Update app.py

app.py

@@ -162,22 +162,32 @@ def convert_model_to_onnx():
     dummy_prompt = "tara: Hello"
     dummy_input = tokenizer(dummy_prompt, return_tensors="pt").input_ids.to(device)
     file_path = "orpheus_model.onnx"
+    
+    # Ensure the model is in evaluation mode and not compiled
+    model.eval()
+    
+    # Reset Torch Dynamo to avoid FX-tracing issues during export.
+    if hasattr(torch, "_dynamo"):
+        try:
+            torch._dynamo.reset()
+            print("Torch Dynamo reset before ONNX export")
+        except Exception as e:
+            print(f"Warning: Torch Dynamo reset failed - {e}")
+            
     try:
-        # Disable Torch Dynamo during export to avoid FX-tracing issues.
-        with torch._dynamo.disable():
-            torch.onnx.export(
-                model,
-                dummy_input,
-                file_path,
-                export_params=True,
-                opset_version=14,
-                input_names=["input_ids"],
-                output_names=["logits"],
-                dynamic_axes={
-                    "input_ids": {0: "batch_size", 1: "sequence_length"},
-                    "logits": {0: "batch_size", 1: "sequence_length"}
-                },
-            )
+        torch.onnx.export(
+            model,
+            dummy_input,
+            file_path,
+            export_params=True,
+            opset_version=14,
+            input_names=["input_ids"],
+            output_names=["logits"],
+            dynamic_axes={
+                "input_ids": {0: "batch_size", 1: "sequence_length"},
+                "logits": {0: "batch_size", 1: "sequence_length"}
+            },
+        )
         return f"Model converted to ONNX and saved as '{file_path}'."
     except Exception as e:
         return f"Error during ONNX conversion: {e}"