ad_discordbot (Fork of Fork of xNul's bot) | Text Generation WebUI | Page 18

valid crypt Sep 27, 2024, 7:52 PM

#

those are ... they are shorter and more 😵‍💫

halcyon quarry Sep 27, 2024, 7:52 PM

#

It looks scarier than it is

#

Just disable the guess imgmodel settings feature if you want to rely 100% on the base settings

#

but you won't be able to dynamically change model settings and tags depending on the model type

#

If you familiarize yourself with that settings handling, you can set up some very good things

valid crypt Sep 27, 2024, 7:54 PM

#

then explain how do i trigger this

halcyon quarry Sep 27, 2024, 7:54 PM

#

I've got dozens of SD15 and SDXL specific LORAs and such

#

when models change, it changes the LORA tags and everything

#

If you have 'xl' in the "filter" tag - it will score models higher that contain "xl" in the filename.

valid crypt Sep 27, 2024, 7:56 PM

#

i started thinking that a command to send in discord with every tag will be useful, if im using bot for myself i may know, but if i let others use it has too much thing to let them know to start

halcyon quarry Sep 27, 2024, 7:56 PM

#

If you have an "exclude" filter, it will also get scored higher if it is missing that from the filename

valid crypt Sep 27, 2024, 7:56 PM

#

halcyon quarry If you have 'xl' in the "filter" tag - it will score models higher that contain ...

so that thing check the file name

halcyon quarry Sep 27, 2024, 7:57 PM

#

When you change imgmodels, it gets the model name and compares it against the presets

#

The preset that scores highest is used

#

this also works with the auto-change imgmodel task, too

valid crypt Sep 27, 2024, 8:09 PM

#

ill try to understand them first

halcyon quarry Sep 27, 2024, 8:13 PM

#

~~If you put your XL models and 15 models into subfolders like /sdxl/ etc - those actually get merged into the filename~~

#

~~So you could more easily manage your filters that way~~

#

actually I was wrong about that last bit

valid crypt Sep 27, 2024, 8:27 PM

#

    filter: ['xl']
    exclude: ['turbo']
    exact_match: ''
    payload:
      width: 1024
      height: 1024
    tags:
      - tag_preset_name: SDXL Tags```
this makes that if i mave a model with xl in the file name generate 1024x1024 imgs and also i can add in payload steps?

#

there is a tag to censor sd is there another for llm?

#

also i consider a bug that after using draw it replies to itself this could be a feature though

#

also got these 22:48:11.866 #5328 ERROR [bot.__main__]: Error merging selected imgmodel data with base imgmodel data: 'payload' 22:48:11.950 #5372 ERROR [bot.__main__]: Error updating settings with the selected imgmodel data: 'width'

#

false alarm

#

from my understanding this should work but doesnt

    filter: ['xl']
    exclude: ['turbo']
    exact_match: ''
    payload:
      width: 1080
      height: 1440
      steps: 20
      sampler_name: "Euler a"
      scheduler: karras
    tags:
      - tag_preset_name: SDXL Tags```

valid crypt Sep 27, 2024, 9:09 PM

#

valid crypt also got these ```22:48:11.866 #5328 ERROR [bot.__main__]: Error merging select...

keep getting this

23:04:34.030 #5328  ERROR [bot.__main__]: Error merging selected imgmodel data with base imgmodel data: 'payload'
23:04:34.116 #5372  ERROR [bot.__main__]: Error updating settings with the selected imgmodel data: 'width'```
sd running from different computer

halcyon quarry Sep 28, 2024, 3:53 AM

#

hmm

#

the second and third errors are likely a result of the first error...

halcyon quarry Sep 28, 2024, 11:11 AM

#

Did you move / rename your image models - and then NOT re-launch your SD client?

#

The SD clients do not collect refreshed list of image models on demand… likely because A1111 coded it that way then everyone forked and didn’t tweak it

halcyon quarry Sep 28, 2024, 11:33 AM

#

I just tested, changing one of my imgmodel names w/o relaunching SD, and it is also erroring same place

#

Ok I see why the subsequent errors are occuring too

#

I can improve the error handling here

valid crypt Sep 28, 2024, 11:53 AM

#

halcyon quarry Did you move / rename your image models - and then NOT re-launch your SD client?

the file is not found just because is not in that computer i think

halcyon quarry Sep 28, 2024, 11:53 AM

#

The SD client thinks it is there

#

otherwise, the model would not appear in the models list (assuming you are using /imgmodel command or auto-change imgmodels)

valid crypt Sep 28, 2024, 11:54 AM

#

the model did change though

halcyon quarry Sep 28, 2024, 11:55 AM

#

The bot said that the model changed, but if you look at the SD cmd window I believe you will see this error:

ERROR [WinError 3] El sistema no puede encontrar la ruta especificada: 'D:\\stable-diffusion-webui\\models\\Stable-diffusion\\AniXL-tPonynai3_v6.safetensors'

valid crypt Sep 28, 2024, 11:57 AM

#

bot says that model did change in discord chat

valid crypt Sep 28, 2024, 11:58 AM

#

halcyon quarry The bot said that the model changed, but if you look at the SD cmd window I beli...

terminal gives the error

#

by generating a image, the model did change

halcyon quarry Sep 28, 2024, 11:58 AM

#

Yes like I said, the bot isn't handling the error well

#

But the error isn't due to the bot, it's because SD is dumb and only collects the model list on startup

#

or if you explicitly click the Refresh Checkpoints button in the UI

valid crypt Sep 28, 2024, 12:01 PM

#

yeah but i didnt change the name or delete file

halcyon quarry Sep 28, 2024, 12:02 PM

#

If you close SD and launch it again, I don't think you'll get an error

valid crypt Sep 28, 2024, 12:05 PM

#

still that error

#

with [[]] tag i could change the steps but everything i ve writteng didnt take effect

halcyon quarry Sep 28, 2024, 12:10 PM

#

updated the bot semi recently?

valid crypt Sep 28, 2024, 12:12 PM

#

updated for the tts

halcyon quarry Sep 28, 2024, 12:17 PM

#

valid crypt terminal gives the error

Did the SD terminal have same error?

#

Are you using the "Per server imgmodels" feature?

valid crypt Sep 28, 2024, 12:21 PM

#

halcyon quarry Did the SD terminal have same error?

sd got no errors

valid crypt Sep 28, 2024, 12:21 PM

#

halcyon quarry Are you using the "Per server imgmodels" feature?

if the default is true

halcyon quarry Sep 28, 2024, 12:24 PM

#

Nah it’s disabled by default

valid crypt Sep 28, 2024, 12:24 PM

#

then im not using that

#

how do i make it only use dict setting parameter?

#

s

halcyon quarry Sep 28, 2024, 12:43 PM

#

Disable the guess imgmodel settings feature in dict imgmodels

valid crypt Sep 28, 2024, 12:46 PM

#

that didnt work

valid crypt Sep 28, 2024, 12:46 PM

#

valid crypt that didnt work

change the steps

halcyon quarry Sep 28, 2024, 12:47 PM

#

You also need to change imgmodels after changing the setting

valid crypt Sep 28, 2024, 12:48 PM

#

rebooted bot

#

still 50 steps

halcyon quarry Sep 28, 2024, 12:50 PM

#

When you do things like change characters, change imgmodel - it retains those settings internally and continues applying them until you change them via the bot functions

#

Such as /imgmodel command

#

In this case the Tags preset it still in effect

#

After disabling the setting - and changing imgmodel - it will not add a tags preset to internal settings

halcyon quarry Sep 28, 2024, 1:20 PM

#

valid crypt sd got no errors

What SD client are you using?

valid crypt Sep 28, 2024, 1:20 PM

#

a1111

halcyon quarry Sep 28, 2024, 1:21 PM

#

Did you browse to the file path that it showed in error, and was it exact same path?

valid crypt Sep 28, 2024, 1:23 PM

#

path is correct but it is on different computer

#

im using my laptop the bot

#

and my pc for sd

halcyon quarry Sep 28, 2024, 1:23 PM

#

We may be getting somewhere

#

The bot tries to read the file to get its filesize and compare that to the presets. I need to add an exception for if it fails

#

Code currently assumes bot, SD and TGWUI are all on same rig

valid crypt Sep 28, 2024, 1:27 PM

#

[[]] is very good but /image cant use tags ;-;

valid crypt Sep 28, 2024, 1:27 PM

#

halcyon quarry The bot tries to read the file to get its filesize and compare that to the prese...

:O

halcyon quarry Sep 28, 2024, 1:28 PM

#

Will do this now actually

#

unless my wife comes upstairs and gets pissy with me 😛

valid crypt Sep 28, 2024, 1:28 PM

#

:P

halcyon quarry Sep 28, 2024, 1:40 PM

#

I just pushed it - try updating, and also turning the "guess model" feature back on

#

It should now print a warning one time when it fails to read the file.
Then, only use the other scoring methods for matching preset

#

#

I'll need to go through my code when I have some time and check if anything else could fail when using SD remotely

#

It's possible this is the only thing, though

valid crypt Sep 28, 2024, 2:01 PM

#

working

halcyon quarry Sep 28, 2024, 2:21 PM

#

Thanks for bearing with me on this

halcyon quarry Sep 28, 2024, 3:58 PM

#

You should also try the post active settings feature - by enabling it and setting a dedicated channel via /set_server_settings_channel

valid crypt Sep 28, 2024, 5:00 PM

#

what does it do?

#

ohhh

#

you have do deal with the design

#

single colored text is 😵‍💫

#

i suppose that if i use edit history the changes aplies to the bot

halcyon quarry Sep 28, 2024, 5:28 PM

#

Yes

valid crypt Sep 28, 2024, 5:44 PM

#

this is not working very well

#

the file name is RLTurbo_dreamshaper but gets the sd15 settings

halcyon quarry Sep 28, 2024, 5:49 PM

#

exclude turbo from the first preset

#

Or, move the turbo preset above the other one. When there is a tied score the higher one is prioritized

#

You could also use exact match if necessary

valid crypt Sep 28, 2024, 5:52 PM

#

where are setting for like who can use commands

#

like /image its fine but i dont want other to change the model

#

i only saw setting to disable dm

halcyon quarry Sep 28, 2024, 5:56 PM

#

Discord server settings > Integrations

valid crypt Sep 28, 2024, 6:00 PM

#

nice

halcyon quarry Sep 28, 2024, 6:07 PM

#

The settings for DM commands are the only commands that can be used in DM

#

except bot owner can use any cmd in DM

#

They’re disabled by the bot for users by default

valid crypt Sep 28, 2024, 6:30 PM

#

how do i remove voice channel?

halcyon quarry Sep 28, 2024, 6:35 PM

#

Each server can only have one dedicated voice channel. I don’t have a means to remove it currently once set, but you can either change the play mode setting (config.yaml) or /toggle_tts to disable tts temporarily altogether

#

I’ll soon add an option to the /set_server_voice_channel command to clear the current value

halcyon quarry Sep 28, 2024, 6:51 PM

#

valid crypt a1111

You should strongly consider trying Forge or Reforge instead of A1111

#

I’ll recommend Swarm as well once I finish adding support for it in the bot 🤗

#

A1111 is significantly slower and will OOM before the other UIs

valid crypt Sep 28, 2024, 7:04 PM

#

slower?

#

whats oom

#

out of maintenance?

#

i dont update a1111 very frequently

#

3month+

halcyon quarry Sep 28, 2024, 7:29 PM

#

Out of memory

#

Yes A1111 will generate images slower than Forge or Reforge with same settings and same result

#

While using more Vram

valid crypt Sep 28, 2024, 7:48 PM

#

:O

halcyon quarry Sep 28, 2024, 8:29 PM

#

Reforge is more stable of the two but development is likely ended for it

#

Forge needs some new settings to use it via API which I added to the bot recently

valid crypt Sep 28, 2024, 8:40 PM

#

forge is like a addon for a1111 so i can use it like always?

valid crypt Sep 28, 2024, 8:41 PM

#

halcyon quarry Forge needs some new settings to use it via API which I added to the bot recentl...

you are more of a forge user than reforge?

halcyon quarry Sep 28, 2024, 9:03 PM

#

I just switched to Forge full time after making it work with the bot

#

I was using Reforge for API only

#

They are similar interface to A1111 but with many differences on the backend

#

Quick history lesson:

Illyasviel released Forge awhile back. It was very good, however he then disappeared and no one could merge PRs or develop it further.
Someone 5 months later decided to fork it as Reforge, then update it to include A1111 updates
Illyasviel came back to Forge and flipped it upsidedown, made it compatible with Flux, appointed some other co-devs (including the Reforge guy) to ensure the project development no longer stagnates

valid crypt Sep 28, 2024, 9:14 PM

#

that made decision making much easier

#

:)

halcyon quarry Sep 28, 2024, 9:16 PM

#

After making it work with Flux development has mainly been fixing the resulting bugs

#

Forge is also using Gradio 4, there are a number of extensions which have some bugs due to incompatiblity

valid crypt Sep 28, 2024, 9:17 PM

#

D:

halcyon quarry Sep 28, 2024, 9:19 PM

#

I have an extension which I was able to update - other devs did the same or some other ambitious users are forking / making Forge compatible versions

#

They have a Discussions thread on it in the github

halcyon quarry Sep 28, 2024, 9:50 PM

#

You should check it out btw

#

https://github.com/altoiddealer/--sd-webui-ar-plusplus

GitHub

GitHub - altoiddealer/--sd-webui-ar-plusplus: Select img aspect rat...

