tokitai-operator 0.1.0

//! IR Model layer (Phase 2.1 of the 0.7B MoE training project).
//!
//! Provides:
//! - [`parameter::Parameter`] — trainable weight tensor + AdamW state.
//! - [`layer::Layer`] trait + concrete layers (`Linear`, `LayerNorm`,
//!   `GELU`, `Softmax`, `Add`, `Router`, `Expert`, `Embedding`).
//! - [`sequential::Sequential`] — chains layers, holds a
//!   `ForwardCache`, and implements the [`sequential::Model`] trait.
//!
//! All numerical work happens in the existing HIP fp16 kernels in
//! `src/backend/hip_*.rs`. The Model layer's job is purely to
//! orchestrate kernel calls and shuttle fp32 activations through
//! fp16 (de)quantization at the kernel boundary.

pub mod layer;
pub mod parameter;
pub mod sequential;
pub mod util;

pub use layer::{
    Add, Embedding, Expert, GELU, Layer, LayerNorm, Linear, LinearResidual, Router, Softmax,
};
pub use parameter::Parameter;
pub use sequential::{ForwardCache, Model, Sequential};