Update utils.py

utils.py

@@ -271,34 +271,47 @@ def predict_prices(data, model, tokenizer, prediction_days=30):
         input_sequence = prices[-context_length:]
         
         # Create prediction input
-        prediction_input = torch.tensor(input_sequence).unsqueeze(0).float()
+        # CRITICAL FIX: Ensure tensor is on the same device as the model (GPU/auto-mapped)
+        prediction_input = torch.tensor(input_sequence).unsqueeze(0).float().to(model.device)
         
         # Generate predictions
         with torch.no_grad():
+            # CRITICAL FIX: Removed prediction_length keyword which caused warnings/potential internal failure.
+            # Model will use its default configuration for generation length.
             forecast = model.generate(
                 prediction_input,
-                prediction_length=prediction_days
             )
         
-        # FIXED: Handle the case where the output is a tuple (common for Seq2Seq models)
-        if isinstance(forecast, tuple):
-             predictions = forecast[0].numpy()
-        else:
-             # Ensure the output is a numpy array (even if it's already a tensor)
-             predictions = forecast[0].numpy()
+        # CRITICAL FIX: Handle complex Chronos output: [batch_size, num_samples, prediction_length]
         
-        # Calculate prediction statistics
+        # 1. Get the actual tensor from the tuple/list if necessary
+        output_tensor = forecast[0] if isinstance(forecast, (tuple, list)) else forecast
+        
+        # 2. Average across the samples (dim=1) and ensure it is a simple 1D numpy array
+        # output_tensor shape: [1, num_samples, prediction_length]
+        # .mean(dim=1) averages the probabilistic samples
+        # .squeeze() removes unnecessary dimensions (like batch size 1)
+        predictions = output_tensor.mean(dim=1).squeeze().cpu().numpy()
+        
+        # Handle case where predictions is a single scalar (convert to array for safety)
+        if predictions.ndim == 0:
+             predictions = np.array([predictions.item()])
+        
+        # Calculation is now safe as predictions is a 1D array
         last_price = prices[-1]
         predicted_high = np.max(predictions)
         predicted_low = np.min(predictions)
         predicted_mean = np.mean(predictions)
         change_pct = ((predicted_mean - last_price) / last_price) * 100
         
+        # Use actual prediction length from the output tensor
+        pred_len = len(predictions)
+        
         return {
             'values': predictions,
             'dates': pd.date_range(
                 start=data.index[-1] + timedelta(days=1),
-                periods=prediction_days,
+                periods=pred_len, 
                 freq='D'
             ),
             'high_30d': predicted_high,
@@ -307,7 +320,7 @@ def predict_prices(data, model, tokenizer, prediction_days=30):
             'change_pct': change_pct,
             'summary': f"""
 AI Model: Amazon Chronos-Bolt
-Prediction Period: {prediction_days} days
+Prediction Period: {pred_len} days
 Expected Change: {change_pct:.2f}%
 Confidence: Medium (based on historical patterns)
 Note: AI predictions are for reference only and not financial advice
@@ -322,7 +335,7 @@ Note: AI predictions are for reference only and not financial advice
             'low_30d': 0,
             'mean_30d': 0,
             'change_pct': 0,
-            'summary': 'Prediction unavailable due to model error'
+            'summary': f'Prediction unavailable due to model error: {e}'
         }
 
 def create_price_chart(data, indicators):