Select img aspect ratio from presets in sd-webui. Contribute to altoiddealer/--sd-webui-ar-plusplus development by creating an account on GitHub.

valid crypt Sep 28, 2024, 10:04 PM

#

i have a similar extension

#

ill make a backup and see what will happen

valid crypt Sep 28, 2024, 10:38 PM

#

Auto-merging scripts/postprocessing_codeformer.py
CONFLICT (content): Merge conflict in scripts/postprocessing_codeformer.py
Auto-merging scripts/postprocessing_gfpgan.py
CONFLICT (content): Merge conflict in scripts/postprocessing_gfpgan.py
Auto-merging scripts/xyz_grid.py
Auto-merging style.css
CONFLICT (content): Merge conflict in style.css
Automatic merge failed; fix conflicts and then commit the result.``` 🥲

#

hint: Fix them up in the work tree, and then use 'git add/rm <file>'
hint: as appropriate to mark resolution and make a commit.
fatal: Exiting because of an unresolved conflict.```

halcyon quarry Sep 28, 2024, 10:56 PM

#

What is this from?

#

Did you try a fresh install of Forge, and are getting this error?

valid crypt Sep 28, 2024, 10:58 PM

#

i want to keep my a1111

#

settings

halcyon quarry Sep 28, 2024, 10:59 PM

#

You'll definitely need to have two separate installations

#

Forge has settings to specify your checkpoints folder, etc (can be shared).
Personally I use symlinks

#

I have these symlinked from my A1111 folder

valid crypt Sep 28, 2024, 11:23 PM

#

how do i do that

halcyon quarry Sep 28, 2024, 11:52 PM

#

Awhile back I asked ChatGPT to make me a bat file that could do it via drag and drop

📎 create-symlink.bat

#

put the bat file somewhere
make a shortcut to it somewhere like your desktop
right click the shortcut > Properties > Advanced > Check off Run as Administrator
Now you may drag-drop a file or a folder onto the shortcut, and it will prompt for the destination path

#

So you might drop your extensions folder onto it and enter something like
C:\0_SD\stable-diffusion-webui-forge\extensions

valid crypt Sep 29, 2024, 12:04 AM

#

:O

valid crypt Sep 29, 2024, 12:26 AM

#

some how i have 2 sd running in the same port xd

#

192 and 127 are different :P

#

what are these?

valid crypt Sep 29, 2024, 12:58 AM

#

uhhh if i launch forge with --api it crashes my fav extensions...

#

most extensions .-.

#

it is a lot faster though

halcyon quarry Sep 29, 2024, 3:09 AM

#

valid crypt what are these?

These are quick settings presets for the entire UI (cfg, steps, dimensions, etc)

#

allows easily changing all your UI settings in a flash, when changing base model types

valid crypt Sep 29, 2024, 2:07 PM

#

llama 3.2 🤩 im excited to use it in tgwui

halcyon quarry Sep 29, 2024, 3:25 PM

#

Working on the persist tag param

#

it's a can of worms indeed

valid crypt Sep 29, 2024, 3:30 PM

#

🤩

valid crypt Sep 29, 2024, 3:56 PM

#

do i have to change anything to use forge?

#

im getting internal server error

halcyon quarry Sep 29, 2024, 4:10 PM

#

Need to make sure there is a text_encoders subdirectory in the models folder (if you symlinked it from A1111)

#

Otherwise, I’d be interested in knowing the error in Forge console

#

If you are getting out of memory error, then you may need to include a new forge specific param in dict_basesettings (check settings templates), it’s commented

#

May need to guess and check the value for it

#

forge_inference_memory

valid crypt Sep 29, 2024, 4:22 PM

#

halcyon quarry Need to make sure there is a text_encoders subdirectory in the models folder (if...

oh

#

📎 message.txt

valid crypt Sep 29, 2024, 4:28 PM

#

halcyon quarry Need to make sure there is a text_encoders subdirectory in the models folder (if...

got that folder from github

#

but there no files in it

#

idk if there has to be something

#

i did the symlink of model and output and
modified the user bat
set COMMANDLINE_ARGS= --port 7862 --listen --api
set CUDA_VISIBLE_DEVICES=1

#

i just cant figure why everything goes wrong with --api

valid crypt Sep 29, 2024, 4:32 PM

#

valid crypt

this appeard when i accessed the link with chrome

halcyon quarry Sep 29, 2024, 4:35 PM

#

is the url + port correct in the bot's config.yaml?

valid crypt Sep 29, 2024, 4:38 PM

#

bot literally printed this in the console

halcyon quarry Sep 29, 2024, 4:40 PM

#

Enter this in your chrome

#

http://192.168.1.11:7862/docs

#

Let me know if you get the API interface

valid crypt Sep 29, 2024, 4:42 PM

#

that worked

#

i did a 100% clean install with git and with the pack

halcyon quarry Sep 29, 2024, 4:44 PM

#

If you scroll down to the cmd-flags endpoint and test it - does it error?

valid crypt Sep 29, 2024, 4:45 PM

#

halcyon quarry If you scroll down to the cmd-flags endpoint and test it - does it error?

?

halcyon quarry Sep 29, 2024, 4:45 PM

#

In the chrome API page

#

You can click on an endpoint to expand it, then “try it out”

valid crypt Sep 29, 2024, 4:48 PM

#

loading forever

#

wait i think it worked?

halcyon quarry Sep 29, 2024, 4:50 PM

#

If below it says response code 200 then it was successful

valid crypt Sep 29, 2024, 4:51 PM

#

then yes

#

i did sdapi/v1/sd-models

halcyon quarry Sep 29, 2024, 4:52 PM

#

yes the bot uses that endpoint to collect all the model data

#

It uses cmd-flags endpoint on startup only to read one of the values to determine which SD WebUI is being used

valid crypt Sep 29, 2024, 4:53 PM

#

so actually does not matter?

halcyon quarry Sep 29, 2024, 4:53 PM

#

well the bot will automatically fix certain things depending on what UI is being used

valid crypt Sep 29, 2024, 4:54 PM

#

#

🥲

#

maybe it is something with the latest forge?

halcyon quarry Sep 29, 2024, 4:56 PM

#

I just updated and I'm not getting the error

valid crypt Sep 29, 2024, 4:56 PM

#

;-;

#

i could try with my local chrome

#

nope

halcyon quarry Sep 29, 2024, 4:59 PM

#

maybe you have some kind of firewall issue?

valid crypt Sep 29, 2024, 4:59 PM

#

a1111 had no issue

halcyon quarry Sep 29, 2024, 4:59 PM

#

network security setting etc

valid crypt Sep 29, 2024, 4:59 PM

#

my local machine also cant

halcyon quarry Sep 29, 2024, 5:00 PM

#

You may have set some "Allow" permission for A1111 explicitly, or something

#

or maybe the url/port is slightly different from A1111?

#

(if that specific port were "allowed")

valid crypt Sep 29, 2024, 5:03 PM

#

valid crypt i did the symlink of model and output and modified the user bat set COMMANDLINE...

it was the port

#

but just the flag

#

Running on local URL:  http://0.0.0.0:7862``` literally same but automatic

#

let me cry a few secs

halcyon quarry Sep 29, 2024, 5:05 PM

#

resolved then? 😛

valid crypt Sep 29, 2024, 5:06 PM

#

it says that im using a1111

halcyon quarry Sep 29, 2024, 5:07 PM

#

It determines it based on what the installation folder ends with

#

If your installation folder ends with -forge then it will acknowledge that Forge is running

#

such as stable-diffusion-webui-forge

#

Actually I was slightly wrong here is exactly how it determines the UI

#

valid crypt Sep 29, 2024, 5:10 PM

#

the one click install pack does not have -forge at then end :v

halcyon quarry Sep 29, 2024, 5:11 PM

#

Did you modify the directory name? Or is it same as default?

#

If default please share the directory name verbatim

valid crypt Sep 29, 2024, 5:13 PM

#

this is the one click install pack, well its 4 but its really simple

halcyon quarry Sep 29, 2024, 5:13 PM

#

up one level

valid crypt Sep 29, 2024, 5:13 PM

#

thats the content

#

the user bat in in webui

valid crypt Sep 29, 2024, 5:13 PM

#

valid crypt thats the content

o the 7zip

#

of

#

imo more people would install trough the pack

#

the github page

halcyon quarry Sep 29, 2024, 5:16 PM

#

so is it something like...

#

/webui_forge_cu121_torch231/

valid crypt Sep 29, 2024, 5:17 PM

#

that is the folder i created

#

this is what the pack has

#

webui is the sd

halcyon quarry Sep 29, 2024, 5:18 PM

#

I'll rework the logic a little bit...

valid crypt Sep 29, 2024, 5:18 PM

#

now that i gave it another thought, people may change the name so it is prettier or smth

halcyon quarry Sep 29, 2024, 5:19 PM

#

#

just pushed the change.

Should resolve the issue!

valid crypt Sep 29, 2024, 5:22 PM

#

halcyon quarry

https://tenor.com/view/frye-dyslexic-gif-4978104

Tenor

Not sure

▶ Play video

halcyon quarry Sep 29, 2024, 5:23 PM

#

now it just looks for any instance of "reforge" anywhere in the string, then checks if "forge" is anywhere in the string, etc

#

changed the order it checks so that "Reforge" won't accidentally be matched to "Forge"

halcyon quarry Sep 29, 2024, 5:53 PM

#

@valid crypt tried it yet?

valid crypt Sep 29, 2024, 5:55 PM

#

halcyon quarry If your installation folder ends with `-forge` then it will acknowledge that For...

when said that i just cleaned the mess and renamed the folder :P

halcyon quarry Sep 29, 2024, 5:55 PM

#

I don't know if any internal settings are populated with the directory name when you initially install it

#

I recommend reverting back to the forge directory name you had, for that reason

valid crypt Sep 29, 2024, 5:56 PM

#

error! :D

#

that fine right?

#

if its fine consider changing it to warn and i suggest warn in orange

#

red is too scary

halcyon quarry Sep 29, 2024, 6:00 PM

#

While resolving this had to merge in some WIP code I’ve been working on the backburner

#

Will see what I can do about that

halcyon quarry Sep 29, 2024, 6:07 PM

#

valid crypt that fine right?

Yeah it’s expected, for now

valid crypt Sep 29, 2024, 6:51 PM

#

buddy

#

you know that you have a filter for sd right?

#

that doesnt apply to /image

#

😱

halcyon quarry Sep 29, 2024, 6:55 PM

#

You likely need to change tags search mode to user llm

valid crypt Sep 29, 2024, 7:01 PM

#

do that really work with /image...

#

i think it didnt work

#

added the settings, reloaded bot, character, imgmodel still dont work

halcyon quarry Sep 29, 2024, 8:49 PM

#

Will check it out

#

Just censoring not working?

halcyon quarry Sep 29, 2024, 9:11 PM

#

valid crypt you know that you have a filter for sd right?

Clarify what you’re talking about please

valid crypt Sep 29, 2024, 9:12 PM

#

more than a bug it more like a suggestion

#

that does not apply to /image

#

and also it does not delete the message

#

the bot still tell the prompt

#

the image is not generated but like always there is a description of the image, so that is what im refering to

halcyon quarry Sep 29, 2024, 9:27 PM

#

valid crypt the image is not generated but like always there is a description of the image, ...

