Modular backend - LoRA/LyCORIS (#6667)

## Summary

Code for lora patching from #6577.
Additionally made it the way, that lora can patch not only `weight`, but
also `bias`, because saw some loras which doing it.

## Related Issues / Discussions

#6606 

https://invokeai.notion.site/Modular-Stable-Diffusion-Backend-Design-Document-e8952daab5d5472faecdc4a72d377b0d

## QA Instructions

Run with and without set `USE_MODULAR_DENOISE` environment.

## Merge Plan

Replace old lora patcher with new after review done.
If you think that there should be some kind of tests - feel free to add.

## Checklist

- [x] _The PR has a short but descriptive title, suitable for a
changelog_
- [ ] _Tests added / updated (if applicable)_
- [ ] _Documentation added / updated (if applicable)_

invokeai/app/invocations/compel.py

@@ -80,12 +80,12 @@ class CompelInvocation(BaseInvocation):
 
         with (
             # apply all patches while the model is on the target device
-            text_encoder_info.model_on_device() as (model_state_dict, text_encoder),
+            text_encoder_info.model_on_device() as (cached_weights, text_encoder),
             tokenizer_info as tokenizer,
             ModelPatcher.apply_lora_text_encoder(
                 text_encoder,
                 loras=_lora_loader(),
-                model_state_dict=model_state_dict,
+                cached_weights=cached_weights,
             ),
             # Apply CLIP Skip after LoRA to prevent LoRA application from failing on skipped layers.
             ModelPatcher.apply_clip_skip(text_encoder, self.clip.skipped_layers),
@@ -175,13 +175,13 @@ class SDXLPromptInvocationBase:
 
         with (
             # apply all patches while the model is on the target device
-            text_encoder_info.model_on_device() as (state_dict, text_encoder),
+            text_encoder_info.model_on_device() as (cached_weights, text_encoder),
             tokenizer_info as tokenizer,
             ModelPatcher.apply_lora(
                 text_encoder,
                 loras=_lora_loader(),
                 prefix=lora_prefix,
-                model_state_dict=state_dict,
+                cached_weights=cached_weights,
             ),
             # Apply CLIP Skip after LoRA to prevent LoRA application from failing on skipped layers.
             ModelPatcher.apply_clip_skip(text_encoder, clip_field.skipped_layers),

invokeai/app/invocations/denoise_latents.py

@@ -62,6 +62,7 @@ from invokeai.backend.stable_diffusion.extensions.controlnet import ControlNetEx
 from invokeai.backend.stable_diffusion.extensions.freeu import FreeUExt
 from invokeai.backend.stable_diffusion.extensions.inpaint import InpaintExt
 from invokeai.backend.stable_diffusion.extensions.inpaint_model import InpaintModelExt
+from invokeai.backend.stable_diffusion.extensions.lora import LoRAExt
 from invokeai.backend.stable_diffusion.extensions.preview import PreviewExt
 from invokeai.backend.stable_diffusion.extensions.rescale_cfg import RescaleCFGExt
 from invokeai.backend.stable_diffusion.extensions.seamless import SeamlessExt
@@ -845,6 +846,16 @@ class DenoiseLatentsInvocation(BaseInvocation):
         if self.unet.freeu_config:
             ext_manager.add_extension(FreeUExt(self.unet.freeu_config))
 
+        ### lora
+        if self.unet.loras:
+            for lora_field in self.unet.loras:
+                ext_manager.add_extension(
+                    LoRAExt(
+                        node_context=context,
+                        model_id=lora_field.lora,
+                        weight=lora_field.weight,
+                    )
+                )
         ### seamless
         if self.unet.seamless_axes:
             ext_manager.add_extension(SeamlessExt(self.unet.seamless_axes))
@@ -964,14 +975,14 @@ class DenoiseLatentsInvocation(BaseInvocation):
         assert isinstance(unet_info.model, UNet2DConditionModel)
         with (
             ExitStack() as exit_stack,
-            unet_info.model_on_device() as (model_state_dict, unet),
+            unet_info.model_on_device() as (cached_weights, unet),
             ModelPatcher.apply_freeu(unet, self.unet.freeu_config),
             SeamlessExt.static_patch_model(unet, self.unet.seamless_axes),  # FIXME
             # Apply the LoRA after unet has been moved to its target device for faster patching.
             ModelPatcher.apply_lora_unet(
                 unet,
                 loras=_lora_loader(),
-                model_state_dict=model_state_dict,
+                cached_weights=cached_weights,
             ),
         ):
             assert isinstance(unet, UNet2DConditionModel)

