tracel-ai/burn · error

int_cumsum: unsupported dtype {:?}

Error message

int_cumsum: unsupported dtype {:?}

What it means

int_cumsum dispatches to the generic cumsum implementation for each supported integer dtype (I64..U8). The panic arm is only reachable when the tensor's dtype is not one of the eight integer dtypes — i.e. a float or bool tensor was passed to the integer cumulative-sum path. It is a defensive guard indicating the wrong tensor kind reached this op.

Source

Thrown at crates/burn-flex/src/ops/int.rs:696

    fn int_prod_dim(tensor: IntTensor<Flex>, dim: usize) -> IntTensor<Flex> {
        crate::ops::reduce::prod_dim(tensor, dim)
    }

    fn int_mean_dim(tensor: IntTensor<Flex>, dim: usize) -> IntTensor<Flex> {
        crate::ops::reduce::mean_dim(tensor, dim)
    }

    fn int_cumsum(tensor: IntTensor<Flex>, dim: usize) -> IntTensor<Flex> {
        match tensor.dtype() {
            DType::I64 => crate::ops::cumulative::cumsum::<i64>(tensor, dim),
            DType::I32 => crate::ops::cumulative::cumsum::<i32>(tensor, dim),
            DType::I16 => crate::ops::cumulative::cumsum::<i16>(tensor, dim),
            DType::I8 => crate::ops::cumulative::cumsum::<i8>(tensor, dim),
            DType::U64 => crate::ops::cumulative::cumsum::<u64>(tensor, dim),
            DType::U32 => crate::ops::cumulative::cumsum::<u32>(tensor, dim),
            DType::U16 => crate::ops::cumulative::cumsum::<u16>(tensor, dim),
            DType::U8 => crate::ops::cumulative::cumsum::<u8>(tensor, dim),
            dt => panic!("int_cumsum: unsupported dtype {:?}", dt),
        }
    }

    fn int_cumprod(tensor: IntTensor<Flex>, dim: usize) -> IntTensor<Flex> {
        match tensor.dtype() {
            DType::I64 => crate::ops::cumulative::cumprod::<i64>(tensor, dim),
            DType::I32 => crate::ops::cumulative::cumprod::<i32>(tensor, dim),
            DType::I16 => crate::ops::cumulative::cumprod::<i16>(tensor, dim),
            DType::I8 => crate::ops::cumulative::cumprod::<i8>(tensor, dim),
            DType::U64 => crate::ops::cumulative::cumprod::<u64>(tensor, dim),
            DType::U32 => crate::ops::cumulative::cumprod::<u32>(tensor, dim),
            DType::U16 => crate::ops::cumulative::cumprod::<u16>(tensor, dim),
            DType::U8 => crate::ops::cumulative::cumprod::<u8>(tensor, dim),
            dt => panic!("int_cumprod: unsupported dtype {:?}", dt),
        }
    }

    fn int_cummin(tensor: IntTensor<Flex>, dim: usize) -> IntTensor<Flex> {

View on GitHub (pinned to d16f7ba2ed)

Solutions

1. Verify tensor.dtype() is an integer type before calling cumsum; cast to i64/i32 first if needed.
2. Use the float tensor's cumsum/sum_dim path for float tensors instead of the int one.
3. Fix upstream ops (e.g. division, casts) that changed the tensor's dtype unexpectedly.
4. Assert the dtype at pipeline boundaries to fail early with a clearer message.

Example fix

// before
let out = int_tensor.cumsum(dim); // panics if dtype is F32
// after
assert!(matches!(int_tensor.dtype(), DType::I64 | DType::I32 | _ if int_tensor.dtype().is_int()));
let out = int_tensor.cast::<i64>().cumsum(dim);

assert!(t.dtype().is_int(), "cumsum requires an int tensor, got {:?}", t.dtype());

fn is_int_tensor(t: &FlexTensor) -> bool { matches!(t.dtype(), DType::I64 | DType::I32 | DType::I16 | DType::I8 | DType::U64 | DType::U32 | DType::U16 | DType::U8) }

When it happens

Trigger: Calling Tensor::cumsum (or int_cumsum via the backend) on a Flex tensor whose dtype is F32/F64/F16/BF16/Bool instead of an integer dtype.

Common situations: Applying cumsum to a float tensor expecting PyTorch-like generic behavior; dtype drift after an operation (e.g. division produced floats); loading data whose dtype was inferred as float but treated as int downstream.

Related errors

• float_cumsum: unsupported dtype {:?}
• burn-flex does not support Bool(U32) storage (only Native an
• compare_int: unsupported dtype {:?}
• compare_int_elem: unsupported dtype {:?}
• any_float: unsupported dtype {:?}

AI-assisted analysis of tracel-ai/burn@d16f7ba2ed (2026-09-05). Data as JSON: /api/errors/93e64cc9cf17a381. Report an issue: GitHub.