Struct Sgd 

pub struct Sgd {
    pub lr: f32,
    pub momentum: f32,
    pub nesterov: bool,
    pub weight_decay: f32,
    /* private fields */
}

SGD with momentum / Nesterov / L2 weight decay.

All hyperparameters are public so callers can hot-swap them between iterations (e.g. for a warm-up schedule). State is keyed by parameter name; the same Sgd instance can drive every tensor in a model.

Fields

lr: f32

Learning rate. No default — pass it to Sgd::new.

momentum: f32

Polyak momentum coefficient ∈ [0, 1). 0.0 disables momentum entirely (and the per-tensor velocity buffer is still allocated but unused — set via Sgd::with_momentum if you want it on).

nesterov: bool

Use Nesterov-accelerated momentum. Only meaningful when momentum > 0.

weight_decay: f32

L2 weight decay coefficient λ. Folded into the gradient before the momentum EMA (classical, not decoupled). Use crate::AdamW-style decoupling if you need that.

Implementations

impl Sgd

pub fn new(lr: f32) -> Self

Construct with lr and momentum / decay disabled.

Examples found in repository

examples/step_bench.rs (line 76)

fn main() {
    for n in [64 * 1024usize, 1024 * 1024, 8 * 1024 * 1024] {
        println!(
            "\n{} elements ({:.1} MiB per buffer)",
            n,
            (n * 4) as f64 / 1048576.0
        );
        bench("adamw (f64 default)", n, Box::new(AdamW::new(1e-3)));
        bench(
            "adamw f32Math",
            n,
            Box::new(AdamW::new(1e-3).with_f32_math(true)),
        );
        bench("adam (f64 default)", n, Box::new(Adam::new(1e-3)));
        bench(
            "adam f32Math",
            n,
            Box::new(Adam::new(1e-3).with_f32_math(true)),
        );
        bench("sgd+momentum", n, {
            let mut o = Sgd::new(1e-3);
            o.momentum = 0.9;
            Box::new(o)
        });
        bench("lion", n, Box::new(Lion::new(1e-3)));
    }
}

pub fn with_momentum(self, momentum: f32, nesterov: bool) -> Self

Enable Polyak (or Nesterov) momentum.

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

Set the L2 weight-decay coefficient.

Trait Implementations

impl Clone for Sgd

fn clone(&self) -> Self

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for Sgd

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

Formats the value using the given formatter.

impl Optimizer for Sgd

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 state_dict(&self) -> Option<OptimizerState>

Snapshot the optimizer’s state for a checkpoint.

fn load_state_dict(&mut self, state: &OptimizerState) -> bool

Restore a snapshot. Returns false when unsupported or when the state does not belong to this algorithm.

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

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 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 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.