Struct MixedPrecisionConfig 

pub struct MixedPrecisionConfig {
    pub compute_dtype: PrecisionMode,
    pub param_dtype: PrecisionMode,
    pub loss_scaling: LossScalingStrategy,
    pub gradient_checkpointing: bool,
    pub gradient_clipping: bool,
    pub max_gradient_norm: f64,
    pub stability_checks: bool,
    pub skip_on_overflow: bool,
    pub use_master_weights: bool,
}

Mixed precision training configuration.

Fields

compute_dtype: PrecisionMode

Precision mode for computation

param_dtype: PrecisionMode

Precision mode for parameters (usually FP32)

loss_scaling: LossScalingStrategy

Loss scaling strategy

gradient_checkpointing: bool

Enable gradient checkpointing to save memory

gradient_clipping: bool

Enable gradient clipping before unscaling

max_gradient_norm: f64

Maximum gradient norm for clipping

stability_checks: bool

Enable numerical stability checks

skip_on_overflow: bool

Skip optimizer step if overflow detected

use_master_weights: bool

Enable master weights in FP32

Implementations

impl MixedPrecisionConfig

pub fn new(compute_dtype: PrecisionMode, param_dtype: PrecisionMode) -> Self

Create a new mixed precision config.

pub fn with_compute_dtype(self, dtype: PrecisionMode) -> Self

Set the compute dtype.

pub fn with_param_dtype(self, dtype: PrecisionMode) -> Self

Set the parameter dtype.

pub fn with_loss_scaling(self, strategy: LossScalingStrategy) -> Self

Set the loss scaling strategy.

pub fn with_gradient_checkpointing(self, enabled: bool) -> Self

Enable or disable gradient checkpointing.

pub fn with_gradient_clipping(self, enabled: bool, max_norm: f64) -> Self

Enable or disable gradient clipping.

pub fn with_stability_checks(self, enabled: bool) -> Self

Enable or disable stability checks.

pub fn with_master_weights(self, enabled: bool) -> Self

Enable or disable master weights.

pub fn fp16() -> Self

Create FP16 mixed precision config.

pub fn bf16() -> Self

Create BF16 mixed precision config.

pub fn fp8() -> Self

Create FP8 mixed precision config (experimental).

pub fn validate(&self) -> Result<(), MixedPrecisionError>

Validate the configuration.

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

Formats the value using the given formatter.

impl Default for MixedPrecisionConfig

fn default() -> Self

Returns the “default value” for a type.

impl<'de> Deserialize<'de> for MixedPrecisionConfig

fn deserialize<__D>(__deserializer: __D) -> Result<Self, __D::Error>

where __D: Deserializer<'de>,

Deserialize this value from the given Serde deserializer.

impl PartialEq for MixedPrecisionConfig

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

Tests for self and other values to be equal, and is used by ==.

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

Tests for !=. The default implementation is almost always sufficient, and should not be overridden without very good reason.

impl Serialize for MixedPrecisionConfig

fn serialize<__S>(&self, __serializer: __S) -> Result<__S::Ok, __S::Error>

where __S: Serializer,

Serialize this value into the given Serde serializer.

impl StructuralPartialEq for MixedPrecisionConfig