What causes "Tensors are not broadcastable along dimension {}" in tracel-ai/burn?

Make both tensors the same rank and shape before cross, e.g. unsqueeze the single vector to [1,3] so it broadcasts.

Tensors are not broadcastable along dimension {}

Error message

Tensors are not broadcastable along dimension {}

What it means

Broadcast-check panic in the ndarray backend's batched cross product: along every dimension other than the batch dim (dim == 3's axis), the two operand shapes must be equal or one must be 1; otherwise they cannot be broadcast for the batched matmul and this panic fires. The failing input is a pair of tensors with incompatible non-batch dimensions.

Source

Thrown at crates/burn-ndarray/src/ops/matmul.rs:211

    let shape_rhs = rhs.shape();
    let ndims = shape_lhs.num_dims();

    // Broadcast the shapes except along dim
    let mut broadcast_shape = vec![0; ndims];
    for i in 0..ndims {
        if i == dim {
            broadcast_shape[i] = shape_lhs[i]; // already checked to be 3
        } else {
            let l = shape_lhs[i];
            let r = shape_rhs[i];
            if l == r {
                broadcast_shape[i] = l;
            } else if l == 1 {
                broadcast_shape[i] = r;
            } else if r == 1 {
                broadcast_shape[i] = l;
            } else {
                panic!("Tensors are not broadcastable along dimension {}", i);
            }
        }
    }

    // Broadcast lhs and rhs
    let lhs_broadcast = if shape_lhs == broadcast_shape.as_slice() {
        lhs
    } else {
        NdArrayOps::expand(lhs, Shape::from(broadcast_shape.clone()))
    };
    let rhs_broadcast = if shape_rhs == broadcast_shape.as_slice() {
        rhs
    } else {
        NdArrayOps::expand(rhs, Shape::from(broadcast_shape.clone()))
    };

    // Now, move dim to the last dimension
    let mut perm = (0..ndims).collect::<Vec<_>>();

View on GitHub (pinned to d16f7ba2ed)

Solutions

Make both tensors the same rank and shape before cross, e.g. unsqueeze the single vector to [1,3] so it broadcasts.
Reshape/expand one tensor so all leading dims match or equal 1.
Verify both inputs come from the same batched source.

Example fix

// before: a [3], b [4,3]
let c = a.cross(b, -1); // panic along dim 0
// after
let a2 = a.unsqueeze().expand([4, 3]); // or a.unsqueeze::<2>() broadcast
let c = a2.cross(b, -1);

Defensive patterns

Strategy: validation

Validate before calling

fn ensure_cross_shapes(a: &[usize], b: &[usize]) {
    assert_eq!(a.len(), b.len(), "cross requires same rank");
    for (l, r) in a.iter().zip(b) {
        assert!(l == r || *l == 1 || *r == 1, "cross dims {l} vs {r} not broadcastable");
    }
}

Prevention

Unsqueeze single vectors before cross with batched vectors.
Keep cross-product inputs from the same batched source.
Assert shapes equal before geometric ops.

When it happens

Trigger: Calling tensor.cross(other, dim) with tensors whose shapes differ in a leading dimension where both sizes are >1, e.g. [2,3] x [4,3].

Common situations: Cross products of vectors from different batch sizes; mixing single vector [3] with batched vectors [N,3] without reshaping.

Related errors

AI-assisted analysis of tracel-ai/burn@d16f7ba2ed (2026-09-05). Data as JSON: /api/errors/1e8a84a1ab5dee38. Report an issue: GitHub.