invokeai/backend/lora.py

@@ -3,12 +3,13 @@
 
 import bisect
 from pathlib import Path
-from typing import Dict, List, Optional, Tuple, Union
+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
 
@@ -46,9 +47,19 @@ class LoRALayerBase:
         self.rank = None  # set in layer implementation
         self.layer_key = layer_key
 
-    def get_weight(self, orig_weight: Optional[torch.Tensor]) -> torch.Tensor:
+    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]:
@@ -60,6 +71,17 @@ class LoRALayerBase:
         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):
@@ -76,14 +98,19 @@ class LoRALayer(LoRALayerBase):
 
         self.up = values["lora_up.weight"]
         self.down = values["lora_down.weight"]
-        if "lora_mid.weight" in values:
-            self.mid: Optional[torch.Tensor] = values["lora_mid.weight"]
-        else:
-            self.mid = None
+        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: Optional[torch.Tensor]) -> torch.Tensor:
+    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])
@@ -125,20 +152,23 @@ class LoHALayer(LoRALayerBase):
         self.w1_b = values["hada_w1_b"]
         self.w2_a = values["hada_w2_a"]
         self.w2_b = values["hada_w2_b"]
-
-        if "hada_t1" in values:
-            self.t1: Optional[torch.Tensor] = values["hada_t1"]
-        else:
-            self.t1 = None
-
-        if "hada_t2" in values:
-            self.t2: Optional[torch.Tensor] = values["hada_t2"]
-        else:
-            self.t2 = None
+        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: Optional[torch.Tensor]) -> torch.Tensor:
+    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)
 
@@ -186,37 +216,39 @@ class LoKRLayer(LoRALayerBase):
     ):
         super().__init__(layer_key, values)
 
-        if "lokr_w1" in values:
-            self.w1: Optional[torch.Tensor] = values["lokr_w1"]
-            self.w1_a = None
-            self.w1_b = None
-        else:
-            self.w1 = None
+        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"]
 
-        if "lokr_w2" in values:
-            self.w2: Optional[torch.Tensor] = values["lokr_w2"]
-            self.w2_a = None
-            self.w2_b = None
-        else:
-            self.w2 = 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"]
 
-        if "lokr_t2" in values:
-            self.t2: Optional[torch.Tensor] = values["lokr_t2"]
-        else:
-            self.t2 = None
+        self.t2 = values.get("lokr_t2", None)
 
-        if "lokr_w1_b" in values:
-            self.rank = values["lokr_w1_b"].shape[0]
-        elif "lokr_w2_b" in values:
-            self.rank = values["lokr_w2_b"].shape[0]
+        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
 
-    def get_weight(self, orig_weight: Optional[torch.Tensor]) -> torch.Tensor:
+        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
@@ -272,7 +304,9 @@ class LoKRLayer(LoRALayerBase):
 
 
 class FullLayer(LoRALayerBase):
+    # bias handled in LoRALayerBase(calc_size, to)
     # weight: torch.Tensor
+    # bias: Optional[torch.Tensor]
 
     def __init__(
         self,
@@ -282,15 +316,12 @@ class FullLayer(LoRALayerBase):
         super().__init__(layer_key, values)
 
         self.weight = values["diff"]
-
-        if len(values.keys()) > 1:
-            _keys = list(values.keys())
-            _keys.remove("diff")
-            raise NotImplementedError(f"Unexpected keys in lora diff layer: {_keys}")
+        self.bias = values.get("diff_b", None)
 
         self.rank = None  # unscaled
+        self.check_keys(values, {"diff", "diff_b"})
 
