Adapter scale from its config: mlx-lm writes lora_parameters.scale;
PEFT writes lora_alpha + r and optionally use_rslora. For ordinary PEFT
return alpha/r as the scale. For PEFT rsLoRA return alpha and set rsLora,
so the caller's one per-layer division produces alpha/√rank exactly once.
Internal configs already store their pre-√rank scale directly.
Adapter scale from its config: mlx-lm writes lora_parameters.scale; PEFT writes lora_alpha + r and optionally use_rslora. For ordinary PEFT return alpha/r as the scale. For PEFT rsLoRA return alpha and set rsLora, so the caller's one per-layer division produces alpha/√rank exactly once. Internal configs already store their pre-√rank scale directly.