Struct TrainingFreeLoopConfig 

pub struct TrainingFreeLoopConfig {
    pub window_start: usize,
    pub window_end: usize,
    pub loop_count: usize,
    pub strategy: SubStepStrategy,
    pub iteration_mode: IterationMode,
    pub cache_strategy: CacheStrategy,
}

Configuration for training-free loop wrapper (Plan 136).

Pure inference-time retrofit: re-applies a contiguous mid-stack block of layers with ODE-motivated damped sub-stepping. No training needed.

Fields

window_start: usize

Start of the loop window (inclusive layer index).

window_end: usize

End of the loop window (inclusive layer index).

loop_count: usize

Number of loop iterations (K in the paper).

strategy: SubStepStrategy

Sub-step integration strategy.

iteration_mode: IterationMode

Iteration mode (block vs layer-wise).

cache_strategy: CacheStrategy

KV cache write strategy.

Implementations

impl TrainingFreeLoopConfig

pub fn from_config(config: &Config) -> Self

Creates a config with sensible defaults for a given Config.

Window heuristic: center at 48% depth, ±1 layer. For small models (≤4 layers), defaults to (0, n_layer−1). Uses the paper-recommended K-stage RK strategy with β=0.5.

Trait Implementations

impl Clone for TrainingFreeLoopConfig

fn clone(&self) -> TrainingFreeLoopConfig

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for TrainingFreeLoopConfig

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

Formats the value using the given formatter.

impl Default for TrainingFreeLoopConfig

fn default() -> Self

Returns the “default value” for a type.