Resolving the Save-Load Discontinuity Caused by Non-Contiguous Tensors

paddleformers/transformers/conversion_utils.py

@@ -745,7 +745,7 @@ def fn(x, is_column=True, transpose=False, is_old_qkv=False, is_naive_2fuse=Fals
             return None
         if transpose:
             if isinstance(x, paddle.Tensor):
-                x = paddle.transpose(x, [1, 0])
+                x = paddle.transpose(x, [1, 0]).contiguous()
             else:
                 x = np.transpose(x, [1, 0])
         if is_old_qkv:
@@ -1252,7 +1252,7 @@ def convert_transpose_selected_weights(state_dict: dict, transpose_weight_keys:
                     continue
                 for trans_key in transpose_weight_keys:
                     if re.search(f"\.{trans_key}\.weight$", key) or re.fullmatch(f"^{trans_key}\.weight$", key):
-                        state_dict[key] = state_dict.pop(key).transpose([-1, -2])
+                        state_dict[key] = state_dict.pop(key).transpose([-1, -2]).contiguous()
         return state_dict
 
     @classmethod

paddleformers/transformers/model_utils.py

@@ -397,7 +397,7 @@ def _is_need_transpose(key):
 
     def _transpose_hf_weight(key, weight):
         if _is_need_transpose(key):
-            return weight.transpose([-1, -2])
+            return weight.transpose([-1, -2]).contiguous()
         return weight
 
     part_state_dict = {}