ScatterND expects k (indices.shape[-1]) to be at most the ra

Q: What causes "ScatterND expects k (indices.shape[-1]) to be at most the rank of data" in huggingface/candle?

Correct the indices tensor so its last dim equals (or is below) the data rank

Error message

ScatterND expects k (indices.shape[-1]) to be at most the rank of data

What it means

In ScatterND, the last dimension k of the `indices` tensor determines the depth of indexing (number of data dimensions being sliced). ONNX requires k <= rank(data). candle-onnx checks this upfront because indices deeper than the data rank are meaningless and cannot be processed.

Source

Thrown at candle-onnx/src/eval.rs:2441

            "ScatterND" => {
                let data = get(&node.input[0])?;

                let indices = get(&node.input[1])?;
                let indices = indices.to_dtype(DType::I64)?;

                let updates = get(&node.input[2])?;

                let reduction = get_attr_opt::<str>(node, "reduction")?.unwrap_or("none");

                let indices_shape = indices.dims();
                let data_shape = data.dims();
                let _updates_shape = updates.dims();

                // Last dimension of indices represents the depth of indexing
                let k = indices_shape.last().unwrap().clone();

                if k > data.rank() {
                    bail!("ScatterND expects k (indices.shape[-1]) to be at most the rank of data");
                }

                let num_updates = indices_shape[..indices_shape.len() - 1]
                    .iter()
                    .product::<usize>();

                let flat_indices = if indices.rank() == 1 && k == 1 {
                    indices.unsqueeze(0)?
                } else {
                    indices.reshape((num_updates, k))?
                };

                // Calculate the shape of each update element
                let update_element_shape = if k < data_shape.len() {
                    data_shape[k..].to_vec()
                } else {
                    vec![]
                };

View on GitHub (pinned to d5fee525bf)

Solutions

Correct the indices tensor so its last dim equals (or is below) the data rank
Fix upstream reshape/expand ops producing the wrong indices shape
Re-export the model and compare ScatterND node inputs with netron

Example fix

# before: data rank 2, indices shape [4, 3]
# after
indices = indices[..., :2]  # slice trailing dim to match data rank

Defensive patterns

Strategy: validation

Validate before calling

let k = indices.dims()[indices.rank() - 1];
if k > data.rank() {
    return Err(format!("ScatterND k={} > data rank {}", k, data.rank()));
}

Type guard

fn scatternd_indices_valid(indices: &Tensor, data: &Tensor) -> bool {
    indices.rank() > 0 && indices.dims()[indices.rank() - 1] <= data.rank()
}

Try / catch

match eval(...) {
    Err(e) if e.contains("ScatterND expects k") => eprintln!("indices depth exceeds data rank"),
    other => other,
}

Prevention

Verify indices.shape[-1] equals the intended indexing depth
Run onnx.shape_inference on the model before inference
Compare ScatterND inputs in netron after graph edits

When it happens

Trigger: Evaluating a ScatterND node where indices.shape[-1] exceeds data.rank(), e.g. 3-D indices into a 2-D tensor, often caused by mismatched indices/data produced upstream in the graph.

Common situations: Models where indices were built for a different data rank than actually flows into ScatterND; exporter bugs; hand-edited graphs after removing a dimension.

Related errors

AI-assisted analysis of huggingface/candle@d5fee525bf (2026-09-02). Data as JSON: /api/errors/69cb286dd5222c4c. Report an issue: GitHub.