Close to a month an a half ago, our last remaining maintainer for the nataili library dropped out and we were left functional but “rudderless” as far as inference goes. We could continue operations, but we couldn’t onboard new features anymore as neither me nor any of the remaining regulars have ML knowledge.
In desperation, I asked one of our regulars, Jug, who had been helping out with some python work on the worker if he thinks it would be possible to switch to the ComfyUI software as a backend, as it had some good ideas and was modular enough to be of use to us
To my surprise Jug not only thought it was a good idea, but jumped with both legs in the deep and started hacking around to make it work. Not only that, but we managed to suck-in another regular developer in, Tazlin, who also started helping us with design best practices. As a result, the new library we started developing was built from the ground up to have extensive coverage support which will make us discover regression bugs that much easier.
First steps were to develop feature parity, and that required not only to wrangle the comfyUI pipelines to be called from nataili, but also to port features which we were using in the AI Horde Worker, such as clip, over to the comfyUI.
This early phase was were I could still provide some help, as I’m pretty good at porting features and writing tests for them, and then integrating stuff into the AI Horde Worker, but still the lion’s share of the work on hordelib was being done by Jug, with Tazlin making the code much more reliable and maintainable.
A couple of weeks in, we had almost all the features we needed, but this is where the tricky business started. First we noticed is that comfyUI was not handling multi-threading well, which make sense as it’s meant to be used by a single user on a single PC. That added massive amounts of instability, because our AI Horde Worker is using threads for everything, to nullify latency delays.
So the next phase for about two more weeks was stabilizing the thing, which required a much deeper dig into the comfyUI internals to wrangle individual processes into a multi-threaded paradigm.
Finally that was done, about 1 month after I inquired about moving to comfy. Then we discovered the next problem: Due to all the mutex locks to prevent multi-threaded instability, the whole things was now much slower than nataili was. Like significantly so!
So another two weeks were spend of figuring out where to slowdowns occurred in our implementation and tweaking things to work more optimally, and even trying to figure out if there was indeed a slowdown in the first place as comparisons with the nataili was difficult to achieve.
We even built a whole benchmark suite to see overall speeds in inference, without getting confused with HTML and model loading latency.
But beta testers were still informing us of a seemingly lower kudos reward, so then we suspected the old way of calculating kudos was not applying well to the hordelib inference, due to it working differently. For example it has no slowdown for weights, but control-net types gave out different speeds than we expected, even different speeds per control type.
To track this down, Jug trained a new Neural Network for figuring how much time a generation is expected to take, rather than try to time each individual feature. The new model was so successful at 96% accuracy, that we decided to onboard it onto the AI Horde itself, as a way to calculate kudos more accurately.
This investigation did point us to some things that worked unexpectedly within comfyUI, for example longer prompts than 77 tokens tended to be quite slower, which was a quality thing after speaking with the comfyUI devs. We did discover a workaround for the AI Horde but it’s these sort of things that are introducing unexpected slowdowns compared to before. We’re going to continue looking for and tweaking things as we discover them.
The good news is that the overall quality of images using the comfyUI branch has increased across the board. Not only that but weights not only don’t add extra slowdown (so the extra kudos cost is removed), but they can also exceed 1.3 without causing the image to distort, which is how most other UIs are using them anyway.
The big change is that images with the same payload and the same seed, will look different in comfyUI compared to nataili. This is simply due to the way inference works and something we’ll have to live with.
1.0.0
So now we have the three pillars built: Parity, Stability and Speed; it’s time to go live!
The hordelib has been bumped to 1.0.0 and the AI Horde Worker to 21.0.0. When you run update-runtime
next time, you’ll automatically be switched to the new inference backend but you may need to update your bridgeData.yaml
file ahead of time.
Very shortly
- Set the
vram_to_leave_free
andram_to_leave_free
to values that work for you. - rename
nataili_cache_home
to cache_home - You can delete any unused keys (like
disable_voodoo
)
Also as a user of the AI Horde, keep in mind that the new Workers do not yet support tiling
and pix2pix
But not only if the new inference available for the AI Horde, but also for everyone else. Due to the generic way we’ve built it, any python project which needs access to image generation can now import hordelib from pypi, and get access to all the multi-threaded text2img and img2img functionality we provide!
What’s next
With the move to hordelib, we are now effectively outsourcing our inference development upstream, which allow us to get to use new developments in stable diffusion as they get on-boarded into their software. Hopefully development of ComfyUI will continue for the foreseeable future as I am really not looking forward to changing libraries again any time soon >_<
This also means that we now finally have the capability to onboard LoRas and textual inversion as well which have been requested for a long time, but we never had the capability in our backend. Likewise with new Stable Diffusion models and all the exciting new developments happening practically weekly.
It’s been a lot of hard work, but we’re coming out of it stronger than ever, thanks to the invaluable help of Jug, Tazlin and the rest of the AI Horde community!
2 thoughts on “The AI Horde Worker Moves to a Completely New Inference Backend”
Comments are closed.