Repository navigation
feat(hsolver): add PPCG support for PW diagonalization - #8264
cheerly-pku wants to merge 228 commits into
Conversation
Consider the previous contributions made by classmates, I'm only capable to make small difference without disrupting the entire program ---- like such a small "static".
…w_Small-Changes 2025PKUCourseHW5: Case: 1 - Change rank_seed_offset to static const
…ent) Add PPCG iterative diagonalization with two strategies: - CONJUGATE_GRADIENT: band-by-band Polak-Ribiere CG (verified working) - BLOCK_SUBSPACE: block subspace diagonalization Includes potrf retry fix: save/restore original matrix before applying diagonal shift, preventing accumulated shifts from corrupting the matrix. Test: 1D particle-in-a-box (n_dim=10), CG strategy matches exact eigenvalues with error 4.3e-12. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…ormalization Three fixes for numerical stability: 1. potrf: save/restore original matrix before diagonal shift retries, preventing accumulated shifts from corrupting the Cholesky factor. 2. sygvd/syevd: skip workspace query (lwork=-1) and allocate directly. The LAPACK replacement ignores workspace queries, causing the second call to operate on already-transformed data, corrupting eigenvalues. 3. Block subspace: add chol_qr + hpsi/spi recomputation after update_one_block and every rayleigh_ritz, keeping wavefunctions S-orthonormal and preventing numerical drift of H|psi> and S|psi>. Results (1D particle-in-a-box, S=I): - CG nband=1: error 4.3e-12 (unchanged, already working) - BLOCK_SUBSPACE nband=1: no longer NaN, converges (to wrong eigenvalue due to algorithmic limitation with S=I)
… RR steps 1. solve_small_generalized: save/restore M matrix before retry with shifts (prevents accumulation of shifts on sygvd-corrupted M) 2. BLOCK_SUBSPACE: add Krylov fallback for near-collinear p/w vectors When p is nearly parallel to w (cos^2 > 0.99), replace p with H·w to keep the 3-vector subspace [psi, w, p] full rank. This fixes NaN eigenvalues for nband>1 with S=I. 3. LAPACK: use standard workspace query (lwork=-1) pattern for syevd/sygvd More robust with real LAPACK implementations. 4. CG: add periodic Rayleigh-Ritz subspace rotation every rr_step iterations Corrects band ordering and eigenvalue estimates after band-by-band line minimization. Resets PR state after rotation.
The gamma_dot function returns only the real part of inner products, which is correct for Hermitian forms like <psi|H|psi> but wrong for projection coefficients where the imaginary part matters. In orth_gradient and project_against, the projection coefficient <psi_i | v> must use the full complex inner product to correctly remove the overlap. Using only Re(<psi_i | v>) leaves an imaginary component that corrupts the search direction, causing excited-state bands to converge to wrong eigenvalues. This fixes the CG strategy bands 1 and 2 converging to the highest eigenvalue (3.919) instead of the first excited states.
…PACE The chol_qr_active call after update_one_block re-orthonormalized psi but left the p vector in the old basis, creating an inconsistency. The p vector is constructed in update_one_block using the same subspace rotation as psi, so they start consistent. Adding chol_qr_active before the p vector is updated breaks this consistency.
This change was unrelated to the PPCG integration and should not have been included.
H and S are real symmetric operators whose eigenvectors are real. The previous complex random initialization produced complex off-diagonal elements in the H-gram matrix (max |Im| ~ 0.5 for nband=3), causing Re(<psi_i|H|psi_j>) != <psi_i|H|psi_j>. The gamma_dot function only returns the real part, so all subspace Gram matrices (built via gram()) computed wrong off-diagonals, leading to incorrect eigenvalues from sygvd. With real-only psi all inner products are real and gamma_dot is exact, so both BLOCK_SUBSPACE and CONJUGATE_GRADIENT strategies should now converge to the correct eigenvalues.
