← 返回事件
持续讨论AI发版

viable/strict/1789939464: [Inductor] Allow custom-op autotuning without fallback (#197446)

图:PyTorch Releases

发生了什么

Add an include_fallback option to custom-op autotuning and thread it through both standard and range-based lowering. It defaults to True , preserving the existing behavior of including the custom op itself as an external fallback choice. Callers can set it to False when the custom op is used only as an autotuning/rewriting vehicle and should not itself be emitted as a candidate. This is used by facebookexperimental/…

摘要按规则整理自下方来源原文

为什么在扩散

来源