Rename flux_kohya_lora_conversion_utils.py

invokeai/app/invocations/compel.py

@@ -19,8 +19,8 @@ from invokeai.app.invocations.model import CLIPField
 from invokeai.app.invocations.primitives import ConditioningOutput
 from invokeai.app.services.shared.invocation_context import InvocationContext
 from invokeai.app.util.ti_utils import generate_ti_list
-from invokeai.backend.lora import LoRAModelRaw
 from invokeai.backend.model_patcher import ModelPatcher
+from invokeai.backend.peft.lora import LoRAModelRaw
 from invokeai.backend.stable_diffusion.diffusion.conditioning_data import (
     BasicConditioningInfo,
     ConditioningFieldData,

invokeai/app/invocations/denoise_latents.py

@@ -36,9 +36,9 @@ from invokeai.app.invocations.t2i_adapter import T2IAdapterField
 from invokeai.app.services.shared.invocation_context import InvocationContext
 from invokeai.app.util.controlnet_utils import prepare_control_image
 from invokeai.backend.ip_adapter.ip_adapter import IPAdapter
-from invokeai.backend.lora import LoRAModelRaw
 from invokeai.backend.model_manager import BaseModelType, ModelVariantType
 from invokeai.backend.model_patcher import ModelPatcher
+from invokeai.backend.peft.lora import LoRAModelRaw
 from invokeai.backend.stable_diffusion import PipelineIntermediateState
 from invokeai.backend.stable_diffusion.denoise_context import DenoiseContext, DenoiseInputs
 from invokeai.backend.stable_diffusion.diffusers_pipeline import (

invokeai/app/invocations/flux_denoise.py

@@ -1,4 +1,4 @@
-from typing import Callable, Optional
+from typing import Callable, Iterator, Optional, Tuple
 
 import torch
 import torchvision.transforms as tv_transforms
@@ -30,6 +30,8 @@ from invokeai.backend.flux.sampling_utils import (
     pack,
     unpack,
 )
+from invokeai.backend.peft.lora import LoRAModelRaw
+from invokeai.backend.peft.peft_patcher import PeftPatcher
 from invokeai.backend.stable_diffusion.diffusion.conditioning_data import FLUXConditioningInfo
 from invokeai.backend.util.devices import TorchDevice
 
@@ -187,7 +189,16 @@ class FluxDenoiseInvocation(BaseInvocation, WithMetadata, WithBoard):
                 noise=noise,
             )
 
-        with transformer_info as transformer:
+        with (
+            transformer_info.model_on_device() as (cached_weights, transformer),
+            # Apply the LoRA after transformer has been moved to its target device for faster patching.
+            PeftPatcher.apply_peft_patches(
+                model=transformer,
+                patches=self._lora_iterator(context),
+                prefix="",
+                cached_weights=cached_weights,
+            ),
+        ):
             assert isinstance(transformer, Flux)
 
             x = denoise(
@@ -241,6 +252,13 @@ class FluxDenoiseInvocation(BaseInvocation, WithMetadata, WithBoard):
         # `latents`.
         return mask.expand_as(latents)
 
+    def _lora_iterator(self, context: InvocationContext) -> Iterator[Tuple[LoRAModelRaw, float]]:
+        for lora in self.transformer.loras:
+            lora_info = context.models.load(lora.lora)
+            assert isinstance(lora_info.model, LoRAModelRaw)
+            yield (lora_info.model, lora.weight)
+            del lora_info
+
     def _build_step_callback(self, context: InvocationContext) -> Callable[[], None]:
         def step_callback() -> None:
             if context.util.is_canceled():

invokeai/app/invocations/flux_lora_loader.py

@@ -0,0 +1,53 @@
+from invokeai.app.invocations.baseinvocation import BaseInvocation, BaseInvocationOutput, invocation, invocation_output
+from invokeai.app.invocations.fields import FieldDescriptions, Input, InputField, OutputField, UIType
+from invokeai.app.invocations.model import LoRAField, ModelIdentifierField, TransformerField
+from invokeai.app.services.shared.invocation_context import InvocationContext
+
+
+@invocation_output("flux_lora_loader_output")
+class FluxLoRALoaderOutput(BaseInvocationOutput):
+    """FLUX LoRA Loader Output"""
+
+    transformer: TransformerField = OutputField(
+        default=None, description=FieldDescriptions.transformer, title="FLUX Transformer"
+    )
+
+
+@invocation(
+    "flux_lora_loader",
+    title="FLUX LoRA",
+    tags=["lora", "model", "flux"],
+    category="model",
+    version="1.0.0",
+)
+class FluxLoRALoaderInvocation(BaseInvocation):
+    """Apply a LoRA model to a FLUX transformer."""
+
+    lora: ModelIdentifierField = InputField(
+        description=FieldDescriptions.lora_model, title="LoRA", ui_type=UIType.LoRAModel
+    )
+    weight: float = InputField(default=0.75, description=FieldDescriptions.lora_weight)
+    transformer: TransformerField = InputField(
+        description=FieldDescriptions.transformer,
+        input=Input.Connection,
+        title="FLUX Transformer",
+    )
+
+    def invoke(self, context: InvocationContext) -> FluxLoRALoaderOutput:
+        lora_key = self.lora.key
+
+        if not context.models.exists(lora_key):
+            raise ValueError(f"Unknown lora: {lora_key}!")
+
+        if any(lora.lora.key == lora_key for lora in self.transformer.loras):
+            raise Exception(f'LoRA "{lora_key}" already applied to transformer.')
+
+        transformer = self.transformer.model_copy(deep=True)
+        transformer.loras.append(
+            LoRAField(
+                lora=self.lora,
+                weight=self.weight,
+            )
+        )
+
+        return FluxLoRALoaderOutput(transformer=transformer)

invokeai/app/invocations/model.py

@@ -69,6 +69,7 @@ class CLIPField(BaseModel):
 
 class TransformerField(BaseModel):
     transformer: ModelIdentifierField = Field(description="Info to load Transformer submodel")
+    loras: List[LoRAField] = Field(description="LoRAs to apply on model loading")
 
 
 class T5EncoderField(BaseModel):
@@ -202,7 +203,7 @@ class FluxModelLoaderInvocation(BaseInvocation):
         assert isinstance(transformer_config, CheckpointConfigBase)
 
         return FluxModelLoaderOutput(
-            transformer=TransformerField(transformer=transformer),
+            transformer=TransformerField(transformer=transformer, loras=[]),
             clip=CLIPField(tokenizer=tokenizer, text_encoder=clip_encoder, loras=[], skipped_layers=0),
             t5_encoder=T5EncoderField(tokenizer=tokenizer2, text_encoder=t5_encoder),
             vae=VAEField(vae=vae),

invokeai/app/invocations/tiled_multi_diffusion_denoise_latents.py

@@ -22,8 +22,8 @@ from invokeai.app.invocations.fields import (
 from invokeai.app.invocations.model import UNetField
 from invokeai.app.invocations.primitives import LatentsOutput
 from invokeai.app.services.shared.invocation_context import InvocationContext
-from invokeai.backend.lora import LoRAModelRaw
 from invokeai.backend.model_patcher import ModelPatcher
+from invokeai.backend.peft.lora import LoRAModelRaw
 from invokeai.backend.stable_diffusion.diffusers_pipeline import ControlNetData, PipelineIntermediateState
 from invokeai.backend.stable_diffusion.multi_diffusion_pipeline import (
     MultiDiffusionPipeline,

invokeai/backend/lora.py

@@ -1,672 +0,0 @@
-# Copyright (c) 2024 The InvokeAI Development team
-"""LoRA model support."""
-
-import bisect
-from pathlib import Path
-from typing import Dict, List, Optional, Set, Tuple, Union
-
-import torch
-from safetensors.torch import load_file
-from typing_extensions import Self
-
-import invokeai.backend.util.logging as logger
-from invokeai.backend.model_manager import BaseModelType
-from invokeai.backend.raw_model import RawModel
-
-
-class LoRALayerBase:
-    # rank: Optional[int]
-    # alpha: Optional[float]
-    # bias: Optional[torch.Tensor]
-    # layer_key: str
-
-    # @property
-    # def scale(self):
-    #    return self.alpha / self.rank if (self.alpha and self.rank) else 1.0
-
-    def __init__(
-        self,
-        layer_key: str,
-        values: Dict[str, torch.Tensor],
-    ):
-        if "alpha" in values:
-            self.alpha = values["alpha"].item()
-        else:
-            self.alpha = None
-
-        if "bias_indices" in values and "bias_values" in values and "bias_size" in values:
-            self.bias: Optional[torch.Tensor] = torch.sparse_coo_tensor(
-                values["bias_indices"],
-                values["bias_values"],
-                tuple(values["bias_size"]),
-            )
-
-        else:
-            self.bias = None
-
-        self.rank = None  # set in layer implementation
-        self.layer_key = layer_key
-
-    def get_weight(self, orig_weight: torch.Tensor) -> torch.Tensor:
-        raise NotImplementedError()
-
-    def get_bias(self, orig_bias: torch.Tensor) -> Optional[torch.Tensor]:
-        return self.bias
-
-    def get_parameters(self, orig_module: torch.nn.Module) -> Dict[str, torch.Tensor]:
-        params = {"weight": self.get_weight(orig_module.weight)}
-        bias = self.get_bias(orig_module.bias)
-        if bias is not None:
-            params["bias"] = bias
-        return params
-
-    def calc_size(self) -> int:
-        model_size = 0
-        for val in [self.bias]:
-            if val is not None:
-                model_size += val.nelement() * val.element_size()
-        return model_size
-
-    def to(self, device: Optional[torch.device] = None, dtype: Optional[torch.dtype] = None) -> None:
-        if self.bias is not None:
-            self.bias = self.bias.to(device=device, dtype=dtype)
-
-    def check_keys(self, values: Dict[str, torch.Tensor], known_keys: Set[str]):
-        """Log a warning if values contains unhandled keys."""
-        # {"alpha", "bias_indices", "bias_values", "bias_size"} are hard-coded, because they are handled by
-        # `LoRALayerBase`. Sub-classes should provide the known_keys that they handled.
-        all_known_keys = known_keys | {"alpha", "bias_indices", "bias_values", "bias_size"}
-        unknown_keys = set(values.keys()) - all_known_keys
-        if unknown_keys:
-            logger.warning(
-                f"Unexpected keys found in LoRA/LyCORIS layer, model might work incorrectly! Keys: {unknown_keys}"
-            )
-
-
-# TODO: find and debug lora/locon with bias
-class LoRALayer(LoRALayerBase):
-    # up: torch.Tensor
-    # mid: Optional[torch.Tensor]
-    # down: torch.Tensor
-
-    def __init__(
-        self,
-        layer_key: str,
-        values: Dict[str, torch.Tensor],
-    ):
-        super().__init__(layer_key, values)
-
-        self.up = values["lora_up.weight"]
-        self.down = values["lora_down.weight"]
-        self.mid = values.get("lora_mid.weight", None)
-
-        self.rank = self.down.shape[0]
-        self.check_keys(
-            values,
-            {
-                "lora_up.weight",
-                "lora_down.weight",
-                "lora_mid.weight",
-            },
-        )
-
-    def get_weight(self, orig_weight: torch.Tensor) -> torch.Tensor:
-        if self.mid is not None:
-            up = self.up.reshape(self.up.shape[0], self.up.shape[1])
-            down = self.down.reshape(self.down.shape[0], self.down.shape[1])
-            weight = torch.einsum("m n w h, i m, n j -> i j w h", self.mid, up, down)
-        else:
-            weight = self.up.reshape(self.up.shape[0], -1) @ self.down.reshape(self.down.shape[0], -1)
-
-        return weight
-
-    def calc_size(self) -> int:
-        model_size = super().calc_size()
-        for val in [self.up, self.mid, self.down]:
-            if val is not None:
-                model_size += val.nelement() * val.element_size()
-        return model_size
-
-    def to(self, device: Optional[torch.device] = None, dtype: Optional[torch.dtype] = None) -> None:
-        super().to(device=device, dtype=dtype)
-
-        self.up = self.up.to(device=device, dtype=dtype)
-        self.down = self.down.to(device=device, dtype=dtype)
-
-        if self.mid is not None:
-            self.mid = self.mid.to(device=device, dtype=dtype)
-
-
-class LoHALayer(LoRALayerBase):
-    # w1_a: torch.Tensor
-    # w1_b: torch.Tensor
-    # w2_a: torch.Tensor
-    # w2_b: torch.Tensor
-    # t1: Optional[torch.Tensor] = None
-    # t2: Optional[torch.Tensor] = None
-
-    def __init__(self, layer_key: str, values: Dict[str, torch.Tensor]):
-        super().__init__(layer_key, values)
-
-        self.w1_a = values["hada_w1_a"]
-        self.w1_b = values["hada_w1_b"]
-        self.w2_a = values["hada_w2_a"]
-        self.w2_b = values["hada_w2_b"]
-        self.t1 = values.get("hada_t1", None)
-        self.t2 = values.get("hada_t2", None)
-
-        self.rank = self.w1_b.shape[0]
-        self.check_keys(
-            values,
-            {
-                "hada_w1_a",
-                "hada_w1_b",
-                "hada_w2_a",
-                "hada_w2_b",
-                "hada_t1",
-                "hada_t2",
-            },
-        )
-
-    def get_weight(self, orig_weight: torch.Tensor) -> torch.Tensor:
-        if self.t1 is None:
-            weight: torch.Tensor = (self.w1_a @ self.w1_b) * (self.w2_a @ self.w2_b)
-
-        else:
-            rebuild1 = torch.einsum("i j k l, j r, i p -> p r k l", self.t1, self.w1_b, self.w1_a)
-            rebuild2 = torch.einsum("i j k l, j r, i p -> p r k l", self.t2, self.w2_b, self.w2_a)
-            weight = rebuild1 * rebuild2
-
-        return weight
-
-    def calc_size(self) -> int:
-        model_size = super().calc_size()
-        for val in [self.w1_a, self.w1_b, self.w2_a, self.w2_b, self.t1, self.t2]:
-            if val is not None:
-                model_size += val.nelement() * val.element_size()
-        return model_size
-
-    def to(self, device: Optional[torch.device] = None, dtype: Optional[torch.dtype] = None) -> None:
-        super().to(device=device, dtype=dtype)
-
-        self.w1_a = self.w1_a.to(device=device, dtype=dtype)
-        self.w1_b = self.w1_b.to(device=device, dtype=dtype)
-        if self.t1 is not None:
-            self.t1 = self.t1.to(device=device, dtype=dtype)
-
-        self.w2_a = self.w2_a.to(device=device, dtype=dtype)
-        self.w2_b = self.w2_b.to(device=device, dtype=dtype)
-        if self.t2 is not None:
-            self.t2 = self.t2.to(device=device, dtype=dtype)
-
-
-class LoKRLayer(LoRALayerBase):
-    # w1: Optional[torch.Tensor] = None
-    # w1_a: Optional[torch.Tensor] = None
-    # w1_b: Optional[torch.Tensor] = None
-    # w2: Optional[torch.Tensor] = None
-    # w2_a: Optional[torch.Tensor] = None
-    # w2_b: Optional[torch.Tensor] = None
-    # t2: Optional[torch.Tensor] = None
-
-    def __init__(
-        self,
-        layer_key: str,
-        values: Dict[str, torch.Tensor],
-    ):
-        super().__init__(layer_key, values)
-
-        self.w1 = values.get("lokr_w1", None)
-        if self.w1 is None:
-            self.w1_a = values["lokr_w1_a"]
-            self.w1_b = values["lokr_w1_b"]
-        else:
-            self.w1_b = None
-            self.w1_a = None
-
-        self.w2 = values.get("lokr_w2", None)
-        if self.w2 is None:
-            self.w2_a = values["lokr_w2_a"]
-            self.w2_b = values["lokr_w2_b"]
-        else:
-            self.w2_a = None
-            self.w2_b = None
-
-        self.t2 = values.get("lokr_t2", None)
-
-        if self.w1_b is not None:
-            self.rank = self.w1_b.shape[0]
-        elif self.w2_b is not None:
-            self.rank = self.w2_b.shape[0]
-        else:
-            self.rank = None  # unscaled
-
-        self.check_keys(
-            values,
-            {
-                "lokr_w1",
-                "lokr_w1_a",
-                "lokr_w1_b",
-                "lokr_w2",
-                "lokr_w2_a",
-                "lokr_w2_b",
-                "lokr_t2",
-            },
-        )
-
-    def get_weight(self, orig_weight: torch.Tensor) -> torch.Tensor:
-        w1: Optional[torch.Tensor] = self.w1
-        if w1 is None:
-            assert self.w1_a is not None
-            assert self.w1_b is not None
-            w1 = self.w1_a @ self.w1_b
-
-        w2 = self.w2
-        if w2 is None:
-            if self.t2 is None:
-                assert self.w2_a is not None
-                assert self.w2_b is not None
-                w2 = self.w2_a @ self.w2_b
-            else:
-                w2 = torch.einsum("i j k l, i p, j r -> p r k l", self.t2, self.w2_a, self.w2_b)
-
-        if len(w2.shape) == 4:
-            w1 = w1.unsqueeze(2).unsqueeze(2)
-        w2 = w2.contiguous()
-        assert w1 is not None
-        assert w2 is not None
-        weight = torch.kron(w1, w2)
-
-        return weight
-
-    def calc_size(self) -> int:
-        model_size = super().calc_size()
-        for val in [self.w1, self.w1_a, self.w1_b, self.w2, self.w2_a, self.w2_b, self.t2]:
-            if val is not None:
-                model_size += val.nelement() * val.element_size()
-        return model_size
-
-    def to(self, device: Optional[torch.device] = None, dtype: Optional[torch.dtype] = None) -> None:
-        super().to(device=device, dtype=dtype)
-
-        if self.w1 is not None:
-            self.w1 = self.w1.to(device=device, dtype=dtype)
-        else:
-            assert self.w1_a is not None
-            assert self.w1_b is not None
-            self.w1_a = self.w1_a.to(device=device, dtype=dtype)
-            self.w1_b = self.w1_b.to(device=device, dtype=dtype)
-
-        if self.w2 is not None:
-            self.w2 = self.w2.to(device=device, dtype=dtype)
-        else:
-            assert self.w2_a is not None
-            assert self.w2_b is not None
-            self.w2_a = self.w2_a.to(device=device, dtype=dtype)
-            self.w2_b = self.w2_b.to(device=device, dtype=dtype)
-
-        if self.t2 is not None:
-            self.t2 = self.t2.to(device=device, dtype=dtype)
-
-
-class FullLayer(LoRALayerBase):
-    # bias handled in LoRALayerBase(calc_size, to)
-    # weight: torch.Tensor
-    # bias: Optional[torch.Tensor]
-
-    def __init__(
-        self,
-        layer_key: str,
-        values: Dict[str, torch.Tensor],
-    ):
-        super().__init__(layer_key, values)
-
-        self.weight = values["diff"]
-        self.bias = values.get("diff_b", None)
-
-        self.rank = None  # unscaled
-        self.check_keys(values, {"diff", "diff_b"})
-
-    def get_weight(self, orig_weight: torch.Tensor) -> torch.Tensor:
-        return self.weight
-
-    def calc_size(self) -> int:
-        model_size = super().calc_size()
-        model_size += self.weight.nelement() * self.weight.element_size()
-        return model_size
-
-    def to(self, device: Optional[torch.device] = None, dtype: Optional[torch.dtype] = None) -> None:
-        super().to(device=device, dtype=dtype)
-
-        self.weight = self.weight.to(device=device, dtype=dtype)
-
-
-class IA3Layer(LoRALayerBase):
-    # weight: torch.Tensor
-    # on_input: torch.Tensor
-
-    def __init__(
-        self,
-        layer_key: str,
-        values: Dict[str, torch.Tensor],
-    ):
-        super().__init__(layer_key, values)
-
-        self.weight = values["weight"]
-        self.on_input = values["on_input"]
-
-        self.rank = None  # unscaled
-        self.check_keys(values, {"weight", "on_input"})
-
-    def get_weight(self, orig_weight: torch.Tensor) -> torch.Tensor:
-        weight = self.weight
-        if not self.on_input:
-            weight = weight.reshape(-1, 1)
-        assert orig_weight is not None
-        return orig_weight * weight
-
-    def calc_size(self) -> int:
-        model_size = super().calc_size()
-        model_size += self.weight.nelement() * self.weight.element_size()
-        model_size += self.on_input.nelement() * self.on_input.element_size()
-        return model_size
-
-    def to(self, device: Optional[torch.device] = None, dtype: Optional[torch.dtype] = None):
-        super().to(device=device, dtype=dtype)
-
-        self.weight = self.weight.to(device=device, dtype=dtype)
-        self.on_input = self.on_input.to(device=device, dtype=dtype)
-
-
-class NormLayer(LoRALayerBase):
-    # bias handled in LoRALayerBase(calc_size, to)
-    # weight: torch.Tensor
-    # bias: Optional[torch.Tensor]
-
-    def __init__(
-        self,
-        layer_key: str,
-        values: Dict[str, torch.Tensor],
-    ):
-        super().__init__(layer_key, values)
-
-        self.weight = values["w_norm"]
-        self.bias = values.get("b_norm", None)
-
-        self.rank = None  # unscaled
-        self.check_keys(values, {"w_norm", "b_norm"})
-
-    def get_weight(self, orig_weight: torch.Tensor) -> torch.Tensor:
-        return self.weight
-
-    def calc_size(self) -> int:
-        model_size = super().calc_size()
-        model_size += self.weight.nelement() * self.weight.element_size()
-        return model_size
-
-    def to(self, device: Optional[torch.device] = None, dtype: Optional[torch.dtype] = None) -> None:
-        super().to(device=device, dtype=dtype)
-
-        self.weight = self.weight.to(device=device, dtype=dtype)
-
-
-AnyLoRALayer = Union[LoRALayer, LoHALayer, LoKRLayer, FullLayer, IA3Layer, NormLayer]
-
-
-class LoRAModelRaw(RawModel):  # (torch.nn.Module):
-    _name: str
-    layers: Dict[str, AnyLoRALayer]
-
-    def __init__(
-        self,
-        name: str,
-        layers: Dict[str, AnyLoRALayer],
-    ):
-        self._name = name
-        self.layers = layers
-
-    @property
-    def name(self) -> str:
-        return self._name
-
-    def to(self, device: Optional[torch.device] = None, dtype: Optional[torch.dtype] = None) -> None:
-        # TODO: try revert if exception?
-        for _key, layer in self.layers.items():
-            layer.to(device=device, dtype=dtype)
-
-    def calc_size(self) -> int:
-        model_size = 0
-        for _, layer in self.layers.items():
-            model_size += layer.calc_size()
-        return model_size
-
-    @classmethod
-    def _convert_sdxl_keys_to_diffusers_format(cls, state_dict: Dict[str, torch.Tensor]) -> Dict[str, torch.Tensor]:
-        """Convert the keys of an SDXL LoRA state_dict to diffusers format.
-
-        The input state_dict can be in either Stability AI format or diffusers format. If the state_dict is already in
-        diffusers format, then this function will have no effect.
-
-        This function is adapted from:
-        https://github.com/bmaltais/kohya_ss/blob/2accb1305979ba62f5077a23aabac23b4c37e935/networks/lora_diffusers.py#L385-L409
-
-        Args:
-            state_dict (Dict[str, Tensor]): The SDXL LoRA state_dict.
-
-        Raises:
-            ValueError: If state_dict contains an unrecognized key, or not all keys could be converted.
-
-        Returns:
-            Dict[str, Tensor]: The diffusers-format state_dict.
-        """
-        converted_count = 0  # The number of Stability AI keys converted to diffusers format.
-        not_converted_count = 0  # The number of keys that were not converted.
-
-        # Get a sorted list of Stability AI UNet keys so that we can efficiently search for keys with matching prefixes.
-        # For example, we want to efficiently find `input_blocks_4_1` in the list when searching for
-        # `input_blocks_4_1_proj_in`.
-        stability_unet_keys = list(SDXL_UNET_STABILITY_TO_DIFFUSERS_MAP)
-        stability_unet_keys.sort()
-
-        new_state_dict = {}
-        for full_key, value in state_dict.items():
-            if full_key.startswith("lora_unet_"):
-                search_key = full_key.replace("lora_unet_", "")
-                # Use bisect to find the key in stability_unet_keys that *may* match the search_key's prefix.
-                position = bisect.bisect_right(stability_unet_keys, search_key)
-                map_key = stability_unet_keys[position - 1]
-                # Now, check if the map_key *actually* matches the search_key.
-                if search_key.startswith(map_key):
-                    new_key = full_key.replace(map_key, SDXL_UNET_STABILITY_TO_DIFFUSERS_MAP[map_key])
-                    new_state_dict[new_key] = value
-                    converted_count += 1
-                else:
-                    new_state_dict[full_key] = value
-                    not_converted_count += 1
-            elif full_key.startswith("lora_te1_") or full_key.startswith("lora_te2_"):
-                # The CLIP text encoders have the same keys in both Stability AI and diffusers formats.
-                new_state_dict[full_key] = value
-                continue
-            else:
-                raise ValueError(f"Unrecognized SDXL LoRA key prefix: '{full_key}'.")
-
-        if converted_count > 0 and not_converted_count > 0:
-            raise ValueError(
-                f"The SDXL LoRA could only be partially converted to diffusers format. converted={converted_count},"
-                f" not_converted={not_converted_count}"
-            )
-
-        return new_state_dict
-
-    @classmethod
-    def from_checkpoint(
-        cls,
-        file_path: Union[str, Path],
-        device: Optional[torch.device] = None,
-        dtype: Optional[torch.dtype] = None,
-        base_model: Optional[BaseModelType] = None,
-    ) -> Self:
-        device = device or torch.device("cpu")
-        dtype = dtype or torch.float32
-
-        if isinstance(file_path, str):
-            file_path = Path(file_path)
-
-        model = cls(
-            name=file_path.stem,
-            layers={},
-        )
-
-        if file_path.suffix == ".safetensors":
-            sd = load_file(file_path.absolute().as_posix(), device="cpu")
-        else:
-            sd = torch.load(file_path, map_location="cpu")
-
-        state_dict = cls._group_state(sd)
-
-        if base_model == BaseModelType.StableDiffusionXL:
-            state_dict = cls._convert_sdxl_keys_to_diffusers_format(state_dict)
-
-        for layer_key, values in state_dict.items():
-            # Detect layers according to LyCORIS detection logic(`weight_list_det`)
-            # https://github.com/KohakuBlueleaf/LyCORIS/tree/8ad8000efb79e2b879054da8c9356e6143591bad/lycoris/modules
-
-            # lora and locon
-            if "lora_up.weight" in values:
-                layer: AnyLoRALayer = LoRALayer(layer_key, values)
-
-            # loha
-            elif "hada_w1_a" in values:
-                layer = LoHALayer(layer_key, values)
-
-            # lokr
-            elif "lokr_w1" in values or "lokr_w1_a" in values:
-                layer = LoKRLayer(layer_key, values)
-
-            # diff
-            elif "diff" in values:
-                layer = FullLayer(layer_key, values)
-
-            # ia3
-            elif "on_input" in values:
-                layer = IA3Layer(layer_key, values)
-
-            # norms
-            elif "w_norm" in values:
-                layer = NormLayer(layer_key, values)
-
-            else:
-                print(f">> Encountered unknown lora layer module in {model.name}: {layer_key} - {list(values.keys())}")
-                raise Exception("Unknown lora format!")
-
-            # lower memory consumption by removing already parsed layer values
-            state_dict[layer_key].clear()
-
-            layer.to(device=device, dtype=dtype)
-            model.layers[layer_key] = layer
-
-        return model
-
-    @staticmethod
-    def _group_state(state_dict: Dict[str, torch.Tensor]) -> Dict[str, Dict[str, torch.Tensor]]:
-        state_dict_groupped: Dict[str, Dict[str, torch.Tensor]] = {}
-
-        for key, value in state_dict.items():
-            stem, leaf = key.split(".", 1)
-            if stem not in state_dict_groupped:
-                state_dict_groupped[stem] = {}
-            state_dict_groupped[stem][leaf] = value
-
-        return state_dict_groupped
-
-
-# code from
-# https://github.com/bmaltais/kohya_ss/blob/2accb1305979ba62f5077a23aabac23b4c37e935/networks/lora_diffusers.py#L15C1-L97C32
-def make_sdxl_unet_conversion_map() -> List[Tuple[str, str]]:
-    """Create a dict mapping state_dict keys from Stability AI SDXL format to diffusers SDXL format."""
-    unet_conversion_map_layer = []
-
-    for i in range(3):  # num_blocks is 3 in sdxl
-        # loop over downblocks/upblocks
-        for j in range(2):
-            # loop over resnets/attentions for downblocks
-            hf_down_res_prefix = f"down_blocks.{i}.resnets.{j}."
-            sd_down_res_prefix = f"input_blocks.{3*i + j + 1}.0."
-            unet_conversion_map_layer.append((sd_down_res_prefix, hf_down_res_prefix))
-
-            if i < 3:
-                # no attention layers in down_blocks.3
-                hf_down_atn_prefix = f"down_blocks.{i}.attentions.{j}."
-                sd_down_atn_prefix = f"input_blocks.{3*i + j + 1}.1."
-                unet_conversion_map_layer.append((sd_down_atn_prefix, hf_down_atn_prefix))
-
-        for j in range(3):
-            # loop over resnets/attentions for upblocks
-            hf_up_res_prefix = f"up_blocks.{i}.resnets.{j}."
-            sd_up_res_prefix = f"output_blocks.{3*i + j}.0."
-            unet_conversion_map_layer.append((sd_up_res_prefix, hf_up_res_prefix))
-
-            # if i > 0: commentout for sdxl
-            # no attention layers in up_blocks.0
-            hf_up_atn_prefix = f"up_blocks.{i}.attentions.{j}."
-            sd_up_atn_prefix = f"output_blocks.{3*i + j}.1."
-            unet_conversion_map_layer.append((sd_up_atn_prefix, hf_up_atn_prefix))
-
-        if i < 3:
-            # no downsample in down_blocks.3
-            hf_downsample_prefix = f"down_blocks.{i}.downsamplers.0.conv."
-            sd_downsample_prefix = f"input_blocks.{3*(i+1)}.0.op."
-            unet_conversion_map_layer.append((sd_downsample_prefix, hf_downsample_prefix))
-
-            # no upsample in up_blocks.3
-            hf_upsample_prefix = f"up_blocks.{i}.upsamplers.0."
-            sd_upsample_prefix = f"output_blocks.{3*i + 2}.{2}."  # change for sdxl
-            unet_conversion_map_layer.append((sd_upsample_prefix, hf_upsample_prefix))
-
-    hf_mid_atn_prefix = "mid_block.attentions.0."
-    sd_mid_atn_prefix = "middle_block.1."
-    unet_conversion_map_layer.append((sd_mid_atn_prefix, hf_mid_atn_prefix))
-
-    for j in range(2):
-        hf_mid_res_prefix = f"mid_block.resnets.{j}."
-        sd_mid_res_prefix = f"middle_block.{2*j}."
-        unet_conversion_map_layer.append((sd_mid_res_prefix, hf_mid_res_prefix))
-
-    unet_conversion_map_resnet = [
-        # (stable-diffusion, HF Diffusers)
-        ("in_layers.0.", "norm1."),
-        ("in_layers.2.", "conv1."),
-        ("out_layers.0.", "norm2."),
-        ("out_layers.3.", "conv2."),
-        ("emb_layers.1.", "time_emb_proj."),
-        ("skip_connection.", "conv_shortcut."),
-    ]
-
-    unet_conversion_map = []
-    for sd, hf in unet_conversion_map_layer:
-        if "resnets" in hf:
-            for sd_res, hf_res in unet_conversion_map_resnet:
-                unet_conversion_map.append((sd + sd_res, hf + hf_res))
-        else:
-            unet_conversion_map.append((sd, hf))
-
-    for j in range(2):
-        hf_time_embed_prefix = f"time_embedding.linear_{j+1}."
-        sd_time_embed_prefix = f"time_embed.{j*2}."
-        unet_conversion_map.append((sd_time_embed_prefix, hf_time_embed_prefix))
-
-    for j in range(2):
-        hf_label_embed_prefix = f"add_embedding.linear_{j+1}."
-        sd_label_embed_prefix = f"label_emb.0.{j*2}."
-        unet_conversion_map.append((sd_label_embed_prefix, hf_label_embed_prefix))
-
-    unet_conversion_map.append(("input_blocks.0.0.", "conv_in."))
-    unet_conversion_map.append(("out.0.", "conv_norm_out."))
-    unet_conversion_map.append(("out.2.", "conv_out."))
-
-    return unet_conversion_map
-
-
-SDXL_UNET_STABILITY_TO_DIFFUSERS_MAP = {
-    sd.rstrip(".").replace(".", "_"): hf.rstrip(".").replace(".", "_") for sd, hf in make_sdxl_unet_conversion_map()
-}

invokeai/backend/model_manager/load/model_loaders/lora.py

@@ -5,8 +5,10 @@ from logging import Logger
 from pathlib import Path
 from typing import Optional
 
