Module onnx 

ONNX inference via the pure-Rust tract engine (onnx feature).

The model is loaded from a byte buffer, never a path, so the identical code path runs natively (bytes from the super::download cache) and in WASM (bytes fetched by JavaScript and handed back in). All tensors are f32 (NCDHW for the volumetric nets); callers convert to/from the crate’s f64 volumes and handle model-specific normalization and padding.

With the parallel feature, each inference spreads tract’s matrix kernels over a thread pool: a dedicated one natively, and rayon’s global pool on WASM (the one wasm_bindgen_rayon::init_thread_pool sets up, once the host reports it via [onnx::set_wasm_threads_available]). Calls from inside a rayon worker — the tiled drivers, which already run one tile per thread — stay on the calling thread.

On wasm32 the build must enable SIMD128 (-C target-feature=+simd128): since tract 0.23, tract-linalg registers its matmul kernels on wasm only under that target feature, so a build without it compiles fine and then fails on the first convolution at runtime with No matmul found. The guard below turns that into a build error instead.

use qsm_core::models::onnx::{OnnxModel, Tensor};
let model = OnnxModel::load(model_bytes)?;
let out = model.run_single(&Tensor::new(vec![1, 1, d, h, w], field))?;

Structs

OnnxModel

A parsed ONNX model, ready to run at any spatial size.

OnnxPlan

A compiled execution plan for fixed input shapes, built by OnnxModel::plan_for. Running it skips graph optimization, so reusing one plan across many equal-shaped inputs (tiled inference) amortizes the optimizer to a single up-front cost.

Tensor

A dense f32 tensor: row-major data interpreted with shape.

Enums

OnnxError

Error from loading or running an ONNX model.

Functions

run_threaded 🔒

Run one inference with tract’s matrix kernels spread over a thread pool when the parallel feature is on (see tract_executor). Calls made from inside a rayon worker — the tiled drivers, which already run one tile per thread — stay on the calling thread.

tract_executor 🔒

The executor run_threaded installs, or None to stay on the calling thread.