huggingface/candle · error

Unsupported nearest_mode for resize: {}

Error message

Unsupported nearest_mode for resize: {}

What it means

Even in nearest mode, candle-onnx only supports nearest_mode="floor" for the Resize operator. Other ONNX nearest_mode values (`round_prefer_floor`, the default, `round_prefer_ceil`, `ceil`) are rejected with this error. Note the ONNX default is round_prefer_floor, so models that never set the attribute explicitly will still hit this.

Source

Thrown at candle-onnx/src/eval.rs:2354

                        .collect::<Vec<_>>(),
                    (None, None) => bail!("Either scales or sizes should be present"),
                };

                let coordinate_transformation_mode =
                    get_attr_opt::<str>(node, "coordinate_transformation_mode")?
                        .unwrap_or("half_pixel");
                // Interpolation mode: nearest, linear, or cubic.
                let mode = get_attr_opt::<str>(node, "mode")?.unwrap_or("nearest");
                // How to determine the "nearest" pixel in nearest interpolation mode.
                let nearest_mode =
                    get_attr_opt::<str>(node, "nearest_mode")?.unwrap_or("round_prefer_floor");

                if mode != "nearest" {
                    bail!("Unsupported resize mode: {}", mode);
                }

                if nearest_mode != "floor" {
                    bail!("Unsupported nearest_mode for resize: {}", nearest_mode);
                }

                if coordinate_transformation_mode != "asymmetric" {
                    bail!(
                        "Unsupported coordinate_transformation_mode for resize: {}",
                        coordinate_transformation_mode
                    );
                }

                let h = output_dims[2];
                let w = output_dims[3];
                let output = input.upsample_nearest2d(h, w)?;

                values.insert(node.output[0].clone(), output);
            }
            "Trilu" => {
                let input = get(&node.input[0])?;

View on GitHub (pinned to d5fee525bf)

Solutions

1. Patch the graph to set nearest_mode="floor" on Resize nodes (results may differ slightly)
2. Set nearest_mode='floor' at export time (e.g. onnx-surgement or onnx.helper edit)
3. Implement the other nearest modes in candle-onnx
4. Pre-compute the resize outside the graph

Example fix

// before
// attribute nearest_mode absent (defaults to "round_prefer_floor")
// after
node.attribute.push(onnx_attr("nearest_mode", "floor"));

if node.op_type == "Resize" {
    let nm = get_attr_opt::<String>(node, "nearest_mode")?.unwrap_or_else(|| "round_prefer_floor".into());
    if nm != "floor" { return Err(format!("nearest_mode {} unsupported", nm)); }
}

fn uses_floor_nearest(node: &Node) -> bool {
    get_attr_opt::<String>(node, "nearest_mode").ok().flatten().map_or(false, |m| m == "floor")
}

match eval(...) {
    Err(e) if e.contains("Unsupported nearest_mode") => rewrite_nearest_mode_to_floor(model).and_then(eval),
    other => other,
}

When it happens

Trigger: Evaluating a Resize node with mode="nearest" but nearest_mode set to (or defaulting to) anything other than "floor".

Common situations: Models exported with default nearest_mode (round_prefer_floor) since the exporter did not set it explicitly; YOLO/SSD detection heads using round-prefer-floor upsampling.

Related errors

• Unsupported resize mode: {}
• Unsupported coordinate_transformation_mode for resize: {}
• Either scales or sizes should be present
• 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/930e4b34e8922ef4. Report an issue: GitHub.