Update app.py

app.py

@@ -1,30 +1,72 @@
+# import gradio as gr
+# import spaces
+
+# # Initialize the model only once, outside of any function
+# # Ensure that CUDA initialization happens within the worker process
+# model_pipe = None
+
+# @spaces.GPU
+# def generate(model_name, image, text):
+#     global model_pipe
+#     import torch
+#     torch.jit.script = lambda f: f
+    
+#     from t2v_metrics import VQAScore, list_all_vqascore_models
+    
+#     if model_pipe is None:
+#         print("Initializing model...")
+#         model_pipe = VQAScore(model="clip-flant5-xl", device="cuda")  # our recommended scoring model
+#         # model_pipe.to("cuda")
+    
+#     print(list_all_vqascore_models())
+#     print("Image:", image)
+#     print("Text:", text)
+    
+#     print("Generating!")
+#     result = model_pipe(images=[image], texts=[text])
+#     return result
+
 import gradio as gr
 import spaces
+import torch
+import os
 
-# Initialize the model only once, outside of any function
-# Ensure that CUDA initialization happens within the worker process
+# Global model variable, but do not initialize or move to CUDA here
 model_pipe = None
 
 @spaces.GPU
 def generate(model_name, image, text):
     global model_pipe
-    import torch
-    torch.jit.script = lambda f: f
+    
+    # Debugging lines to trace CUDA initialization
+    print(f"PID: {os.getpid()}")
+    print(f"Before import: CUDA available: {torch.cuda.is_available()}")
+    
+    torch.jit.script = lambda f: f  # Avoid script error in lambda
     
     from t2v_metrics import VQAScore, list_all_vqascore_models
     
+    print(f"After import: CUDA available: {torch.cuda.is_available()}")
+    
+    # Worker Process: Perform all GPU-related initializations here
     if model_pipe is None:
-        print("Initializing model...")
+        print("Initializing model in PID:", os.getpid())
         model_pipe = VQAScore(model="clip-flant5-xl", device="cuda")  # our recommended scoring model
-        # model_pipe.to("cuda")
+        print(f"Model initialized: CUDA available: {torch.cuda.is_available()}")
     
-    print(list_all_vqascore_models())
-    print("Image:", image)
-    print("Text:", text)
+    print(list_all_vqascore_models())  # Debug: List available models
+    print("Image:", image)  # Debug: Print image path
+    print("Text:", text)  # Debug: Print text input
     
     print("Generating!")
-    result = model_pipe(images=[image], texts=[text])
-    return result
+    # Wrap the model call in a try-except block to capture and debug CUDA errors
+    try:
+        result = model_pipe(images=[image], texts=[text])  # Perform the model inference
+    except RuntimeError as e:
+        print(f"RuntimeError during model inference: {e}")
+        raise e
+    
+    return result  # Return the result
 
 iface = gr.Interface(
     fn=generate,  # function to call