Crate numr 

numr

High-performance numerical computing for Rust with multi-backend GPU acceleration.

numr provides n-dimensional arrays (tensors), linear algebra, FFT, and automatic differentiation - with the same API across CPU, CUDA, and WebGPU backends.

Why numr?

• Multi-backend: Same code runs on CPU, CUDA, and WebGPU
• No vendor lock-in: Native kernels, not cuBLAS/MKL wrappers
• Pure Rust: No Python runtime, no FFI overhead, single binary deployment
• Autograd included: Reverse-mode automatic differentiation built-in
• Sparse tensors: CSR, CSC, COO formats with GPU support

Features

• Tensors: N-dimensional arrays with broadcasting, slicing, views
• Linear algebra: Matmul, LU, QR, SVD, Cholesky, eigendecomposition
• FFT: Fast Fourier transforms (1D, 2D, ND)
• Element-wise ops: Full set of math functions
• Reductions: Sum, mean, max, min, argmax, argmin along axes
• Multiple dtypes: f64, f32, f16, bf16, fp8, integers, bool

Quick Start

use numr::prelude::*;

let device = CpuDevice;
let a = Tensor::<CpuRuntime>::from_slice(&[1.0, 2.0, 3.0, 4.0], &[2, 2], &device);
let b = Tensor::<CpuRuntime>::from_slice(&[5.0, 6.0, 7.0, 8.0], &[2, 2], &device);

let c = &a + &b;
let d = a.matmul(&b)?;

Feature Flags

• cpu (default): CPU backend
• cuda: NVIDIA CUDA backend
• wgpu: Cross-platform GPU via WebGPU
• rayon (default): Multi-threaded CPU operations
• f16: Half-precision floats (F16, BF16)
• sparse: Sparse tensor formats (CSR, CSC, COO)

Modules

algorithm

Algorithm contracts for runtime backends

autograd

Automatic differentiation (autograd)

dtype

Data type system for numr tensors.

error

Error types for numr

ops

Tensor operations

prelude

Prelude module for convenient imports

runtime

Runtime backends for tensor computation

sparse

Sparse tensor support for numr

tensor

Tensor types and operations

Macros

dispatch_dtype

Macro for runtime dtype dispatch to typed operations.

Type Aliases

DefaultRuntime

Default runtime based on enabled features