svd: run test_josef in the nightly LAPACK validation, with a stated ignore reason

.github/workflows/lapack-validation.yml

@@ -15,7 +15,7 @@ concurrency:
 jobs:
   validation:
     runs-on: ubuntu-latest
-    timeout-minutes: 50
+    timeout-minutes: 55
     outputs:
       unexpected: ${{ steps.summary.outputs.unexpected }}
       stale: ${{ steps.summary.outputs.stale }}
@@ -36,7 +36,7 @@ jobs:
           python3 scripts/test-lapack-validation.py
           node --test scripts/test-lapack-reporter.cjs
       - name: Build once and run all numerical families
-        timeout-minutes: 44 # build <=20m; family limits sum to 20m; setup/headroom
+        timeout-minutes: 49 # build <=20m; family limits sum to 25m; setup/headroom
         run: scripts/run-lapack-validation.sh
       - name: Summarize known and new failures
         id: summary

docs/lapack-validation.md

@@ -291,8 +291,10 @@ fired only for a failure that is not listed - or for a listed test that starts p
 lose its entry in the same change. Adding an entry without an issue is a bug in the entry.
 It caches dependencies on failure and uploads environment, build timing, exact
 selection, family logs and summary for 14 days. Per-family timeouts sum to
-20 minutes; the build limit is 20 minutes, the combined step 44 minutes and the
-validation job 50 minutes. A separate two-minute reporting job creates/updates one
+25 minutes; the build limit is 20 minutes, the combined step 49 minutes and the
+validation job 55 minutes. The family `svd_large` runs upstream's large SVD matrix
+test `test_josef` (n = 8660, f32, full U and V; about 70 s and 2.3 GB locally), which is too slow to
+run with every test and is therefore ignored there. A separate two-minute reporting job creates/updates one
 open issue titled “Scheduled LAPACK numerical validation is failing” with label
 `scheduled-lapack-validation`, retaining the failed status and run link. These are
 upper bounds, not expected runtimes. They allow every family timeout to expire

faer/src/linalg/svd/bidiag_svd.rs

@@ -2147,7 +2147,9 @@ mod tests {
 		}
 	}
 	#[test]
-	#[ignore]
+	// n = 8660 in f32 with full U and V: about a minute even optimised, so it runs in the nightly
+	// LAPACK validation (family `svd_large`, scripts/run-lapack-validation.py), not with every test.
+	#[ignore = "large (n = 8660): runs in the nightly LAPACK validation"]
 	fn test_josef() {
 		for file in std::fs::read_dir(
 			PathBuf::from(env!("CARGO_MANIFEST_DIR")).join("test_data/svd/"),

scripts/run-lapack-validation.py

@@ -12,14 +12,18 @@
 FAMILIES = (
     'common', 'svd', 'hessenberg', 'schur', 'nonsymmetric_evd',
     'schur_reorder', 'balance', 'tridiagonal_evd', 'self_adjoint_evd',
-    'lu', 'cholesky', 'indefinite', 'qr', 'triangular', 'lstsq',
+    'lu', 'cholesky', 'indefinite', 'qr', 'triangular', 'lstsq', 'svd_large',
 )
 # Local measurements are documented in docs/lapack-validation.md. Hosted
 # timings remain unmeasured; these bounds leave several times the local runtime.
 LIMITS = dict(common=30, svd=120, hessenberg=60, schur=90,
               nonsymmetric_evd=60, schur_reorder=30, balance=30,
               tridiagonal_evd=90, self_adjoint_evd=120, lu=180,
-              cholesky=60, indefinite=120, qr=90, triangular=60, lstsq=60)
+              cholesky=60, indefinite=120, qr=90, triangular=60, lstsq=60,
+              svd_large=300)
+# Upstream's large SVD matrix tests: too slow for every test run (about a minute each
+# locally), so they are ignored there and run here, in a family of their own.
+LARGE_SVD = ('linalg::svd::bidiag_svd::tests::test_josef',)
 CASE = re.compile(
     r'LAPACK_CASES family=(\S+) dtype=(\S+) planned=(\d+) attempted=(\d+) '
     r'call_failures=(\d+) diagnostics=(\d+)(?: auxiliary_attempted=(\d+))?'
@@ -28,6 +32,8 @@
 
 
 def classify(name):
+    if name in LARGE_SVD:
+        return 'svd_large'
     if name.startswith('linalg::svd::tests::'):
         return 'svd'
     leaf = name.rsplit('::', 1)[-1]

scripts/test-lapack-validation.py

@@ -18,6 +18,8 @@ def test_classification_and_empty_selection(self):
         self.assertEqual(v.classify('linalg::evd::lapack::f64::lapack_schur_reorder_cases'), 'schur_reorder')
         self.assertEqual(v.classify('linalg::qr::lapack::c64::lapack_qr_reflector_checkers'), 'qr')
         self.assertEqual(v.classify('linalg::svd::tests::lapack_svd_f32::lapack::diagonal_families'), 'svd')
+        self.assertEqual(v.classify('linalg::svd::bidiag_svd::tests::test_josef'), 'svd_large')
+        self.assertIsNone(v.classify('linalg::svd::bidiag_svd::tests::test_deflation43'))
         # a self-adjoint accuracy test beside the case suite (tensor4all/faer#100): classified, not a case record
         name = 'linalg::evd::lapack::c32::lapack_self_adjoint_evd_qr_spectrum_accuracy'
         self.assertEqual(v.classify(name), 'self_adjoint_evd')
@@ -43,7 +45,7 @@ def fake(command, log, timeout):
         self.assertEqual(len(calls), len(v.FAMILIES))
         self.assertEqual(next(r for r in rows if r['family'] == 'svd')['exit_code'], 101)
         self.assertFalse(next(r for r in rows if r['family'] == 'lu')['complete'])
-        self.assertEqual(rows[-1]['family'], 'lstsq')
+        self.assertEqual(rows[-1]['family'], 'svd_large')
         self.assertTrue(rows[-1]['complete'])
         self.assertTrue(all('--exact' in c and '--include-ignored' in c for c in calls))