Feature Request: ONNX Opset 21 Support

[aendk]

Feature request

Opset 21 is out for a while now, and it added int4 quantization back in march of last year.

When do you plan to support it? Are you currently blocked by the PyTorch dynamo onnx exporter which is still in beta?

@gedoensmax

Motivation

Opset 21 would expose cutting-edge features of the ONNX standard like more advanced quantization.
The mission statement of this library is to make production grade Inference available for everyone.
Adding better quantization matches with this.

Your contribution

Currently I do not have the bandwidth to help.