diffusers 0.31.0__py3-none-any.whl → 0.32.1__py3-none-any.whl

diffusers/loaders/lora_base.py

@@ -51,6 +51,9 @@ if is_accelerate_available():
 
 logger = logging.get_logger(__name__)
 
+LORA_WEIGHT_NAME = "pytorch_lora_weights.bin"
+LORA_WEIGHT_NAME_SAFE = "pytorch_lora_weights.safetensors"
+
 
 def fuse_text_encoder_lora(text_encoder, lora_scale=1.0, safe_fusing=False, adapter_names=None):
     """
@@ -181,6 +184,119 @@ def _remove_text_encoder_monkey_patch(text_encoder):
         text_encoder._hf_peft_config_loaded = None
 
 
+def _fetch_state_dict(
+    pretrained_model_name_or_path_or_dict,
+    weight_name,
+    use_safetensors,
+    local_files_only,
+    cache_dir,
+    force_download,
+    proxies,
+    token,
+    revision,
+    subfolder,
+    user_agent,
+    allow_pickle,
+):
+    model_file = None
+    if not isinstance(pretrained_model_name_or_path_or_dict, dict):
+        # Let's first try to load .safetensors weights
+        if (use_safetensors and weight_name is None) or (
+            weight_name is not None and weight_name.endswith(".safetensors")
+        ):
+            try:
+                # Here we're relaxing the loading check to enable more Inference API
+                # friendliness where sometimes, it's not at all possible to automatically
+                # determine `weight_name`.
+                if weight_name is None:
+                    weight_name = _best_guess_weight_name(
+                        pretrained_model_name_or_path_or_dict,
+                        file_extension=".safetensors",
+                        local_files_only=local_files_only,
+                    )
+                model_file = _get_model_file(
+                    pretrained_model_name_or_path_or_dict,
+                    weights_name=weight_name or LORA_WEIGHT_NAME_SAFE,
+                    cache_dir=cache_dir,
+                    force_download=force_download,
+                    proxies=proxies,
+                    local_files_only=local_files_only,
+                    token=token,
+                    revision=revision,
+                    subfolder=subfolder,
+                    user_agent=user_agent,
+                )
+                state_dict = safetensors.torch.load_file(model_file, device="cpu")
+            except (IOError, safetensors.SafetensorError) as e:
+                if not allow_pickle:
+                    raise e
+                # try loading non-safetensors weights
+                model_file = None
+                pass
+
+        if model_file is None:
+            if weight_name is None:
+                weight_name = _best_guess_weight_name(
+                    pretrained_model_name_or_path_or_dict, file_extension=".bin", local_files_only=local_files_only
+                )
+            model_file = _get_model_file(
+                pretrained_model_name_or_path_or_dict,
+                weights_name=weight_name or LORA_WEIGHT_NAME,
+                cache_dir=cache_dir,
+                force_download=force_download,
+                proxies=proxies,
+                local_files_only=local_files_only,
+                token=token,
+                revision=revision,
+                subfolder=subfolder,
+                user_agent=user_agent,
+            )
+            state_dict = load_state_dict(model_file)
+    else:
+        state_dict = pretrained_model_name_or_path_or_dict
+
+    return state_dict
+
+
+def _best_guess_weight_name(
+    pretrained_model_name_or_path_or_dict, file_extension=".safetensors", local_files_only=False
+):
+    if local_files_only or HF_HUB_OFFLINE:
+        raise ValueError("When using the offline mode, you must specify a `weight_name`.")
+
+    targeted_files = []
+
+    if os.path.isfile(pretrained_model_name_or_path_or_dict):
+        return
+    elif os.path.isdir(pretrained_model_name_or_path_or_dict):
+        targeted_files = [f for f in os.listdir(pretrained_model_name_or_path_or_dict) if f.endswith(file_extension)]
+    else:
+        files_in_repo = model_info(pretrained_model_name_or_path_or_dict).siblings
+        targeted_files = [f.rfilename for f in files_in_repo if f.rfilename.endswith(file_extension)]
+    if len(targeted_files) == 0:
+        return
+
+    # "scheduler" does not correspond to a LoRA checkpoint.
+    # "optimizer" does not correspond to a LoRA checkpoint
+    # only top-level checkpoints are considered and not the other ones, hence "checkpoint".
+    unallowed_substrings = {"scheduler", "optimizer", "checkpoint"}
+    targeted_files = list(
+        filter(lambda x: all(substring not in x for substring in unallowed_substrings), targeted_files)
+    )
+
+    if any(f.endswith(LORA_WEIGHT_NAME) for f in targeted_files):
+        targeted_files = list(filter(lambda x: x.endswith(LORA_WEIGHT_NAME), targeted_files))
+    elif any(f.endswith(LORA_WEIGHT_NAME_SAFE) for f in targeted_files):
+        targeted_files = list(filter(lambda x: x.endswith(LORA_WEIGHT_NAME_SAFE), targeted_files))
+
+    if len(targeted_files) > 1:
+        raise ValueError(
+            f"Provided path contains more than one weights file in the {file_extension} format. Either specify `weight_name` in `load_lora_weights` or make sure there's only one  `.safetensors` or `.bin` file in  {pretrained_model_name_or_path_or_dict}."
+        )
+    weight_name = targeted_files[0]
+    return weight_name
+
+
 class LoraBaseMixin:
     """Utility class for handling LoRAs."""
 