Do you have 2 separate tags with img_censoring ? If so, the first one ~~may be getting~~ will be prioritized if both are "matched"

valid crypt Sep 29, 2024, 9:29 PM

#

halcyon quarry Do you have 2 separate tags with `img_censoring` ? If so, the first one ~~may b...

it is about the llm sending the description

#

i say draw p**n

#

bot: i drew a n**ed woman

#

sd:

halcyon quarry Sep 29, 2024, 9:32 PM

#

It's not drawing the image though, yes?

valid crypt Sep 29, 2024, 9:32 PM

#

halcyon quarry It's not drawing the image though, yes?

affirmative

#

not drawing but saying inappropriate stuffs

halcyon quarry Sep 29, 2024, 9:33 PM

#

I'll update this just need to think about the best way to handle this one

valid crypt Sep 29, 2024, 9:34 PM

#

so this could be a suggestion: censor for llm

#

from my understanding if the user does not say any inappropriate words, llm mostly dont, so tags for censorship?

#

same for sd, for llm

halcyon quarry Sep 29, 2024, 9:36 PM

#

Please confirm... it's not sending a message with the problematic prompt, and also not drawing the image, when using /image command, correct?

#

It's only the tag activated prompt that is still presented in chat

valid crypt Sep 29, 2024, 9:36 PM

#

the command ignores the censor

halcyon quarry Sep 29, 2024, 9:36 PM

#

That's interesting...

valid crypt Sep 29, 2024, 9:37 PM

#

to clarify
using draw --> problematic reply but sd blocked
using /image --> simply gives the image

halcyon quarry Sep 29, 2024, 9:42 PM

#

does it say anything in the CMD window about prompt being flagged?

valid crypt Sep 29, 2024, 9:43 PM

#

you could reproduce this "bug" though

#

let me do manual censor

#

ship

#

#

saf

#

e

#

😁

halcyon quarry Sep 29, 2024, 9:45 PM

#

Trying to reproduce it

#

Ok big bug in /image - idk how I missed this but it isn't matching tags at all

#

You are an amazing user btw, you've found a lot of bugs - I'm very thankful

#

I got the /image command squared up, tags all working correctly again.
I'll be adding another censoring tag for LLM.

valid crypt Sep 29, 2024, 10:10 PM

#

😳

#

oh no edge tts disappeared https://github.com/Unorthodox-oddball/text-generation-webui-edge-tts

halcyon quarry Sep 29, 2024, 10:15 PM

#

yikes

valid crypt Sep 29, 2024, 10:15 PM

#

it was a amazin extension 😔

halcyon quarry Sep 29, 2024, 10:15 PM

#

Did you fork it to send that PR? Still have the fork?

valid crypt Sep 29, 2024, 10:15 PM

#

i have it 😁

halcyon quarry Sep 29, 2024, 10:15 PM

#

You are new edge dev 🤓

valid crypt Sep 29, 2024, 10:16 PM

#

bruh

halcyon quarry Sep 29, 2024, 10:16 PM

#

Share fork link bc I want to link it from my repo

valid crypt Sep 29, 2024, 10:16 PM

#

i just wanted to ask if it is ethical to publish the fork :v

valid crypt Sep 29, 2024, 10:16 PM

#

halcyon quarry Did you fork it to send that PR? Still have the fork?

wait no

#

i dont have the fork

#

i have the files :O

halcyon quarry Sep 29, 2024, 10:16 PM

#

Ah

#

Well you could just publish it, only difference is it won’t say “Fork of (link to a 404 page)”

valid crypt Sep 29, 2024, 10:20 PM

#

?

halcyon quarry Sep 29, 2024, 10:21 PM

#

If the original edge TTS project exists maybe just fork that

#

Then replace the files

valid crypt Sep 29, 2024, 10:21 PM

#

nice

#

sir, how do i replace?

halcyon quarry Sep 29, 2024, 10:27 PM

#

Could just go one by one delete everything but the readme

#

Then git push your files

valid crypt Sep 29, 2024, 10:29 PM

#

uhh how do i delete...

#

found it :)

#

how do i change the name?

#

oh

#

if everything is alr then this should work https://github.com/marcos33998/edge_tts/tree/master

GitHub

GitHub - marcos33998/edge_tts: A very simple implementation of edge...

A very simple implementation of edge_tts w/ RVC for oobabooga text-generation-webui. - marcos33998/edge_tts

#

i hope tomorrow i dont receive a copyright warn, law suit and being bullied on twitter

#

ill test if it works tomorrow

#

should work...

halcyon quarry Sep 29, 2024, 11:02 PM

#

Recommend naming it like the other one though- text-generation-webui-edge-tts

halcyon quarry Sep 30, 2024, 2:59 AM

#

@valid crypt just pushed fix for everything

/image tags are now working
Improved handling for censoring
added llm_censoring tag and example and updated Wiki
Suppress the SwarmUI warning on startup

#

Learned how to raise a custom Exception in the process 🤓

#

valid crypt Sep 30, 2024, 6:30 AM

#

halcyon quarry Recommend naming it like the other one though- text-generation-webui-edge-tts

the code itself require edge_tts as its name

halcyon quarry Sep 30, 2024, 11:17 AM

#

Derp

valid crypt Sep 30, 2024, 2:07 PM

#

?

halcyon quarry Sep 30, 2024, 2:07 PM

#

I shouldv'e known

valid crypt Sep 30, 2024, 3:26 PM

#

check this https://github.com/oobabooga/text-generation-webui/commit/9ca0cd7749f734bd5e7083d672dee8d400229257 it is dev though

GitHub

Bump llama-cpp-python to 0.3.1 · oobabooga/text-generation-webui@9c...

sullen plover Sep 30, 2024, 3:31 PM

#

valid crypt check this https://github.com/oobabooga/text-generation-webui/commit/9ca0cd7749f...

this means i should update the main script?

valid crypt Sep 30, 2024, 3:32 PM

#

you shouldnt upgrade

#

as it is in dev

#

what you can do is try it

sullen plover Sep 30, 2024, 3:33 PM

#

still trying to figure out how to stop the bot replying all channels asked before but at least it works

sullen plover Sep 30, 2024, 3:33 PM

#

valid crypt what you can do is try it

thanks

valid crypt Sep 30, 2024, 3:33 PM

#

like creating a separated tgwui

valid crypt Sep 30, 2024, 3:33 PM

#

sullen plover still trying to figure out how to stop the bot replying all channels asked befor...

?

#

i think there is no way to prevent bot from replying if you ping it, when you do /main it always reply

sullen plover Sep 30, 2024, 3:36 PM

#

might be the name of the bots all being zero

#

need to fine tune a new llm 8b or 11b maybe

#

11b works on my cpu about the best i can do 9b might be good 11b can be slow 8b best so far

valid crypt Sep 30, 2024, 3:54 PM

#

found a bug: /toggletts (didnt try to hear) toggles through multiple server but only the bot in the server that you use the command will connect/disconnect

#

example
i disable tts in server A, bot disconnect in server A but still in B
I enable tts in server B, the bot in server A does not reconnect

halcyon quarry Sep 30, 2024, 3:57 PM

#

per_server_settings is false in config.yaml ?

valid crypt Sep 30, 2024, 3:57 PM

#

using the default

halcyon quarry Sep 30, 2024, 3:57 PM

#

I'll check out the code now - I may have accidentally changed it to always behave as if server specific settings

#

while it should disconnect from all servers if that feature is disabled

valid crypt Sep 30, 2024, 3:58 PM

#

terminal says per server setting not found, applying defgault

halcyon quarry Sep 30, 2024, 3:58 PM

#

right, default is false

valid crypt Sep 30, 2024, 3:59 PM

#

halcyon quarry <@323088470241312774> just pushed fix for everything - `/image` tags are now wor...

using before this update

halcyon quarry Sep 30, 2024, 3:59 PM

#

The toggle_tts behavior was changed a month or so ago

#

ok yes, I see the flaw in the logic

valid crypt Sep 30, 2024, 4:00 PM

#

as idk about code tell the version is always better :)

#

just in case :)

halcyon quarry Sep 30, 2024, 4:06 PM

#

My bot is on one server atm

#

Could you test this change?

#

            vc_guild_ids = [self.ictx.guild.id] if config.per_server_settings else [guild.id for guild in client.guilds]
            for vc_guild_id in vc_guild_ids:
                await voice_clients.toggle_voice_client(vc_guild_id, message)

#

in the function toggle_tts_task

#

If 'per-server settings' is enabled, it will only toggle the voice client in the current guild.
Otherwise, it will iterate over all guilds that the bot is in and toggle them

valid crypt Sep 30, 2024, 4:16 PM

#

https://media.discordapp.net/attachments/1103411504109326368/1290346413544312832/image.png?ex=66fc203e&is=66facebe&hm=e80fd8715d8e0d780977cd7bc98d04c0830d0076ea554d8c39338ecd8eb318a3&=&format=webp&quality=lossless&width=1440&height=504

#

ok messed up already :P

halcyon quarry Sep 30, 2024, 4:16 PM

#

You overwrote this line 😛

            if self.embeds.enabled('change'):

valid crypt Sep 30, 2024, 4:17 PM

#

halcyon quarry Sep 30, 2024, 4:17 PM

#

Condition will be true anyway

valid crypt Sep 30, 2024, 4:17 PM

#

my eyes says pass

halcyon quarry Sep 30, 2024, 4:17 PM

#

yep

#

Yes agreed- now when you toggle tts it should apply to all servers

#

If so, I push change 🤗

#

ugh

#

sorry

#

One thing I messed up

valid crypt Sep 30, 2024, 4:19 PM

#

just about to say that

#

didnt work :)

halcyon quarry Sep 30, 2024, 4:19 PM

#

Change the one little snippet to this
if config.is_per_server()

valid crypt Sep 30, 2024, 4:21 PM

#

my eyes says pass

halcyon quarry Sep 30, 2024, 4:21 PM

#

🇵 🇦 🇸 🇸

#

hehe... still may not work....

#

er nvm

#

should work

valid crypt Sep 30, 2024, 4:23 PM

#

worked :)

halcyon quarry Sep 30, 2024, 4:24 PM

#

very cool - there's a number of discord values the bot actually doesn't hold which need to use a .fetch() method to get.
Thought for a sec guild ID could be one

#

Pushed - so just undo those changes and Pull

#

Lmk when you've had a chance to test the improved censoring features, if all seems good on your end

valid crypt Sep 30, 2024, 4:27 PM

#

halcyon quarry Lmk when you've had a chance to test the improved censoring features, if all see...

https://tenor.com/view/mhdinner-dinner-time-mhdnr-dnrmh-dinner-gif-14714090031558288784

Tenor

#

soon

valid crypt Sep 30, 2024, 6:08 PM

#

first impression: working

#

second impression: hard to list every word and typo

#

going for the deep test

#

done
result: bot can bypass the censor

#

checking user✅
checking bot❌
user can make bot say those words

halcyon quarry Sep 30, 2024, 6:15 PM

#

🧐

valid crypt Sep 30, 2024, 6:17 PM

#

valid crypt more than a bug it more like a suggestion

#

the word in the ban list, the forth one from the right

#

halcyon quarry Sep 30, 2024, 6:20 PM

#

I'm working on it now

#

If you don't mind!

