[umbrella] exago / hiop: the packaged build cannot produce a working GPU sparse solver

[huojiangming]

Consolidates the five reports I filed on 2026-09-22 — #6598, #6599, #6600, #6601, #6602 —
into one, and closes those. They are all one story, and five separate issues was the wrong
way to tell it. (LLNL/hiop asked me to consolidate on their tracker for the same reason,
llnl/hiop#801; this is the same medicine applied here.)

The story: the packaged ExaGO and HiOp cannot give you a working GPU sparse solver, and
nothing says so.

1. spack install exago with no variants has no OPFLOW solver — was #6598

exago/package.py has variant("hiop", default=False) and variant("ipopt", default=False).
With neither, the binary builds and installs, and then has no solver for OPFLOW at all. A
default install that cannot solve the package's main problem is a surprising default; at the
least the recipe could say so.

2. +sparse forces a licensed dependency — was #6599

hiop/package.py:

95:  variant("sparse", default=False, description="Enable/Disable Sparse linear algebra")
198: depends_on("coinhsl+blas", when="+sparse")
236: self.define_from_variant("HIOP_SPARSE", "sparse"),
237: self.define_from_variant("HIOP_USE_COINHSL", "sparse"),

so asking for HiOp's sparse interface pulls in COINHSL, which needs a licence and a manually
fetched tarball. HiOp itself does not require it. We build upstream HiOp develop
(fc0e881) like this, and it works:

HIOP_SPARSE:BOOL=ON
HIOP_USE_COINHSL:BOOL=OFF

with STRUMPACK and ReSolve as the sparse backends. The coupling is in this recipe, not in
HiOp. A separate coinhsl variant would decouple them:

variant("coinhsl", default=False, description="Use the licensed COINHSL solvers (MA57/MA27)")
depends_on("coinhsl+blas", when="+coinhsl")
...
self.define_from_variant("HIOP_USE_COINHSL", "coinhsl"),

I have not sent that as a pull request because it changes what +sparse resolves to for
people who already depend on it, which is your call rather than mine. Say the word and it
is a five-line change.

3. The GPU sparse solver is off by default and cannot be enabled on its own — was #6600

Following from 2: the combination that gives a GPU sparse solve (+sparse with a GPU
backend, no COINHSL) is not expressible through the variants as they stand.

4. E4S has never shipped a HiOp with any sparse backend — was #6601

A consequence of 2 and 3 rather than a separate defect: because +sparse drags in a
licensed package, the distributed builds leave it off.

5. MA27 cannot stand in for MA57 — was #6602

Where a build does have COINHSL, MA27 does not substitute for MA57 in HiOp's safe mode: the
code asks for MA57 specifically (hiopKKTLinSysSparse.cpp), and a build with only MA27
behaves as a build with neither — see llnl/hiop#810 for what that does.

---

The recipes list @ryandanehy, @cameronrutherford, @pelesh (exago) and @nychiang, @cnpetra,
@pelesh (hiop) as maintainers — one mention rather than five, and no expectation of a quick
answer. If any of this is intended as it stands, say so and I will close this.

[huojiangming]

@cameronrutherford @pelesh — listed maintainers of the exago package; @cnpetra @nychiang @pelesh — listed maintainers of hiop.

Pinging explicitly because .github/workflows/triage.yml only runs on pull_request_target, so the "Tag Maintainers" step never fires for an issue — without a mention this report reaches nobody, and the stale bot would close it at 180+30 days.

I deliberately did not open a PR for the coinhsl part. The five-line change in the issue body would alter how +sparse resolves for existing users, which is a packaging policy call rather than a bug fix, and it is not mine to make. The rest of the report is measurements, not proposals.

So the only thing I need is a direction:

• if you want the coinhsl variant change, say so and I will open the PR against develop with the measurements attached;
• if +sparse should keep resolving as it does today, say so and I will close this and keep the workaround on our side;
• if some of it belongs in exago's package rather than hiop's, tell me which half and I will split it.

Happy to be told this is the wrong venue for part of it.

[cameronrutherford]

Would have to rely on @pelesh or the HiOp developers (@cnpetra @nychiang) to confirm, but IIRC this is intended behaviour. Last I was aware we require CoinHSL for sparse operations.

Maybe this has changed, and if so, should be reflected in the package.py.

[pelesh]

Would have to rely on @pelesh or the HiOp developers (@cnpetra @nychiang) to confirm, but IIRC this is intended behaviour. Last I was aware we require CoinHSL for sparse operations.

Maybe this has changed, and if so, should be reflected in the package.py.

The submitter appears to be a bot. I suggest closing this issue because most of the reported "issues" are intended behavior as you observed, as well.

[huojiangming]

Thanks both for looking. @cameronrutherford's answer settles the part I actually needed: requiring CoinHSL for HiOp's sparse interface is intended, so the package is doing what it is meant to do and there is nothing here for spack to change. I will carry the variant change locally.

Closing this as the reporter. Sorry for the noise.