Implements Blockwise lora (huggingface#7352)

* Initial commit * Implemented block lora - implemented block lora - updated docs - added tests * Finishing up * Reverted unrelated changes made by make style * Fixed typo * Fixed bug + Made text_encoder_2 scalable * Integrated some review feedback * Incorporated review feedback * Fix tests * Made every module configurable * Adapter to new lora test structure * Final cleanup * Some more final fixes - Included examples in `using_peft_for_inference.md` - Added hint that only attns are scaled - Removed NoneTypes - Added test to check mismatching lens of adapter names / weights raise error * Update using_peft_for_inference.md * Update using_peft_for_inference.md * Make style, quality, fix-copies * Updated tutorial;Warning if scale/adapter mismatch * floats are forwarded as-is; changed tutorial scale * make style, quality, fix-copies * Fixed typo in tutorial * Moved some warnings into `lora_loader_utils.py` * Moved scale/lora mismatch warnings back * Integrated final review suggestions * Empty commit to trigger CI * Reverted emoty commit to trigger CI --------- Co-authored-by: Sayak Paul <[email protected]>
noskill · Apr 5, 2024 · 0a14b9d · 0a14b9d
1 parent 14f6d2d
commit 0a14b9d
Show file tree

Hide file tree

Showing 7 changed files with 553 additions and 21 deletions.
diff --git a/docs/source/en/tutorials/using_peft_for_inference.md b/docs/source/en/tutorials/using_peft_for_inference.md
@@ -133,6 +133,62 @@ image
 
 ![no-lora](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/peft_integration/diffusers_peft_lora_inference_20_1.png)
 
+### Customize adapters strength
+For even more customization, you can control how strongly the adapter affects each part of the pipeline. For this, pass a dictionary with the control strengths (called "scales") to [`~diffusers.loaders.UNet2DConditionLoadersMixin.set_adapters`].
+
+For example, here's how you can turn on the adapter for the `down` parts, but turn it off for the `mid` and `up` parts:
+```python
+pipe.enable_lora()  # enable lora again, after we disabled it above
+prompt = "toy_face of a hacker with a hoodie, pixel art"
+adapter_weight_scales = { "unet": { "down": 1, "mid": 0, "up": 0} }
+pipe.set_adapters("pixel", adapter_weight_scales)
+image = pipe(prompt, num_inference_steps=30, generator=torch.manual_seed(0)).images[0]
+image
+```
+
+![block-lora-text-and-down](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/peft_integration/diffusers_peft_lora_inference_block_down.png)
+
+Let's see how turning off the `down` part and turning on the `mid` and `up` part respectively changes the image.
+```python
+adapter_weight_scales = { "unet": { "down": 0, "mid": 1, "up": 0} }
+pipe.set_adapters("pixel", adapter_weight_scales)
+image = pipe(prompt, num_inference_steps=30, generator=torch.manual_seed(0)).images[0]
+image
+```
+
+![block-lora-text-and-mid](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/peft_integration/diffusers_peft_lora_inference_block_mid.png)
+
+```python
+adapter_weight_scales = { "unet": { "down": 0, "mid": 0, "up": 1} }
+pipe.set_adapters("pixel", adapter_weight_scales)
+image = pipe(prompt, num_inference_steps=30, generator=torch.manual_seed(0)).images[0]
+image
+```
+
+![block-lora-text-and-up](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/peft_integration/diffusers_peft_lora_inference_block_up.png)
+
+Looks cool!
+
+This is a really powerful feature. You can use it to control the adapter strengths down to per-transformer level. And you can even use it for multiple adapters.
+```python
+adapter_weight_scales_toy = 0.5
+adapter_weight_scales_pixel = {
+    "unet": {
+        "down": 0.9,  # all transformers in the down-part will use scale 0.9
+        # "mid"  # because, in this example, "mid" is not given, all transformers in the mid part will use the default scale 1.0
+        "up": {
+            "block_0": 0.6,  # all 3 transformers in the 0th block in the up-part will use scale 0.6
+            "block_1": [0.4, 0.8, 1.0],  # the 3 transformers in the 1st block in the up-part will use scales 0.4, 0.8 and 1.0 respectively
+        }
+    }
+}
+pipe.set_adapters(["toy", "pixel"], [adapter_weight_scales_toy, adapter_weight_scales_pixel])
+image = pipe(prompt, num_inference_steps=30, generator=torch.manual_seed(0)).images[0]
+image
+```
+
+![block-lora-mixed](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/peft_integration/diffusers_peft_lora_inference_block_mixed.png)
+
 ## Manage active adapters
 
