Fuse loras (huggingface#4473)

* Fuse loras

* initial implementation.

* add slow test one.

* styling

* add: test for checking efficiency

* print

* position

* place model offload correctly

* style

* style.

* unfuse test.

* final checks

* remove warning test

* remove warnings altogether

* debugging

* tighten up tests.

* debugging

* debugging

* debugging

* debugging

* debugging

* debugging

* debugging

* debugging

* debugging

* denugging

* debugging

* debugging

* debugging

* debugging

* debugging

* debugging

* debugging

* debugging

* debugging

* debugging

* debugging

* debugging

* debugging

* debugging

* debuging

* debugging

* debugging

* debugging

* suit up the generator initialization a bit.

* remove print

* update assertion.

* debugging

* remove print.

* fix: assertions.

* style

* can generator be a problem?

* generator

* correct tests.

* support text encoder lora fusion.

* tighten up tests.

---------

Co-authored-by: Sayak Paul <[email protected]>

Author: patrickvonplaten

loaders.py

@@ -85,7 +85,49 @@ def __init__(self, regular_linear_layer, lora_scale=1, network_alpha=None, rank=
 
         self.lora_scale = lora_scale
 
+    def _fuse_lora(self):
+        if self.lora_linear_layer is None:
+            return
+
+        dtype, device = self.regular_linear_layer.weight.data.dtype, self.regular_linear_layer.weight.data.device
+        logger.info(f"Fusing LoRA weights for {self.__class__}")
+
+        w_orig = self.regular_linear_layer.weight.data.float()
+        w_up = self.lora_linear_layer.up.weight.data.float()
+        w_down = self.lora_linear_layer.down.weight.data.float()
+
+        if self.lora_linear_layer.network_alpha is not None:
+            w_up = w_up * self.lora_linear_layer.network_alpha / self.lora_linear_layer.rank
+
+        fused_weight = w_orig + torch.bmm(w_up[None, :], w_down[None, :])[0]
+        self.regular_linear_layer.weight.data = fused_weight.to(device=device, dtype=dtype)
+
+        # we can drop the lora layer now
+        self.lora_linear_layer = None
+
+        # offload the up and down matrices to CPU to not blow the memory
+        self.w_up = w_up.cpu()
+        self.w_down = w_down.cpu()
+
+    def _unfuse_lora(self):
+        if not (hasattr(self, "w_up") and hasattr(self, "w_down")):
+            return
+        logger.info(f"Unfusing LoRA weights for {self.__class__}")
+
+        fused_weight = self.regular_linear_layer.weight.data
+        dtype, device = fused_weight.dtype, fused_weight.device
+
+        self.w_up = self.w_up.to(device=device, dtype=dtype)
+        self.w_down = self.w_down.to(device, dtype=dtype)
+        unfused_weight = fused_weight - torch.bmm(self.w_up[None, :], self.w_down[None, :])[0]
+        self.regular_linear_layer.weight.data = unfused_weight.to(device=device, dtype=dtype)
+
+        self.w_up = None
+        self.w_down = None
+
     def forward(self, input):
+        if self.lora_linear_layer is None:
+            return self.regular_linear_layer(input)
         return self.regular_linear_layer(input) + self.lora_scale * self.lora_linear_layer(input)
 
 
@@ -525,6 +567,20 @@ def save_function(weights, filename):
         save_function(state_dict, os.path.join(save_directory, weight_name))
         logger.info(f"Model weights saved in {os.path.join(save_directory, weight_name)}")
 
+    def fuse_lora(self):
+        self.apply(self._fuse_lora_apply)
+
+    def _fuse_lora_apply(self, module):
+        if hasattr(module, "_fuse_lora"):
+            module._fuse_lora()
+
+    def unfuse_lora(self):
+        self.apply(self._unfuse_lora_apply)
+
+    def _unfuse_lora_apply(self, module):
+        if hasattr(module, "_unfuse_lora"):
+            module._unfuse_lora()
+
 
 class TextualInversionLoaderMixin:
     r"""
@@ -1712,6 +1768,83 @@ def unload_lora_weights(self):
         # Safe to call the following regardless of LoRA.
         self._remove_text_encoder_monkey_patch()
 
