Struct QLora 

pub struct QLora { /* private fields */ }

A Reparameterizer implementing QLoRA: it quantizes the (frozen) base weights and attaches full-precision trainable LoRA adapters to 2-D weights.

It is applied via Module::apply_qlora.

The quantized base is kept at rest in its low-bit representation; the adapter contribution is added on top during the forward pass (the base is dequantized on the fly when composed). Module-owned control flags are left unchanged, as with Lora.

Memory limitations

QLoRA reduces stored base-weight memory and trains only the adapter factors, but still composes a dense effective weight as dequantize(base) + scale * (a @ b). For a linear projection whose input requires gradients, autodiff retains that dense weight for backward and computes a full weight-shaped gradient before propagating into the factors. Peak training memory can therefore be substantially higher than the packed weights and adapters alone suggest. Fusion can avoid a separate dequantized base allocation, but does not eliminate all of these dense intermediates.

Module::valid preserves the packed base and adapters. Explicitly calling Module::materialize merges them into dense weights; requantization must be requested separately.

Implementations

impl QLora

pub fn new(lora: Lora, quantizer: Quantizer) -> Self

Create a new QLoRA reparameterizer from LoRA settings and a quantizer.

Trait Implementations

impl Reparameterizer for QLora

type Reparam = LoraAdapter

Reparameterization produced for a parameter.

fn reparameterize<const D: usize>( &mut self, path: &str, param: Param<Tensor<D>>, ) -> (Param<Tensor<D>>, Option<Self::Reparam>)

Prepare a parameter and optionally create a reparameterization for it.