model.py

import torch
from torch import nn
from transformers import AutoTokenizer, AutoModel

class AutoModelForSentenceEmbedding(nn.Module):
  def __init__(self, model):
    super().__init__()

    self.model = model

  def forward(self, **kwargs):
    model_output = self.model(**kwargs)
    embeddings = self.mean_pooling(model_output, kwargs['attention_mask'])
    embeddings = torch.nn.functional.normalize(embeddings, p=2, dim=1)
    return embeddings

  def mean_pooling(self, model_output, attention_mask):
    token_embeddings = model_output[0]  # First element of model_output contains all token embeddings
    input_mask_expanded = attention_mask.unsqueeze(-1).expand(token_embeddings.size()).float()
    return torch.sum(token_embeddings * input_mask_expanded, 1) / torch.clamp(input_mask_expanded.sum(1), min=1e-9)

device = "cuda" if torch.cuda.is_available() else "cpu"

def create_semantic_ranking_model(device=device):
  """Creates a HuggingFace all-MiniLM-L6-v2 model.

  Args:
    device: A torch.device
  Returns:
    A tuple of the model and tokenizer
  """
  tokenizer = AutoTokenizer.from_pretrained('sentence-transformers/all-MiniLM-L6-v2')
  model = AutoModelForSentenceEmbedding(AutoModel.from_pretrained('sentence-transformers/all-MiniLM-L6-v2')).to(device)

  for param in model.model.parameters():
    param.requires_grad = False

  return model, tokenizer

# Example usage
model, tokenizer = create_semantic_ranking_model()