diffusers 0.27.1__py3-none-any.whl → 0.32.2__py3-none-any.whl

diffusers/loaders/textual_inversion.py

@@ -18,6 +18,7 @@ import torch
 from huggingface_hub.utils import validate_hf_hub_args
 from torch import nn
 
+from ..models.modeling_utils import load_state_dict
 from ..utils import _get_model_file, is_accelerate_available, is_transformers_available, logging
 
 
@@ -37,7 +38,6 @@ TEXT_INVERSION_NAME_SAFE = "learned_embeds.safetensors"
 def load_textual_inversion_state_dicts(pretrained_model_name_or_paths, **kwargs):
     cache_dir = kwargs.pop("cache_dir", None)
     force_download = kwargs.pop("force_download", False)
-    resume_download = kwargs.pop("resume_download", False)
     proxies = kwargs.pop("proxies", None)
     local_files_only = kwargs.pop("local_files_only", None)
     token = kwargs.pop("token", None)
@@ -71,7 +71,6 @@ def load_textual_inversion_state_dicts(pretrained_model_name_or_paths, **kwargs)
                         weights_name=weight_name or TEXT_INVERSION_NAME_SAFE,
                         cache_dir=cache_dir,
                         force_download=force_download,
-                        resume_download=resume_download,
                         proxies=proxies,
                         local_files_only=local_files_only,
                         token=token,
@@ -92,7 +91,6 @@ def load_textual_inversion_state_dicts(pretrained_model_name_or_paths, **kwargs)
                     weights_name=weight_name or TEXT_INVERSION_NAME,
                     cache_dir=cache_dir,
                     force_download=force_download,
-                    resume_download=resume_download,
                     proxies=proxies,
                     local_files_only=local_files_only,
                     token=token,
@@ -100,7 +98,7 @@ def load_textual_inversion_state_dicts(pretrained_model_name_or_paths, **kwargs)
                     subfolder=subfolder,
                     user_agent=user_agent,
                 )
-                state_dict = torch.load(model_file, map_location="cpu")
+                state_dict = load_state_dict(model_file)
         else:
             state_dict = pretrained_model_name_or_path
 
@@ -307,9 +305,7 @@ class TextualInversionLoaderMixin:
             force_download (`bool`, *optional*, defaults to `False`):
                 Whether or not to force the (re-)download of the model weights and configuration files, overriding the
                 cached versions if they exist.
-            resume_download (`bool`, *optional*, defaults to `False`):
-                Whether or not to resume downloading the model weights and configuration files. If set to `False`, any
-                incompletely downloaded files are deleted.
+
             proxies (`Dict[str, str]`, *optional*):
                 A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
                 'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
@@ -418,15 +414,20 @@ class TextualInversionLoaderMixin:
         # 7.1 Offload all hooks in case the pipeline was cpu offloaded before make sure, we offload and onload again
         is_model_cpu_offload = False
         is_sequential_cpu_offload = False
-        for _, component in self.components.items():
-            if isinstance(component, nn.Module):
-                if hasattr(component, "_hf_hook"):
-                    is_model_cpu_offload = isinstance(getattr(component, "_hf_hook"), CpuOffload)
-                    is_sequential_cpu_offload = isinstance(getattr(component, "_hf_hook"), AlignDevicesHook)
-                    logger.info(
-                        "Accelerate hooks detected. Since you have called `load_textual_inversion()`, the previous hooks will be first removed. Then the textual inversion parameters will be loaded and the hooks will be applied again."
-                    )
-                    remove_hook_from_module(component, recurse=is_sequential_cpu_offload)
+        if self.hf_device_map is None:
+            for _, component in self.components.items():
+                if isinstance(component, nn.Module):
+                    if hasattr(component, "_hf_hook"):
+                        is_model_cpu_offload = isinstance(getattr(component, "_hf_hook"), CpuOffload)
+                        is_sequential_cpu_offload = (
+                            isinstance(getattr(component, "_hf_hook"), AlignDevicesHook)
+                            or hasattr(component._hf_hook, "hooks")
+                            and isinstance(component._hf_hook.hooks[0], AlignDevicesHook)
+                        )
+                        logger.info(
+                            "Accelerate hooks detected. Since you have called `load_textual_inversion()`, the previous hooks will be first removed. Then the textual inversion parameters will be loaded and the hooks will be applied again."
+                        )
+                        remove_hook_from_module(component, recurse=is_sequential_cpu_offload)
 
         # 7.2 save expected device and dtype
         device = text_encoder.device
@@ -486,20 +487,35 @@ class TextualInversionLoaderMixin:
 
