pytc.solver.xtc_ccsd._compute_large_blocks

pytc.solver.xtc_ccsd._compute_large_blocks(eris, xtc_obj, jastrow_params, Lov_reshaped, L_vv_full, nocc, nvir, nmo, panel_blk)

Compute and write balanced tiled ovvv / vovv blocks to HDF5.

The r virtual index is sliced in chunks of panel_blk (GPU-memory-aware, from resolve_v3o_panel_block_size). Both tensors are tiled in a single interleaved _run_tiled_block_pipeline call so GPUs stay occupied across block boundaries.

Both tensors are written as [:, :, r0:r1, :] slabs on axis 2.

ovvv tile: (occ_all, vir_all, vir_r_blk, vir_all) panel_layout=”pr” vovv tile: (vir_all, occ_all, vir_r_blk, vir_all) panel_layout=”qr”

panel_size=max(nocc, panel_blk) pads the variable-sized dimensions (p=occ for ovvv, q=occ for vovv, and r=vir_r_blk) to a fixed size for JIT shape stability. The large vir_all dimensions remain fixed across all tiles.