 You have attached multiple adapters in this tutorial, and if you're feeling a bit lost on what adapters have been attached to the pipeline's components, use the [`~diffusers.loaders.LoraLoaderMixin.get_active_adapters`] method to check the list of active adapters:

diff --git a/docs/source/en/using-diffusers/loading_adapters.md b/docs/source/en/using-diffusers/loading_adapters.md
@@ -153,18 +153,43 @@ image
     <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/load_attn_proc.png" />
 </div>
 
-<Tip>
-
-For both [`~loaders.LoraLoaderMixin.load_lora_weights`] and [`~loaders.UNet2DConditionLoadersMixin.load_attn_procs`], you can pass the `cross_attention_kwargs={"scale": 0.5}` parameter to adjust how much of the LoRA weights to use. A value of `0` is the same as only using the base model weights, and a value of `1` is equivalent to using the fully finetuned LoRA.
-
-</Tip>
-
 To unload the LoRA weights, use the [`~loaders.LoraLoaderMixin.unload_lora_weights`] method to discard the LoRA weights and restore the model to its original weights:
 
 ```py
 pipeline.unload_lora_weights()
 ```
 
+### Adjust LoRA weight scale
+
+For both [`~loaders.LoraLoaderMixin.load_lora_weights`] and [`~loaders.UNet2DConditionLoadersMixin.load_attn_procs`], you can pass the `cross_attention_kwargs={"scale": 0.5}` parameter to adjust how much of the LoRA weights to use. A value of `0` is the same as only using the base model weights, and a value of `1` is equivalent to using the fully finetuned LoRA.
+
+For more granular control on the amount of LoRA weights used per layer, you can use [`~loaders.LoraLoaderMixin.set_adapters`] and pass a dictionary specifying by how much to scale the weights in each layer by.
+```python
+pipe = ... # create pipeline
+pipe.load_lora_weights(..., adapter_name="my_adapter") 
+scales = {
+    "text_encoder": 0.5,
+    "text_encoder_2": 0.5,  # only usable if pipe has a 2nd text encoder
+    "unet": {
+        "down": 0.9,  # all transformers in the down-part will use scale 0.9
+        # "mid"  # in this example "mid" is not given, therefore all transformers in the mid part will use the default scale 1.0
+        "up": {
+            "block_0": 0.6,  # all 3 transformers in the 0th block in the up-part will use scale 0.6
+            "block_1": [0.4, 0.8, 1.0],  # the 3 transformers in the 1st block in the up-part will use scales 0.4, 0.8 and 1.0 respectively
+        }
+    }
+}
+pipe.set_adapters("my_adapter", scales)
+```
+
+This also works with multiple adapters - see [this guide](https://huggingface.co/docs/diffusers/tutorials/using_peft_for_inference#customize-adapters-strength) for how to do it.
+
+<Tip warning={true}>
+
+Currently, [`~loaders.LoraLoaderMixin.set_adapters`] only supports scaling attention weights. If a LoRA has other parts (e.g., resnets or down-/upsamplers), they will keep a scale of 1.0.
+
+</Tip>
+
 ### Kohya and TheLastBen
 
 Other popular LoRA trainers from the community include those by [Kohya](https://github.com/kohya-ss/sd-scripts/) and [TheLastBen](https://github.com/TheLastBen/fast-stable-diffusion). These trainers create different LoRA checkpoints than those trained by 🤗 Diffusers, but they can still be loaded in the same way.

