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
42
43

// no_std + alloc support: works on ESP32-S3 and other bare-metal targets
// when the default `std` feature is disabled. With `std` (the default),
// the crate behaves exactly as before. The feature is purely additive
// for backwards compatibility — no existing consumer needs to change.
#![cfg_attr(not(feature = "std"), no_std)]

extern crate alloc;

// On no_std targets, f32 method calls like `.exp()` / `.sqrt()` / `.tanh()`
// are unavailable because they need libm. This trait restores them via
// the `libm` crate, so existing math code at the call site stays the same.
// On std builds the trait is not defined and the inherent f32 methods are
// used as before — zero behavioural change for std consumers.
#[cfg(not(feature = "std"))]
pub mod no_std_math {
    pub trait F32Ext {
        fn exp(self) -> Self;
        fn sqrt(self) -> Self;
        fn tanh(self) -> Self;
        fn powi(self, n: i32) -> Self;
    }
    impl F32Ext for f32 {
        #[inline] fn exp(self) -> Self { libm::expf(self) }
        #[inline] fn sqrt(self) -> Self { libm::sqrtf(self) }
        #[inline] fn tanh(self) -> Self { libm::tanhf(self) }
        #[inline] fn powi(self, n: i32) -> Self { libm::powf(self, n as f32) }
    }
}

pub mod attention;
pub mod fastgrnn_gate;
pub mod model;
pub mod tensor;

pub use attention::{
    dense_attention, AttentionBackend, AttentionError, IncrementalLandmarks, KvCache,
    SparseAttentionConfig, SubquadraticSparseAttention,
};
#[cfg(feature = "fp16")]
pub use attention::KvCacheF16;
pub use fastgrnn_gate::{FastGrnnGate, DEFAULT_HIDDEN_DIM as FASTGRNN_DEFAULT_HIDDEN_DIM};
pub use model::{RuvLlmSparseBlock, RuvLlmSparseBlockConfig};
pub use tensor::Tensor3;