Fix torch_compile parameter error

app.py

@@ -20,7 +20,23 @@ TITLE = """
 
 print("🚀 Loading reliable JoyCaption system...")
 
-@spaces.GPU(duration=45)  # Short duration to prevent timeouts
+# Load model and processor at startup (ONCE)
+print("📦 Loading model and processor at startup...")
+processor = AutoProcessor.from_pretrained(
+    MODEL_PATH,
+    low_cpu_mem_usage=True
+)
+
+model = LlavaForConditionalGeneration.from_pretrained(
+    MODEL_PATH,
+    torch_dtype=torch.bfloat16,
+    device_map="auto",
+    low_cpu_mem_usage=True
+)
+model.eval()
+print("✅ Model loaded and ready!")
+
+@spaces.GPU(duration=30)  # Shorter duration since no model loading
 @torch.no_grad()
 def caption_image_optimized(image, style, length):
     """Ultra-optimized JoyCaption that won't get stuck"""
@@ -31,23 +47,7 @@ def caption_image_optimized(image, style, length):
     start_time = time.time()
     
     try:
-        print(f"📸 Loading JoyCaption at {time.time() - start_time:.1f}s...")
-        
-        # Load with maximum optimization
-        processor = AutoProcessor.from_pretrained(
-            MODEL_PATH,
-            low_cpu_mem_usage=True
-        )
-        
-        model = LlavaForConditionalGeneration.from_pretrained(
-            MODEL_PATH,
-            torch_dtype=torch.bfloat16,
-            device_map="auto",
-            low_cpu_mem_usage=True
-        )
-        model.eval()
-        
-        print(f"✅ Model loaded at {time.time() - start_time:.1f}s")
+        print(f"🎯 Starting generation at {time.time() - start_time:.1f}s...")
         
         # Optimized prompts based on length
         if length == "Short":
@@ -125,8 +125,8 @@ def caption_image_optimized(image, style, length):
                 result = result.split(split_marker)[-1].strip()
                 break
         
-        # Clean up IMMEDIATELY and AGGRESSIVELY
-        del model, processor, inputs, output
+        # Clean up inputs and output (but NOT the global model/processor)
+        del inputs, output
         torch.cuda.empty_cache()
         gc.collect()
         
@@ -141,10 +141,10 @@ def caption_image_optimized(image, style, length):
     except Exception as e:
         # Emergency cleanup
         try:
-            if 'model' in locals():
-                del model
-            if 'processor' in locals():
-                del processor
+            if 'inputs' in locals():
+                del inputs
+            if 'output' in locals():
+                del output
             torch.cuda.empty_cache()
             gc.collect()
         except: