optimum_encoder.py

import requests
import time
import os
import numpy as np
from tqdm import tqdm
from typing import Any, List, Optional, Dict
from pydantic.v1 import PrivateAttr
from fastembed.common.utils import normalize
from semantic_router.encoders import BaseEncoder
from semantic_router.utils.logger import logger
from langchain_core.embeddings import Embeddings


class OptimumEncoder(BaseEncoder, Embeddings):
    name: str = "mixedbread-ai/mxbai-embed-large-v1"
    type: str = "huggingface"
    score_threshold: float = 0.5
    tokenizer_kwargs: Dict = {}
    model_kwargs: Dict = {}
    device: Optional[str] = None
    _tokenizer: Any = PrivateAttr()
    _model: Any = PrivateAttr()
    _torch: Any = PrivateAttr()

    def __init__(self, **data):
        super().__init__(**data)
        self._tokenizer, self._model = self._initialize_hf_model()

    def _initialize_hf_model(self):
        try:
            import onnxruntime as ort
            from optimum.onnxruntime import ORTModelForFeatureExtraction
        except ImportError:
            raise ImportError(
                "Please install optimum and onnxruntime to use OptimumEncoder. "
                "You can install it with: "
                "`pip install transformers optimum[onnxruntime-gpu]`"
            )

        try:
            import torch
        except ImportError:
            raise ImportError(
                "Please install Pytorch to use OptimumEncoder. "
                "You can install it with: "
                "`pip install semantic-router[local]`"
            )
        try:
            from transformers import AutoTokenizer
        except ImportError:
            raise ImportError(
                "Please install transformers to use OptimumEncoder. "
                "You can install it with: "
                "`pip install semantic-router[local]`"
            )

        self._torch = torch

        tokenizer = AutoTokenizer.from_pretrained(
            self.name,
            **self.tokenizer_kwargs,
        )
        
        #provider_options = {
        #    "trt_engine_cache_enable": True,
        #    "trt_engine_cache_path": os.getenv('HF_HOME'),
        #    "trt_fp16_enable": True
        #}

        session_options = ort.SessionOptions()
        session_options.log_severity_level = 0

        ort_model = ORTModelForFeatureExtraction.from_pretrained(
            model_id=self.name, 
            file_name='model_fp16.onnx',
            subfolder='onnx',
            provider='CUDAExecutionProvider',
            use_io_binding=True,
            #provider_options=provider_options,
            session_options=session_options,
            **self.model_kwargs
        )

        # print("Building engine for a short sequence...")
        # short_text = ["short"]
        # short_encoded_input = tokenizer(
        #     short_text, padding=True, truncation=True, return_tensors="pt"
        # ).to(self.device)
        # short_output = ort_model(**short_encoded_input)
        
        # print("Building engine for a long sequence...")
        # long_text = ["a very long input just for demo purpose, this is very long" * 10]
        # long_encoded_input = tokenizer(
        #     long_text, padding=True, truncation=True, return_tensors="pt"
        # ).to(self.device)
        # long_output = ort_model(**long_encoded_input)

        # text = ["Replace me by any text you'd like."]
        # encoded_input = tokenizer(
        #     text, padding=True, truncation=True, return_tensors="pt"
        # ).to(self.device)
        
        # for i in range(3):
        #     output = ort_model(**encoded_input)

        return tokenizer, ort_model

    def embed_documents(
        self,
        docs: List[str],
        batch_size: int = 32,
        normalize_embeddings: bool = True,
        pooling_strategy: str = "mean",
        convert_to_numpy: bool = False
    ) -> List[List[float]] | List[np.ndarray]:
        all_embeddings = []
        for i in tqdm(range(0, len(docs), batch_size)):
            batch_docs = docs[i : i + batch_size]

            encoded_input = self._tokenizer(
                batch_docs, padding=True, truncation=True, return_tensors="pt"
            ).to(self.device)

            with self._torch.no_grad():
                model_output = self._model(**encoded_input)

            if pooling_strategy == "mean":
                embeddings = self._mean_pooling(
                    model_output, encoded_input["attention_mask"]
                )
            elif pooling_strategy == "max":
                embeddings = self._max_pooling(
                    model_output, encoded_input["attention_mask"]
                )
            else:
                raise ValueError(
                    "Invalid pooling_strategy. Please use 'mean' or 'max'."
                )

            if normalize_embeddings:
                embeddings = self._torch.nn.functional.normalize(embeddings, p=2, dim=1)
                
            if convert_to_numpy:
                embeddings = embeddings.detach().cpu().numpy()
            else:
                embeddings = embeddings.tolist()
                
            all_embeddings.extend(embeddings)

        return all_embeddings

    def embed_query(
        self,
        docs: str,
        normalize_embeddings: bool = True,
        pooling_strategy: str = "mean"
    ) -> List[float]:
        encoded_input = self._tokenizer(
            docs, padding=True, truncation=True, return_tensors="pt"
        ).to(self.device)

        with self._torch.no_grad():
            model_output = self._model(**encoded_input)

        if pooling_strategy == "mean":
            embeddings = self._mean_pooling(
                model_output, encoded_input["attention_mask"]
            )
        elif pooling_strategy == "max":
            embeddings = self._max_pooling(
                model_output, encoded_input["attention_mask"]
            )
        else:
            raise ValueError(
                "Invalid pooling_strategy. Please use 'mean' or 'max'."
            )

        if normalize_embeddings:
            embeddings = self._torch.nn.functional.normalize(embeddings, p=2, dim=1)
                
        return embeddings

    def _mean_pooling(self, model_output, attention_mask):
        token_embeddings = model_output[0]
        input_mask_expanded = (
            attention_mask.unsqueeze(-1).expand(token_embeddings.size()).float()
        )
        return self._torch.sum(
            token_embeddings * input_mask_expanded, 1
        ) / self._torch.clamp(input_mask_expanded.sum(1), min=1e-9)

    def _max_pooling(self, model_output, attention_mask):
        token_embeddings = model_output[0]
        input_mask_expanded = (
            attention_mask.unsqueeze(-1).expand(token_embeddings.size()).float()
        )
        token_embeddings[input_mask_expanded == 0] = -1e9
        return self._torch.max(token_embeddings, 1)[0]