Open raldone01 opened 3 weeks ago
Can confirm. The real path that gets generated after conversion is based on the full path to the original checkpoint and does not contain the UUID(?) that the conversion script seems to be looking for.
A workaround is to convert, let it fail with the above error, then manually move the converted checkpoint folder to the normal location where Invoke places models of that type, then tell Invokeai to scan for models within its root, and finally install the model from its new location. The model appears to work like diffuser models normally work after that point.
I just downgrade invoke to convert diffusers. Just have to fix the schema version in the invoke.yml
. I had no issues with DB migrations so far...
Operating system
Linux (docker)
GPU vendor
Nvidia (CUDA)
Version number
a3cb5da130280d380cbb7bf4492f8cf50ebd060b
Browser
Firefox
What happened
When I try to convert one of my models to diffusers in the model manager it fails because there seems to be a mismatch in the convert_cache naming. The conversion seems to work.
However the model manager is looking for
/invokeai/models/.convert_cache/8b1c8220-9abd-481f-9827-be64ec67461a
which does not exist.I patched
invokeai/app/api/routers/model_manager.py
with the following:Note: Converting with older releases works fine.