+import torch
+from safetensors.torch import load_file
+
 from invokeai.app.services.config import InvokeAIAppConfig
-from invokeai.backend.lora import LoRAModelRaw
 from invokeai.backend.model_manager import (
     AnyModel,
     AnyModelConfig,
@@ -18,6 +20,11 @@ from invokeai.backend.model_manager import (
 from invokeai.backend.model_manager.load.load_default import ModelLoader
 from invokeai.backend.model_manager.load.model_cache.model_cache_base import ModelCacheBase
 from invokeai.backend.model_manager.load.model_loader_registry import ModelLoaderRegistry
+from invokeai.backend.peft.conversions.flux_kohya_lora_conversion_utils import (
+    lora_model_from_flux_kohya_state_dict,
+)
+from invokeai.backend.peft.conversions.sd_lora_conversion_utils import lora_model_from_sd_state_dict
+from invokeai.backend.peft.conversions.sdxl_lora_conversion_utils import convert_sdxl_keys_to_diffusers_format
 
 
 @ModelLoaderRegistry.register(base=BaseModelType.Any, type=ModelType.LoRA, format=ModelFormat.Diffusers)
@@ -45,14 +52,28 @@ class LoRALoader(ModelLoader):
             raise ValueError("There are no submodels in a LoRA model.")
         model_path = Path(config.path)
         assert self._model_base is not None
-        model = LoRAModelRaw.from_checkpoint(
-            file_path=model_path,
-            dtype=self._torch_dtype,
-            base_model=self._model_base,
-        )
+
+        # Load the state dict from the model file.
+        if model_path.suffix == ".safetensors":
+            state_dict = load_file(model_path.absolute().as_posix(), device="cpu")
+        else:
+            state_dict = torch.load(model_path, map_location="cpu")
+
+        # Apply state_dict key conversions, if necessary.
+        if self._model_base == BaseModelType.StableDiffusionXL:
+            state_dict = convert_sdxl_keys_to_diffusers_format(state_dict)
+            model = lora_model_from_sd_state_dict(state_dict=state_dict)
+        elif self._model_base == BaseModelType.Flux:
+            model = lora_model_from_flux_kohya_state_dict(state_dict=state_dict)
+        elif self._model_base in [BaseModelType.StableDiffusion1, BaseModelType.StableDiffusion2]:
+            # Currently, we don't apply any conversions for SD1 and SD2 LoRA models.
+            model = lora_model_from_sd_state_dict(state_dict=state_dict)
+        else:
+            raise ValueError(f"Unsupported LoRA base model: {self._model_base}")
+
+        model.to(dtype=self._torch_dtype)
         return model
 
-    # override
     def _get_model_path(self, config: AnyModelConfig) -> Path:
         # cheating a little - we remember this variable for using in the subsequent call to _load_model()
         self._model_base = config.base

invokeai/backend/model_manager/load/model_util.py

@@ -15,9 +15,9 @@ from invokeai.backend.image_util.depth_anything.depth_anything_pipeline import D
 from invokeai.backend.image_util.grounding_dino.grounding_dino_pipeline import GroundingDinoPipeline
 from invokeai.backend.image_util.segment_anything.segment_anything_pipeline import SegmentAnythingPipeline
 from invokeai.backend.ip_adapter.ip_adapter import IPAdapter
-from invokeai.backend.lora import LoRAModelRaw
 from invokeai.backend.model_manager.config import AnyModel
 from invokeai.backend.onnx.onnx_runtime import IAIOnnxRuntimeModel
+from invokeai.backend.peft.lora import LoRAModelRaw
 from invokeai.backend.spandrel_image_to_image_model import SpandrelImageToImageModel
 from invokeai.backend.textual_inversion import TextualInversionModelRaw
 

invokeai/backend/model_manager/probe.py

@@ -26,6 +26,7 @@ from invokeai.backend.model_manager.config import (
     SchedulerPredictionType,
 )
 from invokeai.backend.model_manager.util.model_util import lora_token_vector_length, read_checkpoint_meta
+from invokeai.backend.peft.conversions.flux_kohya_lora_conversion_utils import is_state_dict_likely_in_flux_kohya_format
 from invokeai.backend.spandrel_image_to_image_model import SpandrelImageToImageModel
 from invokeai.backend.util.silence_warnings import SilenceWarnings
 
@@ -528,9 +529,11 @@ class LoRACheckpointProbe(CheckpointProbeBase):
         return ModelFormat("lycoris")
 
     def get_base_type(self) -> BaseModelType:
-        checkpoint = self.checkpoint
-        token_vector_length = lora_token_vector_length(checkpoint)
+        if is_state_dict_likely_in_flux_kohya_format(self.checkpoint):
+            return BaseModelType.Flux
 
+        # If we've gotten here, we assume that the model is a Stable Diffusion model.
+        token_vector_length = lora_token_vector_length(self.checkpoint)
         if token_vector_length == 768:
             return BaseModelType.StableDiffusion1
         elif token_vector_length == 1024:

invokeai/backend/model_patcher.py

@@ -13,10 +13,10 @@ from diffusers import OnnxRuntimeModel, UNet2DConditionModel
 from transformers import CLIPTextModel, CLIPTextModelWithProjection, CLIPTokenizer
 
 from invokeai.app.shared.models import FreeUConfig
-from invokeai.backend.lora import LoRAModelRaw
 from invokeai.backend.model_manager import AnyModel
 from invokeai.backend.model_manager.load.optimizations import skip_torch_weight_init
 from invokeai.backend.onnx.onnx_runtime import IAIOnnxRuntimeModel
+from invokeai.backend.peft.lora import LoRAModelRaw
 from invokeai.backend.stable_diffusion.extensions.lora import LoRAExt
 from invokeai.backend.textual_inversion import TextualInversionManager, TextualInversionModelRaw
 from invokeai.backend.util.original_weights_storage import OriginalWeightsStorage

invokeai/backend/peft/conversions/flux_kohya_lora_conversion_utils.py

@@ -0,0 +1,84 @@
+import re
+from typing import Any, Dict, TypeVar
+
+import torch
+
+from invokeai.backend.peft.layers.any_lora_layer import AnyLoRALayer
+from invokeai.backend.peft.layers.utils import peft_layer_from_state_dict
+from invokeai.backend.peft.lora import LoRAModelRaw
+
+# A regex pattern that matches all of the keys in the Kohya FLUX LoRA format.
+# Example keys:
+#   lora_unet_double_blocks_0_img_attn_proj.alpha
+#   lora_unet_double_blocks_0_img_attn_proj.lora_down.weight
+#   lora_unet_double_blocks_0_img_attn_proj.lora_up.weight
+FLUX_KOHYA_KEY_REGEX = (
+    r"lora_unet_(\w+_blocks)_(\d+)_(img_attn|img_mlp|img_mod|txt_attn|txt_mlp|txt_mod|linear1|linear2|modulation)_?(.*)"
+)
+
+
+def is_state_dict_likely_in_flux_kohya_format(state_dict: Dict[str, Any]) -> bool:
+    """Checks if the provided state dict is likely in the Kohya FLUX LoRA format.
+
+    This is intended to be a high-precision detector, but it is not guaranteed to have perfect precision. (A
+    perfect-precision detector would require checking all keys against a whitelist and verifying tensor shapes.)
+    """
+    for k in state_dict.keys():
+        if not re.match(FLUX_KOHYA_KEY_REGEX, k):
+            return False
+    return True
+
+
+def lora_model_from_flux_kohya_state_dict(state_dict: Dict[str, torch.Tensor]) -> LoRAModelRaw:
+    # Group keys by layer.
+    grouped_state_dict: dict[str, dict[str, torch.Tensor]] = {}
+    for key, value in state_dict.items():
+        layer_name, param_name = key.split(".", 1)
+        if layer_name not in grouped_state_dict:
+            grouped_state_dict[layer_name] = {}
+        grouped_state_dict[layer_name][param_name] = value
+
+    # Convert the state dict to the InvokeAI format.
+    grouped_state_dict = convert_flux_kohya_state_dict_to_invoke_format(grouped_state_dict)
+
+    # Create LoRA layers.
+    layers: dict[str, AnyLoRALayer] = {}
+    for layer_key, layer_state_dict in grouped_state_dict.items():
+        layer = peft_layer_from_state_dict(layer_key, layer_state_dict)
+        layers[layer_key] = layer
+
+    # Create and return the LoRAModelRaw.
+    return LoRAModelRaw(layers=layers)
+
+
+T = TypeVar("T")
+
+
+def convert_flux_kohya_state_dict_to_invoke_format(state_dict: Dict[str, T]) -> Dict[str, T]:
+    """Converts a state dict from the Kohya FLUX LoRA format to LoRA weight format used internally by InvokeAI.
+
+    Example key conversions:
+    "lora_unet_double_blocks_0_img_attn_proj" -> "double_blocks.0.img_attn.proj"
+    "lora_unet_double_blocks_0_img_attn_proj" -> "double_blocks.0.img_attn.proj"
+    "lora_unet_double_blocks_0_img_attn_proj" -> "double_blocks.0.img_attn.proj"
+    "lora_unet_double_blocks_0_img_attn_qkv" -> "double_blocks.0.img_attn.qkv"
+    "lora_unet_double_blocks_0_img_attn_qkv" -> "double_blocks.0.img.attn.qkv"
+    "lora_unet_double_blocks_0_img_attn_qkv" -> "double_blocks.0.img.attn.qkv"
+    """
+
+    def replace_func(match: re.Match[str]) -> str:
+        s = f"{match.group(1)}.{match.group(2)}.{match.group(3)}"
+        if match.group(4):
+            s += f".{match.group(4)}"
+        return s
+
+    converted_dict: dict[str, T] = {}
+    for k, v in state_dict.items():
+        match = re.match(FLUX_KOHYA_KEY_REGEX, k)
+        if match:
+            new_key = re.sub(FLUX_KOHYA_KEY_REGEX, replace_func, k)
+            converted_dict[new_key] = v
+        else:
+            raise ValueError(f"Key '{k}' does not match the expected pattern for FLUX LoRA weights.")
+
+    return converted_dict

invokeai/backend/peft/conversions/sd_lora_conversion_utils.py

@@ -0,0 +1,30 @@
+from typing import Dict
+
+import torch
+
+from invokeai.backend.peft.layers.any_lora_layer import AnyLoRALayer
+from invokeai.backend.peft.layers.utils import peft_layer_from_state_dict
+from invokeai.backend.peft.lora import LoRAModelRaw
+
+
+def lora_model_from_sd_state_dict(state_dict: Dict[str, torch.Tensor]) -> LoRAModelRaw:
+    grouped_state_dict: dict[str, dict[str, torch.Tensor]] = _group_state(state_dict)
+
+    layers: dict[str, AnyLoRALayer] = {}
+    for layer_key, values in grouped_state_dict.items():
+        layer = peft_layer_from_state_dict(layer_key, values)
+        layers[layer_key] = layer
+
+    return LoRAModelRaw(layers=layers)
+
+
+def _group_state(state_dict: Dict[str, torch.Tensor]) -> Dict[str, Dict[str, torch.Tensor]]:
+    state_dict_groupped: Dict[str, Dict[str, torch.Tensor]] = {}
+
+    for key, value in state_dict.items():
+        stem, leaf = key.split(".", 1)
+        if stem not in state_dict_groupped:
+            state_dict_groupped[stem] = {}
+        state_dict_groupped[stem][leaf] = value
+
+    return state_dict_groupped

invokeai/backend/peft/conversions/sdxl_lora_conversion_utils.py

@@ -0,0 +1,154 @@
+import bisect
+from typing import Dict, List, Tuple, TypeVar
+
+T = TypeVar("T")
+
+
+def convert_sdxl_keys_to_diffusers_format(state_dict: Dict[str, T]) -> dict[str, T]:
+    """Convert the keys of an SDXL LoRA state_dict to diffusers format.
+
+    The input state_dict can be in either Stability AI format or diffusers format. If the state_dict is already in
+    diffusers format, then this function will have no effect.
+
+    This function is adapted from:
+    https://github.com/bmaltais/kohya_ss/blob/2accb1305979ba62f5077a23aabac23b4c37e935/networks/lora_diffusers.py#L385-L409
+
+    Args:
+        state_dict (Dict[str, Tensor]): The SDXL LoRA state_dict.
+
+    Raises:
+        ValueError: If state_dict contains an unrecognized key, or not all keys could be converted.
+
+    Returns:
+        Dict[str, Tensor]: The diffusers-format state_dict.
+    """
+    converted_count = 0  # The number of Stability AI keys converted to diffusers format.
+    not_converted_count = 0  # The number of keys that were not converted.
+
+    # Get a sorted list of Stability AI UNet keys so that we can efficiently search for keys with matching prefixes.
+    # For example, we want to efficiently find `input_blocks_4_1` in the list when searching for
+    # `input_blocks_4_1_proj_in`.
+    stability_unet_keys = list(SDXL_UNET_STABILITY_TO_DIFFUSERS_MAP)
+    stability_unet_keys.sort()
+
+    new_state_dict: dict[str, T] = {}
+    for full_key, value in state_dict.items():
+        if full_key.startswith("lora_unet_"):
+            search_key = full_key.replace("lora_unet_", "")
+            # Use bisect to find the key in stability_unet_keys that *may* match the search_key's prefix.
+            position = bisect.bisect_right(stability_unet_keys, search_key)
+            map_key = stability_unet_keys[position - 1]
+            # Now, check if the map_key *actually* matches the search_key.
+            if search_key.startswith(map_key):
+                new_key = full_key.replace(map_key, SDXL_UNET_STABILITY_TO_DIFFUSERS_MAP[map_key])
+                new_state_dict[new_key] = value
+                converted_count += 1
+            else:
+                new_state_dict[full_key] = value
+                not_converted_count += 1
+        elif full_key.startswith("lora_te1_") or full_key.startswith("lora_te2_"):
+            # The CLIP text encoders have the same keys in both Stability AI and diffusers formats.
+            new_state_dict[full_key] = value
+            continue
+        else:
+            raise ValueError(f"Unrecognized SDXL LoRA key prefix: '{full_key}'.")
+
+    if converted_count > 0 and not_converted_count > 0:
+        raise ValueError(
+            f"The SDXL LoRA could only be partially converted to diffusers format. converted={converted_count},"
+            f" not_converted={not_converted_count}"
+        )
+
+    return new_state_dict
+
+
+# code from
+# https://github.com/bmaltais/kohya_ss/blob/2accb1305979ba62f5077a23aabac23b4c37e935/networks/lora_diffusers.py#L15C1-L97C32
+def _make_sdxl_unet_conversion_map() -> List[Tuple[str, str]]:
+    """Create a dict mapping state_dict keys from Stability AI SDXL format to diffusers SDXL format."""
+    unet_conversion_map_layer: list[tuple[str, str]] = []
+
+    for i in range(3):  # num_blocks is 3 in sdxl
+        # loop over downblocks/upblocks
+        for j in range(2):
+            # loop over resnets/attentions for downblocks
+            hf_down_res_prefix = f"down_blocks.{i}.resnets.{j}."
+            sd_down_res_prefix = f"input_blocks.{3*i + j + 1}.0."
+            unet_conversion_map_layer.append((sd_down_res_prefix, hf_down_res_prefix))
+
+            if i < 3:
+                # no attention layers in down_blocks.3
+                hf_down_atn_prefix = f"down_blocks.{i}.attentions.{j}."
+                sd_down_atn_prefix = f"input_blocks.{3*i + j + 1}.1."
+                unet_conversion_map_layer.append((sd_down_atn_prefix, hf_down_atn_prefix))
+
+        for j in range(3):
+            # loop over resnets/attentions for upblocks
+            hf_up_res_prefix = f"up_blocks.{i}.resnets.{j}."
+            sd_up_res_prefix = f"output_blocks.{3*i + j}.0."
+            unet_conversion_map_layer.append((sd_up_res_prefix, hf_up_res_prefix))
+
+            # if i > 0: commentout for sdxl
+            # no attention layers in up_blocks.0
+            hf_up_atn_prefix = f"up_blocks.{i}.attentions.{j}."
+            sd_up_atn_prefix = f"output_blocks.{3*i + j}.1."
+            unet_conversion_map_layer.append((sd_up_atn_prefix, hf_up_atn_prefix))
+
+        if i < 3:
+            # no downsample in down_blocks.3
+            hf_downsample_prefix = f"down_blocks.{i}.downsamplers.0.conv."
+            sd_downsample_prefix = f"input_blocks.{3*(i+1)}.0.op."
+            unet_conversion_map_layer.append((sd_downsample_prefix, hf_downsample_prefix))
+
+            # no upsample in up_blocks.3
+            hf_upsample_prefix = f"up_blocks.{i}.upsamplers.0."
+            sd_upsample_prefix = f"output_blocks.{3*i + 2}.{2}."  # change for sdxl
+            unet_conversion_map_layer.append((sd_upsample_prefix, hf_upsample_prefix))
+
+    hf_mid_atn_prefix = "mid_block.attentions.0."
+    sd_mid_atn_prefix = "middle_block.1."
+    unet_conversion_map_layer.append((sd_mid_atn_prefix, hf_mid_atn_prefix))
+
+    for j in range(2):
+        hf_mid_res_prefix = f"mid_block.resnets.{j}."
+        sd_mid_res_prefix = f"middle_block.{2*j}."
+        unet_conversion_map_layer.append((sd_mid_res_prefix, hf_mid_res_prefix))
+
+    unet_conversion_map_resnet = [
+        # (stable-diffusion, HF Diffusers)
+        ("in_layers.0.", "norm1."),
+        ("in_layers.2.", "conv1."),
+        ("out_layers.0.", "norm2."),
+        ("out_layers.3.", "conv2."),
+        ("emb_layers.1.", "time_emb_proj."),
+        ("skip_connection.", "conv_shortcut."),
+    ]
+
+    unet_conversion_map: list[tuple[str, str]] = []
+    for sd, hf in unet_conversion_map_layer:
+        if "resnets" in hf:
+            for sd_res, hf_res in unet_conversion_map_resnet:
+                unet_conversion_map.append((sd + sd_res, hf + hf_res))
+        else:
+            unet_conversion_map.append((sd, hf))
+
+    for j in range(2):
+        hf_time_embed_prefix = f"time_embedding.linear_{j+1}."
+        sd_time_embed_prefix = f"time_embed.{j*2}."
+        unet_conversion_map.append((sd_time_embed_prefix, hf_time_embed_prefix))
+
+    for j in range(2):
+        hf_label_embed_prefix = f"add_embedding.linear_{j+1}."
+        sd_label_embed_prefix = f"label_emb.0.{j*2}."
+        unet_conversion_map.append((sd_label_embed_prefix, hf_label_embed_prefix))
+
+    unet_conversion_map.append(("input_blocks.0.0.", "conv_in."))
+    unet_conversion_map.append(("out.0.", "conv_norm_out."))
+    unet_conversion_map.append(("out.2.", "conv_out."))
+
+    return unet_conversion_map
+
+
+SDXL_UNET_STABILITY_TO_DIFFUSERS_MAP = {
+    sd.rstrip(".").replace(".", "_"): hf.rstrip(".").replace(".", "_") for sd, hf in _make_sdxl_unet_conversion_map()
+}

invokeai/backend/peft/layers/any_lora_layer.py

@@ -0,0 +1,10 @@
+from typing import Union
+
+from invokeai.backend.peft.layers.full_layer import FullLayer
+from invokeai.backend.peft.layers.ia3_layer import IA3Layer
+from invokeai.backend.peft.layers.loha_layer import LoHALayer
+from invokeai.backend.peft.layers.lokr_layer import LoKRLayer
+from invokeai.backend.peft.layers.lora_layer import LoRALayer
+from invokeai.backend.peft.layers.norm_layer import NormLayer
+
+AnyLoRALayer = Union[LoRALayer, LoHALayer, LoKRLayer, FullLayer, IA3Layer, NormLayer]

invokeai/backend/peft/layers/full_layer.py

@@ -0,0 +1,37 @@
+from typing import Dict, Optional
+
+import torch
+
+from invokeai.backend.peft.layers.lora_layer_base import LoRALayerBase
+
+
+class FullLayer(LoRALayerBase):
+    # bias handled in LoRALayerBase(calc_size, to)
+    # weight: torch.Tensor
+    # bias: Optional[torch.Tensor]
+
+    def __init__(
+        self,
+        layer_key: str,
+        values: Dict[str, torch.Tensor],
+    ):
+        super().__init__(layer_key, values)
+
+        self.weight = values["diff"]
+        self.bias = values.get("diff_b", None)
+
+        self.rank = None  # unscaled
+        self.check_keys(values, {"diff", "diff_b"})
+
+    def get_weight(self, orig_weight: torch.Tensor) -> torch.Tensor:
+        return self.weight
+
+    def calc_size(self) -> int:
+        model_size = super().calc_size()
+        model_size += self.weight.nelement() * self.weight.element_size()
+        return model_size
+
+    def to(self, device: Optional[torch.device] = None, dtype: Optional[torch.dtype] = None) -> None:
+        super().to(device=device, dtype=dtype)
+
+        self.weight = self.weight.to(device=device, dtype=dtype)

invokeai/backend/peft/layers/ia3_layer.py

@@ -0,0 +1,42 @@
+from typing import Dict, Optional
+
+import torch
+
+from invokeai.backend.peft.layers.lora_layer_base import LoRALayerBase
+
+
+class IA3Layer(LoRALayerBase):
+    # weight: torch.Tensor
+    # on_input: torch.Tensor
+
+    def __init__(
+        self,
+        layer_key: str,
+        values: Dict[str, torch.Tensor],
+    ):
+        super().__init__(layer_key, values)
+
+        self.weight = values["weight"]
+        self.on_input = values["on_input"]
+
+        self.rank = None  # unscaled
+        self.check_keys(values, {"weight", "on_input"})
+
+    def get_weight(self, orig_weight: torch.Tensor) -> torch.Tensor:
+        weight = self.weight
+        if not self.on_input:
+            weight = weight.reshape(-1, 1)
+        assert orig_weight is not None
+        return orig_weight * weight
+
+    def calc_size(self) -> int:
+        model_size = super().calc_size()
+        model_size += self.weight.nelement() * self.weight.element_size()
+        model_size += self.on_input.nelement() * self.on_input.element_size()
+        return model_size
+
+    def to(self, device: Optional[torch.device] = None, dtype: Optional[torch.dtype] = None):
+        super().to(device=device, dtype=dtype)
+
+        self.weight = self.weight.to(device=device, dtype=dtype)
+        self.on_input = self.on_input.to(device=device, dtype=dtype)

invokeai/backend/peft/layers/loha_layer.py

@@ -0,0 +1,68 @@
+from typing import Dict, Optional
+
+import torch
+
+from invokeai.backend.peft.layers.lora_layer_base import LoRALayerBase
+
+
+class LoHALayer(LoRALayerBase):
+    # w1_a: torch.Tensor
+    # w1_b: torch.Tensor
+    # w2_a: torch.Tensor
+    # w2_b: torch.Tensor
+    # t1: Optional[torch.Tensor] = None
+    # t2: Optional[torch.Tensor] = None
+
+    def __init__(self, layer_key: str, values: Dict[str, torch.Tensor]):
+        super().__init__(layer_key, values)
+
+        self.w1_a = values["hada_w1_a"]
+        self.w1_b = values["hada_w1_b"]
+        self.w2_a = values["hada_w2_a"]
+        self.w2_b = values["hada_w2_b"]
+        self.t1 = values.get("hada_t1", None)
+        self.t2 = values.get("hada_t2", None)
+
+        self.rank = self.w1_b.shape[0]
+        self.check_keys(
+            values,
+            {
+                "hada_w1_a",
+                "hada_w1_b",
+                "hada_w2_a",
+                "hada_w2_b",
+                "hada_t1",
+                "hada_t2",
+            },
+        )
+
+    def get_weight(self, orig_weight: torch.Tensor) -> torch.Tensor:
+        if self.t1 is None:
+            weight: torch.Tensor = (self.w1_a @ self.w1_b) * (self.w2_a @ self.w2_b)
+
+        else:
+            rebuild1 = torch.einsum("i j k l, j r, i p -> p r k l", self.t1, self.w1_b, self.w1_a)
+            rebuild2 = torch.einsum("i j k l, j r, i p -> p r k l", self.t2, self.w2_b, self.w2_a)
+            weight = rebuild1 * rebuild2
+
+        return weight
+
+    def calc_size(self) -> int:
+        model_size = super().calc_size()
+        for val in [self.w1_a, self.w1_b, self.w2_a, self.w2_b, self.t1, self.t2]:
+            if val is not None:
+                model_size += val.nelement() * val.element_size()
+        return model_size
+
+    def to(self, device: Optional[torch.device] = None, dtype: Optional[torch.dtype] = None) -> None:
+        super().to(device=device, dtype=dtype)
+
+        self.w1_a = self.w1_a.to(device=device, dtype=dtype)
+        self.w1_b = self.w1_b.to(device=device, dtype=dtype)
+        if self.t1 is not None:
+            self.t1 = self.t1.to(device=device, dtype=dtype)
+
+        self.w2_a = self.w2_a.to(device=device, dtype=dtype)
+        self.w2_b = self.w2_b.to(device=device, dtype=dtype)
+        if self.t2 is not None:
+            self.t2 = self.t2.to(device=device, dtype=dtype)

