onnx files without Microsoft-contrib ONNX ops?
#4
by TigerHix - opened
The current onnx files use SimplifiedLayerNormalization and MatMulNBits that are from the com.microsoft domain, which are not standard ONNX operators. This makes it incompatible with e.g., Unity's inference engine which only supports native operators.
Would you consider creating a version without the Microsoft contrib operators? I would love to do this myself, but I don't know how to export the original repo tensors into the onnx files as listed here... If you can share a code snippet or some instructions, that'd be really appreciated.
Thank you!