Module service 

Stateless Codec service (Phase 15).

Pipeline mirrors tinyquant_cpu.codec.Codec exactly:

• compress: rotate → quantize → (optional) residual on rotated vs reconstructed
• decompress: dequantize → (optional) add residual → inverse rotate

Phase 26 adds compress_prepared / decompress_prepared_into which accept a pre-built PreparedCodec so the O(dim²) rotation factorization is paid only once per session rather than on every call.

Structs

Codec

Zero-sized stateless codec service. Mirrors Python tinyquant_cpu.codec.Codec.

Constants

GPU_BATCH_THRESHOLD

Minimum batch size below which GPU offload is not attempted.

Traits

GpuComputeBackend

Trait that every TinyQuant GPU compute backend must satisfy.

Functions

compress

Module-level compress free function — mirrors tinyquant_cpu.codec.compress.

decompress

Module-level decompress free function — mirrors tinyquant_cpu.codec.decompress.