-    def get_weight(self, orig_weight: Optional[torch.Tensor]) -> torch.Tensor:
+    def get_weight(self, orig_weight: torch.Tensor) -> torch.Tensor:
         return self.weight
 
     def calc_size(self) -> int:
@@ -319,8 +350,9 @@ class IA3Layer(LoRALayerBase):
         self.on_input = values["on_input"]
 
         self.rank = None  # unscaled
+        self.check_keys(values, {"weight", "on_input"})
 
-    def get_weight(self, orig_weight: Optional[torch.Tensor]) -> torch.Tensor:
+    def get_weight(self, orig_weight: torch.Tensor) -> torch.Tensor:
         weight = self.weight
         if not self.on_input:
             weight = weight.reshape(-1, 1)
@@ -458,16 +490,19 @@ class LoRAModelRaw(RawModel):  # (torch.nn.Module):
             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_down.weight" in values:
+            if "lora_up.weight" in values:
                 layer: AnyLoRALayer = LoRALayer(layer_key, values)
 
             # loha
-            elif "hada_w1_b" in values:
+            elif "hada_w1_a" in values:
                 layer = LoHALayer(layer_key, values)
 
             # lokr
-            elif "lokr_w1_b" in values or "lokr_w1" in values:
+            elif "lokr_w1" in values or "lokr_w1_a" in values:
                 layer = LoKRLayer(layer_key, values)
 
             # diff
@@ -475,7 +510,7 @@ class LoRAModelRaw(RawModel):  # (torch.nn.Module):
                 layer = FullLayer(layer_key, values)
 
             # ia3
-            elif "weight" in values and "on_input" in values:
+            elif "on_input" in values:
                 layer = IA3Layer(layer_key, values)
 
             else:

invokeai/backend/model_patcher.py

@@ -17,8 +17,9 @@ 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.stable_diffusion.extensions.lora import LoRAExt
 from invokeai.backend.textual_inversion import TextualInversionManager, TextualInversionModelRaw
-from invokeai.backend.util.devices import TorchDevice
+from invokeai.backend.util.original_weights_storage import OriginalWeightsStorage
 
 """
 loras = [
@@ -85,13 +86,13 @@ class ModelPatcher:
         cls,
         unet: UNet2DConditionModel,
         loras: Iterator[Tuple[LoRAModelRaw, float]],
-        model_state_dict: Optional[Dict[str, torch.Tensor]] = None,
+        cached_weights: Optional[Dict[str, torch.Tensor]] = None,
     ) -> Generator[None, None, None]:
         with cls.apply_lora(
             unet,
             loras=loras,
             prefix="lora_unet_",
-            model_state_dict=model_state_dict,
+            cached_weights=cached_weights,
         ):
             yield
 
@@ -101,9 +102,9 @@ class ModelPatcher:
         cls,
         text_encoder: CLIPTextModel,
         loras: Iterator[Tuple[LoRAModelRaw, float]],
-        model_state_dict: Optional[Dict[str, torch.Tensor]] = None,
+        cached_weights: Optional[Dict[str, torch.Tensor]] = None,
     ) -> Generator[None, None, None]:
-        with cls.apply_lora(text_encoder, loras=loras, prefix="lora_te_", model_state_dict=model_state_dict):
+        with cls.apply_lora(text_encoder, loras=loras, prefix="lora_te_", cached_weights=cached_weights):
             yield
 