diff --git a/src/diffusers/loaders/lora.py b/src/diffusers/loaders/lora.py
@@ -11,6 +11,7 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
+import copy
 import inspect
 import os
 from pathlib import Path
@@ -985,7 +986,7 @@ def set_adapters_for_text_encoder(
         self,
         adapter_names: Union[List[str], str],
         text_encoder: Optional["PreTrainedModel"] = None,  # noqa: F821
-        text_encoder_weights: List[float] = None,
+        text_encoder_weights: Optional[Union[float, List[float], List[None]]] = None,
     ):
         """
         Sets the adapter layers for the text encoder.
@@ -1003,15 +1004,20 @@ def set_adapters_for_text_encoder(
             raise ValueError("PEFT backend is required for this method.")
 
         def process_weights(adapter_names, weights):
-            if weights is None:
-                weights = [1.0] * len(adapter_names)
-            elif isinstance(weights, float):
-                weights = [weights]
+            # Expand weights into a list, one entry per adapter
+            # e.g. for 2 adapters:  7 -> [7,7] ; [3, None] -> [3, None]
+            if not isinstance(weights, list):
+                weights = [weights] * len(adapter_names)
 
             if len(adapter_names) != len(weights):
                 raise ValueError(
                     f"Length of adapter names {len(adapter_names)} is not equal to the length of the weights {len(weights)}"
                 )
+
+            # Set None values to default of 1.0
+            # e.g. [7,7] -> [7,7] ; [3, None] -> [3,1]
+            weights = [w if w is not None else 1.0 for w in weights]
+
             return weights
 
         adapter_names = [adapter_names] if isinstance(adapter_names, str) else adapter_names
@@ -1059,17 +1065,77 @@ def enable_lora_for_text_encoder(self, text_encoder: Optional["PreTrainedModel"]
     def set_adapters(
         self,
         adapter_names: Union[List[str], str],
-        adapter_weights: Optional[List[float]] = None,
+        adapter_weights: Optional[Union[float, Dict, List[float], List[Dict]]] = None,
     ):
+        adapter_names = [adapter_names] if isinstance(adapter_names, str) else adapter_names
+
+        adapter_weights = copy.deepcopy(adapter_weights)
+
+        # Expand weights into a list, one entry per adapter
+        if not isinstance(adapter_weights, list):
+            adapter_weights = [adapter_weights] * len(adapter_names)
+
+        if len(adapter_names) != len(adapter_weights):
+            raise ValueError(
+                f"Length of adapter names {len(adapter_names)} is not equal to the length of the weights {len(adapter_weights)}"
+            )
+
+        # Decompose weights into weights for unet, text_encoder and text_encoder_2
+        unet_lora_weights, text_encoder_lora_weights, text_encoder_2_lora_weights = [], [], []
+
+        list_adapters = self.get_list_adapters()  # eg {"unet": ["adapter1", "adapter2"], "text_encoder": ["adapter2"]}
+        all_adapters = {
+            adapter for adapters in list_adapters.values() for adapter in adapters
+        }  # eg ["adapter1", "adapter2"]
+        invert_list_adapters = {
+            adapter: [part for part, adapters in list_adapters.items() if adapter in adapters]
+            for adapter in all_adapters
+        }  # eg {"adapter1": ["unet"], "adapter2": ["unet", "text_encoder"]}
+
+        for adapter_name, weights in zip(adapter_names, adapter_weights):
+            if isinstance(weights, dict):
+                unet_lora_weight = weights.pop("unet", None)
+                text_encoder_lora_weight = weights.pop("text_encoder", None)
+                text_encoder_2_lora_weight = weights.pop("text_encoder_2", None)
+
+                if len(weights) > 0:
+                    raise ValueError(
+                        f"Got invalid key '{weights.keys()}' in lora weight dict for adapter {adapter_name}."
+                    )
+
+                if text_encoder_2_lora_weight is not None and not hasattr(self, "text_encoder_2"):
+                    logger.warning(
+                        "Lora weight dict contains text_encoder_2 weights but will be ignored because pipeline does not have text_encoder_2."
+                    )
+
+                # warn if adapter doesn't have parts specified by adapter_weights
+                for part_weight, part_name in zip(
+                    [unet_lora_weight, text_encoder_lora_weight, text_encoder_2_lora_weight],
+                    ["uent", "text_encoder", "text_encoder_2"],
+                ):
+                    if part_weight is not None and part_name not in invert_list_adapters[adapter_name]:
+                        logger.warning(
+                            f"Lora weight dict for adapter '{adapter_name}' contains {part_name}, but this will be ignored because {adapter_name} does not contain weights for {part_name}. Valid parts for {adapter_name} are: {invert_list_adapters[adapter_name]}."
+                        )
+
+            else:
+                unet_lora_weight = weights
+                text_encoder_lora_weight = weights
+                text_encoder_2_lora_weight = weights
+
+            unet_lora_weights.append(unet_lora_weight)
+            text_encoder_lora_weights.append(text_encoder_lora_weight)
+            text_encoder_2_lora_weights.append(text_encoder_2_lora_weight)
+
         unet = getattr(self, self.unet_name) if not hasattr(self, "unet") else self.unet
         # Handle the UNET
-        unet.set_adapters(adapter_names, adapter_weights)
+        unet.set_adapters(adapter_names, unet_lora_weights)
 
