Why is Q4_0_4_4 support removed?

[kjh2159]

I could load a Q4_0_4_4 quantized model before. However, after pulling git repository, I couldn't load Q4_0_4_4 models and cli throws a message, "Q4_0_4_4 is removed, use Q4 instead".

Does anyone know about the reason why Q4_0_4_4 support is removed?

[DavidZyy]

I am also curious about this question. In the previous version that the codes is not deleted, I saw the comment about these codes are "activation aware quantization", but it seems have some differences with the method that this paper described.

[ExtReMLapin]

Next time you go find the code part where it say it's removed and look at the git blame

You'll find this PR with the reasons

[DavidZyy]

Thank you. It is very helpful.

[kjh2159]

Thank you for answering. I figured out the reason why.