from FlagEmbedding import FlagReranker model_name = "OpenBMB/MiniCPM-Reranker-Light" model = FlagReranker(model_name, use_fp16=True, query_instruction_for_rerank="Query: ", trust_remote_code=True) # You can hack the __init__() method of the FlagEmbedding BaseReranker class to use flash_attention_2 for faster inference # self.model = AutoModelForSequenceClassification.from_pretrained( # model_name_or_path, # trust_remote_code=trust_remote_code, # cache_dir=cache_dir, # # torch_dtype=torch.float16, # we need to add this line to use fp16 # # attn_implementation="flash_attention_2", # we need to add this line to use flash_attention_2 # ) model.tokenizer.padding_side = "right" query = "中国的首都是哪里?" # "Where is the capital of China?" passages = ["beijing", "shanghai"] # 北京,上海 sentence_pairs = [[query, doc] for doc in passages] scores = model.compute_score(sentence_pairs,normalize=True) print(scores) # [0.01791734476747132, 0.0002472934613244585]