invokeai/backend/peft/layers/lokr_layer.py

@@ -0,0 +1,114 @@
+from typing import Dict, Optional
+
+import torch
+
+from invokeai.backend.peft.layers.lora_layer_base import LoRALayerBase
+
+
+class LoKRLayer(LoRALayerBase):
+    # w1: Optional[torch.Tensor] = None
+    # w1_a: Optional[torch.Tensor] = None
+    # w1_b: Optional[torch.Tensor] = None
+    # w2: Optional[torch.Tensor] = None
+    # w2_a: Optional[torch.Tensor] = None
+    # w2_b: Optional[torch.Tensor] = None
+    # t2: Optional[torch.Tensor] = None
+
+    def __init__(
+        self,
+        layer_key: str,
+        values: Dict[str, torch.Tensor],
+    ):
+        super().__init__(layer_key, values)
+
+        self.w1 = values.get("lokr_w1", None)
+        if self.w1 is None:
+            self.w1_a = values["lokr_w1_a"]
+            self.w1_b = values["lokr_w1_b"]
+        else:
+            self.w1_b = None
+            self.w1_a = None
+
+        self.w2 = values.get("lokr_w2", None)
+        if self.w2 is None:
+            self.w2_a = values["lokr_w2_a"]
+            self.w2_b = values["lokr_w2_b"]
+        else:
+            self.w2_a = None
+            self.w2_b = None
+
+        self.t2 = values.get("lokr_t2", None)
+
+        if self.w1_b is not None:
+            self.rank = self.w1_b.shape[0]
+        elif self.w2_b is not None:
+            self.rank = self.w2_b.shape[0]
+        else:
+            self.rank = None  # unscaled
+
+        self.check_keys(
+            values,
+            {
+                "lokr_w1",
+                "lokr_w1_a",
+                "lokr_w1_b",
+                "lokr_w2",
+                "lokr_w2_a",
+                "lokr_w2_b",
+                "lokr_t2",
+            },
+        )
+
+    def get_weight(self, orig_weight: torch.Tensor) -> torch.Tensor:
+        w1: Optional[torch.Tensor] = self.w1
+        if w1 is None:
+            assert self.w1_a is not None
+            assert self.w1_b is not None
+            w1 = self.w1_a @ self.w1_b
+
+        w2 = self.w2
+        if w2 is None:
+            if self.t2 is None:
+                assert self.w2_a is not None
+                assert self.w2_b is not None
+                w2 = self.w2_a @ self.w2_b
+            else:
+                w2 = torch.einsum("i j k l, i p, j r -> p r k l", self.t2, self.w2_a, self.w2_b)
+
+        if len(w2.shape) == 4:
+            w1 = w1.unsqueeze(2).unsqueeze(2)
+        w2 = w2.contiguous()
+        assert w1 is not None
+        assert w2 is not None
+        weight = torch.kron(w1, w2)
+
+        return weight
+
+    def calc_size(self) -> int:
+        model_size = super().calc_size()
+        for val in [self.w1, self.w1_a, self.w1_b, self.w2, self.w2_a, self.w2_b, self.t2]:
+            if val is not None:
+                model_size += val.nelement() * val.element_size()
+        return model_size
+
+    def to(self, device: Optional[torch.device] = None, dtype: Optional[torch.dtype] = None) -> None:
+        super().to(device=device, dtype=dtype)
+
+        if self.w1 is not None:
+            self.w1 = self.w1.to(device=device, dtype=dtype)
+        else:
+            assert self.w1_a is not None
+            assert self.w1_b is not None
+            self.w1_a = self.w1_a.to(device=device, dtype=dtype)
+            self.w1_b = self.w1_b.to(device=device, dtype=dtype)
+
+        if self.w2 is not None:
+            self.w2 = self.w2.to(device=device, dtype=dtype)
+        else:
+            assert self.w2_a is not None
+            assert self.w2_b is not None
+            self.w2_a = self.w2_a.to(device=device, dtype=dtype)
+            self.w2_b = self.w2_b.to(device=device, dtype=dtype)
+
+        if self.t2 is not None:
+            self.t2 = self.t2.to(device=device, dtype=dtype)

invokeai/backend/peft/layers/lora_layer.py

@@ -0,0 +1,59 @@
+from typing import Dict, Optional
+
+import torch
+
+from invokeai.backend.peft.layers.lora_layer_base import LoRALayerBase
+
+
+# TODO: find and debug lora/locon with bias
+class LoRALayer(LoRALayerBase):
+    # up: torch.Tensor
+    # mid: Optional[torch.Tensor]
+    # down: torch.Tensor
+
+    def __init__(
+        self,
+        layer_key: str,
+        values: Dict[str, torch.Tensor],
+    ):
+        super().__init__(layer_key, values)
+
+        self.up = values["lora_up.weight"]
+        self.down = values["lora_down.weight"]
+        self.mid = values.get("lora_mid.weight", None)
+
+        self.rank = self.down.shape[0]
+        self.check_keys(
+            values,
+            {
+                "lora_up.weight",
+                "lora_down.weight",
+                "lora_mid.weight",
+            },
+        )
+
+    def get_weight(self, orig_weight: torch.Tensor) -> torch.Tensor:
+        if self.mid is not None:
+            up = self.up.reshape(self.up.shape[0], self.up.shape[1])
+            down = self.down.reshape(self.down.shape[0], self.down.shape[1])
+            weight = torch.einsum("m n w h, i m, n j -> i j w h", self.mid, up, down)
+        else:
+            weight = self.up.reshape(self.up.shape[0], -1) @ self.down.reshape(self.down.shape[0], -1)
+
+        return weight
+
+    def calc_size(self) -> int:
+        model_size = super().calc_size()
+        for val in [self.up, self.mid, self.down]:
+            if val is not None:
+                model_size += val.nelement() * val.element_size()
+        return model_size
+
+    def to(self, device: Optional[torch.device] = None, dtype: Optional[torch.dtype] = None) -> None:
+        super().to(device=device, dtype=dtype)
+
+        self.up = self.up.to(device=device, dtype=dtype)
+        self.down = self.down.to(device=device, dtype=dtype)
+
+        if self.mid is not None:
+            self.mid = self.mid.to(device=device, dtype=dtype)

invokeai/backend/peft/layers/lora_layer_base.py

@@ -0,0 +1,74 @@
+from typing import Dict, Optional, Set
+
+import torch
+
+import invokeai.backend.util.logging as logger
+
+
+class LoRALayerBase:
+    # rank: Optional[int]
+    # alpha: Optional[float]
+    # bias: Optional[torch.Tensor]
+    # layer_key: str
+
+    # @property
+    # def scale(self):
+    #    return self.alpha / self.rank if (self.alpha and self.rank) else 1.0
+
+    def __init__(
+        self,
+        layer_key: str,
+        values: Dict[str, torch.Tensor],
+    ):
+        if "alpha" in values:
+            self.alpha = values["alpha"].item()
+        else:
+            self.alpha = None
+
+        if "bias_indices" in values and "bias_values" in values and "bias_size" in values:
+            self.bias: Optional[torch.Tensor] = torch.sparse_coo_tensor(
+                values["bias_indices"],
+                values["bias_values"],
+                tuple(values["bias_size"]),
+            )
+
+        else:
+            self.bias = None
+
+        self.rank = None  # set in layer implementation
+        self.layer_key = layer_key
+
+    def get_weight(self, orig_weight: torch.Tensor) -> torch.Tensor:
+        raise NotImplementedError()
+
+    def get_bias(self, orig_bias: torch.Tensor) -> Optional[torch.Tensor]:
+        return self.bias
+
+    def get_parameters(self, orig_module: torch.nn.Module) -> Dict[str, torch.Tensor]:
+        params = {"weight": self.get_weight(orig_module.weight)}
+        bias = self.get_bias(orig_module.bias)
+        if bias is not None:
+            params["bias"] = bias
+        return params
+
+    def calc_size(self) -> int:
+        model_size = 0
+        for val in [self.bias]:
+            if val is not None:
+                model_size += val.nelement() * val.element_size()
+        return model_size
+
+    def to(self, device: Optional[torch.device] = None, dtype: Optional[torch.dtype] = None) -> None:
+        if self.bias is not None:
+            self.bias = self.bias.to(device=device, dtype=dtype)
+
+    def check_keys(self, values: Dict[str, torch.Tensor], known_keys: Set[str]):
+        """Log a warning if values contains unhandled keys."""
+        # {"alpha", "bias_indices", "bias_values", "bias_size"} are hard-coded, because they are handled by
+        # `LoRALayerBase`. Sub-classes should provide the known_keys that they handled.
+        all_known_keys = known_keys | {"alpha", "bias_indices", "bias_values", "bias_size"}
+        unknown_keys = set(values.keys()) - all_known_keys
+        if unknown_keys:
+            logger.warning(
+                f"Unexpected keys found in LoRA/LyCORIS layer, model might work incorrectly! Keys: {unknown_keys}"
+            )

invokeai/backend/peft/layers/norm_layer.py

@@ -0,0 +1,37 @@
+from typing import Dict, Optional
+
+import torch
+
+from invokeai.backend.peft.layers.lora_layer_base import LoRALayerBase
+
+
+class NormLayer(LoRALayerBase):
+    # bias handled in LoRALayerBase(calc_size, to)
+    # weight: torch.Tensor
+    # bias: Optional[torch.Tensor]
+
+    def __init__(
+        self,
+        layer_key: str,
+        values: Dict[str, torch.Tensor],
+    ):
+        super().__init__(layer_key, values)
+
+        self.weight = values["w_norm"]
+        self.bias = values.get("b_norm", None)
+
+        self.rank = None  # unscaled
+        self.check_keys(values, {"w_norm", "b_norm"})
+
+    def get_weight(self, orig_weight: torch.Tensor) -> torch.Tensor:
+        return self.weight
+
+    def calc_size(self) -> int:
+        model_size = super().calc_size()
+        model_size += self.weight.nelement() * self.weight.element_size()
+        return model_size
+
+    def to(self, device: Optional[torch.device] = None, dtype: Optional[torch.dtype] = None) -> None:
+        super().to(device=device, dtype=dtype)
+
+        self.weight = self.weight.to(device=device, dtype=dtype)

invokeai/backend/peft/layers/utils.py

@@ -0,0 +1,33 @@
+from typing import Dict
+
+import torch
+
+from invokeai.backend.peft.layers.any_lora_layer import AnyLoRALayer
+from invokeai.backend.peft.layers.full_layer import FullLayer
+from invokeai.backend.peft.layers.ia3_layer import IA3Layer
+from invokeai.backend.peft.layers.loha_layer import LoHALayer
+from invokeai.backend.peft.layers.lokr_layer import LoKRLayer
+from invokeai.backend.peft.layers.lora_layer import LoRALayer
+from invokeai.backend.peft.layers.norm_layer import NormLayer
+
+
+def peft_layer_from_state_dict(layer_key: str, state_dict: Dict[str, torch.Tensor]) -> AnyLoRALayer:
+    # Detect layers according to LyCORIS detection logic(`weight_list_det`)
+    # https://github.com/KohakuBlueleaf/LyCORIS/tree/8ad8000efb79e2b879054da8c9356e6143591bad/lycoris/modules
+
+    if "lora_up.weight" in state_dict:
+        # LoRA a.k.a LoCon
+        return LoRALayer(layer_key, state_dict)
+    elif "hada_w1_a" in state_dict:
+        return LoHALayer(layer_key, state_dict)
+    elif "lokr_w1" in state_dict or "lokr_w1_a" in state_dict:
+        return LoKRLayer(layer_key, state_dict)
+    elif "diff" in state_dict:
+        # Full a.k.a Diff
+        return FullLayer(layer_key, state_dict)
+    elif "on_input" in state_dict:
+        return IA3Layer(layer_key, state_dict)
+    elif "w_norm" in state_dict:
+        return NormLayer(layer_key, state_dict)
+    else:
+        raise ValueError(f"Unsupported lora format: {state_dict.keys()}")

invokeai/backend/peft/lora.py

@@ -0,0 +1,22 @@
+# Copyright (c) 2024 The InvokeAI Development team
+from typing import Dict, Optional
+
+import torch
+
+from invokeai.backend.peft.layers.any_lora_layer import AnyLoRALayer
+from invokeai.backend.raw_model import RawModel
+
+
+class LoRAModelRaw(RawModel):  # (torch.nn.Module):
+    def __init__(self, layers: Dict[str, AnyLoRALayer]):
+        self.layers = layers
+
+    def to(self, device: Optional[torch.device] = None, dtype: Optional[torch.dtype] = None) -> None:
+        for _key, layer in self.layers.items():
+            layer.to(device=device, dtype=dtype)
+
+    def calc_size(self) -> int:
+        model_size = 0
+        for _, layer in self.layers.items():
+            model_size += layer.calc_size()
+        return model_size

invokeai/backend/peft/peft_patcher.py

@@ -0,0 +1,102 @@
+from contextlib import contextmanager
+from typing import Dict, Iterator, Optional, Tuple
+
+import torch
+
+from invokeai.backend.peft.lora import LoRAModelRaw
+from invokeai.backend.util.devices import TorchDevice
+from invokeai.backend.util.original_weights_storage import OriginalWeightsStorage
+
+
+class PeftPatcher:
+    @classmethod
+    @torch.no_grad()
+    @contextmanager
+    def apply_peft_patches(
+        cls,
+        model: torch.nn.Module,
+        patches: Iterator[Tuple[LoRAModelRaw, float]],
+        prefix: str,
+        cached_weights: Optional[Dict[str, torch.Tensor]] = None,
+    ):
+        """Apply one or more PEFT patches to a model.
+
+        :param model: The model to patch.
+        :param loras: An iterator that returns tuples of PEFT patches and associated weights. An iterator is used so
+            that the PEFT patches do not need to be loaded into memory all at once.
+        :param prefix: The keys in the patches will be filtered to only include weights with this prefix.
+        :cached_weights: Read-only copy of the model's state dict in CPU, for efficient unpatching purposes.
+        """
+        original_weights = OriginalWeightsStorage(cached_weights)
+        try:
+            for patch, patch_weight in patches:
+                cls._apply_peft_patch(
+                    model=model,
+                    prefix=prefix,
+                    patch=patch,
+                    patch_weight=patch_weight,
+                    original_weights=original_weights,
+                )
+
+            yield
+        finally:
+            for param_key, weight in original_weights.get_changed_weights():
+                model.get_parameter(param_key).copy_(weight)
+
+    @classmethod
+    @torch.no_grad()
+    def _apply_peft_patch(
+        cls,
+        model: torch.nn.Module,
+        prefix: str,
+        patch: LoRAModelRaw,
+        patch_weight: float,
+        original_weights: OriginalWeightsStorage,
+    ):
+        """
+        Apply one a LoRA to a model.
+        :param model: The model to patch.
+        :param patch: LoRA model to patch in.
+        :param patch_weight: LoRA patch weight.
+        :param prefix: A string prefix that precedes keys used in the LoRAs weight layers.
+        :param original_weights: Storage with original weights, filled by weights which lora patches, used for unpatching.
+        """
+
+        if patch_weight == 0:
+            return
+
+        for layer_key, layer in patch.layers.items():
+            if not layer_key.startswith(prefix):
+                continue
+
+            module = model.get_submodule(layer_key)
+
+            # All of the LoRA weight calculations will be done on the same device as the module weight.
+            # (Performance will be best if this is a CUDA device.)
+            device = module.weight.device
+            dtype = module.weight.dtype
+
+            layer_scale = layer.alpha / layer.rank if (layer.alpha and layer.rank) else 1.0
+
+            # We intentionally move to the target device first, then cast. Experimentally, this was found to
+            # be significantly faster for 16-bit CPU tensors being moved to a CUDA device than doing the
+            # same thing in a single call to '.to(...)'.
+            layer.to(device=device)
+            layer.to(dtype=torch.float32)
+
+            # TODO(ryand): Using torch.autocast(...) over explicit casting may offer a speed benefit on CUDA
+            # devices here. Experimentally, it was found to be very slow on CPU. More investigation needed.
+            for param_name, lora_param_weight in layer.get_parameters(module).items():
+                param_key = layer_key + "." + param_name
+                module_param = module.get_parameter(param_name)
+
+                # Save original weight
+                original_weights.save(param_key, module_param)
+
+                if module_param.shape != lora_param_weight.shape:
+                    lora_param_weight = lora_param_weight.reshape(module_param.shape)
+
+                lora_param_weight *= patch_weight * layer_scale
+                module_param += lora_param_weight.to(dtype=dtype)
+
+            layer.to(device=TorchDevice.CPU_DEVICE)

invokeai/backend/stable_diffusion/extensions/lora.py

@@ -12,7 +12,7 @@ from invokeai.backend.util.devices import TorchDevice
 if TYPE_CHECKING:
     from invokeai.app.invocations.model import ModelIdentifierField
     from invokeai.app.services.shared.invocation_context import InvocationContext
-    from invokeai.backend.lora import LoRAModelRaw
+    from invokeai.backend.peft.lora import LoRAModelRaw
     from invokeai.backend.util.original_weights_storage import OriginalWeightsStorage
 
 

tests/backend/model_manager/test_lora.py

@@ -1,12 +1,9 @@
-# test that if the model's device changes while the lora is applied, the weights can still be restored
-
-# test that LoRA patching works on both CPU and CUDA
-
 import pytest
 import torch
 
-from invokeai.backend.lora import LoRALayer, LoRAModelRaw
 from invokeai.backend.model_patcher import ModelPatcher
+from invokeai.backend.peft.layers.lora_layer import LoRALayer
+from invokeai.backend.peft.lora import LoRAModelRaw
 
 
 @pytest.mark.parametrize(
@@ -38,7 +35,7 @@ def test_apply_lora(device):
             },
         )
     }
-    lora = LoRAModelRaw("lora_name", lora_layers)
+    lora = LoRAModelRaw(lora_layers)
 
     lora_weight = 0.5
     orig_linear_weight = model["linear_layer_1"].weight.data.detach().clone()
@@ -82,7 +79,7 @@ def test_apply_lora_change_device():
             },
         )
     }
-    lora = LoRAModelRaw("lora_name", lora_layers)
+    lora = LoRAModelRaw(lora_layers)
 
     orig_linear_weight = model["linear_layer_1"].weight.data.detach().clone()
 

