search

LostRuins / koboldcpp

Run GGUF models easily with a KoboldAI UI. One File. Zero Install.
https://github.com/lostruins/koboldcpp
GNU Affero General Public License v3.0
5.32k stars 363 forks source link

sd3.5_medium-Q8_0.gguf Crash after execution #1230

Open windkwbs opened 2 days ago

windkwbs commented 2 days ago

sd3.5_medium-Q8_0.gguf Crash after execution koboldcpp_cu12-1.78.exe 微信截图_20241120123222

LostRuins commented 2 days ago

Stable Diffusion 3.5 models have many components. You need to download and load a Clip-L, Clip-G and T5-XXL encoder as well (and maybe VAE)

Grab them all here: https://huggingface.co/Comfy-Org/stable-diffusion-3.5-fp8/tree/main/text_encoders Then load them into the respective files below

windkwbs commented 2 days ago

Stable Diffusion 3.5 models have many components. You need to download and load a Clip-L, Clip-G and T5-XXL encoder as well (and maybe VAE)

Grab them all here: https://huggingface.co/Comfy-Org/stable-diffusion-3.5-fp8/tree/main/text_encoders Then load them into the respective files below

Inference is slow, maybe some stuff is in cpu and not all in GPU

LostRuins commented 2 days ago

What GPU do you have? Did you select the Cublas option when loading?

windkwbs commented 2 days ago

What GPU do you have? Did you select the Cublas option when loading?

微信截图_20241121002712 why？

windkwbs commented 2 days ago

微信截图_20241121003750

LostRuins commented 1 day ago

Okay maybe that gguf is defective. Try this, just load the all-in-one model https://huggingface.co/Comfy-Org/stable-diffusion-3.5-fp8/resolve/main/sd3.5_medium_incl_clips_t5xxlfp8scaled.safetensors

Don't need to select anything else.