pytyche.bcf.kernels

Low-level GPU primitives shared across the GPU BCF modules.

This module sits at the bottom of the GPU BCF import graph: it imports only from JAX and numpy. Sibling modules (lml, gfr, the orchestrator, etc.) import FROM this module, never the other way around. Keeping these primitives here breaks the otherwise-circular relationship between the log-marginal/GFR modules and the orchestrator’s MCMC kernel, both of which call the same scatter-add and dual-forest evaluation routines.

Contents

_DEFAULT_NUM_BATCHES — bartz-matched default for contention-reduction batches. _fused_scatter_add — single scatter-add over K stacked value arrays sharing one index. _hurdle_evaluate_dual_forest — forest evaluation for the conversion+severity dual-leaf forests.