@@ -234,124 +350,16 @@ class LoraBaseMixin:
         return (is_model_cpu_offload, is_sequential_cpu_offload)
 
     @classmethod
-    def _fetch_state_dict(
-        cls,
-        pretrained_model_name_or_path_or_dict,
-        weight_name,
-        use_safetensors,
-        local_files_only,
-        cache_dir,
-        force_download,
-        proxies,
-        token,
-        revision,
-        subfolder,
-        user_agent,
-        allow_pickle,
-    ):
-        from .lora_pipeline import LORA_WEIGHT_NAME, LORA_WEIGHT_NAME_SAFE
-
-        model_file = None
-        if not isinstance(pretrained_model_name_or_path_or_dict, dict):
-            # Let's first try to load .safetensors weights
-            if (use_safetensors and weight_name is None) or (
-                weight_name is not None and weight_name.endswith(".safetensors")
-            ):
-                try:
-                    # Here we're relaxing the loading check to enable more Inference API
-                    # friendliness where sometimes, it's not at all possible to automatically
-                    # determine `weight_name`.
-                    if weight_name is None:
-                        weight_name = cls._best_guess_weight_name(
-                            pretrained_model_name_or_path_or_dict,
-                            file_extension=".safetensors",
-                            local_files_only=local_files_only,
-                        )
-                    model_file = _get_model_file(
-                        pretrained_model_name_or_path_or_dict,
-                        weights_name=weight_name or LORA_WEIGHT_NAME_SAFE,
-                        cache_dir=cache_dir,
-                        force_download=force_download,
-                        proxies=proxies,
-                        local_files_only=local_files_only,
-                        token=token,
-                        revision=revision,
-                        subfolder=subfolder,
-                        user_agent=user_agent,
-                    )
-                    state_dict = safetensors.torch.load_file(model_file, device="cpu")
-                except (IOError, safetensors.SafetensorError) as e:
-                    if not allow_pickle:
-                        raise e
-                    # try loading non-safetensors weights
-                    model_file = None
-                    pass
-
-            if model_file is None:
-                if weight_name is None:
-                    weight_name = cls._best_guess_weight_name(
-                        pretrained_model_name_or_path_or_dict, file_extension=".bin", local_files_only=local_files_only
-                    )
-                model_file = _get_model_file(
-                    pretrained_model_name_or_path_or_dict,
-                    weights_name=weight_name or LORA_WEIGHT_NAME,
-                    cache_dir=cache_dir,
-                    force_download=force_download,
-                    proxies=proxies,
-                    local_files_only=local_files_only,
-                    token=token,
-                    revision=revision,
-                    subfolder=subfolder,
-                    user_agent=user_agent,
-                )
-                state_dict = load_state_dict(model_file)
-        else:
-            state_dict = pretrained_model_name_or_path_or_dict
-
-        return state_dict
+    def _fetch_state_dict(cls, *args, **kwargs):
+        deprecation_message = f"Using the `_fetch_state_dict()` method from {cls} has been deprecated and will be removed in a future version. Please use `from diffusers.loaders.lora_base import _fetch_state_dict`."
+        deprecate("_fetch_state_dict", "0.35.0", deprecation_message)
+        return _fetch_state_dict(*args, **kwargs)
 
