Struct YarnScaling 

pub struct YarnScaling {
    pub factor: f32,
    pub beta_fast: f32,
    pub beta_slow: f32,
    pub orig_max_pos: usize,
    pub truncate: bool,
}

YaRN RoPE scaling exactly as a checkpoint declares it, in the shape the reference reads out of rope_scaling (FreeToken python/freetoken/layers/rotary.py:139, the "yarn" arm of _get_rope). beta_fast / beta_slow / truncate carry that arm’s own defaults, because a real YaRN checkpoint usually declares only factor and original_max_position_embeddings.

This type is only the declaration. The frequency rewrite it implies is yarn_freq_factors, whose output feeds apply_rope_with_freq_factors / apply_rope_interleaved_with_freq_factors like any other per-band correction.

Fields

factor: f32

rope_scaling["factor"]: how much longer the served context is than the context the checkpoint was trained at. 1.0 makes the whole rewrite a no-op (every band’s divisor is 1.0).

beta_fast: f32

rope_scaling["beta_fast"], the number of full rotations that marks the high-frequency end of the correction ramp – bands faster than this are left extrapolated. Reference default 32.0.

beta_slow: f32

rope_scaling["beta_slow"], the rotation count marking the low-frequency end – bands slower than this are fully interpolated. Reference default 1.0.

orig_max_pos: usize

rope_scaling["original_max_position_embeddings"]: the context the checkpoint was actually trained at, which is the length the rotation counts above are counted against.

truncate: bool

rope_scaling["truncate"], reference default true: floor the low end and ceil the high end of the correction range to whole band indices. false keeps the fractional range, which is the case the low == high nudge in yarn_correction_range exists for.

Implementations

impl YarnScaling

pub fn new(factor: f32, orig_max_pos: usize) -> Self

The reference’s defaults for everything a checkpoint may omit (rope_scaling.get("beta_fast", 32.0), .get("beta_slow", 1.0), .get("truncate", True)). Using anything else here silently moves the correction range and therefore which dims extrapolate.

Trait Implementations

impl Clone for YarnScaling

fn clone(&self) -> YarnScaling

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Copy for YarnScaling

impl Debug for YarnScaling

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

Formats the value using the given formatter.

impl PartialEq for YarnScaling

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

Equality operator ==.

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

Inequality operator !=.

impl StructuralPartialEq for YarnScaling