RUST / CUDA NEURAL RUNTIME

Bring models
closer to the GPU.

A runtime for inference and training, powered by CUDA-Oxide. No DSL, no host/device FFI boundary — just Rust to PTX and a clear data path.

EXPLORE THE RUNTIME
inference.rs● PTX ready
let runtime = CudaRuntime::new()?;

let input = runtime.new_matrix(
  InitType::Random, 256, 128
);

let output = projection.forward(
  &input, None, &runtime
);
runtime.sync();
POWERED BY CUDA-OXIDEF32 · ROW-MAJOR · EXPLICIT

THE RUNTIME, NOT A FRAMEWORK

The runtime,
not a framework.

Built for neural networks with known shapes and controlled layouts. The performance path does not detour through a general tensor system.

THE EXECUTION CONTRACT

ONE HEAD. FULL PATH.

A Transformer’s
full round trip.

A single-head Post-LN Transformer covers inference and training execution. The training path keeps only what backward computation genuinely needs.