Struct RtTurboConfig 

pub struct RtTurboConfig {
    pub low_dim: usize,
    pub sliding_window: usize,
    pub sink_tokens: usize,
    pub block_size: usize,
    pub retrieval_head_ratio: f32,
    pub top_p: f32,
    pub calibration_mode: CalibrationMode,
}

Configuration for RTPurbo retrieval head sparse decode.

Feature gate: rt_turbo (opt-in, requires dash_attn). Adds head-wise retrieval/local classification + dynamic top-p token selection for decode-phase sparse attention. Complements DashAttention’s α-entmax block routing with per-head specialization.

Must pass 6/6 GOAT proofs before default-on promotion.

Fields ordered by descending alignment to minimize padding: usize (8B) → f32 (4B) → CalibrationMode (1B) — no padding between groups.

Calibration mode (Plan 358)

CalibrationMode::AttentionMass is the default (cheaper: 1 forward pass). CalibrationMode::CausalNecessity is opt-in — strictly stronger on workloads with correlated bystanders but ~10–100× more expensive to calibrate. See calibrate_from_causal_scores in rt_turbo::calibration.

Fields

low_dim: usize

Low-dimensional projection size for pre-RoPE scoring (default: 16). Paper ablation: dim=16 is the sweet spot for low-frequency retrieval.

sliding_window: usize

Sliding window size for local heads (default: 8192).

sink_tokens: usize

Number of attention sink tokens always retained for local heads (default: 4).

block_size: usize

Block size for block-level top-p variant (default: 64). Should match DashAttnConfig::chunk_size for consistent routing.

retrieval_head_ratio: f32

Fraction of heads classified as retrieval heads (default: 0.15). Paper ablation: 15% is optimal balance of accuracy vs sparsity.

top_p: f32

Cumulative attention mass threshold for dynamic top-p selection (default: 0.9). Paper ablation: top-p=0.9 preserves >93% attention mass at 97% sparsity.

calibration_mode: CalibrationMode

Which score semantics to use for head calibration (Plan 358). Default: AttentionMass (cheaper). CausalNecessity is opt-in — strictly stronger on bystander-heavy workloads but ~10–100× more expensive. AdaptiveCausal (Proposal 004) is opt-in — cheap-proxy escalate, unvalidated, requires per-head OV norms from the caller.

Trait Implementations

impl Clone for RtTurboConfig

fn clone(&self) -> RtTurboConfig

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Copy for RtTurboConfig

impl Debug for RtTurboConfig

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

Formats the value using the given formatter.

impl Default for RtTurboConfig

fn default() -> Self

Returns the “default value” for a type.