from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_path = "D:\\hf_cache\\Hy-MT2-1.8B"

if __name__ == "__main__":
    # 从https://ai.gitcode.com/tencent_hunyuan/Hy-MT2-1.8B下载
    # Load tokenizer
    tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)

    # Load model
    model = AutoModelForCausalLM.from_pretrained(
        model_path,
        dtype=torch.bfloat16,
        device_map="auto",
        trust_remote_code=True,
    )

    model.eval()

    # Example inference
    prompt = "将以下文本翻译成英语,注意只需要输出翻译后的结果,不要额外解释:\n\n今天天气真好。"
    messages = [{"role": "user", "content": prompt}]
    inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)

    with torch.no_grad():
        outputs = model.generate(
            **inputs,
            max_new_tokens=4096,
        )
    response = tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True)
    print(response)