Struct OptItem 

pub struct OptItem<'a> {
    pub name: &'a str,
    pub shape: &'a [usize],
    pub param: &'a mut [f32],
    pub grad: &'a [f32],
}

Common parameter-update interface.

name keys the per-parameter state (moments, preconditioners), shape is the parameter’s logical shape (used by matrix-aware algorithms like Adafactor / SOAP / Muon — ignored by elementwise ones), param is updated in place from grad. grad is treated as read-only; callers that need gradient clipping should pre-scale it (see global_grad_clip_scale).

Implementing for a backend

Every algorithm in this crate provides a CPU reference impl. A backend (e.g. rlx-metal, rlx-cuda) is free to write its own fused step kernel and impl Optimizer for a wrapper struct that owns device buffers — the trait places no requirement on where the state lives, only on the entry-point signature. The rlx-metal::splat_adam kernel is the canonical example of a backend that bypasses this crate entirely; you can wrap it with a 5-line impl Optimizer if you want a uniform interface from a generic trainer.

Per-tensor learning rate

For optimizers that don’t need per-tensor LR variation (most transformer pre-training), set lr_scale to return 1.0 (the default). For domain-specific use cases — e.g. 3D Gaussian splatting, where different attributes need wildly different step sizes — override lr_scale to multiply the base lr by a per-name factor. The provided method on the trait does NOT scale automatically; algorithms are free to consult it via Optimizer::lr_scale inside their step. One parameter’s data for a batched optimizer step (Optimizer::step_batch): its name, static shape, mutable data slice (owned by the caller), and gradient.

Fields

name: &'a str

shape: &'a [usize]

param: &'a mut [f32]

grad: &'a [f32]