeros(pad_length, dtype=torch.long)))119        position_ids = torch.cat((position_ids, position_ids[..., -1:].expand(-1, pad_length)), dim=-1)120        return tokens, position_ids, attention_mask121 122    def _collate(self, samples):123        TILE = 1124        length_to_pad = (max(map(lambda spl: len(spl["input_ids"]), samples)) + TILE - 1) // TILE * TILE125 126        token_batch, position_id_batch, attention_mask_batch = [], [], []127        choices_batch, choice_target_ids_batch = [], []128 129        for sample in samples:130            token, position_id, attention_mask = self._pad_batch(131                sample["input_ids"], sample["position_ids"], sample["attention_mask"], length_to_pad132            )133            token_batch.append(token)134            position_id_batch.append(position_id)135            attention_mask_batch.append(attention_mask)136            choices_batch.append(sample["choice_ids"])137            choice_target_ids_batch.append(sample["choice_indices"])138        return {139            "input_ids": torch.stack(token_batch),140            "position_ids": torch.stack(position_id_batch),141            "attention_mask": torch.stack(attention_mask_batch).unsqueeze(1),142            "choice_ids": choices_batch,143            "choice_indices": choice_target_ids_batch,144        }145 146    def build_inputs_for_multiple_choice(self, model_input: BatchEncoding, choices, max_length=None):147        samples = [{key: value[i] for key, value in model_input.items()} for i in range(len(model_input["input_ids"]))]148        samples = [self._build_input_for_multiple_choice(sample, choice) for sample, choice in149                   zip(samples, choices)]150        inputs = self._collate(samples)151        return GLMBatchEncoding(inputs)152 153    def build_inputs_for_generation(self, model_input: BatchEncoding, max_gen_length=512, targets=None, padding=False):154        mask_ids = self.mask_token_ids155        input_ids = model_input.input_ids156        batch_size, seq_length = input_ids.shape[:2]157        position_id, block_position_id = list(range(seq_length)), [0 for _ in range(seq_length)]158        position_ids, block_position_ids = [], []159        labels = None160        if targets is not None:161            is_batched = isinstance(targets, (list, tuple))162            targets = self(targets, add_special_tokens=False, padding=False).input_ids163            if not is_batched:164                targets = [targets]165            assert len(targets) == len(input_ids)166            targets = [(target + [self.eop_token_id])[:max_gen_length] for target in targets]167            if not padding:168                max_gen_length = max(map(len, targets))169            targets = [[self.sop_token_id] + target for target in targets]170            labels = [target[1:] for target in targets]171            targets = [target + [self.pad_token_id] * (max_gen_length + 1 - len(target)) for target in targets]172            labels = [label + [-100] * (max_gen_length - len(label)) for label in labels]173            targets = torch.tensor(targets, dtype=input_ids.dtype, device=input_ids.device)174            labels = torch.tensor(labels, dtype=input_ids.dtype, device=input_ids.device)175            labels = torch.cat((input_ids.new_full((batch_size, seq_length), -100), labels), dim=1)176        for i in range(batch_size):177            mask_positions = []178            for mask_id in mask_ids:179                mask_positions += (input_ids[i] == mask_id).nonzero(as_tuple=True)[0].tolist()180            if not mask_positions:181                raise ValueError("Cannot find mask token in the input")182            mask_positions.sort()183            mask_pos = mask_positions[0]184            position_ids.append(position_id + [mask_pos] * max_gen_length)185            block_position_ids.append(block_position_id + list(range(1, max_gen_length + 1)))186        position_ids = torch.tensor(position_ids, dtype=input_ids.dtype, device=input_ids.device)187        block_position_ids = torch.tensor(block_position_ids, dtype=input_ids.dtype, device=input_ids.device)188        position_ids = torch.stack((position_ids, block_position_ids), dim=1)189        attention_mask = model_input.attention_mask190        attention_mask = attention_mask.unsqueeze(1).expand(-1, seq_length + max_gen_length, -1)191        generation_attention_mask = torch.cat([attention_mask.new_zeros((seq_length, max_gen_length)),192                                               torch.tril(attention_mask.new_ones((max_gen_length, max_gen_length)))],193                                              dim=0).unsqueeze(0).expand(batch_size, -1, -1)194        attention_mask = torch.cat((attention_mask, generation_attention_mask), dim=2)195        attention_mask = attention_mask.unsqueeze(1)196        if targets is None:197            input_ids = torch.cat((input_ids, input_ids.new_full((batch_size, 1), self.sop_token_id)), dim=-1)198        else:199            input_ids = torch.cat((input_ids, targets[:, :-1]), dim=1)200        batch = {"input_ids": input_ids, "position_ids": position_ids}201        if labels is None:202            batch["generation_attention_mask"] = attention_mask203        else:204            batch["attention_mask"] = attention_mask205            batch["labels"] = labels206        return BatchEncoding(batch)207 208 209class GLMRobertaTokenizer(RobertaTokenizer, GLMTokenizerMixin):210    model_input_names = ["input_ids", "position_ids", "attention_mask"]211    truncation_side: str = "left"212 213    @property214    def gmask_token_id(self) -> int:215        raise NotImplementedError("The model doesn't support gMASK")216 217    @property218    def smask_token_id(self) -> int:219        raise NotImplementedError("The model doesn't support sMASK")220 221    @property222    def mask_token_ids(self):223        return [self.mask_token_id]224 225 226class GLMChineseTokenizer(PreTrainedTokenizer, GLMTokenizerMixin):227    vocab_files_names = {"vocab_file": "cog-pretrain.model"}228    truncation_side: str = "left"229 230    def __init__(self, vocab_file, **kwargs):231        super().