#

... could you just slap this code in and test it?

valid crypt Sep 30, 2024, 6:21 PM

#

😁

halcyon quarry Sep 30, 2024, 6:21 PM

#

                    elif key == 'llm_censoring' and value == True:
                        censor_text = tag.get('matched_trigger', '')
                        censor_message = f' (text match: {censor_text})'
                        log.info(f"[TAGS] Censoring: LLM response was blocked{censor_message if censor_text else ''}")
                        await self.embeds.send('img_send', "LLM response was flagged as inappropriate", "Further task processing has been cancelled.", delete_after=5)
                        raise TaskCensored

valid crypt Sep 30, 2024, 6:21 PM

#

alr

halcyon quarry Sep 30, 2024, 6:21 PM

#

In the collect_img_tag_values() function

#

(different from collect_llm_tag_values)

valid crypt Sep 30, 2024, 6:22 PM

#

wait

valid crypt Sep 30, 2024, 6:22 PM

#

halcyon quarry In the `collect_img_tag_values()` function

cofussion

#

?

halcyon quarry Sep 30, 2024, 6:23 PM

#

ah yes

#

hmm

#

right

valid crypt Sep 30, 2024, 6:23 PM

#

halcyon quarry ``` elif key == 'llm_censoring' and value == True: ...

i take this as the correct answer

#

wait

#

huh

halcyon quarry Sep 30, 2024, 6:23 PM

#

hold plz 🙂

#

yeah, forget adding that code... I need to add a similar check right before it sends the response

valid crypt Sep 30, 2024, 6:26 PM

#

halcyon quarry yeah, forget adding that code... I need to add a similar check right before it s...

https://tenor.com/view/digilah-digifun-malaysia-largest-network-digi-emoji-mix-emoji-gif-17801560

Tenor

#

i tell you an idea later :V

halcyon quarry Sep 30, 2024, 6:27 PM

#

I'll have to analyze the text as it's generating really (while determining whether to send response chunks)

#

before it could trigger TTS, etc

valid crypt Sep 30, 2024, 6:27 PM

#

true

halcyon quarry Sep 30, 2024, 6:30 PM

#

this is getting complicated hehe

valid crypt Sep 30, 2024, 6:35 PM

#

feeling it

#

remember to change the comment

halcyon quarry Sep 30, 2024, 6:36 PM

#

which comment?

valid crypt Sep 30, 2024, 6:37 PM

#

says tag for image censoring feature

#

just a little detail

#

or im blind

#

im not blind, its a little detail

valid crypt Sep 30, 2024, 6:39 PM

#

valid crypt says tag for image censoring feature

dict_tags

halcyon quarry Sep 30, 2024, 6:39 PM

#

right, the code is not going to be there in the image censoring block

valid crypt Sep 30, 2024, 6:40 PM

#

either another block or make it image/llm

halcyon quarry Sep 30, 2024, 6:41 PM

#

I need to make a special little tag matching function.
When I build the full list of tags, I'm going to add llm_censoring tags to a separate list, in order to quickly check the LLM's reply against them

valid crypt Sep 30, 2024, 6:44 PM

#

https://media.discordapp.net/attachments/1017863023349747792/1107444186321539142/ezgif.com-gif-maker.gif

halcyon quarry Sep 30, 2024, 7:00 PM

#

So I think the logic should be to block the text string that contains the censored text? Rather than nuke the whole response?

#

As in, the response chunk

#

I can replace the text chunk with:
LLM response was flagged as inappropriate.

#

actually, may be easier to just nuke it after detecting lol

#

I'm mostly doing this for you so whichever logic you think makes the most sense, I will apply here

#

either one sounds good to me

valid crypt Sep 30, 2024, 7:15 PM

#

halcyon quarry I can replace the text chunk with: **LLM response was flagged as inappropriate.*...

i was think on another filter to replace swear words to "filtered"

valid crypt Sep 30, 2024, 7:16 PM

#

halcyon quarry So I think the logic should be to block the text string that contains the censor...

this is pretty good for now

#

nuke

#

https://tenor.com/view/fallout76-nuke-friends-bomb-nuclear-gif-14693420697390417365

Tenor

#

what im thinking is there is some words that are too dangerous so nuke
some like sh*t then replace

valid crypt Sep 30, 2024, 7:18 PM

#

halcyon quarry I can replace the text chunk with: **LLM response was flagged as inappropriate.*...

this amazing too

#

just custom text is better :)

#

this is a piece of sh-potato

halcyon quarry Sep 30, 2024, 7:21 PM

#

text replacement can already happen for user prompt as well as LLM's response (except currently only when the LLM response is used for image prompt).

#

So I may just work in text replacement on all LLM responses as well... need to think on that.

valid crypt Sep 30, 2024, 7:23 PM

#

imagine a server getting wiped because of a bot sending I love n(for my own security)z and also the audio

#

funny

halcyon quarry Sep 30, 2024, 7:23 PM

#

😆

valid crypt Sep 30, 2024, 7:50 PM

#

my edgetts is working >_<

halcyon quarry Sep 30, 2024, 8:06 PM

#

testing the new censor code...

#

It's working

valid crypt Sep 30, 2024, 8:14 PM

#

🤩

halcyon quarry Sep 30, 2024, 8:14 PM

#

Pushed int to main

#

It collects all "llm_censoring" tags to a separate variable.
Whenever text is chunked, or complete response is finished - it will do a quick check against those tag triggers.
If it finds a match, it nukes the process, send the embed and logs it

valid crypt Sep 30, 2024, 8:19 PM

#

dict tag remain the same?

#

i suppose

halcyon quarry Sep 30, 2024, 8:20 PM

#

Yep

#

It's just going to check the same stuff on user text and LLM text

valid crypt Sep 30, 2024, 8:21 PM

#

🥳

halcyon quarry Sep 30, 2024, 11:23 PM

#

I’ll be updating it to only apply that logic for search mode “llm” (or “userllm”)

halcyon quarry Oct 1, 2024, 12:13 AM

#

Aaaand that's done. Also noticed I had left a debug print statement now removed

valid crypt Oct 1, 2024, 2:41 PM

#

👌

sullen plover Oct 1, 2024, 10:17 PM

#

Zero
APP
— Today at 11:16 PM
@ZeroGroq pauses for input

User Input: Change Zero's reply_to_itself attribute from 0.5 to 1.0

#

the ai is telling me the issue it is a setting in one of the files i need to edit to stop it replying all channels?

halcyon quarry Oct 1, 2024, 10:30 PM

#

If you don’t want the bot to talk in a channel you’ll need to use /main to toggle whether it replies to everything or only @ mentions

#

Or remove its join/ View permissions for those channels

#

reply_to_itself is the chance that it replies to its own message

sullen plover Oct 2, 2024, 10:57 AM

#

halcyon quarry reply_to_itself is the chance that it replies to its own message

hmm and i can do that using slash commands sir?

sullen plover Oct 2, 2024, 11:18 AM

#

there is a file i edited changed this setting the ai is telling me the issue but i forgot where the file is: Change Zero's reply_to_itself attribute from 0.5 to 1.0

halcyon quarry Oct 2, 2024, 12:05 PM

#

Its a “behavior” setting - can be defined in base settings and/or character card

#

If in character card will override base settings

#

See Minty for example usage

sullen plover Oct 2, 2024, 1:01 PM

#

ok thank you 🙂

valid crypt Oct 2, 2024, 1:56 PM

#

halcyon quarry See Minty for example usage

what about clone the repo and call it minty_discordbot

#

ad is 💀

valid crypt Oct 2, 2024, 5:43 PM

#

pretty sure this is a bug

#

from my understanding the bot cant be reloaded unless i close it and reopen

#

but i left it there and yeah...

halcyon quarry Oct 2, 2024, 7:03 PM

#

I’ve seen this before I think discord issues can cause this

#

They occasionally have hiccups in their service

terse folio Oct 2, 2024, 7:22 PM

#

the on_ready event can be triggered multiple times as the bot reconnects.

if you have code you only want to run once, create a separate on_ready function like on_first_ready

async def on_ready(bot..?):
    if self.first_ready:
        self.first_ready = False
        await self.on_first_ready()

and ofc init self.first_ready somewhere in the bot

as you don't create your own bot class, any variable outside the bot should work.
But be careful with python's local scope variables.

first_ready = True
def test():
    first_ready = False
test()
print(first_ready)```

should still yield `True`.... so this is a case where using a class will work best!
perhaps active_settings!

#

Ahh, you already got one!

#

i'll move the background task initialisations there because it could cause bugs having multiple copies of "process background task" going on at once

halcyon quarry Oct 2, 2024, 8:58 PM

#

Well the first_run() was not actually made for initial bot launch - it was for first every deployment of the bot on machine - sends welcome message

#

I understand what you're saying though

terse folio Oct 2, 2024, 9:05 PM

#

oh okay,
will update my code ^^

#

#

i've added a context manager to wrap sending messages to channels.

#

that can replace the need to do try:catch for the 403 error if the bot has no access to a channel

halcyon quarry Oct 2, 2024, 9:48 PM

#

Thank you for the contribution gowron2

#

I made a lot of neat changes you should check out, like Embeds()

#

I think the way I did Settings() is pretty neat too

valid crypt Oct 3, 2024, 1:11 PM

#

no img generating message after changing model

valid crypt Oct 3, 2024, 1:25 PM

#

valid crypt no img generating message after changing model

only the first request

halcyon quarry Oct 3, 2024, 2:57 PM

#

Known issue. This is something that @terse folio might be able to fix... I've tried a number of times with ChatGPT but can't figure it out

#

It has something to do with this. The track_progress() task likes to go kaput.

#

In the past the original code to handle these processes would often result in the whole generation getting screwed

#

I settled on the progress bar sometimes disappearing as... good enough

halcyon quarry Oct 3, 2024, 3:04 PM

#

valid crypt only the first request

This is a good observation though, I'll need to see if this happens as predictably as you suggest

#

I'll never be able to get away from NeuralBeagle14-7B I love the output from this model

#

Also, the "SDXL Ghost Style" LORA has very cool results https://civitai.com/models/169671?modelVersionId=190677

      - trigger: 'ghost style,ghostly'
        positive_prompt: 'GhostlyStyle, <lora:SDXLGhostStyle:1.0>'

valid crypt Oct 3, 2024, 3:26 PM

#

halcyon quarry Also, the "SDXL Ghost Style" LORA has very cool results <https://civitai.com/mod...

cool

#

sd take the natural language response as prompt?

#

funny

halcyon quarry Oct 3, 2024, 3:49 PM

#

NeuralBeagle is just great all around IMO, good for RP and good for image prompts

#

XL loras don’t work well with Pony and vice versa, also - idk what model you’re trying out

valid crypt Oct 3, 2024, 4:54 PM

#

reload the model and fixed it

#

idk what happened

halcyon quarry Oct 3, 2024, 5:05 PM

#

the progress bar thing? I think its a bit random

valid crypt Oct 3, 2024, 5:36 PM

#

valid crypt reload the model and fixed it

the model was drunk

#

se reloaded it and got normal outputs

#

so

valid crypt Oct 3, 2024, 9:50 PM

#

sullen plover thanks

the update is in main, your models are probably working now

#

i think 3.1 is fully supported but still have to wait for 3.2

terse folio Oct 4, 2024, 12:28 AM

#

halcyon quarry Known issue. This is something that <@226121791670583296> ***might*** be able t...

All seems to work for me.
I'm using a pretty outdated version of stable diffusion though.

not going to update it as i use some extensions that are no longer maintained.

What's the issue you're seeing?

halcyon quarry Oct 4, 2024, 12:38 AM

#

Occasionally, the status bar will only display with 0% or 1% for a moment before the message disappears.
Generation continues and the results post to channel, but the status bar embed gets terminated

terse folio Oct 4, 2024, 12:39 AM

#

my guess is it's requesting status before a job is created in stable diffusion and it quits

#

add some logging and check the time to see what's being sent first

#

asyncio.gather doesn't start the tasks, that's telling the code to wait until all provided tasks complete

so as you're creating the tasks, they're starting in the background out of order (should be expected with async)

so what I would do is create the background task for the progress bar inside the generate image code after the initial request to generate

#

other option (easier) would be to just add an asyncio.sleep(0.1) at the start of the progress bar task

but I would not recommend this as the stable solution.
For example if someone is hosting SD across the network, you will still run into the same timing issue.

#

see if that helps!

halcyon quarry Oct 4, 2024, 2:43 AM

#

I'll merge your PR tomorrow - and hopefully fix this progress bar thing... it's been bugging me for awhile too

terse folio Oct 4, 2024, 2:54 AM

#

reading more into the code, you already handle this by doing 5 retries to try accessing the progress endpoint.
And this information would be logged if failed.

I'm not sure what the issue is now,
Maybe that generation finishes before 1 second, (but this should still log the warning to console)

halcyon quarry Oct 4, 2024, 2:55 AM

#

yep, nothing logs to console when it poofs

#

nah the generation takes considerably longer, when the bar disappears it always reads 0% (or maybe 1%)

#

the embed message goes away entirely

terse folio Oct 4, 2024, 2:59 AM

#

is it possible to run 2 image tasks at once?
Is there a semaphore or queue preventing that somewhere?

#

though, assuming a new instance of Task() is created for each, there should be no conflicts of one task deleting the other's embed by accident

halcyon quarry Oct 4, 2024, 3:01 AM

#

The semaphore is gone - the queue is working great

#

This issue doesn't occur due to queueing, usually happens when nothing queued

terse folio Oct 4, 2024, 3:01 AM

#

does it happen on fresh start?
or after some use?

halcyon quarry Oct 4, 2024, 3:03 AM

#

seems like it does like to happen after an image model change...

