#custom model not loading
14 messages · Page 1 of 1 (latest)
sorry, didn't see your question here. The specific model was a two level mix following the recipe for "berrymix". the first step used 3 models merged via auto1111's checkpoint merger and the "add difference" option. the second layer took that model and merged it 50/50 with a fourth model using "weighted sum". The FP16 box was not checked
i suspect it may have something to do with whether or not the FP16 option is selected
one of the first models I got to work was my Lead Sled model, and I know I had used the FP16 option when creating that one.
I'll see if I can ferret that one out.
@silver orbit also, let me know if this reply got to you, I deleted the original topic as it got answered in another thread as to how to fix it.
Ok I’ve got it now. Thank you!
Ok, an update: It does not seem to be related to the use of the 'save as float 16" checkbox. I merged two models using weighted sum, one with f16, one without, and both get the "state_dict" error. The merge was between 1.5 and Hassan 3 in both instances. Could be a change to the checkpoint merger in the current version of auto1111, as the models I have that don't get the state_dict error were all merged a while ago, or were the results of custom training.
It looks to me like auto1111's merge process changes the structure of the .ckpt file. Is there a way you could post one of the non-working files so that I can make compensatory code changes?
I cannot copy (or even really look) at auto1111's code because of the licensing issues in his repo, so I need to reverse engineer.
hm. I will see if I can get a .ckpt that is small but also causes the error
well, I'm dead in the water on this for a bit. Latest Auto1111 update apparently attempted to address something related to this issue, as I'm getting the "state_dict" error on all my models there now, and, unlike Invoke AI, this causes it to fail to launch completely.
Ok, I think I've got Auto1111 running again. Really appreciate that Invoke is a bit better when encountering model issues in that it reverts to a model that it knows will work, instead of just crashing/freezing.
anyway, do you still need testing, or would you prefer I wait until 2.2.4 is released?