tests/backend/peft/conversions/lora_state_dicts/flux_lora_diffusers_format.py

@@ -0,0 +1,990 @@
+state_dict_keys = [
+    "transformer.single_transformer_blocks.0.attn.to_k.lora_A.weight",
+    "transformer.single_transformer_blocks.0.attn.to_k.lora_B.weight",
+    "transformer.single_transformer_blocks.0.attn.to_q.lora_A.weight",
+    "transformer.single_transformer_blocks.0.attn.to_q.lora_B.weight",
+    "transformer.single_transformer_blocks.0.attn.to_v.lora_A.weight",
+    "transformer.single_transformer_blocks.0.attn.to_v.lora_B.weight",
+    "transformer.single_transformer_blocks.0.norm.linear.lora_A.weight",
+    "transformer.single_transformer_blocks.0.norm.linear.lora_B.weight",
+    "transformer.single_transformer_blocks.0.proj_mlp.lora_A.weight",
+    "transformer.single_transformer_blocks.0.proj_mlp.lora_B.weight",
+    "transformer.single_transformer_blocks.0.proj_out.lora_A.weight",
+    "transformer.single_transformer_blocks.0.proj_out.lora_B.weight",
+    "transformer.single_transformer_blocks.1.attn.to_k.lora_A.weight",
+    "transformer.single_transformer_blocks.1.attn.to_k.lora_B.weight",
+    "transformer.single_transformer_blocks.1.attn.to_q.lora_A.weight",
+    "transformer.single_transformer_blocks.1.attn.to_q.lora_B.weight",
+    "transformer.single_transformer_blocks.1.attn.to_v.lora_A.weight",
+    "transformer.single_transformer_blocks.1.attn.to_v.lora_B.weight",
+    "transformer.single_transformer_blocks.1.norm.linear.lora_A.weight",
+    "transformer.single_transformer_blocks.1.norm.linear.lora_B.weight",
+    "transformer.single_transformer_blocks.1.proj_mlp.lora_A.weight",
+    "transformer.single_transformer_blocks.1.proj_mlp.lora_B.weight",
+    "transformer.single_transformer_blocks.1.proj_out.lora_A.weight",
+    "transformer.single_transformer_blocks.1.proj_out.lora_B.weight",
+    "transformer.single_transformer_blocks.10.attn.to_k.lora_A.weight",
+    "transformer.single_transformer_blocks.10.attn.to_k.lora_B.weight",
+    "transformer.single_transformer_blocks.10.attn.to_q.lora_A.weight",
+    "transformer.single_transformer_blocks.10.attn.to_q.lora_B.weight",
+    "transformer.single_transformer_blocks.10.attn.to_v.lora_A.weight",
+    "transformer.single_transformer_blocks.10.attn.to_v.lora_B.weight",
+    "transformer.single_transformer_blocks.10.norm.linear.lora_A.weight",
+    "transformer.single_transformer_blocks.10.norm.linear.lora_B.weight",
+    "transformer.single_transformer_blocks.10.proj_mlp.lora_A.weight",
+    "transformer.single_transformer_blocks.10.proj_mlp.lora_B.weight",
+    "transformer.single_transformer_blocks.10.proj_out.lora_A.weight",
+    "transformer.single_transformer_blocks.10.proj_out.lora_B.weight",
+    "transformer.single_transformer_blocks.11.attn.to_k.lora_A.weight",
+    "transformer.single_transformer_blocks.11.attn.to_k.lora_B.weight",
+    "transformer.single_transformer_blocks.11.attn.to_q.lora_A.weight",
+    "transformer.single_transformer_blocks.11.attn.to_q.lora_B.weight",
+    "transformer.single_transformer_blocks.11.attn.to_v.lora_A.weight",
+    "transformer.single_transformer_blocks.11.attn.to_v.lora_B.weight",
+    "transformer.single_transformer_blocks.11.norm.linear.lora_A.weight",
+    "transformer.single_transformer_blocks.11.norm.linear.lora_B.weight",
+    "transformer.single_transformer_blocks.11.proj_mlp.lora_A.weight",
+    "transformer.single_transformer_blocks.11.proj_mlp.lora_B.weight",
+    "transformer.single_transformer_blocks.11.proj_out.lora_A.weight",
+    "transformer.single_transformer_blocks.11.proj_out.lora_B.weight",
+    "transformer.single_transformer_blocks.12.attn.to_k.lora_A.weight",
+    "transformer.single_transformer_blocks.12.attn.to_k.lora_B.weight",
+    "transformer.single_transformer_blocks.12.attn.to_q.lora_A.weight",
+    "transformer.single_transformer_blocks.12.attn.to_q.lora_B.weight",
+    "transformer.single_transformer_blocks.12.attn.to_v.lora_A.weight",
+    "transformer.single_transformer_blocks.12.attn.to_v.lora_B.weight",
+    "transformer.single_transformer_blocks.12.norm.linear.lora_A.weight",
+    "transformer.single_transformer_blocks.12.norm.linear.lora_B.weight",
+    "transformer.single_transformer_blocks.12.proj_mlp.lora_A.weight",
+    "transformer.single_transformer_blocks.12.proj_mlp.lora_B.weight",
+    "transformer.single_transformer_blocks.12.proj_out.lora_A.weight",
+    "transformer.single_transformer_blocks.12.proj_out.lora_B.weight",
+    "transformer.single_transformer_blocks.13.attn.to_k.lora_A.weight",
+    "transformer.single_transformer_blocks.13.attn.to_k.lora_B.weight",
+    "transformer.single_transformer_blocks.13.attn.to_q.lora_A.weight",
+    "transformer.single_transformer_blocks.13.attn.to_q.lora_B.weight",
+    "transformer.single_transformer_blocks.13.attn.to_v.lora_A.weight",
+    "transformer.single_transformer_blocks.13.attn.to_v.lora_B.weight",
+    "transformer.single_transformer_blocks.13.norm.linear.lora_A.weight",
+    "transformer.single_transformer_blocks.13.norm.linear.lora_B.weight",
+    "transformer.single_transformer_blocks.13.proj_mlp.lora_A.weight",
+    "transformer.single_transformer_blocks.13.proj_mlp.lora_B.weight",
+    "transformer.single_transformer_blocks.13.proj_out.lora_A.weight",
+    "transformer.single_transformer_blocks.13.proj_out.lora_B.weight",
+    "transformer.single_transformer_blocks.14.attn.to_k.lora_A.weight",
+    "transformer.single_transformer_blocks.14.attn.to_k.lora_B.weight",
+    "transformer.single_transformer_blocks.14.attn.to_q.lora_A.weight",
+    "transformer.single_transformer_blocks.14.attn.to_q.lora_B.weight",
+    "transformer.single_transformer_blocks.14.attn.to_v.lora_A.weight",
+    "transformer.single_transformer_blocks.14.attn.to_v.lora_B.weight",
+    "transformer.single_transformer_blocks.14.norm.linear.lora_A.weight",
+    "transformer.single_transformer_blocks.14.norm.linear.lora_B.weight",
+    "transformer.single_transformer_blocks.14.proj_mlp.lora_A.weight",
+    "transformer.single_transformer_blocks.14.proj_mlp.lora_B.weight",
+    "transformer.single_transformer_blocks.14.proj_out.lora_A.weight",
+    "transformer.single_transformer_blocks.14.proj_out.lora_B.weight",
+    "transformer.single_transformer_blocks.15.attn.to_k.lora_A.weight",
+    "transformer.single_transformer_blocks.15.attn.to_k.lora_B.weight",
+    "transformer.single_transformer_blocks.15.attn.to_q.lora_A.weight",
+    "transformer.single_transformer_blocks.15.attn.to_q.lora_B.weight",
+    "transformer.single_transformer_blocks.15.attn.to_v.lora_A.weight",
+    "transformer.single_transformer_blocks.15.attn.to_v.lora_B.weight",
+    "transformer.single_transformer_blocks.15.norm.linear.lora_A.weight",
+    "transformer.single_transformer_blocks.15.norm.linear.lora_B.weight",
+    "transformer.single_transformer_blocks.15.proj_mlp.lora_A.weight",
+    "transformer.single_transformer_blocks.15.proj_mlp.lora_B.weight",
+    "transformer.single_transformer_blocks.15.proj_out.lora_A.weight",
+    "transformer.single_transformer_blocks.15.proj_out.lora_B.weight",
+    "transformer.single_transformer_blocks.16.attn.to_k.lora_A.weight",
+    "transformer.single_transformer_blocks.16.attn.to_k.lora_B.weight",
+    "transformer.single_transformer_blocks.16.attn.to_q.lora_A.weight",
+    "transformer.single_transformer_blocks.16.attn.to_q.lora_B.weight",
+    "transformer.single_transformer_blocks.16.attn.to_v.lora_A.weight",
+    "transformer.single_transformer_blocks.16.attn.to_v.lora_B.weight",
+    "transformer.single_transformer_blocks.16.norm.linear.lora_A.weight",
+    "transformer.single_transformer_blocks.16.norm.linear.lora_B.weight",
+    "transformer.single_transformer_blocks.16.proj_mlp.lora_A.weight",
+    "transformer.single_transformer_blocks.16.proj_mlp.lora_B.weight",
+    "transformer.single_transformer_blocks.16.proj_out.lora_A.weight",
+    "transformer.single_transformer_blocks.16.proj_out.lora_B.weight",
+    "transformer.single_transformer_blocks.17.attn.to_k.lora_A.weight",
+    "transformer.single_transformer_blocks.17.attn.to_k.lora_B.weight",
+    "transformer.single_transformer_blocks.17.attn.to_q.lora_A.weight",
+    "transformer.single_transformer_blocks.17.attn.to_q.lora_B.weight",
+    "transformer.single_transformer_blocks.17.attn.to_v.lora_A.weight",
+    "transformer.single_transformer_blocks.17.attn.to_v.lora_B.weight",
+    "transformer.single_transformer_blocks.17.norm.linear.lora_A.weight",
+    "transformer.single_transformer_blocks.17.norm.linear.lora_B.weight",
+    "transformer.single_transformer_blocks.17.proj_mlp.lora_A.weight",
+    "transformer.single_transformer_blocks.17.proj_mlp.lora_B.weight",
+    "transformer.single_transformer_blocks.17.proj_out.lora_A.weight",
+    "transformer.single_transformer_blocks.17.proj_out.lora_B.weight",
+    "transformer.single_transformer_blocks.18.attn.to_k.lora_A.weight",
+    "transformer.single_transformer_blocks.18.attn.to_k.lora_B.weight",
+    "transformer.single_transformer_blocks.18.attn.to_q.lora_A.weight",
+    "transformer.single_transformer_blocks.18.attn.to_q.lora_B.weight",
+    "transformer.single_transformer_blocks.18.attn.to_v.lora_A.weight",
+    "transformer.single_transformer_blocks.18.attn.to_v.lora_B.weight",
+    "transformer.single_transformer_blocks.18.norm.linear.lora_A.weight",
+    "transformer.single_transformer_blocks.18.norm.linear.lora_B.weight",
+    "transformer.single_transformer_blocks.18.proj_mlp.lora_A.weight",
+    "transformer.single_transformer_blocks.18.proj_mlp.lora_B.weight",
+    "transformer.single_transformer_blocks.18.proj_out.lora_A.weight",
+    "transformer.single_transformer_blocks.18.proj_out.lora_B.weight",
+    "transformer.single_transformer_blocks.19.attn.to_k.lora_A.weight",
+    "transformer.single_transformer_blocks.19.attn.to_k.lora_B.weight",
+    "transformer.single_transformer_blocks.19.attn.to_q.lora_A.weight",
+    "transformer.single_transformer_blocks.19.attn.to_q.lora_B.weight",
+    "transformer.single_transformer_blocks.19.attn.to_v.lora_A.weight",
+    "transformer.single_transformer_blocks.19.attn.to_v.lora_B.weight",
+    "transformer.single_transformer_blocks.19.norm.linear.lora_A.weight",
+    "transformer.single_transformer_blocks.19.norm.linear.lora_B.weight",
+    "transformer.single_transformer_blocks.19.proj_mlp.lora_A.weight",
+    "transformer.single_transformer_blocks.19.proj_mlp.lora_B.weight",
+    "transformer.single_transformer_blocks.19.proj_out.lora_A.weight",
+    "transformer.single_transformer_blocks.19.proj_out.lora_B.weight",
+    "transformer.single_transformer_blocks.2.attn.to_k.lora_A.weight",
+    "transformer.single_transformer_blocks.2.attn.to_k.lora_B.weight",
+    "transformer.single_transformer_blocks.2.attn.to_q.lora_A.weight",
+    "transformer.single_transformer_blocks.2.attn.to_q.lora_B.weight",
+    "transformer.single_transformer_blocks.2.attn.to_v.lora_A.weight",
+    "transformer.single_transformer_blocks.2.attn.to_v.lora_B.weight",
+    "transformer.single_transformer_blocks.2.norm.linear.lora_A.weight",
+    "transformer.single_transformer_blocks.2.norm.linear.lora_B.weight",
+    "transformer.single_transformer_blocks.2.proj_mlp.lora_A.weight",
+    "transformer.single_transformer_blocks.2.proj_mlp.lora_B.weight",
+    "transformer.single_transformer_blocks.2.proj_out.lora_A.weight",
+    "transformer.single_transformer_blocks.2.proj_out.lora_B.weight",
+    "transformer.single_transformer_blocks.20.attn.to_k.lora_A.weight",
+    "transformer.single_transformer_blocks.20.attn.to_k.lora_B.weight",
+    "transformer.single_transformer_blocks.20.attn.to_q.lora_A.weight",
+    "transformer.single_transformer_blocks.20.attn.to_q.lora_B.weight",
+    "transformer.single_transformer_blocks.20.attn.to_v.lora_A.weight",
+    "transformer.single_transformer_blocks.20.attn.to_v.lora_B.weight",
+    "transformer.single_transformer_blocks.20.norm.linear.lora_A.weight",
+    "transformer.single_transformer_blocks.20.norm.linear.lora_B.weight",
+    "transformer.single_transformer_blocks.20.proj_mlp.lora_A.weight",
+    "transformer.single_transformer_blocks.20.proj_mlp.lora_B.weight",
+    "transformer.single_transformer_blocks.20.proj_out.lora_A.weight",
+    "transformer.single_transformer_blocks.20.proj_out.lora_B.weight",
+    "transformer.single_transformer_blocks.21.attn.to_k.lora_A.weight",
+    "transformer.single_transformer_blocks.21.attn.to_k.lora_B.weight",
+    "transformer.single_transformer_blocks.21.attn.to_q.lora_A.weight",
+    "transformer.single_transformer_blocks.21.attn.to_q.lora_B.weight",
+    "transformer.single_transformer_blocks.21.attn.to_v.lora_A.weight",
+    "transformer.single_transformer_blocks.21.attn.to_v.lora_B.weight",
+    "transformer.single_transformer_blocks.21.norm.linear.lora_A.weight",
+    "transformer.single_transformer_blocks.21.norm.linear.lora_B.weight",
+    "transformer.single_transformer_blocks.21.proj_mlp.lora_A.weight",
+    "transformer.single_transformer_blocks.21.proj_mlp.lora_B.weight",
+    "transformer.single_transformer_blocks.21.proj_out.lora_A.weight",
+    "transformer.single_transformer_blocks.21.proj_out.lora_B.weight",
+    "transformer.single_transformer_blocks.22.attn.to_k.lora_A.weight",
+    "transformer.single_transformer_blocks.22.attn.to_k.lora_B.weight",
+    "transformer.single_transformer_blocks.22.attn.to_q.lora_A.weight",
+    "transformer.single_transformer_blocks.22.attn.to_q.lora_B.weight",
+    "transformer.single_transformer_blocks.22.attn.to_v.lora_A.weight",
+    "transformer.single_transformer_blocks.22.attn.to_v.lora_B.weight",
+    "transformer.single_transformer_blocks.22.norm.linear.lora_A.weight",
+    "transformer.single_transformer_blocks.22.norm.linear.lora_B.weight",
+    "transformer.single_transformer_blocks.22.proj_mlp.lora_A.weight",
+    "transformer.single_transformer_blocks.22.proj_mlp.lora_B.weight",
+    "transformer.single_transformer_blocks.22.proj_out.lora_A.weight",
+    "transformer.single_transformer_blocks.22.proj_out.lora_B.weight",
+    "transformer.single_transformer_blocks.23.attn.to_k.lora_A.weight",
+    "transformer.single_transformer_blocks.23.attn.to_k.lora_B.weight",
+    "transformer.single_transformer_blocks.23.attn.to_q.lora_A.weight",
+    "transformer.single_transformer_blocks.23.attn.to_q.lora_B.weight",
+    "transformer.single_transformer_blocks.23.attn.to_v.lora_A.weight",
+    "transformer.single_transformer_blocks.23.attn.to_v.lora_B.weight",
+    "transformer.single_transformer_blocks.23.norm.linear.lora_A.weight",
+    "transformer.single_transformer_blocks.23.norm.linear.lora_B.weight",
+    "transformer.single_transformer_blocks.23.proj_mlp.lora_A.weight",
+    "transformer.single_transformer_blocks.23.proj_mlp.lora_B.weight",
+    "transformer.single_transformer_blocks.23.proj_out.lora_A.weight",
+    "transformer.single_transformer_blocks.23.proj_out.lora_B.weight",
+    "transformer.single_transformer_blocks.24.attn.to_k.lora_A.weight",
+    "transformer.single_transformer_blocks.24.attn.to_k.lora_B.weight",
+    "transformer.single_transformer_blocks.24.attn.to_q.lora_A.weight",
+    "transformer.single_transformer_blocks.24.attn.to_q.lora_B.weight",
+    "transformer.single_transformer_blocks.24.attn.to_v.lora_A.weight",
+    "transformer.single_transformer_blocks.24.attn.to_v.lora_B.weight",
+    "transformer.single_transformer_blocks.24.norm.linear.lora_A.weight",
+    "transformer.single_transformer_blocks.24.norm.linear.lora_B.weight",
+    "transformer.single_transformer_blocks.24.proj_mlp.lora_A.weight",
+    "transformer.single_transformer_blocks.24.proj_mlp.lora_B.weight",
+    "transformer.single_transformer_blocks.24.proj_out.lora_A.weight",
+    "transformer.single_transformer_blocks.24.proj_out.lora_B.weight",
+    "transformer.single_transformer_blocks.25.attn.to_k.lora_A.weight",
+    "transformer.single_transformer_blocks.25.attn.to_k.lora_B.weight",
+    "transformer.single_transformer_blocks.25.attn.to_q.lora_A.weight",
+    "transformer.single_transformer_blocks.25.attn.to_q.lora_B.weight",
+    "transformer.single_transformer_blocks.25.attn.to_v.lora_A.weight",
+    "transformer.single_transformer_blocks.25.attn.to_v.lora_B.weight",
+    "transformer.single_transformer_blocks.25.norm.linear.lora_A.weight",
+    "transformer.single_transformer_blocks.25.norm.linear.lora_B.weight",
+    "transformer.single_transformer_blocks.25.proj_mlp.lora_A.weight",
+    "transformer.single_transformer_blocks.25.proj_mlp.lora_B.weight",
+    "transformer.single_transformer_blocks.25.proj_out.lora_A.weight",
+    "transformer.single_transformer_blocks.25.proj_out.lora_B.weight",
+    "transformer.single_transformer_blocks.26.attn.to_k.lora_A.weight",
+    "transformer.single_transformer_blocks.26.attn.to_k.lora_B.weight",
+    "transformer.single_transformer_blocks.26.attn.to_q.lora_A.weight",
+    "transformer.single_transformer_blocks.26.attn.to_q.lora_B.weight",
+    "transformer.single_transformer_blocks.26.attn.to_v.lora_A.weight",
+    "transformer.single_transformer_blocks.26.attn.to_v.lora_B.weight",
+    "transformer.single_transformer_blocks.26.norm.linear.lora_A.weight",
+    "transformer.single_transformer_blocks.26.norm.linear.lora_B.weight",
+    "transformer.single_transformer_blocks.26.proj_mlp.lora_A.weight",
+    "transformer.single_transformer_blocks.26.proj_mlp.lora_B.weight",
+    "transformer.single_transformer_blocks.26.proj_out.lora_A.weight",
+    "transformer.single_transformer_blocks.26.proj_out.lora_B.weight",
+    "transformer.single_transformer_blocks.27.attn.to_k.lora_A.weight",
+    "transformer.single_transformer_blocks.27.attn.to_k.lora_B.weight",
+    "transformer.single_transformer_blocks.27.attn.to_q.lora_A.weight",
+    "transformer.single_transformer_blocks.27.attn.to_q.lora_B.weight",
+    "transformer.single_transformer_blocks.27.attn.to_v.lora_A.weight",
+    "transformer.single_transformer_blocks.27.attn.to_v.lora_B.weight",
+    "transformer.single_transformer_blocks.27.norm.linear.lora_A.weight",
+    "transformer.single_transformer_blocks.27.norm.linear.lora_B.weight",
+    "transformer.single_transformer_blocks.27.proj_mlp.lora_A.weight",
+    "transformer.single_transformer_blocks.27.proj_mlp.lora_B.weight",
+    "transformer.single_transformer_blocks.27.proj_out.lora_A.weight",
+    "transformer.single_transformer_blocks.27.proj_out.lora_B.weight",
+    "transformer.single_transformer_blocks.28.attn.to_k.lora_A.weight",
+    "transformer.single_transformer_blocks.28.attn.to_k.lora_B.weight",
+    "transformer.single_transformer_blocks.28.attn.to_q.lora_A.weight",
+    "transformer.single_transformer_blocks.28.attn.to_q.lora_B.weight",
+    "transformer.single_transformer_blocks.28.attn.to_v.lora_A.weight",
+    "transformer.single_transformer_blocks.28.attn.to_v.lora_B.weight",
+    "transformer.single_transformer_blocks.28.norm.linear.lora_A.weight",
+    "transformer.single_transformer_blocks.28.norm.linear.lora_B.weight",
+    "transformer.single_transformer_blocks.28.proj_mlp.lora_A.weight",
+    "transformer.single_transformer_blocks.28.proj_mlp.lora_B.weight",
+    "transformer.single_transformer_blocks.28.proj_out.lora_A.weight",
+    "transformer.single_transformer_blocks.28.proj_out.lora_B.weight",
+    "transformer.single_transformer_blocks.29.attn.to_k.lora_A.weight",
+    "transformer.single_transformer_blocks.29.attn.to_k.lora_B.weight",
+    "transformer.single_transformer_blocks.29.attn.to_q.lora_A.weight",
+    "transformer.single_transformer_blocks.29.attn.to_q.lora_B.weight",
+    "transformer.single_transformer_blocks.29.attn.to_v.lora_A.weight",
+    "transformer.single_transformer_blocks.29.attn.to_v.lora_B.weight",
+    "transformer.single_transformer_blocks.29.norm.linear.lora_A.weight",
+    "transformer.single_transformer_blocks.29.norm.linear.lora_B.weight",
+    "transformer.single_transformer_blocks.29.proj_mlp.lora_A.weight",
+    "transformer.single_transformer_blocks.29.proj_mlp.lora_B.weight",
+    "transformer.single_transformer_blocks.29.proj_out.lora_A.weight",
+    "transformer.single_transformer_blocks.29.proj_out.lora_B.weight",
+    "transformer.single_transformer_blocks.3.attn.to_k.lora_A.weight",
+    "transformer.single_transformer_blocks.3.attn.to_k.lora_B.weight",
+    "transformer.single_transformer_blocks.3.attn.to_q.lora_A.weight",
+    "transformer.single_transformer_blocks.3.attn.to_q.lora_B.weight",
+    "transformer.single_transformer_blocks.3.attn.to_v.lora_A.weight",
+    "transformer.single_transformer_blocks.3.attn.to_v.lora_B.weight",
+    "transformer.single_transformer_blocks.3.norm.linear.lora_A.weight",
+    "transformer.single_transformer_blocks.3.norm.linear.lora_B.weight",
+    "transformer.single_transformer_blocks.3.proj_mlp.lora_A.weight",
+    "transformer.single_transformer_blocks.3.proj_mlp.lora_B.weight",
+    "transformer.single_transformer_blocks.3.proj_out.lora_A.weight",
+    "transformer.single_transformer_blocks.3.proj_out.lora_B.weight",
+    "transformer.single_transformer_blocks.30.attn.to_k.lora_A.weight",
+    "transformer.single_transformer_blocks.30.attn.to_k.lora_B.weight",
+    "transformer.single_transformer_blocks.30.attn.to_q.lora_A.weight",
+    "transformer.single_transformer_blocks.30.attn.to_q.lora_B.weight",
+    "transformer.single_transformer_blocks.30.attn.to_v.lora_A.weight",
+    "transformer.single_transformer_blocks.30.attn.to_v.lora_B.weight",
+    "transformer.single_transformer_blocks.30.norm.linear.lora_A.weight",
+    "transformer.single_transformer_blocks.30.norm.linear.lora_B.weight",
+    "transformer.single_transformer_blocks.30.proj_mlp.lora_A.weight",
+    "transformer.single_transformer_blocks.30.proj_mlp.lora_B.weight",
+    "transformer.single_transformer_blocks.30.proj_out.lora_A.weight",
+    "transformer.single_transformer_blocks.30.proj_out.lora_B.weight",
+    "transformer.single_transformer_blocks.31.attn.to_k.lora_A.weight",
+    "transformer.single_transformer_blocks.31.attn.to_k.lora_B.weight",
+    "transformer.single_transformer_blocks.31.attn.to_q.lora_A.weight",
+    "transformer.single_transformer_blocks.31.attn.to_q.lora_B.weight",
+    "transformer.single_transformer_blocks.31.attn.to_v.lora_A.weight",
+    "transformer.single_transformer_blocks.31.attn.to_v.lora_B.weight",
+    "transformer.single_transformer_blocks.31.norm.linear.lora_A.weight",
+    "transformer.single_transformer_blocks.31.norm.linear.lora_B.weight",
+    "transformer.single_transformer_blocks.31.proj_mlp.lora_A.weight",
+    "transformer.single_transformer_blocks.31.proj_mlp.lora_B.weight",
+    "transformer.single_transformer_blocks.31.proj_out.lora_A.weight",
+    "transformer.single_transformer_blocks.31.proj_out.lora_B.weight",
+    "transformer.single_transformer_blocks.32.attn.to_k.lora_A.weight",
+    "transformer.single_transformer_blocks.32.attn.to_k.lora_B.weight",
+    "transformer.single_transformer_blocks.32.attn.to_q.lora_A.weight",
+    "transformer.single_transformer_blocks.32.attn.to_q.lora_B.weight",
+    "transformer.single_transformer_blocks.32.attn.to_v.lora_A.weight",
+    "transformer.single_transformer_blocks.32.attn.to_v.lora_B.weight",
+    "transformer.single_transformer_blocks.32.norm.linear.lora_A.weight",
+    "transformer.single_transformer_blocks.32.norm.linear.lora_B.weight",
+    "transformer.single_transformer_blocks.32.proj_mlp.lora_A.weight",
+    "transformer.single_transformer_blocks.32.proj_mlp.lora_B.weight",
+    "transformer.single_transformer_blocks.32.proj_out.lora_A.weight",
+    "transformer.single_transformer_blocks.32.proj_out.lora_B.weight",
+    "transformer.single_transformer_blocks.33.attn.to_k.lora_A.weight",
+    "transformer.single_transformer_blocks.33.attn.to_k.lora_B.weight",
+    "transformer.single_transformer_blocks.33.attn.to_q.lora_A.weight",
+    "transformer.single_transformer_blocks.33.attn.to_q.lora_B.weight",
+    "transformer.single_transformer_blocks.33.attn.to_v.lora_A.weight",
+    "transformer.single_transformer_blocks.33.attn.to_v.lora_B.weight",
+    "transformer.single_transformer_blocks.33.norm.linear.lora_A.weight",
+    "transformer.single_transformer_blocks.33.norm.linear.lora_B.weight",
+    "transformer.single_transformer_blocks.33.proj_mlp.lora_A.weight",
+    "transformer.single_transformer_blocks.33.proj_mlp.lora_B.weight",
+    "transformer.single_transformer_blocks.33.proj_out.lora_A.weight",
+    "transformer.single_transformer_blocks.33.proj_out.lora_B.weight",
+    "transformer.single_transformer_blocks.34.attn.to_k.lora_A.weight",
+    "transformer.single_transformer_blocks.34.attn.to_k.lora_B.weight",
+    "transformer.single_transformer_blocks.34.attn.to_q.lora_A.weight",
+    "transformer.single_transformer_blocks.34.attn.to_q.lora_B.weight",
+    "transformer.single_transformer_blocks.34.attn.to_v.lora_A.weight",
+    "transformer.single_transformer_blocks.34.attn.to_v.lora_B.weight",
+    "transformer.single_transformer_blocks.34.norm.linear.lora_A.weight",
+    "transformer.single_transformer_blocks.34.norm.linear.lora_B.weight",
+    "transformer.single_transformer_blocks.34.proj_mlp.lora_A.weight",
+    "transformer.single_transformer_blocks.34.proj_mlp.lora_B.weight",
+    "transformer.single_transformer_blocks.34.proj_out.lora_A.weight",
+    "transformer.single_transformer_blocks.34.proj_out.lora_B.weight",
+    "transformer.single_transformer_blocks.35.attn.to_k.lora_A.weight",
+    "transformer.single_transformer_blocks.35.attn.to_k.lora_B.weight",
+    "transformer.single_transformer_blocks.35.attn.to_q.lora_A.weight",
+    "transformer.single_transformer_blocks.35.attn.to_q.lora_B.weight",
+    "transformer.single_transformer_blocks.35.attn.to_v.lora_A.weight",
+    "transformer.single_transformer_blocks.35.attn.to_v.lora_B.weight",
+    "transformer.single_transformer_blocks.35.norm.linear.lora_A.weight",
+    "transformer.single_transformer_blocks.35.norm.linear.lora_B.weight",
+    "transformer.single_transformer_blocks.35.proj_mlp.lora_A.weight",
+    "transformer.single_transformer_blocks.35.proj_mlp.lora_B.weight",
+    "transformer.single_transformer_blocks.35.proj_out.lora_A.weight",
+    "transformer.single_transformer_blocks.35.proj_out.lora_B.weight",
+    "transformer.single_transformer_blocks.36.attn.to_k.lora_A.weight",
+    "transformer.single_transformer_blocks.36.attn.to_k.lora_B.weight",
+    "transformer.single_transformer_blocks.36.attn.to_q.lora_A.weight",
+    "transformer.single_transformer_blocks.36.attn.to_q.lora_B.weight",
+    "transformer.single_transformer_blocks.36.attn.to_v.lora_A.weight",
+    "transformer.single_transformer_blocks.36.attn.to_v.lora_B.weight",
+    "transformer.single_transformer_blocks.36.norm.linear.lora_A.weight",
+    "transformer.single_transformer_blocks.36.norm.linear.lora_B.weight",
+    "transformer.single_transformer_blocks.36.proj_mlp.lora_A.weight",
+    "transformer.single_transformer_blocks.36.proj_mlp.lora_B.weight",
+    "transformer.single_transformer_blocks.36.proj_out.lora_A.weight",
+    "transformer.single_transformer_blocks.36.proj_out.lora_B.weight",
+    "transformer.single_transformer_blocks.37.attn.to_k.lora_A.weight",
+    "transformer.single_transformer_blocks.37.attn.to_k.lora_B.weight",
+    "transformer.single_transformer_blocks.37.attn.to_q.lora_A.weight",
+    "transformer.single_transformer_blocks.37.attn.to_q.lora_B.weight",
+    "transformer.single_transformer_blocks.37.attn.to_v.lora_A.weight",
+    "transformer.single_transformer_blocks.37.attn.to_v.lora_B.weight",
+    "transformer.single_transformer_blocks.37.norm.linear.lora_A.weight",
+    "transformer.single_transformer_blocks.37.norm.linear.lora_B.weight",
+    "transformer.single_transformer_blocks.37.proj_mlp.lora_A.weight",
+    "transformer.single_transformer_blocks.37.proj_mlp.lora_B.weight",
+    "transformer.single_transformer_blocks.37.proj_out.lora_A.weight",
+    "transformer.single_transformer_blocks.37.proj_out.lora_B.weight",
+    "transformer.single_transformer_blocks.4.attn.to_k.lora_A.weight",
+    "transformer.single_transformer_blocks.4.attn.to_k.lora_B.weight",
+    "transformer.single_transformer_blocks.4.attn.to_q.lora_A.weight",
+    "transformer.single_transformer_blocks.4.attn.to_q.lora_B.weight",
+    "transformer.single_transformer_blocks.4.attn.to_v.lora_A.weight",
+    "transformer.single_transformer_blocks.4.attn.to_v.lora_B.weight",
+    "transformer.single_transformer_blocks.4.norm.linear.lora_A.weight",
+    "transformer.single_transformer_blocks.4.norm.linear.lora_B.weight",
+    "transformer.single_transformer_blocks.4.proj_mlp.lora_A.weight",
+    "transformer.single_transformer_blocks.4.proj_mlp.lora_B.weight",
+    "transformer.single_transformer_blocks.4.proj_out.lora_A.weight",
+    "transformer.single_transformer_blocks.4.proj_out.lora_B.weight",
+    "transformer.single_transformer_blocks.5.attn.to_k.lora_A.weight",
+    "transformer.single_transformer_blocks.5.attn.to_k.lora_B.weight",
+    "transformer.single_transformer_blocks.5.attn.to_q.lora_A.weight",
+    "transformer.single_transformer_blocks.5.attn.to_q.lora_B.weight",
+    "transformer.single_transformer_blocks.5.attn.to_v.lora_A.weight",
+    "transformer.single_transformer_blocks.5.attn.to_v.lora_B.weight",
+    "transformer.single_transformer_blocks.5.norm.linear.lora_A.weight",
+    "transformer.single_transformer_blocks.5.norm.linear.lora_B.weight",
+    "transformer.single_transformer_blocks.5.proj_mlp.lora_A.weight",
+    "transformer.single_transformer_blocks.5.proj_mlp.lora_B.weight",
+    "transformer.single_transformer_blocks.5.proj_out.lora_A.weight",
+    "transformer.single_transformer_blocks.5.proj_out.lora_B.weight",
+    "transformer.single_transformer_blocks.6.attn.to_k.lora_A.weight",
+    "transformer.single_transformer_blocks.6.attn.to_k.lora_B.weight",
+    "transformer.single_transformer_blocks.6.attn.to_q.lora_A.weight",
+    "transformer.single_transformer_blocks.6.attn.to_q.lora_B.weight",
+    "transformer.single_transformer_blocks.6.attn.to_v.lora_A.weight",
+    "transformer.single_transformer_blocks.6.attn.to_v.lora_B.weight",
+    "transformer.single_transformer_blocks.6.norm.linear.lora_A.weight",
+    "transformer.single_transformer_blocks.6.norm.linear.lora_B.weight",
+    "transformer.single_transformer_blocks.6.proj_mlp.lora_A.weight",
+    "transformer.single_transformer_blocks.6.proj_mlp.lora_B.weight",
+    "transformer.single_transformer_blocks.6.proj_out.lora_A.weight",
+    "transformer.single_transformer_blocks.6.proj_out.lora_B.weight",
+    "transformer.single_transformer_blocks.7.attn.to_k.lora_A.weight",
+    "transformer.single_transformer_blocks.7.attn.to_k.lora_B.weight",
+    "transformer.single_transformer_blocks.7.attn.to_q.lora_A.weight",
+    "transformer.single_transformer_blocks.7.attn.to_q.lora_B.weight",
+    "transformer.single_transformer_blocks.7.attn.to_v.lora_A.weight",
+    "transformer.single_transformer_blocks.7.attn.to_v.lora_B.weight",
+    "transformer.single_transformer_blocks.7.norm.linear.lora_A.weight",
+    "transformer.single_transformer_blocks.7.norm.linear.lora_B.weight",
+    "transformer.single_transformer_blocks.7.proj_mlp.lora_A.weight",
+    "transformer.single_transformer_blocks.7.proj_mlp.lora_B.weight",
+    "transformer.single_transformer_blocks.7.proj_out.lora_A.weight",
+    "transformer.single_transformer_blocks.7.proj_out.lora_B.weight",
+    "transformer.single_transformer_blocks.8.attn.to_k.lora_A.weight",
+    "transformer.single_transformer_blocks.8.attn.to_k.lora_B.weight",
+    "transformer.single_transformer_blocks.8.attn.to_q.lora_A.weight",
+    "transformer.single_transformer_blocks.8.attn.to_q.lora_B.weight",
+    "transformer.single_transformer_blocks.8.attn.to_v.lora_A.weight",
+    "transformer.single_transformer_blocks.8.attn.to_v.lora_B.weight",
+    "transformer.single_transformer_blocks.8.norm.linear.lora_A.weight",
+    "transformer.single_transformer_blocks.8.norm.linear.lora_B.weight",
+    "transformer.single_transformer_blocks.8.proj_mlp.lora_A.weight",
+    "transformer.single_transformer_blocks.8.proj_mlp.lora_B.weight",
+    "transformer.single_transformer_blocks.8.proj_out.lora_A.weight",
+    "transformer.single_transformer_blocks.8.proj_out.lora_B.weight",
+    "transformer.single_transformer_blocks.9.attn.to_k.lora_A.weight",
+    "transformer.single_transformer_blocks.9.attn.to_k.lora_B.weight",
+    "transformer.single_transformer_blocks.9.attn.to_q.lora_A.weight",
+    "transformer.single_transformer_blocks.9.attn.to_q.lora_B.weight",
+    "transformer.single_transformer_blocks.9.attn.to_v.lora_A.weight",
+    "transformer.single_transformer_blocks.9.attn.to_v.lora_B.weight",
+    "transformer.single_transformer_blocks.9.norm.linear.lora_A.weight",
+    "transformer.single_transformer_blocks.9.norm.linear.lora_B.weight",
+    "transformer.single_transformer_blocks.9.proj_mlp.lora_A.weight",
+    "transformer.single_transformer_blocks.9.proj_mlp.lora_B.weight",
+    "transformer.single_transformer_blocks.9.proj_out.lora_A.weight",
+    "transformer.single_transformer_blocks.9.proj_out.lora_B.weight",
+    "transformer.transformer_blocks.0.attn.add_k_proj.lora_A.weight",
+    "transformer.transformer_blocks.0.attn.add_k_proj.lora_B.weight",
+    "transformer.transformer_blocks.0.attn.add_q_proj.lora_A.weight",
+    "transformer.transformer_blocks.0.attn.add_q_proj.lora_B.weight",
+    "transformer.transformer_blocks.0.attn.add_v_proj.lora_A.weight",
+    "transformer.transformer_blocks.0.attn.add_v_proj.lora_B.weight",
+    "transformer.transformer_blocks.0.attn.to_add_out.lora_A.weight",
+    "transformer.transformer_blocks.0.attn.to_add_out.lora_B.weight",
+    "transformer.transformer_blocks.0.attn.to_k.lora_A.weight",
+    "transformer.transformer_blocks.0.attn.to_k.lora_B.weight",
+    "transformer.transformer_blocks.0.attn.to_out.0.lora_A.weight",
+    "transformer.transformer_blocks.0.attn.to_out.0.lora_B.weight",
+    "transformer.transformer_blocks.0.attn.to_q.lora_A.weight",
+    "transformer.transformer_blocks.0.attn.to_q.lora_B.weight",
+    "transformer.transformer_blocks.0.attn.to_v.lora_A.weight",
+    "transformer.transformer_blocks.0.attn.to_v.lora_B.weight",
+    "transformer.transformer_blocks.0.ff.net.0.proj.lora_A.weight",
+    "transformer.transformer_blocks.0.ff.net.0.proj.lora_B.weight",
+    "transformer.transformer_blocks.0.ff.net.2.lora_A.weight",
+    "transformer.transformer_blocks.0.ff.net.2.lora_B.weight",
+    "transformer.transformer_blocks.0.ff_context.net.0.proj.lora_A.weight",
+    "transformer.transformer_blocks.0.ff_context.net.0.proj.lora_B.weight",
+    "transformer.transformer_blocks.0.ff_context.net.2.lora_A.weight",
+    "transformer.transformer_blocks.0.ff_context.net.2.lora_B.weight",
+    "transformer.transformer_blocks.0.norm1.linear.lora_A.weight",
+    "transformer.transformer_blocks.0.norm1.linear.lora_B.weight",
+    "transformer.transformer_blocks.0.norm1_context.linear.lora_A.weight",
+    "transformer.transformer_blocks.0.norm1_context.linear.lora_B.weight",
+    "transformer.transformer_blocks.1.attn.add_k_proj.lora_A.weight",