#

might be something wonky on SD end

terse folio Oct 4, 2024, 3:03 AM

#

okay, will test that out... might not work outdated

#

the restart_client endpoint was non existant for me!

halcyon quarry Oct 4, 2024, 3:04 AM

#

Requires cmd flag 🙂

#

disabled by default in SD WebUIs

terse folio Oct 4, 2024, 3:07 AM

#

all seems normal, changed models a few times

#

is it a specific model type that has issues?

#

i'm testing with Sd1.4/1.5 based models

halcyon quarry Oct 4, 2024, 3:08 AM

#

Could have something to do with it.... haven't used a 1.5 model in like 1 year

#

XL models

terse folio Oct 4, 2024, 3:08 AM

#

okay, will test that

halcyon quarry Oct 4, 2024, 3:09 AM

#

thanks 🙂 I gotta run for now though before wifey knifes me

#

gnight!

terse folio Oct 4, 2024, 3:11 AM

#

night!

#

sdxl with an llm in vram slows down my gpu quite a lot.
The progress bar on discord was stuck on 1% for a few seconds (longer than the 1 second loop that is happening in the code)

But in the SD console it said it was already at 4%.

So it's possible something is locking up in the code causing a timeout.
That could be on SD webui's side or the bot.

#

The progress bar does work, but slowly

terse folio Oct 4, 2024, 3:32 AM

#

it times out (happened on both tests)
once at 90%, and this time at 12%

#

it's 5 minutes by default

#

because the progress bar uses a different endpoint in a loop, it continues to run even though the bot will never send the image

halcyon quarry Oct 4, 2024, 4:01 AM

#

The issue Marcos and I are having isn’t that though

#

But I have had timeout issue before, not often at all though

terse folio Oct 4, 2024, 4:09 AM

#

Hmmm, I see

halcyon quarry Oct 4, 2024, 4:09 AM

#

Might have something to do with triggering image gen via tags activation

#

“draw… “

#

It creates and queues an image gen task

#

I think for starters I may use embeds.create() for that particular embed

#

Maybe a conflict with another embed same value

terse folio Oct 4, 2024, 4:21 AM

#

the embeds should be unique to the task.
That could be tested by giving tasks a random id to log while different embed actions happen

halcyon quarry Oct 4, 2024, 12:22 PM

#

I clone the task so may not be making a unique Embeds

#

Need to check

halcyon quarry Oct 4, 2024, 12:40 PM

#

In any case I’ll add embeds to the list of attributes which should not be copied

halcyon quarry Oct 4, 2024, 1:20 PM

#

I see what you did there with the first on ready

halcyon quarry Oct 4, 2024, 1:37 PM

#

Cloned Embeds() may have been the issue

#

I prompt for a sexy cheerleader using Flux and she has an engagement ring - I have this intuition that all Flux babes are married by default

#

@valid crypt latest commit I think fixes the progress bar

#

more testing will prove it but so far I'm getting it 100% of the time now

valid crypt Oct 4, 2024, 1:42 PM

#

i see a lot of tech talk

valid crypt Oct 4, 2024, 1:42 PM

#

halcyon quarry <@323088470241312774> latest commit I ***think*** fixes the progress bar

:D

halcyon quarry Oct 4, 2024, 1:45 PM

#

Because I have a setting to disable a few Embed types, and because managing Embeds is a little tedious overall, I try to limit the number of Embed variables the bot creates

#

The original bot code used like 2 main Embeds and recycled them constantly, so if not being careful something like delete(embed1) would delete one you didn't want to delete

#

I resolved a lot of this headache by making each bot Task create its own set of unique Embed variables

#

so like embeds.embed1 from one task is completely different from embeds.embed1 in another task

#

The bot also used to have a different behavior when a text request had a tag triggering it to also generate an image

#

It would be greedy and do both. I changed it to basically duplicate all the variables of the current task, and queue it for image generation

halcyon quarry Oct 4, 2024, 8:36 PM

#

Progress bar still bugged 😦

#

need further investigation

valid crypt Oct 4, 2024, 8:41 PM

#

do you know if dicord bot can identify the speaking user in a voice channel

halcyon quarry Oct 4, 2024, 8:41 PM

#

~~TTS~~ STT not yet implemented

valid crypt Oct 4, 2024, 8:42 PM

#

you mean stt

halcyon quarry Oct 4, 2024, 8:42 PM

#

woops, yep 😛

valid crypt Oct 4, 2024, 8:42 PM

#

is a general question to the discord bot

#

not your

#

but as a discord bot

#

discord

#

bot

halcyon quarry Oct 4, 2024, 8:43 PM

#

Yes, a discord bot could know who spoke

valid crypt Oct 4, 2024, 8:44 PM

#

got a "bug"

#

if the computer losses internet connection tts dies

#

basically disconnects and never conect again

#

toggletts didnt work

valid crypt Oct 4, 2024, 8:46 PM

#

valid crypt if the computer losses internet connection tts dies

im 99% sure about that

halcyon quarry Oct 4, 2024, 8:47 PM

#

But other tasks are working?

valid crypt Oct 4, 2024, 8:47 PM

#

working

#

doing a second test

#

now im not very sure if that was the cause

halcyon quarry Oct 4, 2024, 8:50 PM

#

maybe TTS was not enabled when bot was launched?

valid crypt Oct 4, 2024, 8:50 PM

#

nope

#

let me tell you the whole story

halcyon quarry Oct 4, 2024, 8:52 PM

#

Reality updated on_ready() (function that triggers every time bot client established connection with discord) so that certain functions only run on initial bot startup, and not again after connection hiccups, etc

valid crypt Oct 4, 2024, 8:52 PM

#

i left the bot there for a night, and while i was playing games i lost connection because my internet died for a minute, the next time i looked the bot has disconnected

halcyon quarry Oct 4, 2024, 8:53 PM

#

I'll have to add some print statements pertaining to voice clients and stuff, and see what happens when I toggle internet off then on

valid crypt Oct 4, 2024, 8:53 PM

#

valid crypt funny

then i rebooted the bot, to check if loosing connection of disconnecting from discord causes this bug and i reconnected, and i went for a shower.

valid crypt Oct 4, 2024, 8:54 PM

#

valid crypt if the computer losses internet connection tts dies

eventually i sent this message

halcyon quarry Oct 4, 2024, 8:56 PM

#

I can tell that bot will not automatically rejoin voice channels now with the adjusted code

valid crypt Oct 4, 2024, 8:56 PM

#

?

#

now that i tried to hear, the bot is not speaking

#

yep, after rebooting everything is fine, doing a sencond test

terse folio Oct 4, 2024, 9:08 PM

#

valid crypt i left the bot there for a night, and while i was playing games i lost connectio...

bot will lose connection to discord and leave the channel, that's not related to code.

the voice client for the bot will probably error/cancel.

There might be a background task that does something that also errored when the internet went out.
will have to add some logging to see where and why

valid crypt Oct 4, 2024, 9:08 PM

#

valid crypt now that i tried to hear, the bot is not speaking

i think i introduced another variable, which may be connecting to different server's voice channel or sending message in different server

valid crypt Oct 4, 2024, 9:09 PM

#

terse folio bot will lose connection to discord and leave the channel, that's not related to...

i know that losing connection is not bcs of the bot, but once disconnected it cant connect back wihtou a restart

halcyon quarry Oct 4, 2024, 9:09 PM

#

The bot joins voice channels when it does load_character() but this is now skipped on reconnect

#

You could try using /character

#

While I think about rolling back the on ready changes 🙂

terse folio Oct 4, 2024, 9:10 PM

#

valid crypt i know that losing connection is not bcs of the bot, but once disconnected it ca...

the bot might think it's still connected.

as I was testing voice stuff with my own test bot, reloading the code would cause it to bug a little.
where i would have to make it leave and rejoin the channel

#

did the discord bot commands still work after the internet outage?

how long was it off?
maybe there's a timeout that completely kills the bot?

valid crypt Oct 4, 2024, 9:11 PM

#

terse folio did the discord bot commands still work after the internet outage? how long was...

right now i couldnt recreate but the last time only tts died

#

it could sent text and img

#

and /toggletts didnt bring it back

terse folio Oct 4, 2024, 9:12 PM

#

okay will try to replicate when i have time ^^

halcyon quarry Oct 4, 2024, 9:12 PM

#

Same

#

I think we need init_characters() to always run for on_ready()

#

Or maybe I just need to keep separate tabs on what connections are active, and on subsequent calls to on_ready just reestablish them

valid crypt Oct 4, 2024, 9:20 PM

#

valid crypt i think i introduced another variable, which may be connecting to different serv...

pretty sure that losing connection either disconnects the bot forever or no voice

halcyon quarry Oct 4, 2024, 9:20 PM

#

Reloading all character data n such is a bit heavy handed

#

Did you try… toggle tts twice?

#

Might think it’s still connected on the first one

valid crypt Oct 4, 2024, 9:22 PM

#

halcyon quarry Did you try… toggle tts twice?

for the case of the bot disconnected, i tried

halcyon quarry Oct 4, 2024, 9:23 PM

#

Will just need to make a new little routine for reconnecting

#

I believe all other bot variables stay intact…

#

Just voice clients being problematic

valid crypt Oct 4, 2024, 9:29 PM

#

got it yes

#

a few minutes to get it disconnected

#

got more details

#

once disconected the bot requires a message to reconnect

valid crypt Oct 4, 2024, 9:32 PM

#

valid crypt once disconected the bot requires a message to reconnect

display bot has joinet the chat...

#

bot is ready...

#

it does do the api call to the tts

terse folio Oct 4, 2024, 9:40 PM

#

halcyon quarry I think we need init_characters() to always run for on_ready()

make sure to it doesn't start any while loops.
or if it does, have it cancel previously running background tasks of the same type

halcyon quarry Oct 4, 2024, 10:22 PM

#

Won’t be doing that thought I blurted out (init chars again), will just focus on keeping tabs which VCs connected, if they are disconnected for unexpected reasons will reestablish them in on_ready

valid crypt Oct 5, 2024, 1:41 PM

#

what is the prob of doing a split with \n\n

halcyon quarry Oct 5, 2024, 1:42 PM

#

double

valid crypt Oct 5, 2024, 1:42 PM

#

this is

\n\n

#

?

halcyon quarry Oct 5, 2024, 1:42 PM

#

erm, lemme check...

#

if syntax == '\n\n':
    chance_to_chunk = chance_to_chunk * 1.5

#

So if your value is 0.5 then it would actually be 0.75 if \n\n is detected

halcyon quarry Oct 5, 2024, 1:44 PM

#

valid crypt this is \n\n

Your example is correct

valid crypt Oct 5, 2024, 1:44 PM

#

alr

halcyon quarry Oct 5, 2024, 1:45 PM

#

I'm almost done with this TTS reconnection thing I've actually been working on it the past 30 mins or so

valid crypt Oct 5, 2024, 1:47 PM

#

i might be too lucky to not split 5times in a row

halcyon quarry Oct 5, 2024, 1:50 PM

#

As far as I'm aware, the probability checks are working expectedly. Sometimes the dice just roll that way

#

I fixed the VC connection issue 😄

#

Pushed that

#

As VC connections are managed, it now sets a flag for each one.
If the bot enters on_ready() again (after a connection issue) - it now checks those flags and any known connections. If any connection states are unexpected, it resolves them.

#

While handling this - I noticed that requests (bot tasks) hang while the script tries resolving the main client/discord connection after internet is restored, which actually takes a full minute or so.

I think I need to add some client check before executing tasks, to quickly return something like "Bot is currently disconnected and attempting to reconnect. Please try again shortly" etc

valid crypt Oct 5, 2024, 5:03 PM

#

👌

valid crypt Oct 5, 2024, 9:25 PM

#

guess which one is fake

#

quite easy to guess, but it is really good

terse folio Oct 5, 2024, 9:41 PM

#

they all could be fake, 1 and 0 have that "metalic" sound that comes with tts for some reason.

2, has none of the metallic sound, sounds pretty perfect in terms of quality.
though something sounds odd at 5 seconds: "meeting you (cut?) than I"

2 has some strange intonations/expression
"thank you for always having time to chat with me?"
sounds like a question

valid crypt Oct 5, 2024, 10:33 PM

#

we can guess that it is strange to thank someone for be able to chat or just some human confusion or va confusion

#

although it is processed, here is the true audio

terse folio Oct 5, 2024, 10:51 PM

#

That's pretty cool!

valid crypt Oct 6, 2024, 10:28 AM

#

qwen2.5 is much smarter than llama 3.1

#

chinese stereotype applies to llm :P

terse folio Oct 6, 2024, 10:54 AM

#

I found llama3 and 3.1 to not do well with following instructions properly on their own.
But some finetune makes it better, or perhaps that is because I am using a 5/6 bit quant

Qwen2.5 and Gemma2 have been good for with my testing so far iirc.
I found a really small 2bit quant of gemma2 27b that can barely fit on my gpu!
And it seems a lot more understanding compared to the 12-14b range models.

valid crypt Oct 6, 2024, 11:20 AM

#

try this question
in a family every son has triple sisters than brothers, and every daughter has doble sisters than brothers. how many sons and daughter are there?

