is the current NNC able to compress an already quantized model?

[ColdCodeCool]

hello, as the title, my question is how can I use nnc to compress an already quantized model?

[phaase-hhi]

Generally, yes! Basically there are different ways, depending on how the model is represented. If "quantized" means, that you have a model that is represented as integer, NNCodec will detect the integer tensors and just compress the integer values losslessly.

For a model that is quantized but still in float precision it depends on how the model is quantized. If it is quantized uniformly, you could simply try to find a combination of qp_density and qp that yields a quantization step size that matches your quantization step size. Usually, this should result in a more or less lossless compression.

Another way would be to use the "codebook" method. Here, NNCodec will derive a codebook from each tensor, that holds all unqiue values within a tensor an assigns an index. However, it is important to mention, that the values within the codebook are quantized to integer values with a step size that is derived from the qp value. But if you choose a very small qp value (e.g. -75) this should be again more or less lossless.

BTW: Sorry for the late answer. I have been very busy lately.