GRU Quantization

[Rikorose]

So the current tooling state for quantization of recurrent layers is not perfect. Many frameworks (pytorch, onnx, tf light) only support

• LSTMs, no GRU
• dynamic quantization (for LSTMs), no static quantization.

Since GRU quantization has been done before, has anyone an idea how a practical approach could be eventually enabling an export to tract?