基础模型:

• Qwen/Qwen2.5-Coder-7B-Instruct 库名称: transformers 许可证: mit 评估指标:
• 准确率 流水线标签: 文本生成

[📜 论文] • [🤗 HF模型] • [🐱 GitHub]

模型详情

要了解思维迁移模式的启动方法，请参考 https://github.com/efficientscaling/Z1。

评估

Gradio演示

import copy
from typing import List
from dataclasses import dataclass

import gradio as gr
from vllm import LLM, SamplingParams
from transformers import  AutoTokenizer

BOX=r"\boxed{}"
ANSWER_WITH_BOX=f"\n\n经过深入思考，最终答案{BOX}应为：\n\n"
ANSWER_WITHOUT_BOX=f"\n\n经过深入思考，最终答案应为：\n\n"

model_name = "efficientscaling/Z1-7B"

@dataclass
class ThinkingLLM(LLM):

    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)

    def thinking_generate(self, prompts: List[str], sampling_params: SamplingParams = None, max_tokens_for_thinking: int = None):

        if sampling_params is None:
            raise ValueError("采样参数不能为空！")
        else:
            all_max_tokens = sampling_params.max_tokens
            sampling_params.max_tokens = max_tokens_for_thinking
            print(f"总token数: {all_max_tokens}")
            print(f"思考token数: {max_tokens_for_thinking}")

        trajectories = self.generate(prompts, sampling_params)

        rethinking_str = ANSWER_WITHOUT_BOX
        sampling_params.max_tokens = all_max_tokens

        answers = copy.deepcopy(trajectories)

        unfinished_id = []
        thinking_token = 0
        new_prompts = []

        for id, traj in enumerate(trajectories):
            if traj.outputs[0].finish_reason == 'length':
                unfinished_id.append(id)
                new_prompts.append(prompts[id] + traj.outputs[0].text + rethinking_str)
            thinking_token += len(traj.outputs[0].token_ids)

        avg_thinking_token = thinking_token / len(prompts)

        if new_prompts:
            print(new_prompts[0])

            o = self.generate(
                new_prompts,
                sampling_params=sampling_params,
            )
            
        for i, uid in enumerate(unfinished_id):
            answers[uid] = o[i]

        return new_prompts, answers


def generate_text(prompt, max_tokens, max_tokens_for_thinking, temperature, top_p):

    sampling_params = SamplingParams(
        temperature=temperature,
        max_tokens=max_tokens,
        top_p=top_p,
        skip_special_tokens=False,
    )

    trajectories, outputs = llm.thinking_generate(prompt, sampling_params, max_tokens_for_thinking=max_tokens_for_thinking)
    return trajectories[0] + '\n\n' + outputs[0].outputs[0].text if trajectories else outputs[0].outputs[0].text


llm = ThinkingLLM(
    model=model_name,
    tensor_parallel_size=1,
    gpu_memory_utilization=0.96,
    )


with gr.Blocks() as demo:
    gr.Markdown("# 基于思维迁移的推理")
    
    with gr.Row():
        with gr.Column():
            prompt_input = gr.Textbox(
                label="输入提示",
                placeholder="请输入内容",
                lines=5,
            )
            max_tokens_for_thinking_input = gr.Slider(
                label="思维迁移窗口大小",
                minimum=1,
                maximum=32786,
                value=4000,
                step=1,
            )
            max_tokens_input = gr.Slider(
                label="最大token数",
                minimum=1,
                maximum=32786,
                value=32786,
                step=1,
            )
            temperature_input = gr.Slider(
                label="温度参数",
                minimum=00,
                maximum=2.0,
                value=0,
                step=0.1,
            )
            top_p_input = gr.Slider(
                label="Top-p采样",
                minimum=0.0,
                maximum=1.0,
                value=1,
                step=0.01,
            )
            generate_button = gr.Button("生成")
        
        with gr.Column():
            output_text = gr.Textbox(
                label="思维迁移窗口",
                placeholder="生成文本将显示在此处...",
                lines=10,
            )
    
    generate_button.click(
        fn=generate_text,
        inputs=[prompt_input, max_tokens_for_thinking_input,max_tokens_input,  temperature_input, top_p_input],
        outputs=output_text, 
    )

if __name__ == "__main__":
    demo.launch()

引用

@misc{yu2025efficientscaling,
      title={Z1: 基于代码的高效测试时扩展}, 
      author={于兆健 and 吴英豪 and 赵一伦 and Arman Cohan and 张小平},
      year={2025},
      eprint={2504.00810},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2504.00810}, 
}