huggingface/candle · error

only exclusive == 0 is supported in CumSum

Error message

only exclusive == 0 is supported in CumSum

What it means

The CumSum node evaluation only supports the default inclusive, forward-mode cumulative sum. If the node's optional 'exclusive' attribute is set to 1 (exclusive cumsum), evaluation aborts with this error.

Source

Thrown at candle-onnx/src/eval.rs:1138

                        None => {
                            bail!("unsupported 'to' value {dt:?} for cast {}", node.name)
                        }
                    },
                    Err(_) => {
                        bail!("unsupported 'to' value {dt:?} for cast {}", node.name)
                    }
                };
                let output = input.to_dtype(dtype)?;
                values.insert(node.output[0].clone(), output);
            }
            // https://github.com/onnx/onnx/blob/main/docs/Operators.md#CumSum
            "CumSum" => {
                let exclusive = get_attr_opt::<i64>(node, "exclusive")?
                    .copied()
                    .unwrap_or(0);
                let reverse = get_attr_opt::<i64>(node, "reverse")?.copied().unwrap_or(0);
                if exclusive != 0 {
                    bail!("only exclusive == 0 is supported in CumSum")
                }
                if reverse != 0 {
                    bail!("only reverse == 0 is supported in CumSum")
                }
                let input = get(&node.input[0])?;
                let axis = to_vec0_flexible::<u32>(&get(&node.input[1])?.to_dtype(DType::U32)?)?;
                let output = input.cumsum(axis as usize)?;
                values.insert(node.output[0].clone(), output);
            }
            //  https://github.com/onnx/onnx/blob/main/docs/Operators.md#flatten
            "Flatten" => {
                let axis = get_attr_opt::<i64>(node, "axis")?.copied().unwrap_or(1) as usize;
                let input = get(&node.input[0])?;
                let first_part: usize = input.shape().dims().iter().take(axis).product();
                let end_index = input.shape().dims().iter().product::<usize>();
                let new_shape = (first_part, end_index / first_part);
                let output = input.reshape(new_shape)?;
                values.insert(node.output[0].clone(), output);

View on GitHub (pinned to d5fee525bf)

Solutions

1. Set the CumSum node's 'exclusive' attribute to 0 or remove it (exclusive=0 is the default)
2. Pre-compute the exclusive variant outside the graph: exclusive_cumsum(x) = inclusive_cumsum(x) - x
3. Rewrite the export so cumsum is inclusive (shift/add logic in source framework)
4. Patch candle-onnx to implement exclusive cumsum (subtract input from result)

Example fix

// before (python, onnx graph surgery)
node.attribute.append(onnx.helper.make_attribute('exclusive', 1))
// after
node.attribute.append(onnx.helper.make_attribute('exclusive', 0))

for node in &graph.node {
    if node.op_type == "CumSum" {
        let exclusive = get_attr_opt::<i64>(node, "exclusive").unwrap_or(None).copied().unwrap_or(0);
        assert_eq!(exclusive, 0, "CumSum {} uses exclusive=1", node.name);
    }
}

When it happens

Trigger: Evaluating a model with a CumSum node having attribute exclusive=1.

Common situations: Models exported from frameworks that use exclusive cumulative sums (e.g. certain sequence/probability models); handwritten ONNX graphs using the full CumSum spec.

Related errors

• unsupported 'mode' value {mode:?} for Pad node {:?}
• more dilations than expected in conv1d {s:?} {}
• cannot find 'value' attr in 'Constant' for {}
• attribute {} was of type TENSOR, but no tensor was found
• attribute {} of type TENSOR was an invalid data_type number

AI-assisted analysis of huggingface/candle@d5fee525bf (2026-09-02). Data as JSON: /api/errors/e4dae7915ac47fb7. Report an issue: GitHub.