LoRA patching optimization (#6439)

* allow model patcher to optimize away the unpatching step when feasible

* remove lazy_offloading functionality

* allow model patcher to optimize away the unpatching step when feasible

* remove lazy_offloading functionality

* do not save original weights if there is a CPU copy of state dict

* Update invokeai/backend/model_manager/load/load_base.py

Co-authored-by: Ryan Dick <ryanjdick3@gmail.com>

* documentation fixes added during penultimate review

---------

Co-authored-by: Lincoln Stein <lstein@gmail.com>
Co-authored-by: Kent Keirsey <31807370+hipsterusername@users.noreply.github.com>
Co-authored-by: Ryan Dick <ryanjdick3@gmail.com>

Author: 4 people

docs/contributing/MODEL_MANAGER.md

@@ -1367,30 +1367,54 @@ the in-memory loaded model:
 | `model`        | AnyModel               | The instantiated model (details below) |
 | `locker`       | ModelLockerBase        | A context manager that mediates the movement of the model into VRAM |
 
-Because the loader can return multiple model types, it is typed to
-return `AnyModel`, a Union `ModelMixin`, `torch.nn.Module`,
-`IAIOnnxRuntimeModel`, `IPAdapter`, `IPAdapterPlus`, and
-`EmbeddingModelRaw`. `ModelMixin` is the base class of all diffusers
-models, `EmbeddingModelRaw` is used for LoRA and TextualInversion
-models. The others are obvious.
+### get_model_by_key(key, [submodel]) -> LoadedModel
+
+The `get_model_by_key()` method will retrieve the model using its
+unique database key. For example:
+
+loaded_model = loader.get_model_by_key('f13dd932c0c35c22dcb8d6cda4203764', SubModelType('vae'))
+
+`get_model_by_key()` may raise any of the following exceptions:
+
+* `UnknownModelException`   -- key not in database
+* `ModelNotFoundException`  -- key in database but model not found at path
+* `NotImplementedException` -- the loader doesn't know how to load this type of model
+
+### Using the Loaded Model in Inference
 
 `LoadedModel` acts as a context manager. The context loads the model
 into the execution device (e.g. VRAM on CUDA systems), locks the model
 in the execution device for the duration of the context, and returns
 the model. Use it like this:
 
 ```
-model_info = loader.get_model_by_key('f13dd932c0c35c22dcb8d6cda4203764', SubModelType('vae'))
-with model_info as vae:
+loaded_model_= loader.get_model_by_key('f13dd932c0c35c22dcb8d6cda4203764', SubModelType('vae'))
+with loaded_model as vae:
  image = vae.decode(latents)[0]
 ```
 
-`get_model_by_key()` may raise any of the following exceptions:
+The object returned by the LoadedModel context manager is an
+`AnyModel`, which is a Union of `ModelMixin`, `torch.nn.Module`,
+`IAIOnnxRuntimeModel`, `IPAdapter`, `IPAdapterPlus`, and
+`EmbeddingModelRaw`. `ModelMixin` is the base class of all diffusers
+models, `EmbeddingModelRaw` is used for LoRA and TextualInversion
+models. The others are obvious.
+
+In addition, you may call `LoadedModel.model_on_device()`, a context
+manager that returns a tuple of the model's state dict in CPU and the
+model itself in VRAM. It is used to optimize the LoRA patching and
+unpatching process:
+
+```
+loaded_model_= loader.get_model_by_key('f13dd932c0c35c22dcb8d6cda4203764', SubModelType('vae'))
+with loaded_model.model_on_device() as (state_dict, vae):
+ image = vae.decode(latents)[0]
+```
+
+Since not all models have state dicts, the `state_dict` return value
+can be None.
+
 
-* `UnknownModelException`   -- key not in database
-* `ModelNotFoundException`  -- key in database but model not found at path
-* `NotImplementedException` -- the loader doesn't know how to load this type of model
-  
 ### Emitting model loading events
 
 When the `context` argument is passed to `load_model_*()`, it will

invokeai/app/invocations/compel.py

@@ -81,9 +81,13 @@ def _lora_loader() -> Iterator[Tuple[LoRAModelRaw, float]]:
 
         with (
             # apply all patches while the model is on the target device
-            text_encoder_info as text_encoder,
+            text_encoder_info.model_on_device() as (model_state_dict, text_encoder),
             tokenizer_info as tokenizer,
-            ModelPatcher.apply_lora_text_encoder(text_encoder, _lora_loader()),
+            ModelPatcher.apply_lora_text_encoder(
+                text_encoder,
+                loras=_lora_loader(),
+                model_state_dict=model_state_dict,
+            ),
             # Apply CLIP Skip after LoRA to prevent LoRA application from failing on skipped layers.
             ModelPatcher.apply_clip_skip(text_encoder, self.clip.skipped_layers),
             ModelPatcher.apply_ti(tokenizer, text_encoder, ti_list) as (
@@ -172,9 +176,14 @@ def _lora_loader() -> Iterator[Tuple[LoRAModelRaw, float]]:
 
