OpenCompass/opencompass/models/huggingface.py

import os
from typing import Dict, List, Optional, Union

import numpy as np
import torch

from opencompass.models.base import BaseModel
from opencompass.registry import MODELS
from opencompass.utils.logging import get_logger
from opencompass.utils.prompt import PromptList

PromptType = Union[PromptList, str]


@MODELS.register_module()
class HuggingFace(BaseModel):
    """Model wrapper around HuggingFace models.

    Args:
        path (str): The name or path to HuggingFace's model.
        hf_cache_dir: Set the cache dir to HF model cache dir. If None, it will
            use the env variable HF_MODEL_HUB. Defaults to None.
        max_seq_len (int): The maximum length of the input sequence. Defaults
            to 2048.
        tokenizer_path (str): The path to the tokenizer. Defaults to None.
        tokenizer_kwargs (dict): Keyword arguments for the tokenizer.
            Defaults to {}.
        peft_path (str, optional): The name or path to the HuggingFace's PEFT
            model. If None, the original model will not be converted to PEFT.
            Defaults to None.
        tokenizer_only (bool): If True, only the tokenizer will be initialized.
            Defaults to False.
        model_kwargs (dict): Keyword arguments for the model, used in loader.
            Defaults to dict(device_map='auto').
        meta_template (Dict, optional): The model's meta prompt
            template if needed, in case the requirement of injecting or
            wrapping of any meta instructions.
        extract_pred_after_decode (bool): Whether to extract the prediction
            string from the decoded output string, instead of extract the
            prediction tokens before decoding. Defaults to False.
        batch_padding (bool): If False, inference with be performed in for-loop
            without batch padding.
        pad_token_id (int): The id of the padding token. Defaults to None. Use
            (#vocab + pad_token_id) if get negative value.
        mode (str, optional): The method of input truncation when input length
            exceeds max_seq_len. 'mid' represents the part of input to
            truncate. Defaults to 'none'.

    Note:
        About ``extract_pred_after_decode``: Commonly, we should extract the
        the prediction tokens before decoding. But for some tokenizers using
        ``sentencepiece``, like LLaMA,  this behavior may change the number of
        whitespaces, which is harmful for Python programming tasks.
    """

    def __init__(self,
                 path: str,
                 hf_cache_dir: Optional[str] = None,
                 max_seq_len: int = 2048,
                 tokenizer_path: Optional[str] = None,
                 tokenizer_kwargs: dict = dict(),
                 peft_path: Optional[str] = None,
                 tokenizer_only: bool = False,
                 model_kwargs: dict = dict(device_map='auto'),
                 meta_template: Optional[Dict] = None,
                 extract_pred_after_decode: bool = False,
                 batch_padding: bool = False,
                 pad_token_id: Optional[int] = None,
                 mode: str = 'none'):
        super().__init__(path=path,
                         max_seq_len=max_seq_len,
                         tokenizer_only=tokenizer_only,
                         meta_template=meta_template)
        from opencompass.utils.fileio import patch_hf_auto_model
        if hf_cache_dir is None:
            hf_cache_dir = os.getenv('HF_MODEL_HUB', None)
        patch_hf_auto_model(hf_cache_dir)
        self.logger = get_logger()
        self.pad_token_id = pad_token_id
        assert mode in ['none', 'mid']
        self.mode = mode
        self._load_tokenizer(path=path,
                             tokenizer_path=tokenizer_path,
                             tokenizer_kwargs=tokenizer_kwargs)
        self.batch_padding = batch_padding
        self.extract_pred_after_decode = extract_pred_after_decode
        if not tokenizer_only:
            self._load_model(path=path,
                             model_kwargs=model_kwargs,
                             peft_path=peft_path)

    def _load_tokenizer(self, path: str, tokenizer_path: Optional[str],
                        tokenizer_kwargs: dict):
        from transformers import AutoTokenizer
        self.tokenizer = AutoTokenizer.from_pretrained(
            tokenizer_path if tokenizer_path else path, **tokenizer_kwargs)

