Handling of .alpha in from-comfy converters

Describe the bug

This is a question and maybe a bug report:
LoRAs in diffusers/peft-format don't support an alpha scaling factor. when converting another format to diffusers/peft, .alpha can be folded into lora_B]

kohya format LoRAs do hgave .alpha and for those models that diffusers implement a from-kohya converter, .alpha is supported.

For recent models, diffusers seems to have switched to supporting comfy format rather than kohya - which is mostly: diffusion_model prefix, original model keys.
And there is the issue: some from-comfy converters in diffusers have .alpha supports, others don't.

The table below is AI generated from static analysis so I cannot guarantee its correctness, but I tripped over this for Anima comfy-format LoRAs, so that one I can confirm.

And here it becomes a question: are these converters supposed to support .alpha, or do you consider comfy-format to be without alpha?
Con: comfy-format isn't really defined. it's just whatever they support, and that varies wildly between models. They do support .alpha for all models though.
Pro: it's in use. For example, the kohya-to-comfy converter that ships with kohya-ss does output .alpha

Reproduction

Theoretical reprodution:
Load a LoRA generated by kohya-ss, converted to comfy-format by their to-comfy converter script, load that LoRA into diffusers.

If you need actual reproduction code or a file let me know

Logs

System Info

eb0a900ca

Who can help?

Questions on LoRA: @sayakpaul

Überprüfe die in der Tabelle beschriebenen betroffenen from-comfy-Konverter und vergleiche deren Verarbeitung mit den entsprechenden from-kohya-Konvertern. Reproduziere den gemeldeten kohya-to-comfy-Ladefall und stelle anschließend ein konsistentes .alpha-Verhalten über die betroffenen Modelle hinweg her; abgeschlossen ist die Aufgabe, wenn die erwartete Skalierung erhalten bleibt und durch geeignete Konvertertests abgedeckt ist.