         # Handle the Text Encoder
         if hasattr(self, "text_encoder"):
-            self.set_adapters_for_text_encoder(adapter_names, self.text_encoder, adapter_weights)
+            self.set_adapters_for_text_encoder(adapter_names, self.text_encoder, text_encoder_lora_weights)
         if hasattr(self, "text_encoder_2"):
-            self.set_adapters_for_text_encoder(adapter_names, self.text_encoder_2, adapter_weights)
+            self.set_adapters_for_text_encoder(adapter_names, self.text_encoder_2, text_encoder_2_lora_weights)
 
     def disable_lora(self):
         if not USE_PEFT_BACKEND:

diff --git a/src/diffusers/loaders/unet.py b/src/diffusers/loaders/unet.py
@@ -47,6 +47,7 @@
     infer_stable_cascade_single_file_config,
     load_single_file_model_checkpoint,
 )
+from .unet_loader_utils import _maybe_expand_lora_scales
 from .utils import AttnProcsLayers
 
 
@@ -564,7 +565,7 @@ def _unfuse_lora_apply(self, module):
     def set_adapters(
         self,
         adapter_names: Union[List[str], str],
-        weights: Optional[Union[List[float], float]] = None,
+        weights: Optional[Union[float, Dict, List[float], List[Dict], List[None]]] = None,
     ):
         """
         Set the currently active adapters for use in the UNet.
@@ -597,16 +598,23 @@ def set_adapters(
 
         adapter_names = [adapter_names] if isinstance(adapter_names, str) else adapter_names
 
-        if weights is None:
-            weights = [1.0] * len(adapter_names)
-        elif isinstance(weights, float):
+        # Expand weights into a list, one entry per adapter
+        # examples for e.g. 2 adapters:  [{...}, 7] -> [7,7] ; None -> [None, None]
+        if not isinstance(weights, list):
             weights = [weights] * len(adapter_names)
 
         if len(adapter_names) != len(weights):
             raise ValueError(
                 f"Length of adapter names {len(adapter_names)} is not equal to the length of their weights {len(weights)}."
             )
 
+        # Set None values to default of 1.0
+        # e.g. [{...}, 7] -> [{...}, 7] ; [None, None] -> [1.0, 1.0]
+        weights = [w if w is not None else 1.0 for w in weights]
+
+        # e.g. [{...}, 7] -> [{expanded dict...}, 7]
+        weights = _maybe_expand_lora_scales(self, weights)
+
         set_weights_and_activate_adapters(self, adapter_names, weights)
 
     def disable_lora(self):