langchain/libs/community/langchain_community/embeddings/ascend.py

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

from langchain_core.embeddings import Embeddings
from pydantic import BaseModel, ConfigDict, model_validator


class AscendEmbeddings(Embeddings, BaseModel):
    """
    Ascend NPU accelerate Embedding model

    Please ensure that you have installed CANN and torch_npu.

    Example:

    from langchain_community.embeddings import AscendEmbeddings
    model = AscendEmbeddings(model_path=<path_to_model>,
        device_id=0,
        query_instruction="Represent this sentence for searching relevant passages: "
    )
    """

    """model path"""
    model_path: str
    """Ascend NPU device id."""
    device_id: int = 0
    """Unstruntion to used for embedding query."""
    query_instruction: str = ""
    """Unstruntion to used for embedding document."""
    document_instruction: str = ""
    use_fp16: bool = True
    pooling_method: Optional[str] = "cls"
    batch_size: int = 32
    model: Any
    tokenizer: Any

    model_config = ConfigDict(protected_namespaces=())

    def __init__(self, *args: Any, **kwargs: Any) -> None:
        super().__init__(*args, **kwargs)
        try:
            from transformers import AutoModel, AutoTokenizer
        except ImportError as e:
            raise ImportError(
                "Unable to import transformers, please install with "
                "`pip install -U transformers`."
            ) from e
        try:
            self.model = AutoModel.from_pretrained(self.model_path).npu().eval()
            self.tokenizer = AutoTokenizer.from_pretrained(self.model_path)
        except Exception as e:
            raise Exception(
                f"Failed to load model [self.model_path], due to following error:{e}"
            )

        if self.use_fp16:
            self.model.half()
        self.encode([f"warmup {i} times" for i in range(10)])

    @model_validator(mode="before")
    @classmethod
    def validate_environment(cls, values: Dict) -> Any:
        if "model_path" not in values:
            raise ValueError("model_path is required")
        if not os.access(values["model_path"], os.F_OK):
            raise FileNotFoundError(
                f"Unable to find valid model path in [{values['model_path']}]"
            )
        try:
            import torch_npu
        except ImportError:
            raise ModuleNotFoundError("torch_npu not found, please install torch_npu")
        except Exception as e:
            raise e
        try:
            torch_npu.npu.set_device(values["device_id"])
        except Exception as e:
            raise Exception(f"set device failed due to {e}")
        return values

    def encode(self, sentences: Any) -> Any:
        inputs = self.tokenizer(
            sentences,
            padding=True,
            truncation=True,
            return_tensors="pt",
            max_length=512,
        )
        try:
            import torch
        except ImportError as e:
            raise ImportError(
                "Unable to import torch, please install with `pip install -U torch`."
            ) from e
        last_hidden_state = self.model(
            inputs.input_ids.npu(), inputs.attention_mask.npu(), return_dict=True
        ).last_hidden_state
        tmp = self.pooling(last_hidden_state, inputs["attention_mask"].npu())
        embeddings = torch.nn.functional.normalize(tmp, dim=-1)
        return embeddings.cpu().detach().numpy()

    def pooling(self, last_hidden_state: Any, attention_mask: Any = None) -> Any:
        try:
            import torch
        except ImportError as e:
            raise ImportError(
                "Unable to import torch, please install with `pip install -U torch`."
            ) from e
        if self.pooling_method == "cls":
            return last_hidden_state[:, 0]
        elif self.pooling_method == "mean":
            s = torch.sum(
                last_hidden_state * attention_mask.unsqueeze(-1).float(), dim=-1
            )
            d = attention_mask.sum(dim=1, keepdim=True).float()
            return s / d
        else:
            raise NotImplementedError(
                f"Pooling method [{self.pooling_method}] not implemented"
            )

    def embed_documents(self, texts: List[str]) -> List[List[float]]:
        try:
            import numpy as np
        except ImportError as e:
            raise ImportError(
                "Unable to import numpy, please install with `pip install -U numpy`."
            ) from e
        embedding_list = []
        for i in range(0, len(texts), self.batch_size):
            texts_ = texts[i : i + self.batch_size]
            emb = self.encode([self.document_instruction + text for text in texts_])
            embedding_list.append(emb)
        return np.concatenate(embedding_list)

    def embed_query(self, text: str) -> List[float]:
        return self.encode([self.query_instruction + text])[0]