tracel-ai/burn · error

float_scatter with {other:?} update is not implemented

Error message

float_scatter with {other:?} update is not implemented

What it means

burn-flex's `float_scatter` matches on the dtype category of the `value` (update) tensor. Float dtypes (F32/BF16, etc.) are supported, but any non-float update dtype falls into the `other` arm and panics with `unimplemented!("float_scatter with {other:?} update is not implemented")`.

Source

Thrown at crates/burn-flex/src/ops/float.rs:338

                }
                _ => panic!("float_scatter: unsupported dtype {:?}", tensor.dtype()),
            },
            burn_backend::tensor::IndexingUpdateOp::Mul => match tensor.dtype() {
                DType::F32 => {
                    crate::ops::gather_scatter::scatter_mul::<f32>(tensor, dim, indices, value)
                }
                DType::F64 => {
                    crate::ops::gather_scatter::scatter_mul::<f64>(tensor, dim, indices, value)
                }
                DType::F16 => {
                    crate::ops::gather_scatter::scatter_mul::<f16>(tensor, dim, indices, value)
                }
                DType::BF16 => {
                    crate::ops::gather_scatter::scatter_mul::<bf16>(tensor, dim, indices, value)
                }
                _ => panic!("float_scatter: unsupported dtype {:?}", tensor.dtype()),
            },
            other => unimplemented!("float_scatter with {other:?} update is not implemented"),
        }
    }

    fn float_scatter_nd(
        data: FloatTensor<Flex>,
        indices: IntTensor<Flex>,
        values: FloatTensor<Flex>,
        reduction: burn_backend::tensor::IndexingUpdateOp,
    ) -> FloatTensor<Flex> {
        match data.dtype() {
            DType::F32 => {
                crate::ops::gather_scatter::scatter_nd::<f32>(data, indices, values, reduction)
            }
            DType::F64 => {
                crate::ops::gather_scatter::scatter_nd::<f64>(data, indices, values, reduction)
            }
            DType::F16 => {
                crate::ops::gather_scatter::scatter_nd::<f16>(data, indices, values, reduction)

View on GitHub (pinned to d16f7ba2ed)

Solutions

1. Cast the update tensor to the data tensor's float dtype first: `values.cast(DType::F32)` (or `.float()` on the Tensor API)
2. Ensure scatter values come from float ops, not int/bool ops
3. Check the dtype of `values` at the call site before scattering
4. Use `select_assign`/`scatter_nd` only with dtype-matched tensors

Example fix

// before
let updated = tensor.scatter(dim, indices, indices_i32);
// after
let updated = tensor.scatter(dim, indices, indices_i32.cast(DType::F32));

assert_eq!(values.dtype(), tensor.dtype(), "float_scatter update dtype must match data tensor dtype");
debug_assert!(matches!(tensor.dtype(), DType::F32 | DType::F64 | DType::BF16 | DType::F16));

fn is_float_update(t: &TensorData) -> bool {
    matches!(t.dtype, DType::F32 | DType::F64 | DType::BF16 | DType::F16)
}

// panics are not catchable in Rust; guard before call
if !is_float_update(&values_data) {
    values = values.cast(DType::F32);
}
let out = tensor.scatter(dim, indices, values);

When it happens

Trigger: Calling `Tensor::scatter` (or `float_scatter`) on a float tensor with an integer/bool `values` tensor as the update, instead of a float tensor of the same dtype.

Common situations: Passing an index/integer tensor as the update by mistake; results of an `arange` or comparison reused directly as scatter values; dtype mismatches after refactors.

Related errors

• int_scatter: unsupported dtype {:?}
• int_scatter with {other:?} update is not implemented
• float_scatter: unsupported dtype {:?}
• int_gather: unsupported dtype {:?}
• int_scatter_nd: unsupported dtype {:?}

AI-assisted analysis of tracel-ai/burn@d16f7ba2ed (2026-09-05). Data as JSON: /api/errors/711f4950d29e263c. Report an issue: GitHub.