        # A patch for some models without pad_token_id
        if self.pad_token_id is not None:
            if self.pad_token_id < 0:
                self.pad_token_id += self.tokenizer.vocab_size
            if self.tokenizer.pad_token_id is None:
                self.logger.warning(
                    f'Using {self.pad_token_id} as pad_token_id')
            elif self.tokenizer.pad_token_id != self.pad_token_id:
                self.logger.warning(
                    f'pad_token_id is not consistent with the tokenizer. Using {self.pad_token_id} as pad_token_id'  # noqa
                )
            self.tokenizer.pad_token_id = self.pad_token_id
        elif self.tokenizer.pad_token_id is None:
            self.logger.warning('pad_token_id is not set for the tokenizer.')
            if self.tokenizer.eos_token is not None:
                self.logger.warning('Using eos_token_id as pad_token_id.')
                self.logger.warning(
                    f'{self.tokenizer.eos_token} la {self.tokenizer.eos_token is None}'  # noqa
                )
                self.tokenizer.pad_token = self.tokenizer.eos_token
            else:
                raise ValueError(
                    'pad_token_id is not set for this tokenizer. Try to set pad_token_id via passing `pad_token_id={PAD_TOKEN_ID}` in model_cfg. You may find pad_token_id in `generation.json`'  # noqa
                )

        # A patch for llama when batch_padding = True
        if 'decapoda-research/llama' in path or \
                (tokenizer_path and
                 'decapoda-research/llama' in tokenizer_path):
            self.logger.warning('We set new pad_token_id for LLaMA model')
            # keep consistent with official LLaMA repo
            # https://github.com/google/sentencepiece/blob/master/python/sentencepiece_python_module_example.ipynb  # noqa
            self.tokenizer.bos_token = '<s>'
            self.tokenizer.eos_token = '</s>'
            self.tokenizer.pad_token_id = 0

    def _load_model(self,
                    path: str,
                    model_kwargs: dict,
                    peft_path: Optional[str] = None):
        from transformers import AutoModel, AutoModelForCausalLM

        model_kwargs.setdefault('torch_dtype', torch.float16)
        try:
            self.model = AutoModelForCausalLM.from_pretrained(
                path, **model_kwargs)
        except ValueError:
            self.model = AutoModel.from_pretrained(path, **model_kwargs)

        if peft_path is not None:
            from peft import PeftModel
            self.model = PeftModel.from_pretrained(self.model,
                                                   peft_path,
                                                   is_trainable=False)
        self.model.eval()

        # A patch for llama when batch_padding = True
        if 'decapoda-research/llama' in path:
            self.model.config.bos_token_id = 1
            self.model.config.eos_token_id = 2
            self.model.config.pad_token_id = self.tokenizer.pad_token_id

    def generate(self, inputs: List[str], max_out_len: int,
                 **kwargs) -> List[str]:
        """Generate results given a list of inputs.

        Args:
            inputs (List[str]): A list of strings.
            max_out_len (int): The maximum length of the output.

        Returns:
            List[str]: A list of generated strings.
        """
        if self.batch_padding and len(inputs) > 1:
            return self._batch_generate(inputs=inputs,
                                        max_out_len=max_out_len,
                                        **kwargs)
        else:
            return sum((self._single_generate(
                inputs=[input_], max_out_len=max_out_len, **kwargs)
                        for input_ in inputs), [])

    def _batch_generate(self, inputs: List[str], max_out_len: int,
                        **kwargs) -> List[str]:
        """Support for batch prompts inference.

        Args:
            inputs (List[str]): A list of strings.
            max_out_len (int): The maximum length of the output.

