search

ORNL / hykkt

HyKKT is an experimental linear solver for Karush-Kuhn-Tucker type of systems.
Other
12 stars 2 forks source link

Support different CMAKE_CUDA_ARCHITECTURES depending on Platforms #12

Open ryandanehy opened 1 year ago

ryandanehy commented 1 year ago

Should be a relative easy fix to Pipeline ->CMAKE -> build script. Looking to support 60,70,75,80.

ryandanehy commented 1 year ago

Instead of partition=$(scontrol show jobid -dd $jobid | grep Partition)

which provides an output that will always result in a100 (due Alphabetical order)

Use partition=$(scontrol show jobid -dd $jobid | grep NodeList)

which provides an output like "dlv06"

pelesh commented 1 year ago

Instead of partition=$(scontrol show jobid -dd $jobid | grep Partition)

which provides an output that will always result in a100 (due Alphabetical order)

Use partition=$(scontrol show jobid -dd $jobid | grep NodeList)

which provides an output like "dlv06"

@ryandanehy: How about building a "fat" CUDA binary with support for all four arches compiled? Then you can deploy the same binary on each CI pipeline.