fix

Author: Qiaolin-Yu

python/sglang/srt/lora/layers.py

@@ -219,9 +219,11 @@ def set_lora_info(
             self.B_buffer_qkv[:, output_dim_q + output_dim_kv :, :].copy_(
                 B_buffer_kv[1]
             )
-
-            # Offsets of q/k/v in output dimension
-            self.output_offset = torch.tensor(
+            if not hasattr(self, "output_offset") or self.output_offset is None:
+                self.output_offset = torch.empty(
+                    4, dtype=torch.int32, device=B_buffer_q.device
+                )
+            self.output_offset[:4] = torch.tensor(
                 [
                     0,
                     output_dim_q,