#

this is reflection llama 3.1 q2 L though

#

70b

#

this is qwen 32b q6
and its answer is correct

valid crypt Oct 6, 2024, 12:22 PM

#

try this
Kiwi, Tardis and Spaghetti are playing with chips. They have started with 12, 11 and 10 chips respectively and, in each turn, the player who has the most chips distributes one to each of the others and discards another. The game ends when someone runs out of chips. How many turns will the game last?

#

the answer is 28

halcyon quarry Oct 6, 2024, 1:06 PM

#

double brothers

#

and you usually express these kinds of expressions like “three times the number of” or “twice as many”

#

Prompts need improvement IMO unless you expect good results from
bad prompts

#

I wouldn’t trust an LLM with any important math, and these aren’t even logic problems really just confusing math

#

To solve these I’d instead be looking for a coding LLM that can write the code to execute the logic, and output the number of iterations it takes to fulfill the task

valid crypt Oct 6, 2024, 1:47 PM

#

valid crypt try this Kiwi, Tardis and Spaghetti are playing with chips. They have started wi...

this is meant to be calculated by logic and not simulate the process

valid crypt Oct 6, 2024, 1:47 PM

#

halcyon quarry double brothers

not good at english

halcyon quarry Oct 6, 2024, 1:48 PM

#

key word, calculated

valid crypt Oct 6, 2024, 1:48 PM

#

idk how to express

#

the idea is to use the reverse thinking

halcyon quarry Oct 6, 2024, 1:52 PM

#

Most of the LLM testing logic examples I’ve seen are reasoning, and don’t imply “track three variables, on each iteration the variable with the maximum value is reduced by 2 while the others are incremented by one” It’s usually like Jack and Jill go to the park, Jack is sick, Jill is pregnant. Something something who cooks dinner?

valid crypt Oct 6, 2024, 1:57 PM

#

valid crypt try this Kiwi, Tardis and Spaghetti are playing with chips. They have started wi...

a simplified example of that question would be:
In a school chess tournament with 317 participants, each match is a one-on-one elimination, where there are no ties, and a player is eliminated after each loss. Players who win continue to compete against other winners until only one player remains. If a participant has no available opponent in a round, they do not play that round. How many total matches must be played to determine the overall winner?

#

llm tried to simulate the process and gave the wrong answer

#

@halcyon quarry how would you make llm solve:
We have a 3x5 board and a piece placed on one of the squares. On each move, the piece can move to an adjacent square either horizontally or vertically, but not diagonally. From how many squares on the board is it possible to visit all squares within 14 moves?

#

i think that qwen2.5 q4m has the same math level as gpt4o

halcyon quarry Oct 6, 2024, 2:26 PM

#

My point is our use cases don’t involve any math… or does yours?

valid crypt Oct 6, 2024, 2:31 PM

#

halcyon quarry My point is our use cases don’t involve any math… or does yours?

just for fun

halcyon quarry Oct 6, 2024, 2:36 PM

#

I have a puzzle game which involves a game board with a 3x5 tile configuration. The puzzle is solved by placing a piece onto one of the game board tiles, then moving it 14 times to visit every tile. However, the piece cannot be moved vertically. How many solutions are possible for this puzzle game?

#

I don’t see how an LLM is expected to solve something like this unless it’s training data includes variations of the same puzzle with successful responses

valid crypt Oct 6, 2024, 2:43 PM

#

halcyon quarry I don’t see how an LLM is expected to solve something like this unless it’s trai...

that is what LLM should do

#

solve problems that it has never seen before

#

i ve never seen that question, but i can solve it

halcyon quarry Oct 6, 2024, 2:49 PM

#

Uh huh, but LLMs don’t think really. The big boys like ChatGPT likely have additional background processing tools, a miriad of finetuned models working in the background, variables for storing and manipulating data, etc

valid crypt Oct 6, 2024, 2:54 PM

#

😔

halcyon quarry Oct 6, 2024, 2:54 PM

#

Probably runs code with math functions and crap

#

Probably works similar to my Flows tag… background prompt to a specialized context: Determine if prompt is conversational, puzzle, x, y, z…. Answers puzzle. Next bg prompt to specialized context; is the puzzle a riddle, math related, reasoning, combination, etc

valid crypt Oct 6, 2024, 2:56 PM

#

halcyon quarry I have a puzzle game which involves a game board with a 3x5 tile configuration. ...

ps you got the idea wrong and a mistake, although with your prompt it identified that it is a Hamiltonian path like problem

#

do you have access to o1?

#

im curious

valid crypt Oct 6, 2024, 2:57 PM

#

valid crypt do you have access to o1?

preview

halcyon quarry Oct 6, 2024, 2:57 PM

#

I believe so

valid crypt Oct 6, 2024, 2:57 PM

#

valid crypt <@670018869418786816> how would you make llm solve: We have a 3x5 board and a pi...

can you try? plz

#

ppl say that o1 is strong as hll at math beating 90% of dr blahblahblah

halcyon quarry Oct 6, 2024, 2:59 PM

#

says answer is 15

valid crypt Oct 6, 2024, 2:59 PM

#

😢

#

so you are telling me that o1 preview is just full of marketing and lies

#

https://tenor.com/view/genshin-impact-genshin-qiqi-genshin-qiqi-sad-gif-23361676

Tenor

halcyon quarry Oct 6, 2024, 3:01 PM

#

From my question, it says 8 solutions

valid crypt Oct 6, 2024, 3:02 PM

#

halcyon quarry From my question, it says 8 solutions

nice

halcyon quarry Oct 6, 2024, 3:02 PM

#

I had to tweak it slightly because I wrote "vertically" instead of "diagonally" and I also changed my last question to: How many starting positions can result in solving the puzzle?

#

https://chatgpt.com/share/6702a687-7a18-8013-9556-b7a0441e119f

#

From your question copy pasted https://chatgpt.com/share/6702a69f-93b4-8013-8751-778f0479d50c

#

In this case it didn't do any math or anything, training data included specific puzzles, identifying puzzle types, rules of puzzles, etc etc. The answer was already trained into it

valid crypt Oct 6, 2024, 3:09 PM

#

halcyon quarry I had to tweak it slightly because I wrote "vertically" instead of "diagonally" ...

lemme try

#

its because of o1 ._ . chatgpt gave 16 which is absurd and qwen want to code

#

qwen's code gave 0 F

#

try to solve this, it's in spanish because this is tough to express as it is much easier by looking at a img

valid crypt Oct 6, 2024, 3:21 PM

#

valid crypt try to solve this, it's in spanish because this is tough to express as it is muc...

Alargando los lados del triángulo ABC de la figura construimos el A'B'C'. Si sabemos que AC=AC', BA=BA' y CB=CB', Cual es el cociente entre las superficies S(A'B'C')/S(ABC)

valid crypt Oct 6, 2024, 4:31 PM

#

@halcyon quarry are you working on something of the bot?

halcyon quarry Oct 6, 2024, 4:48 PM

#

Not right now, just being a dad

halcyon quarry Oct 6, 2024, 5:22 PM

#

Plans:

add a check for discord client has stable connection
finish persist tag code
finish the SwarmUI support code I began

valid crypt Oct 6, 2024, 5:34 PM

#

i see

halcyon quarry Oct 6, 2024, 6:02 PM

#

remaining items on todo list

valid crypt Oct 6, 2024, 6:10 PM

#

bros prompt is god

valid crypt Oct 6, 2024, 6:11 PM

#

valid crypt bros prompt is god

found a dedicated qwen model for math 🥳

halcyon quarry Oct 6, 2024, 6:13 PM

#

It’s just better clarity on what the situation is, rules and objective

#

A 3x5 board is a piece of lumber

halcyon quarry Oct 6, 2024, 6:33 PM

#

halcyon quarry Plans: - add a check for discord client has stable connection - finish persist t...

Actually this first item isn’t solvable lol Bot would need connection

valid crypt Oct 6, 2024, 6:44 PM

#

why

halcyon quarry Oct 6, 2024, 6:54 PM

#

Eh… idk, need to do some more testing 🤓

valid crypt Oct 6, 2024, 6:55 PM

#

i dont see any logical paradox

halcyon quarry Oct 6, 2024, 6:56 PM

#

It’s odd that the bot will receive requests immediately after internet is restored, but it takes a minute or so before functions accessing the client will work… the paradox is that the bot is getting the requests at all

valid crypt Oct 6, 2024, 6:57 PM

#

you mean that is discord who communicate first so if there is no connection there is nothing

halcyon quarry Oct 6, 2024, 6:58 PM

#

If I disable internet, wait 10 seconds, enable internet, then wait 10 seconds, then send a request to bot it will begin the task then get hung up

valid crypt Oct 6, 2024, 6:59 PM

#

as loosing connection is not very frequent and only tts get impacted, the fix could be manual and be a command

halcyon quarry Oct 6, 2024, 7:00 PM

#

Until a minute or so later when the bot runs on_ready() again meaning it’s established a new connection to discord

valid crypt Oct 6, 2024, 7:06 PM

#

https://tenor.com/view/i-have-faith-james-greer-tom-clancys-jack-ryan-i-believe-in-something-i-trust-in-it-gif-27281977

Tenor

terse folio Oct 6, 2024, 7:23 PM

#

halcyon quarry Most of the LLM testing logic examples I’ve seen are reasoning, and don’t imply ...

yea, it would be better to have the llm code something that solves the solution rather than trying to do it itself.

terse folio Oct 6, 2024, 7:25 PM

#

valid crypt try this question in a family every son has triple sisters than brothers, and ev...

I'm not really talking about math problems, more just sticking to character... simple things! which are surprizingly hard for small models.

I'll try your word problems next time I have it online though

terse folio Oct 6, 2024, 7:27 PM

#

valid crypt found a dedicated qwen model for math 🥳

oh nice, how does that work out?

valid crypt Oct 6, 2024, 7:28 PM

#

terse folio I'm not really talking about math problems, more just sticking to character... s...

you remind me that i have to try qwen 7b

valid crypt Oct 6, 2024, 7:28 PM

#

terse folio oh nice, how does that work out?

idk https://huggingface.co/Qwen/Qwen2.5-Math-7B-Instruct

Qwen/Qwen2.5-Math-7B-Instruct · Hugging Face

#

the gguf is buggy

terse folio Oct 6, 2024, 7:34 PM

#

valid crypt the gguf is buggy

hmm, wonder what causes that.
had some buggy experiences with some models crashing Cuda even!

valid crypt Oct 6, 2024, 7:34 PM

#

with buggy i mean that after certain amount of token it goes crazy

#

mixture of chinese english and symbols and nonsens

terse folio Oct 6, 2024, 7:36 PM

#

yea similar experience

"Illegal memory access" + keeping old prompts in cache somehow.

#

often not following the instructions, and making up nonsense from all the combined chaged prompts

valid crypt Oct 6, 2024, 7:36 PM

#

owww the first one is buggy

#

should have always chosen quantfactory 😔

#

i lied

#

they are buggy

#

so far this is the best prompt to test LLM


Point A' is located along the extension of line AB such that B is the midpoint of AA'.

Point B' is located along the extension of line BC such that C is the midpoint of BB'.

Point C' is located along the extension of line CA such that A is the midpoint of CC'.

This results in a larger triangle A'B'C' surrounding the original triangle ABC.

