Implement rest of the test cases (LoRA tests)

src/diffusers/models/unet_3d_blocks.py

@@ -251,7 +251,9 @@ def forward(
                 encoder_hidden_states=encoder_hidden_states,
                 cross_attention_kwargs=cross_attention_kwargs,
             ).sample
-            hidden_states = temp_attn(hidden_states, num_frames=num_frames).sample
+            hidden_states = temp_attn(
+                hidden_states, num_frames=num_frames, cross_attention_kwargs=cross_attention_kwargs
+            ).sample
             hidden_states = resnet(hidden_states, temb)
             hidden_states = temp_conv(hidden_states, num_frames=num_frames)
 
@@ -376,7 +378,9 @@ def forward(
                 encoder_hidden_states=encoder_hidden_states,
                 cross_attention_kwargs=cross_attention_kwargs,
             ).sample
-            hidden_states = temp_attn(hidden_states, num_frames=num_frames).sample
+            hidden_states = temp_attn(
+                hidden_states, num_frames=num_frames, cross_attention_kwargs=cross_attention_kwargs
+            ).sample
 
             output_states += (hidden_states,)
 
@@ -587,7 +591,9 @@ def forward(
                 encoder_hidden_states=encoder_hidden_states,
                 cross_attention_kwargs=cross_attention_kwargs,
             ).sample
-            hidden_states = temp_attn(hidden_states, num_frames=num_frames).sample
+            hidden_states = temp_attn(
+                hidden_states, num_frames=num_frames, cross_attention_kwargs=cross_attention_kwargs
+            ).sample
 
         if self.upsamplers is not None:
             for upsampler in self.upsamplers:

src/diffusers/models/unet_3d_condition.py

@@ -20,6 +20,7 @@
 import torch.utils.checkpoint
 
 from ..configuration_utils import ConfigMixin, register_to_config
+from ..loaders import UNet2DConditionLoadersMixin
 from ..utils import BaseOutput, logging
 from .attention_processor import AttentionProcessor, AttnProcessor
 from .embeddings import TimestepEmbedding, Timesteps
@@ -50,7 +51,7 @@ class UNet3DConditionOutput(BaseOutput):
     sample: torch.FloatTensor
 
 
-class UNet3DConditionModel(ModelMixin, ConfigMixin):
+class UNet3DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin):
     r"""
     UNet3DConditionModel is a conditional 2D UNet model that takes in a noisy sample, conditional state, and a timestep
     and returns sample shaped output.
@@ -465,7 +466,9 @@ def forward(
         sample = sample.permute(0, 2, 1, 3, 4).reshape((sample.shape[0] * num_frames, -1) + sample.shape[3:])
         sample = self.conv_in(sample)
 
-        sample = self.transformer_in(sample, num_frames=num_frames).sample
+        sample = self.transformer_in(
+            sample, num_frames=num_frames, cross_attention_kwargs=cross_attention_kwargs
+        ).sample
 
         # 3. down
         down_block_res_samples = (sample,)

tests/models/test_models_unet_2d_condition.py

@@ -41,7 +41,7 @@
 torch.backends.cuda.matmul.allow_tf32 = False
 
 
-def create_lora_layers(model):
+def create_lora_layers(model, mock_weights: bool = True):
     lora_attn_procs = {}
     for name in model.attn_processors.keys():
         cross_attention_dim = None if name.endswith("attn1.processor") else model.config.cross_attention_dim
@@ -57,12 +57,13 @@ def create_lora_layers(model):
         lora_attn_procs[name] = LoRAAttnProcessor(hidden_size=hidden_size, cross_attention_dim=cross_attention_dim)
         lora_attn_procs[name] = lora_attn_procs[name].to(model.device)
 
-        # add 1 to weights to mock trained weights
-        with torch.no_grad():
-            lora_attn_procs[name].to_q_lora.up.weight += 1
-            lora_attn_procs[name].to_k_lora.up.weight += 1
-            lora_attn_procs[name].to_v_lora.up.weight += 1
-            lora_attn_procs[name].to_out_lora.up.weight += 1
+        if mock_weights:
+            # add 1 to weights to mock trained weights
+            with torch.no_grad():
+                lora_attn_procs[name].to_q_lora.up.weight += 1
+                lora_attn_procs[name].to_k_lora.up.weight += 1
+                lora_attn_procs[name].to_v_lora.up.weight += 1
+                lora_attn_procs[name].to_out_lora.up.weight += 1
 
     return lora_attn_procs
 
@@ -378,26 +379,7 @@ def test_lora_processors(self):
         with torch.no_grad():
             sample1 = model(**inputs_dict).sample
 
