Model Restrictions

Supported Inference Format

Our file storage network can support models of any type and format. Currently, the OpenGradient network only supports inference for models in ONNX format. You can convert your custom models into ONNX format by using the ONNX converter. The converter supports formats described here.

NOTE

You can store and upload models in any format; however if you want to use it for on-chain inference, it must be ONNX.

Supported Model Types for ZK Inference

Using the EZKL library there are currently some limitations on what models can be supported for ZK Inference. The main constrictions are:

• The supported ONNX operator set is 9 - 18 inclusive for the latest EZKL version. This may require you to downgrade some operators in order run a ZKML inference.
• Inputs must be of numerical typing (float, int, etc.). Any string type input must be encoded into a numerical type in order to be processed correctly.
• Models cannot have a mix of static and dynamic (e.g. batched) shaped inputs. Multiple static shaped inputs, and singular batched inputs are accepted. The reccomendation for this scenario is to hard-code the constant values into your model beforehand.