speed up perform_gsn: device-aware fast backend + numerical fixes#27

Open

jacob-prince wants to merge 24 commits into

mainfrom

speedup-perform-gsn

Commits on May 27, 2026

add equivalence test 10 exercising 1e-6*I ridge bug in 2D data-cov path
jacob-prince
committed

Commits on May 28, 2026

remove 1e-6*I ridge regularization in calc_shrunken_covariance
jacob-prince
committed
replace pinv with solve_triangular in calc_mv_gaussian_pdf
jacob-prince
committed
use eigh instead of svd in construct_nearest_psd_covariance
jacob-prince
committed
lazy-import matplotlib in rsa_noise_ceiling
jacob-prince
committed
batched-Cholesky shrinkage evaluation via optional torch dependency
jacob-prince
committed
add device option to batched shrinkage NLL path
jacob-prince
committed
add Python-side stress tests for the GSN speedup changes
jacob-prince
committed
matlab: use eig instead of svd in constructnearestpsdcovariance
jacob-prince
committed
add tests/test_speedup_magnitude.py — runtime benchmark across backends
jacob-prince
committed
add cluster/ Kempner H100 pipeline + GPU edge-case tests
jacob-prince
committed
untrack cluster/ + relax GPU test tolerances
jacob-prince
committed
relax GPU test tolerances to relative + add argmin-agreement check
jacob-prince
committed
device-native end-to-end fast_perform_gsn
jacob-prince
committed
update speedup_magnitude figure to single-panel target style
jacob-prince
committed
chunk batched_nll shrinkage stack to bound device memory
jacob-prince
committed
batched_nll: relax chunking trigger to 95% of free device memory
jacob-prince
committed

Commits on May 30, 2026

Commits on May 31, 2026

scrub dataset-specific references from docstrings
jacob-prince
committed