+    "transformer.transformer_blocks.1.attn.add_k_proj.lora_B.weight",
+    "transformer.transformer_blocks.1.attn.add_q_proj.lora_A.weight",
+    "transformer.transformer_blocks.1.attn.add_q_proj.lora_B.weight",
+    "transformer.transformer_blocks.1.attn.add_v_proj.lora_A.weight",
+    "transformer.transformer_blocks.1.attn.add_v_proj.lora_B.weight",
+    "transformer.transformer_blocks.1.attn.to_add_out.lora_A.weight",
+    "transformer.transformer_blocks.1.attn.to_add_out.lora_B.weight",
+    "transformer.transformer_blocks.1.attn.to_k.lora_A.weight",
+    "transformer.transformer_blocks.1.attn.to_k.lora_B.weight",
+    "transformer.transformer_blocks.1.attn.to_out.0.lora_A.weight",
+    "transformer.transformer_blocks.1.attn.to_out.0.lora_B.weight",
+    "transformer.transformer_blocks.1.attn.to_q.lora_A.weight",
+    "transformer.transformer_blocks.1.attn.to_q.lora_B.weight",
+    "transformer.transformer_blocks.1.attn.to_v.lora_A.weight",
+    "transformer.transformer_blocks.1.attn.to_v.lora_B.weight",
+    "transformer.transformer_blocks.1.ff.net.0.proj.lora_A.weight",
+    "transformer.transformer_blocks.1.ff.net.0.proj.lora_B.weight",
+    "transformer.transformer_blocks.1.ff.net.2.lora_A.weight",
+    "transformer.transformer_blocks.1.ff.net.2.lora_B.weight",
+    "transformer.transformer_blocks.1.ff_context.net.0.proj.lora_A.weight",
+    "transformer.transformer_blocks.1.ff_context.net.0.proj.lora_B.weight",
+    "transformer.transformer_blocks.1.ff_context.net.2.lora_A.weight",
+    "transformer.transformer_blocks.1.ff_context.net.2.lora_B.weight",
+    "transformer.transformer_blocks.1.norm1.linear.lora_A.weight",
+    "transformer.transformer_blocks.1.norm1.linear.lora_B.weight",
+    "transformer.transformer_blocks.1.norm1_context.linear.lora_A.weight",
+    "transformer.transformer_blocks.1.norm1_context.linear.lora_B.weight",
+    "transformer.transformer_blocks.10.attn.add_k_proj.lora_A.weight",
+    "transformer.transformer_blocks.10.attn.add_k_proj.lora_B.weight",
+    "transformer.transformer_blocks.10.attn.add_q_proj.lora_A.weight",
+    "transformer.transformer_blocks.10.attn.add_q_proj.lora_B.weight",
+    "transformer.transformer_blocks.10.attn.add_v_proj.lora_A.weight",
+    "transformer.transformer_blocks.10.attn.add_v_proj.lora_B.weight",
+    "transformer.transformer_blocks.10.attn.to_add_out.lora_A.weight",
+    "transformer.transformer_blocks.10.attn.to_add_out.lora_B.weight",
+    "transformer.transformer_blocks.10.attn.to_k.lora_A.weight",
+    "transformer.transformer_blocks.10.attn.to_k.lora_B.weight",
+    "transformer.transformer_blocks.10.attn.to_out.0.lora_A.weight",
+    "transformer.transformer_blocks.10.attn.to_out.0.lora_B.weight",
+    "transformer.transformer_blocks.10.attn.to_q.lora_A.weight",
+    "transformer.transformer_blocks.10.attn.to_q.lora_B.weight",
+    "transformer.transformer_blocks.10.attn.to_v.lora_A.weight",
+    "transformer.transformer_blocks.10.attn.to_v.lora_B.weight",
+    "transformer.transformer_blocks.10.ff.net.0.proj.lora_A.weight",
+    "transformer.transformer_blocks.10.ff.net.0.proj.lora_B.weight",
+    "transformer.transformer_blocks.10.ff.net.2.lora_A.weight",
+    "transformer.transformer_blocks.10.ff.net.2.lora_B.weight",
+    "transformer.transformer_blocks.10.ff_context.net.0.proj.lora_A.weight",
+    "transformer.transformer_blocks.10.ff_context.net.0.proj.lora_B.weight",
+    "transformer.transformer_blocks.10.ff_context.net.2.lora_A.weight",
+    "transformer.transformer_blocks.10.ff_context.net.2.lora_B.weight",
+    "transformer.transformer_blocks.10.norm1.linear.lora_A.weight",
+    "transformer.transformer_blocks.10.norm1.linear.lora_B.weight",
+    "transformer.transformer_blocks.10.norm1_context.linear.lora_A.weight",
+    "transformer.transformer_blocks.10.norm1_context.linear.lora_B.weight",
+    "transformer.transformer_blocks.11.attn.add_k_proj.lora_A.weight",
+    "transformer.transformer_blocks.11.attn.add_k_proj.lora_B.weight",
+    "transformer.transformer_blocks.11.attn.add_q_proj.lora_A.weight",
+    "transformer.transformer_blocks.11.attn.add_q_proj.lora_B.weight",
+    "transformer.transformer_blocks.11.attn.add_v_proj.lora_A.weight",
+    "transformer.transformer_blocks.11.attn.add_v_proj.lora_B.weight",
+    "transformer.transformer_blocks.11.attn.to_add_out.lora_A.weight",
+    "transformer.transformer_blocks.11.attn.to_add_out.lora_B.weight",
+    "transformer.transformer_blocks.11.attn.to_k.lora_A.weight",
+    "transformer.transformer_blocks.11.attn.to_k.lora_B.weight",
+    "transformer.transformer_blocks.11.attn.to_out.0.lora_A.weight",
+    "transformer.transformer_blocks.11.attn.to_out.0.lora_B.weight",
+    "transformer.transformer_blocks.11.attn.to_q.lora_A.weight",
+    "transformer.transformer_blocks.11.attn.to_q.lora_B.weight",
+    "transformer.transformer_blocks.11.attn.to_v.lora_A.weight",
+    "transformer.transformer_blocks.11.attn.to_v.lora_B.weight",
+    "transformer.transformer_blocks.11.ff.net.0.proj.lora_A.weight",
+    "transformer.transformer_blocks.11.ff.net.0.proj.lora_B.weight",
+    "transformer.transformer_blocks.11.ff.net.2.lora_A.weight",
+    "transformer.transformer_blocks.11.ff.net.2.lora_B.weight",
+    "transformer.transformer_blocks.11.ff_context.net.0.proj.lora_A.weight",
+    "transformer.transformer_blocks.11.ff_context.net.0.proj.lora_B.weight",
+    "transformer.transformer_blocks.11.ff_context.net.2.lora_A.weight",
+    "transformer.transformer_blocks.11.ff_context.net.2.lora_B.weight",
+    "transformer.transformer_blocks.11.norm1.linear.lora_A.weight",
+    "transformer.transformer_blocks.11.norm1.linear.lora_B.weight",
+    "transformer.transformer_blocks.11.norm1_context.linear.lora_A.weight",
+    "transformer.transformer_blocks.11.norm1_context.linear.lora_B.weight",
+    "transformer.transformer_blocks.12.attn.add_k_proj.lora_A.weight",
+    "transformer.transformer_blocks.12.attn.add_k_proj.lora_B.weight",
+    "transformer.transformer_blocks.12.attn.add_q_proj.lora_A.weight",
+    "transformer.transformer_blocks.12.attn.add_q_proj.lora_B.weight",
+    "transformer.transformer_blocks.12.attn.add_v_proj.lora_A.weight",
+    "transformer.transformer_blocks.12.attn.add_v_proj.lora_B.weight",
+    "transformer.transformer_blocks.12.attn.to_add_out.lora_A.weight",
+    "transformer.transformer_blocks.12.attn.to_add_out.lora_B.weight",
+    "transformer.transformer_blocks.12.attn.to_k.lora_A.weight",
+    "transformer.transformer_blocks.12.attn.to_k.lora_B.weight",
+    "transformer.transformer_blocks.12.attn.to_out.0.lora_A.weight",
+    "transformer.transformer_blocks.12.attn.to_out.0.lora_B.weight",
+    "transformer.transformer_blocks.12.attn.to_q.lora_A.weight",
+    "transformer.transformer_blocks.12.attn.to_q.lora_B.weight",
+    "transformer.transformer_blocks.12.attn.to_v.lora_A.weight",
+    "transformer.transformer_blocks.12.attn.to_v.lora_B.weight",
+    "transformer.transformer_blocks.12.ff.net.0.proj.lora_A.weight",
+    "transformer.transformer_blocks.12.ff.net.0.proj.lora_B.weight",
+    "transformer.transformer_blocks.12.ff.net.2.lora_A.weight",
+    "transformer.transformer_blocks.12.ff.net.2.lora_B.weight",
+    "transformer.transformer_blocks.12.ff_context.net.0.proj.lora_A.weight",
+    "transformer.transformer_blocks.12.ff_context.net.0.proj.lora_B.weight",
+    "transformer.transformer_blocks.12.ff_context.net.2.lora_A.weight",
+    "transformer.transformer_blocks.12.ff_context.net.2.lora_B.weight",
+    "transformer.transformer_blocks.12.norm1.linear.lora_A.weight",
+    "transformer.transformer_blocks.12.norm1.linear.lora_B.weight",
+    "transformer.transformer_blocks.12.norm1_context.linear.lora_A.weight",
+    "transformer.transformer_blocks.12.norm1_context.linear.lora_B.weight",
+    "transformer.transformer_blocks.13.attn.add_k_proj.lora_A.weight",
+    "transformer.transformer_blocks.13.attn.add_k_proj.lora_B.weight",
+    "transformer.transformer_blocks.13.attn.add_q_proj.lora_A.weight",
+    "transformer.transformer_blocks.13.attn.add_q_proj.lora_B.weight",
+    "transformer.transformer_blocks.13.attn.add_v_proj.lora_A.weight",
+    "transformer.transformer_blocks.13.attn.add_v_proj.lora_B.weight",
+    "transformer.transformer_blocks.13.attn.to_add_out.lora_A.weight",
+    "transformer.transformer_blocks.13.attn.to_add_out.lora_B.weight",
+    "transformer.transformer_blocks.13.attn.to_k.lora_A.weight",
+    "transformer.transformer_blocks.13.attn.to_k.lora_B.weight",
+    "transformer.transformer_blocks.13.attn.to_out.0.lora_A.weight",
+    "transformer.transformer_blocks.13.attn.to_out.0.lora_B.weight",
+    "transformer.transformer_blocks.13.attn.to_q.lora_A.weight",
+    "transformer.transformer_blocks.13.attn.to_q.lora_B.weight",
+    "transformer.transformer_blocks.13.attn.to_v.lora_A.weight",
+    "transformer.transformer_blocks.13.attn.to_v.lora_B.weight",
+    "transformer.transformer_blocks.13.ff.net.0.proj.lora_A.weight",
+    "transformer.transformer_blocks.13.ff.net.0.proj.lora_B.weight",
+    "transformer.transformer_blocks.13.ff.net.2.lora_A.weight",
+    "transformer.transformer_blocks.13.ff.net.2.lora_B.weight",
+    "transformer.transformer_blocks.13.ff_context.net.0.proj.lora_A.weight",
+    "transformer.transformer_blocks.13.ff_context.net.0.proj.lora_B.weight",
+    "transformer.transformer_blocks.13.ff_context.net.2.lora_A.weight",
+    "transformer.transformer_blocks.13.ff_context.net.2.lora_B.weight",
+    "transformer.transformer_blocks.13.norm1.linear.lora_A.weight",
+    "transformer.transformer_blocks.13.norm1.linear.lora_B.weight",
+    "transformer.transformer_blocks.13.norm1_context.linear.lora_A.weight",
+    "transformer.transformer_blocks.13.norm1_context.linear.lora_B.weight",
+    "transformer.transformer_blocks.14.attn.add_k_proj.lora_A.weight",
+    "transformer.transformer_blocks.14.attn.add_k_proj.lora_B.weight",
+    "transformer.transformer_blocks.14.attn.add_q_proj.lora_A.weight",
+    "transformer.transformer_blocks.14.attn.add_q_proj.lora_B.weight",
+    "transformer.transformer_blocks.14.attn.add_v_proj.lora_A.weight",
+    "transformer.transformer_blocks.14.attn.add_v_proj.lora_B.weight",
+    "transformer.transformer_blocks.14.attn.to_add_out.lora_A.weight",
+    "transformer.transformer_blocks.14.attn.to_add_out.lora_B.weight",
+    "transformer.transformer_blocks.14.attn.to_k.lora_A.weight",
+    "transformer.transformer_blocks.14.attn.to_k.lora_B.weight",
+    "transformer.transformer_blocks.14.attn.to_out.0.lora_A.weight",
+    "transformer.transformer_blocks.14.attn.to_out.0.lora_B.weight",
+    "transformer.transformer_blocks.14.attn.to_q.lora_A.weight",
+    "transformer.transformer_blocks.14.attn.to_q.lora_B.weight",
+    "transformer.transformer_blocks.14.attn.to_v.lora_A.weight",
+    "transformer.transformer_blocks.14.attn.to_v.lora_B.weight",
+    "transformer.transformer_blocks.14.ff.net.0.proj.lora_A.weight",
+    "transformer.transformer_blocks.14.ff.net.0.proj.lora_B.weight",
+    "transformer.transformer_blocks.14.ff.net.2.lora_A.weight",
+    "transformer.transformer_blocks.14.ff.net.2.lora_B.weight",
+    "transformer.transformer_blocks.14.ff_context.net.0.proj.lora_A.weight",
+    "transformer.transformer_blocks.14.ff_context.net.0.proj.lora_B.weight",
+    "transformer.transformer_blocks.14.ff_context.net.2.lora_A.weight",
+    "transformer.transformer_blocks.14.ff_context.net.2.lora_B.weight",
+    "transformer.transformer_blocks.14.norm1.linear.lora_A.weight",
+    "transformer.transformer_blocks.14.norm1.linear.lora_B.weight",
+    "transformer.transformer_blocks.14.norm1_context.linear.lora_A.weight",
+    "transformer.transformer_blocks.14.norm1_context.linear.lora_B.weight",
+    "transformer.transformer_blocks.15.attn.add_k_proj.lora_A.weight",
+    "transformer.transformer_blocks.15.attn.add_k_proj.lora_B.weight",
+    "transformer.transformer_blocks.15.attn.add_q_proj.lora_A.weight",
+    "transformer.transformer_blocks.15.attn.add_q_proj.lora_B.weight",
+    "transformer.transformer_blocks.15.attn.add_v_proj.lora_A.weight",
+    "transformer.transformer_blocks.15.attn.add_v_proj.lora_B.weight",
+    "transformer.transformer_blocks.15.attn.to_add_out.lora_A.weight",
+    "transformer.transformer_blocks.15.attn.to_add_out.lora_B.weight",
+    "transformer.transformer_blocks.15.attn.to_k.lora_A.weight",
+    "transformer.transformer_blocks.15.attn.to_k.lora_B.weight",
+    "transformer.transformer_blocks.15.attn.to_out.0.lora_A.weight",
+    "transformer.transformer_blocks.15.attn.to_out.0.lora_B.weight",
+    "transformer.transformer_blocks.15.attn.to_q.lora_A.weight",
+    "transformer.transformer_blocks.15.attn.to_q.lora_B.weight",
+    "transformer.transformer_blocks.15.attn.to_v.lora_A.weight",
+    "transformer.transformer_blocks.15.attn.to_v.lora_B.weight",
+    "transformer.transformer_blocks.15.ff.net.0.proj.lora_A.weight",
+    "transformer.transformer_blocks.15.ff.net.0.proj.lora_B.weight",
+    "transformer.transformer_blocks.15.ff.net.2.lora_A.weight",
+    "transformer.transformer_blocks.15.ff.net.2.lora_B.weight",
+    "transformer.transformer_blocks.15.ff_context.net.0.proj.lora_A.weight",
+    "transformer.transformer_blocks.15.ff_context.net.0.proj.lora_B.weight",
+    "transformer.transformer_blocks.15.ff_context.net.2.lora_A.weight",
+    "transformer.transformer_blocks.15.ff_context.net.2.lora_B.weight",
+    "transformer.transformer_blocks.15.norm1.linear.lora_A.weight",
+    "transformer.transformer_blocks.15.norm1.linear.lora_B.weight",
+    "transformer.transformer_blocks.15.norm1_context.linear.lora_A.weight",
+    "transformer.transformer_blocks.15.norm1_context.linear.lora_B.weight",
+    "transformer.transformer_blocks.16.attn.add_k_proj.lora_A.weight",
+    "transformer.transformer_blocks.16.attn.add_k_proj.lora_B.weight",
+    "transformer.transformer_blocks.16.attn.add_q_proj.lora_A.weight",
+    "transformer.transformer_blocks.16.attn.add_q_proj.lora_B.weight",
+    "transformer.transformer_blocks.16.attn.add_v_proj.lora_A.weight",
+    "transformer.transformer_blocks.16.attn.add_v_proj.lora_B.weight",
+    "transformer.transformer_blocks.16.attn.to_add_out.lora_A.weight",
+    "transformer.transformer_blocks.16.attn.to_add_out.lora_B.weight",
+    "transformer.transformer_blocks.16.attn.to_k.lora_A.weight",
+    "transformer.transformer_blocks.16.attn.to_k.lora_B.weight",
+    "transformer.transformer_blocks.16.attn.to_out.0.lora_A.weight",
+    "transformer.transformer_blocks.16.attn.to_out.0.lora_B.weight",
+    "transformer.transformer_blocks.16.attn.to_q.lora_A.weight",
+    "transformer.transformer_blocks.16.attn.to_q.lora_B.weight",
+    "transformer.transformer_blocks.16.attn.to_v.lora_A.weight",
+    "transformer.transformer_blocks.16.attn.to_v.lora_B.weight",
+    "transformer.transformer_blocks.16.ff.net.0.proj.lora_A.weight",
+    "transformer.transformer_blocks.16.ff.net.0.proj.lora_B.weight",
+    "transformer.transformer_blocks.16.ff.net.2.lora_A.weight",
+    "transformer.transformer_blocks.16.ff.net.2.lora_B.weight",
+    "transformer.transformer_blocks.16.ff_context.net.0.proj.lora_A.weight",
+    "transformer.transformer_blocks.16.ff_context.net.0.proj.lora_B.weight",
+    "transformer.transformer_blocks.16.ff_context.net.2.lora_A.weight",
+    "transformer.transformer_blocks.16.ff_context.net.2.lora_B.weight",
+    "transformer.transformer_blocks.16.norm1.linear.lora_A.weight",
+    "transformer.transformer_blocks.16.norm1.linear.lora_B.weight",
+    "transformer.transformer_blocks.16.norm1_context.linear.lora_A.weight",
+    "transformer.transformer_blocks.16.norm1_context.linear.lora_B.weight",
+    "transformer.transformer_blocks.17.attn.add_k_proj.lora_A.weight",
+    "transformer.transformer_blocks.17.attn.add_k_proj.lora_B.weight",
+    "transformer.transformer_blocks.17.attn.add_q_proj.lora_A.weight",
+    "transformer.transformer_blocks.17.attn.add_q_proj.lora_B.weight",
+    "transformer.transformer_blocks.17.attn.add_v_proj.lora_A.weight",
+    "transformer.transformer_blocks.17.attn.add_v_proj.lora_B.weight",
+    "transformer.transformer_blocks.17.attn.to_add_out.lora_A.weight",
+    "transformer.transformer_blocks.17.attn.to_add_out.lora_B.weight",
+    "transformer.transformer_blocks.17.attn.to_k.lora_A.weight",
+    "transformer.transformer_blocks.17.attn.to_k.lora_B.weight",
+    "transformer.transformer_blocks.17.attn.to_out.0.lora_A.weight",
+    "transformer.transformer_blocks.17.attn.to_out.0.lora_B.weight",
+    "transformer.transformer_blocks.17.attn.to_q.lora_A.weight",
+    "transformer.transformer_blocks.17.attn.to_q.lora_B.weight",
+    "transformer.transformer_blocks.17.attn.to_v.lora_A.weight",
+    "transformer.transformer_blocks.17.attn.to_v.lora_B.weight",
+    "transformer.transformer_blocks.17.ff.net.0.proj.lora_A.weight",
+    "transformer.transformer_blocks.17.ff.net.0.proj.lora_B.weight",
+    "transformer.transformer_blocks.17.ff.net.2.lora_A.weight",
+    "transformer.transformer_blocks.17.ff.net.2.lora_B.weight",
+    "transformer.transformer_blocks.17.ff_context.net.0.proj.lora_A.weight",
+    "transformer.transformer_blocks.17.ff_context.net.0.proj.lora_B.weight",
+    "transformer.transformer_blocks.17.ff_context.net.2.lora_A.weight",
+    "transformer.transformer_blocks.17.ff_context.net.2.lora_B.weight",
+    "transformer.transformer_blocks.17.norm1.linear.lora_A.weight",
+    "transformer.transformer_blocks.17.norm1.linear.lora_B.weight",
+    "transformer.transformer_blocks.17.norm1_context.linear.lora_A.weight",
+    "transformer.transformer_blocks.17.norm1_context.linear.lora_B.weight",
+    "transformer.transformer_blocks.18.attn.add_k_proj.lora_A.weight",
+    "transformer.transformer_blocks.18.attn.add_k_proj.lora_B.weight",
+    "transformer.transformer_blocks.18.attn.add_q_proj.lora_A.weight",
+    "transformer.transformer_blocks.18.attn.add_q_proj.lora_B.weight",
+    "transformer.transformer_blocks.18.attn.add_v_proj.lora_A.weight",
+    "transformer.transformer_blocks.18.attn.add_v_proj.lora_B.weight",
+    "transformer.transformer_blocks.18.attn.to_add_out.lora_A.weight",
+    "transformer.transformer_blocks.18.attn.to_add_out.lora_B.weight",
+    "transformer.transformer_blocks.18.attn.to_k.lora_A.weight",
+    "transformer.transformer_blocks.18.attn.to_k.lora_B.weight",
+    "transformer.transformer_blocks.18.attn.to_out.0.lora_A.weight",
+    "transformer.transformer_blocks.18.attn.to_out.0.lora_B.weight",
+    "transformer.transformer_blocks.18.attn.to_q.lora_A.weight",
+    "transformer.transformer_blocks.18.attn.to_q.lora_B.weight",
+    "transformer.transformer_blocks.18.attn.to_v.lora_A.weight",
+    "transformer.transformer_blocks.18.attn.to_v.lora_B.weight",
+    "transformer.transformer_blocks.18.ff.net.0.proj.lora_A.weight",
+    "transformer.transformer_blocks.18.ff.net.0.proj.lora_B.weight",
+    "transformer.transformer_blocks.18.ff.net.2.lora_A.weight",
+    "transformer.transformer_blocks.18.ff.net.2.lora_B.weight",
+    "transformer.transformer_blocks.18.ff_context.net.0.proj.lora_A.weight",
+    "transformer.transformer_blocks.18.ff_context.net.0.proj.lora_B.weight",
+    "transformer.transformer_blocks.18.ff_context.net.2.lora_A.weight",
+    "transformer.transformer_blocks.18.ff_context.net.2.lora_B.weight",
+    "transformer.transformer_blocks.18.norm1.linear.lora_A.weight",
+    "transformer.transformer_blocks.18.norm1.linear.lora_B.weight",
+    "transformer.transformer_blocks.18.norm1_context.linear.lora_A.weight",
+    "transformer.transformer_blocks.18.norm1_context.linear.lora_B.weight",
+    "transformer.transformer_blocks.2.attn.add_k_proj.lora_A.weight",
+    "transformer.transformer_blocks.2.attn.add_k_proj.lora_B.weight",
+    "transformer.transformer_blocks.2.attn.add_q_proj.lora_A.weight",
+    "transformer.transformer_blocks.2.attn.add_q_proj.lora_B.weight",
+    "transformer.transformer_blocks.2.attn.add_v_proj.lora_A.weight",
+    "transformer.transformer_blocks.2.attn.add_v_proj.lora_B.weight",
+    "transformer.transformer_blocks.2.attn.to_add_out.lora_A.weight",
+    "transformer.transformer_blocks.2.attn.to_add_out.lora_B.weight",
+    "transformer.transformer_blocks.2.attn.to_k.lora_A.weight",
+    "transformer.transformer_blocks.2.attn.to_k.lora_B.weight",
+    "transformer.transformer_blocks.2.attn.to_out.0.lora_A.weight",
+    "transformer.transformer_blocks.2.attn.to_out.0.lora_B.weight",
+    "transformer.transformer_blocks.2.attn.to_q.lora_A.weight",
+    "transformer.transformer_blocks.2.attn.to_q.lora_B.weight",
+    "transformer.transformer_blocks.2.attn.to_v.lora_A.weight",
+    "transformer.transformer_blocks.2.attn.to_v.lora_B.weight",
+    "transformer.transformer_blocks.2.ff.net.0.proj.lora_A.weight",
+    "transformer.transformer_blocks.2.ff.net.0.proj.lora_B.weight",
+    "transformer.transformer_blocks.2.ff.net.2.lora_A.weight",
+    "transformer.transformer_blocks.2.ff.net.2.lora_B.weight",
+    "transformer.transformer_blocks.2.ff_context.net.0.proj.lora_A.weight",
+    "transformer.transformer_blocks.2.ff_context.net.0.proj.lora_B.weight",
+    "transformer.transformer_blocks.2.ff_context.net.2.lora_A.weight",
+    "transformer.transformer_blocks.2.ff_context.net.2.lora_B.weight",
+    "transformer.transformer_blocks.2.norm1.linear.lora_A.weight",
+    "transformer.transformer_blocks.2.norm1.linear.lora_B.weight",
+    "transformer.transformer_blocks.2.norm1_context.linear.lora_A.weight",
+    "transformer.transformer_blocks.2.norm1_context.linear.lora_B.weight",
+    "transformer.transformer_blocks.3.attn.add_k_proj.lora_A.weight",
+    "transformer.transformer_blocks.3.attn.add_k_proj.lora_B.weight",
+    "transformer.transformer_blocks.3.attn.add_q_proj.lora_A.weight",
+    "transformer.transformer_blocks.3.attn.add_q_proj.lora_B.weight",
+    "transformer.transformer_blocks.3.attn.add_v_proj.lora_A.weight",
+    "transformer.transformer_blocks.3.attn.add_v_proj.lora_B.weight",
+    "transformer.transformer_blocks.3.attn.to_add_out.lora_A.weight",
+    "transformer.transformer_blocks.3.attn.to_add_out.lora_B.weight",
+    "transformer.transformer_blocks.3.attn.to_k.lora_A.weight",
+    "transformer.transformer_blocks.3.attn.to_k.lora_B.weight",
+    "transformer.transformer_blocks.3.attn.to_out.0.lora_A.weight",
+    "transformer.transformer_blocks.3.attn.to_out.0.lora_B.weight",
+    "transformer.transformer_blocks.3.attn.to_q.lora_A.weight",
+    "transformer.transformer_blocks.3.attn.to_q.lora_B.weight",
+    "transformer.transformer_blocks.3.attn.to_v.lora_A.weight",
+    "transformer.transformer_blocks.3.attn.to_v.lora_B.weight",
+    "transformer.transformer_blocks.3.ff.net.0.proj.lora_A.weight",
+    "transformer.transformer_blocks.3.ff.net.0.proj.lora_B.weight",
+    "transformer.transformer_blocks.3.ff.net.2.lora_A.weight",
+    "transformer.transformer_blocks.3.ff.net.2.lora_B.weight",
+    "transformer.transformer_blocks.3.ff_context.net.0.proj.lora_A.weight",
+    "transformer.transformer_blocks.3.ff_context.net.0.proj.lora_B.weight",
+    "transformer.transformer_blocks.3.ff_context.net.2.lora_A.weight",
+    "transformer.transformer_blocks.3.ff_context.net.2.lora_B.weight",
+    "transformer.transformer_blocks.3.norm1.linear.lora_A.weight",
+    "transformer.transformer_blocks.3.norm1.linear.lora_B.weight",
+    "transformer.transformer_blocks.3.norm1_context.linear.lora_A.weight",
+    "transformer.transformer_blocks.3.norm1_context.linear.lora_B.weight",
+    "transformer.transformer_blocks.4.attn.add_k_proj.lora_A.weight",
+    "transformer.transformer_blocks.4.attn.add_k_proj.lora_B.weight",
+    "transformer.transformer_blocks.4.attn.add_q_proj.lora_A.weight",
+    "transformer.transformer_blocks.4.attn.add_q_proj.lora_B.weight",
+    "transformer.transformer_blocks.4.attn.add_v_proj.lora_A.weight",
+    "transformer.transformer_blocks.4.attn.add_v_proj.lora_B.weight",
+    "transformer.transformer_blocks.4.attn.to_add_out.lora_A.weight",
+    "transformer.transformer_blocks.4.attn.to_add_out.lora_B.weight",
+    "transformer.transformer_blocks.4.attn.to_k.lora_A.weight",
+    "transformer.transformer_blocks.4.attn.to_k.lora_B.weight",
+    "transformer.transformer_blocks.4.attn.to_out.0.lora_A.weight",
+    "transformer.transformer_blocks.4.attn.to_out.0.lora_B.weight",
+    "transformer.transformer_blocks.4.attn.to_q.lora_A.weight",
+    "transformer.transformer_blocks.4.attn.to_q.lora_B.weight",
+    "transformer.transformer_blocks.4.attn.to_v.lora_A.weight",
+    "transformer.transformer_blocks.4.attn.to_v.lora_B.weight",
+    "transformer.transformer_blocks.4.ff.net.0.proj.lora_A.weight",
+    "transformer.transformer_blocks.4.ff.net.0.proj.lora_B.weight",
+    "transformer.transformer_blocks.4.ff.net.2.lora_A.weight",
+    "transformer.transformer_blocks.4.ff.net.2.lora_B.weight",
+    "transformer.transformer_blocks.4.ff_context.net.0.proj.lora_A.weight",
+    "transformer.transformer_blocks.4.ff_context.net.0.proj.lora_B.weight",
+    "transformer.transformer_blocks.4.ff_context.net.2.lora_A.weight",
+    "transformer.transformer_blocks.4.ff_context.net.2.lora_B.weight",
+    "transformer.transformer_blocks.4.norm1.linear.lora_A.weight",
+    "transformer.transformer_blocks.4.norm1.linear.lora_B.weight",
+    "transformer.transformer_blocks.4.norm1_context.linear.lora_A.weight",
+    "transformer.transformer_blocks.4.norm1_context.linear.lora_B.weight",
+    "transformer.transformer_blocks.5.attn.add_k_proj.lora_A.weight",
+    "transformer.transformer_blocks.5.attn.add_k_proj.lora_B.weight",
+    "transformer.transformer_blocks.5.attn.add_q_proj.lora_A.weight",
+    "transformer.transformer_blocks.5.attn.add_q_proj.lora_B.weight",
+    "transformer.transformer_blocks.5.attn.add_v_proj.lora_A.weight",
+    "transformer.transformer_blocks.5.attn.add_v_proj.lora_B.weight",
+    "transformer.transformer_blocks.5.attn.to_add_out.lora_A.weight",
+    "transformer.transformer_blocks.5.attn.to_add_out.lora_B.weight",
+    "transformer.transformer_blocks.5.attn.to_k.lora_A.weight",
+    "transformer.transformer_blocks.5.attn.to_k.lora_B.weight",
+    "transformer.transformer_blocks.5.attn.to_out.0.lora_A.weight",
+    "transformer.transformer_blocks.5.attn.to_out.0.lora_B.weight",
+    "transformer.transformer_blocks.5.attn.to_q.lora_A.weight",
+    "transformer.transformer_blocks.5.attn.to_q.lora_B.weight",
+    "transformer.transformer_blocks.5.attn.to_v.lora_A.weight",
+    "transformer.transformer_blocks.5.attn.to_v.lora_B.weight",
+    "transformer.transformer_blocks.5.ff.net.0.proj.lora_A.weight",
+    "transformer.transformer_blocks.5.ff.net.0.proj.lora_B.weight",
+    "transformer.transformer_blocks.5.ff.net.2.lora_A.weight",
+    "transformer.transformer_blocks.5.ff.net.2.lora_B.weight",
+    "transformer.transformer_blocks.5.ff_context.net.0.proj.lora_A.weight",
+    "transformer.transformer_blocks.5.ff_context.net.0.proj.lora_B.weight",
+    "transformer.transformer_blocks.5.ff_context.net.2.lora_A.weight",
+    "transformer.transformer_blocks.5.ff_context.net.2.lora_B.weight",
+    "transformer.transformer_blocks.5.norm1.linear.lora_A.weight",
+    "transformer.transformer_blocks.5.norm1.linear.lora_B.weight",
+    "transformer.transformer_blocks.5.norm1_context.linear.lora_A.weight",
+    "transformer.transformer_blocks.5.norm1_context.linear.lora_B.weight",
+    "transformer.transformer_blocks.6.attn.add_k_proj.lora_A.weight",
+    "transformer.transformer_blocks.6.attn.add_k_proj.lora_B.weight",
+    "transformer.transformer_blocks.6.attn.add_q_proj.lora_A.weight",
+    "transformer.transformer_blocks.6.attn.add_q_proj.lora_B.weight",
+    "transformer.transformer_blocks.6.attn.add_v_proj.lora_A.weight",
+    "transformer.transformer_blocks.6.attn.add_v_proj.lora_B.weight",
+    "transformer.transformer_blocks.6.attn.to_add_out.lora_A.weight",
+    "transformer.transformer_blocks.6.attn.to_add_out.lora_B.weight",
+    "transformer.transformer_blocks.6.attn.to_k.lora_A.weight",
+    "transformer.transformer_blocks.6.attn.to_k.lora_B.weight",
+    "transformer.transformer_blocks.6.attn.to_out.0.lora_A.weight",
+    "transformer.transformer_blocks.6.attn.to_out.0.lora_B.weight",
+    "transformer.transformer_blocks.6.attn.to_q.lora_A.weight",
+    "transformer.transformer_blocks.6.attn.to_q.lora_B.weight",
+    "transformer.transformer_blocks.6.attn.to_v.lora_A.weight",
+    "transformer.transformer_blocks.6.attn.to_v.lora_B.weight",
+    "transformer.transformer_blocks.6.ff.net.0.proj.lora_A.weight",
+    "transformer.transformer_blocks.6.ff.net.0.proj.lora_B.weight",
+    "transformer.transformer_blocks.6.ff.net.2.lora_A.weight",
+    "transformer.transformer_blocks.6.ff.net.2.lora_B.weight",
+    "transformer.transformer_blocks.6.ff_context.net.0.proj.lora_A.weight",
+    "transformer.transformer_blocks.6.ff_context.net.0.proj.lora_B.weight",
+    "transformer.transformer_blocks.6.ff_context.net.2.lora_A.weight",
+    "transformer.transformer_blocks.6.ff_context.net.2.lora_B.weight",
+    "transformer.transformer_blocks.6.norm1.linear.lora_A.weight",
+    "transformer.transformer_blocks.6.norm1.linear.lora_B.weight",
+    "transformer.transformer_blocks.6.norm1_context.linear.lora_A.weight",
+    "transformer.transformer_blocks.6.norm1_context.linear.lora_B.weight",
+    "transformer.transformer_blocks.7.attn.add_k_proj.lora_A.weight",
+    "transformer.transformer_blocks.7.attn.add_k_proj.lora_B.weight",
+    "transformer.transformer_blocks.7.attn.add_q_proj.lora_A.weight",
+    "transformer.transformer_blocks.7.attn.add_q_proj.lora_B.weight",
+    "transformer.transformer_blocks.7.attn.add_v_proj.lora_A.weight",
+    "transformer.transformer_blocks.7.attn.add_v_proj.lora_B.weight",
+    "transformer.transformer_blocks.7.attn.to_add_out.lora_A.weight",
+    "transformer.transformer_blocks.7.attn.to_add_out.lora_B.weight",
+    "transformer.transformer_blocks.7.attn.to_k.lora_A.weight",
+    "transformer.transformer_blocks.7.attn.to_k.lora_B.weight",
+    "transformer.transformer_blocks.7.attn.to_out.0.lora_A.weight",
+    "transformer.transformer_blocks.7.attn.to_out.0.lora_B.weight",
+    "transformer.transformer_blocks.7.attn.to_q.lora_A.weight",
+    "transformer.transformer_blocks.7.attn.to_q.lora_B.weight",
+    "transformer.transformer_blocks.7.attn.to_v.lora_A.weight",
+    "transformer.transformer_blocks.7.attn.to_v.lora_B.weight",
+    "transformer.transformer_blocks.7.ff.net.0.proj.lora_A.weight",
+    "transformer.transformer_blocks.7.ff.net.0.proj.lora_B.weight",
+    "transformer.transformer_blocks.7.ff.net.2.lora_A.weight",
+    "transformer.transformer_blocks.7.ff.net.2.lora_B.weight",
+    "transformer.transformer_blocks.7.ff_context.net.0.proj.lora_A.weight",
+    "transformer.transformer_blocks.7.ff_context.net.0.proj.lora_B.weight",
+    "transformer.transformer_blocks.7.ff_context.net.2.lora_A.weight",
+    "transformer.transformer_blocks.7.ff_context.net.2.lora_B.weight",
+    "transformer.transformer_blocks.7.norm1.linear.lora_A.weight",
+    "transformer.transformer_blocks.7.norm1.linear.lora_B.weight",
+    "transformer.transformer_blocks.7.norm1_context.linear.lora_A.weight",
+    "transformer.transformer_blocks.7.norm1_context.linear.lora_B.weight",
+    "transformer.transformer_blocks.8.attn.add_k_proj.lora_A.weight",
+    "transformer.transformer_blocks.8.attn.add_k_proj.lora_B.weight",
+    "transformer.transformer_blocks.8.attn.add_q_proj.lora_A.weight",
+    "transformer.transformer_blocks.8.attn.add_q_proj.lora_B.weight",
+    "transformer.transformer_blocks.8.attn.add_v_proj.lora_A.weight",
+    "transformer.transformer_blocks.8.attn.add_v_proj.lora_B.weight",
+    "transformer.transformer_blocks.8.attn.to_add_out.lora_A.weight",
+    "transformer.transformer_blocks.8.attn.to_add_out.lora_B.weight",
+    "transformer.transformer_blocks.8.attn.to_k.lora_A.weight",
+    "transformer.transformer_blocks.8.attn.to_k.lora_B.weight",
+    "transformer.transformer_blocks.8.attn.to_out.0.lora_A.weight",
+    "transformer.transformer_blocks.8.attn.to_out.0.lora_B.weight",
+    "transformer.transformer_blocks.8.attn.to_q.lora_A.weight",
+    "transformer.transformer_blocks.8.attn.to_q.lora_B.weight",
+    "transformer.transformer_blocks.8.attn.to_v.lora_A.weight",
+    "transformer.transformer_blocks.8.attn.to_v.lora_B.weight",
+    "transformer.transformer_blocks.8.ff.net.0.proj.lora_A.weight",
+    "transformer.transformer_blocks.8.ff.net.0.proj.lora_B.weight",
+    "transformer.transformer_blocks.8.ff.net.2.lora_A.weight",
+    "transformer.transformer_blocks.8.ff.net.2.lora_B.weight",
+    "transformer.transformer_blocks.8.ff_context.net.0.proj.lora_A.weight",
+    "transformer.transformer_blocks.8.ff_context.net.0.proj.lora_B.weight",
+    "transformer.transformer_blocks.8.ff_context.net.2.lora_A.weight",
+    "transformer.transformer_blocks.8.ff_context.net.2.lora_B.weight",
+    "transformer.transformer_blocks.8.norm1.linear.lora_A.weight",
+    "transformer.transformer_blocks.8.norm1.linear.lora_B.weight",
+    "transformer.transformer_blocks.8.norm1_context.linear.lora_A.weight",
+    "transformer.transformer_blocks.8.norm1_context.linear.lora_B.weight",
+    "transformer.transformer_blocks.9.attn.add_k_proj.lora_A.weight",
+    "transformer.transformer_blocks.9.attn.add_k_proj.lora_B.weight",
+    "transformer.transformer_blocks.9.attn.add_q_proj.lora_A.weight",
+    "transformer.transformer_blocks.9.attn.add_q_proj.lora_B.weight",
+    "transformer.transformer_blocks.9.attn.add_v_proj.lora_A.weight",
+    "transformer.transformer_blocks.9.attn.add_v_proj.lora_B.weight",
+    "transformer.transformer_blocks.9.attn.to_add_out.lora_A.weight",
+    "transformer.transformer_blocks.9.attn.to_add_out.lora_B.weight",
+    "transformer.transformer_blocks.9.attn.to_k.lora_A.weight",
+    "transformer.transformer_blocks.9.attn.to_k.lora_B.weight",
+    "transformer.transformer_blocks.9.attn.to_out.0.lora_A.weight",
+    "transformer.transformer_blocks.9.attn.to_out.0.lora_B.weight",
+    "transformer.transformer_blocks.9.attn.to_q.lora_A.weight",
+    "transformer.transformer_blocks.9.attn.to_q.lora_B.weight",
+    "transformer.transformer_blocks.9.attn.to_v.lora_A.weight",
+    "transformer.transformer_blocks.9.attn.to_v.lora_B.weight",
+    "transformer.transformer_blocks.9.ff.net.0.proj.lora_A.weight",
+    "transformer.transformer_blocks.9.ff.net.0.proj.lora_B.weight",
+    "transformer.transformer_blocks.9.ff.net.2.lora_A.weight",
+    "transformer.transformer_blocks.9.ff.net.2.lora_B.weight",
+    "transformer.transformer_blocks.9.ff_context.net.0.proj.lora_A.weight",
+    "transformer.transformer_blocks.9.ff_context.net.0.proj.lora_B.weight",
+    "transformer.transformer_blocks.9.ff_context.net.2.lora_A.weight",
+    "transformer.transformer_blocks.9.ff_context.net.2.lora_B.weight",
+    "transformer.transformer_blocks.9.norm1.linear.lora_A.weight",
+    "transformer.transformer_blocks.9.norm1.linear.lora_B.weight",
+    "transformer.transformer_blocks.9.norm1_context.linear.lora_A.weight",
+    "transformer.transformer_blocks.9.norm1_context.linear.lora_B.weight",
+]

