🐛 Describe the bug

Since triton-lang/triton@76ed94d AMD backend has removed stream pipeliner v1 in favour of v2. This change impacts PyTorch as we should no longer default to num_stages=0 for mm configs as this will now error out.

assert options.num_stages != 0, ("Triton AMD backend pipeliner has been updated. "
                                 "We used to trigger software pipelining with "
                                 "num_stages == 0. Now it will not happen anymore; "
                                 "please update to use num_stages == 2 for "
                                 "equivalent behavior in the past.")

Opening an issue to track updating the PyTorch code to match the new pipeliner as part of the triton upgrade for release/2.6 (#139175)

Ideally we should resolve this in a way that allows older triton to still use num_stages==0

Versions

Triton commit: 3549db850c7daf86624a624378e12d2a5a2bd236
PyTorch commit: 4dd04db

cc @jeffdaily @sunway513 @jithunnair-amd @pruthvistony @ROCmSupport @dllehr-amd @hongxiayang @naromero77amd @bertmaher @int3 @davidberard98 @nmacchioni @chenyang78 @embg @peterbell10 @aakhundov