//! Candle (GPU) backend for box embeddings.
//!
//! Provides differentiable box and Gumbel box embeddings using Candle's
//! autograd. Training is done via `CandleBoxTrainer`.
//!
//! # Missing geometries
//!
//! The Candle backend currently only supports **Box** and **GumbelBox**.
//! The following geometries are available only via the ndarray (CPU) backend:
//!
//! - **Cone** — Candle autograd for angular representations requires careful
//! handling of periodic boundary conditions (`sin`/`cos` gradients).
//! - **Octagon** — Diagonal constraints are expensive to differentiate;
//! the CPU implementation uses polygon clipping which doesn't translate
//! directly to tensor ops.
//! - **Gaussian** — Complex number representations need custom Candle kernels.
//!
//! If you need GPU training for these geometries, consider contributing
//! implementations to the `candle_backend` module.
pub use CandleBox;
pub use CandleGumbelBox;
pub use ;