[CI] Lanczos SM tests fail cuda 13.2

[aamijar]

Observed here:

https://github.com/rapidsai/raft/actions/runs/25983993593
https://github.com/rapidsai/raft/actions/runs/25983959550
https://github.com/rapidsai/raft/actions/runs/25955468286
https://github.com/rapidsai/raft/actions/runs/25955311798

https://github.com/rapidsai/raft/actions/runs/25972693380/job/76351882397?pr=3019#step:12:748

Most prevalent on a100, arm64

[----------] 1 test from LanczosTests/LanczosTestD_SM
[ RUN      ] LanczosTests/LanczosTestD_SM.Result/0
/tmp/conda-bld-output/bld/rattler-build_libraft-headers-only/work/cpp/tests/sparse/solver/lanczos.cu:351: Failure
Value of: raft::devArrMatch<ValueType>(eigenvalues.data_handle(), expected_eigenvalues.data_handle(), params.n_components, raft::CompareApprox<ValueType>(1e-5), stream)
  Actual: false (actual=0.013678239999999999 != expected=0.013691776200317895 @1)
Expected: true

[divyegala]

@aamijar what is the absolute diff in all the failures that you are tracking at the moment?

[aamijar]

Hi @divyegala, it is just this one failure in the issue description

Actual: false (actual=0.013678239999999999 != expected=0.013691776200317895 @1)

Abs diff = 0.0000135362

So we are above threshold by a tiny amount 1.3536e-5 > 1.0e-5

[divyegala]

@aamijar and in the other links you posted? What is the error diff?

[aamijar]

@divyegala, I think all the errors are the same. Same actual and expected values.

[cjnolet]

Anupam, The results are very close. I would suggest temporarily increasing the threshold by just enough to get this test to pass. That will at least unblock CI while still rendering the test useful. After that, I would suggest testing based on the characteristics / constraints of knowing the outputs are eigenpairs, instead of hard coding them.

…

On Thu, May 28, 2026 at 1:18 AM Anupam ***@***.***> wrote: *aamijar* left a comment (NVIDIA/raft#3021) <#3021 (comment)> @divyegala <https://github.com/divyegala>, I think all the errors are the same. Same actual and expected values. — Reply to this email directly, view it on GitHub <#3021>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAJPKYELB54IPNV72EFZUK3447DZLAVCNFSM6AAAAACZB2LAC2VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHM2DKNRRGAZTIOJWG4> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[divyegala]

@aamijar when you temporarily increase the threshold as Corey suggested, could you also bring nightly check tolerance back to 14 days?