← 返回事件
持续讨论AI发版2.0

viable/strict/1791287563: [inductor-perf] allow fusion into template epilogue inputs. (#198809)

发生了什么

Template prologue fusion lets a pointwise producer be fused into the code where a template loads an input (load_input()). Inputs that a template consumes only in its epilogue, like the addmm/baddbmm bias, could not take a fused producer. A computed bias was always written to memory by a separate kernel and read back in the template epilogue: tmp = bias * 2.0 - 1.0 addmm(tmp, a, b) This PR lets producers of those inp…

摘要按规则整理自下方来源原文

为什么在扩散

来源