        Returns:
            List[str]: A list of generated strings.
        """
        if self.extract_pred_after_decode:
            prompt_lens = [len(input_) for input_ in inputs]

        # step-1: tokenize the input with batch_encode_plus
        tokens = self.tokenizer.batch_encode_plus(inputs,
                                                  padding=True,
                                                  truncation=True,
                                                  max_length=self.max_seq_len -
                                                  max_out_len)
        tokens = {
            k: torch.tensor(np.array(tokens[k]), device=self.model.device)
            for k in tokens if k in ['input_ids', 'attention_mask']
        }

        # step-2: conduct model forward to generate output
        outputs = self.model.generate(**tokens,
                                      max_new_tokens=max_out_len,
                                      **kwargs)

        if not self.extract_pred_after_decode:
            outputs = outputs[:, tokens['input_ids'].shape[1]:]

        decodeds = self.tokenizer.batch_decode(outputs,
                                               skip_special_tokens=True)

        if self.extract_pred_after_decode:
            decodeds = [
                token[len_:] for token, len_ in zip(decodeds, prompt_lens)
            ]

        return decodeds

    def _single_generate(self, inputs: List[str], max_out_len: int,
                         **kwargs) -> List[str]:
        """Support for single prompt inference.

        Args:
            inputs (List[str]): A list of strings.
            max_out_len (int): The maximum length of the output.

        Returns:
            List[str]: A list of generated strings.
        """
        if self.extract_pred_after_decode:
            prompt_lens = [len(input_) for input_ in inputs]

        if self.mode == 'mid':
            input_ids = self.tokenizer(inputs, truncation=False)['input_ids']
            input_ids = torch.tensor(input_ids, device=self.model.device)
            if len(input_ids[0]) > self.max_seq_len - max_out_len:
                half = int((self.max_seq_len - max_out_len) / 2)
                inputs = [
                    self.tokenizer.decode(input_ids[0][:half],
                                          skip_special_tokens=True) +
                    self.tokenizer.decode(input_ids[0][-half:],
                                          skip_special_tokens=True)
                ]

        input_ids = self.tokenizer(inputs,
                                   truncation=True,
                                   max_length=self.max_seq_len -
                                   max_out_len)['input_ids']
        input_ids = torch.tensor(input_ids, device=self.model.device)
        # To accommodate the PeftModel, parameters should be passed in
        # key-value format for generate.
        outputs = self.model.generate(input_ids=input_ids,
                                      max_new_tokens=max_out_len,
                                      **kwargs)

        if not self.extract_pred_after_decode:
            outputs = outputs[:, input_ids.shape[1]:]

        decodeds = self.tokenizer.batch_decode(outputs,
                                               skip_special_tokens=True)

        if self.extract_pred_after_decode:
            decodeds = [
                token[len_:] for token, len_ in zip(decodeds, prompt_lens)
            ]

        return decodeds

    def get_logits(self, inputs: List[str]):

        if self.batch_padding and len(inputs) > 1:
            # batch inference
            tokens = self.tokenizer(inputs,
                                    padding=True,
                                    truncation=True,
                                    max_length=self.max_seq_len)

            tokens = {
                k: torch.tensor(np.array(tokens[k]), device=self.model.device)
                for k in tokens if k in ['input_ids', 'attention_mask']
            }
            outputs = self.model(**tokens)

        else:
            input_ids = self.tokenizer(
                inputs,
                padding=False,
                truncation=True,
                max_length=self.max_seq_len)['input_ids']
            input_ids = torch.tensor(input_ids, device=self.model.device)
            tokens = {'input_ids': input_ids}

            outputs = self.model(input_ids)
        return outputs[0], {'tokens': tokens}

    def get_ppl(self,
                inputs: List[str],
                mask_length: Optional[List[int]] = None) -> List[float]:
        """Get perplexity scores given a list of inputs.

        Args:
            inputs (List[str]): A list of strings.
            mask_length (Optional[List[int]]): A list of mask lengths. If
                provided, the perplexity scores will be calculated with the
                first mask_length[i] tokens masked out. It's okay to skip
                its implementation if advanced features in PPLInfernecer is
                not needed.

