huggingface/candle · error

Slice node is invalid, expected 3-5 inputs, got {}: {:?}

Error message

Slice node is invalid, expected 3-5 inputs, got {}: {:?}

What it means

The Slice node in ONNX takes data, starts, ends and optional axes and steps (3-5 inputs). Candle-onnx enforces this arity and bails for any other count (fewer than 3 or more than 5), since it cannot resolve which tensors are starts/ends/axes/steps.

Source

Thrown at candle-onnx/src/eval.rs:1272

                match node.input.len() {
                    3 => {
                        let len = starts.dims()[0];
                        default_axes = Some(Tensor::arange(0, len as i64, starts.device())?);
                        axes = default_axes.as_ref().unwrap();
                        default_steps = Some(Tensor::ones((len,), DType::I64, starts.device())?);
                        steps = default_steps.as_ref().unwrap();
                    }
                    4 => {
                        let len = starts.dims()[0];
                        axes = get(&node.input[3])?;
                        default_steps = Some(Tensor::ones((len,), DType::I64, starts.device())?);
                        steps = default_steps.as_ref().unwrap();
                    }
                    5 => {
                        steps = get(&node.input[4])?;
                        axes = get(&node.input[3])?;
                    }
                    _ => bail!(
                        "Slice node is invalid, expected 3-5 inputs, got {}: {:?}",
                        node.input.len(),
                        node
                    ),
                }

                let mut out = data.clone();
                for (i, axis) in axes.to_vec1::<i64>()?.into_iter().enumerate() {
                    // All negative elements of axes are made non-negative by
                    // adding r to them, where r = rank(input).
                    let axis = if axis < 0 {
                        axis + data.rank() as i64
                    } else {
                        axis
                    } as usize;

                    let data_dim = data.dims()[axis] as i64;
                    let mut s = to_scalar_flexible::<i64>(&starts.get(i)?)?;

View on GitHub (pinned to d5fee525bf)

Solutions

1. Re-export the model with opset >= 10 so Slice uses inputs (data, starts, ends[, axes][, steps])
2. Convert attribute-based Slice to input-based Slice via onnx version converter to opset 10+
3. Supply all required inputs: starts and ends (and axes/steps as needed) as 1-D i64 constants
4. Fix graph edits that dropped required Slice inputs

Example fix

// before (opset 1, attribute-based)
Slice(data) with attrs starts/ends
// after (opset 10+)
Slice(data, starts, ends, axes)  # 3-5 inputs

for node in &graph.node {
    if node.op_type == "Slice" {
        let n = node.input.len();
        assert!((3..=5).contains(&n), "Slice {} has {} inputs", node.name, n);
    }
}

fn slice_arity_ok(node: &NodeProto) -> bool {
    (3..=5).contains(&node.input.len())
}

When it happens

Trigger: Evaluating a Slice node with 0-2 inputs (e.g. old opset-1 Slice using attributes) or 6+ inputs, i.e. any count outside 3..=5.

Common situations: Very old models exporting Slice as an attribute-based op (opset < 10) instead of input-based; an optimizer removing 'axes'/'steps' inputs improperly; hand-written graphs missing starts/ends.

Related errors

• unsupported number of inputs {} for Pad node {:?}, expected
• attribute {} was of type TENSOR, but no tensor was found
• attribute {} of type TENSOR was an invalid data_type number
• attribute {} of type TENSOR has an unsupported data_type {}
• attribute {} of type TENSOR has a negative dimension, which

AI-assisted analysis of huggingface/candle@d5fee525bf (2026-09-02). Data as JSON: /api/errors/7d0398bbe658e1ad. Report an issue: GitHub.