Neural network weights stored as IEEE 754 floats have a structure that standard entropy coders (Huffman, ANS) cannot exploit directly: the sign, exponent, and mantissa bits are interleaved across byte boundaries, making every byte look near-random (~8 bits of entropy). This module applies two transformations before Huffman coding: ## Stage 1: Bit reorder (`bits_mode = 1`) Rearranges bits within each float so the exponent occupies a dedicated byte position. For fp32: ```text IEEE 754: [S EEEEEEEE MMMMMMMMMMMMMMMMMMMMMMM] (sign, exponent, mantissa) Reordered: [EEEEEEEE S MMMMMMMMMMMMMMMMMMMMMMM] (exponent, sign, mantissa) ``` In little-endian memory, this moves the exponent into byte 3 (the MSB). Without this step, the sign bit straddles the exponent/mantissa byte boundary, contaminating both bytes and reducing compressibility. The same principle applies to fp16/bf16, where the exponent is shifted into byte 1 of each 2-byte element. ## Stage 2: Byte split (round-robin deinterleaving) Distributes bytes across `num_buf` planes via `byte[i] → plane[i % num_buf]`. For fp32 (`bytes_mode = 220`, `num_buf = 4`), this produces: - **Plane 0**: byte 0 of every float (mantissa low bits) - **Plane 1**: byte 1 of every float (mantissa mid bits) - **Plane 2**: byte 2 of every float (mantissa high + sign) - **Plane 3**: byte 3 of every float (**exponent** — after bit reorder) Each plane is then Huffman-compressed independently. The exponent plane compresses extremely well because trained weights cluster in a narrow exponent range (typically ~20 unique values out of 256, giving ~2.6 bits/byte entropy vs ~8.0 for mantissa planes). This is where nearly all compression savings come from. ## Why bit reorder matters The byte split (stage 2) alone is a standard byte-shuffle preprocessing step, well known in scientific-array compression. The bit reorder (stage 1) is what makes the exponent plane cleanly separable from the sign bit. Without it, the sign bit bleeds across the byte 2/byte 3 boundary, increasing the exponent byte's entropy.