…tioning The 3-block subspace method builds a generalized eigenvalue problem with basis V = [psi, w, p] where p is constructed from the previous subspace eigenvectors (p_new += w_l * cw in update_one_block). This makes p a linear combination of the w vectors, causing the [w, p] block of the S-gram matrix M to become nearly rank-deficient. With nband=3 and sbsize=4 the 9x9 M matrix has condition number large enough that dsygvd produces negative eigenvalues for the positive-definite problem (observed: -0.26 at iter=2), and the eigenvalues diverge exponentially thereafter. Setting use_p=false reduces the subspace to [psi, w] (2-block), which is a preconditioned Davidson-like method. It converges robustly: the BLOCK_SUBSPACE test now passes in 57 ms with all 3 eigenvalues within 1e-8 of the exact values. The 3-block code path is preserved for future re-enablement once a more robust p-vector construction is implemented.
With rr_step=4, the non-RR iterations use Cholesky orthonormalization
which mixes bands through the upper-triangular U^{-1}, causing high-energy
bands to contaminate low-energy ones. This drives CG eigenvalues to the
spectrum maximum [3.31, 3.68, 3.92] instead of the correct lowest values
[0.081, 0.317, 0.690].
Using rr_step=1 forces Rayleigh-Ritz every iteration, which correctly
diagonalizes the subspace and preserves band ordering.
The orth_cholesky call before rayleigh_ritz mixes bands through the
upper-triangular U^{-1} factor, contaminating low-energy bands with
high-energy components. This drives CG eigenvalues to the spectrum
maximum instead of the minimum.
rayleigh_ritz solves the generalized eigenvalue problem K v = λ M v
via dsygvd, which correctly handles non-S-orthogonal bases. The
orth_cholesky is not needed and is actively harmful.
This makes the CG RR path consistent with BLOCK_SUBSPACE, which
calls rayleigh_ritz without prior orth_cholesky.
BLOCK_SUBSPACE starts with rayleigh_ritz (line 1085) which finds correct eigenvalues and rotates psi before the iteration loop. CG was using diagonal Rayleigh quotients instead — these are poor approximations for random initial guesses, producing wrong gradients that drive the band-by-band line_minimize toward high-energy eigenstates. With rr_step=1 (every-iteration RR), the CG loop itself is now correct, but without an initial RR the first line_minimize step already pushes psi in the wrong direction, and subsequent RR steps cannot fully recover.
Backup preserved at diago_ppcg_test.cpp.bak
The linear approximation α = -C/B drops the α² term from the Rayleigh quotient derivative dR/dα = 0. This picks one of the two stationary points (minimum or maximum) arbitrarily. For bands far from convergence it can select the MAXIMUM, driving ψ toward high-energy states instead of the desired lowest eigenvalues. Solve the full quadratic Aα² + Bα + C = 0, evaluate R(α) for both roots (and the linear guess), and pick the one with the lowest R. Also restore the CG unit test (rr_step=1, initial rayleigh_ritz).
…e use_p
Three changes to make both PPCG strategies correctly converge with rr_step=4:
1. CG non-RR path: After orth_cholesky, solve the nband x nband subspace
generalized eigenvalue problem instead of using diagonal Rayleigh quotients.
The upper-triangular U^{-1} from Cholesky mixes high-energy components into
low-energy bands, making diagonal RQs overestimate the eigenvalues. The
subspace solve gives correct Ritz values without rotating the states,
preserving Polak-Ribiere conjugate-direction accumulators.
2. BLOCK_SUBSPACE: Re-enable use_p=true (3-block [psi, w, p] subspace).
The Krylov fallback (replace p with H·w when p ~ w) was already in place
but dead because use_p was hardcoded to false. Now it activates on the
first iteration (p is zero-initialized) and whenever p becomes collinear
with w after update_one_block.