So what the ratio of the area of both triangles?```

#

it got the answer but then...

terse folio Oct 6, 2024, 8:00 PM

#

hmm, I know the most recent models are supposed to be 8k or higher nowdays.
But this seems like the kind of bug you could have going over the context limit in the past

valid crypt Oct 6, 2024, 11:07 PM

#

i just thought a little bit about stt, and oof idk

#

1.bot must be always listening
2.if we want to make bot reply faster --> few seconds of silence to send the msg
3.beable to cut the response as user might want to add something or others reason like not interested or want the bot to stay silence
4.wrap the user's input if user is trying to add
5.more stuffs if multi speaker...

halcyon quarry Oct 6, 2024, 11:15 PM

#

Will overcome whatever obstacles there are, whenever I'm ambitious enough to try tackling it

valid crypt Oct 6, 2024, 11:16 PM

#

the other side, choosing a good stt that has quality and speed

halcyon quarry Oct 6, 2024, 11:16 PM

#

the bot is already listening for a lot of things

valid crypt Oct 6, 2024, 11:17 PM

#

text is much easier to track as user can think before hitting enter

#

if stt is too sensitive will send un letter as input like a
or a noise

terse folio Oct 6, 2024, 11:18 PM

#

valid crypt 1.bot must be always listening 2.if we want to make bot reply faster --> few sec...

thankfully multi speaker is easy, discord gives you access to separate voice streams.
inference speed is another thing though.
Yes Whisper STT is fast, but to do streaming in almost real time, you have to re-run multiple times on a handful of previous audio each second/time interval you decide on

valid crypt Oct 6, 2024, 11:19 PM

#

terse folio thankfully multi speaker is easy, discord gives you access to separate voice str...

the order of the conversation is important too

#

i think that is better to wrap them

terse folio Oct 6, 2024, 11:19 PM

#

of course!
that's something I haven't thought of yet..

hmmm, I think Its possible as whisper gives you timestamps for text output

valid crypt Oct 6, 2024, 11:25 PM

#

ooo also chucking the voice to not wait for the entire sentence before sending data to the llm

terse folio Oct 6, 2024, 11:26 PM

#

yea, that's a complicated one

valid crypt Oct 6, 2024, 11:26 PM

#

could be by silence like always

terse folio Oct 6, 2024, 11:27 PM

#

you'll want some system prompt that can indicate if a user has been cut off mid sentence

Reality: like --
(later)
Reality: -- this

valid crypt Oct 6, 2024, 11:27 PM

#

and some improvements to the tts, right now i suggest adding a toggle to always chunk the first . or \n...

terse folio Oct 6, 2024, 11:28 PM

#

mhmm, too small of a chunk can cause audio hallucinations

#

if tts tries to just generate a word

valid crypt Oct 6, 2024, 11:28 PM

#

that is why only the first one

#

with the time of the first one, its pretty easy to process the rest in time

terse folio Oct 6, 2024, 11:32 PM

#

yea, that makes sense

valid crypt Oct 6, 2024, 11:36 PM

#

terse folio you'll want some system prompt that can indicate if a user has been cut off mid ...

for a human sometime we have to think for example:
tell me a story about a knight and .... (sent to bot)... and evil magician (interrupt and wrap)

terse folio Oct 6, 2024, 11:36 PM

#

yup

#

to solve this, if the user is still speaking --
have the bot generate slower, like a few tokens at a time and constantly checking for new user inputs

#

so in the end, it might only generate a few words going in the wrong direction, then correcting itself

valid crypt Oct 6, 2024, 11:39 PM

#

as there is a feature about chunking the output text might be a good idea to add it to the prompt like
you have been cut off by the user and you was say "blahblahblah"

valid crypt Oct 6, 2024, 11:40 PM

#

terse folio to solve this, if the user is still speaking -- have the bot generate slower, li...

instead of generate slower it could be only delaying tts

#

https://tenor.com/view/spongebob-long-list-list-infinite-list-gif-7915371

Tenor

terse folio Oct 6, 2024, 11:41 PM

#

getting stt -> text gen tuned is the most important.
because that's your input to output.
Tts is just the method of output.

Like you were saying, if the bot goes generating some thing based on a half finished sentence, it wont matter if tts is delayed, it's already generated the wrong output

#

hmmm speaking of tts,
are there any decent models that can be finetuned/do cloning that don't use 6gb+ of vram!

valid crypt Oct 6, 2024, 11:44 PM

#

i think mine is taking only 1~gb

terse folio Oct 6, 2024, 11:44 PM

#

O:

valid crypt Oct 6, 2024, 11:44 PM

#

https://github.com/Arondight/vits_api_tts

GitHub

GitHub - Arondight/vits_api_tts: This is an extension of text-gener...

This is an extension of text-generation-webui in order to generate audio using vits-simple-api. - Arondight/vits_api_tts

terse folio Oct 6, 2024, 11:44 PM

#

I've been using Xtts at fp32, i think it should be possible to half that, but would have to fork the lib

valid crypt Oct 6, 2024, 11:45 PM

#

mine is not for zero shot

valid crypt Oct 6, 2024, 11:45 PM

#

valid crypt https://github.com/Arondight/vits_api_tts

this is a extension that call https://github.com/Artrajz/vits-simple-api through api

GitHub

GitHub - Artrajz/vits-simple-api: A simple VITS HTTP API, developed...

A simple VITS HTTP API, developed by extending Moegoe with additional features. - Artrajz/vits-simple-api

terse folio Oct 6, 2024, 11:46 PM

#

I used to use a Vits tts a while ago iirc
it was okay, but couldn't pronounce some words, I had to make a lookup table for words that would swap out phonemes before passing them to the tts

#

Will try it out!

valid crypt Oct 6, 2024, 11:47 PM

#

valid crypt this is a extension that call https://github.com/Artrajz/vits-simple-api through...

https://github.com/Plachtaa/VITS-fast-fine-tuning you can train model easily with this

GitHub

GitHub - Plachtaa/VITS-fast-fine-tuning: This repo is a pipeline of...

This repo is a pipeline of VITS finetuning for fast speaker adaptation TTS, and many-to-many voice conversion - Plachtaa/VITS-fast-fine-tuning

terse folio Oct 6, 2024, 11:48 PM

#

That'll be epic if it all goes smoothly!

valid crypt Oct 6, 2024, 11:48 PM

#

valid crypt this is a extension that call https://github.com/Artrajz/vits-simple-api through...

this is not just about vits but a lot of variants, just the extension only can use the basic vits, ill try to make it use others

valid crypt Oct 6, 2024, 11:49 PM

#

valid crypt guess which one is fake

the 1 is gpt sovits, im not sure of how much vram but i dont think it is a lot

#

0 is a zero shot without finetuning

terse folio Oct 6, 2024, 11:49 PM

#

Yea, it was really minimal and could even be ran on cpu at real time

valid crypt Oct 6, 2024, 11:50 PM

#

in my case with cpu i could generate 3~5s in 1s

#

pretty good

halcyon quarry Oct 7, 2024, 2:58 AM

#

This persistent tags... tag... the juice just isn't worth the squeeze

#

I'm not liking all the changes that need to be made to accomodate it

#

eh... idk...

valid crypt Oct 7, 2024, 2:18 PM

#

😔

halcyon quarry Oct 7, 2024, 2:19 PM

#

I'm likely going to add this feature after all, but I'm not thrilled about some things I had to change up

#

debugging one little issue and it should be working expectedly

valid crypt Oct 7, 2024, 2:35 PM

#

i wonder why you are not thrilled about the changes ¯_(ツ)_/¯

halcyon quarry Oct 7, 2024, 2:38 PM

#

"all tags" are compiled for each request. While the code was checking them for "LLM" related tag values, I was removing data that was not applicable in the "Img" matching phase

#

But when a tag is triggered with "persist", I need to ensure that tag is applied again. And not applied twice

#

So for next requests I need to compare tags to any triggered persistent tags and they need to be identical

#

So, I can't remove data from tags while checking them

#

just means it has to iterate over irrelavent data later on

#

aaaaanyway, I have it working now

#

Some additional logic for it...

Only applies to same channel the persistent tag was triggered in
Only re-applies on the same match phase that it was originally triggered in. If it was triggered after LLM phase (post-LLM / pre-Img gen), then it would trigger in the same match phase on the next request

#

@valid crypt if you could test out this branch a little bit I would appreciate it
https://github.com/altoiddealer/ad_discordbot/tree/persistent-tags

I expect that it will work as you had suggested

#

Here's an example tag that uses it:

  - trigger: persist test
    persist: 2
    should_send_text: false

valid crypt Oct 7, 2024, 3:02 PM

#

reading...

#

so you want me to feel if the tag persist

halcyon quarry Oct 7, 2024, 3:07 PM

#

Just change branches for a bit (as am I), let me know if everything is still working expectedly, and if the reason you were interested in having this added is fulfilled

#

I don't think I messed anything up in the process but I had to change a lot of lines to plug this in

valid crypt Oct 7, 2024, 3:12 PM

#

halcyon quarry Here's an example tag that uses it: ``` - trigger: persist test persist: 2...

should_send_text: false?

halcyon quarry Oct 7, 2024, 3:13 PM

#

You could use anything, like prefix_context: etc

#

when match phases occur, it fetches any captured 'persistent tags' that are relavent for the current channel and match phase.
Before it checks a tag for trigger matches, it will first compare to any of those persistent tags, and will automatically "match" it

valid crypt Oct 7, 2024, 3:19 PM

#

i created a new folder with the new branch and i pasted all the config files

#

reboot fixed it

halcyon quarry Oct 7, 2024, 3:25 PM

#

Even if you don't consider yourself a "developer" I recommend getting the GitHub Desktop app

valid crypt Oct 7, 2024, 3:26 PM

#

how so?

halcyon quarry Oct 7, 2024, 3:26 PM

#

To switch branches I just select from this drop down here, and this doesn't affect any user files

#

It will show a little "down arrow" icon next to repositories that have new commits in this dropdown (I already went through and "Pulled" them).

#

So, it's just great for managing any github content on the PC

valid crypt Oct 7, 2024, 3:37 PM

#

pretty much working

#

it told me the prompt after a few interactions

#

and instantly found a bug

halcyon quarry Oct 7, 2024, 3:39 PM

#

lemme guess...

#

it's bypassing censoring?

valid crypt Oct 7, 2024, 3:39 PM

#

valid crypt

hahaha

#

yes

#

and not just the time it got triggered

#

but all the time got bypassed

#

second problem

halcyon quarry Oct 7, 2024, 3:42 PM

#

Apparently I must not realize something about condition logic, I have a bunch of if statements, I'd have thought that since the censor check is higher up that the lower one for if persist would not trigger, but apparently they are all checked simultaneously or something

valid crypt Oct 7, 2024, 3:43 PM

#

valid crypt second problem

first message

#

second

halcyon quarry Oct 7, 2024, 3:43 PM

#

Did it censor the tag at the same time?

valid crypt Oct 7, 2024, 3:43 PM

#

halcyon quarry Did it censor the tag at the same time?

?

halcyon quarry Oct 7, 2024, 3:44 PM

#

valid crypt second problem

When it said "a persistent tag was matched" did it also block that generation?

valid crypt Oct 7, 2024, 3:44 PM

#

valid crypt it told me the prompt after a few interactions

you think so?, although i can check terminal

halcyon quarry Oct 7, 2024, 3:45 PM

#

If it didn't censor that initial tag match then I would be pretty stumped

#

I can see how it could bypass on the next requests

#

(until I fix that... working on it now)

valid crypt Oct 7, 2024, 3:47 PM

#

nothing about blocking of censor,...

halcyon quarry Oct 7, 2024, 3:47 PM

#

request before that one

#

er

#

I see

valid crypt Oct 7, 2024, 3:48 PM

#

before the 5th i reloaded the character to change the trigger word to funtiontest

#

a missing screenshot

#

the forever came out of no where

halcyon quarry Oct 7, 2024, 3:50 PM

#

That's an oversight by me, not actually repeating forever in this case 😛

#

bad print statement

valid crypt Oct 7, 2024, 3:50 PM

#

oki

halcyon quarry Oct 7, 2024, 3:50 PM

#

now this thing with the censoring...

#

You basically created a workaround to avoid it, but I don't think this workaround could be created by user

#

erm

#

OK I know the issue

#

sort of

#

Fixed the "Forever" message

valid crypt Oct 7, 2024, 6:42 PM

#

https://www.ebay.com/itm/256533572643?itmmeta=01J9KSMKXAMQPN2957MTBTW9W2&hash=item3bba97ac23:g:bH8AAOSwgFBl5dm6 💀

eBay

2080 Ti 22G Memory. Custom Upgraded. Stable Diffusion, LLAMA2, Deep...

Why do you need a 22Gb 2080ti?. Large language model. Note: you do not need an upgraded 2080ti if you only want to do regular gaming. This upgraded GPU is for deep learning. Perfect for deep learning training!

#

💀 💀 💀 💀

valid crypt Oct 7, 2024, 9:36 PM

#

do you use the default windows file explorer?