+    def fuse_lora(self, fuse_unet: bool = True, fuse_text_encoder: bool = True):
+        r"""
+        Fuses the LoRA parameters into the original parameters of the corresponding blocks.
+
+        <Tip warning={true}>
+
+        This is an experimental API.
+
+        </Tip>
+
+        Args:
+            fuse_unet (`bool`, defaults to `True`): Whether to fuse the UNet LoRA parameters.
+            fuse_text_encoder (`bool`, defaults to `True`):
+                Whether to fuse the text encoder LoRA parameters. If the text encoder wasn't monkey-patched with the
+                LoRA parameters then it won't have any effect.
+        """
+        if fuse_unet:
+            self.unet.fuse_lora()
+
+        def fuse_text_encoder_lora(text_encoder):
+            for _, attn_module in text_encoder_attn_modules(text_encoder):
+                if isinstance(attn_module.q_proj, PatchedLoraProjection):
+                    attn_module.q_proj._fuse_lora()
+                    attn_module.k_proj._fuse_lora()
+                    attn_module.v_proj._fuse_lora()
+                    attn_module.out_proj._fuse_lora()
+
+            for _, mlp_module in text_encoder_mlp_modules(text_encoder):
+                if isinstance(mlp_module.fc1, PatchedLoraProjection):
+                    mlp_module.fc1._fuse_lora()
+                    mlp_module.fc2._fuse_lora()
+
+        if fuse_text_encoder:
+            if hasattr(self, "text_encoder"):
+                fuse_text_encoder_lora(self.text_encoder)
+            if hasattr(self, "text_encoder_2"):
+                fuse_text_encoder_lora(self.text_encoder_2)
+
+    def unfuse_lora(self, unfuse_unet: bool = True, unfuse_text_encoder: bool = True):
+        r"""
+        Reverses the effect of
+        [`pipe.fuse_lora()`](https://huggingface.co/docs/diffusers/main/en/api/loaders#diffusers.loaders.LoraLoaderMixin.fuse_lora).
+
+        <Tip warning={true}>
+
+        This is an experimental API.
+
+        </Tip>
+
+        Args:
+            unfuse_unet (`bool`, defaults to `True`): Whether to unfuse the UNet LoRA parameters.
+            unfuse_text_encoder (`bool`, defaults to `True`):
+                Whether to unfuse the text encoder LoRA parameters. If the text encoder wasn't monkey-patched with the
+                LoRA parameters then it won't have any effect.
+        """
+        if unfuse_unet:
+            self.unet.unfuse_lora()
+
+        def unfuse_text_encoder_lora(text_encoder):
+            for _, attn_module in text_encoder_attn_modules(text_encoder):
+                if isinstance(attn_module.q_proj, PatchedLoraProjection):
+                    attn_module.q_proj._unfuse_lora()
+                    attn_module.k_proj._unfuse_lora()
+                    attn_module.v_proj._unfuse_lora()
+                    attn_module.out_proj._unfuse_lora()
+
+            for _, mlp_module in text_encoder_mlp_modules(text_encoder):
+                if isinstance(mlp_module.fc1, PatchedLoraProjection):
+                    mlp_module.fc1._unfuse_lora()
+                    mlp_module.fc2._unfuse_lora()
+
+        if unfuse_text_encoder:
+            if hasattr(self, "text_encoder"):
+                unfuse_text_encoder_lora(self.text_encoder)
+            if hasattr(self, "text_encoder_2"):
+                unfuse_text_encoder_lora(self.text_encoder_2)
+
 
 class FromSingleFileMixin:
     """

models/lora.py

@@ -14,9 +14,15 @@
 
 from typing import Optional
 
+import torch
 import torch.nn.functional as F
 from torch import nn
 
+from ..utils import logging
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
 
 class LoRALinearLayer(nn.Module):
     def __init__(self, in_features, out_features, rank=4, network_alpha=None, device=None, dtype=None):
@@ -91,6 +97,51 @@ def __init__(self, *args, lora_layer: Optional[LoRAConv2dLayer] = None, **kwargs
     def set_lora_layer(self, lora_layer: Optional[LoRAConv2dLayer]):
         self.lora_layer = lora_layer
 
