Enum HardwareOptimizerKind 

pub enum HardwareOptimizerKind {
    Sgd,
    Lion,
    Adam,
    Lamb,
}

Optimizer families the hardware analysis can recommend.

The distinguishing property is how much per-parameter optimizer state each family needs, because on the memory-constrained platforms this module models that state — not the parameters themselves — is what decides whether a model fits.

Variants

Sgd

Momentum SGD: one auxiliary buffer per parameter, and none at all when momentum is zero. Recommended when optimizer state has to be paid for over a slow link (CPU offloading) or under a very small power budget.

Lion

Lion: a single momentum buffer and a sign-based update (Chen et al., “Symbolic Discovery of Optimization Algorithms”, arXiv:2302.06675). Half of Adam’s optimizer state, and the sign update is insensitive to gradient scale, which suits the quantized/low-precision arithmetic edge devices use.

Adam

Adam: first and second moments, two buffers per parameter. The default when optimizer-state memory is not the binding constraint.

Lamb

LAMB: Adam’s moments plus a layer-wise trust ratio, for the very large batches TPU and distributed configurations produce.

Implementations

impl HardwareOptimizerKind

pub fn state_buffers_per_parameter(self) -> usize

Number of per-parameter auxiliary buffers this family keeps.

Multiply by the parameter count and the element size to get the optimizer-state footprint that has to fit alongside the model.

pub fn name(self) -> &'static str

Stable short name, for diagnostics and reports.

pub fn recommend_for<A: Float>( platform: &HardwarePlatform, config: &HardwareOptimizationConfig<A>, ) -> Self

Recommend an optimizer family for a platform and the configuration the hardware analysis produced for it.

The rules are ordered from the hardest constraint to the softest:

1. Offloading optimizer state to CPU memory, or a sub-LOW_POWER_BUDGET_WATTS power budget, makes every extra per-parameter buffer expensive: use SGD.
2. Any other edge device: use Lion, which halves Adam’s state and whose sign update tolerates quantized arithmetic.
3. Batches of LARGE_BATCH_THRESHOLD or more (TPU, distributed): use LAMB.
4. Otherwise: Adam.

Trait Implementations

impl Clone for HardwareOptimizerKind

fn clone(&self) -> HardwareOptimizerKind

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Copy for HardwareOptimizerKind

impl Debug for HardwareOptimizerKind

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

Formats the value using the given formatter.

impl Eq for HardwareOptimizerKind

impl Hash for HardwareOptimizerKind

fn hash<__H: Hasher>(&self, state: &mut __H)

Feeds this value into the given Hasher.

fn hash_slice<H>(data: &[Self], state: &mut H)

where H: Hasher, Self: Sized,

Feeds a slice of this type into the given Hasher.

impl PartialEq for HardwareOptimizerKind

fn eq(&self, other: &HardwareOptimizerKind) -> bool

Equality operator ==.

fn ne(&self, other: &Rhs) -> bool

Inequality operator !=.

impl StructuralPartialEq for HardwareOptimizerKind