     @classmethod
-    def _best_guess_weight_name(
-        cls, pretrained_model_name_or_path_or_dict, file_extension=".safetensors", local_files_only=False
-    ):
-        from .lora_pipeline import LORA_WEIGHT_NAME, LORA_WEIGHT_NAME_SAFE
-
-        if local_files_only or HF_HUB_OFFLINE:
-            raise ValueError("When using the offline mode, you must specify a `weight_name`.")
-
-        targeted_files = []
-
-        if os.path.isfile(pretrained_model_name_or_path_or_dict):
-            return
-        elif os.path.isdir(pretrained_model_name_or_path_or_dict):
-            targeted_files = [
-                f for f in os.listdir(pretrained_model_name_or_path_or_dict) if f.endswith(file_extension)
-            ]
-        else:
-            files_in_repo = model_info(pretrained_model_name_or_path_or_dict).siblings
-            targeted_files = [f.rfilename for f in files_in_repo if f.rfilename.endswith(file_extension)]
-        if len(targeted_files) == 0:
-            return
-
-        # "scheduler" does not correspond to a LoRA checkpoint.
-        # "optimizer" does not correspond to a LoRA checkpoint
-        # only top-level checkpoints are considered and not the other ones, hence "checkpoint".
-        unallowed_substrings = {"scheduler", "optimizer", "checkpoint"}
-        targeted_files = list(
-            filter(lambda x: all(substring not in x for substring in unallowed_substrings), targeted_files)
-        )
-
-        if any(f.endswith(LORA_WEIGHT_NAME) for f in targeted_files):
-            targeted_files = list(filter(lambda x: x.endswith(LORA_WEIGHT_NAME), targeted_files))
-        elif any(f.endswith(LORA_WEIGHT_NAME_SAFE) for f in targeted_files):
-            targeted_files = list(filter(lambda x: x.endswith(LORA_WEIGHT_NAME_SAFE), targeted_files))
-
-        if len(targeted_files) > 1:
-            raise ValueError(
-                f"Provided path contains more than one weights file in the {file_extension} format. Either specify `weight_name` in `load_lora_weights` or make sure there's only one  `.safetensors` or `.bin` file in  {pretrained_model_name_or_path_or_dict}."
-            )
-        weight_name = targeted_files[0]
-        return weight_name
+    def _best_guess_weight_name(cls, *args, **kwargs):
+        deprecation_message = f"Using the `_best_guess_weight_name()` method from {cls} has been deprecated and will be removed in a future version. Please use `from diffusers.loaders.lora_base import _best_guess_weight_name`."
+        deprecate("_best_guess_weight_name", "0.35.0", deprecation_message)
+        return _best_guess_weight_name(*args, **kwargs)
 
     def unload_lora_weights(self):
         """
@@ -725,8 +733,6 @@ class LoraBaseMixin:
         save_function: Callable,
         safe_serialization: bool,
     ):
-        from .lora_pipeline import LORA_WEIGHT_NAME, LORA_WEIGHT_NAME_SAFE
-
         if os.path.isfile(save_directory):
             logger.error(f"Provided path ({save_directory}) should be a directory, not a file")
             return

diffusers/loaders/lora_conversion_utils.py

@@ -636,10 +636,19 @@ def _convert_xlabs_flux_lora_to_diffusers(old_state_dict):
             block_num = re.search(r"single_blocks\.(\d+)", old_key).group(1)
             new_key = f"transformer.single_transformer_blocks.{block_num}"
 
