tracel-ai/burn · error

ndarray scatter_nd requires contiguous indices

Error message

ndarray scatter_nd requires contiguous indices

What it means

In the same scatter_nd kernel, the indices tensor is flattened via as_slice, which returns None for non-contiguous storage; the expect panics with 'ndarray scatter_nd requires contiguous indices'. Like the data check, this reflects the backend kernel's contiguous-memory-only implementation.

Source

Thrown at crates/burn-ndarray/src/ops/base.rs:325

        let data_shape: Vec<usize> = data.shape().to_vec();
        let idx_shape: Vec<usize> = indices.shape().to_vec();
        let m = idx_shape.len();
        let k = idx_shape[m - 1];

        // Number of index tuples = product of batch dims (first M-1 dims of indices)
        let num_indices: usize = idx_shape[..m - 1].iter().product();
        // Size of each slice to scatter = product of data.shape[K..]
        let slice_size: usize = data_shape[k..].iter().product();

        let mut output = data.into_owned();
        let output_flat = output
            .as_slice_mut()
            .expect("ndarray scatter_nd requires contiguous data");

        // Flatten indices to [num_indices, K]
        let idx_flat = indices
            .as_slice()
            .expect("ndarray scatter_nd requires contiguous indices");

        // Flatten values to [num_indices, slice_size]
        let val_flat = values
            .as_slice()
            .expect("ndarray scatter_nd requires contiguous values");

        let strides: Vec<usize> = {
            let mut s = vec![0usize; k];
            if k > 0 {
                s[k - 1] = slice_size;
                for i in (0..k - 1).rev() {
                    s[i] = s[i + 1] * data_shape[i + 1];
                }
            }
            s
        };

        for n in 0..num_indices {

View on GitHub (pinned to d16f7ba2ed)

Solutions

1. Copy the indices into a new owned contiguous tensor before calling scatter_nd.
2. Construct index tensors directly (from vec/IntNdArrayElement) instead of slicing existing ones.
3. Report as a bug if produced through burn's public high-level API, which should maintain contiguity.

Example fix

// before
let idx = big_index_tensor.slice(s![.., 0]);
let out = data.scatter(idx, values);
// after
let idx_owned = idx.to_data().convert::<IntNdArrayElement>();
let idx = NdArrayTensor::new(idx_owned.into_ndarray()); // contiguous copy
let out = data.scatter(idx, values);

// Copy indices into a fresh contiguous tensor before scatter
let idx_owned = indices.to_data().convert::<IntNdArrayElement>();
let indices = NdArrayTensor::new(idx_owned.into_ndarray());

When it happens

Trigger: Calling scatter_nd where the indices tensor is a non-contiguous view (e.g. produced by slicing/indexing another tensor) on the ndarray backend.

Common situations: Building index tensors by slicing a larger index buffer; graph-captured models passing strided index intermediates; reusing views after permute operations.

Related errors

• ndarray scatter_nd requires contiguous data
• ndarray gather_nd requires contiguous indices
• Requires autodiff tensor.
• an enabled float tensor must use an autodiff primitive
• Should be float, got int

AI-assisted analysis of tracel-ai/burn@d16f7ba2ed (2026-09-05). Data as JSON: /api/errors/30df85c207c4c785. Report an issue: GitHub.