//! # Embedding blocks
//!
//! - [`RotaryEmbedding`] is rotary position embedding (`RoPE`) as a module.
//! Built by [`RotaryEmbeddingConfig`], it caches the `cos` and `sin` of
//! [`positional_frequency_table`] as bare tensors (no parameters).
//! [`apply`](RotaryEmbedding::apply) rotates a `[B, T, H, D]` query or key,
//! and [`clip_range`](RotaryEmbedding::clip_range) narrows the tables to the
//! positions of a decode step. [`CausalSelfAttention`] takes one in its
//! `forward`; the `NanoChat` GPT holds one as a field of its module tree.
//! - [`RotarySet`] is a stack's rotary tables: one or more, and which one each
//! layer applies (Gemma 3's local and global tables).
//! - [`TokenEmbedding`] looks tokens up in a table, then, as its
//! [`TokenEmbeddingConfig`] says, normalizes and scales them: nanochat's
//! normed embedding, Gemma's scaled one.
//! - `testing` (behind the `testing` feature) holds `iota_embedding` and
//! `identity_embedding`, burn `Embedding`s with weights you can check by
//! hand, as fixtures for tests.
//!
//! The table builders, and [`unembed`] for a tied output head, are operations
//! in [`crate::ops::embedding`].
//!
//! [`positional_frequency_table`]: crate::ops::embedding::positional_frequency_table
//! [`unembed`]: crate::ops::embedding::unembed
//! [`CausalSelfAttention`]: crate::blocks::transformers::attention::csa::CausalSelfAttention
//! [`Embedding`]: burn::nn::Embedding
pub use *;
pub use *;
pub use *;