PR #180285: [ROCm][inductor][UT] Allow 1D bias_addmm candidate generation

• Repository: pytorch/pytorch
• Author: naromero77amd
• State: open | merged: False
• Link: https://github.com/pytorch/pytorch/pull/180285

Description (problem / solution / changelog)

Fixes #180031 Fixes #179967

This follows up on #179087 and the original regression from #177130. #178929 updated tuned_addmm to inspect the original ROCm bias input, but bias_addmm still only stayed eligible when that bias looked like the expanded stride-0 form. #179087 then updated the ROCm bias_addmm heuristic to accept a native 1D bias input, but tuned_addmm could still silently drop bias_addmm before the heuristic ran. This change keeps bias_addmm eligible when ROCm addmm autotuning sees a native 1D bias input so max-autotune can still consider the fused bias path instead of silently dropping that choice.

Made with Cursor

cc @jeffdaily @sunway513 @jithunnair-amd @pruthvistony @ROCmSupport @jataylo @hongxiayang @pragupta @jerrymannil @xinyazhang @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben @mlazos

Changed files

• torch/_inductor/kernel/mm.py (modified, +7/-2)