Upload inference.py

Updated the file to return the input tokens as well

inference.py

@@ -33,5 +33,5 @@ def predict(x, model, max_output_len = 30):
             # if xcol == config.end_token:
             #     break
     
-    return x[:, input_len:] # B, max_output_len
+    return x[:, :] # B, max_output_len