OOM or RAM error
Looks like RTX 3060 (12GB VRAM) + 24GB DRAM is not enough for this to load, whatever I try to offload, and even using GGUF CLIP and GGUF VAE, it doesn't go through.
Is there any chance of having a nunchaku GGUF model? Or is it simply impossible?
This solution worked with RTX 3060 (12GB VRAM)
https://github.com/nunchaku-tech/ComfyUI-nunchaku/issues/527#issuecomment-3258055288
try modify nodes/qwenimage.py line number 183 with
model.model.diffusion_model.set_offload(cpu_offload_enabled, num_blocks_on_gpu=30)
I also have the same issue with 12GB VRAM and 24GB DRAM.
The 11.5GB model + 9.4GB CLIP model + CLIP data (image+conditioning) + Ubuntu don't fit on 24GB DRAM.
Moving the CLIP (model+data) on the VRAM is OKish, but ComfyUI starts by using 3GB when starting, so there's no VRAM available for the Ksampler to process, and for unknown reason ComfyUI crashes when loading the 11.5GB in the DRAM.
ComfyUI-MultiGPU may deal with that (smart offload of blocks) but I don't see a way to mix it with nunchaku models loading.