        Returns:
            List[float]: A list of perplexity scores.
        """

        if self.batch_padding and len(inputs) > 1:
            assert self.tokenizer.pad_token
            return self._get_ppl(inputs, mask_length=mask_length)
        else:
            return np.concatenate([
                self._get_ppl(inputs=[text], mask_length=mask_length)
                for text in inputs
            ])

    def _get_ppl(self,
                 inputs: List[str],
                 mask_length: Optional[List[int]] = None) -> List[float]:
        """Get perplexity scores given a list of inputs.

        Args:
            inputs (List[str]): A list of strings.
            mask_length (Optional[List[int]]): A list of mask lengths. If
                provided, the perplexity scores will be calculated with the
                first mask_length[i] tokens masked out. It's okay to skip
                its implementation if advanced features in PPLInfernecer is
                not needed.

        Returns:
            List[float]: A list of perplexity scores.
        """

        outputs, inputs = self.get_logits(inputs)
        shift_logits = outputs[..., :-1, :].contiguous().float()

        shift_labels = inputs['tokens']['input_ids'][..., 1:].contiguous()

        loss_fct = torch.nn.CrossEntropyLoss(
            reduction='none', ignore_index=self.tokenizer.pad_token_id)
        loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)),
                        shift_labels.view(-1)).view(shift_labels.size())

        if mask_length is not None:
            mask = torch.zeros_like(shift_labels)  # [batch,seqlen]
            for i in range(len(mask)):
                for j in range(mask_length[i] - 1, len(mask[i])):
                    mask[i][j] = 1
            loss = loss * mask

        lens = (inputs['tokens']['input_ids'] !=
                self.tokenizer.pad_token_id).sum(-1).cpu().numpy()
        if mask_length is not None:
            lens -= np.array(mask_length)
        ce_loss = loss.sum(-1).cpu().detach().numpy() / lens
        return ce_loss

    def get_token_len(self, prompt: str) -> int:
        """Get lengths of the tokenized strings.

        Args:
            prompt (str): Input string.

        Returns:
            int: Length of the input tokens
        """
        return len(self.tokenizer.encode(prompt))


@MODELS.register_module()
class HuggingFaceCausalLM(HuggingFace):
    """Model wrapper around HuggingFace CausalLM.

    Args:
        path (str): The name or path to HuggingFace's model.
        hf_cache_dir: Set the cache dir to HF model cache dir. If None, it will
            use the env variable HF_MODEL_HUB. Defaults to None.
        max_seq_len (int): The maximum length of the input sequence. Defaults
            to 2048.
        tokenizer_path (str): The path to the tokenizer. Defaults to None.
        tokenizer_kwargs (dict): Keyword arguments for the tokenizer.
            Defaults to {}.
        peft_path (str, optional): The name or path to the HuggingFace's PEFT
            model. If None, the original model will not be converted to PEFT.
            Defaults to None.
        tokenizer_only (bool): If True, only the tokenizer will be initialized.
            Defaults to False.
        model_kwargs (dict): Keyword arguments for the model, used in loader.
            Defaults to dict(device_map='auto').
        meta_template (Dict, optional): The model's meta prompt
            template if needed, in case the requirement of injecting or
            wrapping of any meta instructions.
        batch_padding (bool): If False, inference with be performed in for-loop
            without batch padding.
    """

    def _load_model(self,
                    path: str,
                    model_kwargs: dict,
                    peft_path: Optional[str] = None):
        from transformers import AutoModelForCausalLM

        model_kwargs.setdefault('torch_dtype', torch.float16)
        self.model = AutoModelForCausalLM.from_pretrained(path, **model_kwargs)
        if peft_path is not None:
            from peft import PeftModel
            self.model = PeftModel.from_pretrained(self.model,
                                                   peft_path,
                                                   is_trainable=False)
        self.model.eval()