app.py

import os
import sys
import subprocess
import torch
from typing import Union
from pydantic import BaseModel
from fastapi import FastAPI
import uvicorn

# gradio相关
import gradio as gr

# transformers相关
from transformers import AutoTokenizer

##############################################################################
# 1) 若本地(或Space)没有 skywork-o1-prm-inference 目录，则 clone 下来，并将其加入 sys.path
##############################################################################
def setup_environment():
    if not os.path.exists("skywork-o1-prm-inference"):
        print("Cloning repository skywork-o1-prm-inference...")
        subprocess.run(
            ["git", "clone", "https://github.com/SkyworkAI/skywork-o1-prm-inference.git"],
            check=True
        )
    repo_path = os.path.abspath("skywork-o1-prm-inference")
    if repo_path not in sys.path:
        sys.path.append(repo_path)
        print(f"Added {repo_path} to Python path")

setup_environment()

##############################################################################
# 2) 导入Skywork项目内的模块
##############################################################################
from model_utils.prm_model import PRM_MODEL
from model_utils.io_utils import prepare_input, prepare_batch_input_for_model, derive_step_rewards

##############################################################################
# 3) 加载模型及其相关资源
##############################################################################
# 你可根据需求更换 model_id
model_id = "Skywork/Skywork-o1-Open-PRM-Qwen-2.5-1.5B"

print(f"Loading tokenizer from {model_id} ...")
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)

print(f"Loading model from {model_id} ...")
model = PRM_MODEL.from_pretrained(model_id).to("cpu").eval()

##############################################################################
# 4) 定义推理函数
##############################################################################
def compute_rewards(problem: str, response: str):
    """核心推理函数：将 problem + response 输入模型，输出 step_rewards。"""
    data = {
        "problem": problem,
        "response": response
    }
    processed_data = [
        prepare_input(
            data["problem"],
            data["response"],
            tokenizer=tokenizer,
            step_token="\n"
        )
    ]
    input_ids, steps, reward_flags = zip(*processed_data)

    # 准备 batch 输入
    input_ids, attention_mask, reward_flags = prepare_batch_input_for_model(
        input_ids,
        reward_flags,
        tokenizer.pad_token_id
    )

    input_ids = input_ids.to("cpu")
    attention_mask = attention_mask.to("cpu")
    if isinstance(reward_flags, torch.Tensor):
        reward_flags = reward_flags.to("cpu")

    # 前向推理
    with torch.no_grad():
        _, _, rewards = model(
            input_ids=input_ids,
            attention_mask=attention_mask,
            return_probs=True
        )

    # 计算 step_rewards
    step_rewards = derive_step_rewards(rewards, reward_flags)
    return step_rewards[0]

##############################################################################
# 5) 准备FastAPI应用：对外暴露 /api/predict 接口
##############################################################################
app = FastAPI(title="PRM Inference Server", description="FastAPI + Gradio App")

# 5.1 定义API输入与输出的数据结构
class InferenceData(BaseModel):
    problem: str
    response: str

class InferenceOutput(BaseModel):
    step_rewards: list

@app.post("/api/predict", response_model=InferenceOutput)
def predict(data: InferenceData):
    """
    直接使用HTTP POST /api/predict，
    body JSON: { "problem": "...", "response": "..." } 即可得到 step_rewards。
    """
    rewards = compute_rewards(data.problem, data.response)
    return InferenceOutput(step_rewards=rewards)

##############################################################################
# 6) 构建 Gradio 界面，并将其挂载到 /gradio 路径
##############################################################################
def build_gradio_interface():
    with gr.Blocks() as demo:
        gr.Markdown(
            "## PRM Reward Calculation\n\n"
            "输入 `Problem` 和 `Response`，点击下方按钮即可获得 step_rewards。"
        )

        problem_input = gr.Textbox(
            label="Problem",
            lines=5,
            value=(
                "Janet's ducks lay 16 eggs per day. She eats three for breakfast every morning "
                "and bakes muffins for her friends every day with four. She sells the remainder "
                "at the farmers' market daily for $2 per fresh duck egg. How much in dollars "
                "does she make every day at the farmers' market?"
            )
        )
        response_input = gr.Textbox(
            label="Response",
            lines=10,
            value=(
                "To determine how much money Janet makes every day at the farmers' market, "
                "we need to follow these steps:\n1. ..."
            )
        )
        output = gr.JSON(label="Step Rewards")

        submit_btn = gr.Button("Compute Rewards")
        submit_btn.click(
            fn=compute_rewards,
            inputs=[problem_input, response_input],
            outputs=output
        )

    return demo

demo = build_gradio_interface()

# 6.1 挂载Gradio到FastAPI的 /gradio 路径
app = gr.mount_gradio_app(app, demo, path="/gradio")

##############################################################################
# 7) 在 main 中用 uvicorn 启动
##############################################################################
# 注意：在Hugging Face Spaces中，一般只需要 `python app.py` 即可开始监听。
# 若Spaces是自动检测并执行 `app.py`，则不一定会执行 `if __name__ == "__main__"` 分支。
# 因此你只要确保上面定义的 `app` 变量存在即可。
# 这里写一个可选的 main，方便本地调试。
if __name__ == "__main__":
    uvicorn.run(app, host="0.0.0.0", port=7860)