lightweight.py

import torch
import sys
import warnings
import numpy as np
from tqdm import trange
from typing import Any, List, Union, Tuple, Optional
from peft import PeftModel
from torch import Tensor
from transformers import AutoModelForCausalLM, AutoTokenizer

from FlagEmbedding.abc.inference import AbsReranker
from FlagEmbedding.inference.reranker.encoder_only.base import sigmoid


def last_logit_pool_lightweight(logits: Tensor,
                    attention_mask: Tensor) -> Tensor:
    """Pool the last logit.

    Args:
        logits (torch.Tensor): The output logits of the model.
        attention_mask (torch.Tensor): Attention mask.

    Returns:
        torch.Tensor: The tensor after pooling.
    """
    left_padding = (attention_mask[:, -1].sum() == attention_mask.shape[0])
    if left_padding:
        return logits[:, -1]
    else:
        sequence_lengths = attention_mask.sum(dim=1) - 1
        batch_size = logits.shape[0]
        return torch.stack([logits[i, sequence_lengths[i]] for i in range(batch_size)], dim=0)


class Collater_for_lightweight:
    """
    Collator of the lightweight LLM reranker.
    
    Args:
        tokenizer (transformers.AutoTokenizer): The tokenizer for reranker.
        max_len (int): Maximum length of tokens.
    """
    def __init__(self, tokenizer, max_len):
        self.tokenizer = tokenizer
        self.max_len = max_len
        self.pad_to_multiple_of = 8
        self.label_pad_token_id = -100
        warnings.filterwarnings("ignore",
                                message="`max_length` is ignored when `padding`=`True` and there is no truncation strategy.")

    def __call__(self, data):
        features = data[0]
        query_lengths = data[1]
        prompt_lengths = data[2]

        labels = [feature["labels"] for feature in features] if "labels" in features[0].keys() else None
        # We have to pad the labels before calling `tokenizer.pad` as this method won't pad them and needs them of the
        # same length to return tensors.
        if labels is not None:
            max_label_length = max(len(l) for l in labels)
            if self.pad_to_multiple_of is not None:
                max_label_length = (
                        (max_label_length + self.pad_to_multiple_of - 1)
                        // self.pad_to_multiple_of
                        * self.pad_to_multiple_of
                )

            padding_side = self.tokenizer.padding_side
            for feature in features:
                remainder = [self.label_pad_token_id] * (max_label_length - len(feature["labels"]))
                if isinstance(feature["labels"], list):
                    feature["labels"] = (
                        feature["labels"] + remainder if padding_side == "right" else remainder + feature["labels"]
                    )
                elif padding_side == "right":
                    feature["labels"] = np.concatenate([feature["labels"], remainder]).astype(np.int64)
                else:
                    feature["labels"] = np.concatenate([remainder, feature["labels"]]).astype(np.int64)

        collected = self.tokenizer.pad(
            features,
            padding=True,
            pad_to_multiple_of=8,
            return_tensors='pt',
        )

        return collected, query_lengths, prompt_lengths


class LightweightLLMReranker(AbsReranker):
    """Base reranker class for light weight LLM like decoder only models.

