Improve RAM<->VRAM memory copy performance in LoRA patching and elsewhere (#6490)

* 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 requested during penultimate review

* add non-blocking=True parameters to several torch.nn.Module.to() calls, for slight performance increases

* fix ruff errors

* prevent crash on non-cuda-enabled systems

---------

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

invokeai/backend/ip_adapter/ip_adapter.py

@@ -125,13 +125,16 @@ def __init__(
             self.device, dtype=self.dtype
         )
 
-    def to(self, device: torch.device, dtype: Optional[torch.dtype] = None):
-        self.device = device
+    def to(
+        self, device: Optional[torch.device] = None, dtype: Optional[torch.dtype] = None, non_blocking: bool = False
+    ):
+        if device is not None:
+            self.device = device
         if dtype is not None:
             self.dtype = dtype
 
-        self._image_proj_model.to(device=self.device, dtype=self.dtype)
-        self.attn_weights.to(device=self.device, dtype=self.dtype)
+        self._image_proj_model.to(device=self.device, dtype=self.dtype, non_blocking=non_blocking)
+        self.attn_weights.to(device=self.device, dtype=self.dtype, non_blocking=non_blocking)
 
     def calc_size(self):
         # workaround for circular import

invokeai/backend/lora.py

@@ -61,9 +61,10 @@ def to(
         self,
         device: Optional[torch.device] = None,
         dtype: Optional[torch.dtype] = None,
+        non_blocking: bool = False,
     ) -> None:
         if self.bias is not None:
-            self.bias = self.bias.to(device=device, dtype=dtype)
+            self.bias = self.bias.to(device=device, dtype=dtype, non_blocking=non_blocking)
 
 
 # TODO: find and debug lora/locon with bias
@@ -109,14 +110,15 @@ def to(
         self,
         device: Optional[torch.device] = None,
         dtype: Optional[torch.dtype] = None,
+        non_blocking: bool = False,
     ) -> None:
-        super().to(device=device, dtype=dtype)
+        super().to(device=device, dtype=dtype, non_blocking=non_blocking)
 
-        self.up = self.up.to(device=device, dtype=dtype)
-        self.down = self.down.to(device=device, dtype=dtype)
+        self.up = self.up.to(device=device, dtype=dtype, non_blocking=non_blocking)
+        self.down = self.down.to(device=device, dtype=dtype, non_blocking=non_blocking)
 
         if self.mid is not None:
-            self.mid = self.mid.to(device=device, dtype=dtype)
+            self.mid = self.mid.to(device=device, dtype=dtype, non_blocking=non_blocking)
 
 
 class LoHALayer(LoRALayerBase):
@@ -169,18 +171,19 @@ def to(
         self,
         device: Optional[torch.device] = None,
         dtype: Optional[torch.dtype] = None,
+        non_blocking: bool = False,
     ) -> 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)
+        self.w1_a = self.w1_a.to(device=device, dtype=dtype, non_blocking=non_blocking)
+        self.w1_b = self.w1_b.to(device=device, dtype=dtype, non_blocking=non_blocking)
         if self.t1 is not None:
-            self.t1 = self.t1.to(device=device, dtype=dtype)
+            self.t1 = self.t1.to(device=device, dtype=dtype, non_blocking=non_blocking)
 
-        self.w2_a = self.w2_a.to(device=device, dtype=dtype)
-        self.w2_b = self.w2_b.to(device=device, dtype=dtype)
+        self.w2_a = self.w2_a.to(device=device, dtype=dtype, non_blocking=non_blocking)
+        self.w2_b = self.w2_b.to(device=device, dtype=dtype, non_blocking=non_blocking)
         if self.t2 is not None:
-            self.t2 = self.t2.to(device=device, dtype=dtype)
+            self.t2 = self.t2.to(device=device, dtype=dtype, non_blocking=non_blocking)
 
 
 class LoKRLayer(LoRALayerBase):
@@ -265,6 +268,7 @@ def to(
         self,
         device: Optional[torch.device] = None,
         dtype: Optional[torch.dtype] = None,
+        non_blocking: bool = False,
     ) -> None:
         super().to(device=device, dtype=dtype)
 
@@ -273,19 +277,19 @@ def to(
         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)
+            self.w1_a = self.w1_a.to(device=device, dtype=dtype, non_blocking=non_blocking)
+            self.w1_b = self.w1_b.to(device=device, dtype=dtype, non_blocking=non_blocking)
 
         if self.w2 is not None:
-            self.w2 = self.w2.to(device=device, dtype=dtype)
+            self.w2 = self.w2.to(device=device, dtype=dtype, non_blocking=non_blocking)
         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)
+            self.w2_a = self.w2_a.to(device=device, dtype=dtype, non_blocking=non_blocking)
+            self.w2_b = self.w2_b.to(device=device, dtype=dtype, non_blocking=non_blocking)
 
         if self.t2 is not None:
-            self.t2 = self.t2.to(device=device, dtype=dtype)
+            self.t2 = self.t2.to(device=device, dtype=dtype, non_blocking=non_blocking)
 
 
 class FullLayer(LoRALayerBase):
@@ -319,10 +323,11 @@ def to(
         self,
         device: Optional[torch.device] = None,
         dtype: Optional[torch.dtype] = None,
+        non_blocking: bool = False,
     ) -> None:
         super().to(device=device, dtype=dtype)
 
-        self.weight = self.weight.to(device=device, dtype=dtype)
+        self.weight = self.weight.to(device=device, dtype=dtype, non_blocking=non_blocking)
 
 
 class IA3Layer(LoRALayerBase):
@@ -358,11 +363,12 @@ def to(
         self,
         device: Optional[torch.device] = None,
         dtype: Optional[torch.dtype] = None,
+        non_blocking: bool = False,
     ):
         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)
+        self.weight = self.weight.to(device=device, dtype=dtype, non_blocking=non_blocking)
+        self.on_input = self.on_input.to(device=device, dtype=dtype, non_blocking=non_blocking)
 
