Hi, trying to use a particular LoRA and Swarm doesn't do anything! It just kinda hangs and has to be restarted. This is the LoRA https://civitai.com/models/1899191?modelVersionId=2149758 and here are the logs, https://paste.denizenscript.com/View/135942
#No image
clear surge
hybrid locust
are you sure the lora's even related there?
looks like the backend just froze up on the nunchaku model load
clear surge
It worked before using the LoRA, and works after a restart without it.
hybrid locust
actually did it freeze up? you submitted logs ~1 second after clicking gen
can you wait a bit longer before submitting the log? I'd expect more messages after that point
clear surge
However, slight update. It eventually worked, but was incredibly slow, like, 8 mins prep, then failed on the next log, https://paste.denizenscript.com/View/135943
clear surge
what's an OOM?
clear surge
Ahh, oh yeah!!
Any way to fix that?
Also, any idea why it's so slow with that LoRA?
hybrid locust
oh yeah you had a whole long list of loras
you might have better luck with an fp8 copy of flux instead of nunchaku
nunchaku is very unstable code
fp8 is heavier but it's stable core code that's hard to break
clear surge
Wasn't it the Flux model that failed, not the Nunchaku?
hybrid locust
oh wait that was a not-nunchaku that errored i didn't even realize
... how did you gpu oom an fp8 flux
uhh go to Server -> Backends -> edit the comfy backend -> in ExtraArgs add --reserve-vram 2 -> save that
clear surge
Do I need the '2 ->' or was that a typo?
Ok, it's fixed!! I think it was the ridiculously long trigger phase, deleted that and it works fine now.
grizzled temple
Is there a correlation at any point in the flow between longer trigger phrase length or longer prompt length and higher vram or ram usage?
hybrid locust
yes longer prompt uses more vram, by how much is quite variable on the model in use, eg SDXL only changes when the prompt length is multiples of 77 tokens