[Inductor] Limit fusions to a node distance of 64 (#154688)

fix for https://github.com/pytorch/pytorch/issues/154652 and https://fb.workplace.com/groups/1075192433118967/permalink/1484799079148049/

[window 128 dashboard run here w/ no regressions](https://hud.pytorch.org/benchmark/compilers?dashboard=torchinductor&startTime=Sun%2C%2001%20Jun%202025%2006%3A38%3A41%20GMT&stopTime=Sun%2C%2008%20Jun%202025%2006%3A38%3A41%20GMT&granularity=hour&mode=inference&dtype=bfloat16&deviceName=cuda%20(a100)&lBranch=mlazos/fuse-window&lCommit=8576f00ebfa53567d7bddc89d9882df9eb990561&rBranch=main&rCommit=9d59b516e9b3026948918e3ff8c2ef55a33d13ad)

Pull Request resolved: https://github.com/pytorch/pytorch/pull/154688
Approved by: https://github.com/eellison, https://github.com/Raymo111

torch/_inductor/config.py

@@ -582,6 +582,9 @@ max_epilogue_benchmarked_choices = 1
 # how many nodes to allow into a single fusion
 max_fusion_size = 64
 
+# how many nodes to attempt pairwise fusion with in a buffer group
+max_fusion_buffer_group_pairwise_attempts = 64
+
 # max number of inputs to generate cat as a pointwise op with masked laods
 max_pointwise_cat_inputs = 8
 

torch/_inductor/scheduler.py

@@ -3229,7 +3229,11 @@ class Scheduler:
 
         def check_all_pairs(nodes: list[BaseSchedulerNode]) -> None:
             for node1_index, node1 in enumerate(nodes):
-                for node2 in nodes[node1_index + 1 :]:
+                for node2 in nodes[
+                    node1_index + 1 : node1_index
+                    + 1
+                    + config.max_fusion_buffer_group_pairwise_attempts
+                ]:
                     key = (node1, node2)
                     if key in seen:
                         continue