Accelerated LoRA is optimized during training for a specific sampler scheduling strategy and model weights; when the actual sampler's step size distribution, noise schedule, or model quantization method is inconsistent with the training conditions, the shortcut paths learned by the LoRA may fail entirely.