Hi, I don't use Lora.jl, but opening up the repo on GitHub, I see the tests on master are failing. It's probably a good idea to stick with the latest tagged release.
The README also has this warning: "Lora has undergone a major upgrade. Some of its recent changes include: Models are represented internally by graphs. Memory allocation and garbage collection have been reduced by using mutating functions associated with targets. It is possible to select storing output in memory or in file at runtime. Automatic differentiation is available allowing to choose between forward mode and reverse mode (the latter relying on source transformation). To run the current version of Lora, it is needed to Pkg.checkout() both Lora and ReverseDiffSource. Both packages will be pushed to METADATA very soon. Some of the old one has not been fully ported. The full porting of old functionality, as well as further developments, will be completed shortly. Progress is being tracked systematically via issues and milestones. The documentation is out of date, but will be brought up-to-date fairly soon. In the meantime, this README file provides a few examples of the new interface, explaining how to get up to speed with the new face of Lora." Hope this helps.
