update

python/sglang/srt/sampling/penaltylib/penalizers/dry_penalty.py

@@ -74,18 +74,27 @@ def _apply(self, logits: torch.Tensor) -> torch.Tensor:
         max_back_length = 50  # Limit the backward match to 50 to prevent overflow
         for i in range(1):
             if self.input_ids is None:
-                # Initialize self.input_ids as needed, for example:
                 self.input_ids = []
 
             if self.output_ids is not None:
+                # Convert lists to tensors if necessary
+                if isinstance(self.output_ids, list):
+                    self.output_ids = torch.tensor(self.output_ids)
+
                 if i < len(self.input_ids):
+                    if isinstance(self.input_ids[i], list):
+                        self.input_ids[i] = torch.tensor(self.input_ids[i])
+
                     self.input_ids[i] = torch.cat(
                         [self.input_ids[i], self.output_ids], dim=0
                     )
                 else:
-                    self.input_ids.append(
-                        torch.cat([self.input_ids, self.output_ids], dim=0)
-                    )
+                    if isinstance(self.input_ids, list):
+                        self.input_ids = torch.tensor(self.input_ids)
+
+                    new_tensor = torch.cat([self.input_ids, self.output_ids], dim=0)
+                    self.input_ids.append(new_tensor)
+
                 input_ids = self.input_ids[i]
             else:
                 if i < len(self.input_ids) and self.input_ids[i] is not None: