Provide unsurprising behavior for tensile threads

Open davidd-amd opened this issue 1 year ago • 2 comments

On two occasions now I've helped users untangle surprising behavior when attempting to pass parallelization settings to TensileCreateLibrary. This PR corrects some of the issues that cause the surprising behavior such as a cap at 64 threads. We have users that want to request 128 and 256 parallel jobs.

Having made these changes, we need to confirm that the CI configuration is still throttling to avoid OOM issues (see https://github.com/ROCm/rocJenkins/pull/919). The rocJenkins PR was merged so we should be ok.

FYI @loggrado @amd-aakash

Dec 19 '24 15:12 davidd-amd

should we notify build team to limit the number of parallel jobs at their end?

Dec 26 '24 09:12 jichangjichang

should we notify build team to limit the number of parallel jobs at their end?

I don't think this will be a problem because we are using cpu nodes for builds.

Jan 10 '25 19:01 davidd-amd