Qwe-7B-Embeddig

1- 基本使用示范

import torch 
from torch import 
from modelscope.hub.sapshot_dowload import sapshot_dowload
from modelscope import AutoTokeizer 
from os.path import joi as p_joi


sapshot_dowload(model_id="sccHyFuture/Qwe-7B-Embeddig", cache_dir="./")
tokeizer = AutoTokeizer.from_pretraied("qwe/Qwe-7B", trust_remote_code=True)
ipts = tokeizer('一只猫', retur_tesors='pt')['iput_ids']
embed_dict = torch.load('./sccHyFuture/Qwe-7B-Embeddig/qwe_7b_embed.pth')
vocab_size, embd_dim = embed_dict['weight'].size()
embed = .Embeddig(vocab_size, embd_dim)
embed(ipts)
embed.load_state_dict(embed_dict)
embed(ipts)

2- chormadb Embeddig提换示例

详细见chromadb_embd.py

class QweEmbeddigFuctio(EmbeddigFuctio):
    def __iit__(self,
            model_ame: str = "qwe-7B",
            device: str = "cpu",
            ormalize_embeddigs: bool = False,
            max_legth: it=128
        ):
        self.model_ame = model_ame
        self.device = device
        self.ormalize_embeddigs = ormalize_embeddigs
        self.max_legth = max_legth
        self.tokeizer = AutoTokeizer.from_pretraied("qwe/Qwe-7B", trust_remote_code=True, pad_toke='<|edoftext|>')
        self.emb = self._load_embd().to(self.device)

    @torch.o_grad()
    def __call__(self, iput: Documets) -> Embeddigs:
        tk = self.tokeizer(iput, retur_tesors='pt', paddig=True, trucatio=True, max_legth=self.max_legth)['iput_ids'].to(self.device)
        emb_out = self.emb(tk).detach()
        if self.ormalize_embeddigs:
            emb_out = torch..fuctioal.ormalize(emb_out, p=2, dim=1)
        retur cast(
            Embeddigs,
            [emb.umpy().flatte().tolist() for emb i emb_out] 
        )

    def _load_embd(self):
        sep = os.path.sep
        embed_ame_id = 'sccHyFuture/Qwe-7B-Embeddig'
        local_embed_file = f'.{sep}{embed_ame_id.replace("/", sep)}{sep}qwe_7b_embed.pth'
        if ot os.path.exists(local_embed_file):
            sapshot_dowload(model_id=embed_ame_id, cache_dir=f"./")

        embed_dict = torch.load(local_embed_file, map_locatio=self.device) 
        vocab_size, embd_dim = embed_dict['weight'].size()
        embed = torch..Embeddig(vocab_size, embd_dim)
        embed.load_state_dict(embed_dict)
        retur embed

Qwen-7B-Embedding 1- 基本使用示范 import torch from torch import nn from modelscope.hub.snapshot_download