fix default sdpa

huggingface · Dec 16, 2024 · 9bd6c94 · 9bd6c94
1 parent a28ad19
commit 9bd6c94
Showing 1 changed file with 4 additions and 4 deletions.
diff --git a/src/transformers/modeling_utils.py b/src/transformers/modeling_utils.py
@@ -1531,10 +1531,10 @@ def _autoset_attn_implementation(
             config = cls._check_and_enable_flex_attn(config, hard_check_only=True)
         elif requested_attn_implementation in [None, "sdpa"] and not is_torch_xla_available():
             # use_flash_attention_2 takes priority over SDPA, hence SDPA treated in this elif.
-            # config = cls._check_and_enable_sdpa(
-            #     config,
-            #     hard_check_only=False if requested_attn_implementation is None else True,
-            # )
+            config = cls._check_and_enable_sdpa(
+                config,
+                hard_check_only=False if requested_attn_implementation is None else True,
+            )
 
             if (
                 torch.version.hip is not None