lora.py

# Copyright 2023 The HuggingFace Team. All rights reserved.
#
# 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.

from typing import Optional

import torch.nn.functional as F
from torch import nn


class LoRALinearLayer(nn.Module):
    def __init__(self, in_features, out_features, rank=4, network_alpha=None, device=None, dtype=None):
        super().__init__()

        if rank > min(in_features, out_features):
            raise ValueError(f"LoRA rank {rank} must be less or equal than {min(in_features, out_features)}")

        self.down = nn.Linear(in_features, rank, bias=False, device=device, dtype=dtype)
        self.up = nn.Linear(rank, out_features, bias=False, device=device, dtype=dtype)
        # This value has the same meaning as the `--network_alpha` option in the kohya-ss trainer script.
        # See https://github.com/darkstorm2150/sd-scripts/blob/main/docs/train_network_README-en.md#execute-learning
        self.network_alpha = network_alpha
        self.rank = rank

        nn.init.normal_(self.down.weight, std=1 / rank)
        nn.init.zeros_(self.up.weight)

    def forward(self, hidden_states):
        orig_dtype = hidden_states.dtype
        dtype = self.down.weight.dtype

        down_hidden_states = self.down(hidden_states.to(dtype))
        up_hidden_states = self.up(down_hidden_states)

        if self.network_alpha is not None:
            up_hidden_states *= self.network_alpha / self.rank

        return up_hidden_states.to(orig_dtype)


class LoRAConv2dLayer(nn.Module):
    def __init__(self, in_features, out_features, rank=4, network_alpha=None):
        super().__init__()

        if rank > min(in_features, out_features):
            raise ValueError(f"LoRA rank {rank} must be less or equal than {min(in_features, out_features)}")

        self.down = nn.Conv2d(in_features, rank, (1, 1), (1, 1), bias=False)
        self.up = nn.Conv2d(rank, out_features, (1, 1), (1, 1), bias=False)
        # This value has the same meaning as the `--network_alpha` option in the kohya-ss trainer script.
        # See https://github.com/darkstorm2150/sd-scripts/blob/main/docs/train_network_README-en.md#execute-learning
        self.network_alpha = network_alpha
        self.rank = rank

        nn.init.normal_(self.down.weight, std=1 / rank)
        nn.init.zeros_(self.up.weight)

    def forward(self, hidden_states):
        orig_dtype = hidden_states.dtype
        dtype = self.down.weight.dtype

        down_hidden_states = self.down(hidden_states.to(dtype))
        up_hidden_states = self.up(down_hidden_states)

        if self.network_alpha is not None:
            up_hidden_states *= self.network_alpha / self.rank

        return up_hidden_states.to(orig_dtype)


class LoRACompatibleConv(nn.Conv2d):
    """
    A convolutional layer that can be used with LoRA.
    """

    def __init__(self, *args, lora_layer: Optional[LoRAConv2dLayer] = None, **kwargs):
        super().__init__(*args, **kwargs)
        self.lora_layer = lora_layer

    def set_lora_layer(self, lora_layer: Optional[LoRAConv2dLayer]):
        self.lora_layer = lora_layer

    def forward(self, x):
        if self.lora_layer is None:
            # make sure to the functional Conv2D function as otherwise torch.compile's graph will break
            # see: https://github.com/huggingface/diffusers/pull/4315
            return F.conv2d(x, self.weight, self.bias, self.stride, self.padding, self.dilation, self.groups)
        else:
            return super().forward(x) + self.lora_layer(x)


class LoRACompatibleLinear(nn.Linear):
    """
    A Linear layer that can be used with LoRA.
    """

    def __init__(self, *args, lora_layer: Optional[LoRALinearLayer] = None, **kwargs):
        super().__init__(*args, **kwargs)
        self.lora_layer = lora_layer

    def set_lora_layer(self, lora_layer: Optional[LoRAConv2dLayer]):
        self.lora_layer = lora_layer

    def forward(self, x):
        if self.lora_layer is None:
            return super().forward(x)
        else:
            return super().forward(x) + self.lora_layer(x)