Provides a deterministic, CPU-only fixed-point inference path using i16 tensors with a global scale (scale_log2). Includes pure integer arithmetic functions (requantize with widene i64 accumulation, ties-round-away-from-zero), a FixedLinear struct that implements integer dense layers, and exact integer ReLU. This module is self-contained in its arithmetic logic and can be extracted for any integer-only inference backend, demonstrating a clean pattern for producing bit-exact results suitable for zero-knowledge proof systems.

Implements a custom binary format (.rrec) for high-throughput data loading using memory-mapped files (memmap2). Provides a RecordWriter that packs tensor metadata, data, and padding for alignment, and a RecordReader that yields tensors from the mmapped region. The format is simple and reusable for any machine learning pipeline requiring efficient I/O, and the code is well-structured with proper error handling and explicit unsafe for mmap only.

Implements a #[derive(Module)] proc macro that auto-generates methods like train(), eval(), parameters(), state_dict() for named-field structs. Uses syn/quote for token manipulation and handles recursive field traversal. This macro is a clean example of Rust metaprogramming for reducing boilerplate in ML frameworks and can be adapted (with trait changes) for similar derive patterns in other Rust projects.