M2M100 to ONNX

[sorgfresser]

I am currently trying to, with slight modifications, apply fastT5 to M2M100. While the conversion itself is working like a charm, I am getting a lot of

Ignore MatMul due to non constant B: /[MatMul_(insert int here)] e.G. Ignore MatMul due to non constant B: /[MatMul_2256]

errors during quantization. After some research I dug up multiple colabs that are simply ignoring this warning. Is any treatment necessary and are there known ways to handle it?

Apart from that: do you plan on expanding this repo to different models (for example in different branches) or do you rather want forks that provide support for other models?