__init__(**kwargs)232        self.vocab_file = vocab_file233        self.sp_model = spm.SentencePieceProcessor()234        self.sp_model.Load(vocab_file)235 236    @property237    def vocab_size(self):238        return len(self.sp_model)239 240    def get_vocab(self):241        vocab = {self.convert_ids_to_tokens(i): i for i in range(self.vocab_size)}242        vocab.update(self.added_tokens_encoder)243        return vocab244 245    def _tokenize(self, text, **kwargs):246        return self.sp_model.encode(text, out_type=str)247 248    def _convert_token_to_id(self, token):249        """Converts a token (str) in an id using the vocab."""250        return self.sp_model.PieceToId(token)251 252    def _convert_id_to_token(self, index):253        """Converts an index (integer) in a token (str) using the vocab."""254        return self.sp_model.IdToPiece(index)255 256    def convert_tokens_to_string(self, tokens):257        return self.sp_model.decode(tokens)258 259    def save_vocabulary(self, save_directory: str, filename_prefix: Optional[str] = None) -> Tuple[str]:260        if not os.path.isdir(save_directory):261            logger.error(f"Vocabulary path ({save_directory}) should be a directory")262            return263        out_vocab_file = os.path.join(264            save_directory, (filename_prefix + "-" if filename_prefix else "") + self.vocab_files_names["vocab_file"]265        )266 267        if os.path.abspath(self.vocab_file) != os.path.abspath(out_vocab_file) and os.path.isfile(self.vocab_file):268            copyfile(self.vocab_file, out_vocab_file)269        elif not os.path.isfile(self.vocab_file):270            with open(out_vocab_file, "wb") as fi:271                content_spiece_model = self.sp_model.serialized_model_proto()272                fi.write(content_spiece_model)273 274        return (out_vocab_file,)275 276    def build_inputs_with_special_tokens(277            self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None278    ) -> List[int]:279        """280        Build model inputs from a sequence or a pair of sequence for sequence classification tasks by concatenating and281        adding special tokens. A BERT sequence has the following format:282 283        - single sequence: ``[CLS] X [SEP]``284        - pair of sequences: ``[CLS] A [SEP] B [SEP]``285 286        Args:287            token_ids_0 (:obj:`List[int]`):288                List of IDs to which the special tokens will be added.289            token_ids_1 (:obj:`List[int]`, `optional`):290                Optional second list of IDs for sequence pairs.291 292        Returns:293            :obj:`List[int]`: List of `input IDs <../glossary.html#input-ids>`__ with the appropriate special tokens.294        """295        assert token_ids_1 is None296        cls = [self.cls_token_id]297        eos = [self.eos_token_id]298        return cls + token_ids_0 + eos299 300 301class GLMGPT2Tokenizer(GPT2Tokenizer, GLMTokenizerMixin):302    model_input_names = ["input_ids", "position_ids", "attention_mask"]303    truncation_side: str = "left"304 305    def build_inputs_with_special_tokens(306            self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None307    ) -> List[int]:308        """309        Build model inputs from a sequence or a pair of sequence for sequence classification tasks by concatenating and310        adding special tokens. A BERT sequence has the following format:311 312        - single sequence: ``[CLS] X [SEP]``313        - pair of sequences: ``[CLS] A [SEP] B [SEP]``314 315        Args:316            token_ids_0 (:obj:`List[int]`):317                List of IDs to which the special tokens will be added.318            token_ids_1 (:obj:`List[int]`, `optional`):319                Optional second list of IDs for sequence pairs.320 321        Returns:322            :obj:`List[int]`: List of `input IDs <../glossary.html#input-ids>`__ with the appropriate special tokens.323        """324        assert token_ids_1 is None325        cls = [self.cls_token_id]326        eos = [self.eos_token_id]327        return cls + token_ids_0 + eos328 329 330class GLMBertTokenizer(BertTokenizer, GLMTokenizerMixin):331    model_input_names = ["input_ids", "position_ids", "attention_mask"]332    truncation_side: str = "left"333 334    @property335    def gmask_token_id(self) -> int:336        raise NotImplementedError("The model doesn't support gMASK")337 338    @property339    def smask_token_id(self) -> int:340        raise NotImplementedError("The model doesn't support sMASK")341 342    @property343    def mask_token_ids(self):344        return [self.mask_token_id]345 346 347class GLMTokenizer:348    @classmethod349    def from_pretrained(cls, pretrained_model_name_or_path, *inputs, **kwargs):350        tokenizer_config = get_tokenizer_config(pretrained_model_name_or_path, **kwargs)351        config_tokenizer_class = tokenizer_config.get("tokenizer_class")352        if config_tokenizer_class == "GLMRobertaTokenizer":353            tokenizer_class = GLMRobertaTokenizer354        elif config_tokenizer_class == "GLMChineseTokenizer":355            tokenizer_class = GLMChineseTokenizer356        elif config_tokenizer_class == "GLMGPT2Tokenizer":357            tokenizer_class = GLMGPT2Tokenizer358        elif config_tokenizer_class == "GLMBertTokenizer":359            tokenizer_class = GLMBertTokenizer360        else:361            raise NotImplementedError("Not implemented tokenizer type:", config_tokenizer_class)362        return tokenizer_class.from_pretrained(pretrained_model_name_or_path, *inputs, **kwargs)363