Crate oxionnx_cuda 

oxionnx-cuda

CUDA-accelerated dispatch for ONNX ops via the OxiCUDA GPU stack.

This crate provides:

• CudaContext — a wrapper around a CUDA device context + DNN handle, constructed lazily via CudaContext::try_new.
• CudaError — error type returned by the CUDA dispatch layer.
• try_cuda_dispatch — the top-level dispatch function called from oxionnx::session::run_sequential_inner when the cuda feature is enabled.

Dispatch flow

CUDA (highest priority)
  └─ try_cuda_dispatch → Ok(Some(results))   ← GPU handled it
     └─ Ok(None)                              ← fall back to wgpu / CPU
wgpu GPU dispatch
CPU dispatch

Graceful degradation

On any CUDA error during dispatch, the function returns Err(...) which the caller maps to OnnxError::Internal. If CUDA is not available at session build time, CudaContext::try_new() returns None and no CUDA dispatch is attempted.

Re-exports

pub use context::CudaContext;

pub use error::CudaDispatchError as CudaError;

Modules

context

CUDA context wrapper for oxionnx-cuda.

conv

CUDA-accelerated 2-D convolution dispatch.

elementwise

CUDA-accelerated elementwise operator dispatch.

error

Error types for CUDA-accelerated ONNX dispatch.

matmul

CUDA-accelerated MatMul / Gemm dispatch.

reduce

CUDA-accelerated ReduceSum / ReduceMax dispatch.

softmax

CUDA-accelerated Softmax dispatch.

Functions

try_cuda_dispatch

Attempt to dispatch a single ONNX node to the CUDA backend.