tracel-ai/burn · error

Unsupported dtype for `bool_from_data` {:?}

Error message

Unsupported dtype for `bool_from_data` {:?}

What it means

bool_from_data converts host TensorData into a GPU boolean tensor. The CubeCL backend only stores bools as U8 or U32 (BoolStore::U8/U32); any other dtype in the incoming data cannot be interpreted and triggers this unimplemented! panic.

Source

Thrown at crates/burn-cubecl/src/ops/bool_tensor.rs:52

    fn bool_zeros(shape: Shape, device: &Device<Self>, dtype: BoolDType) -> BoolTensor<Self> {
        numeric::zeros(device.clone(), shape, dtype.into())
    }

    fn bool_ones(shape: Shape, device: &Device<Self>, dtype: BoolDType) -> BoolTensor<Self> {
        numeric::ones(device.clone(), shape, dtype.into())
    }

    async fn bool_into_data(tensor: BoolTensor<Self>) -> Result<TensorData, ExecutionError> {
        super::into_data(tensor).await
    }

    fn bool_from_data(data: TensorData, device: &Device<Self>) -> BoolTensor<Self> {
        if !matches!(
            data.dtype,
            DType::Bool(BoolStore::U8) | DType::Bool(BoolStore::U32)
        ) {
            unimplemented!("Unsupported dtype for `bool_from_data` {:?}", data.dtype);
        }
        super::from_data(data, device)
    }

    fn bool_into_int(tensor: BoolTensor<Self>, out_dtype: IntDType) -> IntTensor<Self> {
        kernel::bool_cast(tensor, out_dtype.into())
    }

    fn bool_to_device(tensor: BoolTensor<Self>, device: &Device<Self>) -> BoolTensor<Self> {
        super::to_device(tensor, device)
    }

    fn bool_reshape(tensor: BoolTensor<Self>, shape: Shape) -> BoolTensor<Self> {
        super::reshape(tensor, shape)
    }

    fn bool_slice(tensor: BoolTensor<Self>, slices: &[Slice]) -> BoolTensor<Self> {
        // Check if all steps are 1

View on GitHub (pinned to d16f7ba2ed)

Solutions

1. Ensure TensorData.dtype is DType::Bool(BoolStore::U8) or BoolStore::U32 before creating the tensor
2. Convert data bytes to u8/u32 storage when building TensorData (e.g. cast Vec<bool> to Vec<u8>)
3. Re-export the tensor from a compatible path or use the backend's conversion APIs
4. Upgrade burn — conversion layers for native bool may be added

Example fix

// before
let data = TensorData::new(bools_vec, shape); // dtype Bool(Native)
// after
let bytes: Vec<u8> = bools_vec.iter().map(|&b| b as u8).collect();
let data = TensorData::new(bytes, shape).convert::<DTypeBoolU8>(); // bool u8 storage

fn bool_data_supported(data: &TensorData) -> bool {
    matches!(data.dtype, DType::Bool(BoolStore::U8) | DType::Bool(BoolStore::U32))
}

fn is_storable_bool(dtype: DType) -> bool {
    matches!(dtype, DType::Bool(BoolStore::U8 | BoolStore::U32))
}

// Panic-based; convert first:
if !bool_data_supported(&data) { let data = to_u8_bool_storage(data); }
let t = Tensor::<B, D>::from_data(data, &device);

When it happens

Trigger: Calling bool_from_data (directly or via TensorData-backed bool tensor creation, e.g. Tensor::from_data with bool data tagged with a wrong DType) with data.dtype not DType::Bool(BoolStore::U8) or BoolStore::U32 — e.g. DType::Bool(BoolStore::Native) or numeric dtypes.

Common situations: Serializing tensors on CPU (native bool storage) and loading on the CubeCL backend; constructing TensorData manually with an incorrect dtype; cross-backend tensor migration.

Related errors

• Unsupported dtype for `bool_from_data` {other:?}
• Unsupported dtype for `int_from_data`
• ctc_loss_backward: 2 * max_target_len + 1 = {} exceeds the k
• Should be float, got bool
• not supported for sorting operations

AI-assisted analysis of tracel-ai/burn@d16f7ba2ed (2026-09-05). Data as JSON: /api/errors/f2e1a3ed614ebbd1. Report an issue: GitHub.