Update spack_build.yml to fix cmake pulling llvm #698
There are no files selected for viewing
| Original file line number | Diff line number | Diff line change |
|---|---|---|
|
|
@@ -99,9 +99,9 @@ jobs: | |
| strategy: | ||
| matrix: | ||
| spack_spec: | ||
| - hiop@develop+mpi~raja~shared~kron~sparse ^openmpi | ||
| - hiop@develop~mpi~raja~shared~kron~sparse | ||
| - hiop@develop~mpi+raja~shared~kron~sparse | ||
|
|
||
| # We will need coinhsl for this, but what are the rules for using | ||
| # a coinhsl tarball? | ||
|
|
@@ -130,8 +130,11 @@ jobs: | |
| spack: | ||
| specs: | ||
| - ${{ matrix.spack_spec }} target=x86_64_v2 | ||
| - openblas | ||
| - libevent~openssl | ||
| concretizer: | ||
| reuse: true | ||
| unify: true | ||
|
There was a problem hiding this comment. @cameronrutherford how did you figure out these? There was a problem hiding this comment. reuse true is just a sane default. It makes sure you always pull a version of a package from a build cache when possible. There are different options as detailed here. I was having trouble with the blas provider specification, so I decided to put it in the core spack spec. Then, by specifying unify to be true, I was able to ensure the version of hiop was also built with the version of blas that I wanted. |
||
| config: | ||
| source_cache: $SPACK_CACHE/source_cache | ||
| misc_cache: $SPACK_CACHE/misc_cache | ||
|
|
@@ -141,10 +144,13 @@ jobs: | |
| padded_length: False | ||
| mirrors: | ||
| local-buildcache: oci://${{ env.REGISTRY }}/${{ env.IMAGE_NAME }} | ||
| spack: https://binaries.spack.io/develop | ||
|
There was a problem hiding this comment. Maybe this is a bad idea |
||
| packages: | ||
| all: | ||
| require: "%gcc" | ||
| # https://spack.readthedocs.io/en/latest/packages_yaml.html#setting-requirements-on-virtual-specs | ||
| cmake: | ||
| require: "~qtgui~ncurses" | ||
|
There was a problem hiding this comment. @cameronrutherford how did you figure out these parameters? There was a problem hiding this comment. So this was part of my initial debugging. As we are using This forces any build of CMake to have these options disabled. See this job's concretization step for an example of what I mean. Compare that to a later passing pipeline here. Notice how the updated config results in a much smaller build. Sure we could probably get the more complicated builds to work too, but a minimal build is faster, and less error prone - we shouldn't build things we don't need. There was a problem hiding this comment. What pointed me to this was noticing the delta in the concretizations. I looked through the logs, noticed that we were having builds fail on a dependency we don't normally build, and then figured out what package was to blame. Then I worked with that package's spack build config to figure out what to enable / disable. https://packages.spack.io is a great way to get this information quickly. |
||
| EOF | ||
|
|
||
| - name: Configure GHCR mirror | ||
|
|
||
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
@nychiang this build is failing because for some reason it's trying to build
libflame, and notopenblas.Looks at the concretization comparison between builds, and you'll see that a 3 hour build would make no sense for the smallest build of the batch here.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Why would that spec possibly need python for this build?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
python is required via
hiop -> amdlibflame -> aocl-utils -> doxygen -> graphviz -> gtkplus -> meson -> python-xxx
why do we need amdlibflame? Concretization shows that it installs a lot of packages.
@cameronrutherford