pytc.df.thc.exact_df_panelled¶

pytc.df.thc.exact_df_panelled(b, t2, aux_panel)[source]¶

Exact current-DF sandwich, panelled over the auxiliary axis.

The (nocc^2, nvir, nvir, q) intermediate inside _exact_accumulate() costs nocc^2 * nvir^2 * q * 8 bytes – ~4.9 GiB per aux column at large systems, so the caller’s aux_panel=32 would demand ~157 GiB. The panel step is therefore clamped so the intermediate stays under PYTC_EXACT_PANEL_CAP_GB (default 8 GiB, read at call time); at q=1 the GEMM shapes are unchanged (the batch is the occupied-pair axis, not q), so the clamp costs no efficiency. Small cases are unaffected.

Parameters:: aux_panel (int)