-            if "proj_lora1" in old_key or "proj_lora2" in old_key:
+            if "proj_lora" in old_key:
                 new_key += ".proj_out"
-            elif "qkv_lora1" in old_key or "qkv_lora2" in old_key:
-                new_key += ".norm.linear"
+            elif "qkv_lora" in old_key and "up" not in old_key:
+                handle_qkv(
+                    old_state_dict,
+                    new_state_dict,
+                    old_key,
+                    [
+                        f"transformer.single_transformer_blocks.{block_num}.attn.to_q",
+                        f"transformer.single_transformer_blocks.{block_num}.attn.to_k",
+                        f"transformer.single_transformer_blocks.{block_num}.attn.to_v",
+                    ],
+                )
 
             if "down" in old_key:
                 new_key += ".lora_A.weight"
@@ -658,3 +667,309 @@ def _convert_xlabs_flux_lora_to_diffusers(old_state_dict):
         raise ValueError(f"`old_state_dict` should be at this point but has: {list(old_state_dict.keys())}.")
 
     return new_state_dict
+
+
+def _convert_bfl_flux_control_lora_to_diffusers(original_state_dict):
+    converted_state_dict = {}
+    original_state_dict_keys = list(original_state_dict.keys())
+    num_layers = 19
+    num_single_layers = 38
+    inner_dim = 3072
+    mlp_ratio = 4.0
+
+    def swap_scale_shift(weight):
+        shift, scale = weight.chunk(2, dim=0)
+        new_weight = torch.cat([scale, shift], dim=0)
+        return new_weight
+
+    for lora_key in ["lora_A", "lora_B"]:
+        ## time_text_embed.timestep_embedder <-  time_in
+        converted_state_dict[
+            f"time_text_embed.timestep_embedder.linear_1.{lora_key}.weight"
+        ] = original_state_dict.pop(f"time_in.in_layer.{lora_key}.weight")
+        if f"time_in.in_layer.{lora_key}.bias" in original_state_dict_keys:
+            converted_state_dict[
+                f"time_text_embed.timestep_embedder.linear_1.{lora_key}.bias"
+            ] = original_state_dict.pop(f"time_in.in_layer.{lora_key}.bias")
+
+        converted_state_dict[
+            f"time_text_embed.timestep_embedder.linear_2.{lora_key}.weight"
+        ] = original_state_dict.pop(f"time_in.out_layer.{lora_key}.weight")
+        if f"time_in.out_layer.{lora_key}.bias" in original_state_dict_keys:
+            converted_state_dict[
+                f"time_text_embed.timestep_embedder.linear_2.{lora_key}.bias"
+            ] = original_state_dict.pop(f"time_in.out_layer.{lora_key}.bias")
+
+        ## time_text_embed.text_embedder <- vector_in
+        converted_state_dict[f"time_text_embed.text_embedder.linear_1.{lora_key}.weight"] = original_state_dict.pop(
+            f"vector_in.in_layer.{lora_key}.weight"
+        )
+        if f"vector_in.in_layer.{lora_key}.bias" in original_state_dict_keys:
+            converted_state_dict[f"time_text_embed.text_embedder.linear_1.{lora_key}.bias"] = original_state_dict.pop(
+                f"vector_in.in_layer.{lora_key}.bias"
+            )
+
+        converted_state_dict[f"time_text_embed.text_embedder.linear_2.{lora_key}.weight"] = original_state_dict.pop(
+            f"vector_in.out_layer.{lora_key}.weight"
+        )
+        if f"vector_in.out_layer.{lora_key}.bias" in original_state_dict_keys:
+            converted_state_dict[f"time_text_embed.text_embedder.linear_2.{lora_key}.bias"] = original_state_dict.pop(
+                f"vector_in.out_layer.{lora_key}.bias"
+            )
+
+        # guidance
+        has_guidance = any("guidance" in k for k in original_state_dict)