    Args:
        model_name_or_path (str): If it's a path to a local model, it loads the model from the path. Otherwise tries to download and
            load a model from HuggingFace Hub with the name.
        peft_path (Optional[str], optional): Path to the PEFT config. Defaults to :data:`None`.
        use_fp16 (bool, optional): If true, use half-precision floating-point to speed up computation with a slight performance 
            degradation. Defaults to :data:`False`. Defaults to :data:`False`.
        use_bf16 (bool, optional): Another type of half-precision floating-point, you can use bf16 if the hardware supports. 
            Defaults to :data:False.
        query_instruction_for_rerank (str, optional): Query instruction for retrieval tasks, which will be used with
            with :attr:`query_instruction_format`. Defaults to :data:`"A: "`.
        query_instruction_format (str, optional): The template for :attr:`query_instruction_for_rerank`. Defaults to :data:`"{}{}"`.
        passage_instruction_for_rerank (str, optional): Passage instruction for retrieval tasks, which will be used with
            with :attr:`passage_instruction_format`. Defaults to :data:`"B: "`.
        passage_instruction_format (str, optional): The template for passage. Defaults to "{}{}".
        cache_dir (Optional[str], optional): Cache directory for the model. Defaults to :data:`None`.
        trust_remote_code (bool, optional): trust_remote_code. Defaults to :data:`False`.
        devices (Union[str, List[str], List[int]], optional): Devices to use for model inference, such as ["cuda:0"] or ["0"].
            Defaults to :data:`None`.
        cutoff_layers (Optional[List[int]]): Pick which layers are used for computing the score. Defaults to :data:`None`.
        compress_layers (List[int], optional): Choose the layers to compress. Defaults to :data:`[8]`.
        compress_ratio (int, optional): Ratio to compress the selected layers, supported ratios: :data:`[1, 2, 4, 8]`. 
            Defaults to :data:`1`.
        prompt (Optional[str], optional): Prompt for the specific task. Defaults to :data:`None`.
        batch_size (int, optional): Batch size for inference. Defaults to :data:`128`.
        query_max_length (int, optional): Maximum length for queries. If not specified, will be 3/4 of :attr:`max_length`.
            Defaults to :data:`None`.
        max_length (int, optional): Maximum length of passages. Defaults to :data`512`.
        normalize (bool, optional): If True, use Sigmoid to normalize the results. Defaults to :data:`False`.
    """
    def __init__(
        self,
        model_name_or_path: str,
        peft_path: Optional[str] = None,
        use_fp16: bool = False,
        use_bf16: bool = False,
        query_instruction_for_rerank: str = "A: ",
        query_instruction_format: str = "{}{}", # specify the format of query_instruction_for_rerank
        passage_instruction_for_rerank: str = "B: ",
        passage_instruction_format: str = "{}{}", # specify the format of passage_instruction_for_rerank
        cache_dir: Optional[str] = None,
        trust_remote_code: bool = False,
        devices: Union[str, List[str], List[int]] = None, # specify devices, such as ["cuda:0"] or ["0"]
        # inference
        cutoff_layers: Optional[List[int]] = None,
        compress_layers: List[int] = [8],
        compress_ratio: int = 1,
        prompt: Optional[str] = None,
        batch_size: int = 128,
        query_max_length: Optional[int] = None,
        max_length: int = 512,
        normalize: bool = False,
        **kwargs: Any,
    ) -> None:
        try:
            from .models.gemma_model import CostWiseGemmaForCausalLM
        except:
            print('*') * 20
            print('*') * 20
            print('error for load lightweight reranker, please install transformers==4.46.0')
            print('*') * 20
            print('*') * 20
            sys.exit()

        super().__init__(
            model_name_or_path=model_name_or_path,
            use_fp16=use_fp16,
            query_instruction_for_rerank=query_instruction_for_rerank,
            query_instruction_format=query_instruction_format,
            passage_instruction_for_rerank=passage_instruction_for_rerank,
            passage_instruction_format=passage_instruction_format,
            devices=devices,
            batch_size=batch_size,
            query_max_length=query_max_length,
            max_length=max_length,
            normalize=normalize,
            **kwargs
        )

        self.cutoff_layers = cutoff_layers
        self.compress_layers = compress_layers
        self.compress_ratio = compress_ratio
        self.prompt = prompt

        self.tokenizer = AutoTokenizer.from_pretrained(
            model_name_or_path,
            cache_dir=cache_dir,
            trust_remote_code=trust_remote_code
        )
        self.tokenizer.padding_side = 'right'

        if use_bf16 is False and use_fp16 is False:
            warnings.warn("Due to model constraints, `use_bf16` and `use_fp16` cannot both be `False`. Here, `use_fp16` is set to `True` by default.", UserWarning)
            use_fp16 = True

        try:
            self.model = CostWiseGemmaForCausalLM.from_pretrained(
                model_name_or_path,
                cache_dir=cache_dir,
                trust_remote_code=trust_remote_code,
                torch_dtype=torch.bfloat16 if use_bf16 else torch.float32
            )
        except:
            self.model = AutoModelForCausalLM.from_pretrained(
                model_name_or_path,
                cache_dir=cache_dir,
                trust_remote_code=trust_remote_code,
                torch_dtype=torch.bfloat16 if use_bf16 else torch.float32
            )
        if peft_path:
            self.model = PeftModel.from_pretrained(self.model,peft_path)
            self.model = self.model.merge_and_unload()

    @torch.no_grad()
    def compute_score_single_gpu(
        self,
        sentence_pairs: Union[List[Tuple[str, str]], Tuple[str, str]],
        batch_size: Optional[int] = None,
        query_max_length: Optional[int] = None,
        max_length: Optional[int] = None,
        cutoff_layers: Optional[List[int]] = None,
        compress_layer: Optional[List[int]] = None,
        compress_layers: Optional[List[int]] = None,
        compress_ratio: Optional[int] = None,
        prompt: Optional[str] = None,
        normalize: Optional[bool] = None,
        device: Optional[str] = None,
        **kwargs: Any
    ) -> List[float]:
        """Compute the relevance scores using a single GPU.

