WIP - still need to figure out why the keys are wrong.

invokeai/app/invocations/compel.py

@@ -9,7 +9,8 @@ from invokeai.app.invocations.fields import FieldDescriptions, Input, InputField
 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.lora_model_patcher import LoraModelPatcher
+from invokeai.backend.lora_model_raw import LoRAModelRaw
 from invokeai.backend.model_patcher import ModelPatcher
 from invokeai.backend.stable_diffusion.diffusion.conditioning_data import (
     BasicConditioningInfo,
@@ -80,7 +81,8 @@ class CompelInvocation(BaseInvocation):
             ),
             text_encoder_info as text_encoder,
             # Apply the LoRA after text_encoder has been moved to its target device for faster patching.
-            ModelPatcher.apply_lora_text_encoder(text_encoder, _lora_loader()),
+            # ModelPatcher.apply_lora_text_encoder(text_encoder, _lora_loader()),
+            LoraModelPatcher.apply_lora_to_text_encoder(text_encoder, _lora_loader(), "text_encoder"),
             # Apply CLIP Skip after LoRA to prevent LoRA application from failing on skipped layers.
             ModelPatcher.apply_clip_skip(text_encoder_model, self.clip.skipped_layers),
         ):
@@ -181,7 +183,8 @@ class SDXLPromptInvocationBase:
             ),
             text_encoder_info as text_encoder,
             # Apply the LoRA after text_encoder has been moved to its target device for faster patching.
-            ModelPatcher.apply_lora(text_encoder, _lora_loader(), lora_prefix),
+            # ModelPatcher.apply_lora(text_encoder, _lora_loader(), lora_prefix),
+            LoraModelPatcher.apply_lora_to_text_encoder(text_encoder, _lora_loader(), lora_prefix),
             # Apply CLIP Skip after LoRA to prevent LoRA application from failing on skipped layers.
             ModelPatcher.apply_clip_skip(text_encoder_model, clip_field.skipped_layers),
         ):
@@ -259,15 +262,15 @@ class SDXLCompelPromptInvocation(BaseInvocation, SDXLPromptInvocationBase):
     @torch.no_grad()
     def invoke(self, context: InvocationContext) -> ConditioningOutput:
         c1, c1_pooled, ec1 = self.run_clip_compel(
-            context, self.clip, self.prompt, False, "lora_te1_", zero_on_empty=True
+            context, self.clip, self.prompt, False, "text_encoder", zero_on_empty=True
         )
         if self.style.strip() == "":
             c2, c2_pooled, ec2 = self.run_clip_compel(
-                context, self.clip2, self.prompt, True, "lora_te2_", zero_on_empty=True
+                context, self.clip2, self.prompt, True, "text_encoder_2", zero_on_empty=True
             )
         else:
             c2, c2_pooled, ec2 = self.run_clip_compel(
-                context, self.clip2, self.style, True, "lora_te2_", zero_on_empty=True
+                context, self.clip2, self.style, True, "text_encoder_2", zero_on_empty=True
             )
 
         original_size = (self.original_height, self.original_width)

invokeai/app/invocations/latent.py

@@ -52,7 +52,8 @@ 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, IPAdapterPlus
-from invokeai.backend.lora import LoRAModelRaw
+from invokeai.backend.lora_model_patcher import LoraModelPatcher
+from invokeai.backend.lora_model_raw import LoRAModelRaw
 from invokeai.backend.model_manager import BaseModelType, LoadedModel
 from invokeai.backend.model_patcher import ModelPatcher
 from invokeai.backend.stable_diffusion import PipelineIntermediateState, set_seamless
@@ -739,7 +740,8 @@ class DenoiseLatentsInvocation(BaseInvocation):
                 set_seamless(unet_info.model, self.unet.seamless_axes),  # FIXME
                 unet_info as unet,
                 # Apply the LoRA after unet has been moved to its target device for faster patching.
-                ModelPatcher.apply_lora_unet(unet, _lora_loader()),
+                # ModelPatcher.apply_lora_unet(unet, _lora_loader()),
+                LoraModelPatcher.apply_lora_to_unet(unet, _lora_loader()),
             ):
                 assert isinstance(unet, UNet2DConditionModel)
                 latents = latents.to(device=unet.device, dtype=unet.dtype)

