Merge branch 'main' into integrations/ltx-098

Author: a-r-r-o-w

src/diffusers/__init__.py

@@ -139,6 +139,7 @@
             "AutoGuidance",
             "ClassifierFreeGuidance",
             "ClassifierFreeZeroStarGuidance",
+            "FrequencyDecoupledGuidance",
             "PerturbedAttentionGuidance",
             "SkipLayerGuidance",
             "SmoothedEnergyGuidance",
@@ -804,6 +805,7 @@
             AutoGuidance,
             ClassifierFreeGuidance,
             ClassifierFreeZeroStarGuidance,
+            FrequencyDecoupledGuidance,
             PerturbedAttentionGuidance,
             SkipLayerGuidance,
             SmoothedEnergyGuidance,

src/diffusers/guiders/__init__.py

@@ -22,6 +22,7 @@
     from .auto_guidance import AutoGuidance
     from .classifier_free_guidance import ClassifierFreeGuidance
     from .classifier_free_zero_star_guidance import ClassifierFreeZeroStarGuidance
+    from .frequency_decoupled_guidance import FrequencyDecoupledGuidance
     from .perturbed_attention_guidance import PerturbedAttentionGuidance
     from .skip_layer_guidance import SkipLayerGuidance
     from .smoothed_energy_guidance import SmoothedEnergyGuidance
@@ -32,6 +33,7 @@
         AutoGuidance,
         ClassifierFreeGuidance,
         ClassifierFreeZeroStarGuidance,
+        FrequencyDecoupledGuidance,
         PerturbedAttentionGuidance,
         SkipLayerGuidance,
         SmoothedEnergyGuidance,

src/diffusers/guiders/frequency_decoupled_guidance.py

@@ -133,6 +133,7 @@ def _register_attention_processors_metadata():
             skip_processor_output_fn=_skip_proc_output_fn_Attention_WanAttnProcessor2_0,
         ),
     )
+
     # FluxAttnProcessor
     AttentionProcessorRegistry.register(
         model_class=FluxAttnProcessor,

src/diffusers/hooks/_helpers.py

@@ -245,7 +245,6 @@ def _offload_to_memory(self):
                 param.data = self.cpu_param_dict[param]
             for buffer in self.buffers:
                 buffer.data = self.cpu_param_dict[buffer]
-
         else:
             for group_module in self.modules:
                 group_module.to(self.offload_device, non_blocking=False)
@@ -303,9 +302,23 @@ def pre_forward(self, module: torch.nn.Module, *args, **kwargs):
         if self.group.onload_leader == module:
             if self.group.onload_self:
                 self.group.onload_()
-            if self.next_group is not None and not self.next_group.onload_self:
+
+            should_onload_next_group = self.next_group is not None and not self.next_group.onload_self
+            if should_onload_next_group:
                 self.next_group.onload_()
 
+            should_synchronize = (
+                not self.group.onload_self and self.group.stream is not None and not should_onload_next_group
+            )
+            if should_synchronize:
+                # If this group didn't onload itself, it means it was asynchronously onloaded by the
+                # previous group. We need to synchronize the side stream to ensure parameters
+                # are completely loaded to proceed with forward pass. Without this, uninitialized
+                # weights will be used in the computation, leading to incorrect results
+                # Also, we should only do this synchronization if we don't already do it from the sync call in
+                # self.next_group.onload_, hence the `not should_onload_next_group` check.
+                self.group.stream.synchronize()
+
         args = send_to_device(args, self.group.onload_device, non_blocking=self.group.non_blocking)
         kwargs = send_to_device(kwargs, self.group.onload_device, non_blocking=self.group.non_blocking)
         return args, kwargs

src/diffusers/hooks/group_offloading.py

@@ -0,0 +1,43 @@
+# Copyright 2025 The HuggingFace Team. All rights reserved.
+#
+# 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 torch
+
+from ._common import _ALL_TRANSFORMER_BLOCK_IDENTIFIERS, _ATTENTION_CLASSES, _FEEDFORWARD_CLASSES
+
+
+def _get_identifiable_transformer_blocks_in_module(module: torch.nn.Module):
+    module_list_with_transformer_blocks = []
+    for name, submodule in module.named_modules():
+        name_endswith_identifier = any(name.endswith(identifier) for identifier in _ALL_TRANSFORMER_BLOCK_IDENTIFIERS)
+        is_modulelist = isinstance(submodule, torch.nn.ModuleList)
+        if name_endswith_identifier and is_modulelist:
+            module_list_with_transformer_blocks.append((name, submodule))
+    return module_list_with_transformer_blocks
+
+
+def _get_identifiable_attention_layers_in_module(module: torch.nn.Module):
+    attention_layers = []
+    for name, submodule in module.named_modules():
+        if isinstance(submodule, _ATTENTION_CLASSES):
+            attention_layers.append((name, submodule))
+    return attention_layers
+
+
+def _get_identifiable_feedforward_layers_in_module(module: torch.nn.Module):
+    feedforward_layers = []
+    for name, submodule in module.named_modules():
+        if isinstance(submodule, _FEEDFORWARD_CLASSES):
+            feedforward_layers.append((name, submodule))
+    return feedforward_layers

src/diffusers/hooks/utils.py

@@ -384,7 +384,7 @@ def forward(
         temb: torch.Tensor,
         image_rotary_emb: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
         joint_attention_kwargs: Optional[Dict[str, Any]] = None,
-    ) -> torch.Tensor:
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
         text_seq_len = encoder_hidden_states.shape[1]
         hidden_states = torch.cat([encoder_hidden_states, hidden_states], dim=1)
 

src/diffusers/models/transformers/transformer_flux.py

@@ -310,7 +310,7 @@ def _get_clip_prompt_embeds(
     def encode_prompt(
         self,
         prompt: Union[str, List[str]],
-        prompt_2: Union[str, List[str]],
+        prompt_2: Optional[Union[str, List[str]]] = None,
         device: Optional[torch.device] = None,
         num_images_per_prompt: int = 1,
         prompt_embeds: Optional[torch.FloatTensor] = None,

src/diffusers/pipelines/flux/pipeline_flux.py

@@ -324,7 +324,7 @@ def _get_clip_prompt_embeds(
     def encode_prompt(
         self,
         prompt: Union[str, List[str]],
-        prompt_2: Union[str, List[str]],
+        prompt_2: Optional[Union[str, List[str]]] = None,
         device: Optional[torch.device] = None,
         num_images_per_prompt: int = 1,
         prompt_embeds: Optional[torch.FloatTensor] = None,

src/diffusers/pipelines/flux/pipeline_flux_control.py

@@ -335,7 +335,7 @@ def _get_clip_prompt_embeds(
     def encode_prompt(
         self,
         prompt: Union[str, List[str]],
-        prompt_2: Union[str, List[str]],
+        prompt_2: Optional[Union[str, List[str]]] = None,
         device: Optional[torch.device] = None,
         num_images_per_prompt: int = 1,
         prompt_embeds: Optional[torch.FloatTensor] = None,