Obtain quantized weights for entropy coding

[aegroto]

Hello everyone! I have just began to use NNI in my projects and it seems awesome to compress trained models out-of-the-box. However, after following tutorials on QAT I still struggle to find out how to save the quantized weights in a PyTorch state dict that uses less than 32-bits.

My purpose is to entropy code the quantized weight to obtain dimensionality reduction on the original model and I have been able to find that ModelSpeedup is supposed to do so for pruned models: https://nni.readthedocs.io/zh/latest/tutorials/pruning_speedup.html. However, I cannot find an analogous object for quantized weight tensors or any similar references in the documentation.