huggingface/candle · error

unsupported 'value' data-type {dt:?} for {}

Error message

unsupported 'value' data-type {dt:?} for {}

What it means

During input validation in simple_eval_, the declared elem_type maps to a valid ONNX DataType but candle's dtype() returns None — meaning ONNX knows the element type but candle-onnx has no corresponding DType (or it is intentionally unsupported). The library refuses to build an input tensor of that type.

Source

Thrown at candle-onnx/src/eval.rs:280

        };
        let input_type = match &input_type.value {
            Some(input_type) => input_type,
            None => continue,
        };
        let tensor_type = match input_type {
            onnx::type_proto::Value::TensorType(tt) => tt,
            _ => continue,
        };

        let tensor = match values.get(&input.name) {
            None => bail!("missing input {}", input.name),
            Some(tensor) => tensor,
        };
        let dt = match DataType::try_from(tensor_type.elem_type) {
            Ok(dt) => match dtype(dt) {
                Some(dt) => dt,
                None => {
                    bail!("unsupported 'value' data-type {dt:?} for {}", input.name)
                }
            },
            type_ => bail!("unsupported input type {type_:?}"),
        };
        match &tensor_type.shape {
            None => continue,
            Some(shape) => {
                if shape.dim.len() != tensor.rank() {
                    bail!(
                        "unexpected rank for {}, got {:?}, expected {:?}",
                        input.name,
                        shape.dim,
                        tensor.shape()
                    )
                }
                for (idx, (d, &dim)) in shape.dim.iter().zip(tensor.dims().iter()).enumerate() {
                    match &d.value {
                        Some(onnx::tensor_shape_proto::dimension::Value::DimValue(v)) => {

View on GitHub (pinned to d5fee525bf)

Solutions

1. Cast the input tensor to a supported dtype (f32/i64) and re-run eval.
2. Re-export the model with a supported elem_type for its inputs.
3. Check the candle-onnx version's supported dtype list; upgrade if a newer version adds the type.
4. If the input is unused by the graph, remove it from the model's input list during export (non-optional inputs are validated).

Example fix

// before
let v = Tensor::from_vec(bf16_data, shape)?; // bfloat16 input
// after
let v = Tensor::from_vec(bf16_data.to_f32v(), shape)?; // cast to f32

fn input_dtypes_supported(model: &onnx::ModelProto) -> bool {
    model.graph.as_ref().map_or(true, |g| g.input.iter().all(|i| {
        match &i.r#type.value {
            Some(onnx::type_proto::Value::TensorType(tt)) =>
                candle_onnx::DataType::try_from(tt.elem_type)
                    .ok()
                    .and_then(candle_onnx::eval::dtype)
                    .is_some(),
            _ => true,
        }
    }))
}

fn has_candle_dtype(elem_type: i32) -> bool {
    use candle_onnx::DataType;
    DataType::try_from(elem_type).ok()
        .and_then(candle_onnx::eval::dtype)
        .is_some()
}

match simple_eval(&model, inputs) {
    Err(e) if e.to_string().contains("unsupported 'value' data-type") => {
        anyhow::bail!("convert inputs to a supported dtype (f32/i64) before eval")
    }
    r => r?,
}

When it happens

Trigger: Supplying inputs to simple_eval where a graph input is declared with an element type known to ONNX but unsupported by candle-onnx (e.g. float8, bfloat16 depending on version, string, complex types).

Common situations: Feeding bf16 tensors to a candle-onnx build lacking bf16 support; string inputs (tokenizers sometimes declare string inputs); models from exporters using newer type sets.

Related errors

• unsupported 'value' data-type {} for {name}
• attribute {} of type TENSOR was an invalid data_type number
• attribute {} of type TENSOR has an unsupported data_type {}
• unsupported 'value' data-type {dt:?} for {name}
• missing input {}

AI-assisted analysis of huggingface/candle@d5fee525bf (2026-09-02). Data as JSON: /api/errors/ac2006fb5b710f73. Report an issue: GitHub.