mamba-rs 0.7.3

Mamba SSM and Mamba-3 SISO in Rust with optional CUDA acceleration: inference and training (BPTT through the SSM state, AdamW) on CPU and GPU, custom NVRTC-compiled kernels, CUDA Graph capture, f32 / bf16 / f16 storage, deterministic batch-invariant GEMMs by default with explicit cuBLAS Fast and Pedantic modes.
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17

//! CPU backend for Mamba SSM.
//!
//! - `inference` — zero-alloc T=1 step (single + batch + sequence)
//! - `forward` — training forward pass (batched SGEMM)
//! - `backward` — training backward pass (BPTT)
//! - `target` — target network forward (no activation saves)

pub mod backward;
pub mod backward_ops;
pub mod flat;
pub mod forward;
pub mod inference;
pub mod parallel;
pub mod prefill;
pub mod scratch;
pub mod target;
pub mod weights;