Struct QHAdamW 

pub struct QHAdamW {
    pub lr: f32,
    pub beta1: f32,
    pub beta2: f32,
    pub nu1: f32,
    pub nu2: f32,
    pub eps: f32,
    pub weight_decay: f32,
    /* private fields */
}

Quasi-hyperbolic AdamW. Per-tensor state: two f32 buffers.

Fields

lr: f32

Learning rate.

beta1: f32

First-moment EMA decay β₁. Ma & Yarats recommend 0.995 (much closer to 1 than vanilla Adam) — the QH interpolation already keeps current-gradient weight in the numerator.

beta2: f32

Second-moment EMA decay β₂. Default 0.999.

nu1: f32

First-moment QH interpolation coefficient ν₁ ∈ [0, 1]. 1.0 = pure EMA (standard Adam first moment); 0.0 = pure current gradient (no momentum). Default 0.7.

nu2: f32

Second-moment QH interpolation coefficient ν₂. 1.0 = standard Adam denominator. Default 1.0.

eps: f32

Denominator stability constant. Default 1e-8.

weight_decay: f32

Decoupled weight-decay coefficient λ. Default 0.01.

Implementations

impl QHAdamW

pub fn new(lr: f32) -> Self

Construct with (β₁, β₂, ν₁, ν₂, ε, λ) = (0.995, 0.999, 0.7, 1.0, 1e-8, 0.01).

pub fn with_betas(self, b1: f32, b2: f32) -> Self

Override (β₁, β₂).

pub fn with_nus(self, n1: f32, n2: f32) -> Self

Override the quasi-hyperbolic coefficients (ν₁, ν₂).

pub fn with_weight_decay(self, wd: f32) -> Self

Override the decoupled-decay coefficient.

Trait Implementations

impl Clone for QHAdamW

fn clone(&self) -> QHAdamW

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for QHAdamW

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Optimizer for QHAdamW

fn set_lr(&mut self, lr: f32)

Set the base learning rate (for LR schedules / warmup). Default is a no-op for algorithms without a scalar lr (e.g. Adafactor’s relative step size); every algorithm in this crate that has an lr field overrides this to update it.

fn step( &mut self, name: &str, _shape: &[usize], param: &mut [f32], grad: &[f32], )

fn end_iteration(&mut self)

Advance the global step counter. Most algorithms increment per call to [step], so most implementations leave this a no-op.

fn step_batch(&mut self, items: &mut [OptItem<'_>])

Batched step over ALL parameters in one call. Default: sequential step per item — bit-identical to the per-parameter loop. Optimizers whose parameter groups are independent (e.g. Muon on the 2-D weight matrices vs AdamW on the embeddings/biases/norms) can override this to run the groups on separate threads; because the groups touch disjoint parameters and disjoint optimizer state, the result is bit-for-bit the same as the serial loop — only the wall-clock (the CPU-side optimizer bubble) shrinks toward max(group_times) instead of their sum.

fn lr_scale(&self, _name: &str) -> f32

Per-tensor multiplier on the effective learning rate. Default is 1.0 for every name. Override when wrapping this crate to support per-name LR schedules (e.g. embedding-vs-attention splits, or the Gaussian-splat attribute-typed LR setup). The CPU impls in this crate currently honor this only when the caller passes a pre-scaled lr for the relevant call — backends are encouraged to consult it inside their fused kernel.