SD3 and clip_on_cpu

[handshape]

I'm curious about the logic expressed here:

stable-diffusion.cpp/stable-diffusion.cpp

Line 231 in 697d000

if (!ggml_backend_is_cpu(backend) && version == VERSION_3_2B && model_data_type != GGML_TYPE_F32) {

If I'm reading this right, SD3 only does CLIP on the GPU if the model has the 32-bit weights. Is there some special constraint that necessitates this?