alf.pretrained_models.model_adapters

alf.pretrained_models.model_adapters.lora

class Conv2dAdapter(m, rank=16, weight=1.0, name='LoRA')

Bases: alf.pretrained_models.model_adapters.lora.LoRA

Adapter for Conv2d layers.

The most natural way of a LoRA decomposition for Conv2d is

(rank, kernel_size[0] * kernel_size[1] * in_channels) x (out_channels, rank)

However, since out_channels is usually small, this decomposition is not low-rank actually and it won’t save much memory.

We can first reinterpret the weight matrix as a shape of

(out_channels * kernel_size[0], in_channels * kernel_size[1])

to make the in- and out-dimensions balanced, and then decompose.

Parameters

• m (Module) – the module to be adapted

• rank (int) – the low rank

• weight (float) – weight for the low-rank weight matrix

classmethod can_adapt(m)

Check if the adapter class can adapt a given module.

Return type

bool

forward(input)

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

training: bool

class EmbeddingAdapter(m, rank=16, weight=1.0, name='LoRA')

Bases: alf.pretrained_models.model_adapters.lora.LoRA

Adapter for embedding layers.

Parameters

• m (Module) – the module to be adapted

• rank (int) – the low rank

• weight (float) – weight for the low-rank weight matrix

classmethod can_adapt(m)

Check if the adapter class can adapt a given module.

Return type

bool

forward(input)

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

training: bool

class LinearAdapter(m, rank=16, weight=1.0, name='LoRA')

Bases: alf.pretrained_models.model_adapters.lora.LoRA

Adapter for linear layers.

Parameters

• m (Module) – the module to be adapted

• rank (int) – the low rank

• weight (float) – weight for the low-rank weight matrix

classmethod can_adapt(m)

Check if the adapter class can adapt a given module.

Return type

bool

forward(input)

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

training: bool

class LoRA(m, rank=16, weight=1.0, name='LoRA')

Bases: torch.nn.modules.module.Module

Base class for LoRA (Low-Rank Adaptation).

For any model layer expressed as a matrix multiplication of the form \(h=W_0x\), it performs a reparameterization such that:

\[h = W_0x + \frac{\alpha}{r} BAx\]

where \(A\in\mathbb{R}^{r\times k}\) and \(B\in\mathbb{R}^{d\times r}\) are the decomposition matrices and \(r\) is the low-dim rank of the decomposition.

The original LoRA paper doesn’t adapt biases: “We leave the empirical investigation of adapting the MLP layers, LayerNorm layers, and biases to a future work.”

If a module has a weight matrix with dimensionality greater than 2 (e.g., conv), we might need to first reinterpret it as a 2D matrix W_0. This reinterpretation will differ for different modules.

Parameters

• m (Module) – the module to be adapted

• rank (int) – the low rank

• weight (float) – weight for the low-rank weight matrix

classmethod can_adapt(m)

Check if the adapter class can adapt a given module.

Return type

bool

detach()

Detach from and recover the base module.

Note that this operation is irreversible. Once detached, there is no way to add the adapter back to the module.

We still keep the adapter weights after detachment.

forward(input)

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

merge()

Merge the adapter and base module weights.

This operation will change the base module weight in place and restore its original forward(). It should only be called in the inference mode.

reset_parameters()

property scaling: float

Return type

float

training: bool

unmerge()

Unmerge the adapter weights. The base module is still adapted, and the adapter weights can be trained after unmerge.