Quantization layer filter

[dxqb]

• Better quantization quality
• Better training quality for new quant types: New quant types, activation quantization #1034

float 8:
full 12.5 GB vram
blocks 12.8 GB
attn-mlp > 16GB, offloading required

nf4:
full 9.4
blocks 9.8
attn-mlp 14.5

Unfortunately, many changes across all models required only to pass on the configuration to the relevant code. The model loader architecture should be generalized, but I've decided against doing that in this PR.

• add "blocks" preset for all models, because that can probably recommended for most models. GGUFs seem to use this filter

[dxqb]

Commit of only the UI changes:
eaf4fe2

[dxqb]

• what type to use for layers that don't match the filter and are therefore not quantized? for most models, this ends up correctly by accident, because recent models are published and loaded in bf16. But it can go wrong, for example for SDXL checkpoints

Train dtype

[dxqb]

merged with #1139: 606b7a8
@O-J1

[O-J1]

This was just merging master in right

[dxqb]

This was just merging master in right

No. This PR moves and refactors all layer filter code from TrainingTab to components.py.

I merged it separately for easy review. Here is #1139 merge: 606b7a8
Here is upstream merge: 8ce4604

[O-J1]

Let me reclarify then:

• 1139 was merged into master last week already no?
• So isnt this just merging master into your PR?

[dxqb]

it is technically a merge, but resolving conflicts and moving your code around. that's why I pinged you.

[O-J1]

I see, alright