There is an optimization opportunity for the offload test suite in testing different configs.
By splitting out a job into a build job and a test job, and establishing an artifact transfer system, machines that are using the GPU for testing can also be utilized for building, and machines using the CPU for building can also take a separate testing job, so we can maximally utilize hardware.
Separately, we can take advantage of parallel builds, since all we truly need is for tests to be run on a specific GPU, but any machine with the right architecture can build compilers. We should change the workflow logic so that other vendor machines with say, AMD or Intel compilers, can build the compiler for say, an NVidia run. Since AMD and Nvidia aren't part of the default non-test-all label workflows, using those two machines will help lighten the load on Intel / QC.
There is an optimization opportunity for the offload test suite in testing different configs.
By splitting out a job into a build job and a test job, and establishing an artifact transfer system, machines that are using the GPU for testing can also be utilized for building, and machines using the CPU for building can also take a separate testing job, so we can maximally utilize hardware.
Separately, we can take advantage of parallel builds, since all we truly need is for tests to be run on a specific GPU, but any machine with the right architecture can build compilers. We should change the workflow logic so that other vendor machines with say, AMD or Intel compilers, can build the compiler for say, an NVidia run. Since AMD and Nvidia aren't part of the default non-test-all label workflows, using those two machines will help lighten the load on Intel / QC.