Question

In QLoRA, how does the integration of 4-bit NormalFloat quantization with low-rank matrices specifically reduce the memory footprint required for the backward pass?

Accepted Answer

In QLoRA, the backward pass memory footprint is reduced through three core mechanisms: 4-bit NormalFloat quantization, double quantization, and the frozen state of the main weights. First, 4-bit NormalFloat (NF4) is a data type that assumes neural network weights follow a normal distribution. By mapping these weights to a 4-bit space such that each value has an equal probability of occurring, NF4 maximizes information density, allowing the massive pre-trained model weights to be stored in significantly less memory than the standard 16-bit float representation. Second, double quantization applies a second level of quantization to the quantization constants themselves, which further shrinks the auxiliary memory needed to keep track of how the 4-bit values map back to high-precision numbers. Third, and most importantly for the backward pass, the pre-trained model weights are kept frozen in 4-bit, meaning the system does not need to store gradients for these weights. Because the gradients are only computed for the small, trainable low-rank adapter matrices—which are typically less than 1 percent of the total parameter count—the memory required to hold these gradients during backpropagation is drastically lowered. During the backward pass, the frozen weights are dequantized on-the-fly to 16-bit precision only as needed to perform matrix multiplications with activations, meaning the model avoids the memory overhead of storing full-precision weight gradients while maintaining the performance of a much larger model.

Home → All Courses → Programming Courses → Large Language Model (LLM) Engineering → Flashcard

In QLoRA, how does the integration of 4-bit NormalFloat quantization with low-rank matrices specifically reduce the memory footprint required for the backward pass?