Crate axonml_core 

Foundation layer for the AxonML deep learning framework.

Provides the Device enum (CPU, CUDA, Vulkan, Metal, WebGPU) with runtime capability queries, the Scalar/Numeric/Float trait hierarchy for generic type-safe dispatch, reference-counted Storage<T> with pooled GPU allocations, and five compute backends: CPU (rayon-parallel GEMM/GEMV via matrixmultiply), CUDA (cuBLAS + 15 custom PTX kernel modules covering elementwise ops, activations, attention, Q4_K/Q6_K dequant-in-shader matmul, softmax, layernorm, RMSNorm, transpose, and embedding gather), Vulkan (ash + gpu-allocator, full buffer/pipeline/dispatch), Metal (full buffer/pipeline/dispatch on Apple Silicon), and WebGPU (wgpu, full buffer/pipeline/dispatch for browser targets).

File

crates/axonml-core/src/lib.rs

Author

Andrew Jewell Sr. — AutomataNexus LLC ORCID: 0009-0005-2158-7060

Updated

April 14, 2026 11:15 PM EST

Disclaimer

Use at own risk. This software is provided “as is”, without warranty of any kind, express or implied. The author and AutomataNexus shall not be held liable for any damages arising from the use of this software.

Re-exports

pub use allocator::Allocator;

pub use allocator::DefaultAllocator;

pub use device::Device;

pub use dtype::DType;

pub use dtype::Float;

pub use dtype::Numeric;

pub use dtype::Scalar;

pub use error::Error;

pub use error::Result;

pub use storage::Storage;

Modules

allocator

Memory allocation traits and the default CPU allocator.

backends

Compute backend modules and the Backend trait.

device

Device abstraction and hardware backend management.

dtype

Type system for AxonML’s generic tensor operations.

error

Error types for the AxonML core crate.

prelude

Convenient imports for common usage.

storage

Reference-counted raw memory management for tensors.