-        lora_attn_procs = {}
-        for name in model.attn_processors.keys():
-            cross_attention_dim = None if name.endswith("attn1.processor") else model.config.cross_attention_dim
-            if name.startswith("mid_block"):
-                hidden_size = model.config.block_out_channels[-1]
-            elif name.startswith("up_blocks"):
-                block_id = int(name[len("up_blocks.")])
-                hidden_size = list(reversed(model.config.block_out_channels))[block_id]
-            elif name.startswith("down_blocks"):
-                block_id = int(name[len("down_blocks.")])
-                hidden_size = model.config.block_out_channels[block_id]
-
-            lora_attn_procs[name] = LoRAAttnProcessor(hidden_size=hidden_size, cross_attention_dim=cross_attention_dim)
-
-            # add 1 to weights to mock trained weights
-            with torch.no_grad():
-                lora_attn_procs[name].to_q_lora.up.weight += 1
-                lora_attn_procs[name].to_k_lora.up.weight += 1
-                lora_attn_procs[name].to_v_lora.up.weight += 1
-                lora_attn_procs[name].to_out_lora.up.weight += 1
+        lora_attn_procs = create_lora_layers(model)
 
         # make sure we can set a list of attention processors
         model.set_attn_processor(lora_attn_procs)
@@ -465,28 +447,7 @@ def test_lora_save_load_safetensors(self):
         with torch.no_grad():
             old_sample = model(**inputs_dict).sample
 
-        lora_attn_procs = {}
-        for name in model.attn_processors.keys():
-            cross_attention_dim = None if name.endswith("attn1.processor") else model.config.cross_attention_dim
-            if name.startswith("mid_block"):
-                hidden_size = model.config.block_out_channels[-1]
-            elif name.startswith("up_blocks"):
-                block_id = int(name[len("up_blocks.")])
-                hidden_size = list(reversed(model.config.block_out_channels))[block_id]
-            elif name.startswith("down_blocks"):
-                block_id = int(name[len("down_blocks.")])
-                hidden_size = model.config.block_out_channels[block_id]
-
-            lora_attn_procs[name] = LoRAAttnProcessor(hidden_size=hidden_size, cross_attention_dim=cross_attention_dim)
-            lora_attn_procs[name] = lora_attn_procs[name].to(model.device)
-
-            # add 1 to weights to mock trained weights
-            with torch.no_grad():
-                lora_attn_procs[name].to_q_lora.up.weight += 1
-                lora_attn_procs[name].to_k_lora.up.weight += 1
-                lora_attn_procs[name].to_v_lora.up.weight += 1
-                lora_attn_procs[name].to_out_lora.up.weight += 1
-
+        lora_attn_procs = create_lora_layers(model)
         model.set_attn_processor(lora_attn_procs)
 
         with torch.no_grad():
@@ -518,21 +479,7 @@ def test_lora_save_safetensors_load_torch(self):
         model = self.model_class(**init_dict)
         model.to(torch_device)
 
-        lora_attn_procs = {}
-        for name in model.attn_processors.keys():
-            cross_attention_dim = None if name.endswith("attn1.processor") else model.config.cross_attention_dim
-            if name.startswith("mid_block"):
-                hidden_size = model.config.block_out_channels[-1]
-            elif name.startswith("up_blocks"):
-                block_id = int(name[len("up_blocks.")])
-                hidden_size = list(reversed(model.config.block_out_channels))[block_id]
-            elif name.startswith("down_blocks"):
-                block_id = int(name[len("down_blocks.")])
-                hidden_size = model.config.block_out_channels[block_id]
-
-            lora_attn_procs[name] = LoRAAttnProcessor(hidden_size=hidden_size, cross_attention_dim=cross_attention_dim)
-            lora_attn_procs[name] = lora_attn_procs[name].to(model.device)
-
+        lora_attn_procs = create_lora_layers(model, mock_weights=False)
         model.set_attn_processor(lora_attn_procs)
         # Saving as torch, properly reloads with directly filename
         with tempfile.TemporaryDirectory() as tmpdirname:
@@ -553,21 +500,7 @@ def test_lora_save_torch_force_load_safetensors_error(self):
         model = self.model_class(**init_dict)
         model.to(torch_device)
 
-        lora_attn_procs = {}
-        for name in model.attn_processors.keys():
-            cross_attention_dim = None if name.endswith("attn1.processor") else model.config.cross_attention_dim
-            if name.startswith("mid_block"):
-                hidden_size = model.config.block_out_channels[-1]
-            elif name.startswith("up_blocks"):
-                block_id = int(name[len("up_blocks.")])
-                hidden_size = list(reversed(model.config.block_out_channels))[block_id]
-            elif name.startswith("down_blocks"):
-                block_id = int(name[len("down_blocks.")])
-                hidden_size = model.config.block_out_channels[block_id]
-
-            lora_attn_procs[name] = LoRAAttnProcessor(hidden_size=hidden_size, cross_attention_dim=cross_attention_dim)
-            lora_attn_procs[name] = lora_attn_procs[name].to(model.device)
-
+        lora_attn_procs = create_lora_layers(model, mock_weights=False)
         model.set_attn_processor(lora_attn_procs)
         # Saving as torch, properly reloads with directly filename
         with tempfile.TemporaryDirectory() as tmpdirname: