Realm

oa::ml::nn::CausalLinearAttention

Multi-head linear attention over packed [B*S, D] values. Uses the Katharopoulos (2020) ELU+1 kernel rather than softmax. Causal heads up to 64 use fused forward and recomputed prefix/suffix adjoints. Saved projections/context scale as O(BHSd), with bounded O(d²) scan state per head. Larger heads retain the compositional path, whose training tape saves O(BHSd²) prefixes. Bidirectional mode uses the global KV composition. Standard attention stores S² probabilities; Flash recomputes them and does not store a global S² score matrix. Parameter layout Four biased projections (q_proj, k_proj, v_proj, out_proj) with the same names and seed stride as MultiHeadAttention, so checkpoints are transferable between the two modules.

StRust public API

Declaration

pub struct CausalLinearAttention { /* private fields */ }
Public Rust declaration.

Public Methods

pub fn forward(&self, input: &Matrix) -> Result<Matrix>
pub fn set_mode(&self, mode: LinearAttentionMode)
pub fn mode(&self) -> LinearAttentionMode
pub const fn model_width(&self) -> usize
pub const fn num_heads(&self) -> usize
pub fn sequence_length(&self) -> usize
pub fn set_sequence_length(&self, sequence_length: usize) -> Result<()>
pub fn has_bias(&self) -> bool
fn forward(&self, input: &Matrix) -> Result<Matrix>
fn registry(&self) -> &ModuleRegistry
fn parameters(&self) -> Vec<Parameter>
fn named_parameters(&self) -> Vec<NamedParameter>
fn buffers(&self) -> Vec<Matrix>
fn named_buffers(&self) -> Vec<NamedBuffer>
fn all_parameters(&self) -> Result<Vec<Parameter>>
fn all_named_parameters(&self) -> Result<Vec<NamedParameter>>
fn all_named_buffers(&self) -> Result<Vec<NamedBuffer>>
fn all_named_state_u32(&self) -> Result<Vec<NamedStateU32>>
fn num_parameters(&self) -> Result<usize>
fn train(&self, training: bool)
fn eval(&self)
fn is_training(&self) -> bool
fn scoped_eval(&self) -> ScopedEval<'_>

Associated Functions

pub fn new_with_seed( engine: &Engine, model_width: usize, num_heads: usize, sequence_length: usize, seed: u64, ) -> Result<Self>
pub fn new_with_seed_and_bias( engine: &Engine, model_width: usize, num_heads: usize, sequence_length: usize, bias: bool, seed: u64, ) -> Result<Self>
pub fn new_with_seed_and_options( engine: &Engine, model_width: usize, num_heads: usize, sequence_length: usize, bias: bool, seed: u64, ) -> Result<Self>

Public Method Documentation

pub fn forward( &self, input: &Matrix ) -> Result<Matrix>
No public source comment is attached to this declaration.
pub fn set_mode( &self, mode: LinearAttentionMode )
No public source comment is attached to this declaration.
pub fn mode( &self ) -> LinearAttentionMode
No public source comment is attached to this declaration.
pub const fn model_width( &self ) -> usize
Construct four biased projections with causal visibility. Errors Returns an error unless dimensions are positive, model_width is divisible by num_heads, or child construction/registration fails.
pub const fn num_heads( &self ) -> usize
Construct four biased projections with causal visibility. Errors Returns an error unless dimensions are positive, model_width is divisible by num_heads, or child construction/registration fails.
pub fn sequence_length( &self ) -> usize
Construct four biased projections with causal visibility. Errors Returns an error unless dimensions are positive, model_width is divisible by num_heads, or child construction/registration fails.
pub fn set_sequence_length(&self, sequence_length: usize) -> Result<()>
No public source comment is attached to this declaration.
pub fn has_bias( &self ) -> bool
No public source comment is attached to this declaration.
fn forward( &self, input: &Matrix ) -> Result<Matrix>
No public source comment is attached to this declaration.
fn registry( &self ) -> &ModuleRegistry
No public source comment is attached to this declaration.
fn parameters( &self ) -> Vec<Parameter>
No public source comment is attached to this declaration.
fn named_parameters( &self ) -> Vec<NamedParameter>
No public source comment is attached to this declaration.
fn buffers( &self ) -> Vec<Matrix>
No public source comment is attached to this declaration.
fn named_buffers( &self ) -> Vec<NamedBuffer>
No public source comment is attached to this declaration.
fn all_parameters( &self ) -> Result<Vec<Parameter>>
No public source comment is attached to this declaration.
fn all_named_parameters( &self ) -> Result<Vec<NamedParameter>>
No public source comment is attached to this declaration.
fn all_named_buffers( &self ) -> Result<Vec<NamedBuffer>>
No public source comment is attached to this declaration.
fn all_named_state_u32( &self ) -> Result<Vec<NamedStateU32>>
No public source comment is attached to this declaration.
fn num_parameters( &self ) -> Result<usize>
No public source comment is attached to this declaration.
fn train( &self, training: bool )
No public source comment is attached to this declaration.
fn eval( &self )
No public source comment is attached to this declaration.
fn is_training( &self ) -> bool
No public source comment is attached to this declaration.
fn scoped_eval( &self ) -> ScopedEval<'_>
No public source comment is attached to this declaration.
pub fn new_with_seed( engine: &Engine, model_width: usize, num_heads: usize, sequence_length: usize, seed: u64 ) -> Result<Self>
No public source comment is attached to this declaration.
pub fn new_with_seed_and_bias( engine: &Engine, model_width: usize, num_heads: usize, sequence_length: usize, bias: bool, seed: u64 ) -> Result<Self>
No public source comment is attached to this declaration.
pub fn new_with_seed_and_options( engine: &Engine, model_width: usize, num_heads: usize, sequence_length: usize, bias: bool, seed: u64 ) -> Result<Self>
No public source comment is attached to this declaration.