Update

fffrog · fffrog · commit 3871a3fd9fdc · 2025-07-16T22:24:26.000+08:00
[ghstack-poisoned]
diff --git a/torch/_inductor/fx_passes/fuse_attention.py b/torch/_inductor/fx_passes/fuse_attention.py
@@ -584,7 +584,7 @@ def _sfdp_replacement_20(query, key, value, attn_mask, dropout_p):
 def _sfdp_pattern_24(query, key, value, attention_mask):
     """
     this pattern is for MBartForCausalLM/PLBartForCausalLM.
-    attn_mask has a differnt dtype with QKV.
+    attn_mask has a different dtype with QKV.
     there is no scale in sdpa.
     """
     bs = query.size(0)
diff --git a/torch/fx/traceback.py b/torch/fx/traceback.py
@@ -81,8 +81,8 @@ def __init__(
             self.from_node = []
 
         # cache the action string and dict representation for performance.
-        self._action_string = None
-        self._dict = None
+        self._action_string: Optional[str] = None
+        self._dict: Optional[dict[str, Any]] = None
 
     @property
     def name(self) -> str:
@@ -130,6 +130,7 @@ def to_dict(self) -> dict:
                 "from_node": [node.to_dict() for node in self.from_node],
             }
 
+        assert self._dict is not None
         return self._dict
 
     def __eq__(self, other: object):