tests/backend/peft/conversions/lora_state_dicts/flux_lora_kohya_format.py

@@ -0,0 +1,914 @@
+state_dict_keys = [
+    "lora_unet_double_blocks_0_img_attn_proj.alpha",
+    "lora_unet_double_blocks_0_img_attn_proj.lora_down.weight",
+    "lora_unet_double_blocks_0_img_attn_proj.lora_up.weight",
+    "lora_unet_double_blocks_0_img_attn_qkv.alpha",
+    "lora_unet_double_blocks_0_img_attn_qkv.lora_down.weight",
+    "lora_unet_double_blocks_0_img_attn_qkv.lora_up.weight",
+    "lora_unet_double_blocks_0_img_mlp_0.alpha",
+    "lora_unet_double_blocks_0_img_mlp_0.lora_down.weight",
+    "lora_unet_double_blocks_0_img_mlp_0.lora_up.weight",
+    "lora_unet_double_blocks_0_img_mlp_2.alpha",
+    "lora_unet_double_blocks_0_img_mlp_2.lora_down.weight",
+    "lora_unet_double_blocks_0_img_mlp_2.lora_up.weight",
+    "lora_unet_double_blocks_0_img_mod_lin.alpha",
+    "lora_unet_double_blocks_0_img_mod_lin.lora_down.weight",
+    "lora_unet_double_blocks_0_img_mod_lin.lora_up.weight",
+    "lora_unet_double_blocks_0_txt_attn_proj.alpha",
+    "lora_unet_double_blocks_0_txt_attn_proj.lora_down.weight",
+    "lora_unet_double_blocks_0_txt_attn_proj.lora_up.weight",
+    "lora_unet_double_blocks_0_txt_attn_qkv.alpha",
+    "lora_unet_double_blocks_0_txt_attn_qkv.lora_down.weight",
+    "lora_unet_double_blocks_0_txt_attn_qkv.lora_up.weight",
+    "lora_unet_double_blocks_0_txt_mlp_0.alpha",
+    "lora_unet_double_blocks_0_txt_mlp_0.lora_down.weight",
+    "lora_unet_double_blocks_0_txt_mlp_0.lora_up.weight",
+    "lora_unet_double_blocks_0_txt_mlp_2.alpha",
+    "lora_unet_double_blocks_0_txt_mlp_2.lora_down.weight",
+    "lora_unet_double_blocks_0_txt_mlp_2.lora_up.weight",
+    "lora_unet_double_blocks_0_txt_mod_lin.alpha",
+    "lora_unet_double_blocks_0_txt_mod_lin.lora_down.weight",
+    "lora_unet_double_blocks_0_txt_mod_lin.lora_up.weight",
+    "lora_unet_double_blocks_10_img_attn_proj.alpha",
+    "lora_unet_double_blocks_10_img_attn_proj.lora_down.weight",
+    "lora_unet_double_blocks_10_img_attn_proj.lora_up.weight",
+    "lora_unet_double_blocks_10_img_attn_qkv.alpha",
+    "lora_unet_double_blocks_10_img_attn_qkv.lora_down.weight",
+    "lora_unet_double_blocks_10_img_attn_qkv.lora_up.weight",
+    "lora_unet_double_blocks_10_img_mlp_0.alpha",
+    "lora_unet_double_blocks_10_img_mlp_0.lora_down.weight",
+    "lora_unet_double_blocks_10_img_mlp_0.lora_up.weight",
+    "lora_unet_double_blocks_10_img_mlp_2.alpha",
+    "lora_unet_double_blocks_10_img_mlp_2.lora_down.weight",
+    "lora_unet_double_blocks_10_img_mlp_2.lora_up.weight",
+    "lora_unet_double_blocks_10_img_mod_lin.alpha",
+    "lora_unet_double_blocks_10_img_mod_lin.lora_down.weight",
+    "lora_unet_double_blocks_10_img_mod_lin.lora_up.weight",
+    "lora_unet_double_blocks_10_txt_attn_proj.alpha",
+    "lora_unet_double_blocks_10_txt_attn_proj.lora_down.weight",
+    "lora_unet_double_blocks_10_txt_attn_proj.lora_up.weight",
+    "lora_unet_double_blocks_10_txt_attn_qkv.alpha",
+    "lora_unet_double_blocks_10_txt_attn_qkv.lora_down.weight",
+    "lora_unet_double_blocks_10_txt_attn_qkv.lora_up.weight",
+    "lora_unet_double_blocks_10_txt_mlp_0.alpha",
+    "lora_unet_double_blocks_10_txt_mlp_0.lora_down.weight",
+    "lora_unet_double_blocks_10_txt_mlp_0.lora_up.weight",
+    "lora_unet_double_blocks_10_txt_mlp_2.alpha",
+    "lora_unet_double_blocks_10_txt_mlp_2.lora_down.weight",
+    "lora_unet_double_blocks_10_txt_mlp_2.lora_up.weight",
+    "lora_unet_double_blocks_10_txt_mod_lin.alpha",
+    "lora_unet_double_blocks_10_txt_mod_lin.lora_down.weight",
+    "lora_unet_double_blocks_10_txt_mod_lin.lora_up.weight",
+    "lora_unet_double_blocks_11_img_attn_proj.alpha",
+    "lora_unet_double_blocks_11_img_attn_proj.lora_down.weight",
+    "lora_unet_double_blocks_11_img_attn_proj.lora_up.weight",
+    "lora_unet_double_blocks_11_img_attn_qkv.alpha",
+    "lora_unet_double_blocks_11_img_attn_qkv.lora_down.weight",
+    "lora_unet_double_blocks_11_img_attn_qkv.lora_up.weight",
+    "lora_unet_double_blocks_11_img_mlp_0.alpha",
+    "lora_unet_double_blocks_11_img_mlp_0.lora_down.weight",
+    "lora_unet_double_blocks_11_img_mlp_0.lora_up.weight",
+    "lora_unet_double_blocks_11_img_mlp_2.alpha",
+    "lora_unet_double_blocks_11_img_mlp_2.lora_down.weight",
+    "lora_unet_double_blocks_11_img_mlp_2.lora_up.weight",
+    "lora_unet_double_blocks_11_img_mod_lin.alpha",
+    "lora_unet_double_blocks_11_img_mod_lin.lora_down.weight",
+    "lora_unet_double_blocks_11_img_mod_lin.lora_up.weight",
+    "lora_unet_double_blocks_11_txt_attn_proj.alpha",
+    "lora_unet_double_blocks_11_txt_attn_proj.lora_down.weight",
+    "lora_unet_double_blocks_11_txt_attn_proj.lora_up.weight",
+    "lora_unet_double_blocks_11_txt_attn_qkv.alpha",
+    "lora_unet_double_blocks_11_txt_attn_qkv.lora_down.weight",
+    "lora_unet_double_blocks_11_txt_attn_qkv.lora_up.weight",
+    "lora_unet_double_blocks_11_txt_mlp_0.alpha",
+    "lora_unet_double_blocks_11_txt_mlp_0.lora_down.weight",
+    "lora_unet_double_blocks_11_txt_mlp_0.lora_up.weight",
+    "lora_unet_double_blocks_11_txt_mlp_2.alpha",
+    "lora_unet_double_blocks_11_txt_mlp_2.lora_down.weight",
+    "lora_unet_double_blocks_11_txt_mlp_2.lora_up.weight",
+    "lora_unet_double_blocks_11_txt_mod_lin.alpha",
+    "lora_unet_double_blocks_11_txt_mod_lin.lora_down.weight",
+    "lora_unet_double_blocks_11_txt_mod_lin.lora_up.weight",
+    "lora_unet_double_blocks_12_img_attn_proj.alpha",
+    "lora_unet_double_blocks_12_img_attn_proj.lora_down.weight",
+    "lora_unet_double_blocks_12_img_attn_proj.lora_up.weight",
+    "lora_unet_double_blocks_12_img_attn_qkv.alpha",
+    "lora_unet_double_blocks_12_img_attn_qkv.lora_down.weight",
+    "lora_unet_double_blocks_12_img_attn_qkv.lora_up.weight",
+    "lora_unet_double_blocks_12_img_mlp_0.alpha",
+    "lora_unet_double_blocks_12_img_mlp_0.lora_down.weight",
+    "lora_unet_double_blocks_12_img_mlp_0.lora_up.weight",
+    "lora_unet_double_blocks_12_img_mlp_2.alpha",
+    "lora_unet_double_blocks_12_img_mlp_2.lora_down.weight",
+    "lora_unet_double_blocks_12_img_mlp_2.lora_up.weight",
+    "lora_unet_double_blocks_12_img_mod_lin.alpha",
+    "lora_unet_double_blocks_12_img_mod_lin.lora_down.weight",
+    "lora_unet_double_blocks_12_img_mod_lin.lora_up.weight",
+    "lora_unet_double_blocks_12_txt_attn_proj.alpha",
+    "lora_unet_double_blocks_12_txt_attn_proj.lora_down.weight",
+    "lora_unet_double_blocks_12_txt_attn_proj.lora_up.weight",
+    "lora_unet_double_blocks_12_txt_attn_qkv.alpha",
+    "lora_unet_double_blocks_12_txt_attn_qkv.lora_down.weight",
+    "lora_unet_double_blocks_12_txt_attn_qkv.lora_up.weight",
+    "lora_unet_double_blocks_12_txt_mlp_0.alpha",
+    "lora_unet_double_blocks_12_txt_mlp_0.lora_down.weight",
+    "lora_unet_double_blocks_12_txt_mlp_0.lora_up.weight",
+    "lora_unet_double_blocks_12_txt_mlp_2.alpha",
+    "lora_unet_double_blocks_12_txt_mlp_2.lora_down.weight",
+    "lora_unet_double_blocks_12_txt_mlp_2.lora_up.weight",
+    "lora_unet_double_blocks_12_txt_mod_lin.alpha",
+    "lora_unet_double_blocks_12_txt_mod_lin.lora_down.weight",
+    "lora_unet_double_blocks_12_txt_mod_lin.lora_up.weight",
+    "lora_unet_double_blocks_13_img_attn_proj.alpha",
+    "lora_unet_double_blocks_13_img_attn_proj.lora_down.weight",
+    "lora_unet_double_blocks_13_img_attn_proj.lora_up.weight",
+    "lora_unet_double_blocks_13_img_attn_qkv.alpha",
+    "lora_unet_double_blocks_13_img_attn_qkv.lora_down.weight",
+    "lora_unet_double_blocks_13_img_attn_qkv.lora_up.weight",
+    "lora_unet_double_blocks_13_img_mlp_0.alpha",
+    "lora_unet_double_blocks_13_img_mlp_0.lora_down.weight",
+    "lora_unet_double_blocks_13_img_mlp_0.lora_up.weight",
+    "lora_unet_double_blocks_13_img_mlp_2.alpha",
+    "lora_unet_double_blocks_13_img_mlp_2.lora_down.weight",
+    "lora_unet_double_blocks_13_img_mlp_2.lora_up.weight",
+    "lora_unet_double_blocks_13_img_mod_lin.alpha",
+    "lora_unet_double_blocks_13_img_mod_lin.lora_down.weight",
+    "lora_unet_double_blocks_13_img_mod_lin.lora_up.weight",
+    "lora_unet_double_blocks_13_txt_attn_proj.alpha",
+    "lora_unet_double_blocks_13_txt_attn_proj.lora_down.weight",
+    "lora_unet_double_blocks_13_txt_attn_proj.lora_up.weight",
+    "lora_unet_double_blocks_13_txt_attn_qkv.alpha",
+    "lora_unet_double_blocks_13_txt_attn_qkv.lora_down.weight",
+    "lora_unet_double_blocks_13_txt_attn_qkv.lora_up.weight",
+    "lora_unet_double_blocks_13_txt_mlp_0.alpha",
+    "lora_unet_double_blocks_13_txt_mlp_0.lora_down.weight",
+    "lora_unet_double_blocks_13_txt_mlp_0.lora_up.weight",
+    "lora_unet_double_blocks_13_txt_mlp_2.alpha",
+    "lora_unet_double_blocks_13_txt_mlp_2.lora_down.weight",
+    "lora_unet_double_blocks_13_txt_mlp_2.lora_up.weight",
+    "lora_unet_double_blocks_13_txt_mod_lin.alpha",
+    "lora_unet_double_blocks_13_txt_mod_lin.lora_down.weight",
+    "lora_unet_double_blocks_13_txt_mod_lin.lora_up.weight",
+    "lora_unet_double_blocks_14_img_attn_proj.alpha",
+    "lora_unet_double_blocks_14_img_attn_proj.lora_down.weight",
+    "lora_unet_double_blocks_14_img_attn_proj.lora_up.weight",
+    "lora_unet_double_blocks_14_img_attn_qkv.alpha",
+    "lora_unet_double_blocks_14_img_attn_qkv.lora_down.weight",
+    "lora_unet_double_blocks_14_img_attn_qkv.lora_up.weight",
+    "lora_unet_double_blocks_14_img_mlp_0.alpha",
+    "lora_unet_double_blocks_14_img_mlp_0.lora_down.weight",
+    "lora_unet_double_blocks_14_img_mlp_0.lora_up.weight",
+    "lora_unet_double_blocks_14_img_mlp_2.alpha",
+    "lora_unet_double_blocks_14_img_mlp_2.lora_down.weight",
+    "lora_unet_double_blocks_14_img_mlp_2.lora_up.weight",
+    "lora_unet_double_blocks_14_img_mod_lin.alpha",
+    "lora_unet_double_blocks_14_img_mod_lin.lora_down.weight",
+    "lora_unet_double_blocks_14_img_mod_lin.lora_up.weight",
+    "lora_unet_double_blocks_14_txt_attn_proj.alpha",
+    "lora_unet_double_blocks_14_txt_attn_proj.lora_down.weight",
+    "lora_unet_double_blocks_14_txt_attn_proj.lora_up.weight",
+    "lora_unet_double_blocks_14_txt_attn_qkv.alpha",
+    "lora_unet_double_blocks_14_txt_attn_qkv.lora_down.weight",
+    "lora_unet_double_blocks_14_txt_attn_qkv.lora_up.weight",
+    "lora_unet_double_blocks_14_txt_mlp_0.alpha",
+    "lora_unet_double_blocks_14_txt_mlp_0.lora_down.weight",
+    "lora_unet_double_blocks_14_txt_mlp_0.lora_up.weight",
+    "lora_unet_double_blocks_14_txt_mlp_2.alpha",
+    "lora_unet_double_blocks_14_txt_mlp_2.lora_down.weight",
+    "lora_unet_double_blocks_14_txt_mlp_2.lora_up.weight",
+    "lora_unet_double_blocks_14_txt_mod_lin.alpha",
+    "lora_unet_double_blocks_14_txt_mod_lin.lora_down.weight",
+    "lora_unet_double_blocks_14_txt_mod_lin.lora_up.weight",
+    "lora_unet_double_blocks_15_img_attn_proj.alpha",
+    "lora_unet_double_blocks_15_img_attn_proj.lora_down.weight",
+    "lora_unet_double_blocks_15_img_attn_proj.lora_up.weight",
+    "lora_unet_double_blocks_15_img_attn_qkv.alpha",
+    "lora_unet_double_blocks_15_img_attn_qkv.lora_down.weight",
+    "lora_unet_double_blocks_15_img_attn_qkv.lora_up.weight",
+    "lora_unet_double_blocks_15_img_mlp_0.alpha",
+    "lora_unet_double_blocks_15_img_mlp_0.lora_down.weight",
+    "lora_unet_double_blocks_15_img_mlp_0.lora_up.weight",
+    "lora_unet_double_blocks_15_img_mlp_2.alpha",
+    "lora_unet_double_blocks_15_img_mlp_2.lora_down.weight",
+    "lora_unet_double_blocks_15_img_mlp_2.lora_up.weight",
+    "lora_unet_double_blocks_15_img_mod_lin.alpha",
+    "lora_unet_double_blocks_15_img_mod_lin.lora_down.weight",
+    "lora_unet_double_blocks_15_img_mod_lin.lora_up.weight",
+    "lora_unet_double_blocks_15_txt_attn_proj.alpha",
+    "lora_unet_double_blocks_15_txt_attn_proj.lora_down.weight",
+    "lora_unet_double_blocks_15_txt_attn_proj.lora_up.weight",
+    "lora_unet_double_blocks_15_txt_attn_qkv.alpha",
+    "lora_unet_double_blocks_15_txt_attn_qkv.lora_down.weight",
+    "lora_unet_double_blocks_15_txt_attn_qkv.lora_up.weight",
+    "lora_unet_double_blocks_15_txt_mlp_0.alpha",
+    "lora_unet_double_blocks_15_txt_mlp_0.lora_down.weight",
+    "lora_unet_double_blocks_15_txt_mlp_0.lora_up.weight",
+    "lora_unet_double_blocks_15_txt_mlp_2.alpha",
+    "lora_unet_double_blocks_15_txt_mlp_2.lora_down.weight",
+    "lora_unet_double_blocks_15_txt_mlp_2.lora_up.weight",
+    "lora_unet_double_blocks_15_txt_mod_lin.alpha",
+    "lora_unet_double_blocks_15_txt_mod_lin.lora_down.weight",
+    "lora_unet_double_blocks_15_txt_mod_lin.lora_up.weight",
+    "lora_unet_double_blocks_16_img_attn_proj.alpha",
+    "lora_unet_double_blocks_16_img_attn_proj.lora_down.weight",
+    "lora_unet_double_blocks_16_img_attn_proj.lora_up.weight",
+    "lora_unet_double_blocks_16_img_attn_qkv.alpha",
+    "lora_unet_double_blocks_16_img_attn_qkv.lora_down.weight",
+    "lora_unet_double_blocks_16_img_attn_qkv.lora_up.weight",
+    "lora_unet_double_blocks_16_img_mlp_0.alpha",
+    "lora_unet_double_blocks_16_img_mlp_0.lora_down.weight",
+    "lora_unet_double_blocks_16_img_mlp_0.lora_up.weight",
+    "lora_unet_double_blocks_16_img_mlp_2.alpha",
+    "lora_unet_double_blocks_16_img_mlp_2.lora_down.weight",
+    "lora_unet_double_blocks_16_img_mlp_2.lora_up.weight",
+    "lora_unet_double_blocks_16_img_mod_lin.alpha",
+    "lora_unet_double_blocks_16_img_mod_lin.lora_down.weight",
+    "lora_unet_double_blocks_16_img_mod_lin.lora_up.weight",
+    "lora_unet_double_blocks_16_txt_attn_proj.alpha",
+    "lora_unet_double_blocks_16_txt_attn_proj.lora_down.weight",
+    "lora_unet_double_blocks_16_txt_attn_proj.lora_up.weight",
+    "lora_unet_double_blocks_16_txt_attn_qkv.alpha",
+    "lora_unet_double_blocks_16_txt_attn_qkv.lora_down.weight",
+    "lora_unet_double_blocks_16_txt_attn_qkv.lora_up.weight",
+    "lora_unet_double_blocks_16_txt_mlp_0.alpha",
+    "lora_unet_double_blocks_16_txt_mlp_0.lora_down.weight",
+    "lora_unet_double_blocks_16_txt_mlp_0.lora_up.weight",
+    "lora_unet_double_blocks_16_txt_mlp_2.alpha",
+    "lora_unet_double_blocks_16_txt_mlp_2.lora_down.weight",
+    "lora_unet_double_blocks_16_txt_mlp_2.lora_up.weight",
+    "lora_unet_double_blocks_16_txt_mod_lin.alpha",
+    "lora_unet_double_blocks_16_txt_mod_lin.lora_down.weight",
+    "lora_unet_double_blocks_16_txt_mod_lin.lora_up.weight",
+    "lora_unet_double_blocks_17_img_attn_proj.alpha",
+    "lora_unet_double_blocks_17_img_attn_proj.lora_down.weight",
+    "lora_unet_double_blocks_17_img_attn_proj.lora_up.weight",
+    "lora_unet_double_blocks_17_img_attn_qkv.alpha",
+    "lora_unet_double_blocks_17_img_attn_qkv.lora_down.weight",
+    "lora_unet_double_blocks_17_img_attn_qkv.lora_up.weight",
+    "lora_unet_double_blocks_17_img_mlp_0.alpha",
+    "lora_unet_double_blocks_17_img_mlp_0.lora_down.weight",
+    "lora_unet_double_blocks_17_img_mlp_0.lora_up.weight",
+    "lora_unet_double_blocks_17_img_mlp_2.alpha",
+    "lora_unet_double_blocks_17_img_mlp_2.lora_down.weight",
+    "lora_unet_double_blocks_17_img_mlp_2.lora_up.weight",
+    "lora_unet_double_blocks_17_img_mod_lin.alpha",
+    "lora_unet_double_blocks_17_img_mod_lin.lora_down.weight",
+    "lora_unet_double_blocks_17_img_mod_lin.lora_up.weight",
+    "lora_unet_double_blocks_17_txt_attn_proj.alpha",
+    "lora_unet_double_blocks_17_txt_attn_proj.lora_down.weight",
+    "lora_unet_double_blocks_17_txt_attn_proj.lora_up.weight",
+    "lora_unet_double_blocks_17_txt_attn_qkv.alpha",
+    "lora_unet_double_blocks_17_txt_attn_qkv.lora_down.weight",
+    "lora_unet_double_blocks_17_txt_attn_qkv.lora_up.weight",
+    "lora_unet_double_blocks_17_txt_mlp_0.alpha",
+    "lora_unet_double_blocks_17_txt_mlp_0.lora_down.weight",
+    "lora_unet_double_blocks_17_txt_mlp_0.lora_up.weight",
+    "lora_unet_double_blocks_17_txt_mlp_2.alpha",
+    "lora_unet_double_blocks_17_txt_mlp_2.lora_down.weight",
+    "lora_unet_double_blocks_17_txt_mlp_2.lora_up.weight",
+    "lora_unet_double_blocks_17_txt_mod_lin.alpha",
+    "lora_unet_double_blocks_17_txt_mod_lin.lora_down.weight",
+    "lora_unet_double_blocks_17_txt_mod_lin.lora_up.weight",
+    "lora_unet_double_blocks_18_img_attn_proj.alpha",
+    "lora_unet_double_blocks_18_img_attn_proj.lora_down.weight",
+    "lora_unet_double_blocks_18_img_attn_proj.lora_up.weight",
+    "lora_unet_double_blocks_18_img_attn_qkv.alpha",
+    "lora_unet_double_blocks_18_img_attn_qkv.lora_down.weight",
+    "lora_unet_double_blocks_18_img_attn_qkv.lora_up.weight",
+    "lora_unet_double_blocks_18_img_mlp_0.alpha",
+    "lora_unet_double_blocks_18_img_mlp_0.lora_down.weight",
+    "lora_unet_double_blocks_18_img_mlp_0.lora_up.weight",
+    "lora_unet_double_blocks_18_img_mlp_2.alpha",
+    "lora_unet_double_blocks_18_img_mlp_2.lora_down.weight",
+    "lora_unet_double_blocks_18_img_mlp_2.lora_up.weight",
+    "lora_unet_double_blocks_18_img_mod_lin.alpha",
+    "lora_unet_double_blocks_18_img_mod_lin.lora_down.weight",
+    "lora_unet_double_blocks_18_img_mod_lin.lora_up.weight",
+    "lora_unet_double_blocks_18_txt_attn_proj.alpha",
+    "lora_unet_double_blocks_18_txt_attn_proj.lora_down.weight",
+    "lora_unet_double_blocks_18_txt_attn_proj.lora_up.weight",
+    "lora_unet_double_blocks_18_txt_attn_qkv.alpha",
+    "lora_unet_double_blocks_18_txt_attn_qkv.lora_down.weight",
+    "lora_unet_double_blocks_18_txt_attn_qkv.lora_up.weight",
+    "lora_unet_double_blocks_18_txt_mlp_0.alpha",
+    "lora_unet_double_blocks_18_txt_mlp_0.lora_down.weight",
+    "lora_unet_double_blocks_18_txt_mlp_0.lora_up.weight",
+    "lora_unet_double_blocks_18_txt_mlp_2.alpha",
+    "lora_unet_double_blocks_18_txt_mlp_2.lora_down.weight",
+    "lora_unet_double_blocks_18_txt_mlp_2.lora_up.weight",
+    "lora_unet_double_blocks_18_txt_mod_lin.alpha",
+    "lora_unet_double_blocks_18_txt_mod_lin.lora_down.weight",
+    "lora_unet_double_blocks_18_txt_mod_lin.lora_up.weight",
+    "lora_unet_double_blocks_1_img_attn_proj.alpha",
+    "lora_unet_double_blocks_1_img_attn_proj.lora_down.weight",
+    "lora_unet_double_blocks_1_img_attn_proj.lora_up.weight",
+    "lora_unet_double_blocks_1_img_attn_qkv.alpha",
+    "lora_unet_double_blocks_1_img_attn_qkv.lora_down.weight",
+    "lora_unet_double_blocks_1_img_attn_qkv.lora_up.weight",
+    "lora_unet_double_blocks_1_img_mlp_0.alpha",
+    "lora_unet_double_blocks_1_img_mlp_0.lora_down.weight",
+    "lora_unet_double_blocks_1_img_mlp_0.lora_up.weight",
+    "lora_unet_double_blocks_1_img_mlp_2.alpha",
+    "lora_unet_double_blocks_1_img_mlp_2.lora_down.weight",
+    "lora_unet_double_blocks_1_img_mlp_2.lora_up.weight",
+    "lora_unet_double_blocks_1_img_mod_lin.alpha",
+    "lora_unet_double_blocks_1_img_mod_lin.lora_down.weight",
+    "lora_unet_double_blocks_1_img_mod_lin.lora_up.weight",
+    "lora_unet_double_blocks_1_txt_attn_proj.alpha",
+    "lora_unet_double_blocks_1_txt_attn_proj.lora_down.weight",
+    "lora_unet_double_blocks_1_txt_attn_proj.lora_up.weight",
+    "lora_unet_double_blocks_1_txt_attn_qkv.alpha",
+    "lora_unet_double_blocks_1_txt_attn_qkv.lora_down.weight",
+    "lora_unet_double_blocks_1_txt_attn_qkv.lora_up.weight",
+    "lora_unet_double_blocks_1_txt_mlp_0.alpha",
+    "lora_unet_double_blocks_1_txt_mlp_0.lora_down.weight",
+    "lora_unet_double_blocks_1_txt_mlp_0.lora_up.weight",
+    "lora_unet_double_blocks_1_txt_mlp_2.alpha",
+    "lora_unet_double_blocks_1_txt_mlp_2.lora_down.weight",
+    "lora_unet_double_blocks_1_txt_mlp_2.lora_up.weight",
+    "lora_unet_double_blocks_1_txt_mod_lin.alpha",
+    "lora_unet_double_blocks_1_txt_mod_lin.lora_down.weight",
+    "lora_unet_double_blocks_1_txt_mod_lin.lora_up.weight",
+    "lora_unet_double_blocks_2_img_attn_proj.alpha",
+    "lora_unet_double_blocks_2_img_attn_proj.lora_down.weight",
+    "lora_unet_double_blocks_2_img_attn_proj.lora_up.weight",
+    "lora_unet_double_blocks_2_img_attn_qkv.alpha",
+    "lora_unet_double_blocks_2_img_attn_qkv.lora_down.weight",
+    "lora_unet_double_blocks_2_img_attn_qkv.lora_up.weight",
+    "lora_unet_double_blocks_2_img_mlp_0.alpha",
+    "lora_unet_double_blocks_2_img_mlp_0.lora_down.weight",
+    "lora_unet_double_blocks_2_img_mlp_0.lora_up.weight",
+    "lora_unet_double_blocks_2_img_mlp_2.alpha",
+    "lora_unet_double_blocks_2_img_mlp_2.lora_down.weight",
+    "lora_unet_double_blocks_2_img_mlp_2.lora_up.weight",
+    "lora_unet_double_blocks_2_img_mod_lin.alpha",
+    "lora_unet_double_blocks_2_img_mod_lin.lora_down.weight",
+    "lora_unet_double_blocks_2_img_mod_lin.lora_up.weight",
+    "lora_unet_double_blocks_2_txt_attn_proj.alpha",
+    "lora_unet_double_blocks_2_txt_attn_proj.lora_down.weight",
+    "lora_unet_double_blocks_2_txt_attn_proj.lora_up.weight",
+    "lora_unet_double_blocks_2_txt_attn_qkv.alpha",
+    "lora_unet_double_blocks_2_txt_attn_qkv.lora_down.weight",
+    "lora_unet_double_blocks_2_txt_attn_qkv.lora_up.weight",
+    "lora_unet_double_blocks_2_txt_mlp_0.alpha",
+    "lora_unet_double_blocks_2_txt_mlp_0.lora_down.weight",
+    "lora_unet_double_blocks_2_txt_mlp_0.lora_up.weight",
+    "lora_unet_double_blocks_2_txt_mlp_2.alpha",
+    "lora_unet_double_blocks_2_txt_mlp_2.lora_down.weight",
+    "lora_unet_double_blocks_2_txt_mlp_2.lora_up.weight",
+    "lora_unet_double_blocks_2_txt_mod_lin.alpha",
+    "lora_unet_double_blocks_2_txt_mod_lin.lora_down.weight",
+    "lora_unet_double_blocks_2_txt_mod_lin.lora_up.weight",
+    "lora_unet_double_blocks_3_img_attn_proj.alpha",
+    "lora_unet_double_blocks_3_img_attn_proj.lora_down.weight",
+    "lora_unet_double_blocks_3_img_attn_proj.lora_up.weight",
+    "lora_unet_double_blocks_3_img_attn_qkv.alpha",
+    "lora_unet_double_blocks_3_img_attn_qkv.lora_down.weight",
+    "lora_unet_double_blocks_3_img_attn_qkv.lora_up.weight",
+    "lora_unet_double_blocks_3_img_mlp_0.alpha",
+    "lora_unet_double_blocks_3_img_mlp_0.lora_down.weight",
+    "lora_unet_double_blocks_3_img_mlp_0.lora_up.weight",
+    "lora_unet_double_blocks_3_img_mlp_2.alpha",
+    "lora_unet_double_blocks_3_img_mlp_2.lora_down.weight",
+    "lora_unet_double_blocks_3_img_mlp_2.lora_up.weight",
+    "lora_unet_double_blocks_3_img_mod_lin.alpha",
+    "lora_unet_double_blocks_3_img_mod_lin.lora_down.weight",
+    "lora_unet_double_blocks_3_img_mod_lin.lora_up.weight",
+    "lora_unet_double_blocks_3_txt_attn_proj.alpha",
+    "lora_unet_double_blocks_3_txt_attn_proj.lora_down.weight",
+    "lora_unet_double_blocks_3_txt_attn_proj.lora_up.weight",
+    "lora_unet_double_blocks_3_txt_attn_qkv.alpha",
+    "lora_unet_double_blocks_3_txt_attn_qkv.lora_down.weight",
+    "lora_unet_double_blocks_3_txt_attn_qkv.lora_up.weight",
+    "lora_unet_double_blocks_3_txt_mlp_0.alpha",
+    "lora_unet_double_blocks_3_txt_mlp_0.lora_down.weight",
+    "lora_unet_double_blocks_3_txt_mlp_0.lora_up.weight",
+    "lora_unet_double_blocks_3_txt_mlp_2.alpha",
+    "lora_unet_double_blocks_3_txt_mlp_2.lora_down.weight",
+    "lora_unet_double_blocks_3_txt_mlp_2.lora_up.weight",
+    "lora_unet_double_blocks_3_txt_mod_lin.alpha",
+    "lora_unet_double_blocks_3_txt_mod_lin.lora_down.weight",
+    "lora_unet_double_blocks_3_txt_mod_lin.lora_up.weight",
+    "lora_unet_double_blocks_4_img_attn_proj.alpha",
+    "lora_unet_double_blocks_4_img_attn_proj.lora_down.weight",
+    "lora_unet_double_blocks_4_img_attn_proj.lora_up.weight",
+    "lora_unet_double_blocks_4_img_attn_qkv.alpha",
+    "lora_unet_double_blocks_4_img_attn_qkv.lora_down.weight",
+    "lora_unet_double_blocks_4_img_attn_qkv.lora_up.weight",
+    "lora_unet_double_blocks_4_img_mlp_0.alpha",
+    "lora_unet_double_blocks_4_img_mlp_0.lora_down.weight",
+    "lora_unet_double_blocks_4_img_mlp_0.lora_up.weight",
+    "lora_unet_double_blocks_4_img_mlp_2.alpha",
+    "lora_unet_double_blocks_4_img_mlp_2.lora_down.weight",
+    "lora_unet_double_blocks_4_img_mlp_2.lora_up.weight",
+    "lora_unet_double_blocks_4_img_mod_lin.alpha",
+    "lora_unet_double_blocks_4_img_mod_lin.lora_down.weight",
+    "lora_unet_double_blocks_4_img_mod_lin.lora_up.weight",
+    "lora_unet_double_blocks_4_txt_attn_proj.alpha",
+    "lora_unet_double_blocks_4_txt_attn_proj.lora_down.weight",
+    "lora_unet_double_blocks_4_txt_attn_proj.lora_up.weight",
+    "lora_unet_double_blocks_4_txt_attn_qkv.alpha",
+    "lora_unet_double_blocks_4_txt_attn_qkv.lora_down.weight",
+    "lora_unet_double_blocks_4_txt_attn_qkv.lora_up.weight",
+    "lora_unet_double_blocks_4_txt_mlp_0.alpha",
+    "lora_unet_double_blocks_4_txt_mlp_0.lora_down.weight",
+    "lora_unet_double_blocks_4_txt_mlp_0.lora_up.weight",
+    "lora_unet_double_blocks_4_txt_mlp_2.alpha",
+    "lora_unet_double_blocks_4_txt_mlp_2.lora_down.weight",
+    "lora_unet_double_blocks_4_txt_mlp_2.lora_up.weight",
+    "lora_unet_double_blocks_4_txt_mod_lin.alpha",
+    "lora_unet_double_blocks_4_txt_mod_lin.lora_down.weight",
+    "lora_unet_double_blocks_4_txt_mod_lin.lora_up.weight",
+    "lora_unet_double_blocks_5_img_attn_proj.alpha",
+    "lora_unet_double_blocks_5_img_attn_proj.lora_down.weight",
+    "lora_unet_double_blocks_5_img_attn_proj.lora_up.weight",
+    "lora_unet_double_blocks_5_img_attn_qkv.alpha",
+    "lora_unet_double_blocks_5_img_attn_qkv.lora_down.weight",
+    "lora_unet_double_blocks_5_img_attn_qkv.lora_up.weight",
+    "lora_unet_double_blocks_5_img_mlp_0.alpha",
+    "lora_unet_double_blocks_5_img_mlp_0.lora_down.weight",
+    "lora_unet_double_blocks_5_img_mlp_0.lora_up.weight",
+    "lora_unet_double_blocks_5_img_mlp_2.alpha",
+    "lora_unet_double_blocks_5_img_mlp_2.lora_down.weight",
+    "lora_unet_double_blocks_5_img_mlp_2.lora_up.weight",
+    "lora_unet_double_blocks_5_img_mod_lin.alpha",
+    "lora_unet_double_blocks_5_img_mod_lin.lora_down.weight",
+    "lora_unet_double_blocks_5_img_mod_lin.lora_up.weight",
+    "lora_unet_double_blocks_5_txt_attn_proj.alpha",
+    "lora_unet_double_blocks_5_txt_attn_proj.lora_down.weight",
+    "lora_unet_double_blocks_5_txt_attn_proj.lora_up.weight",
+    "lora_unet_double_blocks_5_txt_attn_qkv.alpha",
+    "lora_unet_double_blocks_5_txt_attn_qkv.lora_down.weight",
+    "lora_unet_double_blocks_5_txt_attn_qkv.lora_up.weight",
+    "lora_unet_double_blocks_5_txt_mlp_0.alpha",
+    "lora_unet_double_blocks_5_txt_mlp_0.lora_down.weight",
+    "lora_unet_double_blocks_5_txt_mlp_0.lora_up.weight",
+    "lora_unet_double_blocks_5_txt_mlp_2.alpha",
+    "lora_unet_double_blocks_5_txt_mlp_2.lora_down.weight",
+    "lora_unet_double_blocks_5_txt_mlp_2.lora_up.weight",
+    "lora_unet_double_blocks_5_txt_mod_lin.alpha",
+    "lora_unet_double_blocks_5_txt_mod_lin.lora_down.weight",
+    "lora_unet_double_blocks_5_txt_mod_lin.lora_up.weight",
+    "lora_unet_double_blocks_6_img_attn_proj.alpha",
