Robert Knight

Build rten-simd docs for Wasm

2

The [rten-simd documentation](https://docs.rs/rten-simd) is missing docs for architecture-specific types except for x64. The [Isa page](https://docs.rs/rten-simd/latest/rten_simd/isa/index.html) for example is missing the structs for Isa types used on Arm and WASM. I...

`Model::total_params` undercounts parameters when model uses 4-bit weights

1

The parameter count reported when running `rten model.onnx` is smaller than it should be when a model uses 4-bit weights. This is because `Model::total_params` returns the total number of tensor...

This is a tracking issue for various improvements that will make it possible to evolve RTen's functionality without breaking API changes. - [x] Change `LoadError` from an enum to a...

Remove internal operator implementation functions from public API

1

Internal operator implementations are currently exposed in the public API of the rten crate in two ways: 1. Via extension traits for the `Tensor` type 2. Via standalone functions in...

Model loading/conversion implements some fallbacks for unsupported features. These fallbacks may impact accuracy (but often do not). When converting a model to rten format, the converter prints a warning, but...

In ONNX Runtime, some operators treat single-element vectors and scalars as interchangeable. DequantizeLinear for example accepts `x_zero_point=[x], x_scale=y` even though the spec says that `x_zero_point` and `x_scale` should have the...

Optimize int8 kernels for symmetric quantization of weights

1

For each output element the int8 GEMM kernels compute: ``` C += dot(a, b) - sum(a) * b_zero_point - sum(b) * a_zero_point + k * a_zero_point * b_zero_point ``` Where...

The comments for [`GraphOptimizer::optimize`](https://github.com/robertknight/rten/blob/c756ff56aa465e645b982c69e69af8f4f62fe9a8/src/optimize.rs#L329) state that input IDs and output IDs are preserved, but transformations are inconsistent about whether they ensure the _IDs_ of output nodes are preserved, or only...

STFT with 2D signal input

1

The [STFT documentation](https://onnx.ai/onnx/operators/onnx__STFT.html) states that the shape of the `signal` input must be 3D: > Input tensor representing a real or complex valued signal. For real input, the following shape...

Internal server error in h when fetching public annotation

3

Visiting https://hypothes.is/a/2ZcSQKo7EfCeAUP1uxAeig while not logged in shows a "this annotation is not available" message, even though the annotation is public. Inspecting the browser console shows that a 500 error occurred...