Crate oxionnx 

oxionnx — Pure Rust ONNX inference engine.

Zero C/C++ dependencies. Supports a wide subset of ONNX operators sufficient to run transformer-based models (BERT, T5, GPT-2 family).

Quick start

use std::collections::HashMap;
use oxionnx::{Session, Tensor};

fn main() -> Result<(), Box<dyn std::error::Error>> {
    let session = Session::from_file("model.onnx".as_ref())?;
    let mut inputs = HashMap::new();
    inputs.insert("input_ids", Tensor::new(vec![1.0, 2.0, 3.0], vec![1, 3]));
    let outputs = session.run(&inputs)?;
    Ok(())
}

Re-exports

pub use benchmark_select::benchmark_elementwise;

pub use benchmark_select::benchmark_matmul;

pub use benchmark_select::benchmark_suite;

pub use benchmark_select::format_results;

pub use benchmark_select::BenchmarkResult;

pub use benchmark_select::ExecutionPath;

pub use compat::GraphOptimizationLevel;

pub use compat::SessionOutputs;

pub use execution_providers::CPUExecutionProvider;

pub use execution_providers::CUDAExecutionProvider;

pub use execution_providers::CoreMLExecutionProvider;

pub use execution_providers::DirectMLExecutionProvider;

pub use execution_providers::ExecutionProviderDispatch;

pub use execution_providers::OpenVINOExecutionProvider;

pub use execution_providers::TensorRTExecutionProvider;

pub use io_binding::IoBinding;

pub use memory::BufferPool;

pub use memory::MemoryPlan;

pub use memory::PoolStats;

pub use memory::SizeClassPool;

pub use memory_tracker::MemoryTracker;

pub use session::ModelInfo;

pub use session::ModelMetadata;

pub use session::NodeProfile;

pub use session::OptLevel;

pub use session::Session;

pub use session::SessionBuilder;

pub use tolerance::compare_tensors;

pub use tolerance::tensors_close;

pub use tolerance::ToleranceReport;

pub use typed_session::Shape;

pub use typed_session::TypedSession;

Modules

benchmark_select

Benchmark-based execution path selection.

compat

Compatibility layer for easy migration from the ort crate.

execution_providers

Execution Provider compatibility layer for migration from ort.

graph

io_binding

Pre-allocated I/O buffer management for zero-copy inference.

macros

Convenience macros for building inference input maps.

memory

Auto-generated module structure

memory_tracker

Peak memory usage tracking for inference runs.

model

ops

opset_coverage

Opset version coverage report generator.

optimizer

Graph optimization passes for ONNX inference. Applied after model loading but before topological sort and execution.

proto

session

tensor

tolerance

Numerical tolerance validation for inference results.

typed_session

Typed inference API with compile-time shape checking.

Macros

define_shape

Macro to define a shape type.

inputs

Build a Result<HashMap<&str, Tensor>, OnnxError> input map.

inputs_typed

Build a typed input map for use with crate::Session::run_typed.

Structs

Attributes

Parsed attributes for a node.

Graph

Compute graph produced by parsing an ONNX model.

Node

A single computation node in the ONNX graph.

OpContext

Runtime context passed to every operator during execution.

OperatorRegistry

Maps ONNX op_type strings to operator implementations.

Tensor

N-dimensional tensor with f32 data and a shape vector. Layout: row-major (C order), last dimension varies fastest.

TensorInfo

Metadata for a model input or output tensor (parsed from ValueInfoProto).

TypedTensor

A multi-dtype tensor with shape and typed storage. This complements the existing Tensor (which is f32-only) for advanced use cases.

Enums

DType

ONNX data types.

Dim

A tensor dimension: either a concrete size, a symbolic name, or unknown.

Error

Typed error returned by all public Session methods.

OnnxError

Typed error returned by all public Session methods.

OpKind

All ONNX operators currently supported by oxionnx.

TensorStorage

Type-erased tensor storage.

Traits

Operator

Trait for ONNX operator implementations. Operators are stateless – all runtime state comes through OpContext.

Functions

default_registry

Build the default operator registry containing all supported ONNX operators.