rev app

app.py

@@ -38,34 +38,41 @@ def model_context():
     _tokenizer = None
     _model = None
     
+    def init_models():
+        nonlocal _tokenizer, _model
+        try:
+            print("กำลังโหลด tokenizer...")
+            from unsloth import FastVisionModel
+            base_model, _tokenizer = FastVisionModel.from_pretrained(
+                "unsloth/Llama-3.2-11B-Vision-Instruct",
+                use_gradient_checkpointing = "unsloth"
+            )
+            print("โหลด tokenizer สำเร็จ")
+
+            print("กำลังโหลดโมเดล fine-tuned...")
+            from transformers import AutoModelForVision2Seq
+            _model = AutoModelForVision2Seq.from_pretrained(
+                "Aekanun/Llama-3.2-11B-Vision-Instruct-XRay",
+                load_in_4bit=True,
+                torch_dtype=torch.float16
+            ).to('cuda')
+            FastVisionModel.for_inference(_model)
+            print("โหลดโมเดลสำเร็จ!")
+            return True
+        except Exception as e:
+            print(f"เกิดข้อผิดพลาดในการโหลดโมเดล: {str(e)}")
+            return False
+
     def decorator(func):
         def wrapper(*args, **kwargs):
-            nonlocal _tokenizer, _model
-            if _tokenizer is None:
-                print("กำลังโหลด tokenizer...")
-                from unsloth import FastVisionModel
-                base_model, _tokenizer = FastVisionModel.from_pretrained(
-                    "unsloth/Llama-3.2-11B-Vision-Instruct",
-                    use_gradient_checkpointing = "unsloth"
-                )
-                print("โหลด tokenizer สำเร็จ")
-
-            if _model is None:
-                print("กำลังโหลดโมเดล fine-tuned...")
-                from transformers import AutoModelForVision2Seq
-                _model = AutoModelForVision2Seq.from_pretrained(
-                    "Aekanun/Llama-3.2-11B-Vision-Instruct-XRay",
-                    load_in_4bit=True,
-                    torch_dtype=torch.float16
-                ).to('cuda')
-                FastVisionModel.for_inference(_model)
-                print("โหลดโมเดลสำเร็จ!")
-
             return func(_model, _tokenizer, *args, **kwargs)
         return wrapper
-    return decorator
 
-@model_context()
+    return init_models, decorator
+
+init_models, model_decorator = model_context()
+
+@model_decorator
 @spaces.GPU(duration=30)
 def process_image(_model, _tokenizer, image):
     if image is None:
@@ -107,12 +114,15 @@ def process_image(_model, _tokenizer, image):
         return f"เกิดข้อผิดพลาด: {str(e)}"
 
 print("กำลังเริ่มต้นแอปพลิเคชัน...")
-demo = gr.Interface(
-    fn=process_image,
-    inputs=gr.Image(type="pil"),
-    outputs=gr.Textbox(),
-    title="Medical Vision Analysis"
-)
-
-if __name__ == "__main__":
-    demo.launch()
+if init_models():
+    demo = gr.Interface(
+        fn=process_image,
+        inputs=gr.Image(type="pil"),
+        outputs=gr.Textbox(),
+        title="Medical Vision Analysis"
+    )
+    
+    if __name__ == "__main__":
+        demo.launch()
+else:
+    print("ไม่สามารถเริ่มต้นแอปพลิเคชันได้")