3. CG test: Change rr_step from 1 back to 4 so the non-RR Cholesky path
is exercised, validating the true Polak-Ribiere CG mechanism.
The 3-block [psi, w, p] subspace generalized eigenproblem becomes ill-conditioned when residuals are small (near convergence). The [w, p] Gram block shrinks, the M matrix approaches singularity, and dsygvd produces garbage eigenvectors that drive eigenvalues to catastrophic values (e.g., -137775 instead of 0.081). The p-bad H·w Krylov fallback fixes p~w collinearity but does not address the small-residual ill-conditioning, which is fundamental to the 3-block construction. Keep use_p=false for robust convergence.
The [w,p] block of the Gram matrix M shrinks as residuals converge, making M nearly singular and causing sygvd to produce garbage eigenvectors. Scaling w and p to unit S-norm keeps M well-conditioned (diagonal ~1) without changing the subspace — Ritz values are identical and Ritz vector coefficients cancel in update_one_block. This enables the full 3-block [psi,w,p] subspace (use_p=true) by addressing the fundamental ill-conditioning that the p-bad Krylov fallback alone could not handle.
The Krylov fallback (replace p with Hw when p~w) was flawed: when w is approximately an eigenvector (Hw ≈ λw), the replacement does not fix collinearity. After S-norm scaling, p ≈ w still, M_wp ≈ [1,1;1,1] is rank-1, and dsygvd fails. Instead, simply skip p for this iteration (use_p_now=false). update_one_block still produces a valid p for the next iteration from the w Ritz-vector contribution.
Add tests for: - 2x2 matrix (smallest non-trivial case) - Degenerate eigenvalues (H = I + J, multiplicity-3 degeneracy) - Larger 20x20 tridiagonal with 5 bands - Dense 8x8 matrix via Givens rotations (addresses full-matrix coverage) All use CONJUGATE_GRADIENT strategy which has sygvd fallback. BLOCK_SUBSPACE tests deferred due to dsygvd instability with some LAPACK builds. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Covers diagonal, tridiagonal, dense, pentadiagonal, degenerate, Neumann, S≠I, gamma_g0, single-band, all-band, many-band, bad preconditioner, tight threshold, scaled, gapped spectrum, rr_step=1, 1x1, and eigenvector quality checks. Adds QuickBenchmark (CI-friendly) and DISABLED_FullBenchmark.
Keep the PPCG implementation, explicit Rayleigh-Ritz cadence, focused unit tests, and dedicated integration fixtures on the current develop base. Remove unrelated solver, XC, charge-mixing, output, comparator, and reference calibrations from the old PR scope.
# Conflicts: # source/source_hsolver/CMakeLists.txt # source/source_hsolver/unittests/CMakeLists.txt # source/source_pw/module_stodft/test/CMakeLists.txt
Use the existing parallel reduction wrapper for trace writer selection and keep the production PPCG path on the conservative full Ritz cadence. Remove the unvalidated user-facing cadence parameter from the focused PR.
ABACUS-PR-BOT
left a comment
There was a problem hiding this comment.
PPCG's normal-return failure status needs to reach the PW caller, and concurrent test invocations need separate trace files. Static review at b37124c; builds, runtime behavior and reference values were not independently verified.
Propagate PPCG convergence diagnostics to the PW caller, stop uncertified NSCF results, and test both SCF and NSCF dispatch. Isolate concurrent MPI trace files and report resolved paths through an explicit running-log stream. Correct the poor-preconditioner regression and scoped help/fixture descriptions.
|
Fresh verification report for the current head Key results:
All remote checks for this head now pass, including CUDA build/test, GNU/Intel builds, Test, governance, pre-commit, PyTest, abacuslite, and documentation. The seven earlier inline bot findings each have a separate reply and are resolved; no new unanswered inline finding was present during this update. |
ABACUS-PR-BOT
left a comment
There was a problem hiding this comment.
