ciflow/torchtitan/196847: [Inductor] Account NonOwningLayout storage in peak-memory planning
- PyTorch 近 90 天出现 821 次
- PyTorch 近 90 天第 810 次发版
- 上一次：同一天稍早 · trunk/84e524623ea4754a748936bf1ba6ecaaa92c3ae6: add kernel shared memory get_graph_data() (#196527)
发生了什么
Summary: NonOwningLayout buffers do not allocate independent storage. Count scheduler-owned aliases as zero bytes, transfer their consumers to the ultimate storage owner, and keep owners of graph-output aliases live. Preserve conservative sizing when the owner is a graph input. Test Plan: python3 -m py_compile fbcode/caffe2/torch/_inductor/memory.py fbcode/caffe2/test/inductor/test_memory.py arc lint fbcode/caffe2/t…
摘要按规则整理自下方来源原文