# Copyright 2026-present the HuggingFace Inc. team. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import math import warnings from typing import Any, Optional import torch from torch import nn from transformers.activations import ACT2FN from peft.tuners.tuners_utils import BaseTunerLayer, check_adapters_to_merge class PeanutLayer(BaseTunerLayer): # All names of layers that may contain (trainable) adapter weights adapter_layer_names: tuple[str, ...] = ("peanut_A", "peanut_B", "peanut_encoders", "peanut_decoders") # All names of other parameters that may contain adapter-related parameters other_param_names: tuple[str, ...] = ("r", "depth", "scaling", "act_fn", "res_num") def __init__(self, base_layer: nn.Module, **kwargs) -> None: self.base_layer = base_layer self.r = {} self.depth = {} self.res_num = {} self.scaling = {} self.act_fn = {} self.peanut_A = nn.ModuleDict({}) self.peanut_B = nn.ModuleDict({}) self.peanut_encoders = nn.ModuleDict({}) self.peanut_decoders = nn.ModuleDict({}) self.kwargs = kwargs self._disable_adapters = False self.merged_adapters = [] self._cached_delta_weights = {} base_layer = self.get_base_layer() if isinstance(base_layer, nn.Linear): in_features, out_features = base_layer.in_features, base_layer.out_features else: if hasattr(base_layer, "in_features") and hasattr(base_layer, "out_features"): in_features, out_features = base_layer.in_features, base_layer.out_features else: in_features, out_features = None, None warnings.warn( f"Unsupported layer type '{type(base_layer)}' encountered, proceed at your own risk.", UserWarning ) self.in_features = in_features self.out_features = out_features def update_layer( self, adapter_name, r, depth, scaling, act_fn, init_weights: bool = True, inference_mode: bool = False, ): self.r[adapter_name] = r self.depth[adapter_name] = depth self.res_num[adapter_name] = depth self.scaling[adapter_name] = scaling self.act_fn[adapter_name] = act_fn self.peanut_A[adapter_name] = nn.Linear(self.out_features, r, bias=False) self.peanut_encoders[adapter_name] = nn.ModuleList([nn.Linear(r, r, bias=False) for _ in range(depth)]) self.peanut_decoders[adapter_name] = nn.ModuleList([nn.Linear(r, r, bias=False) for _ in range(depth)]) self.peanut_B[adapter_name] = nn.Linear(r, self.out_features, bias=False) self.reset_peanut_parameters(adapter_name, init_weights=init_weights) self._move_adapter_to_device_of_base_layer(adapter_name) self.set_adapter(self.active_adapters, inference_mode=inference_mode) def reset_peanut_parameters(self, adapter_name: str, init_weights: bool = True): if adapter_name not in self.peanut_A: return nn.init.kaiming_uniform_(self.peanut_A[adapter_name].weight, a=math.sqrt(5)) for encoder in self.peanut_encoders[adapter_name]: nn.init.kaiming_uniform_(encoder.weight, a=math.sqrt(5)) for decoder in self.peanut_decoders[adapter_name]: nn.init.kaiming_uniform_(decoder.weight, a=math.sqrt(5)) if init_weights: nn.init.zeros_(self.peanut_B[adapter_name].weight) else: nn.init.kaiming_uniform_(self.peanut_B[adapter_name].weight, a=math.sqrt(5)) class Linear(nn.Module, PeanutLayer): # PEANuT implemented in a dense layer def __init__( self, base_layer, adapter_name: str, r: int = 32, depth: int = 0, scaling: float = 1.0, act_fn: str = "relu", init_weights: bool = True, **kwargs, ) -> None: super().__init__() PeanutLayer.__init__(self, base_layer, **kwargs) self._active_adapter = adapter_name self.update_layer( adapter_name, r, depth=depth, scaling=scaling, act_fn=act_fn, init_weights=init_weights, ) def _compute_delta_weight(self, adapter: str, base_weight: torch.Tensor) -> torch.Tensor: if adapter not in self.peanut_A: raise ValueError(f"Adapter {adapter} not found.") peanut_A = self.peanut_A[adapter] peanut_B = self.peanut_B[adapter] non_linear = ACT2FN[self.act_fn[adapter]] scaling = self.scaling[adapter] res_num = self.res_num[adapter] peanut_encoders = self.peanut_encoders[adapter] peanut_decoders = self.peanut_decoders[adapter] base_weight_t = base_weight.transpose(0, 1).to(peanut_A.weight.dtype) delta_w = non_linear(torch.matmul(base_weight_t, peanut_A.weight.t())) residuals = [] for i in range(res_num): residuals.append(delta_w) encoder = peanut_encoders[i] delta_w = non_linear(encoder(delta_w)) for i in range(res_num): decoder = peanut_decoders[i] delta_w = non_linear(decoder(delta_w)) delta_w = delta_w + residuals[res_num - 1 - i] delta_w = peanut_B(delta_w) return (delta_w * scaling).transpose(0, 1) def get_delta_weight(self, adapter) -> torch.Tensor: base_weight = self.get_base_layer().weight return self._compute_delta_weight(adapter, base_weight) def merge(self, safe_merge: bool = False, adapter_names: Optional[list[str]] = None) -> None: adapter_names = check_adapters_to_merge(self, adapter_names) if not adapter_names: return base_layer = self.get_base_layer() merge_base_weight = base_layer.weight.data.detach().clone() for active_adapter in adapter_names: if active_adapter not in self.peanut_A: continue with torch.no_grad(): delta_weight = self._compute_delta_weight(active_adapter, merge_base_weight) delta_weight = delta_weight.to(dtype=base_layer.weight.dtype, device=base_layer.weight.device) if safe_merge: orig_weights = base_layer.weight.data.clone() orig_weights = orig_weights + delta_weight if not torch.isfinite(orig_weights).all(): raise ValueError( f"NaNs detected in the merged weights. The adapter {active_adapter} seems to be broken" ) base_layer.weight.data = orig_weights else: base_layer.weight.data.add_(delta_weight) if delta_weight.device.type != "cpu": cached_delta_weight = delta_weight.detach().to("cpu") else: cached_delta_weight = delta_weight.detach() self._cached_delta_weights[active_adapter] = cached_delta_weight self.merged_adapters.append(active_adapter) def unmerge(self) -> None: if not self.merged: warnings.warn("Already unmerged. Nothing to do.") return base_layer = self.get_base_layer() while len(self.merged_adapters) > 0: active_adapter = self.merged_adapters.pop() if active_adapter not in self.peanut_A: continue delta_weight = self._cached_delta_weights.pop(active_adapter, None) if delta_weight is None: raise ValueError(f"Cached delta weight for adapter '{active_adapter}' is missing; cannot unmerge.") base_layer.weight.data.sub_( delta_weight.to(dtype=base_layer.weight.dtype, device=base_layer.weight.device) ) def forward(self, x: torch.Tensor, *args: Any, **kwargs: Any) -> torch.Tensor: if self.disable_adapters: if self.merged: self.unmerge() result = self.base_layer(x, *args, **kwargs) elif self.merged: result = self.base_layer(x, *args, **kwargs) else: result = self.base_layer(x, *args, **kwargs) if self.active_adapters: torch_result_dtype = result.dtype for active_adapter in self.active_adapters: if active_adapter not in self.peanut_A: continue delta_weight = self.get_delta_weight(active_adapter) x_cast = self._cast_input_dtype(x, delta_weight.dtype) delta = torch.matmul(x_cast, delta_weight.transpose(0, 1)) result = result + delta.to(torch_result_dtype) return result def __repr__(self) -> str: rep = super().__repr__() return "peanut." + rep