[LoRA] Enabling limited LoRA support for text encoder (huggingface#2918)

* add: first draft for a better LoRA enabler.

* make fix-copies.

* feat: backward compatibility.

* add: entry to the docs.

* add: tests.

* fix: docs.

* fix: norm group test for UNet3D.

* feat: add support for flat dicts.

* add depcrcation message instead of warning.

docs/source/en/api/loaders.mdx

@@ -28,3 +28,11 @@ API to load such adapter neural networks via the [`loaders.py` module](https://g
 ### UNet2DConditionLoadersMixin
 
 [[autodoc]] loaders.UNet2DConditionLoadersMixin
+
+### TextualInversionLoaderMixin
+
+[[autodoc]] loaders.TextualInversionLoaderMixin
+
+### LoraLoaderMixin
+
+[[autodoc]] loaders.LoraLoaderMixin

src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion.py

@@ -20,7 +20,7 @@
 from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer
 
 from ...configuration_utils import FrozenDict
-from ...loaders import TextualInversionLoaderMixin
+from ...loaders import LoraLoaderMixin, TextualInversionLoaderMixin
 from ...models import AutoencoderKL, UNet2DConditionModel
 from ...schedulers import KarrasDiffusionSchedulers
 from ...utils import (
@@ -53,7 +53,7 @@
 """
 
 
-class StableDiffusionPipeline(DiffusionPipeline, TextualInversionLoaderMixin):
+class StableDiffusionPipeline(DiffusionPipeline, TextualInversionLoaderMixin, LoraLoaderMixin):
     r"""
     Pipeline for text-to-image generation using Stable Diffusion.
 

src/diffusers/utils/__init__.py

@@ -30,6 +30,7 @@
     ONNX_EXTERNAL_WEIGHTS_NAME,
     ONNX_WEIGHTS_NAME,
     SAFETENSORS_WEIGHTS_NAME,
+    TEXT_ENCODER_TARGET_MODULES,
     WEIGHTS_NAME,
 )
 from .deprecation_utils import deprecate

src/diffusers/utils/constants.py

@@ -30,3 +30,4 @@
 DIFFUSERS_DYNAMIC_MODULE_NAME = "diffusers_modules"
 HF_MODULES_CACHE = os.getenv("HF_MODULES_CACHE", os.path.join(hf_cache_home, "modules"))
 DEPRECATED_REVISION_ARGS = ["fp16", "non-ema"]
+TEXT_ENCODER_TARGET_MODULES = ["q_proj", "v_proj", "k_proj", "out_proj"]