+        if has_guidance:
+            converted_state_dict[
+                f"time_text_embed.guidance_embedder.linear_1.{lora_key}.weight"
+            ] = original_state_dict.pop(f"guidance_in.in_layer.{lora_key}.weight")
+            if f"guidance_in.in_layer.{lora_key}.bias" in original_state_dict_keys:
+                converted_state_dict[
+                    f"time_text_embed.guidance_embedder.linear_1.{lora_key}.bias"
+                ] = original_state_dict.pop(f"guidance_in.in_layer.{lora_key}.bias")
+
+            converted_state_dict[
+                f"time_text_embed.guidance_embedder.linear_2.{lora_key}.weight"
+            ] = original_state_dict.pop(f"guidance_in.out_layer.{lora_key}.weight")
+            if f"guidance_in.out_layer.{lora_key}.bias" in original_state_dict_keys:
+                converted_state_dict[
+                    f"time_text_embed.guidance_embedder.linear_2.{lora_key}.bias"
+                ] = original_state_dict.pop(f"guidance_in.out_layer.{lora_key}.bias")
+
+        # context_embedder
+        converted_state_dict[f"context_embedder.{lora_key}.weight"] = original_state_dict.pop(
+            f"txt_in.{lora_key}.weight"
+        )
+        if f"txt_in.{lora_key}.bias" in original_state_dict_keys:
+            converted_state_dict[f"context_embedder.{lora_key}.bias"] = original_state_dict.pop(
+                f"txt_in.{lora_key}.bias"
+            )
+
+        # x_embedder
+        converted_state_dict[f"x_embedder.{lora_key}.weight"] = original_state_dict.pop(f"img_in.{lora_key}.weight")
+        if f"img_in.{lora_key}.bias" in original_state_dict_keys:
+            converted_state_dict[f"x_embedder.{lora_key}.bias"] = original_state_dict.pop(f"img_in.{lora_key}.bias")
+
+    # double transformer blocks
+    for i in range(num_layers):
+        block_prefix = f"transformer_blocks.{i}."
+
+        for lora_key in ["lora_A", "lora_B"]:
+            # norms
+            converted_state_dict[f"{block_prefix}norm1.linear.{lora_key}.weight"] = original_state_dict.pop(
+                f"double_blocks.{i}.img_mod.lin.{lora_key}.weight"
+            )
+            if f"double_blocks.{i}.img_mod.lin.{lora_key}.bias" in original_state_dict_keys:
+                converted_state_dict[f"{block_prefix}norm1.linear.{lora_key}.bias"] = original_state_dict.pop(
+                    f"double_blocks.{i}.img_mod.lin.{lora_key}.bias"
+                )
+
+            converted_state_dict[f"{block_prefix}norm1_context.linear.{lora_key}.weight"] = original_state_dict.pop(
+                f"double_blocks.{i}.txt_mod.lin.{lora_key}.weight"
+            )
+            if f"double_blocks.{i}.txt_mod.lin.{lora_key}.bias" in original_state_dict_keys:
+                converted_state_dict[f"{block_prefix}norm1_context.linear.{lora_key}.bias"] = original_state_dict.pop(
+                    f"double_blocks.{i}.txt_mod.lin.{lora_key}.bias"
+                )
+
+            # Q, K, V
+            if lora_key == "lora_A":
+                sample_lora_weight = original_state_dict.pop(f"double_blocks.{i}.img_attn.qkv.{lora_key}.weight")
+                converted_state_dict[f"{block_prefix}attn.to_v.{lora_key}.weight"] = torch.cat([sample_lora_weight])
+                converted_state_dict[f"{block_prefix}attn.to_q.{lora_key}.weight"] = torch.cat([sample_lora_weight])
+                converted_state_dict[f"{block_prefix}attn.to_k.{lora_key}.weight"] = torch.cat([sample_lora_weight])
+
+                context_lora_weight = original_state_dict.pop(f"double_blocks.{i}.txt_attn.qkv.{lora_key}.weight")
