v0.3.0

Third pre-release. CPU-side API + GPU baseline infrastructure.

Added

• inv_sqrt_spd_ns(A) — Newton-Schulz iteration for A^{-1/2}. All-GEMM, returns (X, iters, residual). Progress on #14.
• scipy.linalg CPU baselines in the benchmark suite
• benchmarks/bench_cuda.py + infra/terraform/gpu.tf + scripts/run_cuda_tests.sh — vintage-matched CUDA/cuSOLVER baseline track (trn1 ↔ A10G on g5.xlarge; trn2 ↔ H100 on p5.4xlarge). Progress on #13.

Changed

• docs rewrite with vintage-matching methodology and real CPU numbers
• CI bumped to checkout@v6 / setup-python@v6; standalone docs deploy removed (docs served from trnsci.dev)
• pyproject normalized across the trnsci suite

Milestones

Milestone numbering shifted up one to reflect the new tag:

• old v0.3.0 (NKI hardware validation) → now v0.4.0
• old v0.4.0 (production polish) → now v0.5.0

See CHANGELOG.md.

v0.2.0

Second pre-release. Closes the v0.2.0 milestone.

Added

• jacobi_preconditioner(A) for CG (preview of #16)
• Benchmark suite (benchmarks/bench_solver.py) covering eigh, factorizations, solve, CG, GMRES (#8)
• Status badges in README (#4)
• Manual Neuron hardware CI workflow via workflow_dispatch + OIDC (#5)

Changed

• cg(A, b, M=tensor) now applies M as M@r instead of inv(M)@r
• Repository transferred from scttfrdmn/trnsolver to trnsci/trnsolver

See CHANGELOG.md for full details.

Note on PyPI publishing: #3 remains open — trusted-publisher registration is pending.