     @classmethod
@@ -113,7 +114,7 @@ class ModelPatcher:
         model: AnyModel,
         loras: Iterator[Tuple[LoRAModelRaw, float]],
         prefix: str,
-        model_state_dict: Optional[Dict[str, torch.Tensor]] = None,
+        cached_weights: Optional[Dict[str, torch.Tensor]] = None,
     ) -> Generator[None, None, None]:
         """
         Apply one or more LoRAs to a model.
@@ -121,66 +122,26 @@ class ModelPatcher:
         :param model: The model to patch.
         :param loras: An iterator that returns the LoRA to patch in and its patch weight.
         :param prefix: A string prefix that precedes keys used in the LoRAs weight layers.
-        :model_state_dict: Read-only copy of the model's state dict in CPU, for unpatching purposes.
+        :cached_weights: Read-only copy of the model's state dict in CPU, for unpatching purposes.
         """
-        original_weights = {}
+        original_weights = OriginalWeightsStorage(cached_weights)
         try:
-            with torch.no_grad():
-                for lora, lora_weight in loras:
-                    # assert lora.device.type == "cpu"
-                    for layer_key, layer in lora.layers.items():
-                        if not layer_key.startswith(prefix):
-                            continue
+            for lora_model, lora_weight in loras:
+                LoRAExt.patch_model(
+                    model=model,
+                    prefix=prefix,
+                    lora=lora_model,
+                    lora_weight=lora_weight,
+                    original_weights=original_weights,
+                )
+                del lora_model
 
-                        # TODO(ryand): A non-negligible amount of time is currently spent resolving LoRA keys. This
-                        # should be improved in the following ways:
-                        # 1. The key mapping could be more-efficiently pre-computed. This would save time every time a
-                        #    LoRA model is applied.
-                        # 2. From an API perspective, there's no reason that the `ModelPatcher` should be aware of the
-                        #    intricacies of Stable Diffusion key resolution. It should just expect the input LoRA
-                        #    weights to have valid keys.
-                        assert isinstance(model, torch.nn.Module)
-                        module_key, module = cls._resolve_lora_key(model, layer_key, prefix)
-
-                        # 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
-
-                        if module_key not in original_weights:
-                            if model_state_dict is not None:  # we were provided with the CPU copy of the state dict
-                                original_weights[module_key] = model_state_dict[module_key + ".weight"]
-                            else:
-                                original_weights[module_key] = module.weight.detach().to(device="cpu", copy=True)
-
-                        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.
-                        layer_weight = layer.get_weight(module.weight) * (lora_weight * layer_scale)
-                        layer.to(device=TorchDevice.CPU_DEVICE)
-
-                        assert isinstance(layer_weight, torch.Tensor)  # mypy thinks layer_weight is a float|Any ??!
-                        if module.weight.shape != layer_weight.shape:
-                            # TODO: debug on lycoris
-                            assert hasattr(layer_weight, "reshape")
-                            layer_weight = layer_weight.reshape(module.weight.shape)
-
-                        assert isinstance(layer_weight, torch.Tensor)  # mypy thinks layer_weight is a float|Any ??!
-                        module.weight += layer_weight.to(dtype=dtype)
-
-            yield  # wait for context manager exit
+            yield
 
         finally:
-            assert hasattr(model, "get_submodule")  # mypy not picking up fact that torch.nn.Module has get_submodule()
             with torch.no_grad():
-                for module_key, weight in original_weights.items():
-                    model.get_submodule(module_key).weight.copy_(weight)
+                for param_key, weight in original_weights.get_changed_weights():
+                    model.get_parameter(param_key).copy_(weight)
 
     @classmethod
     @contextmanager

invokeai/backend/stable_diffusion/extensions/base.py

@@ -2,14 +2,14 @@ from __future__ import annotations
 
 from contextlib import contextmanager
 from dataclasses import dataclass
-from typing import TYPE_CHECKING, Callable, Dict, List, Optional
+from typing import TYPE_CHECKING, Callable, Dict, List
 
-import torch
 from diffusers import UNet2DConditionModel
 
 if TYPE_CHECKING:
     from invokeai.backend.stable_diffusion.denoise_context import DenoiseContext
     from invokeai.backend.stable_diffusion.extension_callback_type import ExtensionCallbackType
+    from invokeai.backend.util.original_weights_storage import OriginalWeightsStorage
 