+                converted_state_dict[f"{block_prefix}attn.add_q_proj.{lora_key}.weight"] = torch.cat(
+                    [context_lora_weight]
+                )
+                converted_state_dict[f"{block_prefix}attn.add_k_proj.{lora_key}.weight"] = torch.cat(
+                    [context_lora_weight]
+                )
+                converted_state_dict[f"{block_prefix}attn.add_v_proj.{lora_key}.weight"] = torch.cat(
+                    [context_lora_weight]
+                )
+            else:
+                sample_q, sample_k, sample_v = torch.chunk(
+                    original_state_dict.pop(f"double_blocks.{i}.img_attn.qkv.{lora_key}.weight"), 3, dim=0
+                )
+                converted_state_dict[f"{block_prefix}attn.to_q.{lora_key}.weight"] = torch.cat([sample_q])
+                converted_state_dict[f"{block_prefix}attn.to_k.{lora_key}.weight"] = torch.cat([sample_k])
+                converted_state_dict[f"{block_prefix}attn.to_v.{lora_key}.weight"] = torch.cat([sample_v])
+
+                context_q, context_k, context_v = torch.chunk(
+                    original_state_dict.pop(f"double_blocks.{i}.txt_attn.qkv.{lora_key}.weight"), 3, dim=0
+                )
+                converted_state_dict[f"{block_prefix}attn.add_q_proj.{lora_key}.weight"] = torch.cat([context_q])
+                converted_state_dict[f"{block_prefix}attn.add_k_proj.{lora_key}.weight"] = torch.cat([context_k])
+                converted_state_dict[f"{block_prefix}attn.add_v_proj.{lora_key}.weight"] = torch.cat([context_v])
+
+            if f"double_blocks.{i}.img_attn.qkv.{lora_key}.bias" in original_state_dict_keys:
+                sample_q_bias, sample_k_bias, sample_v_bias = torch.chunk(
+                    original_state_dict.pop(f"double_blocks.{i}.img_attn.qkv.{lora_key}.bias"), 3, dim=0
+                )
+                converted_state_dict[f"{block_prefix}attn.to_q.{lora_key}.bias"] = torch.cat([sample_q_bias])
+                converted_state_dict[f"{block_prefix}attn.to_k.{lora_key}.bias"] = torch.cat([sample_k_bias])
+                converted_state_dict[f"{block_prefix}attn.to_v.{lora_key}.bias"] = torch.cat([sample_v_bias])
+
+            if f"double_blocks.{i}.txt_attn.qkv.{lora_key}.bias" in original_state_dict_keys:
+                context_q_bias, context_k_bias, context_v_bias = torch.chunk(
+                    original_state_dict.pop(f"double_blocks.{i}.txt_attn.qkv.{lora_key}.bias"), 3, dim=0
+                )
+                converted_state_dict[f"{block_prefix}attn.add_q_proj.{lora_key}.bias"] = torch.cat([context_q_bias])
+                converted_state_dict[f"{block_prefix}attn.add_k_proj.{lora_key}.bias"] = torch.cat([context_k_bias])
+                converted_state_dict[f"{block_prefix}attn.add_v_proj.{lora_key}.bias"] = torch.cat([context_v_bias])
+
+            # ff img_mlp
+            converted_state_dict[f"{block_prefix}ff.net.0.proj.{lora_key}.weight"] = original_state_dict.pop(
+                f"double_blocks.{i}.img_mlp.0.{lora_key}.weight"
+            )
+            if f"double_blocks.{i}.img_mlp.0.{lora_key}.bias" in original_state_dict_keys:
+                converted_state_dict[f"{block_prefix}ff.net.0.proj.{lora_key}.bias"] = original_state_dict.pop(
+                    f"double_blocks.{i}.img_mlp.0.{lora_key}.bias"
+                )
+
+            converted_state_dict[f"{block_prefix}ff.net.2.{lora_key}.weight"] = original_state_dict.pop(
+                f"double_blocks.{i}.img_mlp.2.{lora_key}.weight"
+            )
+            if f"double_blocks.{i}.img_mlp.2.{lora_key}.bias" in original_state_dict_keys:
