veloxquant_mlx.handlers.bit_pack_handler module¶

Handler stage that packs/unpacks codebook indices via BitPackBuffer.

Wraps veloxquant_mlx.dsa.bit_pack.BitPackBuffer as a QuantizationHandler pipeline stage: on encode it bit-packs ctx.indices down to b bits per element into ctx.packed_bits; on decode it unpacks ctx.packed_bits back into ctx.indices. Typically the last stage in an encode chain, after scalar quantization has produced codebook indices.

class veloxquant_mlx.handlers.bit_pack_handler.BitPackingHandler(b, d)[source]¶

Bases: QuantizationHandler

Pack codebook indices into b-bit compact storage.

On encode:: Reads ctx.indices (batch, d) uint8. Packs each row independently using BitPackBuffer. ctx.packed_bits = packed byte array.
On decode:: Reads ctx.packed_bits and the original n (d). Unpacks back to (batch, d) uint8 and stores in ctx.indices.

Parameters:

b (int) – Bit-width per index.
d (int) – Vector dimension (needed for unpack).

handle(ctx)[source]¶

Pack or unpack codebook indices.

Parameters:: ctx (QuantizationContext) – Quantization context.
Return type:: QuantizationContext
Returns:: Updated context.

property handler_name: str¶: Registry name for this handler (‘bit_packing’).