Made flux generation times better for Dante (60s->36s and 540s->300s) and improved maw. Maw can now auto-join voice chats when someone joins and types while sending a reply. As always, everything runs entirely locally.
made a gui for 20 questions in Java!
Added a new model, Flux (both dev and schnell), to Dante. This ones really good (though also pretty slow). As always, everything runs locally. Also, I made a mini space sim! But its only 3 sessions so I don't think it deserves its own scrapbook post
Maw now works in VC! Ask it questions and it will give you answers. You can even ask it to make images
It's obviously a little buggy right now but it's getting there!
As always, everything runs locallymade my portfolio website! poke around a bit!
ednamode.xyz (www.ednamode.xyz doesn't work right now, weird certificate stuff. trying to get it fixed soon)
github.com/Ednaordinary/Ednaordinary.github.io
I switched maw to llama 3.1 8b and in the process switched engines! The new engine is more than twice as fast in the majority of scenarios, and I was able to fit the entire 128000k context window in my gpu (24gb) with it! The old engine was barely able to fit 8k, and was already topping out my vram
nickolas for discord! It's currently incomplete but I just reached the point where I've successfully made a model from a command. you upload a video and it makes a 3D scene that can be viewed in your browser!!
did a bunch of bugfixing! i dont remember half of what I did though. What I do remember is a bunch of work on how maw characters work and adding a proof of concept for a new transformers engine. also added DM usage and some integrations like dante and pings
added some bug fixes to a bunch of stuff, mostly EdnaMusic and Maw. New maw features for:
• editing system prompt and environment prompt
• locking characters
• button persistence (they still work after the bot restarts)
there's some stuff for EdnaMusic too but I don't remember, it was a little while ago
This is a collection of bug fixes that I didn't think were important enough individually to be a scrapbook post. Attached is an image of one of my favorite new fixes/features: buttons on EdnaMusic individual song messages. the "next" button stays on the currently playing song, while the "remove" button is only on songs playing in the future
All fixes: started work on letting maw understand images, added a command to remove context from maw for the server, buttons on individual ednamusic songs, a fix (or two?) for cross service model manager, and possible some other fixes I forgot about.
Cross Service Model Manager
github.com/Ednaordinary/ModelManager
What is this? I run 3 bots that use cuda based models (all are open source). Unfortunately, that means they clash when multiple try to use memory at once. This fixes that by ensuring only one bots models run at a time
(see hackclub.slack.com/archives/C01504DCLVD/p1720075345755789)
With Dante4 comes a new experience: DanteLive
DanteLive returns images to you as you type, letting you iterate instantly instead of sending a bunch of commandsgithub.com/Ednaordinary/DanteMode
This is Dante4, my 4th revision of a discord bot I've been working on for the same amount of years. Dante was born before image generation became popular, when I found a small repo of an emerging technology, lucidrains/big-sleep. Since then, Dante has grown with the technology into what is now an extremely optimized discord bot for both loading and execution. As I moved on from Dante3, I focused on a few things: maintaining modularity, increasing performance, and making something cool. Dante now features the following features to meet these goals:
An asynchronous framework: Just like Dante3, Dante4 handles requests asynchronously between request (discord command), and execution (making the image). This ensures the user will never be left hanging during the process. Only now, Dante takes that to the next level by adding a third layer in there: Model loading. More on that below
Model Passthrough: If a request from discord comes that shares the same model as the currently executing request, that model passes itself to the next request, completely eliminating model loading latency
Model Preloading: If a request from discord comes with a different model than the one currently executing, that model will preload itself to the cpu (ram), then load itself quickly to the gpu (vram) when its time for execution. This reduces model loading latency since the model does not have to load itself from disk.
Latent Previewing: During a request, the bot will asynchronously decode the current image to show to the user, so they know where in the generating process the bot is.
Embedded Processing: Models are run right in the script, eliminating import time for large modules like torch. With this comes a modular class based system, which allows all of the other optimizations to happen.
Live Status: Models, both with and without latent previews, will update the user with the current generation status with both percents and seconds since start.
Exception Handling: Just like Dante3, even if something goes incredibly wrong, the bot recovers and continues working.
This isn't all though, more in a second.
Maw for discord
Both a chatbot and a character messager. It allows you to redo message generations and edit character messages, while dynamically making sure the right message has the right buttons.
github.com/Ednaordinary/MawDiscord
My chatbot now has terminal colors, emotion, and different modes for higher quality models (at the cost of latency)Music bot for discord! It supports multiple servers, both spotify and youtube, playlists from both, and is discord thread based
github.com/Ednaordinary/EdnaMusicReverse (semantic) image search in discord! Find pictures from your server with just a text prompt.
This milestone gets it to the 1.0 state and does the following:
• Terminal colors for the indexer
• Fix a bunch of model loading stuff
• Allow the indexer to resume after bot downtime
• Increase indexer thread count
• Add event handlers for everything (new messages, edits, deletions post bot start)
• Add message links for searched images
github.com/Ednaordinary/Vesta/commit/7e0c5af8436ac8d3302d79c2e001db20ca8cb0c4
Update to Vesta: my semantic image search for discord servers bot.
It is now way faster at scraping discord attachments and uses a larger and better model. No image ever touches storage (SSD), staying in ram while it is processed. This makes it extremely private as only semantic embeddings are saved to storage. The scraper now uses many threads to process everything as fast as possible, getting as many semantic embeddings as it can at onceSemantic image search in discord: search every image in a server with a single text prompt. This is an initial commit with a "working" state that could be better, I will be improving it. github.com/Ednaordinary/Vesta
Finished adding the initial llama model itself to the chatbot. Doesn't look too impressive from a user perspective, but it's handled in a multithreaded way
