protomotions.agents.peft.adapters module

Conditioned LoRA/DoRA adapters for TransformerEncoder stacks.

This module is model-agnostic: it only knows how to wrap transformer layers, freeze base weights, and expose adapter parameters for training.

class protomotions.agents.peft.adapters.LoRALayer(c_dim, in_dim, out_dim, rank, alpha)

Bases: <Mock object at 0x7fd694ba4990>[]

Low-rank adaptation with FiLM-style gating from a conditioning vector.

__init__(c_dim, in_dim, out_dim, rank, alpha)

forward(x, c_mul, c_add)

class protomotions.agents.peft.adapters.TransformerLayerWithLoRA(transformer_layer, c_dim, rank, alpha)

Bases: <Mock object at 0x7fd694bc8310>[]

__init__(

transformer_layer,

c_dim,

rank,

alpha,

)

forward(c, x, **kwargs)

class protomotions.agents.peft.adapters.TransformerLayerWithDoRA(transformer_layer, c_dim, rank, alpha)

Bases: <Mock object at 0x7fd694ba3050>[]

__init__(

transformer_layer,

c_dim,

rank,

alpha,

)

forward(c, x, **kwargs)

class protomotions.agents.peft.adapters.TransformerEncoderWithConditioning(layers)

Bases: <Mock object at 0x7fd694bca1d0>[]

Passes task conditioning to PEFT layers, standard forward to others.

__init__(layers)

forward(

x,

task_c=None,

mask=None,

src_key_padding_mask=None,

)

protomotions.agents.peft.adapters.freeze_base_and_enable_peft(module)

Freeze base weights and leave adapter/conditioning parameters trainable.

protomotions.agents.peft.adapters.set_peft_layers_train_mode(module, mode)

Set PEFT wrappers to train/eval while keeping wrapped base layers eval.

protomotions.agents.peft.adapters.inject_transformer_peft(

transformer,

conditioning_dim,

rank,

alpha,

peft_type,

)

Wrap TransformerEncoder layers with conditioned LoRA/DoRA adapters.