         with (
             # apply all patches while the model is on the target device
-            text_encoder_info as text_encoder,
+            text_encoder_info.model_on_device() as (state_dict, text_encoder),
             tokenizer_info as tokenizer,
-            ModelPatcher.apply_lora(text_encoder, _lora_loader(), lora_prefix),
+            ModelPatcher.apply_lora(
+                text_encoder,
+                loras=_lora_loader(),
+                prefix=lora_prefix,
+                model_state_dict=state_dict,
+            ),
             # Apply CLIP Skip after LoRA to prevent LoRA application from failing on skipped layers.
             ModelPatcher.apply_clip_skip(text_encoder, clip_field.skipped_layers),
             ModelPatcher.apply_ti(tokenizer, text_encoder, ti_list) as (

invokeai/app/invocations/latent.py

@@ -952,11 +952,15 @@ def _lora_loader() -> Iterator[Tuple[LoRAModelRaw, float]]:
             assert isinstance(unet_info.model, UNet2DConditionModel)
             with (
                 ExitStack() as exit_stack,
-                unet_info as unet,
+                unet_info.model_on_device() as (model_state_dict, unet),
                 ModelPatcher.apply_freeu(unet, self.unet.freeu_config),
                 set_seamless(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, _lora_loader()),
+                ModelPatcher.apply_lora_unet(
+                    unet,
+                    loras=_lora_loader(),
+                    model_state_dict=model_state_dict,
+                ),
             ):
                 assert isinstance(unet, UNet2DConditionModel)
                 latents = latents.to(device=unet.device, dtype=unet.dtype)

invokeai/backend/model_manager/load/load_base.py

@@ -4,10 +4,13 @@
 """
 
 from abc import ABC, abstractmethod
+from contextlib import contextmanager
 from dataclasses import dataclass
 from logging import Logger
 from pathlib import Path
-from typing import Any, Optional
+from typing import Any, Dict, Generator, Optional, Tuple
+
+import torch
 
 from invokeai.app.services.config import InvokeAIAppConfig
 from invokeai.backend.model_manager.config import (
@@ -21,7 +24,42 @@
 
 @dataclass
 class LoadedModel:
-    """Context manager object that mediates transfer from RAM<->VRAM."""
+    """
+    Context manager object that mediates transfer from RAM<->VRAM.
+
+    This is a context manager object that has two distinct APIs:
+
+    1. Older API (deprecated):
+    Use the LoadedModel object directly as a context manager.
+    It will move the model into VRAM (on CUDA devices), and
+    return the model in a form suitable for passing to torch.
+    Example:
+    ```
+    loaded_model_= loader.get_model_by_key('f13dd932', SubModelType('vae'))
+    with loaded_model as vae:
+      image = vae.decode(latents)[0]
+    ```
+
+    2. Newer API (recommended):
+    Call the LoadedModel's `model_on_device()` method in a
+    context. It returns a tuple consisting of a copy of
+    the model's state dict in CPU RAM followed by a copy
+    of the model in VRAM. The state dict is provided to allow
+    LoRAs and other model patchers to return the model to
+    its unpatched state without expensive copy and restore
+    operations.
+
+    Example:
+    ```
+    loaded_model_= loader.get_model_by_key('f13dd932', SubModelType('vae'))
+    with loaded_model.model_on_device() as (state_dict, vae):
+        image = vae.decode(latents)[0]
+    ```
+
+    The state_dict should be treated as a read-only object and
+    never modified. Also be aware that some loadable models do
+    not have a state_dict, in which case this value will be None.
+    """
 
     config: AnyModelConfig
     _locker: ModelLockerBase
@@ -35,6 +73,16 @@ def __exit__(self, *args: Any, **kwargs: Any) -> None:
         """Context exit."""
         self._locker.unlock()
 
+    @contextmanager
+    def model_on_device(self) -> Generator[Tuple[Optional[Dict[str, torch.Tensor]], AnyModel], None, None]:
+        """Return a tuple consisting of the model's state dict (if it exists) and the locked model on execution device."""
+        locked_model = self._locker.lock()
+        try:
+            state_dict = self._locker.get_state_dict()
+            yield (state_dict, locked_model)
+        finally:
+            self._locker.unlock()
+
     @property
     def model(self) -> AnyModel:
         """Return the model without locking it."""

invokeai/backend/model_manager/load/model_cache/model_cache_base.py

@@ -30,6 +30,11 @@ def unlock(self) -> None:
         """Unlock the contained model, and remove it from VRAM."""
         pass
 
