tracel-ai/burn · error

Can't add a new input that is already used in an index opera

Error message

Can't add a new input that is already used in an index operation

What it means

Fusion-engine resource conflict: a tensor ID already registered as an indexed resource (used as an index operand in a gather/scatter-like op) is being re-registered as a new fusion input via `input_unhandled`. Fusion compilation cannot represent a tensor that is simultaneously an index source and a plain data input, so the panic fires; the offending input is the tensor whose ID appears in `resources.indexed`.

Source

Thrown at crates/burn-cubecl-fusion/src/engine/trace/fuser.rs:189

    pub fn input_unhandled(&mut self, tensor: &TensorIr) -> FuseArg {
        if self.resources.indexed.contains_key(&tensor.id) {
            panic!("Can't add a new input that is already used in an index operation");
        }

        self.resources.outputs.update(tensor);

        let precision = tensor.dtype.into();
        let new_input = self.resources.inputs.insert(precision, tensor.clone());
        let arg = FuseArg::Input(new_input, precision, LayoutInfo::Unknown);

        self.resources.inputs_unhandled.push(tensor.id);
        arg
    }

    /// Register an input tensor.
    pub fn input_quantized_unhandled(&mut self, tensor: &TensorIr) -> Option<(FuseArg, FuseArg)> {
        if self.resources.indexed.contains_key(&tensor.id) {
            panic!("Can't add a new input that is already used in an index operation");
        }

        let (precision, precision_scales) = match tensor.dtype {
            DType::QFloat(scheme) => (
                FuseType::from_quant_scheme(scheme)?,
                FuseType::from_scale_dtype(scheme.scale_dtype())?,
            ),
            _ => return None,
        };

        self.resources.outputs.update(tensor);

        let (new_input, q_index) = self.resources.inputs.insert_quant(tensor.clone());
        let input = FuseArg::Input(new_input, precision, LayoutInfo::Unknown);
        let scales = FuseArg::Input(q_index, precision_scales, LayoutInfo::Unknown);

        self.resources.inputs_unhandled.push(tensor.id);
        Some((input, scales))

View on GitHub (pinned to d16f7ba2ed)

Solutions

1. Materialize the index tensor outside the fused kernel (clone or detach it into a separate operation) so it is not registered twice.
2. Check the trace builder so a tensor previously used for indexing is routed through the indexed-input path, not `input_unhandled`.
3. Update the fusion engine to reuse the existing indexed argument instead of inserting a new input for that ID.

When it happens

Trigger: Thrown at crates/burn-cubecl-fusion/src/engine/trace/fuser.rs:189 when the library encounters an invalid state.

Common situations: See trigger scenarios.

AI-assisted analysis of tracel-ai/burn@d16f7ba2ed (2026-09-05). Data as JSON: /api/errors/f26fc62e6f751e73. Report an issue: GitHub.