[@cene555][Kandinsky 3.0] Add Kandinsky 3.0 (huggingface#5913)

* finalize

* finalize

* finalize

* add slow test

* add slow test

* add slow test

* Fix more

* add slow test

* fix more

* fix more

* fix more

* fix more

* fix more

* fix more

* fix more

* fix more

* fix more

* Better

* Fix more

* Fix more

* add slow test

* Add auto pipelines

* add slow test

* Add all

* add slow test

* add slow test

* add slow test

* add slow test

* add slow test

* Apply suggestions from code review

* add slow test

* add slow test

Author: patrickvonplaten

__init__.py

@@ -79,6 +79,7 @@
             "AutoencoderTiny",
             "ConsistencyDecoderVAE",
             "ControlNetModel",
+            "Kandinsky3UNet",
             "ModelMixin",
             "MotionAdapter",
             "MultiAdapter",
@@ -214,6 +215,8 @@
             "IFPipeline",
             "IFSuperResolutionPipeline",
             "ImageTextPipelineOutput",
+            "Kandinsky3Img2ImgPipeline",
+            "Kandinsky3Pipeline",
             "KandinskyCombinedPipeline",
             "KandinskyImg2ImgCombinedPipeline",
             "KandinskyImg2ImgPipeline",
@@ -446,6 +449,7 @@
             AutoencoderTiny,
             ConsistencyDecoderVAE,
             ControlNetModel,
+            Kandinsky3UNet,
             ModelMixin,
             MotionAdapter,
             MultiAdapter,
@@ -560,6 +564,8 @@
             IFPipeline,
             IFSuperResolutionPipeline,
             ImageTextPipelineOutput,
+            Kandinsky3Img2ImgPipeline,
+            Kandinsky3Pipeline,
             KandinskyCombinedPipeline,
             KandinskyImg2ImgCombinedPipeline,
             KandinskyImg2ImgPipeline,

models/__init__.py

@@ -36,6 +36,7 @@
     _import_structure["unet_2d"] = ["UNet2DModel"]
     _import_structure["unet_2d_condition"] = ["UNet2DConditionModel"]
     _import_structure["unet_3d_condition"] = ["UNet3DConditionModel"]
+    _import_structure["unet_kandi3"] = ["Kandinsky3UNet"]
     _import_structure["unet_motion_model"] = ["MotionAdapter", "UNetMotionModel"]
     _import_structure["vq_model"] = ["VQModel"]
 
@@ -63,6 +64,7 @@
         from .unet_2d import UNet2DModel
         from .unet_2d_condition import UNet2DConditionModel
         from .unet_3d_condition import UNet3DConditionModel
+        from .unet_kandi3 import Kandinsky3UNet
         from .unet_motion_model import MotionAdapter, UNetMotionModel
         from .vq_model import VQModel
 

models/attention_processor.py

@@ -16,7 +16,7 @@
 
 import torch
 import torch.nn.functional as F
-from torch import nn
+from torch import einsum, nn
 
 from ..utils import USE_PEFT_BACKEND, deprecate, logging
 from ..utils.import_utils import is_xformers_available
@@ -2219,6 +2219,44 @@ def __call__(
         return hidden_states
 
 
+# TODO(Yiyi): This class should not exist, we can replace it with a normal attention processor I believe
+# this way torch.compile and co. will work as well
+class Kandi3AttnProcessor:
+    r"""
+    Default kandinsky3 proccesor for performing attention-related computations.
+    """
+
+    @staticmethod
+    def _reshape(hid_states, h):
+        b, n, f = hid_states.shape
+        d = f // h
+        return hid_states.unsqueeze(-1).reshape(b, n, h, d).permute(0, 2, 1, 3)
+
+    def __call__(
+        self,
+        attn,
+        x,
+        context,
+        context_mask=None,
+    ):
+        query = self._reshape(attn.to_q(x), h=attn.num_heads)
+        key = self._reshape(attn.to_k(context), h=attn.num_heads)
+        value = self._reshape(attn.to_v(context), h=attn.num_heads)
+
+        attention_matrix = einsum("b h i d, b h j d -> b h i j", query, key)
+
+        if context_mask is not None:
+            max_neg_value = -torch.finfo(attention_matrix.dtype).max
+            context_mask = context_mask.unsqueeze(1).unsqueeze(1)
+            attention_matrix = attention_matrix.masked_fill(~(context_mask != 0), max_neg_value)
+        attention_matrix = (attention_matrix * attn.scale).softmax(dim=-1)
+
+        out = einsum("b h i j, b h j d -> b h i d", attention_matrix, value)
+        out = out.permute(0, 2, 1, 3).reshape(out.shape[0], out.shape[2], -1)
+        out = attn.to_out[0](out)
+        return out
+
+
 LORA_ATTENTION_PROCESSORS = (
     LoRAAttnProcessor,
     LoRAAttnProcessor2_0,
@@ -2244,6 +2282,7 @@ def __call__(
     LoRAXFormersAttnProcessor,
     IPAdapterAttnProcessor,
     IPAdapterAttnProcessor2_0,
+    Kandi3AttnProcessor,
 )
 
 AttentionProcessor = Union[