tests/test_lora_layers.py

@@ -0,0 +1,213 @@
+# coding=utf-8
+# Copyright 2023 HuggingFace Inc.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# 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 os
+import tempfile
+import unittest
+
+import torch
+import torch.nn as nn
+from transformers import CLIPTextConfig, CLIPTextModel, CLIPTokenizer
+
+from diffusers import AutoencoderKL, DDIMScheduler, StableDiffusionPipeline, UNet2DConditionModel
+from diffusers.loaders import AttnProcsLayers, LoraLoaderMixin
+from diffusers.models.attention_processor import LoRAAttnProcessor
+from diffusers.utils import TEXT_ENCODER_TARGET_MODULES, floats_tensor, torch_device
+
+
+def create_unet_lora_layers(unet: nn.Module):
+    lora_attn_procs = {}
+    for name in unet.attn_processors.keys():
+        cross_attention_dim = None if name.endswith("attn1.processor") else unet.config.cross_attention_dim
+        if name.startswith("mid_block"):
+            hidden_size = unet.config.block_out_channels[-1]
+        elif name.startswith("up_blocks"):
+            block_id = int(name[len("up_blocks.")])
+            hidden_size = list(reversed(unet.config.block_out_channels))[block_id]
+        elif name.startswith("down_blocks"):
+            block_id = int(name[len("down_blocks.")])
+            hidden_size = unet.config.block_out_channels[block_id]
+        lora_attn_procs[name] = LoRAAttnProcessor(hidden_size=hidden_size, cross_attention_dim=cross_attention_dim)
+    unet_lora_layers = AttnProcsLayers(lora_attn_procs)
+    return lora_attn_procs, unet_lora_layers
+
+
+def create_text_encoder_lora_layers(text_encoder: nn.Module):
+    text_lora_attn_procs = {}
+    for name, module in text_encoder.named_modules():
+        if any([x in name for x in TEXT_ENCODER_TARGET_MODULES]):
+            text_lora_attn_procs[name] = LoRAAttnProcessor(hidden_size=module.out_features, cross_attention_dim=None)
+    text_encoder_lora_layers = AttnProcsLayers(text_lora_attn_procs)
+    return text_encoder_lora_layers
+
+
+class LoraLoaderMixinTests(unittest.TestCase):
+    def get_dummy_components(self):
+        torch.manual_seed(0)
+        unet = UNet2DConditionModel(
+            block_out_channels=(32, 64),
+            layers_per_block=2,
+            sample_size=32,
+            in_channels=4,
+            out_channels=4,
+            down_block_types=("DownBlock2D", "CrossAttnDownBlock2D"),
+            up_block_types=("CrossAttnUpBlock2D", "UpBlock2D"),
+            cross_attention_dim=32,
+        )
+        scheduler = DDIMScheduler(
+            beta_start=0.00085,
+            beta_end=0.012,
+            beta_schedule="scaled_linear",
+            clip_sample=False,
+            set_alpha_to_one=False,
+        )
+        torch.manual_seed(0)
+        vae = AutoencoderKL(
+            block_out_channels=[32, 64],
+            in_channels=3,
+            out_channels=3,
+            down_block_types=["DownEncoderBlock2D", "DownEncoderBlock2D"],
+            up_block_types=["UpDecoderBlock2D", "UpDecoderBlock2D"],
+            latent_channels=4,
+        )
+        text_encoder_config = CLIPTextConfig(
+            bos_token_id=0,
+            eos_token_id=2,
+            hidden_size=32,
+            intermediate_size=37,
+            layer_norm_eps=1e-05,
+            num_attention_heads=4,
+            num_hidden_layers=5,
+            pad_token_id=1,
+            vocab_size=1000,
+        )
+        text_encoder = CLIPTextModel(text_encoder_config)
+        tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")
+
+        unet_lora_attn_procs, unet_lora_layers = create_unet_lora_layers(unet)
+        text_encoder_lora_layers = create_text_encoder_lora_layers(text_encoder)
+
+        pipeline_components = {
+            "unet": unet,
+            "scheduler": scheduler,
+            "vae": vae,
+            "text_encoder": text_encoder,
+            "tokenizer": tokenizer,
+            "safety_checker": None,
+            "feature_extractor": None,
+        }
+        lora_components = {
+            "unet_lora_layers": unet_lora_layers,
+            "text_encoder_lora_layers": text_encoder_lora_layers,
+            "unet_lora_attn_procs": unet_lora_attn_procs,
+        }
+        return pipeline_components, lora_components
+
+    def get_dummy_inputs(self):