        Args:
            sentence_pairs (Union[List[Tuple[str, str]], Tuple[str, str]]): Input sentence pairs to compute scores.
            batch_size (Optional[int], optional): Number of inputs for each iter. Defaults to :data:`None`.
            query_max_length (Optional[int], optional): Maximum length of tokens of queries. Defaults to :data:`None`.
            max_length (Optional[int], optional): Maximum length of tokens. Defaults to :data:`None`.
            cutoff_layers (Optional[List[int]], optional): Pick which layers are used for computing the score. Defaults to :data:`None`.
            compress_layer (Optional[List[int]]): Deprecated, use :attr:`compress_layers` instead. Defaults to :data:`None`.
            compress_layers (Optional[List[int]]): Selected layers to compress. Defaults to :data:`None`.
            compress_ratio (Optional[int]): Ratio to compress the selected layers, supported ratios: :data:`[1, 2, 4, 8]`. 
                Defaults to :data:`None`.
            prompt (Optional[str], optional): Prompt for the specific task. Defaults to :data:`None`.
            normalize (Optional[bool], optional): If True, use Sigmoid to normalize the results. Defaults to :data:`None`.
            device (Optional[str], optional): Device to use for computation. Defaults to :data:`None`.

        Returns:
            List[float]: The computed scores.
        """

        if cutoff_layers is None: cutoff_layers = self.cutoff_layers
        if compress_layers is None: compress_layers = self.compress_layers
        if compress_layer is not None:
            print('Try not to use the parameter `compress_layer`; use `compress_layers` instead.')
            compress_layers = compress_layer
        if compress_ratio is None: compress_ratio = self.compress_ratio
        if prompt is None: prompt = self.prompt
        if batch_size is None: batch_size = self.batch_size
        if max_length is None: max_length = self.max_length
        if query_max_length is None:
            if self.query_max_length is not None:
                query_max_length = self.query_max_length
            else:
                query_max_length = max_length * 3 // 4
        if normalize is None: normalize = self.normalize

        if device is None:
            device = self.target_devices[0]

        if device == "cpu": self.use_fp16 = False
        if self.use_fp16 and next(self.model.parameters()).dtype != torch.float16:
            self.model.half()

        self.model.to(device)
        self.model.eval()

        assert isinstance(sentence_pairs, list)
        if isinstance(sentence_pairs[0], str):
            sentence_pairs = [sentence_pairs]

        # tokenize without padding to get the correct length
        all_queries_inputs = []
        all_passages_inputs = []
        for start_index in trange(0, len(sentence_pairs), batch_size, desc="pre tokenize",
                                  disable=len(sentence_pairs) < batch_size):
            sentences_batch = sentence_pairs[start_index:start_index + batch_size]
            queries = [s[0] for s in sentences_batch]
            passages = [s[1] for s in sentences_batch]
            queries_inputs_batch = self.tokenizer(
                queries,
                return_tensors=None,
                add_special_tokens=False,
                max_length=query_max_length,
                truncation=True,
                **kwargs
            )
            passages_inputs_batch = self.tokenizer(
                passages,
                return_tensors=None,
                add_special_tokens=False,
                max_length=max_length,
                truncation=True,
                **kwargs
            )
            queries_inputs_batch = [{
                k: queries_inputs_batch[k][i] for k in queries_inputs_batch.keys()
            } for i in range(len(sentences_batch))]
            passages_inputs_batch = [{
                k: passages_inputs_batch[k][i] for k in passages_inputs_batch.keys()
            } for i in range(len(sentences_batch))]

            all_queries_inputs.extend(queries_inputs_batch)
            all_passages_inputs.extend(passages_inputs_batch)

        # sort by length for less padding
        length_sorted_idx = np.argsort([-len(x['input_ids']) - len(y['input_ids']) for (x, y) in zip(all_queries_inputs, all_passages_inputs)])
        all_queries_inputs_sorted = [all_queries_inputs[i] for i in length_sorted_idx]
        all_passages_inputs_sorted = [all_passages_inputs[i] for i in length_sorted_idx]

        # other inputs
        if prompt is None:
            prompt = "Predict whether passage B contains an answer to query A."
        prompt_inputs = self.tokenizer(
            prompt,
            return_tensors=None,
            add_special_tokens=False
        )['input_ids']
        sep = "\n"
        sep_inputs = self.tokenizer(
            sep,
            return_tensors=None,
            add_special_tokens=False
        )['input_ids']
        encode_max_length = max_length + len(sep_inputs) + len(prompt_inputs)

