Dont pairwise check unfusable nodes in scheduler (pytorch#136682)

Gives 8% wall time speedup on n=1000 benchmark in pytorch#136429

Pull Request resolved: pytorch#136682
Approved by: https://github.com/ezyang, https://github.com/jansel, https://github.com/shunting314

torch/_inductor/scheduler.py

@@ -2672,6 +2672,8 @@ def check_all_pairs(nodes: List[BaseSchedulerNode]) -> None:
 
         buffer_names_grouping = collections.defaultdict(list)
         for node in nodes:
+            if self.unfusable_node(node):
+                continue
             for buf in node.used_buffer_names():
                 buffer_names_grouping[buf].append(node)
         for node_grouping in buffer_names_grouping.values():
@@ -2895,6 +2897,15 @@ def has_shared_data_after_reordering_loop(
 
         return self.score_fusion_memory(node1, node2) > 0
 
+    def unfusable_node(self, node: BaseSchedulerNode) -> bool:
+        """
+        Is this node unfusable under any conditions.
+        """
+        return (
+            isinstance(node, (ExternKernelSchedulerNode, NopKernelSchedulerNode))
+            and not node.is_template()
+        )
+
     def can_fuse(self, node1: BaseSchedulerNode, node2: BaseSchedulerNode) -> bool:
         """
         Determine if it is possible to combine node1 and node2 into a