Output dtype of DequantizeLinear v19 is not aligned with its doc #5704
Description
Bug Report
The document description of DequantizeLinear v19 is:
| y (heterogeneous) - T2: N-D full precision output tensor. It has same shape as input ‘x’.
| T2 in ( tensor(bfloat16), tensor(float), tensor(float16) ): ‘x_scale’ determines the output type.
So the dtype of y should be the same as x_scale. However in the code it's fixed to FLOAT:
.TypeAndShapeInferenceFunction([](ONNX_NAMESPACE::InferenceContext& ctx) {
auto y_type = ctx.getOutputType(0);
// only float is supported
y_type->mutable_tensor_type()->set_elem_type(ONNX_NAMESPACE::TensorProto::FLOAT);
if (!hasInputShape(ctx, 0))
return;
auto& input_shape = getInputShape(ctx, 0);
updateOutputShape(ctx, 0, input_shape);
}));So if x_scale dtype is float16, onnx.shape_inference.infer_shapes will deduce a wrong output data type.