        # adjust batch size
        flag = False
        while flag is False:
            try:
                batch_inputs = []
                query_lengths = []
                prompt_lengths = []
                for query_inputs, passage_inputs in zip(
                    all_queries_inputs_sorted[:min(len(all_queries_inputs_sorted), batch_size)], 
                    all_passages_inputs_sorted[:min(len(all_passages_inputs_sorted), batch_size)]
                ):
                    item = self.tokenizer.prepare_for_model(
                        [self.tokenizer.bos_token_id] + query_inputs['input_ids'],
                        sep_inputs + passage_inputs['input_ids'],
                        truncation='only_second',
                        max_length=encode_max_length,
                        padding=False,
                        return_attention_mask=False,
                        return_token_type_ids=False,
                        add_special_tokens=False
                    )
                    item['input_ids'] = item['input_ids'] + sep_inputs + prompt_inputs
                    item['attention_mask'] = [1] * len(item['input_ids'])
                    item.pop('token_type_ids') if 'token_type_ids' in item.keys() else None
                    if 'position_ids' in item.keys():
                        item['position_ids'] = list(range(len(item['input_ids'])))
                    batch_inputs.append(item)
                    query_lengths.append(len([self.tokenizer.bos_token_id] + query_inputs['input_ids'] + sep_inputs))
                    prompt_lengths.append(len(sep_inputs + prompt_inputs))

                collater_instance = Collater_for_lightweight(self.tokenizer, max_length)
                batch_inputs = collater_instance([
                    [{
                        'input_ids': item['input_ids'],
                        'attention_mask': item['attention_mask']
                    } for item in batch_inputs],
                    query_lengths,
                    prompt_lengths
                ])[0]

                batch_inputs = {key: val.to(device) for key, val in batch_inputs.items()}

                self.model(
                    **batch_inputs,
                    output_hidden_states=True,
                    compress_layer=compress_layers,
                    compress_ratio=compress_ratio,
                    query_lengths=query_lengths,
                    prompt_lengths=prompt_lengths,
                    cutoff_layers=cutoff_layers
                )
                flag = True
            except RuntimeError as e:
                batch_size = batch_size * 3 // 4
            except torch.cuda.OutOfMemoryError as e:
                batch_size = batch_size * 3 // 4

        all_scores = []
        for batch_start in trange(0, len(all_queries_inputs_sorted), batch_size):
            queries_inputs = all_queries_inputs_sorted[batch_start:batch_start+batch_size]
            passages_inputs = all_passages_inputs_sorted[batch_start:batch_start+batch_size]

            batch_inputs = []
            query_lengths = []
            prompt_lengths = []
            for query_inputs, passage_inputs in zip(queries_inputs, passages_inputs):
                item = self.tokenizer.prepare_for_model(
                    [self.tokenizer.bos_token_id] + query_inputs['input_ids'],
                    sep_inputs + passage_inputs['input_ids'],
                    truncation='only_second',
                    max_length=encode_max_length,
                    padding=False,
                    return_attention_mask=False,
                    return_token_type_ids=False,
                    add_special_tokens=False
                )
                item['input_ids'] = item['input_ids'] + sep_inputs + prompt_inputs
                item['attention_mask'] = [1] * len(item['input_ids'])
                item.pop('token_type_ids') if 'token_type_ids' in item.keys() else None
                if 'position_ids' in item.keys():
                    item['position_ids'] = list(range(len(item['input_ids'])))
                batch_inputs.append(item)
                query_lengths.append(len([self.tokenizer.bos_token_id] + query_inputs['input_ids'] + sep_inputs))
                prompt_lengths.append(len(sep_inputs + prompt_inputs))

            collater_instance = Collater_for_lightweight(self.tokenizer, max_length)
            batch_inputs = collater_instance([
                [{
                    'input_ids': item['input_ids'],
                    'attention_mask': item['attention_mask']
                } for item in batch_inputs],
                query_lengths,
                prompt_lengths
            ])[0]

            batch_inputs = {key: val.to(device) for key, val in batch_inputs.items()}

            outputs = self.model(
                **batch_inputs,
                output_hidden_states=True,
                compress_layer=compress_layers,
                compress_ratio=compress_ratio,
                query_lengths=query_lengths,
                prompt_lengths=prompt_lengths,
                cutoff_layers=cutoff_layers
            )
            scores = []
            for i in range(len(outputs.logits)):
                logits = last_logit_pool_lightweight(outputs.logits[i], outputs.attention_masks[i])
                scores.append(logits.cpu().float().tolist())
            if len(all_scores) == 0:
                for i in range(len(scores)):
                    all_scores.append([])
            for i in range(len(scores)):
                all_scores[i].extend(scores[i])

        for i in range(len(all_scores)):
            all_scores[i] = [all_scores[i][idx] for idx in np.argsort(length_sorted_idx)]
            if normalize:
                all_scores[i] = [sigmoid(score) for score in all_scores[i]]
    
        if len(all_scores) == 1 and isinstance(all_scores[0], list):
            all_scores = all_scores[0]

        return all_scores