invokeai/backend/lora_model_patcher.py

@@ -0,0 +1,65 @@
+from contextlib import contextmanager
+from typing import Iterator, Tuple, Union
+
+from diffusers.loaders.lora import LoraLoaderMixin
+from diffusers.models.unets.unet_2d_condition import UNet2DConditionModel
+from diffusers.utils.peft_utils import recurse_remove_peft_layers
+from transformers import CLIPTextModel
+
+from invokeai.backend.lora_model_raw import LoRAModelRaw
+
+
+class LoraModelPatcher:
+    @classmethod
+    def unload_lora_from_model(cls, m: Union[UNet2DConditionModel, CLIPTextModel]):
+        """Unload all LoRA models from a UNet or Text Encoder.
+        This implementation is base on LoraLoaderMixin.unload_lora_weights().
+        """
+        recurse_remove_peft_layers(m)
+        if hasattr(m, "peft_config"):
+            del m.peft_config  # type: ignore
+        if hasattr(m, "_hf_peft_config_loaded"):
+            m._hf_peft_config_loaded = None  # type: ignore
+
+    @classmethod
+    @contextmanager
+    def apply_lora_to_unet(cls, unet: UNet2DConditionModel, loras: Iterator[Tuple[LoRAModelRaw, float]]):
+        try:
+            # TODO(ryand): Test speed of low_cpu_mem_usage=True.
+            for lora, lora_weight in loras:
+                LoraLoaderMixin.load_lora_into_unet(
+                    state_dict=lora.state_dict,
+                    network_alphas=lora.network_alphas,
+                    unet=unet,
+                    low_cpu_mem_usage=True,
+                    adapter_name=lora.name,
+                    _pipeline=None,
+                )
+            yield
+        finally:
+            cls.unload_lora_from_model(unet)
+
+    @classmethod
+    @contextmanager
+    def apply_lora_to_text_encoder(
+        cls, text_encoder: CLIPTextModel, loras: Iterator[Tuple[LoRAModelRaw, float]], prefix: str
+    ):
+        assert prefix in ["text_encoder", "text_encoder_2"]
+        try:
+            for lora, lora_weight in loras:
+                # Filter the state_dict to only include the keys that start with the prefix.
+                text_encoder_state_dict = {
+                    key: value for key, value in lora.state_dict.items() if key.startswith(prefix + ".")
+                }
+                if len(text_encoder_state_dict) > 0:
+                    LoraLoaderMixin.load_lora_into_text_encoder(
+                        state_dict=text_encoder_state_dict,
+                        network_alphas=lora.network_alphas,
+                        text_encoder=text_encoder,
+                        low_cpu_mem_usage=True,
+                        adapter_name=lora.name,
+                        _pipeline=None,
+                    )
+            yield
+        finally:
+            cls.unload_lora_from_model(text_encoder)

invokeai/backend/lora_model_raw.py

@@ -0,0 +1,66 @@
+from pathlib import Path
+from typing import Optional, Union
+
+import torch
+from diffusers.loaders.lora import LoraLoaderMixin
+from typing_extensions import Self
+
+
+class LoRAModelRaw:
+    def __init__(
+        self,
+        name: str,
+        state_dict: dict[str, torch.Tensor],
+        network_alphas: Optional[dict[str, float]],
+    ):
+        self._name = name
+        self.state_dict = state_dict
+        self.network_alphas = network_alphas
+
+    @property
+    def name(self) -> str:
+        return self._name
+
+    def to(self, device: Optional[torch.device] = None, dtype: Optional[torch.dtype] = None) -> None:
+        for key, layer in self.state_dict.items():
+            self.state_dict[key] = layer.to(device=device, dtype=dtype)
+
+    def calc_size(self) -> int:
+        """Calculate the size of the model in bytes."""
+        model_size = 0
+        for layer in self.state_dict.values():
+            model_size += layer.numel() * layer.element_size()
+        return model_size
+
+    @classmethod
+    def from_checkpoint(
+        cls, file_path: Union[str, Path], device: Optional[torch.device] = None, dtype: Optional[torch.dtype] = None
+    ) -> Self:
+        """This function is based on diffusers LoraLoaderMixin.load_lora_weights()."""
+
+        file_path = Path(file_path)
+        if file_path.is_dir():
+            raise NotImplementedError("LoRA models from directories are not yet supported.")
+
+        dir_path = file_path.parent
+        file_name = file_path.name
+
+        state_dict, network_alphas = LoraLoaderMixin.lora_state_dict(
+            pretrained_model_name_or_path_or_dict=str(file_path), local_files_only=True, weight_name=str(file_name)
+        )
+
+        is_correct_format = all("lora" in key for key in state_dict.keys())
+        if not is_correct_format:
+            raise ValueError("Invalid LoRA checkpoint.")
+
+        model = cls(
+            # TODO(ryand): Handle both files and directories here?
+            name=Path(file_path).stem,
+            state_dict=state_dict,
+            network_alphas=network_alphas,
+        )
+
+        device = device or torch.device("cpu")
+        dtype = dtype or torch.float32
+        model.to(device=device, dtype=dtype)
+        return model

invokeai/backend/model_manager/config.py

@@ -32,7 +32,7 @@ from typing_extensions import Annotated, Any, Dict
 from invokeai.app.invocations.constants import SCHEDULER_NAME_VALUES
 from invokeai.app.util.misc import uuid_string
 from invokeai.backend.ip_adapter.ip_adapter import IPAdapter
-from invokeai.backend.lora import LoRAModelRaw
+from invokeai.backend.lora_model_raw import LoRAModelRaw
 from invokeai.backend.onnx.onnx_runtime import IAIOnnxRuntimeModel
 from invokeai.backend.textual_inversion import TextualInversionModelRaw
 

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

@@ -6,7 +6,7 @@ from pathlib import Path
 from typing import Optional, Tuple
 
 from invokeai.app.services.config import InvokeAIAppConfig
-from invokeai.backend.lora import LoRAModelRaw
+from invokeai.backend.lora_model_raw import LoRAModelRaw
 from invokeai.backend.model_manager import (
     AnyModel,
     AnyModelConfig,
@@ -51,7 +51,6 @@ class LoRALoader(ModelLoader):
         model = LoRAModelRaw.from_checkpoint(
             file_path=model_path,
             dtype=self._torch_dtype,
-            base_model=self._model_base,
         )
         return model
 

invokeai/backend/model_patcher.py

@@ -17,7 +17,7 @@ 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 .lora import LoRAModelRaw
+from .lora_model_raw import LoRAModelRaw
 from .textual_inversion import TextualInversionManager, TextualInversionModelRaw
 
 """

tests/backend/model_manager/test_lora.py

@@ -5,7 +5,7 @@
 import pytest
 import torch
 
-from invokeai.backend.lora import LoRALayer, LoRAModelRaw
+from invokeai.backend.lora_model_raw import LoRALayer, LoRAModelRaw
 from invokeai.backend.model_patcher import ModelPatcher