ciflow/torchtitan/201326: [Inductor] Reuse identical custom-op layout constraints

发生了什么

Summary: Summary Reuse prepared inputs when the same IR value needs identical exact strides. We observed lower peak memory in a workload using this pattern. Problem Some attention workloads pass the same tensor as both K and V. Independent argument preparation can materialize it twice. Minimal CUDA reproducer on this branch: torch.Tensor: print("shared buffer:", k.data_ptr() == v.data_ptr()) return k + v pair.regist…

摘要按规则整理自下方来源原文

来源