ruvllm_sparse_attention 0.1.1

Subquadratic O(N log N) sparse attention kernel for Rust LLM inference on edge devices, with optional FastGRNN salience gating for near-linear O(N) scaling
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41

use crate::attention::{AttentionBackend, AttentionError, SparseAttentionConfig, SubquadraticSparseAttention};
use crate::tensor::Tensor3;

#[derive(Clone, Debug)]
pub struct RuvLlmSparseBlockConfig {
    pub attention: SparseAttentionConfig,
}

impl Default for RuvLlmSparseBlockConfig {
    fn default() -> Self {
        Self {
            attention: SparseAttentionConfig::default(),
        }
    }
}

#[derive(Clone, Debug)]
pub struct RuvLlmSparseBlock {
    attention: SubquadraticSparseAttention,
}

impl RuvLlmSparseBlock {
    pub fn new(config: RuvLlmSparseBlockConfig) -> Result<Self, AttentionError> {
        Ok(Self {
            attention: SubquadraticSparseAttention::new(config.attention)?,
        })
    }

    pub fn forward_qkv(
        &self,
        q: &Tensor3,
        k: &Tensor3,
        v: &Tensor3,
    ) -> Result<Tensor3, AttentionError> {
        self.attention.forward(q, k, v)
    }

    pub fn attention_backend(&self) -> &SubquadraticSparseAttention {
        &self.attention
    }
}