app.py

import os
os.environ["OMP_NUM_THREADS"] = "1"

import gradio as gr
import torch
# 【关键修改】不再用 AutoTokenizer，改用 GPT2Tokenizer
from transformers import GPT2Tokenizer, SiglipImageProcessor, AutoModelForCausalLM
from PIL import Image
import re

# --- 模型配置 ---
model_id = "starvector/starvector-1b-im2svg"

print(f"正在加载模型: {model_id} ...")

# 1. 加载图片处理器
try:
    image_processor = SiglipImageProcessor.from_pretrained(model_id, trust_remote_code=True)
except:
    from transformers import AutoImageProcessor
    image_processor = AutoImageProcessor.from_pretrained(model_id, trust_remote_code=True)

# 2. 加载 Tokenizer
print("正在加载 Tokenizer (强制使用 GPT2 Python 版)...")
# 【核心修复】
# StarVector-1B 基于 StarCoder (GPT2架构)。
# 直接使用 GPT2Tokenizer 可以避开 AutoTokenizer 强制读取 tokenizer.json 导致的报错。
# 这个类只读取 vocab.json，兼容性最好。
tokenizer = GPT2Tokenizer.from_pretrained(model_id, trust_remote_code=True)

# 确保 pad_token 存在 (GPT2 默认没有 pad_token，StarCoder 通常用 eos 作为 pad)
if tokenizer.pad_token is None:
    tokenizer.pad_token = tokenizer.eos_token

# 3. 加载模型
print("正在加载主模型 (FP16)...")
# 1B 模型很小，直接加载到 CPU 没问题，不会 OOM
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.float16,
    trust_remote_code=True
    # 不使用 device_map，防止 Meta Tensor 报错
)

print("正在移动模型到 GPU...")
model.to("cuda")

print("模型加载完成！")

# --- 推理函数 ---
def convert_image_to_svg(image):
    if image is None:
        return "请上传图片", ""
    
    try:
        # 预处理图片
        image_inputs = image_processor(image, return_tensors="pt")
        # 确保数据在 GPU 上且精度匹配
        pixel_values = image_inputs.pixel_values.to("cuda").to(torch.float16)
        
        # 预处理文本
        text_prompt = "Generate SVG code"
        text_inputs = tokenizer(text_prompt, return_tensors="pt")
        input_ids = text_inputs.input_ids.to("cuda")
        attention_mask = text_inputs.attention_mask.to("cuda")

        # 生成
        with torch.no_grad():
            generated_ids = model.generate(
                input_ids=input_ids,
                attention_mask=attention_mask,
                pixel_values=pixel_values,
                max_new_tokens=2048, 
                do_sample=False,
                pad_token_id=tokenizer.eos_token_id
            )

        # 解码
        generated_text = tokenizer.decode(generated_ids[0], skip_special_tokens=True)
        svg_code = extract_svg_content(generated_text)
        
        return svg_code, svg_code

    except Exception as e:
        import traceback
        traceback.print_exc()
        return f"生成出错: {str(e)}", ""

def extract_svg_content(text):
    pattern = r"(<svg.*?</svg>)"
    match = re.search(pattern, text, re.DOTALL)
    if match:
        return match.group(1)
    
    if "```xml" in text:
        return text.split("```xml")[1].split("```")[0]
    if "```svg" in text:
        return text.split("```svg")[1].split("```")[0]
        
    if "xmlns=" in text and "<svg" in text:
        start = text.find("<svg")
        end = text.rfind("</svg>") + 6
        if start != -1 and end != -1:
            return text[start:end]

    return text

# --- 界面 ---
with gr.Blocks(title="StarVector-1B") as demo:
    gr.Markdown("# StarVector-1B Image to SVG")
    
    with gr.Row():
        with gr.Column():
            input_image = gr.Image(type="pil", label="上传图片")
            run_btn = gr.Button("生成 SVG", variant="primary")
        
        with gr.Column():
            html_output = gr.HTML(label="SVG 预览")
            code_output = gr.Code(language="xml", label="SVG 代码")

    run_btn.click(
        fn=convert_image_to_svg,
        inputs=[input_image],
        outputs=[code_output, html_output]
    )

demo.launch(server_name="0.0.0.0", server_port=7860)