Release v0.5.0: experimental CRT-based SNF fast path#21

Merged

events555 merged 23 commits into

mainfrom

Jun 6, 2026

Commits on Jun 6, 2026

feat: add CRT-style non-recursive SNF path

claude
authored and
events555
committed
chore: add CRT SNF prototype and profiling scripts

claude
authored and
events555
committed
docs: replace PLAN.md with CRT optimization roadmap
events555
committed
perf(crt): i64 fast-path reduction in local_snf hot loops (Phase 1)
events555
committed
perf(crt): split local_snf into two-phase column/row elimination (Phase 2a)
events555
committed
docs: record Phase 2a done + detailed Phase 2b blocked-GEMM design
events555
committed
perf(crt): blocked right-looking LU with GEMM trailing update (Phase 2b)
events555
committed
docs: mark Phase 2b done; note tiled/BLAS GEMM kernel as Phase 2c
events555
committed
perf(crt): packed cache-friendly i64 GEMM micro-kernel (Phase 2c)
events555
committed
docs: mark Phase 2c done (packed i64 GEMM kernel)
events555
committed
fix(snf): keep transforms unimodular for tall rank-deficient inputs
events555
committed
refactor(crt): remove dev scaffolding and tidy the implementation
events555
committed
refactor: collapse the Python package into a thin Rust binding
events555
committed
test: replace Python correctness suite with PyO3 boundary tests
events555
committed
docs: position CRT as experimental fast path; reflect thin wrapper
events555
committed
build: bump version to 0.5.0
events555
committed
merge: reconcile main into dev for v0.5.0 release
events555
committed