Two issues need attention: PPCG executes diagnostic work with tracing disabled, and fresh PW NSCF runs ignore the configured iteration budget. One non-blocking benchmark wording correction is also included. Static source review only; no builds, numerical runs or reported CI results were independently verified.
There was a problem hiding this comment.
it's not necessary to have such a large test file
Reuse the analytical fixture for diagonal, padded, eigenvector and preconditioner tests. Preserve all 54 cases and their seeds, matrix inputs and acceptance thresholds; independently recompute residuals and orthogonality in one helper. Validation: MODULE_HSOLVER_ppcg built; all 54 tests passed in three shuffled repetitions (seed 1109). No documentation change is required because this only refactors tests.
Model analytical test preparation after the CG parameter table: consolidate nine Laplacian solves and the diagonal/tridiagonal overlap fixtures without dropping cases. Retain structured overlap callbacks and independently recompute dense residuals and S orthogonality. Validation: MODULE_HSOLVER_ppcg rebuilt; 54 tests passed in each of five shuffled repetitions (seed 1109). This test-only refactor does not require documentation changes.
Keep both prescribed startup subspaces as named parameter instances and reuse the matrix solve helper. Print analytical case names in GoogleTest diagnostics rather than raw parameter bytes. Validation: focused CTest 6/6, PPCG 54/54 in each of five shuffled repetitions, C++11 syntax check, and a complete 54-to-54 test-inventory mapping. No user-facing documentation change is needed for this test-only refactor.
|
|
||
| if (inp.calculation != "nscf") | ||
| // PPCG also uses the configured iteration budget for NSCF calculations. | ||
| if (inp.calculation != "nscf" || inp.ks_solver == "ppcg") |
There was a problem hiding this comment.
ppcg should not be used together with nscf
There was a problem hiding this comment.
you should not modify this
ABACUS-PR-BOT
left a comment
There was a problem hiding this comment.
The requested PPCG/NSCF exclusion is not enforced by current INPUT handling; the support contract and budget regression need reconciliation in the existing thread below. Static incremental source review at a12fb908e810f263c706933a9ac535d80c8dd2f0 against fcd57f6db2660a2d82bd66dfbc93d78e802600c5; no build, runtime test or CI result was independently verified.
Adds PPCG to PW Hamiltonian diagonalization, with residual-based convergence checks, CPU/MPI tests, and native CPU/GPU fixtures. This supersedes closed #8042; the unrelated Libxc, charge-mixing, CG/Davidson, output-comparator, LR, and ordinary PW fixture changes remain outside this PR. SCAN fixture 820 remains excluded pending independent calibration.
Current head:
a12fb908e810f263c706933a9ac535d80c8dd2f0. Calculation mode (scf/nscf) and eigensolver selection (cg/dav/dav_subspace/ppcg) are handled separately. The PW caller passes the configured iteration budget explicitly to PPCG; common parameter initialization has no PPCG exception. Fresh application and dispatch checks for this head are below. Earlier unit/MPI and efficiency measurements retain their stated commit attribution.Linked Issue
Successor to closed #8042; no separate issue.
What's changed?
Keep PPCG nonconvergence visible at PW dispatch: SCF logs the returned status; uncertified NSCF results and non-finite residuals terminate with a diagnostic.
Require
is_open() && good()before enabling trace-only work. A regression compares diagnostic-helper call counts with tracing unset, file opening unsuccessful, and tracing enabled. It reproduced the defect before the fix.Pass
pw_diag_nmaxexplicitly when constructing the PW solver for PPCG.setup_diago_params_pwretains develop's calculation-mode logic, with no solver-specific exception. Other eigensolvers retain their existing budget path. Fresh application runs verify INPUT handling; unit tests check explicit constructor arguments and shared setup independently.Preserve unique MPI trace paths, single-writer behavior, full-path logging, the nonuniform preconditioner regression, and the corrected help/mesh descriptions from the earlier review.
