Struct MixedPrecisionPass 

pub struct MixedPrecisionPass {
    pub policy: MixedPrecisionPolicy,
}

Mixed-precision quantization pass.

Applies different precisions to different global variables based on a MixedPrecisionPolicy. This allows keeping sensitive layers (like Softmax or LayerNorm) at higher precision while quantizing the bulk of computation to INT8 or F16.

Fields

policy: MixedPrecisionPolicy

The policy dictating per-variable precision.

Trait Implementations

impl Debug for MixedPrecisionPass

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

Formats the value using the given formatter.

impl Pass for MixedPrecisionPass

fn name(&self) -> &str

Human-readable name of the pass.

fn run(&self, module: &mut Module) -> bool

Run the pass on a module. Returns true if anything was modified.