Cross-Version π₯-Transfer of QLoRA-Trained Adapters
Abstract
AbstractTo investigate whether QLoRA-trained adapters transfer across model versions like LoRA-trained adapters, we train LoRA and QLoRA adapters on Llama-3-8B-Instruct and transfer them to Llama-3.1-8B-Instruct with a task-specific merge scale Ξ». Both methods are evaluated on a BF16 target base, isolating the training-time precision difference; NF4 deployment is evaluated separately. Across 20 main tasks, the absolute evaluation-score gap between QLoRA Transfer and LoRA Transfer is at most 0.054 on all tasks and at most 0.03 on 18/20 tasks, so training the adapter on a quantized base does not compromise its cross-version transferability. In a two-task deployment case study, serving QLoRA adapters with an NF4 base uses 5.7β10.0 GB peak allocated inference VRAM, compared with 15.1β19.6 GB for LoRA-merged BF16. Across 44 settings (22 tasks Γ 2 methods), the task-optimal Ξ» consistently exceeds the Frobenius-distance optimum (we call this discrepancy the Ξ» gap), suggesting that weight-space proximity alone is insufficient for identifying it. Replicating the pipeline on two further families (Mistral v0.1βv0.3, Qwen2β2.5) reproduces these findings. We also report a negative case: on CommonsenseQA, QLoRA Transfer falls significantly below direct adaptation.