Module rope 

RoPE-aware bit allocation for KV-cache quantization.

Provides CPU-side, deterministic infrastructure for assigning variable bit-widths to RoPE frequency blocks. The greedy energy heuristic is experimental and inspired by block-level frequency analysis; it does not claim equivalence with any published algorithm.

Structs

RopeBitAllocation

Bit-width assignment for every RoPE block in a head.

RopeBlock

A contiguous pair of head dimensions forming one RoPE frequency block.

RopeBlockEnergy

Energy estimate for one RoPE block.

Functions

allocate_rope_bits

Greedily allocate bits across RoPE blocks.

rope_block_energies

Compute per-block energy from a set of key vectors.

rope_blocks

Partition head_dim into contiguous 2-D RoPE blocks.