Struct MixedPrecisionConfig 

pub struct MixedPrecisionConfig {
    pub compute_precision: PrecisionMode,
    pub master_precision: PrecisionMode,
    pub loss_scale: f32,
    pub dynamic_loss_scale: bool,
    pub min_loss_scale: f32,
    pub max_loss_scale: f32,
    pub scale_growth_factor: f32,
    pub scale_backoff_factor: f32,
    pub scale_growth_interval: usize,
}

Mixed precision training configuration.

Fields

compute_precision: PrecisionMode

Precision mode for computation

master_precision: PrecisionMode

Precision mode for master weights (usually FP32)

loss_scale: f32

Loss scale factor to prevent gradient underflow

dynamic_loss_scale: bool

Enable dynamic loss scaling

min_loss_scale: f32

Minimum loss scale for dynamic scaling

max_loss_scale: f32

Maximum loss scale for dynamic scaling

scale_growth_factor: f32

Growth factor for dynamic loss scaling

scale_backoff_factor: f32

Backoff factor for dynamic loss scaling

scale_growth_interval: usize

Number of consecutive non-overflow steps before increasing scale

Implementations

impl MixedPrecisionConfig

pub fn new(compute_precision: PrecisionMode) -> MixedPrecisionConfig

Create a new mixed precision configuration.

pub fn fp16() -> MixedPrecisionConfig

Create a configuration for FP16 training.

pub fn bf16() -> MixedPrecisionConfig

Create a configuration for BF16 training.

pub fn fp32() -> MixedPrecisionConfig

Create a configuration for FP32 training (no mixed precision).

Trait Implementations

impl Clone for MixedPrecisionConfig

fn clone(&self) -> MixedPrecisionConfig

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for MixedPrecisionConfig

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

Formats the value using the given formatter.

impl Default for MixedPrecisionConfig

fn default() -> MixedPrecisionConfig

Returns the “default value” for a type.