Update utils.py

utils.py

@@ -271,14 +271,19 @@ def predict_prices(data, model, tokenizer, prediction_days=30):
         input_sequence = prices[-context_length:]
         
         # Create prediction input
-        # CRITICAL FIX: Cast input to LongTensor to satisfy model embedding layer expectation.
-        prediction_input = torch.tensor(input_sequence).unsqueeze(0).long().to(model.device)
+        # CRITICAL FIX: Revert to float, but use max_new_tokens for generation.
+        # This forces the model's custom generation logic to handle the raw floats, 
+        # as manual quantization/token mapping is impossible without the Chronos tokenizer.
+        prediction_input = torch.tensor(input_sequence).unsqueeze(0).float().to(model.device)
         
         # Generate predictions
         with torch.no_grad():
-            # NOTE: Removed prediction_length from generate call as it was causing failures.
+            # CRITICAL FIX: Use max_new_tokens, which is standard for Seq2Seq generation length.
+            # Removed prediction_length keyword and added do_sample to align with typical Chronos usage.
             forecast = model.generate(
                 prediction_input,
+                max_new_tokens=prediction_days, 
+                do_sample=True 
             )
         
         # Handle complex Chronos output: [batch_size, num_samples, prediction_length]
@@ -287,8 +292,7 @@ def predict_prices(data, model, tokenizer, prediction_days=30):
         output_tensor = forecast[0] if isinstance(forecast, (tuple, list)) else forecast
         
         # 2. Average across the samples (dim=1) and convert to a simple 1D numpy array
-        # predictions = output_tensor.mean(dim=1).squeeze().cpu().numpy()
-        # To avoid unexpected dimensions, let's simplify averaging:
+        # output_tensor shape: [1, num_samples, prediction_length] (If sampling, the float values are the predictions)
         predictions = output_tensor.float().mean(dim=1).squeeze().cpu().numpy()
         
         # Handle case where predictions is a single scalar (convert to array for safety)