         # Example 3: unload from SDXL
         pipeline = AutoPipelineForText2Image.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0")
-        embedding_path = hf_hub_download(repo_id="linoyts/web_y2k", filename="web_y2k_emb.safetensors", repo_type="model")
+        embedding_path = hf_hub_download(
+            repo_id="linoyts/web_y2k", filename="web_y2k_emb.safetensors", repo_type="model"
+        )
 
         # load embeddings to the text encoders
         state_dict = load_file(embedding_path)
 
         # load embeddings of text_encoder 1 (CLIP ViT-L/14)
-        pipeline.load_textual_inversion(state_dict["clip_l"], token=["<s0>", "<s1>"], text_encoder=pipeline.text_encoder, tokenizer=pipeline.tokenizer)
+        pipeline.load_textual_inversion(
+            state_dict["clip_l"],
+            tokens=["<s0>", "<s1>"],
+            text_encoder=pipeline.text_encoder,
+            tokenizer=pipeline.tokenizer,
+        )
         # load embeddings of text_encoder 2 (CLIP ViT-G/14)
-        pipeline.load_textual_inversion(state_dict["clip_g"], token=["<s0>", "<s1>"], text_encoder=pipeline.text_encoder_2, tokenizer=pipeline.tokenizer_2)
-
-        # Unload explicitly from both text encoders abd tokenizers
-        pipeline.unload_textual_inversion(tokens=["<s0>", "<s1>"], text_encoder=pipeline.text_encoder, tokenizer=pipeline.tokenizer)
-        pipeline.unload_textual_inversion(tokens=["<s0>", "<s1>"], text_encoder=pipeline.text_encoder_2, tokenizer=pipeline.tokenizer_2)
+        pipeline.load_textual_inversion(
+            state_dict["clip_g"],
+            tokens=["<s0>", "<s1>"],
+            text_encoder=pipeline.text_encoder_2,
+            tokenizer=pipeline.tokenizer_2,
+        )
 
+        # Unload explicitly from both text encoders and tokenizers
+        pipeline.unload_textual_inversion(
+            tokens=["<s0>", "<s1>"], text_encoder=pipeline.text_encoder, tokenizer=pipeline.tokenizer
+        )
+        pipeline.unload_textual_inversion(
+            tokens=["<s0>", "<s1>"], text_encoder=pipeline.text_encoder_2, tokenizer=pipeline.tokenizer_2
+        )
         ```
         """
 
@@ -545,6 +561,8 @@ class TextualInversionLoaderMixin:
                 tokenizer._added_tokens_encoder[token.content] = last_special_token_id + key_id
                 key_id += 1
         tokenizer._update_trie()
+        # set correct total vocab size after removing tokens
+        tokenizer._update_total_vocab_size()
 