#

is painful to me to wait for the long loading time and the poor search function...

valid crypt Oct 7, 2024, 10:40 PM

#

ive just come up with an idea, ai alarm or whatever, idk

tag to create tags
2.llm creates tags
3.tags that expires by rl time
a system that sends the current date and time when the date is reached using tags or something

by doing this we can connect the rl to the bot and maybe can help its other aspects, idk

terse folio Oct 7, 2024, 10:56 PM

#

valid crypt https://www.ebay.com/itm/256533572643?itmmeta=01J9KSMKXAMQPN2957MTBTW9W2&hash=it...

I saw something about some people in China modding their 2080tis to double the vram.
Didn't expect to see them sold world wide now!

I wonder if it's worth it, (I have a 2080ti and it's still really fast for all these models nowdays)
But is the inherent risk of Ebay worth it?
That would be cool though.

valid crypt Oct 7, 2024, 11:02 PM

#

4060ti 16gb has 6gb less vram but risk free https://www.newegg.com/msi-rtx-4060-ti-ventus-2x-black-16g-oc-nvidia-geforce-rtx-4060-ti-16gb-gddr6/p/N82E16814137818

halcyon quarry Oct 7, 2024, 11:02 PM

#

Were the censor tags not working at all on the persist branc?

valid crypt Oct 7, 2024, 11:03 PM

#

halcyon quarry Were the censor tags not working at all on the persist branc?

if you didnt update i say no

valid crypt Oct 7, 2024, 11:04 PM

#

valid crypt 4060ti 16gb has 6gb less vram but risk free https://www.newegg.com/msi-rtx-4060-...

ofc 100$ cheaper

terse folio Oct 7, 2024, 11:08 PM

#

valid crypt ofc 100$ cheaper

yea, good point,
And model splitting across gpus has probably gotten better too!

valid crypt Oct 7, 2024, 11:09 PM

#

terse folio yea, good point, And model splitting across gpus has probably gotten better too!

for tgwui yes, its fantastic

#

some motherboard support splitting the x16 pcie into 2x8 which means you can plug 2 4060ti 16gb without bottleneck, and maybe another 4060ti 16gb in the x4 from chipset...

#

to go beyond the price goes

#

https://tenor.com/view/floki-to-the-moon-floki-rocket-flokiviking-to-the-moon-gif-25363840

Tenor

valid crypt Oct 7, 2024, 11:37 PM

#

gonna leave my question here, how do i load transformer in 4bit, checked the option in tgwui and saved... (rn im getting cuda out memory (bot) and tgwui does not)

halcyon quarry Oct 7, 2024, 11:37 PM

#

#

expected behavior...

#

unsure what was happening with your testing earlier

terse folio Oct 7, 2024, 11:39 PM

#

valid crypt some motherboard support splitting the x16 pcie into 2x8 which means you can plu...

Might have to try that!
Got a new powersupply with lots of overhead for the 2080ti a while back.
And the 4060ti draws a lot less power, almost half.
It would be realistic to be able to fit a few in there without further modifications!

terse folio Oct 7, 2024, 11:40 PM

#

halcyon quarry

speaking on censoring,
have you see the censoring code in the openai-api extension?

They load up an embedding model and compare the embedding of the text to some swear words.
(ofc this could be improved upon)
But maybe it would be a good idea for more fluid censoring!

halcyon quarry Oct 7, 2024, 11:41 PM

#

I merged the persistent tags to MAIN

#

You could also use Flows tag to have a specialized character review the text, and return "true" or "false"

halcyon quarry Oct 7, 2024, 11:44 PM

#

valid crypt ive just come up with an idea, ai alarm or whatever, idk 1. tag to create tags 2...

Thanks for the suggestions!

#

Some good ones

terse folio Oct 8, 2024, 5:10 AM

#

halcyon quarry Apparently I must not realize something about condition logic, I have a bunch of...

The current code in collect_llm_tag_values looks like it should be working fine.

The if statement raises an exception if censoring is detected which will stop the loop.

What I did notice though, is the function iterates through a list of tags.
Perhaps there's something that happens before that is not censored but set to persist, then the next tag is censored and exits the loop.

valid crypt Oct 8, 2024, 2:44 PM

#

valid crypt gonna leave my question here, how do i load transformer in 4bit, checked the opt...

?

terse folio Oct 8, 2024, 2:50 PM

#

valid crypt ?

I rarely use transformers.
I think even for LLava I was using gptq (which is slow with recent updates)

there is probably a way to quantized the transformers model to 4bit before hand.
Perhaps tgwui tries to load everything into ram/vram?
Maybe it can quantize on the fly... not sure.

Iirc quantizing to 4bit was a time consuming process.
Long ago I was manually quantizing my own llama1 models because the gptq ones I downloaded kept failing for some reason.
(That was probably for a different loader anyway, it could be faster with transformers!)

valid crypt Oct 8, 2024, 2:51 PM

#

in tgwui there is a option to load in 4bit

#

the problem is that using the qwen math gguf i encountered some problems but with 4bit transformer didnt

terse folio Oct 8, 2024, 2:53 PM

#

Hmmm, maybe you should ask in the general/help chat here, or maybe the transformers github, it could be a them issue

valid crypt Oct 8, 2024, 2:53 PM

#

the thing is the bot is not reading the setting i think

#

i thought the bot could load in 4bit as tgwui can

terse folio Oct 8, 2024, 2:53 PM

#

try loading it without the bot and hitting save settings, the loading with the bot

valid crypt Oct 8, 2024, 2:54 PM

#

that is what i did

terse folio Oct 8, 2024, 2:54 PM

#

Hm!

valid crypt Oct 8, 2024, 2:54 PM

#

bot loaded it in normal

#

and mi little laptop cant load it

#

if it isnt in 4bit

terse folio Oct 8, 2024, 2:55 PM

#

There are 2 cmd_flag text files.

One for tgwui, and one for the bot (that is also for tgwui)

I think... i might be remembering it wrong because I sometimes keep a backup of a file in another dir.

But make sure you put the start flag in the right file ^^

valid crypt Oct 8, 2024, 3:01 PM

#

i simply do /llm

#

and bot loads the model in full weights

#

;-;

#

this is what i have

valid crypt Oct 8, 2024, 3:06 PM

#

halcyon quarry

the llm side

#

the censor works on the user side

halcyon quarry Oct 8, 2024, 3:07 PM

#

I had merged another branch in that may have reverted the code... looking into it now'

#

ok... on the LLM side, llm_censoring tag will only have an effect if it includes triggers

#

oh ok I see the issue...

#

sort of...

terse folio Oct 8, 2024, 3:11 PM

#

valid crypt i simply do /llm

Ahhh, I never really did that.
I have a dedicated 3b model to use for testing the bot that I put in the cmd_flags.txt
That's so I can also load SD or TTS while debugging

halcyon quarry Oct 8, 2024, 3:11 PM

#

actually no I don't see what the issue is

terse folio Oct 8, 2024, 3:11 PM

#

maybe they use different functions to load the model

valid crypt Oct 8, 2024, 3:12 PM

#

halcyon quarry actually no I don't see what the issue is

try it yourself

#

suffix with a word that is prohibited and make it say it

valid crypt Oct 8, 2024, 3:13 PM

#

terse folio maybe they use different functions to load the model

im not sure if the cmdflag would change anything

#

my conversation

halcyon quarry Oct 8, 2024, 3:15 PM

#

Before the LLM sends any messages, it runs that text through a check_censored() function.
What it does is checks if any triggers from any llm_censoring tags match anywhere in the text

#

Ok here is the problem here

#

the trigger regex currently looks for word boundaries only

#

I'm going to update the pattern to include non-word characters such as "xxx" or xxx! or xxx? or 'xxx' etc

terse folio Oct 8, 2024, 3:25 PM

#

censoring is a hard battle, people will xx* or xxxtest or x x x

#

even other lookalike unicode characters 𝓍𝓍𝓍

#

yes there are ways to santize unicode lookalikes to their ascii versions.

#

But don't stress too hard about actual censoring for like the public use!

But yea, I can see how it's a great tool for getting the bot to avoid talking about something if people aren't actively trying to break it 😸

valid crypt Oct 8, 2024, 3:28 PM

#

valid crypt it told me the prompt after a few interactions

im not very sure

halcyon quarry Oct 8, 2024, 3:28 PM

#

yeah something seems broken after all... wtf I know it was working...

valid crypt Oct 8, 2024, 3:29 PM

#

terse folio But don't stress too hard about actual censoring for like the public use! But y...

people can 𝓍𝓍𝓍 but bot is harder

#

just in case, caps :)

halcyon quarry Oct 8, 2024, 3:33 PM

#

found the bug, but the regex was another thing for sure

#

really found it this time 😛

#

code was not actually collecting the "llm_censoring" tags to check later.
I may have had it working, then tweaked it in a way that I thought it would still work but actually stopped working

#

But now its working.

#

In the process noticed that /regenerate does not handle the censoring gracefully - fixing that

halcyon quarry Oct 8, 2024, 3:59 PM

#

@valid crypt The censoring should now be working great on MAIN

valid crypt Oct 8, 2024, 4:24 PM

#

ok

valid crypt Oct 8, 2024, 4:25 PM

#

valid crypt gonna leave my question here, how do i load transformer in 4bit, checked the opt...

do you know something about this?

halcyon quarry Oct 8, 2024, 4:29 PM

#

Are you using the same exact loader?
When I search the TGWUI code for 'load_in_4bit' I see that this value is only recognized in a function called huggingface_loader()

valid crypt Oct 8, 2024, 4:30 PM

#

😵‍💫

halcyon quarry Oct 8, 2024, 4:30 PM

#

I see that Transformers will use this function

valid crypt Oct 8, 2024, 4:30 PM

#

what

valid crypt Oct 8, 2024, 4:31 PM

#

halcyon quarry Are you using the same exact loader? When I search the TGWUI code for 'load_in_4...

can you explain a little bit more

#

i just /llm to load it

halcyon quarry Oct 8, 2024, 4:31 PM

#

Are you sure that the bot is using loader "Transformers" - same as what you are using in the UI?

#

Check your models/config-user.yaml file and ensure that the model in question has these values:
loader: Transformers
load_in_4bit: true

valid crypt Oct 8, 2024, 4:32 PM

#

you somehow fixed it...

#

solved :)

#

im pretty sure that yesterday there wasnt so many parameters...

#

¯_(ツ)_/¯

halcyon quarry Oct 8, 2024, 4:35 PM

#

I think when you click save config, it saves all relavent settings (including defaults)

#

and if you have a user config it will print all

visual dagger Oct 9, 2024, 3:55 AM

#

@halcyon quarry how are you doing and how is the project going?

halcyon quarry Oct 9, 2024, 12:28 PM

#

visual dagger <@670018869418786816> how are you doing and how is the project going?

Good morning thanks for asking the project and myself are both doing quite well

#

Over the past few months I was able to add a few features that I thought would never see the light of day most of which are pertaining to using the bot on multiple servers

#

I didn’t think I would be able to figure out TTS streaming but alas that is also working

#

Development has slowed obit since I am distracted by a game and also due to my contributing towards the SD web UI Forge Project

halcyon quarry Oct 9, 2024, 1:39 PM

#

Pushed minor update, include new "XTC" sampler param defaults

visual dagger Oct 9, 2024, 3:17 PM

#

@halcyon quarry good luck man

#

so the tts streaming is implemented already? and what's the logic behind it?

#

and is it real time?

#

sorry for bombing you with qs haha

halcyon quarry Oct 9, 2024, 3:20 PM

#

visual dagger so the tts streaming is implemented already? and what's the logic behind it?

I modified the TGWUI chat wrapper function, to extract the part which would trigger TTS generation only at the end of the generation

#

Now when the bot is triggered to stream a text response (on a sentence completeion, double line break etc) via that feature - it can now trigger TTS gen on that text

visual dagger Oct 9, 2024, 3:22 PM

#

so it's not real time?

#

maybe you can try triggering the tts on ",. : ;". when it detects a new sentence basically.

halcyon quarry Oct 9, 2024, 3:45 PM

#

visual dagger maybe you can try triggering the tts on ",. : ;". when it detects a new sentenc...

It's user configurable

visual dagger Oct 9, 2024, 3:45 PM

#

is it possible to make it real time tts

#

as it is generating the text

#

not waiting for a full sentence

halcyon quarry Oct 9, 2024, 3:48 PM

#

if you split text on " " space then maybe

#ad_discordbot (Fork of Fork of xNul's bot)

Pushed that

Pushed minor update, include new "XTC" sampler param defaults