pytc.df.thc¶

Production DF/THC algebra and X-store layouts.

The JAX section implements the panelled LS-THC algebra used by the solver. extract_vv_df_factor builds its metric-applied virtual DF input. The X-store section converts ISDF X factors between the rank-innermost store layout (nmo, nmo, rank) and the rank-major layout (rank, nmo, nmo) (panel-contiguous), in place or into a new store.

Classes

ScalableDirectSandwichesJax

Exact, cross, full-THC, and robust production contractions.

Functions

`_as_fp64_host`	FP64 as a HOST array -- for operands the panel loops only ever slice.
`_as_fp64_jax`	Validate rank while landing an operand on device as float64.
`_convert_x_dataset`
`_x_source_stamp`	Provenance stamp for an X dataset: sha256 over shape plus EVERY row slab, so any changed entry changes the stamp (no sampling shortcuts).
`add_rank_major`	Append a rank-major copy of X to an existing store, in place.
`convert_store_to_rank_major`	Whole-store conversion: every dataset copied; X kept AND a rank-major twin appended.
`convert_x_to_rank_major`	Convert one X dataset to rank-major layout, chunked over an nmo axis.
`df_sandwiches_jax`	Evaluate the panelled production DF/THC contractions.
`exact_df_panelled`	Exact current-DF sandwich, panelled over the auxiliary axis.
`extract_vv_df_factor`	Extract PySCF's metric-applied virtual DF factor in PyTC order.
`fit_lsthc_jax`	JAX FP64 normal-equation LS-THC fit with host-resident DF factors.
`full_thc_panelled`	B_tilde[a,c,Q] B_tilde[b,d,Q] t2[ij,c,d] without ever forming B_tilde.
`partial_thc_crosses_panelled`	Both partial-THC crosses with a precontracted DF endpoint.
`require_float64`	Fail closed if JAX is not in x64 mode.