+    "lora_unet_double_blocks_6_img_attn_proj.lora_down.weight",
+    "lora_unet_double_blocks_6_img_attn_proj.lora_up.weight",
+    "lora_unet_double_blocks_6_img_attn_qkv.alpha",
+    "lora_unet_double_blocks_6_img_attn_qkv.lora_down.weight",
+    "lora_unet_double_blocks_6_img_attn_qkv.lora_up.weight",
+    "lora_unet_double_blocks_6_img_mlp_0.alpha",
+    "lora_unet_double_blocks_6_img_mlp_0.lora_down.weight",
+    "lora_unet_double_blocks_6_img_mlp_0.lora_up.weight",
+    "lora_unet_double_blocks_6_img_mlp_2.alpha",
+    "lora_unet_double_blocks_6_img_mlp_2.lora_down.weight",
+    "lora_unet_double_blocks_6_img_mlp_2.lora_up.weight",
+    "lora_unet_double_blocks_6_img_mod_lin.alpha",
+    "lora_unet_double_blocks_6_img_mod_lin.lora_down.weight",
+    "lora_unet_double_blocks_6_img_mod_lin.lora_up.weight",
+    "lora_unet_double_blocks_6_txt_attn_proj.alpha",
+    "lora_unet_double_blocks_6_txt_attn_proj.lora_down.weight",
+    "lora_unet_double_blocks_6_txt_attn_proj.lora_up.weight",
+    "lora_unet_double_blocks_6_txt_attn_qkv.alpha",
+    "lora_unet_double_blocks_6_txt_attn_qkv.lora_down.weight",
+    "lora_unet_double_blocks_6_txt_attn_qkv.lora_up.weight",
+    "lora_unet_double_blocks_6_txt_mlp_0.alpha",
+    "lora_unet_double_blocks_6_txt_mlp_0.lora_down.weight",
+    "lora_unet_double_blocks_6_txt_mlp_0.lora_up.weight",
+    "lora_unet_double_blocks_6_txt_mlp_2.alpha",
+    "lora_unet_double_blocks_6_txt_mlp_2.lora_down.weight",
+    "lora_unet_double_blocks_6_txt_mlp_2.lora_up.weight",
+    "lora_unet_double_blocks_6_txt_mod_lin.alpha",
+    "lora_unet_double_blocks_6_txt_mod_lin.lora_down.weight",
+    "lora_unet_double_blocks_6_txt_mod_lin.lora_up.weight",
+    "lora_unet_double_blocks_7_img_attn_proj.alpha",
+    "lora_unet_double_blocks_7_img_attn_proj.lora_down.weight",
+    "lora_unet_double_blocks_7_img_attn_proj.lora_up.weight",
+    "lora_unet_double_blocks_7_img_attn_qkv.alpha",
+    "lora_unet_double_blocks_7_img_attn_qkv.lora_down.weight",
+    "lora_unet_double_blocks_7_img_attn_qkv.lora_up.weight",
+    "lora_unet_double_blocks_7_img_mlp_0.alpha",
+    "lora_unet_double_blocks_7_img_mlp_0.lora_down.weight",
+    "lora_unet_double_blocks_7_img_mlp_0.lora_up.weight",
+    "lora_unet_double_blocks_7_img_mlp_2.alpha",
+    "lora_unet_double_blocks_7_img_mlp_2.lora_down.weight",
+    "lora_unet_double_blocks_7_img_mlp_2.lora_up.weight",
+    "lora_unet_double_blocks_7_img_mod_lin.alpha",
+    "lora_unet_double_blocks_7_img_mod_lin.lora_down.weight",
+    "lora_unet_double_blocks_7_img_mod_lin.lora_up.weight",
+    "lora_unet_double_blocks_7_txt_attn_proj.alpha",
+    "lora_unet_double_blocks_7_txt_attn_proj.lora_down.weight",
+    "lora_unet_double_blocks_7_txt_attn_proj.lora_up.weight",
+    "lora_unet_double_blocks_7_txt_attn_qkv.alpha",
+    "lora_unet_double_blocks_7_txt_attn_qkv.lora_down.weight",
+    "lora_unet_double_blocks_7_txt_attn_qkv.lora_up.weight",
+    "lora_unet_double_blocks_7_txt_mlp_0.alpha",
+    "lora_unet_double_blocks_7_txt_mlp_0.lora_down.weight",
+    "lora_unet_double_blocks_7_txt_mlp_0.lora_up.weight",
+    "lora_unet_double_blocks_7_txt_mlp_2.alpha",
+    "lora_unet_double_blocks_7_txt_mlp_2.lora_down.weight",
+    "lora_unet_double_blocks_7_txt_mlp_2.lora_up.weight",
+    "lora_unet_double_blocks_7_txt_mod_lin.alpha",
+    "lora_unet_double_blocks_7_txt_mod_lin.lora_down.weight",
+    "lora_unet_double_blocks_7_txt_mod_lin.lora_up.weight",
+    "lora_unet_double_blocks_8_img_attn_proj.alpha",
+    "lora_unet_double_blocks_8_img_attn_proj.lora_down.weight",
+    "lora_unet_double_blocks_8_img_attn_proj.lora_up.weight",
+    "lora_unet_double_blocks_8_img_attn_qkv.alpha",
+    "lora_unet_double_blocks_8_img_attn_qkv.lora_down.weight",
+    "lora_unet_double_blocks_8_img_attn_qkv.lora_up.weight",
+    "lora_unet_double_blocks_8_img_mlp_0.alpha",
+    "lora_unet_double_blocks_8_img_mlp_0.lora_down.weight",
+    "lora_unet_double_blocks_8_img_mlp_0.lora_up.weight",
+    "lora_unet_double_blocks_8_img_mlp_2.alpha",
+    "lora_unet_double_blocks_8_img_mlp_2.lora_down.weight",
+    "lora_unet_double_blocks_8_img_mlp_2.lora_up.weight",
+    "lora_unet_double_blocks_8_img_mod_lin.alpha",
+    "lora_unet_double_blocks_8_img_mod_lin.lora_down.weight",
+    "lora_unet_double_blocks_8_img_mod_lin.lora_up.weight",
+    "lora_unet_double_blocks_8_txt_attn_proj.alpha",
+    "lora_unet_double_blocks_8_txt_attn_proj.lora_down.weight",
+    "lora_unet_double_blocks_8_txt_attn_proj.lora_up.weight",
+    "lora_unet_double_blocks_8_txt_attn_qkv.alpha",
+    "lora_unet_double_blocks_8_txt_attn_qkv.lora_down.weight",
+    "lora_unet_double_blocks_8_txt_attn_qkv.lora_up.weight",
+    "lora_unet_double_blocks_8_txt_mlp_0.alpha",
+    "lora_unet_double_blocks_8_txt_mlp_0.lora_down.weight",
+    "lora_unet_double_blocks_8_txt_mlp_0.lora_up.weight",
+    "lora_unet_double_blocks_8_txt_mlp_2.alpha",
+    "lora_unet_double_blocks_8_txt_mlp_2.lora_down.weight",
+    "lora_unet_double_blocks_8_txt_mlp_2.lora_up.weight",
+    "lora_unet_double_blocks_8_txt_mod_lin.alpha",
+    "lora_unet_double_blocks_8_txt_mod_lin.lora_down.weight",
+    "lora_unet_double_blocks_8_txt_mod_lin.lora_up.weight",
+    "lora_unet_double_blocks_9_img_attn_proj.alpha",
+    "lora_unet_double_blocks_9_img_attn_proj.lora_down.weight",
+    "lora_unet_double_blocks_9_img_attn_proj.lora_up.weight",
+    "lora_unet_double_blocks_9_img_attn_qkv.alpha",
+    "lora_unet_double_blocks_9_img_attn_qkv.lora_down.weight",
+    "lora_unet_double_blocks_9_img_attn_qkv.lora_up.weight",
+    "lora_unet_double_blocks_9_img_mlp_0.alpha",
+    "lora_unet_double_blocks_9_img_mlp_0.lora_down.weight",
+    "lora_unet_double_blocks_9_img_mlp_0.lora_up.weight",
+    "lora_unet_double_blocks_9_img_mlp_2.alpha",
+    "lora_unet_double_blocks_9_img_mlp_2.lora_down.weight",
+    "lora_unet_double_blocks_9_img_mlp_2.lora_up.weight",
+    "lora_unet_double_blocks_9_img_mod_lin.alpha",
+    "lora_unet_double_blocks_9_img_mod_lin.lora_down.weight",
+    "lora_unet_double_blocks_9_img_mod_lin.lora_up.weight",
+    "lora_unet_double_blocks_9_txt_attn_proj.alpha",
+    "lora_unet_double_blocks_9_txt_attn_proj.lora_down.weight",
+    "lora_unet_double_blocks_9_txt_attn_proj.lora_up.weight",
+    "lora_unet_double_blocks_9_txt_attn_qkv.alpha",
+    "lora_unet_double_blocks_9_txt_attn_qkv.lora_down.weight",
+    "lora_unet_double_blocks_9_txt_attn_qkv.lora_up.weight",
+    "lora_unet_double_blocks_9_txt_mlp_0.alpha",
+    "lora_unet_double_blocks_9_txt_mlp_0.lora_down.weight",
+    "lora_unet_double_blocks_9_txt_mlp_0.lora_up.weight",
+    "lora_unet_double_blocks_9_txt_mlp_2.alpha",
+    "lora_unet_double_blocks_9_txt_mlp_2.lora_down.weight",
+    "lora_unet_double_blocks_9_txt_mlp_2.lora_up.weight",
+    "lora_unet_double_blocks_9_txt_mod_lin.alpha",
+    "lora_unet_double_blocks_9_txt_mod_lin.lora_down.weight",
+    "lora_unet_double_blocks_9_txt_mod_lin.lora_up.weight",
+    "lora_unet_single_blocks_0_linear1.alpha",
+    "lora_unet_single_blocks_0_linear1.lora_down.weight",
+    "lora_unet_single_blocks_0_linear1.lora_up.weight",
+    "lora_unet_single_blocks_0_linear2.alpha",
+    "lora_unet_single_blocks_0_linear2.lora_down.weight",
+    "lora_unet_single_blocks_0_linear2.lora_up.weight",
+    "lora_unet_single_blocks_0_modulation_lin.alpha",
+    "lora_unet_single_blocks_0_modulation_lin.lora_down.weight",
+    "lora_unet_single_blocks_0_modulation_lin.lora_up.weight",
+    "lora_unet_single_blocks_10_linear1.alpha",
+    "lora_unet_single_blocks_10_linear1.lora_down.weight",
+    "lora_unet_single_blocks_10_linear1.lora_up.weight",
+    "lora_unet_single_blocks_10_linear2.alpha",
+    "lora_unet_single_blocks_10_linear2.lora_down.weight",
+    "lora_unet_single_blocks_10_linear2.lora_up.weight",
+    "lora_unet_single_blocks_10_modulation_lin.alpha",
+    "lora_unet_single_blocks_10_modulation_lin.lora_down.weight",
+    "lora_unet_single_blocks_10_modulation_lin.lora_up.weight",
+    "lora_unet_single_blocks_11_linear1.alpha",
+    "lora_unet_single_blocks_11_linear1.lora_down.weight",
+    "lora_unet_single_blocks_11_linear1.lora_up.weight",
+    "lora_unet_single_blocks_11_linear2.alpha",
+    "lora_unet_single_blocks_11_linear2.lora_down.weight",
+    "lora_unet_single_blocks_11_linear2.lora_up.weight",
+    "lora_unet_single_blocks_11_modulation_lin.alpha",
+    "lora_unet_single_blocks_11_modulation_lin.lora_down.weight",
+    "lora_unet_single_blocks_11_modulation_lin.lora_up.weight",
+    "lora_unet_single_blocks_12_linear1.alpha",
+    "lora_unet_single_blocks_12_linear1.lora_down.weight",
+    "lora_unet_single_blocks_12_linear1.lora_up.weight",
+    "lora_unet_single_blocks_12_linear2.alpha",
+    "lora_unet_single_blocks_12_linear2.lora_down.weight",
+    "lora_unet_single_blocks_12_linear2.lora_up.weight",
+    "lora_unet_single_blocks_12_modulation_lin.alpha",
+    "lora_unet_single_blocks_12_modulation_lin.lora_down.weight",
+    "lora_unet_single_blocks_12_modulation_lin.lora_up.weight",
+    "lora_unet_single_blocks_13_linear1.alpha",
+    "lora_unet_single_blocks_13_linear1.lora_down.weight",
+    "lora_unet_single_blocks_13_linear1.lora_up.weight",
+    "lora_unet_single_blocks_13_linear2.alpha",
+    "lora_unet_single_blocks_13_linear2.lora_down.weight",
+    "lora_unet_single_blocks_13_linear2.lora_up.weight",
+    "lora_unet_single_blocks_13_modulation_lin.alpha",
+    "lora_unet_single_blocks_13_modulation_lin.lora_down.weight",
+    "lora_unet_single_blocks_13_modulation_lin.lora_up.weight",
+    "lora_unet_single_blocks_14_linear1.alpha",
+    "lora_unet_single_blocks_14_linear1.lora_down.weight",
+    "lora_unet_single_blocks_14_linear1.lora_up.weight",
+    "lora_unet_single_blocks_14_linear2.alpha",
+    "lora_unet_single_blocks_14_linear2.lora_down.weight",
+    "lora_unet_single_blocks_14_linear2.lora_up.weight",
+    "lora_unet_single_blocks_14_modulation_lin.alpha",
+    "lora_unet_single_blocks_14_modulation_lin.lora_down.weight",
+    "lora_unet_single_blocks_14_modulation_lin.lora_up.weight",
+    "lora_unet_single_blocks_15_linear1.alpha",
+    "lora_unet_single_blocks_15_linear1.lora_down.weight",
+    "lora_unet_single_blocks_15_linear1.lora_up.weight",
+    "lora_unet_single_blocks_15_linear2.alpha",
+    "lora_unet_single_blocks_15_linear2.lora_down.weight",
+    "lora_unet_single_blocks_15_linear2.lora_up.weight",
+    "lora_unet_single_blocks_15_modulation_lin.alpha",
+    "lora_unet_single_blocks_15_modulation_lin.lora_down.weight",
+    "lora_unet_single_blocks_15_modulation_lin.lora_up.weight",
+    "lora_unet_single_blocks_16_linear1.alpha",
+    "lora_unet_single_blocks_16_linear1.lora_down.weight",
+    "lora_unet_single_blocks_16_linear1.lora_up.weight",
+    "lora_unet_single_blocks_16_linear2.alpha",
+    "lora_unet_single_blocks_16_linear2.lora_down.weight",
+    "lora_unet_single_blocks_16_linear2.lora_up.weight",
+    "lora_unet_single_blocks_16_modulation_lin.alpha",
+    "lora_unet_single_blocks_16_modulation_lin.lora_down.weight",
+    "lora_unet_single_blocks_16_modulation_lin.lora_up.weight",
+    "lora_unet_single_blocks_17_linear1.alpha",
+    "lora_unet_single_blocks_17_linear1.lora_down.weight",
+    "lora_unet_single_blocks_17_linear1.lora_up.weight",
+    "lora_unet_single_blocks_17_linear2.alpha",
+    "lora_unet_single_blocks_17_linear2.lora_down.weight",
+    "lora_unet_single_blocks_17_linear2.lora_up.weight",
+    "lora_unet_single_blocks_17_modulation_lin.alpha",
+    "lora_unet_single_blocks_17_modulation_lin.lora_down.weight",
+    "lora_unet_single_blocks_17_modulation_lin.lora_up.weight",
+    "lora_unet_single_blocks_18_linear1.alpha",
+    "lora_unet_single_blocks_18_linear1.lora_down.weight",
+    "lora_unet_single_blocks_18_linear1.lora_up.weight",
+    "lora_unet_single_blocks_18_linear2.alpha",
+    "lora_unet_single_blocks_18_linear2.lora_down.weight",
+    "lora_unet_single_blocks_18_linear2.lora_up.weight",
+    "lora_unet_single_blocks_18_modulation_lin.alpha",
+    "lora_unet_single_blocks_18_modulation_lin.lora_down.weight",
+    "lora_unet_single_blocks_18_modulation_lin.lora_up.weight",
+    "lora_unet_single_blocks_19_linear1.alpha",
+    "lora_unet_single_blocks_19_linear1.lora_down.weight",
+    "lora_unet_single_blocks_19_linear1.lora_up.weight",
+    "lora_unet_single_blocks_19_linear2.alpha",
+    "lora_unet_single_blocks_19_linear2.lora_down.weight",
+    "lora_unet_single_blocks_19_linear2.lora_up.weight",
+    "lora_unet_single_blocks_19_modulation_lin.alpha",
+    "lora_unet_single_blocks_19_modulation_lin.lora_down.weight",
+    "lora_unet_single_blocks_19_modulation_lin.lora_up.weight",
+    "lora_unet_single_blocks_1_linear1.alpha",
+    "lora_unet_single_blocks_1_linear1.lora_down.weight",
+    "lora_unet_single_blocks_1_linear1.lora_up.weight",
+    "lora_unet_single_blocks_1_linear2.alpha",
+    "lora_unet_single_blocks_1_linear2.lora_down.weight",
+    "lora_unet_single_blocks_1_linear2.lora_up.weight",
+    "lora_unet_single_blocks_1_modulation_lin.alpha",
+    "lora_unet_single_blocks_1_modulation_lin.lora_down.weight",
+    "lora_unet_single_blocks_1_modulation_lin.lora_up.weight",
+    "lora_unet_single_blocks_20_linear1.alpha",
+    "lora_unet_single_blocks_20_linear1.lora_down.weight",
+    "lora_unet_single_blocks_20_linear1.lora_up.weight",
+    "lora_unet_single_blocks_20_linear2.alpha",
+    "lora_unet_single_blocks_20_linear2.lora_down.weight",
+    "lora_unet_single_blocks_20_linear2.lora_up.weight",
+    "lora_unet_single_blocks_20_modulation_lin.alpha",
+    "lora_unet_single_blocks_20_modulation_lin.lora_down.weight",
+    "lora_unet_single_blocks_20_modulation_lin.lora_up.weight",
+    "lora_unet_single_blocks_21_linear1.alpha",
+    "lora_unet_single_blocks_21_linear1.lora_down.weight",
+    "lora_unet_single_blocks_21_linear1.lora_up.weight",
+    "lora_unet_single_blocks_21_linear2.alpha",
+    "lora_unet_single_blocks_21_linear2.lora_down.weight",
+    "lora_unet_single_blocks_21_linear2.lora_up.weight",
+    "lora_unet_single_blocks_21_modulation_lin.alpha",
+    "lora_unet_single_blocks_21_modulation_lin.lora_down.weight",
+    "lora_unet_single_blocks_21_modulation_lin.lora_up.weight",
+    "lora_unet_single_blocks_22_linear1.alpha",
+    "lora_unet_single_blocks_22_linear1.lora_down.weight",
+    "lora_unet_single_blocks_22_linear1.lora_up.weight",
+    "lora_unet_single_blocks_22_linear2.alpha",
+    "lora_unet_single_blocks_22_linear2.lora_down.weight",
+    "lora_unet_single_blocks_22_linear2.lora_up.weight",
+    "lora_unet_single_blocks_22_modulation_lin.alpha",
+    "lora_unet_single_blocks_22_modulation_lin.lora_down.weight",
+    "lora_unet_single_blocks_22_modulation_lin.lora_up.weight",
+    "lora_unet_single_blocks_23_linear1.alpha",
+    "lora_unet_single_blocks_23_linear1.lora_down.weight",
+    "lora_unet_single_blocks_23_linear1.lora_up.weight",
+    "lora_unet_single_blocks_23_linear2.alpha",
+    "lora_unet_single_blocks_23_linear2.lora_down.weight",
+    "lora_unet_single_blocks_23_linear2.lora_up.weight",
+    "lora_unet_single_blocks_23_modulation_lin.alpha",
+    "lora_unet_single_blocks_23_modulation_lin.lora_down.weight",
+    "lora_unet_single_blocks_23_modulation_lin.lora_up.weight",
+    "lora_unet_single_blocks_24_linear1.alpha",
+    "lora_unet_single_blocks_24_linear1.lora_down.weight",
+    "lora_unet_single_blocks_24_linear1.lora_up.weight",
+    "lora_unet_single_blocks_24_linear2.alpha",
+    "lora_unet_single_blocks_24_linear2.lora_down.weight",
+    "lora_unet_single_blocks_24_linear2.lora_up.weight",
+    "lora_unet_single_blocks_24_modulation_lin.alpha",
+    "lora_unet_single_blocks_24_modulation_lin.lora_down.weight",
+    "lora_unet_single_blocks_24_modulation_lin.lora_up.weight",
+    "lora_unet_single_blocks_25_linear1.alpha",
+    "lora_unet_single_blocks_25_linear1.lora_down.weight",
+    "lora_unet_single_blocks_25_linear1.lora_up.weight",
+    "lora_unet_single_blocks_25_linear2.alpha",
+    "lora_unet_single_blocks_25_linear2.lora_down.weight",
+    "lora_unet_single_blocks_25_linear2.lora_up.weight",
+    "lora_unet_single_blocks_25_modulation_lin.alpha",
+    "lora_unet_single_blocks_25_modulation_lin.lora_down.weight",
+    "lora_unet_single_blocks_25_modulation_lin.lora_up.weight",
+    "lora_unet_single_blocks_26_linear1.alpha",
+    "lora_unet_single_blocks_26_linear1.lora_down.weight",
+    "lora_unet_single_blocks_26_linear1.lora_up.weight",
+    "lora_unet_single_blocks_26_linear2.alpha",
+    "lora_unet_single_blocks_26_linear2.lora_down.weight",
+    "lora_unet_single_blocks_26_linear2.lora_up.weight",
+    "lora_unet_single_blocks_26_modulation_lin.alpha",
+    "lora_unet_single_blocks_26_modulation_lin.lora_down.weight",
+    "lora_unet_single_blocks_26_modulation_lin.lora_up.weight",
+    "lora_unet_single_blocks_27_linear1.alpha",
+    "lora_unet_single_blocks_27_linear1.lora_down.weight",
+    "lora_unet_single_blocks_27_linear1.lora_up.weight",
+    "lora_unet_single_blocks_27_linear2.alpha",
+    "lora_unet_single_blocks_27_linear2.lora_down.weight",
+    "lora_unet_single_blocks_27_linear2.lora_up.weight",
+    "lora_unet_single_blocks_27_modulation_lin.alpha",
+    "lora_unet_single_blocks_27_modulation_lin.lora_down.weight",
+    "lora_unet_single_blocks_27_modulation_lin.lora_up.weight",
+    "lora_unet_single_blocks_28_linear1.alpha",
+    "lora_unet_single_blocks_28_linear1.lora_down.weight",
+    "lora_unet_single_blocks_28_linear1.lora_up.weight",
+    "lora_unet_single_blocks_28_linear2.alpha",
+    "lora_unet_single_blocks_28_linear2.lora_down.weight",
+    "lora_unet_single_blocks_28_linear2.lora_up.weight",
+    "lora_unet_single_blocks_28_modulation_lin.alpha",
+    "lora_unet_single_blocks_28_modulation_lin.lora_down.weight",
+    "lora_unet_single_blocks_28_modulation_lin.lora_up.weight",
+    "lora_unet_single_blocks_29_linear1.alpha",
+    "lora_unet_single_blocks_29_linear1.lora_down.weight",
+    "lora_unet_single_blocks_29_linear1.lora_up.weight",
+    "lora_unet_single_blocks_29_linear2.alpha",
+    "lora_unet_single_blocks_29_linear2.lora_down.weight",
+    "lora_unet_single_blocks_29_linear2.lora_up.weight",
+    "lora_unet_single_blocks_29_modulation_lin.alpha",
+    "lora_unet_single_blocks_29_modulation_lin.lora_down.weight",
+    "lora_unet_single_blocks_29_modulation_lin.lora_up.weight",
+    "lora_unet_single_blocks_2_linear1.alpha",
+    "lora_unet_single_blocks_2_linear1.lora_down.weight",
+    "lora_unet_single_blocks_2_linear1.lora_up.weight",
+    "lora_unet_single_blocks_2_linear2.alpha",
+    "lora_unet_single_blocks_2_linear2.lora_down.weight",
+    "lora_unet_single_blocks_2_linear2.lora_up.weight",
+    "lora_unet_single_blocks_2_modulation_lin.alpha",
+    "lora_unet_single_blocks_2_modulation_lin.lora_down.weight",
+    "lora_unet_single_blocks_2_modulation_lin.lora_up.weight",
+    "lora_unet_single_blocks_30_linear1.alpha",
+    "lora_unet_single_blocks_30_linear1.lora_down.weight",
+    "lora_unet_single_blocks_30_linear1.lora_up.weight",
+    "lora_unet_single_blocks_30_linear2.alpha",
+    "lora_unet_single_blocks_30_linear2.lora_down.weight",
+    "lora_unet_single_blocks_30_linear2.lora_up.weight",
+    "lora_unet_single_blocks_30_modulation_lin.alpha",
+    "lora_unet_single_blocks_30_modulation_lin.lora_down.weight",
+    "lora_unet_single_blocks_30_modulation_lin.lora_up.weight",
+    "lora_unet_single_blocks_31_linear1.alpha",
+    "lora_unet_single_blocks_31_linear1.lora_down.weight",
+    "lora_unet_single_blocks_31_linear1.lora_up.weight",
+    "lora_unet_single_blocks_31_linear2.alpha",
+    "lora_unet_single_blocks_31_linear2.lora_down.weight",
+    "lora_unet_single_blocks_31_linear2.lora_up.weight",
+    "lora_unet_single_blocks_31_modulation_lin.alpha",
+    "lora_unet_single_blocks_31_modulation_lin.lora_down.weight",
+    "lora_unet_single_blocks_31_modulation_lin.lora_up.weight",
+    "lora_unet_single_blocks_32_linear1.alpha",
+    "lora_unet_single_blocks_32_linear1.lora_down.weight",
+    "lora_unet_single_blocks_32_linear1.lora_up.weight",
+    "lora_unet_single_blocks_32_linear2.alpha",
+    "lora_unet_single_blocks_32_linear2.lora_down.weight",
+    "lora_unet_single_blocks_32_linear2.lora_up.weight",
+    "lora_unet_single_blocks_32_modulation_lin.alpha",
+    "lora_unet_single_blocks_32_modulation_lin.lora_down.weight",
+    "lora_unet_single_blocks_32_modulation_lin.lora_up.weight",
+    "lora_unet_single_blocks_33_linear1.alpha",
+    "lora_unet_single_blocks_33_linear1.lora_down.weight",
+    "lora_unet_single_blocks_33_linear1.lora_up.weight",
+    "lora_unet_single_blocks_33_linear2.alpha",
+    "lora_unet_single_blocks_33_linear2.lora_down.weight",
+    "lora_unet_single_blocks_33_linear2.lora_up.weight",
+    "lora_unet_single_blocks_33_modulation_lin.alpha",
+    "lora_unet_single_blocks_33_modulation_lin.lora_down.weight",
+    "lora_unet_single_blocks_33_modulation_lin.lora_up.weight",
+    "lora_unet_single_blocks_34_linear1.alpha",
+    "lora_unet_single_blocks_34_linear1.lora_down.weight",
+    "lora_unet_single_blocks_34_linear1.lora_up.weight",
+    "lora_unet_single_blocks_34_linear2.alpha",
+    "lora_unet_single_blocks_34_linear2.lora_down.weight",
+    "lora_unet_single_blocks_34_linear2.lora_up.weight",
+    "lora_unet_single_blocks_34_modulation_lin.alpha",
+    "lora_unet_single_blocks_34_modulation_lin.lora_down.weight",
+    "lora_unet_single_blocks_34_modulation_lin.lora_up.weight",
+    "lora_unet_single_blocks_35_linear1.alpha",
+    "lora_unet_single_blocks_35_linear1.lora_down.weight",
+    "lora_unet_single_blocks_35_linear1.lora_up.weight",
+    "lora_unet_single_blocks_35_linear2.alpha",
+    "lora_unet_single_blocks_35_linear2.lora_down.weight",
+    "lora_unet_single_blocks_35_linear2.lora_up.weight",
+    "lora_unet_single_blocks_35_modulation_lin.alpha",
+    "lora_unet_single_blocks_35_modulation_lin.lora_down.weight",
+    "lora_unet_single_blocks_35_modulation_lin.lora_up.weight",
+    "lora_unet_single_blocks_36_linear1.alpha",
+    "lora_unet_single_blocks_36_linear1.lora_down.weight",
+    "lora_unet_single_blocks_36_linear1.lora_up.weight",
+    "lora_unet_single_blocks_36_linear2.alpha",
+    "lora_unet_single_blocks_36_linear2.lora_down.weight",
+    "lora_unet_single_blocks_36_linear2.lora_up.weight",
+    "lora_unet_single_blocks_36_modulation_lin.alpha",
+    "lora_unet_single_blocks_36_modulation_lin.lora_down.weight",
+    "lora_unet_single_blocks_36_modulation_lin.lora_up.weight",
+    "lora_unet_single_blocks_37_linear1.alpha",
+    "lora_unet_single_blocks_37_linear1.lora_down.weight",
+    "lora_unet_single_blocks_37_linear1.lora_up.weight",
+    "lora_unet_single_blocks_37_linear2.alpha",
+    "lora_unet_single_blocks_37_linear2.lora_down.weight",
+    "lora_unet_single_blocks_37_linear2.lora_up.weight",
+    "lora_unet_single_blocks_37_modulation_lin.alpha",
+    "lora_unet_single_blocks_37_modulation_lin.lora_down.weight",
+    "lora_unet_single_blocks_37_modulation_lin.lora_up.weight",
+    "lora_unet_single_blocks_3_linear1.alpha",
+    "lora_unet_single_blocks_3_linear1.lora_down.weight",
+    "lora_unet_single_blocks_3_linear1.lora_up.weight",
+    "lora_unet_single_blocks_3_linear2.alpha",
+    "lora_unet_single_blocks_3_linear2.lora_down.weight",
+    "lora_unet_single_blocks_3_linear2.lora_up.weight",
+    "lora_unet_single_blocks_3_modulation_lin.alpha",
+    "lora_unet_single_blocks_3_modulation_lin.lora_down.weight",
+    "lora_unet_single_blocks_3_modulation_lin.lora_up.weight",
+    "lora_unet_single_blocks_4_linear1.alpha",
+    "lora_unet_single_blocks_4_linear1.lora_down.weight",
+    "lora_unet_single_blocks_4_linear1.lora_up.weight",
+    "lora_unet_single_blocks_4_linear2.alpha",
+    "lora_unet_single_blocks_4_linear2.lora_down.weight",
+    "lora_unet_single_blocks_4_linear2.lora_up.weight",
+    "lora_unet_single_blocks_4_modulation_lin.alpha",
+    "lora_unet_single_blocks_4_modulation_lin.lora_down.weight",
+    "lora_unet_single_blocks_4_modulation_lin.lora_up.weight",
+    "lora_unet_single_blocks_5_linear1.alpha",
+    "lora_unet_single_blocks_5_linear1.lora_down.weight",
+    "lora_unet_single_blocks_5_linear1.lora_up.weight",
+    "lora_unet_single_blocks_5_linear2.alpha",
+    "lora_unet_single_blocks_5_linear2.lora_down.weight",
+    "lora_unet_single_blocks_5_linear2.lora_up.weight",
+    "lora_unet_single_blocks_5_modulation_lin.alpha",
+    "lora_unet_single_blocks_5_modulation_lin.lora_down.weight",
+    "lora_unet_single_blocks_5_modulation_lin.lora_up.weight",
+    "lora_unet_single_blocks_6_linear1.alpha",
+    "lora_unet_single_blocks_6_linear1.lora_down.weight",
+    "lora_unet_single_blocks_6_linear1.lora_up.weight",
+    "lora_unet_single_blocks_6_linear2.alpha",
+    "lora_unet_single_blocks_6_linear2.lora_down.weight",
+    "lora_unet_single_blocks_6_linear2.lora_up.weight",
+    "lora_unet_single_blocks_6_modulation_lin.alpha",
+    "lora_unet_single_blocks_6_modulation_lin.lora_down.weight",
+    "lora_unet_single_blocks_6_modulation_lin.lora_up.weight",
+    "lora_unet_single_blocks_7_linear1.alpha",
+    "lora_unet_single_blocks_7_linear1.lora_down.weight",
+    "lora_unet_single_blocks_7_linear1.lora_up.weight",
+    "lora_unet_single_blocks_7_linear2.alpha",
+    "lora_unet_single_blocks_7_linear2.lora_down.weight",
+    "lora_unet_single_blocks_7_linear2.lora_up.weight",
+    "lora_unet_single_blocks_7_modulation_lin.alpha",
+    "lora_unet_single_blocks_7_modulation_lin.lora_down.weight",
+    "lora_unet_single_blocks_7_modulation_lin.lora_up.weight",
+    "lora_unet_single_blocks_8_linear1.alpha",
+    "lora_unet_single_blocks_8_linear1.lora_down.weight",
+    "lora_unet_single_blocks_8_linear1.lora_up.weight",
+    "lora_unet_single_blocks_8_linear2.alpha",
+    "lora_unet_single_blocks_8_linear2.lora_down.weight",
+    "lora_unet_single_blocks_8_linear2.lora_up.weight",
+    "lora_unet_single_blocks_8_modulation_lin.alpha",
+    "lora_unet_single_blocks_8_modulation_lin.lora_down.weight",
+    "lora_unet_single_blocks_8_modulation_lin.lora_up.weight",
+    "lora_unet_single_blocks_9_linear1.alpha",
+    "lora_unet_single_blocks_9_linear1.lora_down.weight",
+    "lora_unet_single_blocks_9_linear1.lora_up.weight",
+    "lora_unet_single_blocks_9_linear2.alpha",
+    "lora_unet_single_blocks_9_linear2.lora_down.weight",
+    "lora_unet_single_blocks_9_linear2.lora_up.weight",
+    "lora_unet_single_blocks_9_modulation_lin.alpha",
+    "lora_unet_single_blocks_9_modulation_lin.lora_down.weight",
+    "lora_unet_single_blocks_9_modulation_lin.lora_up.weight",
+]

