Struct QuantSpec 

pub struct QuantSpec {
    pub inv_grad_scale: f32,
    pub inv_hess_scale: f32,
    pub grad_scale: f64,
    pub hess_scale: f64,
    pub grad_max_q: i32,
    pub hess_max_q: i32,
    pub constant_hessian: bool,
    pub stochastic: bool,
    pub seed: u32,
    pub iteration: u32,
}

What one iteration’s quantization needs; mirrored by QuantParams in the kernels.

Fields

inv_grad_scale: f32

1 / grad_scale, where grad_scale = max |g| / (num_grad_quant_bins / 2) over every rank.

inv_hess_scale: f32

1 / hess_scale, where hess_scale = max h / num_grad_quant_bins; unused with a constant hessian.

grad_scale: f64

hess_scale: f64

grad_max_q: i32

num_grad_quant_bins / 2.

hess_max_q: i32

num_grad_quant_bins.

constant_hessian: bool

Every hessian equals hess_scale, so h_q is 1 and hessian sums are exact row counts.

stochastic: bool

seed: u32

iteration: u32

Implementations

impl QuantSpec

pub fn new( max_abs_grad: f64, max_hess: f64, constant_hessian: Option<f64>, num_bins: u32, stochastic: bool, seed: u32, iteration: u32, ) -> QuantSpec

LightGBM GradientDiscretizer::DiscretizeGradients scales from the global maxima. All-zero gradients get scale 1 so the inverse stays finite.

pub fn max_rows_per_packed_block(&self) -> u32

Rows per GPU threadgroup so a packed i32 (two i16 halves) histogram bin never overflows.

Trait Implementations

impl Clone for QuantSpec

fn clone(&self) -> Self

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Copy for QuantSpec

impl Debug for QuantSpec

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

Formats the value using the given formatter.

impl PartialEq for QuantSpec

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

Equality operator ==.

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

Inequality operator !=.

impl StructuralPartialEq for QuantSpec