[Feature]: Enable SVD independenty for model and TE

[alerikaisattera]

Feature description

At the moment, SVD can only be applied to both the model and TE or neither of them, in contrast with other quantization settings. There should be a way to enable SVD for them independently

Version Platform Description

No response

[vladmandic]

other than "I want to test", what is the use case?
I'm not against it, I just want to avoid more items to settings that will complicate things for normal users without having some benefit.

[alerikaisattera]

Sometimes the TE can be quantized just fine at int8 but the model requires uint4. SVD is pointless with int8

[vladmandic]

"can be" - yes, but why - what is the use case? note the massive complication in the ui that will further confuse most normal users. so need an actual reason, not just "it can"

[Disty0]

You can manually quantize each part outside of SDNext and load the quantized models in SDNext.
SDNQ settings UI is already very crowded as it is right now.

[alerikaisattera]

You can manually quantize each part outside of SDNext

How to do it?

I actually thought of achieving this by saving both SVD and no SVD variants and combining SVD model and no SVD TE. I assume this should work