+        batch_size = 1
+        sequence_length = 10
+        num_channels = 4
+        sizes = (32, 32)
+
+        generator = torch.manual_seed(0)
+        noise = floats_tensor((batch_size, num_channels) + sizes)
+        input_ids = torch.randint(1, sequence_length, size=(batch_size, sequence_length), generator=generator)
+
+        pipeline_inputs = {
+            "prompt": "A painting of a squirrel eating a burger",
+            "generator": generator,
+            "num_inference_steps": 2,
+            "guidance_scale": 6.0,
+            "output_type": "numpy",
+        }
+
+        return noise, input_ids, pipeline_inputs
+
+    def test_lora_save_load(self):
+        pipeline_components, lora_components = self.get_dummy_components()
+        sd_pipe = StableDiffusionPipeline(**pipeline_components)
+        sd_pipe = sd_pipe.to(torch_device)
+        sd_pipe.set_progress_bar_config(disable=None)
+
+        noise, input_ids, pipeline_inputs = self.get_dummy_inputs()
+
+        original_images = sd_pipe(**pipeline_inputs).images
+        orig_image_slice = original_images[0, -3:, -3:, -1]
+
+        with tempfile.TemporaryDirectory() as tmpdirname:
+            LoraLoaderMixin.save_lora_weights(
+                save_directory=tmpdirname,
+                unet_lora_layers=lora_components["unet_lora_layers"],
+                text_encoder_lora_layers=lora_components["text_encoder_lora_layers"],
+            )
+            self.assertTrue(os.path.isfile(os.path.join(tmpdirname, "pytorch_lora_weights.bin")))
+            sd_pipe.load_lora_weights(tmpdirname)
+
+        lora_images = sd_pipe(**pipeline_inputs).images
+        lora_image_slice = lora_images[0, -3:, -3:, -1]
+
+        # Outputs shouldn't match.
+        self.assertFalse(torch.allclose(torch.from_numpy(orig_image_slice), torch.from_numpy(lora_image_slice)))
+
+    def test_lora_save_load_safetensors(self):
+        pipeline_components, lora_components = self.get_dummy_components()
+        sd_pipe = StableDiffusionPipeline(**pipeline_components)
+        sd_pipe = sd_pipe.to(torch_device)
+        sd_pipe.set_progress_bar_config(disable=None)
+
+        noise, input_ids, pipeline_inputs = self.get_dummy_inputs()
+
+        original_images = sd_pipe(**pipeline_inputs).images
+        orig_image_slice = original_images[0, -3:, -3:, -1]
+
+        with tempfile.TemporaryDirectory() as tmpdirname:
+            LoraLoaderMixin.save_lora_weights(
+                save_directory=tmpdirname,
+                unet_lora_layers=lora_components["unet_lora_layers"],
+                text_encoder_lora_layers=lora_components["text_encoder_lora_layers"],
+                safe_serialization=True,
+            )
+            self.assertTrue(os.path.isfile(os.path.join(tmpdirname, "pytorch_lora_weights.safetensors")))
+            sd_pipe.load_lora_weights(tmpdirname)
+
+        lora_images = sd_pipe(**pipeline_inputs).images
+        lora_image_slice = lora_images[0, -3:, -3:, -1]
+
+        # Outputs shouldn't match.
+        self.assertFalse(torch.allclose(torch.from_numpy(orig_image_slice), torch.from_numpy(lora_image_slice)))
+
+    def test_lora_save_load_legacy(self):
+        pipeline_components, lora_components = self.get_dummy_components()
+        unet_lora_attn_procs = lora_components["unet_lora_attn_procs"]
+        sd_pipe = StableDiffusionPipeline(**pipeline_components)
+        sd_pipe = sd_pipe.to(torch_device)
+        sd_pipe.set_progress_bar_config(disable=None)
+
+        noise, input_ids, pipeline_inputs = self.get_dummy_inputs()
+
+        original_images = sd_pipe(**pipeline_inputs).images
+        orig_image_slice = original_images[0, -3:, -3:, -1]
+
+        with tempfile.TemporaryDirectory() as tmpdirname:
+            unet = sd_pipe.unet
+            unet.set_attn_processor(unet_lora_attn_procs)
+            unet.save_attn_procs(tmpdirname)
+            self.assertTrue(os.path.isfile(os.path.join(tmpdirname, "pytorch_lora_weights.bin")))
+            sd_pipe.load_lora_weights(tmpdirname)
+
+        lora_images = sd_pipe(**pipeline_inputs).images
+        lora_image_slice = lora_images[0, -3:, -3:, -1]
+
+        # Outputs shouldn't match.
+        self.assertFalse(torch.allclose(torch.from_numpy(orig_image_slice), torch.from_numpy(lora_image_slice)))