Remove all 10 hand-written file-scope data objects from
diago_ppcg.cpp: delete one unused constant and put the remaining nine at their function or block of use. OpenMP work thresholds remain protected by_OPENMP. Norm floors, overlap cutoffs, diagonal shifts and refresh/restart intervals retain their values and scalar conversions.Keep analytical π values and parameter tables local to their test functions or parameter generators. GoogleTest copies the local parameter arrays. PW dispatch fixtures now save and restore the existing shared
avg_iterstatistic along with their configuration.Consolidate analytical matrix fixtures and initialization helpers using the CG test's preparation-table pattern, and keep exploratory timing outside the unit tests. Split the former 2,436-line test into four source files under the same
MODULE_HSOLVER_ppcgtarget:source/source_hsolver/unittests/test_diago_ppcg.cpptest_diago_ppcg_roots.cpptest_diago_ppcg_iteration.cpptest_diago_ppcg_subspace.cppShared matrix storage and independent eigenpair checks occupy the 132-line test-only
diago_ppcg_test_support.h; the two shared free-function definitions are intest_diago_ppcg.cpp. All 54 runtime cases remain in CTest. The source extraction preserved all test bodies, parameters and acceptance thresholds; the subsequent scope cleanup only relocates constants and parameter tables. It reduces the largest source file rather than removing regression coverage. One MPI-aware main initializes and finalizes the complete executable. The test-file split changes test organization; the production scope cleanup is described separately above. Efficiency measurements below run the complete application separately.avg_itermeans average subspace iterations per band.Unit Tests and/or Case Tests for my changes
Calculation-mode / solver separation verification (2026-11-13,
a12fb908e).GNU 12.3, Release CPU, OpenMPI 4.1.2, existing
build-review, ABACUSv3.11.0-beta13; runtime threads fixed toOMP_NUM_THREADS=1,OPENBLAS_NUM_THREADS=1,MKL_NUM_THREADS=1. Commands run:cmake --build build-review --target MODULE_HSOLVER_pw -j4 cmake --build build-review --target abacus_pw_para -j4 OMP_NUM_THREADS=1 OPENBLAS_NUM_THREADS=1 MKL_NUM_THREADS=1 ctest --test-dir build-review -V -R '^MODULE_HSOLVER_pw$' python3 /mnt/data/Cheerly/ppcg-nscf-layer-20261113/run_nscf.py git diff --check python3 tools/03_code_analysis/agent_governance_check.py --stagedmpirun --bind-to core --map-by core -np 2 /mnt/data/Cheerly/abacus-develop-v3/build-review/abacus_pw_para. Budget 1 reported nonconvergence at iteration 1 and exited nonzero as expected. Budget 77 converged at iteration 36, exceeding the old static budget of 30.tests/01_PW, the command wasbash ../integrate/Autotest.sh -a /mnt/data/Cheerly/abacus-develop-v3/build-review/abacus_pw_para -n 2 -o 1 -j 1 -f CASES_CPU.txt, selecting only those three fixtures, withABACUS_MPIEXEC_PREFLAGS='--bind-to core --map-by core'.--version,-h pw_diag_nmax, and--check-inputfor the three fresh SCF/NSCF verifier inputs succeeded. Diff checks passed; staged governance reported only the documentation-sync warning. No documentation change is needed: INPUT definitions and documented PPCG budget behavior are preserved.Records:
/mnt/data/Cheerly/ppcg-nscf-layer-20261113/.validated-commit.jsonconfirms the tested patch exactly matches this commit. Executable SHA-256:81679b16c7f6bdd2f2f22bb74a415d883cc2e54e4cac9a1f097221ecf50e71c8. Efficiency was not remeasured for this parameter-routing change.Inline cleanup (2026-11-13,
7b95cc341)Removed redundant explicit
inlinespecifiers from the PPCG implementation and test support. The two shared test helpers now have declarations in the support header and definitions in one existing test source file, so they no longer need header-level inline definitions. The class helperidxremains defined in the class body, where C++ already treats it as inline; there is no explicit specifier. This is a clarity and linkage cleanup, not a performance optimization: these edits do not target a runtime hot path and are not expected to improve PPCG speed.GNU 12.3 Release CPU builds completed for MPI and non-MPI configurations. Commands run:
MPI CTest passed 5/5 (54 double-precision and four float PPCG cases, eight PW tests, plus MPI configurations with 2/3/6 ranks and two pools). Non-MPI CTest passed 2/2 (54 double-precision and four float PPCG cases).