 
 AnyLoRALayer = Union[LoRALayer, LoHALayer, LoKRLayer, FullLayer, IA3Layer]
@@ -388,10 +394,11 @@ def to(
         self,
         device: Optional[torch.device] = None,
         dtype: Optional[torch.dtype] = None,
+        non_blocking: bool = False,
     ) -> None:
         # TODO: try revert if exception?
         for _key, layer in self.layers.items():
-            layer.to(device=device, dtype=dtype)
+            layer.to(device=device, dtype=dtype, non_blocking=non_blocking)
 
     def calc_size(self) -> int:
         model_size = 0
@@ -514,7 +521,7 @@ def from_checkpoint(
             # lower memory consumption by removing already parsed layer values
             state_dict[layer_key].clear()
 
-            layer.to(device=device, dtype=dtype)
+            layer.to(device=device, dtype=dtype, non_blocking=True)
             model.layers[layer_key] = layer
 
         return model

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

@@ -285,9 +285,9 @@ def move_model_to_device(self, cache_entry: CacheRecord[AnyModel], target_device
                 else:
                     new_dict: Dict[str, torch.Tensor] = {}
                     for k, v in cache_entry.state_dict.items():
-                        new_dict[k] = v.to(torch.device(target_device), copy=True)
+                        new_dict[k] = v.to(torch.device(target_device), copy=True, non_blocking=True)
                     cache_entry.model.load_state_dict(new_dict, assign=True)
-            cache_entry.model.to(target_device)
+            cache_entry.model.to(target_device, non_blocking=True)
             cache_entry.device = target_device
         except Exception as e:  # blow away cache entry
             self._delete_cache_entry(cache_entry)

invokeai/backend/model_patcher.py

@@ -67,7 +67,7 @@ def apply_lora_unet(
         unet: UNet2DConditionModel,
         loras: Iterator[Tuple[LoRAModelRaw, float]],
         model_state_dict: Optional[Dict[str, torch.Tensor]] = None,
-    ) -> None:
+    ) -> Generator[None, None, None]:
         with cls.apply_lora(
             unet,
             loras=loras,
@@ -83,7 +83,7 @@ def apply_lora_text_encoder(
         text_encoder: CLIPTextModel,
         loras: Iterator[Tuple[LoRAModelRaw, float]],
         model_state_dict: Optional[Dict[str, torch.Tensor]] = None,
-    ) -> None:
+    ) -> Generator[None, None, None]:
         with cls.apply_lora(text_encoder, loras=loras, prefix="lora_te_", model_state_dict=model_state_dict):
             yield
 
@@ -95,7 +95,7 @@ def apply_lora(
         loras: Iterator[Tuple[LoRAModelRaw, float]],
         prefix: str,
         model_state_dict: Optional[Dict[str, torch.Tensor]] = None,
-    ) -> Generator[Any, None, None]:
+    ) -> Generator[None, None, None]:
         """
         Apply one or more LoRAs to a model.
 
@@ -139,12 +139,12 @@ def apply_lora(
                         # 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)
+                        layer.to(device=device, non_blocking=True)
+                        layer.to(dtype=torch.float32, non_blocking=True)
                         # 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=torch.device("cpu"))
+                        layer.to(device=torch.device("cpu"), non_blocking=True)
 
                         assert isinstance(layer_weight, torch.Tensor)  # mypy thinks layer_weight is a float|Any ??!
                         if module.weight.shape != layer_weight.shape:
@@ -153,15 +153,15 @@ def apply_lora(
                             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)
+                        module.weight += layer_weight.to(dtype=dtype, non_blocking=True)
 
             yield  # wait for context manager exit
 
         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)
+                    model.get_submodule(module_key).weight.copy_(weight, non_blocking=True)
 
     @classmethod
     @contextmanager

invokeai/backend/onnx/onnx_runtime.py

@@ -6,6 +6,7 @@
 
 import numpy as np
 import onnx
+import torch
 from onnx import numpy_helper
 from onnxruntime import InferenceSession, SessionOptions, get_available_providers
 
@@ -188,6 +189,15 @@ def __call__(self, **kwargs):
         # return self.io_binding.copy_outputs_to_cpu()
         return self.session.run(None, inputs)
 
+    # compatability with RawModel ABC
+    def to(
+        self,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+        non_blocking: bool = False,
+    ) -> None:
+        pass
+
     # compatability with diffusers load code
     @classmethod
     def from_pretrained(

invokeai/backend/raw_model.py

@@ -10,6 +10,20 @@
 that adds additional methods and attributes.
 """
 
+from abc import ABC, abstractmethod
+from typing import Optional
 
-class RawModel:
-    """Base class for 'Raw' model wrappers."""
+import torch
+
+
+class RawModel(ABC):
+    """Abstract base class for 'Raw' model wrappers."""
+
+    @abstractmethod
+    def to(
+        self,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+        non_blocking: bool = False,
+    ) -> None:
+        pass

invokeai/backend/textual_inversion.py

@@ -65,6 +65,18 @@ def from_checkpoint(
 
         return result
 
+    def to(
+        self,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+        non_blocking: bool = False,
+    ) -> None:
+        if not torch.cuda.is_available():
+            return
+        for emb in [self.embedding, self.embedding_2]:
+            if emb is not None:
+                emb.to(device=device, dtype=dtype, non_blocking=non_blocking)
+
 
 class TextualInversionManager(BaseTextualInversionManager):
     """TextualInversionManager implements the BaseTextualInversionManager ABC from the compel library."""