from modelscope import AutoModelForCausalLM, AutoTokenizer
from accelerate import Accelerator
import torch
import os

modelName = "SageGen_qwen2.5_14B_instruct"
accelerator = Accelerator()

tokenizers = AutoTokenizer.from_pretrained(modelName,trust_remote_code=True)

model = AutoModelForCausalLM.from_pretrained(
    modelName,
    torch_dtype='auto',
    device_map='auto',
    trust_remote_code=True
)

messages = [
    {"role": "system", "content": system_example},
    {"role": "user", "content": user_example},
    {"role": "assistant", "content": assistance_example}

]

userInput_query = input("用户描述内容为:")

messages.append({"role":"user","content":userInput_query})

user_inputs = tokenizers.apply_chat_template(
    messages,
    add_generation_prompt=True,
    tokenize=True,
    return_tensors="pt",
    return_dict=True
)
model, user_inputs = accelerator.prepare(model, user_inputs)
user_inputs = user_inputs.to(model.device)

with torch.no_grad():
    print("开始推理\n\n")
    torch.cuda.empty_cache()  # 如果使用GPU
    generated_ids = model.generate(
        **user_inputs,
        max_new_tokens=20000,
        do_sample=True,
        temperature=0.8,
        top_p=0.9,
        repetition_penalty=1.1, 

    )

generated_ids = [output_ids[len(input_ids):] for input_ids, output_ids in zip(user_inputs.input_ids, generated_ids)]
response = tokenizers.batch_decode(generated_ids, skip_special_tokens=True)[0]

print(response)