Closed eetgeenappels closed 1 year ago
The developers are aware of this issue. Thank you for including the crash report. That info is helpful to the team.
Here is some useful info:
A build at this commit appears stable: 09 July - export sort option strings for localization e2ec72b49a38ef2c9226de35772fe9fc5cb70ad0
A build at the next commit crashes when I change the model selection in the sidebar: 12 July - Update packages e6e79cd0841867edcff9e55f36fff8187bacc613
The "Update packages" commit was only a change to the package dependency hash for the ml-stable-diffusion build. There were no changes in the Mochi code. (There was a Crowdin and a I10n update in betwen, but I don't think those are relevant here)
The screen cap is the Xcode display at the crash.
Tomorrow, I will try to establish the Swift files that were changed at ml-stable-diffusion between the two hashes listed for that package in Mochi. I'm pretty sure there are only going to be 2 possible culprit files. If someone beats me to this next step, that's fine.
This is the diff for the 2 versions of ml-stable-diffusion involved here:
"identity" : "ml-stable-diffusion", "kind" : "remoteSourceControl", "location" : "https://github.com/apple/ml-stable-diffusion", "state" : { "branch" : "main", "revision" : "e3875a584b173774357c76beda659df9f62527ec" ## changed from "revision" : "7f9c58a6c1483a6da1f398cad49eee2f9fa3bee4" ## changed to
ml-stable-diffusion had just that single commit change between the hash we used that was stable, and the hash we used that introduced the crash. ("changed from" ---> "changed to").
So the error has to be from that commit:
7f9c58a6c1483a6da1f398cad49eee2f9fa3bee4 WWDC23 update This changed: 16 files, screen cap attached.
More tomorrow.
@jrittvo i think its some thing about self.pipeline because i added a line before pipline init and set self.pipeline = nil And also crashed on this line. It can be some bug of asynchronization or something about threads.
Yes. @vzsg has been saying that there are async glitches in Mochi for a while. There have been random model load crashes coming and going for a few months. And when I was experimenting with the new split-einsum-v2 models a month or so ago, there was a intermittent crash in Mochi where it tried to unload a model that had just finished unloading. So that whole set of transitions needs looking at.
But some change in ml-stable-difusion seems to make the bug present all over the interface now, when it used to be pretty rare before.
And honestly, I am in way over my head at this point. I was just trying to follow calls from one place in the code to another a few minutes ago, and I can't make sense of it. I don't know what I'm looking at half the time when it gets down to that level.
In brief testing, this is indeed fixed by #283.
Running Latest Version
Processor
M1 (or later)
Intel processor
No response
Memory
8GB
What happened?
Go to the setting of the app and click apply
Crash Log