trunk/ff9d60a3e6a3002b92dd89bd9949c70a82e3688f: [inductor] Reuse formatted origin nodes across kernel comments (#194010)
- PyTorch 近 90 天出现 211 次
- PyTorch 近 90 天第 208 次发版
- 上一次：同一天稍早 · trunk/e2d110a75ba155e0699f6de302a055154551e460: [distributed] Stabilize NCCL timeout tests (#194954)
发生了什么
Every generated kernel carries a comment naming the graph fragment it came from, and building it re-formats each of the kernel's origin nodes with Node.format_node(include_tensor_metadata=True). The same node is an origin of many kernels, so on a large graph this rebuilds the same strings over and over: 258k format_node calls to emit comments for 1266 kernels on one production model, which cProfile puts at a larger…
摘要按规则整理自下方来源原文