RUNTIME_EXCEPTION occurs when running a per-channel quantized model in QDQ format

[feihongfiona]

Describe the issue

I have a per-channel quantized model in QDQ format. Running the model gets the following terminal interrupt :

onnxruntime.capi.onnxruntime_pybind11_state.RuntimeException: [ONNXRuntimeError] : 6 : RUNTIME_EXCEPTION : Non-zero status code returned while running DequantizeLinear node. Name:'W_DequantizeLinear/duplicated' Status Message: /private/var/folders/nz/j6p8yfhx1mv_0grj5xl4650h0000gp/T/abs_a92wmqsrpo/croot/onnxruntime_1689679134480/work/onnxruntime/core/providers/cpu/quantization/quantize_linear.cc:36 void onnxruntime::PrepareForQDQ(const onnxruntime::TensorShape &, const onnxruntime::Tensor &, const onnxruntime::Tensor *, int64_t, int64_t &, int64_t &, int64_t &) scale.Shape().NumDimensions() == 1 && scale.Shape()[0] == broadcast_dim was false. scale must be 1D tensor with size 525

I trace the source code bellow:

template <typename T>
class DequantizeLinear final : public OpKernel {
 public:
  DequantizeLinear(const OpKernelInfo& info) : OpKernel(info) {
    if (!info.GetAttr<int64_t>("axis", &axis_).IsOK()) {
      axis_ = 1;   
    }
  }

  Status Compute(OpKernelContext* context) const override;

 private:
  int64_t axis_;
};

I think the axis_ should default to 0

To reproduce

Conv_layer_quant_U8U8_perch_sym01_QDQ.onnx.zip

Urgency

No response

Platform

Mac

OS Version

13.5

ONNX Runtime Installation

Released Package

ONNX Runtime Version or Commit ID

1.15.1

ONNX Runtime API

Python

Architecture

X64

Execution Provider

Default CPU

Execution Provider Library Version

No response

[wschin]

Default is 1 per ONNX standard, so onnxruntime shouldn't change its behavior. If you think this behavior is wrong, please file an issue to ONNX. If your model is generated by a quantization tool, please file a bug to that tool since it generates an invalid model. Or, as a workaround, you can edit your model file and fix that attribute.