Oneliner

TinyML model inference with one-line code. Focus on no_std embedded targets.

For the verified host and embedded targets, see Target support.

Why Oneliner?

• One-line model deployment: Replace conversion scripts, native linking setup, tensor declarations, and dispatch glue with #[model(...)].
• Embedded-ready: The runtime supports no_std and is demonstrated with Ariel OS and Embassy on ARM Cortex-M targets.

Oneliner turns a model file into a callable Rust type with oneline code:

#[model("models/model.tflite")]
struct MyModel;

Quick Start

1. Install the host model compilation toolchain.

2. Add the crate to your Cargo.toml:

   [dependencies]
   oneliner = "0.2"

3. Bind and run a model:

   use oneliner::model;
   use oneliner::runtime::ModelInference;
   
   #[model("models/model.tflite")]
   struct MyModel;
   
   let mut model = MyModel::new();
   let mut input = MyModel::create_input_tensor();
   input.as_slice_mut().copy_from_slice(&input_data);
   
   let output = model.run(&input);
   let values = output.as_slice();

Oneliner generates the input and output tensor types directly from the model. The application does not need to repeat their data types or dimensions.

Supported Models

Oneliner accepts:

• TFLite
• ONNX
• PyTorch ExportedProgram (.pt2)
• TensorFlow SavedModel v2 directories
• MLIR accepted by IREE

See Model formats for per-format guides and Memory model for the owned/shared arena modes.

The #[model] attribute

The attribute takes the model path as its first positional argument, followed by optional named parameters:

#[model("models/model.tflite", backend = "iree", arena = "shared", format = "tflite")]
struct MyModel;

Parameter	Values	Default	Description
"<path>" (positional, required)	string	—	Model file or directory path, resolved relative to the application's Cargo.toml.
backend	"iree"	"iree"	Execution backend. Only IREE is currently available.
arena	"owned", "shared"	"owned"	Workspace memory mode, see Memory model.
format	"mlir", "onnx", "pytorch"/"pt2", "tensorflow"/"tf", "tflite"	auto-detected from the file extension	Explicit model format override. Required for TensorFlow SavedModel v2 directories (no extension).

The format is normally inferred from the extension (.mlir, .onnx, .pt2, .tflite); pass format explicitly to override it, which is mandatory for TensorFlow SavedModel directories. Duplicate or unknown options are rejected at compile time.

Profiling

Measure inference latency with the optional oneliner-profiler crate and get an automatic flash/RAM footprint report on every model build. See Profiling and footprint reporting and the benchmark numbers.

Examples

Each example is an independent Cargo project. Run its commands from the example directory with the Python environment activated.

Example	What it demonstrates	Active model
Desktop Std	The shortest end-to-end validation path on a standard host	Quantized MCUNet visual wake word
Ariel OS	no_std, Ariel OS as environment	Quantized LeNet5 and MCUNet
Embassy on Rasperry Pi Pico	Bare-metal RP2040, static input storage	Quantized LeNet5
Ariel OS + Profiler	no_std latency profiling with Profiler	Quantized LeNet5 and MCUNet
Embassy on Rasperry Pi Pico + Profiler	Bare-metal RP2040 latency profiling	Quantized LeNet5

Start with the desktop example to confirm the model toolchain, then move to the operating system or board example that matches your target.

Project Status

Oneliner is currently at version 0.2.0. The project focuses on making fixed-shape, single-input, single-output inference straightforward across desktop Rust and memory-constrained no_std targets.

The examples are intentionally small and explicit. They are designed to help you validate the toolchain, understand the memory trade-offs, and replace the bundled model with your own.

Acknowledgements

Oneliner evolved from ariel-ml: it keeps the mature IREE compiler for model compilation and optimization, while dropping the bulky IREE runtime that was difficult to integrate and compile with Rust. The interface design is inspired by microflow-rs. We thank all involved projects for their work.

License

Licensed under either of Apache License, Version 2.0 or MIT license at your option.

Unless you explicitly state otherwise, any contribution intentionally submitted for inclusion in the work by you shall be dual-licensed as above, without any additional terms or conditions.

Other languages: 简体中文