fix qwen2 tokenizer

Author: akaitsuki-ii

diffsynth_engine/tokenizers/qwen2.py

@@ -197,8 +197,8 @@ def __call__(
         encoded.fill_(self.pad_token_id)
         attention_mask = torch.zeros(len(texts), max_length, dtype=torch.long)
         for i, ids in enumerate(batch_ids):
-            if len(ids) > self.model_max_length:
-                ids = ids[: self.model_max_length]
+            if len(ids) > max_length:
+                ids = ids[:max_length]
                 ids[-1] = self.eos_token_id
             if padding_side == "right":
                 encoded[i, : len(ids)] = torch.tensor(ids)

diffsynth_engine/utils/parallel.py

@@ -304,12 +304,14 @@ def wrap_for_parallel(module: PreTrainedModel):
             if rank == 0:
                 queue_out.put(res)
             dist.barrier()
-    except Exception as e:
+    except Exception:
         import traceback
 
-        traceback.print_exc()
-        logger.error(f"Error in worker loop (rank {rank}): {e}")
-        queue_out.put(e)  # any exception caught in the worker will be raised to the main process
+        msg = traceback.format_exc()
+        err = RuntimeError(msg)
+        logger.error(f"Error in worker loop (rank {rank}): {msg}")
+        if rank == 0:
+            queue_out.put(err)  # any exception caught in the worker will be raised to the main process
     finally:
         del module
         torch.cuda.synchronize()