Release 0.2.1

PyTC 0.2.1 adds validated efficiency paths for constructing and consuming ISDF xTC intermediates.

Added

  • Deterministic blocked pivot selection for molecular ISDF.

  • Exact auxiliary recovery of K1 and K3, including streamed out-of-core construction with cache provenance.

  • An opt-in rank-\(M\) orbital Tucker representation of X and direct T2-U-Z contraction in the factorized ISDF xTC-CCSD solver.

  • An opt-in persistent XLA compilation cache.

  • A standalone rank-\(M\) / factor-direct example: 06_rank_m_x_factor_direct_ccsd.py.

Changed and fixed

  • Accelerator-memory probes fail closed when capacity cannot be measured reliably.

  • Disk-backed X panels remain bounded during the JAX CCSD direct-tile path.

  • Delta-U dispatch sizing no longer double-counts an already resident D kernel.

  • Rank-major X-store conversion retains the source layout and adds a provenance-stamped twin for legacy compatibility.

Validation scope

On H10/R=1.6/cc-pVTZ/grid2, exact auxiliary recovery reduced K1/K3 construction from 324.90 s to 89.62 s on one V100, with K1/K3 relative errors below \(1.2\times10^{-15}\).

On the corresponding one-A100 gate, blocked pivot selection was 9.515 times faster in the relaxed-energy run and changed the full-X total energy by +0.000267 mHa. The opt-in \(M=80\) orbital-X approximation changed the exact-pivot total energy by -0.844607 mHa. Transfer beyond this H10 gate and a matched rank-\(M\) construction speedup have not been established.

See ISDF Efficiency Paths for API usage and validation guidance.