         # Delete from text encoder
         text_embedding_dim = text_encoder.get_input_embeddings().embedding_dim

diffusers/loaders/transformer_flux.py

@@ -0,0 +1,181 @@
+# Copyright 2024 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.
+from contextlib import nullcontext
+
+from ..models.embeddings import (
+    ImageProjection,
+    MultiIPAdapterImageProjection,
+)
+from ..models.modeling_utils import load_model_dict_into_meta
+from ..utils import (
+    is_accelerate_available,
+    is_torch_version,
+    logging,
+)
+
+
+if is_accelerate_available():
+    pass
+
+logger = logging.get_logger(__name__)
+
+
+class FluxTransformer2DLoadersMixin:
+    """
+    Load layers into a [`FluxTransformer2DModel`].
+    """
+
+    def _convert_ip_adapter_image_proj_to_diffusers(self, state_dict, low_cpu_mem_usage=False):
+        if low_cpu_mem_usage:
+            if is_accelerate_available():
+                from accelerate import init_empty_weights
+
+            else:
+                low_cpu_mem_usage = False
+                logger.warning(
+                    "Cannot initialize model with low cpu memory usage because `accelerate` was not found in the"
+                    " environment. Defaulting to `low_cpu_mem_usage=False`. It is strongly recommended to install"
+                    " `accelerate` for faster and less memory-intense model loading. You can do so with: \n```\npip"
+                    " install accelerate\n```\n."
+                )
+
+        if low_cpu_mem_usage is True and not is_torch_version(">=", "1.9.0"):
+            raise NotImplementedError(
+                "Low memory initialization requires torch >= 1.9.0. Please either update your PyTorch version or set"
+                " `low_cpu_mem_usage=False`."
+            )
+
+        updated_state_dict = {}
+        image_projection = None
+        init_context = init_empty_weights if low_cpu_mem_usage else nullcontext
+
+        if "proj.weight" in state_dict:
+            # IP-Adapter
+            num_image_text_embeds = 4
+            if state_dict["proj.weight"].shape[0] == 65536:
+                num_image_text_embeds = 16
+            clip_embeddings_dim = state_dict["proj.weight"].shape[-1]
+            cross_attention_dim = state_dict["proj.weight"].shape[0] // num_image_text_embeds
+
+            with init_context():
+                image_projection = ImageProjection(
+                    cross_attention_dim=cross_attention_dim,
+                    image_embed_dim=clip_embeddings_dim,
+                    num_image_text_embeds=num_image_text_embeds,
+                )
+
+            for key, value in state_dict.items():
+                diffusers_name = key.replace("proj", "image_embeds")
+                updated_state_dict[diffusers_name] = value
+
+        if not low_cpu_mem_usage:
+            image_projection.load_state_dict(updated_state_dict, strict=True)
+        else:
+            load_model_dict_into_meta(image_projection, updated_state_dict, device=self.device, dtype=self.dtype)
+
+        return image_projection
+
+    def _convert_ip_adapter_attn_to_diffusers(self, state_dicts, low_cpu_mem_usage=False):
+        from ..models.attention_processor import (
+            FluxIPAdapterJointAttnProcessor2_0,
+        )
+
+        if low_cpu_mem_usage:
+            if is_accelerate_available():
+                from accelerate import init_empty_weights
+
+            else:
+                low_cpu_mem_usage = False
+                logger.warning(
+                    "Cannot initialize model with low cpu memory usage because `accelerate` was not found in the"
+                    " environment. Defaulting to `low_cpu_mem_usage=False`. It is strongly recommended to install"
+                    " `accelerate` for faster and less memory-intense model loading. You can do so with: \n```\npip"
+                    " install accelerate\n```\n."
+                )
+
+        if low_cpu_mem_usage is True and not is_torch_version(">=", "1.9.0"):
+            raise NotImplementedError(
+                "Low memory initialization requires torch >= 1.9.0. Please either update your PyTorch version or set"
+                " `low_cpu_mem_usage=False`."
+            )
+
+        # set ip-adapter cross-attention processors & load state_dict
+        attn_procs = {}
+        key_id = 0
+        init_context = init_empty_weights if low_cpu_mem_usage else nullcontext
+        for name in self.attn_processors.keys():
+            if name.startswith("single_transformer_blocks"):
+                attn_processor_class = self.attn_processors[name].__class__
+                attn_procs[name] = attn_processor_class()
+            else:
+                cross_attention_dim = self.config.joint_attention_dim
+                hidden_size = self.inner_dim
+                attn_processor_class = FluxIPAdapterJointAttnProcessor2_0
+                num_image_text_embeds = []
+                for state_dict in state_dicts:
+                    if "proj.weight" in state_dict["image_proj"]:
+                        num_image_text_embed = 4
+                        if state_dict["image_proj"]["proj.weight"].shape[0] == 65536:
+                            num_image_text_embed = 16
+                        # IP-Adapter
+                        num_image_text_embeds += [num_image_text_embed]
+
+                with init_context():
+                    attn_procs[name] = attn_processor_class(
+                        hidden_size=hidden_size,
+                        cross_attention_dim=cross_attention_dim,
+                        scale=1.0,
+                        num_tokens=num_image_text_embeds,
+                        dtype=self.dtype,
+                        device=self.device,
+                    )
+
+                value_dict = {}
+                for i, state_dict in enumerate(state_dicts):
+                    value_dict.update({f"to_k_ip.{i}.weight": state_dict["ip_adapter"][f"{key_id}.to_k_ip.weight"]})
+                    value_dict.update({f"to_v_ip.{i}.weight": state_dict["ip_adapter"][f"{key_id}.to_v_ip.weight"]})
+                    value_dict.update({f"to_k_ip.{i}.bias": state_dict["ip_adapter"][f"{key_id}.to_k_ip.bias"]})
+                    value_dict.update({f"to_v_ip.{i}.bias": state_dict["ip_adapter"][f"{key_id}.to_v_ip.bias"]})
+
+                if not low_cpu_mem_usage:
+                    attn_procs[name].load_state_dict(value_dict)
+                else:
+                    device = self.device
+                    dtype = self.dtype
+                    load_model_dict_into_meta(attn_procs[name], value_dict, device=device, dtype=dtype)
+
+                key_id += 1
+
+        return attn_procs
+
+    def _load_ip_adapter_weights(self, state_dicts, low_cpu_mem_usage=False):
+        if not isinstance(state_dicts, list):
+            state_dicts = [state_dicts]
+
+        self.encoder_hid_proj = None
+
+        attn_procs = self._convert_ip_adapter_attn_to_diffusers(state_dicts, low_cpu_mem_usage=low_cpu_mem_usage)
+        self.set_attn_processor(attn_procs)
+
+        image_projection_layers = []
+        for state_dict in state_dicts:
+            image_projection_layer = self._convert_ip_adapter_image_proj_to_diffusers(
+                state_dict["image_proj"], low_cpu_mem_usage=low_cpu_mem_usage
+            )
+            image_projection_layers.append(image_projection_layer)
+
+        self.encoder_hid_proj = MultiIPAdapterImageProjection(image_projection_layers)
+        self.config.encoder_hid_dim_type = "ip_image_proj"
+
+        self.to(dtype=self.dtype, device=self.device)

