viable/strict/1789161800: [BE][MPS] Route add/sub through the common add_stub (#196661)
- PyTorch 近 90 天出现 814 次
- PyTorch 近 90 天第 803 次发版
- 上一次：同一天稍早 · trunk/599651c7aef6f85740f997fcc4aed9642a5f9a03: [ROCm][CD] Add gfx1250 (MI450) to the nightly wheel arch list (#196610)
发生了什么
MPS had its own add_out_mps / sub_out_mps pair backed by a shared add_sub_lerp_template , which predated the Metal binary-kernel machinery. Registering add_stub from BinaryKernel.mm allows one to unify add and sub together, though still keep shortcut for non-alpha version as it's signiicantly faster thanks to ILP. Slightly changes behavior when alpha == 0 - before it just copied self , but now it runs the ops, which…
摘要按规则整理自下方来源原文