+    @abstractmethod
+    def get_state_dict(self) -> Optional[Dict[str, torch.Tensor]]:
+        """Return the state dict (if any) for the cached model."""
+        pass
+
     @property
     @abstractmethod
     def model(self) -> AnyModel:
@@ -56,6 +61,11 @@ class CacheRecord(Generic[T]):
     and then injected into the model. When the model is finished, the VRAM
     copy of the state dict is deleted, and the RAM version is reinjected
     into the model.
+
+    The state_dict should be treated as a read-only attribute. Do not attempt
+    to patch or otherwise modify it. Instead, patch the copy of the state_dict
+    after it is loaded into the execution device (e.g. CUDA) using the `LoadedModel`
+    context manager call `model_on_device()`.
     """
 
     key: str

invokeai/backend/model_manager/load/model_cache/model_locker.py

@@ -2,6 +2,8 @@
 Base class and implementation of a class that moves models in and out of VRAM.
 """
 
+from typing import Dict, Optional
+
 import torch
 
 from invokeai.backend.model_manager import AnyModel
@@ -27,6 +29,10 @@ def model(self) -> AnyModel:
         """Return the model without moving it around."""
         return self._cache_entry.model
 
+    def get_state_dict(self) -> Optional[Dict[str, torch.Tensor]]:
+        """Return the state dict (if any) for the cached model."""
+        return self._cache_entry.state_dict
+
     def lock(self) -> AnyModel:
         """Move the model into the execution device (GPU) and lock it."""
         if not hasattr(self.model, "to"):
@@ -37,10 +43,8 @@ def lock(self) -> AnyModel:
         try:
             if self._cache.lazy_offloading:
                 self._cache.offload_unlocked_models(self._cache_entry.size)
-
             self._cache.move_model_to_device(self._cache_entry, self._cache.execution_device)
             self._cache_entry.loaded = True
-
             self._cache.logger.debug(f"Locking {self._cache_entry.key} in {self._cache.execution_device}")
             self._cache.print_cuda_stats()
         except torch.cuda.OutOfMemoryError:

invokeai/backend/model_patcher.py

@@ -5,7 +5,7 @@
 
 import pickle
 from contextlib import contextmanager
-from typing import Any, Dict, Iterator, List, Optional, Tuple, Union
+from typing import Any, Dict, Generator, Iterator, List, Optional, Tuple, Union
 
 import numpy as np
 import torch
@@ -66,8 +66,14 @@ def apply_lora_unet(
         cls,
         unet: UNet2DConditionModel,
         loras: Iterator[Tuple[LoRAModelRaw, float]],
+        model_state_dict: Optional[Dict[str, torch.Tensor]] = None,
     ) -> None:
-        with cls.apply_lora(unet, loras, "lora_unet_"):
+        with cls.apply_lora(
+            unet,
+            loras=loras,
+            prefix="lora_unet_",
+            model_state_dict=model_state_dict,
+        ):
             yield
 
     @classmethod
@@ -76,28 +82,9 @@ def apply_lora_text_encoder(
         cls,
         text_encoder: CLIPTextModel,
         loras: Iterator[Tuple[LoRAModelRaw, float]],
+        model_state_dict: Optional[Dict[str, torch.Tensor]] = None,
     ) -> None:
-        with cls.apply_lora(text_encoder, loras, "lora_te_"):
-            yield
-
-    @classmethod
-    @contextmanager
-    def apply_sdxl_lora_text_encoder(
-        cls,
-        text_encoder: CLIPTextModel,
-        loras: List[Tuple[LoRAModelRaw, float]],
-    ) -> None:
-        with cls.apply_lora(text_encoder, loras, "lora_te1_"):
-            yield
-
-    @classmethod
-    @contextmanager
-    def apply_sdxl_lora_text_encoder2(
-        cls,
-        text_encoder: CLIPTextModel,
-        loras: List[Tuple[LoRAModelRaw, float]],
-    ) -> None:
-        with cls.apply_lora(text_encoder, loras, "lora_te2_"):
+        with cls.apply_lora(text_encoder, loras=loras, prefix="lora_te_", model_state_dict=model_state_dict):
             yield
 
     @classmethod
@@ -107,7 +94,16 @@ def apply_lora(
         model: AnyModel,
         loras: Iterator[Tuple[LoRAModelRaw, float]],
         prefix: str,
-    ) -> None:
+        model_state_dict: Optional[Dict[str, torch.Tensor]] = None,
+    ) -> Generator[Any, None, None]:
+        """
+        Apply one or more LoRAs to a model.
+
+        :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.
+        """
         original_weights = {}
         try:
             with torch.no_grad():
@@ -133,7 +129,10 @@ def apply_lora(
                         dtype = module.weight.dtype
 
                         if module_key not in original_weights:
-                            original_weights[module_key] = module.weight.detach().to(device="cpu", copy=True)
+                            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