git diff --checkpassed. The staged governance check reported only its documentation-sync warning; this cleanup changes no user-facing behavior, solver settings or documentation contract.MPI infrastructure cleanup verification (2026-11-13,
0ff3b7411).The trace writer now queries the existing communication domain only when
ABACUS_PPCG_RESIDUAL_TRACEis set. No MPI primitive, global variable, or public solver interface is added. The numerical reductions continue to use the existingParallel_Reduce::reduce_pool. The MPI test checks that the trace-write message appears on world rank zero only, alongside the existing CSV completeness and numerical checks.GNU 12.3, Release, CPU, OpenMP enabled, LCAO disabled. Reused
build-review(ENABLE_MPI=ON, OpenMPI 4.1.2) andbuild-review-nompi(ENABLE_MPI=OFF). Commands run:Both builds succeeded. MPI CTest: 5/5 passed, including 2/3/6 ranks and four ranks with two pools. Non-MPI CTest: 2/2 passed. Both configurations include 54 double-precision and four float cases; the MPI configuration also includes eight PW setup/dispatch cases. The disabled-trace and failed-open regression passed.
git diff --checkpassed; staged governance reported only the documentation-sync warning addressed by the no-docs-needed explanation below. Existing hwloc/X authority startup messages remain in the MPI logs.No documentation update required for this cleanup: it replaces an implementation detail with an existing API while retaining trace behavior and numerical settings. The application and efficiency measurements below were not rerun for this cleanup and retain their original commits.
Earlier scope-cleanup verification (2026-11-09,
a98134e07).The production cleanup and test cleanup were each built, run and committed separately. GNU 12.3, Release, Intel Xeon Gold 6242, OpenMPI 4.1.2, OpenMP enabled, LCAO/CUDA disabled, ELPA and OpenBLAS; the existing
build-reviewcache was reused. Runtime variables:OMP_NUM_THREADS=1,OPENBLAS_NUM_THREADS=1,MKL_NUM_THREADS=1, with residual tracing unset except in trace-specific tests.tests/01_PW, the command wasbash ../integrate/Autotest.sh -a /mnt/data/Cheerly/abacus-develop-v3/build-review/abacus_pw_para -n 2 -o 1 -j 1 -f CASES_CPU.txt, with the selection containing only these three fixtures. All three SCF runs converged without PPCG nonconvergence messages.740483d43. The earlier test-file extraction also passed its independent source-body audit at that commit.GlobalV,GlobalCorPARAMaccesses. Existing multi-solverDiagoIterAssiststatic interfaces are retained; PW fixtures save and restore the shared values they use.git diff --checkpasses; the governance check has no blockers. Its 12 header dependency warnings are reviewed: the production header owns scalar aliases, callbacks and workspace vectors; the test support header owns matrix vectors and implements seeded initialization and independent analytical checks. Their includes provide complete types and the functions used in these inline definitions. The support header is test-only.diago_ppcg.cppremains below that threshold because of its file length and existing long/complex routines. The production file has the same 12 findings and score as before the scope cleanup. Those structural findings remain open and are recorded separately from the passing correctness and governance checks..cppfiles are explicitly listed in the existing CMake target.source/Makefile.Objectslists only the production solver object; pyabacus has no source list for these unit tests./mnt/data/Cheerly/ppcg-local-scope-20261109/contains build/runtime logs, native-case results, source snapshots, test inventories and C++11 checks. The earlier split audit remains in/mnt/data/Cheerly/ppcg-test-split-20261109/. The 2026-11-05 NSCF and efficiency batches below retain their original commit attribution.Application correctness checks (2026-11-05,
