Forward and backward computation passes through the quantized base weights, while adapter parameters retain suitable training precision. The method reduces memory use, though quantization choices can affect speed and fidelity.
Quantized low-rank adaptation trains LoRA adapters while the frozen base model is stored in a lower-precision representation.
Forward and backward computation passes through the quantized base weights, while adapter parameters retain suitable training precision. The method reduces memory use, though quantization choices can affect speed and fidelity.