#Basic Newbie Questions

1 messages · Page 1 of 1 (latest)

regal halo
#

I've been using SD for months but am just starting to play with llama/ooba. Installed Ooba today and I had played with a llama install before. I've been playing with wizard vicuna 30B Q5 1. Using it to generate stories. I seem to get much better, more intelligent seeming results with Ooba with the same prompt and model. The llama install just gets fed the prompt, model name, and n=4000 through a json. On ooba I use the same prompt and model. The only parameters I changed were max new tokens = 2000 and min_length = 2000. So: why is the ooba version so much better? Or am I imagining things?

Question 2: the llama install will keep generating different results with the same prompt until it crashes or I kill it. Is there a way to get Ooba to keep generating off the same prompt?

Question 3: I have the seed in the model tab set to 0. In stable diffusion random seed would mean that I could keep generating with the same prompt to get different results. But with Ooba when I generate again I get the same thing with the same prompt. I can change the prompt slightly and get different results. But do I have to? Or is there a way to get different results from the same prompt/model combo?

Question 4: Is there a way to have Ooba automatically log what it generates to a text file? Or do I need to copy/past each time to save?

cobalt sentinel
#

many of your answers are on the page itself

#

-1 is random, and it's not set on the model tab, among other things

#

i'd encourage you to do some more research

regal halo
cobalt sentinel
#

the same stuff each time is because of the generation parameters

#

llama-precise is precise

#

change it to sphinx moth

#

logging happens in chat, don't think in other modes

regal halo
cobalt sentinel
#

it's a copy-paste

regal halo
#

I can cut and paste

regal halo
# cobalt sentinel it's a copy-paste

Thanks again. With the subreddit down, is there another good resource for beginner questions you could point me to? Most of the documentation I saw for the WebUI was about installation and performance. Like I said I'm coming from stable diffusion which has a bit more robust documentation and settings tweaking setup and, at least as far as I can find, more "how to" pages out there. So if there's a good place to start (that's not the private subreddit), I'd love the pointer. I know these were super basic questions - I'm just having a heck of a time finding decent resources for beginners.

cobalt sentinel
#

not really - this stuff is all in cataclysmic flux still

#

first GPTQ-for-LLaMa was the hotness; then AutoGPTQ. both now appear dead thanks to exllama

#

documentation will come when this stuff solidifies. if you want to be on the cutting edge, you'll have to bring your wits and tools

regal halo
#

ha. fair enough.

cobalt sentinel
#

hate to put it that candidly but :3

regal halo
#

is there any way to run the thing in a loop - like keep generating from thes same promt x times

#

I'm used to running batches of pictures in stable diffusion and picking the best one

cobalt sentinel
#

sure, use the API and your own script

#

within the GUI? no

regal halo
#

hah again, fair enough

#

cool.

cobalt sentinel
#

if you develop a script like that, consider making a PR if it's quality

regal halo
#

I couldn't tell whether this stuff was there and I didn't know where to look of if it wasn't. You've been super helpful.

cobalt sentinel
#

enjoy the ride