 
 @dataclass
@@ -56,5 +56,17 @@ class ExtensionBase:
         yield None
 
     @contextmanager
-    def patch_unet(self, unet: UNet2DConditionModel, cached_weights: Optional[Dict[str, torch.Tensor]] = None):
-        yield None
+    def patch_unet(self, unet: UNet2DConditionModel, original_weights: OriginalWeightsStorage):
+        """A context manager for applying patches to the UNet model. The context manager's lifetime spans the entire
+        diffusion process. Weight unpatching is handled upstream, and is achieved by saving unchanged weights by
+        `original_weights.save` function. Note that this enables some performance optimization by avoiding redundant
+        operations. All other patches (e.g. changes to tensor shapes, function monkey-patches, etc.) should be unpatched
+        by this context manager.
+
+        Args:
+            unet (UNet2DConditionModel): The UNet model on execution device to patch.
+            original_weights (OriginalWeightsStorage): A storage with copy of the model's original weights in CPU, for
+                unpatching purposes. Extension should save tensor which being modified in this storage, also extensions
+                can access original weights values.
+        """
+        yield

invokeai/backend/stable_diffusion/extensions/freeu.py

@@ -1,15 +1,15 @@
 from __future__ import annotations
 
 from contextlib import contextmanager
-from typing import TYPE_CHECKING, Dict, Optional
+from typing import TYPE_CHECKING
 
-import torch
 from diffusers import UNet2DConditionModel
 
 from invokeai.backend.stable_diffusion.extensions.base import ExtensionBase
 
 if TYPE_CHECKING:
     from invokeai.app.shared.models import FreeUConfig
+    from invokeai.backend.util.original_weights_storage import OriginalWeightsStorage
 
 
 class FreeUExt(ExtensionBase):
@@ -21,7 +21,7 @@ class FreeUExt(ExtensionBase):
         self._freeu_config = freeu_config
 
     @contextmanager
-    def patch_unet(self, unet: UNet2DConditionModel, cached_weights: Optional[Dict[str, torch.Tensor]] = None):
+    def patch_unet(self, unet: UNet2DConditionModel, original_weights: OriginalWeightsStorage):
         unet.enable_freeu(
             b1=self._freeu_config.b1,
             b2=self._freeu_config.b2,

invokeai/backend/stable_diffusion/extensions/lora.py

@@ -0,0 +1,137 @@
+from __future__ import annotations
+
+from contextlib import contextmanager
+from typing import TYPE_CHECKING, Tuple
+
+import torch
+from diffusers import UNet2DConditionModel
+
+from invokeai.backend.stable_diffusion.extensions.base import ExtensionBase
+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.util.original_weights_storage import OriginalWeightsStorage
+
+
+class LoRAExt(ExtensionBase):
+    def __init__(
+        self,
+        node_context: InvocationContext,
+        model_id: ModelIdentifierField,
+        weight: float,
+    ):
+        super().__init__()
+        self._node_context = node_context
+        self._model_id = model_id
+        self._weight = weight
+
+    @contextmanager
+    def patch_unet(self, unet: UNet2DConditionModel, original_weights: OriginalWeightsStorage):
+        lora_model = self._node_context.models.load(self._model_id).model
+        self.patch_model(
+            model=unet,
+            prefix="lora_unet_",
+            lora=lora_model,
+            lora_weight=self._weight,
+            original_weights=original_weights,
+        )
+        del lora_model
+
+        yield
+
+    @classmethod
+    @torch.no_grad()
+    def patch_model(
+        cls,
+        model: torch.nn.Module,
+        prefix: str,
+        lora: LoRAModelRaw,
+        lora_weight: float,
+        original_weights: OriginalWeightsStorage,
+    ):
+        """
+        Apply one or more LoRAs to a model.
+        :param model: The model to patch.
+        :param lora: LoRA model to patch in.
+        :param lora_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 lora_weight == 0:
+            return
+
+        # assert lora.device.type == "cpu"
+        for layer_key, layer in lora.layers.items():
+            if not layer_key.startswith(prefix):
+                continue
+
+            # TODO(ryand): A non-negligible amount of time is currently spent resolving LoRA keys. This
+            # should be improved in the following ways:
+            # 1. The key mapping could be more-efficiently pre-computed. This would save time every time a
+            #    LoRA model is applied.
+            # 2. From an API perspective, there's no reason that the `ModelPatcher` should be aware of the
+            #    intricacies of Stable Diffusion key resolution. It should just expect the input LoRA
+            #    weights to have valid keys.
+            assert isinstance(model, torch.nn.Module)
+            module_key, module = cls._resolve_lora_key(model, layer_key, prefix)
+
+            # 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 = module_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:
+                    # TODO: debug on lycoris
+                    lora_param_weight = lora_param_weight.reshape(module_param.shape)
+
+                lora_param_weight *= lora_weight * layer_scale
+                module_param += lora_param_weight.to(dtype=dtype)
+
+            layer.to(device=TorchDevice.CPU_DEVICE)
+
+    @staticmethod
+    def _resolve_lora_key(model: torch.nn.Module, lora_key: str, prefix: str) -> Tuple[str, torch.nn.Module]:
+        assert "." not in lora_key
+
+        if not lora_key.startswith(prefix):
+            raise Exception(f"lora_key with invalid prefix: {lora_key}, {prefix}")
+
+        module = model
+        module_key = ""
+        key_parts = lora_key[len(prefix) :].split("_")
+
+        submodule_name = key_parts.pop(0)
+
+        while len(key_parts) > 0:
+            try:
+                module = module.get_submodule(submodule_name)
+                module_key += "." + submodule_name
+                submodule_name = key_parts.pop(0)
+            except Exception:
+                submodule_name += "_" + key_parts.pop(0)
+
+        module = module.get_submodule(submodule_name)
+        module_key = (module_key + "." + submodule_name).lstrip(".")
+
+        return (module_key, module)