tests/backend/peft/conversions/test_flux_lora_conversion_utils.py

@@ -0,0 +1,97 @@
+import pytest
+import torch
+
+from invokeai.backend.flux.model import Flux
+from invokeai.backend.flux.util import params
+from invokeai.backend.peft.conversions.flux_kohya_lora_conversion_utils import (
+    convert_flux_kohya_state_dict_to_invoke_format,
+    is_state_dict_likely_in_flux_kohya_format,
+    lora_model_from_flux_kohya_state_dict,
+)
+from tests.backend.peft.conversions.lora_state_dicts.flux_lora_kohya_format import state_dict_keys
+
+
+def test_is_state_dict_likely_in_flux_kohya_format_true():
+    """Test that is_state_dict_likely_in_flux_kohya_format() can identify a state dict in the Kohya FLUX LoRA format."""
+    # Construct a state dict that is in the Kohya FLUX LoRA format.
+    state_dict: dict[str, torch.Tensor] = {}
+    for k in state_dict_keys:
+        state_dict[k] = torch.empty(1)
+    assert is_state_dict_likely_in_flux_kohya_format(state_dict)
+
+
+def test_is_state_dict_likely_in_flux_kohya_format_false():
+    """Test that is_state_dict_likely_in_flux_kohya_format() returns False for a state dict that is not in the Kohya FLUX LoRA format."""
+    state_dict: dict[str, torch.Tensor] = {
+        "unexpected_key.lora_up.weight": torch.empty(1),
+    }
+    assert not is_state_dict_likely_in_flux_kohya_format(state_dict)
+
+
+def test_convert_flux_kohya_state_dict_to_invoke_format():
+    # Construct state_dict from state_dict_keys.
+    state_dict: dict[str, torch.Tensor] = {}
+    for k in state_dict_keys:
+        state_dict[k] = torch.empty(1)
+
+    converted_state_dict = convert_flux_kohya_state_dict_to_invoke_format(state_dict)
+
+    # Extract the prefixes from the converted state dict (i.e. without the .lora_up.weight, .lora_down.weight, and
+    # .alpha suffixes).
+    converted_key_prefixes: list[str] = []
+    for k in converted_state_dict.keys():
+        k = k.replace(".lora_up.weight", "")
+        k = k.replace(".lora_down.weight", "")
+        k = k.replace(".alpha", "")
+        converted_key_prefixes.append(k)
+
+    # Initialize a FLUX model on the meta device.
+    with torch.device("meta"):
+        model = Flux(params["flux-dev"])
+    model_keys = set(model.state_dict().keys())
+
+    # Assert that the converted state dict matches the keys in the actual model.
+    for converted_key_prefix in converted_key_prefixes:
+        found_match = False
+        for model_key in model_keys:
+            if model_key.startswith(converted_key_prefix):
+                found_match = True
+                break
+        if not found_match:
+            raise AssertionError(f"Could not find a match for the converted key prefix: {converted_key_prefix}")
+
+
+def test_convert_flux_kohya_state_dict_to_invoke_format_error():
+    """Test that an error is raised by convert_flux_kohya_state_dict_to_invoke_format() if the input state_dict contains
+    unexpected keys.
+    """
+    state_dict = {
+        "unexpected_key.lora_up.weight": torch.empty(1),
+    }
+
+    with pytest.raises(ValueError):
+        convert_flux_kohya_state_dict_to_invoke_format(state_dict)
+
+
+def test_lora_model_from_flux_kohya_state_dict():
+    """Test that a LoRAModelRaw can be created from a state dict in the Kohya FLUX LoRA format."""
+    # Construct state_dict from state_dict_keys.
+    state_dict: dict[str, torch.Tensor] = {}
+    for k in state_dict_keys:
+        state_dict[k] = torch.empty(1)
+
+    lora_model = lora_model_from_flux_kohya_state_dict(state_dict)
+
+    # Prepare expected layer keys.
+    expected_layer_keys: set[str] = set()
+    for k in state_dict_keys:
+        k = k.replace("lora_unet_", "")
+        k = k.replace(".lora_up.weight", "")
+        k = k.replace(".lora_down.weight", "")
+        k = k.replace(".alpha", "")
+        expected_layer_keys.add(k)
+
+    # Assert that the lora_model has the expected layers.
+    lora_model_keys = set(lora_model.layers.keys())
+    lora_model_keys = {k.replace(".", "_") for k in lora_model_keys}
+    assert lora_model_keys == expected_layer_keys