9de3fb707)GNU 12.3, Release, Intel Xeon Gold 6242, OpenMPI 4.1.2, OpenMP enabled, LCAO/CUDA disabled, ELPA and OpenBLAS. The existing
build-reviewcache was reused. Runtime variables wereOMP_NUM_THREADS=1,OPENBLAS_NUM_THREADS=1,MKL_NUM_THREADS=1; residual tracing was unset except in trace-specific checks.tests/01_PW, wasbash ../integrate/Autotest.sh -a <absolute-path-to-abacus_pw_para> -n 2 -o 1 -j 1 -f CASES_CPU.txt; that selection contained only these three fixtures.pw_diag_nmax=1exited as expected at iteration 1;pw_diag_nmax=77completed successfully, with the last k-point trace reaching iteration 36. This exercises a budget beyond the previous static value of 30.abacus_pw_para --versionreportedv3.11.0-beta12;-h pw_diag_nmaxand--check-inputin a valid case directory succeeded.git diff --checkand the diff-oriented governance check passed. Existing OpenMPI/hwloc startup warnings remain in the logs.The native final energies (817/818/819) were
-4862.3309704994053391,-227.6796325580887697,-1882.1779472344287569eV.Application efficiency measurements (2026-11-05,
9de3fb707)Each case/solver used a clean copy of the same committed fixture; only
ks_solverdiffered within each group. Runs were sequential, trace disabled, usingmpirun --bind-to core --map-by core -np 2 abacus_pw_para. There was one warm-up and five measured runs per solver, with rotated solver order. Wall time surrounds the MPI subprocess;solve_psikis the ABACUS solver timer.solve_psikmedian [min, max] sOf the 60 measured runs, 55 converged. The five excluded samples are all 819/CG, which reached the fixture's 50-step SCF limit. PPCG is slower than the Davidson variants on these small regression cases. Large-system performance remains to be measured.
A separate 12-run correctness comparison used the same overrides for every solver:
scf_thr_type=2,scf_thr=1e-10,scf_nmax=150,pw_diag_thr=1e-10,pw_diag_nmax=200. Davidson, Davidson-subspace and PPCG converged in all three cases. CG reached the 150-step SCF limit in all three and is excluded from the numerical comparison. Across the six converged Davidson/PPCG comparisons, the largest differences were 2.729E-12 eV in total energy, 2.060E-8 eV/Å in a force component, and 5.424E-7 kbar in a stress component. These runs are used only for numerical agreement.The local artifact directory is
/mnt/data/Cheerly/ppcg-bot-followup-20261105/:independent-validation.md/.jsonaudit the raw logs, 84 INPUTs, 336 other tracked files and pseudopotentials; protocols record the command, threads, commit and executable hash. The fresh native and NSCF records are innative/andnscf-cli/. Executable SHA-256:27666bef7165a9086d6f13d063a39d25ab80d40959bfaa731574f8430e6fc8ac. The earlier batch with a stale configure-time version banner is archived underbefore-version-refresh/and supplies none of the table's numbers.Remote checks and documentation
The
a12fb908epush starts a new CI run; its status is shown in this PR's checks. The current local checks cover MPI and non-MPI CPU builds and the listed unit/runtime tests. CUDA and the other toolchains rely on their corresponding CI jobs.No documentation update required for these follow-up fixes: they restore the documented iteration-budget and opt-in trace behavior without changing the INPUT schema, defaults or acceptance tolerances. The PR retains its PPCG help metadata and solver documentation. No new global dependency or default interface argument is introduced by these follow-ups.