+                converted_state_dict[f"{block_prefix}ff.net.2.{lora_key}.bias"] = original_state_dict.pop(
+                    f"double_blocks.{i}.img_mlp.2.{lora_key}.bias"
+                )
+
+            converted_state_dict[f"{block_prefix}ff_context.net.0.proj.{lora_key}.weight"] = original_state_dict.pop(
+                f"double_blocks.{i}.txt_mlp.0.{lora_key}.weight"
+            )
+            if f"double_blocks.{i}.txt_mlp.0.{lora_key}.bias" in original_state_dict_keys:
+                converted_state_dict[f"{block_prefix}ff_context.net.0.proj.{lora_key}.bias"] = original_state_dict.pop(
+                    f"double_blocks.{i}.txt_mlp.0.{lora_key}.bias"
+                )
+
+            converted_state_dict[f"{block_prefix}ff_context.net.2.{lora_key}.weight"] = original_state_dict.pop(
+                f"double_blocks.{i}.txt_mlp.2.{lora_key}.weight"
+            )
+            if f"double_blocks.{i}.txt_mlp.2.{lora_key}.bias" in original_state_dict_keys:
+                converted_state_dict[f"{block_prefix}ff_context.net.2.{lora_key}.bias"] = original_state_dict.pop(
+                    f"double_blocks.{i}.txt_mlp.2.{lora_key}.bias"
+                )
+
+            # output projections.
+            converted_state_dict[f"{block_prefix}attn.to_out.0.{lora_key}.weight"] = original_state_dict.pop(
+                f"double_blocks.{i}.img_attn.proj.{lora_key}.weight"
+            )
+            if f"double_blocks.{i}.img_attn.proj.{lora_key}.bias" in original_state_dict_keys:
+                converted_state_dict[f"{block_prefix}attn.to_out.0.{lora_key}.bias"] = original_state_dict.pop(
+                    f"double_blocks.{i}.img_attn.proj.{lora_key}.bias"
+                )
+            converted_state_dict[f"{block_prefix}attn.to_add_out.{lora_key}.weight"] = original_state_dict.pop(
+                f"double_blocks.{i}.txt_attn.proj.{lora_key}.weight"
+            )
+            if f"double_blocks.{i}.txt_attn.proj.{lora_key}.bias" in original_state_dict_keys:
+                converted_state_dict[f"{block_prefix}attn.to_add_out.{lora_key}.bias"] = original_state_dict.pop(
+                    f"double_blocks.{i}.txt_attn.proj.{lora_key}.bias"
+                )
+
+        # qk_norm
+        converted_state_dict[f"{block_prefix}attn.norm_q.weight"] = original_state_dict.pop(
+            f"double_blocks.{i}.img_attn.norm.query_norm.scale"
+        )
+        converted_state_dict[f"{block_prefix}attn.norm_k.weight"] = original_state_dict.pop(
+            f"double_blocks.{i}.img_attn.norm.key_norm.scale"
+        )
+        converted_state_dict[f"{block_prefix}attn.norm_added_q.weight"] = original_state_dict.pop(
+            f"double_blocks.{i}.txt_attn.norm.query_norm.scale"
+        )
+        converted_state_dict[f"{block_prefix}attn.norm_added_k.weight"] = original_state_dict.pop(
+            f"double_blocks.{i}.txt_attn.norm.key_norm.scale"
+        )
+
+    # single transfomer blocks
+    for i in range(num_single_layers):
+        block_prefix = f"single_transformer_blocks.{i}."
+
+        for lora_key in ["lora_A", "lora_B"]:
+            # norm.linear  <- single_blocks.0.modulation.lin
+            converted_state_dict[f"{block_prefix}norm.linear.{lora_key}.weight"] = original_state_dict.pop(
+                f"single_blocks.{i}.modulation.lin.{lora_key}.weight"
+            )
+            if f"single_blocks.{i}.modulation.lin.{lora_key}.bias" in original_state_dict_keys:
+                converted_state_dict[f"{block_prefix}norm.linear.{lora_key}.bias"] = original_state_dict.pop(
