huggingface/candle · error

more pads than expected in conv2d {pads:?} {}

Error message

more pads than expected in conv2d {pads:?} {}

What it means

For rank-2 (2D convolution) weights, candle-onnx expects the Conv node's `pads` attribute to be absent (pads=0) or contain exactly one or two values; a pad list of any other length (e.g. 4 entries: begin/end per axis, or 3+) cannot be interpreted, so simple_eval_ bails with this message naming the node.

Source

Thrown at candle-onnx/src/eval.rs:933

                        xs.conv1d(ws, pads, strides, dilations, groups as usize)?
                    }
                    4 => {
                        let (pads, xs) = match pads {
                            None => (0, xs.clone()),
                            Some([p]) => (*p as usize, xs.clone()),
                            Some(&[p1, p2, p3, p4]) => {
                                let p1 = p1 as usize;
                                let p2 = p2 as usize;
                                let p3 = p3 as usize;
                                let p4 = p4 as usize;
                                if p1 != p2 || p1 != p3 || p1 != p4 {
                                    (0, xs.pad_with_zeros(2, p1, p3)?.pad_with_zeros(3, p2, p4)?)
                                } else {
                                    (p1, xs.clone())
                                }
                            }
                            Some(pads) => {
                                bail!("more pads than expected in conv2d {pads:?} {}", node.name)
                            }
                        };
                        let strides = match strides {
                            None => 1,
                            Some([p]) => *p as usize,
                            Some([p1, p2]) => {
                                if p1 != p2 {
                                    bail!(
                                        "strides have to be the same on both axis {pads:?} {}",
                                        node.name
                                    )
                                }
                                *p1 as usize
                            }
                            Some(s) => {
                                bail!("more strides than expected in conv2d {s:?} {}", node.name)
                            }
                        };

View on GitHub (pinned to d5fee525bf)

Solutions

  1. Rewrite the pads attribute in the model to the 2-value form [pad_h, pad_w] (requires symmetric padding) using onnx python tooling
  2. Use symmetric begin/end padding in the source model before export so the exporter emits compact pads
  3. Pre-pad the input tensor yourself, set pads to 0/absent, and re-export
  4. Patch candle-onnx to expand 4-element [b1,b2,e1,e2] pads into pad_with_zeros calls

Example fix

// before: pads: [1, 1, 1, 1]
// after:
attr = onnx.helper.make_attribute('pads', [1, 1])
node.attr.remove(next(a for a in node.attr if a.name == 'pads'))
node.attr.append(attr)
Defensive patterns

Strategy: validation

Validate before calling

for node in &model.graph.node {
    if node.op_type == "Conv" {
        if let Some(p) = node.attribute.iter().find(|a| a.name == "pads") {
            let n = p.ints.len();
            if n > 2 { panic!("node {}: conv2d pads arity {} unsupported (need 0-2 values)", node.name, n); }
        }
    }
}

Type guard

fn has_compact_pads(attr: &Attribute) -> bool {
    attr.name == "pads" && attr.ints.len() <= 2
}

Try / catch

match candle_onnx::simple_eval(&model, &inputs) {
    Err(e) if e.to_string().contains("more pads than expected in conv2d") => {
        eprintln!("rewrite pads attribute to [h, w]: {e}");
    }
    other => other?,
}

Prevention

When it happens

Trigger: Calling simple_eval on an ONNX model where a Conv node with 4D weights carries a `pads` attribute with more than two elements, such as pads=[1,1,1,1].

Common situations: Models exported from PyTorch/TensorFlow where the ONNX spec pads format is [x1_begin, x2_begin, x1_end, x2_end] (4 values) rather than the compact [h, w] form this evaluator expects; hand-authored or converted graphs.

Related errors

AI-assisted analysis of huggingface/candle@d5fee525bf (2026-09-02). Data as JSON: /api/errors/f88130c49804bee7. Report an issue: GitHub.