Coming from Microsoft.ML.OnnxRuntime · onemodel docs

OneModel runs the same ONNX files, in managed code, with no native library. This page maps the ONNX Runtime C# API onto OneModel's. The ONNX Runtime snippets are shown as plain text for comparison (they are not built here); the OneModel snippets are built and run in CI.

The same job in both

ONNX Runtime:

Text

using var options = new SessionOptions { IntraOpNumThreads = 4 };
using var session = new InferenceSession("model.onnx", options);
 
var data = new float[32];
using var x = OrtValue.CreateTensorValueFromMemory(data, [2, 16]);
var inputs = new Dictionary<string, OrtValue> { ["x"] = x };
using var runOptions = new RunOptions();
using IDisposableReadOnlyCollection<OrtValue> results = session.Run(runOptions, inputs, session.OutputNames);
 
float[] y = results[0].GetTensorDataAsSpan<float>().ToArray();

OneModel:

using var session = Session.Load(modelPath, new ModelOptions { IntraOpThreads = 4 });
 
using Tensor x = Tensor.FromSpan<float>(data, 2, 16);
using SessionOutputs results = session.Run(new Dictionary<string, Tensor> { ["x"] = x });
 
float[] y = results["y"].ToArray<float>();

For the common case of one array in and one array out, Model is shorter still:

Text

using var model = Model.Load("model.onnx");
float[] y = model.Run(data);

Mapping

ONNX Runtime	OneModel	Notes
`new InferenceSession(path \| bytes, options)`	`Session.Load(path \| stream, options)`	OneModel also reads its own format; the file kind is detected.
`SessionOptions`	`ModelOptions`	Renamed so both libraries can be used in one file without aliases. Immutable once built (`init` properties).
`SessionOptions.IntraOpNumThreads`	`ModelOptions.IntraOpThreads`	0 means one per core. There is no inter-op thread setting: one `Run` is one graph walk, and many threads can call `Run` at once.
`GraphOptimizationLevel`	`ModelOptions.OptimizationLevel`	`None`, `Basic`, `All`.
`ExecutionMode` (`ORT_SEQUENTIAL`/`ORT_PARALLEL`)	not needed	OneModel's `ArithmeticMode` (`Fast`/`Exact`) is a different thing: exact gives bit-identical results everywhere.
execution providers	none	CPU only.
`OrtValue` / `DenseTensor<T>` / `NamedOnnxValue`	`Tensor`	One tensor type for inputs, outputs and eager computation.
`OrtValue.CreateTensorValueFromMemory(data, shape)`	`Tensor.FromSpan<T>(data, dims)`	Copies the data. For no copy at all use the span overload of `Run`.
`session.Run(runOptions, inputs, outputNames)` returning `IDisposableReadOnlyCollection<OrtValue>`	`session.Run(inputs, outputNames)` returning `SessionOutputs`	Dispose the result. `outputNames` may be omitted to get every output.
parallel arrays of names and values	`session.Run(inputNames, inputs, outputNames, outputs)`	The allocation-free form.
`RunOptions` (tag, termination)	`CancellationToken` on load	A run is short and not cancellable midway.
`session.InputMetadata` / `OutputMetadata`	`session.Inputs` / `session.Outputs`	`TensorInfo`: name, element type, declared shape (`Dims`, symbolic extents have a name).
`OnnxRuntimeException`	`OneModelException` and subclasses	The kind is a property (`Kind`), not text in the message. File errors are the BCL's own.
`GetTensorDataAsSpan<T>()`	`Data<T>()`	Contiguous tensors only; `ToArray<T>()` copies.
`session.Dispose()` then `Run` crashes the process	`ObjectDisposedException`
`OrtEnv`, logging	none	No global environment.
custom operators	not supported yet	A model with an unknown operator fails at load, naming every one.
`SessionOptions.OptimizedModelFilePath`	`ModelFormat.Convert`	Saves the imported graph in the OneModel format.

Names to remember when two libraries meet: ONNX Runtime's SessionOptions and ExecutionMode do not exist here (they are ModelOptions and ArithmeticMode), so using Microsoft.ML.OnnxRuntime; and using RecursiveAI.OneModel.Graph; can sit in the same file.

Looking at the model

foreach (TensorInfo input in session.Inputs)
{
    Console.WriteLine($"{input.Name} {input.DType} rank {input.Rank}");   // names, element types and declared shapes
}

Things that behave differently

Ownership. A Tensor you get from SessionOutputs is owned by the container. Disposing a tensor is always safe and idempotent; see tensors.md.
Shapes. Declared shapes with symbolic extents are checked, and a model can be run at any size that fits them. Each distinct set of input shapes builds an execution plan once and caches it.
Determinism. The default Fast mode gives the same output on one machine whatever the thread count; Exact gives the same bits everywhere.
Errors for missing inputs are raised before anything runs, with the input names the model has.