kanana_embedding_wrapper.py

import os
import json
import torch
from torch import nn
import torch.nn.functional as F
from typing import List, Dict

from transformers import AutoModel

os.environ["TOKENIZERS_PARALLELISM"] = "false"

class KananaEmbeddingWrapper(nn.Module):
    def __init__(self, model_name_or_path: str, trust_remote_code=True, device: str = "cpu", max_seq_length:int=None):
        """
        Initialize the KananaEmbeddingWrapper.
        
        Args:
            model_name_or_path: Path or name of the pretrained model
            trust_remote_code: Whether to trust remote code when loading the model
            device: Device to load the model on (e.g., 'cpu', 'cuda')
        """
        super(KananaEmbeddingWrapper, self).__init__()
        self.model_name_or_path = model_name_or_path
        self.trust_remote_code = trust_remote_code
        self.device = device
        self.kanana2vec = AutoModel.from_pretrained(
            model_name_or_path, trust_remote_code=trust_remote_code
        ).to(self.device)

        self.max_seq_length = max_seq_length if max_seq_length is not None else self.kanana2vec.config.max_position_embeddings

    def get_sentence_embedding_dimension(self) -> int:
        """
        Returns the dimension of the sentence embeddings.
        
        Returns:
            Dimensionality of the sentence embeddings
        """
        return self.kanana2vec.config.hidden_size
    
    def get_max_seq_length(self) -> int:
        """
        Returns the maximum sequence length this module can process.
        
        Returns:
            Maximum sequence length
        """
        return self.max_seq_length
    
    def tokenize(self, texts: List[str]) -> Dict[str, torch.Tensor]:
        """
        Tokenize input texts.
        
        Args:
            texts: List of input texts to tokenize
            
        Returns:
            Dictionary containing tokenized inputs
        """
        return self.kanana2vec.tokenizer(
            texts,
            padding=True,
            return_token_type_ids=False,
            return_tensors="pt",
            truncation=True,
            max_length=self.max_seq_length
        ).to(self.device)

    def forward(self, features: Dict[str, torch.Tensor]) -> Dict[str, torch.Tensor]:
        """
        Forward pass of the module.
        
        Args:
            features: Dictionary with inputs including 'input_ids', 'attention_mask', etc.
        
        Returns:
            Dictionary with updated features including 'sentence_embedding'
        """
        # Extract only the required features for the model
        model_inputs = self._extract_model_inputs(features)
        
        # Create pool mask considering prompt length if available
        model_inputs["pool_mask"] = self._create_pool_mask(features)
        
        # Get embeddings from the model and normalize
        embedding = self.kanana2vec.forward(**model_inputs).embedding
        normalized_embedding = F.normalize(embedding, p=2, dim=1)
        
        # Update features with sentence embedding
        features['sentence_embedding'] = normalized_embedding
        return features
    
    def _extract_model_inputs(self, features: Dict[str, torch.Tensor]) -> Dict[str, torch.Tensor]:
        """
        Extract only the inputs needed for the model.
        
        Args:
            features: Complete feature dictionary
            
        Returns:
            Dictionary with only the required keys for the model
        """
        return {k: v for k, v in features.items() if k in ['input_ids', 'attention_mask']}
    
    def _create_pool_mask(self, features: Dict[str, torch.Tensor]) -> torch.Tensor:
        """
        Create a pool mask based on attention mask and prompt length.
        
        Args:
            features: Feature dictionary containing attention_mask and optionally prompt_length
            
        Returns:
            Pool mask tensor
        """
        pool_mask = features['attention_mask'].clone()
        if "prompt_length" in features:
            pool_mask[:, :features['prompt_length']] = 0
        return pool_mask
    
    def get_config_dict(self) -> Dict:
        """
        Returns a dictionary with the module's configuration.
        
        Returns:
            Dictionary with module configuration
        """
        return {
            "model_name_or_path": self.model_name_or_path,
            "trust_remote_code": self.trust_remote_code,
            "device": self.device,
            "hidden_size": self.get_sentence_embedding_dimension(),
            "max_seq_length": self.get_max_seq_length()
        }
    
    def save(self, save_dir: str) -> None:
        """
        Saves the module's configuration and model to the specified directory.
        
        Args:
            save_dir: Directory to save the module configuration
        """
        os.makedirs(save_dir, exist_ok=True)
        
        # Save model configuration
        config_path = os.path.join(save_dir, "kanana_embedding_config.json")
        with open(config_path, 'w', encoding='utf-8') as f:
            json.dump(self.get_config_dict(), f, ensure_ascii=False, indent=2)
        
        # Save the underlying model
        model_save_path = os.path.join(save_dir, "kanana2vec")
        self.kanana2vec.save_pretrained(model_save_path)
        
        print(f"KananaEmbeddingWrapper model saved to {save_dir}")
    
    @staticmethod
    def load(load_dir: str, device: str = "cpu") -> 'KananaEmbeddingWrapper':
        """
        Loads a KananaEmbeddingWrapper model from the specified directory.
        
        Args:
            load_dir: Directory containing the saved module
            device: Device to load the model on
            
        Returns:
            Initialized KananaEmbeddingWrapper
        """
        # Load configuration
        config_path = os.path.join(load_dir, "kanana_embedding_config.json")
        with open(config_path, 'r', encoding='utf-8') as f:
            config = json.load(f)
        
        # Use the saved model path
        model_load_path = os.path.join(load_dir, "kanana2vec")
        
        # Create instance with saved configuration
        instance = KananaEmbeddingWrapper(
            model_name_or_path=model_load_path,
            trust_remote_code=config.get("trust_remote_code", True),
            device=device  # Use the provided device or default
        )
        
        return instance