Update gradio1.py

gradio1.py

@@ -1,13 +1,3 @@
-"""
-Gradio 推理界面 - 多模态 Dense Transformer (适配 Qwen Tokenizer 版)
-
-用法:
-pip install -r requirements.txt
-# requirements.txt 至少包含:
-# torch>=1.12, transformers, pillow, gradio
-python app_gradio.py --checkpoint /path/to/final_model.pt --tokenizer Qwen/Qwen2.5-7B-Instruct --port 7860 --share False
-"""
-
 import os
 import argparse
 from pathlib import Path
@@ -20,15 +10,11 @@ from transformers import AutoTokenizer
 
 # UI
 import gradio as gr
-
-# 本项目代码引用（按你的工程结构调整）
 from model import MultiModalDenseTransformer
 from continual_learning import UnifiedMultiModalPreprocessor
 
-# 设置国内镜像（如需要）
 os.environ["HF_ENDPOINT"] = "https://hf-mirror.com"
 
-# ---- 与你原来保持一致的图像预处理 ----
 from torchvision import transforms
 image_transform = transforms.Compose([
     transforms.Resize((224, 224)),
@@ -37,7 +23,6 @@ image_transform = transforms.Compose([
                          std=[0.229, 0.224, 0.225]),
 ])
 
-# -------- ModelInference 类（轻微改写） --------
 class ModelInference:
     def __init__(self, checkpoint_path: str, tokenizer_name: str, config_path: Optional[str] = None, device: str = 'cuda' if torch.cuda.is_available() else 'cpu'):
         self.device = torch.device(device)
@@ -52,7 +37,6 @@ class ModelInference:
             with open(config_path, 'r') as f:
                 self.config = json.load(f)
         else:
-            # 采用你原始脚本中的默认 config（可按需调整）
             self.config = {
                 'model_dim': 1536,
                 'vocab_size': len(self.tokenizer),
@@ -77,7 +61,6 @@ class ModelInference:
 
         print(f"Loading checkpoint from {checkpoint_path}...")
         checkpoint = torch.load(checkpoint_path, map_location=self.device)
-        # 支持 checkpoint 包含 'model_state_dict' 的情况
         state_dict = checkpoint.get('model_state_dict', checkpoint) if isinstance(checkpoint, dict) else checkpoint
 
         new_state_dict = {}
@@ -158,12 +141,11 @@ class ModelInference:
             traceback.print_exc()
             return f"Error: {e}"
 
-# -------- Gradio UI 部分 --------
 def build_ui(model_instance):
     with gr.Blocks(title="MultiModal Dense Transformer - Gradio", css="""
         .gradio-container { max-width: 900px; margin: auto; }
     """) as demo:
-        gr.Markdown("## 🚀 多模态在线推理（文本 + 图片）")
+        gr.Markdown("##  多模态在线推理（文本 + 图片）")
         with gr.Row():
             with gr.Column(scale=3):
                 txt = gr.Textbox(label="Prompt (Instruction)", placeholder="请输入指令或问题...", lines=5)
@@ -198,7 +180,6 @@ def build_ui(model_instance):
 
     return demo
 
-# -------- CLI / main --------
 def main():
     parser = argparse.ArgumentParser()
     parser.add_argument("--checkpoint", type=str, default="/root/multimodal/checkpoints/posttrain/final_model.pt")
@@ -208,7 +189,6 @@ def main():
     parser.add_argument("--share", type=lambda x: x.lower() in ("true","1","yes"), default=True)
     args = parser.parse_args()
 
-    # 如果 default 的 final_model 不存在，尝试寻找最近 step
     if not Path(args.checkpoint).exists():
         possible = list(Path("checkpoints/pretrain").glob("step_*.pt"))
         if possible:
@@ -220,7 +200,6 @@ def main():
     global model_instance
     model_instance = ModelInference(args.checkpoint, args.tokenizer, args.config)
 
-    # 启动 Gradio（使用 share 参数决定是否创建公网链接）
     demo = build_ui(model_instance)
     demo.launch(server_port=args.port, share=args.share)