Open ali0une opened 10 months ago
Hello! This is probably a VRAM management issue, on the part of the script in textgen-webui.
What's your GPU specs? How much VRAM do you use?
I see you use the new Mixtral model, what's its size? Can you try disabling the 'Manage VRAM' checkbox?
Thanks for the quick answer!
i'm using a NViDIA 3060 RTX with 12G VRAM. The model.safetensors is 7,3G Once charged it's about 11G in VRAM.
if i disable 'Manage VRAM' i just can't load both models due to limited VRAM i have OOM even with a 2G SD-1.5 model. Maybe i should try with a tiny Text generation model like 7B?
Let me know if you need more informations.
Hi!
Just tested with a smaller model TheBloke_Wizard-Vicuna-7B-Uncensored-GPTQ_gptq-4bit-128g-actorder_True
Automatic1111 doesn't seem to be the culprit as i can load a sd-1.5 model just fine, only takes 2G of 12G VRAM. Then Wizard-Vicuna-7B takes 4G more, i'm at 6G of 12G VRAM.
i can tell Text Generation web UI to "send selfie" and it works at about 9G of VRAM usage with a final peak at 12G (VAE i guess) but only if i uncheck the 'Manage VRAM' checkbox.
So you were right as it is probably a VRAM management issue.
Hope that helps.
Very nice extension btw, hope you could fix this issue later and maybe i could use a 13B or 7Bx2 model with an SDXL model in SD.
Yeah, that helps immensely!
I'm trying to narrow the problem down and it seems like the GGUF loader somewhy can't correctly handle the unload-reload cycle (which is unavoidable when Manage VRAM
is on). IDK if I can fix it on my end, but if I manage to, I'll let ooba know with a PR
Hi there!
i'd like to have this extension running on my Text generation web UI but have this error and being new to this local LLaMA AND having limited understanding of python i have this error i can't solve at the moment. Tried to use the script .py of this repository instead of the one in oobabooga repo but still no luck.
With the script.py of this repository i get : On the Automatic1111 side : Total progress:
100%|█████████████████████████████████████████████████████████████████████████████| 32/32 [00:07<00:00, 4.46it/s]
So the image is created.
But then on the text-generation-webui side :
And then the model that was in use is not loaded. Could someone explain why it says "AssertionError: Can't find models/None" please?
Thanks!