invokeai/backend/stable_diffusion/extensions_manager.py

@@ -7,6 +7,7 @@ import torch
 from diffusers import UNet2DConditionModel
 
 from invokeai.app.services.session_processor.session_processor_common import CanceledException
+from invokeai.backend.util.original_weights_storage import OriginalWeightsStorage
 
 if TYPE_CHECKING:
     from invokeai.backend.stable_diffusion.denoise_context import DenoiseContext
@@ -67,9 +68,15 @@ class ExtensionsManager:
         if self._is_canceled and self._is_canceled():
             raise CanceledException
 
-        # TODO: create weight patch logic in PR with extension which uses it
+        original_weights = OriginalWeightsStorage(cached_weights)
+        try:
             with ExitStack() as exit_stack:
                 for ext in self._extensions:
-                exit_stack.enter_context(ext.patch_unet(unet, cached_weights))
+                    exit_stack.enter_context(ext.patch_unet(unet, original_weights))
 
                 yield None
+
+        finally:
+            with torch.no_grad():
+                for param_key, weight in original_weights.get_changed_weights():
+                    unet.get_parameter(param_key).copy_(weight)

invokeai/backend/util/original_weights_storage.py

@@ -0,0 +1,39 @@
+from __future__ import annotations
+
+from typing import Dict, Iterator, Optional, Tuple
+
+import torch
+
+from invokeai.backend.util.devices import TorchDevice
+
+
+class OriginalWeightsStorage:
+    """A class for tracking the original weights of a model for patch/unpatch operations."""
+
+    def __init__(self, cached_weights: Optional[Dict[str, torch.Tensor]] = None):
+        # The original weights of the model.
+        self._weights: dict[str, torch.Tensor] = {}
+        # The keys of the weights that have been changed (via `save()`) during the lifetime of this instance.
+        self._changed_weights: set[str] = set()
+        if cached_weights:
+            self._weights.update(cached_weights)
+
+    def save(self, key: str, weight: torch.Tensor, copy: bool = True):
+        self._changed_weights.add(key)
+        if key in self._weights:
+            return
+
+        self._weights[key] = weight.detach().to(device=TorchDevice.CPU_DEVICE, copy=copy)
+
+    def get(self, key: str, copy: bool = False) -> Optional[torch.Tensor]:
+        weight = self._weights.get(key, None)
+        if weight is not None and copy:
+            weight = weight.clone()
+        return weight
+
+    def contains(self, key: str) -> bool:
+        return key in self._weights
+
+    def get_changed_weights(self) -> Iterator[Tuple[str, torch.Tensor]]:
+        for key in self._changed_weights:
+            yield key, self._weights[key]