+    def _fuse_lora(self):
+        if self.lora_layer is None:
+            return
+
+        dtype, device = self.weight.data.dtype, self.weight.data.device
+        logger.info(f"Fusing LoRA weights for {self.__class__}")
+
+        w_orig = self.weight.data.float()
+        w_up = self.lora_layer.up.weight.data.float()
+        w_down = self.lora_layer.down.weight.data.float()
+
+        if self.lora_layer.network_alpha is not None:
+            w_up = w_up * self.lora_layer.network_alpha / self.lora_layer.rank
+
+        fusion = torch.mm(w_up.flatten(start_dim=1), w_down.flatten(start_dim=1))
+        fusion = fusion.reshape((w_orig.shape))
+        fused_weight = w_orig + fusion
+        self.weight.data = fused_weight.to(device=device, dtype=dtype)
+
+        # we can drop the lora layer now
+        self.lora_layer = None
+
+        # offload the up and down matrices to CPU to not blow the memory
+        self.w_up = w_up.cpu()
+        self.w_down = w_down.cpu()
+
+    def _unfuse_lora(self):
+        if not (hasattr(self, "w_up") and hasattr(self, "w_down")):
+            return
+        logger.info(f"Unfusing LoRA weights for {self.__class__}")
+
+        fused_weight = self.weight.data
+        dtype, device = fused_weight.data.dtype, fused_weight.data.device
+
+        self.w_up = self.w_up.to(device=device, dtype=dtype)
+        self.w_down = self.w_down.to(device, dtype=dtype)
+
+        fusion = torch.mm(self.w_up.flatten(start_dim=1), self.w_down.flatten(start_dim=1))
+        fusion = fusion.reshape((fused_weight.shape))
+        unfused_weight = fused_weight - fusion
+        self.weight.data = unfused_weight.to(device=device, dtype=dtype)
+
+        self.w_up = None
+        self.w_down = None
+
     def forward(self, x):
         if self.lora_layer is None:
             # make sure to the functional Conv2D function as otherwise torch.compile's graph will break
@@ -109,9 +160,49 @@ def __init__(self, *args, lora_layer: Optional[LoRALinearLayer] = None, **kwargs
         super().__init__(*args, **kwargs)
         self.lora_layer = lora_layer
 
-    def set_lora_layer(self, lora_layer: Optional[LoRAConv2dLayer]):
+    def set_lora_layer(self, lora_layer: Optional[LoRALinearLayer]):
         self.lora_layer = lora_layer
 
+    def _fuse_lora(self):
+        if self.lora_layer is None:
+            return
+
+        dtype, device = self.weight.data.dtype, self.weight.data.device
+        logger.info(f"Fusing LoRA weights for {self.__class__}")
+
+        w_orig = self.weight.data.float()
+        w_up = self.lora_layer.up.weight.data.float()
+        w_down = self.lora_layer.down.weight.data.float()
+
+        if self.lora_layer.network_alpha is not None:
+            w_up = w_up * self.lora_layer.network_alpha / self.lora_layer.rank
+
+        fused_weight = w_orig + torch.bmm(w_up[None, :], w_down[None, :])[0]
+        self.weight.data = fused_weight.to(device=device, dtype=dtype)
+
+        # we can drop the lora layer now
+        self.lora_layer = None
+
+        # offload the up and down matrices to CPU to not blow the memory
+        self.w_up = w_up.cpu()
+        self.w_down = w_down.cpu()
+
+    def _unfuse_lora(self):
+        if not (hasattr(self, "w_up") and hasattr(self, "w_down")):
+            return
+        logger.info(f"Unfusing LoRA weights for {self.__class__}")
+
+        fused_weight = self.weight.data
+        dtype, device = fused_weight.dtype, fused_weight.device
+
+        self.w_up = self.w_up.to(device=device, dtype=dtype)
+        self.w_down = self.w_down.to(device, dtype=dtype)
+        unfused_weight = fused_weight - torch.bmm(self.w_up[None, :], self.w_down[None, :])[0]
+        self.weight.data = unfused_weight.to(device=device, dtype=dtype)
+
+        self.w_up = None
+        self.w_down = None
+
     def forward(self, hidden_states, lora_scale: int = 1):
         if self.lora_layer is None:
             return super().forward(hidden_states)