-
Notifications
You must be signed in to change notification settings - Fork 40
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
feat: latest torch/comfyui; perf improvments; fix: SSL cert issues #309
base: main
Are you sure you want to change the base?
Conversation
This already functionally happens in hordelib. We can use the bytestream passed from it directly.
This is tied to a bug of unclear root-cause but whose practical effect is that the root signing certificate was not being found on a relatively fresh windows 10 pro machine. `certifi` should already be being pulled in, but I've marked it as an explicit requirement and I anticipate that there should not be side effects on machines which were previously running fine.
The new config option `very_fast_disk_mode` overrides this.
The recent change to preloading being single-model only causes the logic here to think that the model isn't loaded when it should be. If there is one holding up the line, we'll wait to see what happens.
@CodiumAI-Agent /describe |
Titlefeat: latest torch/comfyui; perf improvments; fix: SSL cert issues User descriptiontodo PR Typeenhancement, bug fix, documentation Description
Changes walkthrough 📝
|
This is response to some observed behavior that control+c exiting would lead to all processes getting killed before all jobs were finished (with threads=2)
This was incomplete and masked by 22.04 also using 3.11 as it's Python version. python3.11-dev is technically optional, but included because it's needed on the AMD side. venv creation HAS to be called with the full version, otherwise the dist default is used. pip only needs to be updated inside the venv.
New Features/Updates
very_fast_disk_mode
configuration option for concurrent model loading.false
.very_fast_disk_mode: false
to only load one model at a time when it is being explicitly preloaded. There are some cases where it still might attempt to load more than one but it should be far less often.3bb4dec
; torch >=2.4.1 and cu124 by default hordelib#343Fixes and Improvements
high_performance_mode
.max_threads
values greater than one.max_threads: 2
and a bit of tuning.high_performance_mode
if you have a xx90 card.high_memory_mode
can still lead to additional instability with threads at 2.max_threads: 2
in SD1.5-only setups without controlnets/post-processing or in other conservative configurations.rawpng
directly, reducing redundant operations.PIL.Image.open(...)
was highly inefficient, especially for very large images.certifi
to resolve certificate resolution issues.download_models.py
would not exit if the compvis models failed to download. This would cause the worker to crash unexpectedly as it expects the models to be available on worker start.