diffusers/loaders/transformer_sd3.py

@@ -0,0 +1,89 @@
+# Copyright 2024 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.
+from typing import Dict
+
+from ..models.attention_processor import SD3IPAdapterJointAttnProcessor2_0
+from ..models.embeddings import IPAdapterTimeImageProjection
+from ..models.modeling_utils import _LOW_CPU_MEM_USAGE_DEFAULT, load_model_dict_into_meta
+
+
+class SD3Transformer2DLoadersMixin:
+    """Load IP-Adapters and LoRA layers into a `[SD3Transformer2DModel]`."""
+
+    def _load_ip_adapter_weights(self, state_dict: Dict, low_cpu_mem_usage: bool = _LOW_CPU_MEM_USAGE_DEFAULT) -> None:
+        """Sets IP-Adapter attention processors, image projection, and loads state_dict.
+
+        Args:
+            state_dict (`Dict`):
+                State dict with keys "ip_adapter", which contains parameters for attention processors, and
+                "image_proj", which contains parameters for image projection net.
+            low_cpu_mem_usage (`bool`, *optional*, defaults to `True` if torch version >= 1.9.0 else `False`):
+                Speed up model loading only loading the pretrained weights and not initializing the weights. This also
+                tries to not use more than 1x model size in CPU memory (including peak memory) while loading the model.
+                Only supported for PyTorch >= 1.9.0. If you are using an older version of PyTorch, setting this
+                argument to `True` will raise an error.
+        """
+        # IP-Adapter cross attention parameters
+        hidden_size = self.config.attention_head_dim * self.config.num_attention_heads
+        ip_hidden_states_dim = self.config.attention_head_dim * self.config.num_attention_heads
+        timesteps_emb_dim = state_dict["ip_adapter"]["0.norm_ip.linear.weight"].shape[1]
+
+        # Dict where key is transformer layer index, value is attention processor's state dict
+        # ip_adapter state dict keys example: "0.norm_ip.linear.weight"
+        layer_state_dict = {idx: {} for idx in range(len(self.attn_processors))}
+        for key, weights in state_dict["ip_adapter"].items():
+            idx, name = key.split(".", maxsplit=1)
+            layer_state_dict[int(idx)][name] = weights
+
+        # Create IP-Adapter attention processor
+        attn_procs = {}
+        for idx, name in enumerate(self.attn_processors.keys()):
+            attn_procs[name] = SD3IPAdapterJointAttnProcessor2_0(
+                hidden_size=hidden_size,
+                ip_hidden_states_dim=ip_hidden_states_dim,
+                head_dim=self.config.attention_head_dim,
+                timesteps_emb_dim=timesteps_emb_dim,
+            ).to(self.device, dtype=self.dtype)
+
+            if not low_cpu_mem_usage:
+                attn_procs[name].load_state_dict(layer_state_dict[idx], strict=True)
+            else:
+                load_model_dict_into_meta(
+                    attn_procs[name], layer_state_dict[idx], device=self.device, dtype=self.dtype
+                )
+
+        self.set_attn_processor(attn_procs)
+
+        # Image projetion parameters
+        embed_dim = state_dict["image_proj"]["proj_in.weight"].shape[1]
+        output_dim = state_dict["image_proj"]["proj_out.weight"].shape[0]
+        hidden_dim = state_dict["image_proj"]["proj_in.weight"].shape[0]
+        heads = state_dict["image_proj"]["layers.0.attn.to_q.weight"].shape[0] // 64
+        num_queries = state_dict["image_proj"]["latents"].shape[1]
+        timestep_in_dim = state_dict["image_proj"]["time_embedding.linear_1.weight"].shape[1]
+
+        # Image projection
+        self.image_proj = IPAdapterTimeImageProjection(
+            embed_dim=embed_dim,
+            output_dim=output_dim,
+            hidden_dim=hidden_dim,
+            heads=heads,
+            num_queries=num_queries,
+            timestep_in_dim=timestep_in_dim,
+        ).to(device=self.device, dtype=self.dtype)
+
+        if not low_cpu_mem_usage:
+            self.image_proj.load_state_dict(state_dict["image_proj"], strict=True)
+        else:
+            load_model_dict_into_meta(self.image_proj, state_dict["image_proj"], device=self.device, dtype=self.dtype)