What causes "unsupported op_type {op_type} for op {node:?}" in huggingface/candle?

Check the listed op_type against candle-onnx's supported ops and implement it in eval.rs

unsupported op_type {op_type} for op {node:?}

Error message

unsupported op_type {op_type} for op {node:?}

What it means

simple_eval_ dispatches on node op_type inside a match; the final arm catches any operator type the evaluator does not implement and bails with this message, including the node debug dump. It means the ONNX model uses an operator candle-onnx has no eval support for.

Source

Thrown at candle-onnx/src/eval.rs:2567

                                flat_output = flat_output.slice_scatter(
                                    &update_slice.unsqueeze(0)?,
                                    0,
                                    flat_idx,
                                )?;
                            } else {
                                flat_output =
                                    flat_output.slice_scatter(&update_slice, 0, flat_idx)?;
                            }
                        }
                    }
                }

                // Reshape flat output back to original shape
                output = flat_output.reshape(data_shape.to_vec())?;

                values.insert(node.output[0].clone(), output);
            }
            op_type => bail!("unsupported op_type {op_type} for op {node:?}"),
        }
    }
    graph
        .output
        .iter()
        .map(|output| match values.remove(&output.name) {
            None => bail!("cannot find output {}", output.name),
            Some(value) => Ok((output.name.clone(), value)),
        })
        .collect()
}

fn broadcast_shape(shape_a: &[usize], shape_b: &[usize]) -> Result<Vec<usize>> {
    let (longest, shortest) = if shape_a.len() > shape_b.len() {
        (shape_a, shape_b)
    } else {
        (shape_b, shape_a)
    };

View on GitHub (pinned to d5fee525bf)

Solutions

Check the listed op_type against candle-onnx's supported ops and implement it in eval.rs
Simplify/export the model avoiding unsupported ops (opset downgrade, operator replacement via onnx-surgement)
Run inference with onnxruntime for models with unsupported ops
Contribute/patch the missing op and rebuild candle-onnx

Example fix

// eval.rs
op_type => bail!("unsupported op_type {op_type} for op {node:?}"),
// after: add an arm
"Clip" => { /* implement clip */ }

Defensive patterns

Strategy: try-catch

Validate before calling

const SUPPORTED: &[&str] = &["Add","Mul","Resize","Trilu","ScatterND" /* ... */];
for node in &graph.node {
    if !SUPPORTED.contains(&node.op_type.as_str()) {
        return Err(format!("unsupported op: {}", node.op_type));
    }
}

Type guard

fn all_ops_supported(graph: &Graph) -> bool {
    graph.node.iter().all(|n| SUPPORTED.contains(&n.op_type.as_str()))
}

Try / catch

match eval(model, inputs) {
    Err(e) if e.starts_with("unsupported op_type") => run_with_onnxruntime(model, inputs),
    other => other,
}

Prevention

Audit model opsets against candle-onnx's supported ops list before inference
Avoid control-flow ops (If/Loop/Scan) when targeting candle-onnx
Keep the supported-op list in CI as a lint step

When it happens

Trigger: Running any model containing an op_type not covered by the match arms in simple_eval_ (e.g. newer or less common ONNX ops like RandomNormalLike, Loop, If, custom domains).

Common situations: Using models with control-flow ops (Loop/If/Scan) or ops added in newer opsets than the evaluator supports; custom-operator domains from framework-specific exporters.

Related errors

AI-assisted analysis of huggingface/candle@d5fee525bf (2026-09-02). Data as JSON: /api/errors/e150492964eb0c64. Report an issue: GitHub.