SDXL Clip doesn't get quantized and is set to FP32

[rmatif]

I've noticed that with the release master-9578fdc , and specifically with this commit, the ability to quantize CLIP on SDXL models has been removed, causing it to always load as FP32. And it's still the case in the current release

log 9578fdc

log_9578fdc.txt

log b5f4932 (the just previous release)

log_b5f4932.txt

I'm wondering if this change was intentional. If not, I'm surprised that no one has noticed it yet

[stduhpf]

That's definitely not intentional. Only the VAE is supposed to be forced as f32 for SDXL. I'll look into it.

[rmatif]

@stduhpf Thanks for taking a look!

[stduhpf]

Interesting, it seems to only be happening when using a .safetensors file. With a .gguf, the clip models are quantized as expected, which might explain why it got unnoticed before...