+                    f"single_blocks.{i}.modulation.lin.{lora_key}.bias"
+                )
+
+            # Q, K, V, mlp
+            mlp_hidden_dim = int(inner_dim * mlp_ratio)
+            split_size = (inner_dim, inner_dim, inner_dim, mlp_hidden_dim)
+
+            if lora_key == "lora_A":
+                lora_weight = original_state_dict.pop(f"single_blocks.{i}.linear1.{lora_key}.weight")
+                converted_state_dict[f"{block_prefix}attn.to_q.{lora_key}.weight"] = torch.cat([lora_weight])
+                converted_state_dict[f"{block_prefix}attn.to_k.{lora_key}.weight"] = torch.cat([lora_weight])
+                converted_state_dict[f"{block_prefix}attn.to_v.{lora_key}.weight"] = torch.cat([lora_weight])
+                converted_state_dict[f"{block_prefix}proj_mlp.{lora_key}.weight"] = torch.cat([lora_weight])
+
+                if f"single_blocks.{i}.linear1.{lora_key}.bias" in original_state_dict_keys:
+                    lora_bias = original_state_dict.pop(f"single_blocks.{i}.linear1.{lora_key}.bias")
+                    converted_state_dict[f"{block_prefix}attn.to_q.{lora_key}.bias"] = torch.cat([lora_bias])
+                    converted_state_dict[f"{block_prefix}attn.to_k.{lora_key}.bias"] = torch.cat([lora_bias])
+                    converted_state_dict[f"{block_prefix}attn.to_v.{lora_key}.bias"] = torch.cat([lora_bias])
+                    converted_state_dict[f"{block_prefix}proj_mlp.{lora_key}.bias"] = torch.cat([lora_bias])
+            else:
+                q, k, v, mlp = torch.split(
+                    original_state_dict.pop(f"single_blocks.{i}.linear1.{lora_key}.weight"), split_size, dim=0
+                )
+                converted_state_dict[f"{block_prefix}attn.to_q.{lora_key}.weight"] = torch.cat([q])
+                converted_state_dict[f"{block_prefix}attn.to_k.{lora_key}.weight"] = torch.cat([k])
+                converted_state_dict[f"{block_prefix}attn.to_v.{lora_key}.weight"] = torch.cat([v])
+                converted_state_dict[f"{block_prefix}proj_mlp.{lora_key}.weight"] = torch.cat([mlp])
+
+                if f"single_blocks.{i}.linear1.{lora_key}.bias" in original_state_dict_keys:
+                    q_bias, k_bias, v_bias, mlp_bias = torch.split(
+                        original_state_dict.pop(f"single_blocks.{i}.linear1.{lora_key}.bias"), split_size, dim=0
+                    )
+                    converted_state_dict[f"{block_prefix}attn.to_q.{lora_key}.bias"] = torch.cat([q_bias])
+                    converted_state_dict[f"{block_prefix}attn.to_k.{lora_key}.bias"] = torch.cat([k_bias])
+                    converted_state_dict[f"{block_prefix}attn.to_v.{lora_key}.bias"] = torch.cat([v_bias])
+                    converted_state_dict[f"{block_prefix}proj_mlp.{lora_key}.bias"] = torch.cat([mlp_bias])
+
+            # output projections.
+            converted_state_dict[f"{block_prefix}proj_out.{lora_key}.weight"] = original_state_dict.pop(
+                f"single_blocks.{i}.linear2.{lora_key}.weight"
+            )
+            if f"single_blocks.{i}.linear2.{lora_key}.bias" in original_state_dict_keys:
+                converted_state_dict[f"{block_prefix}proj_out.{lora_key}.bias"] = original_state_dict.pop(
+                    f"single_blocks.{i}.linear2.{lora_key}.bias"
+                )
+
+        # qk norm
+        converted_state_dict[f"{block_prefix}attn.norm_q.weight"] = original_state_dict.pop(
+            f"single_blocks.{i}.norm.query_norm.scale"
+        )
+        converted_state_dict[f"{block_prefix}attn.norm_k.weight"] = original_state_dict.pop(
+            f"single_blocks.{i}.norm.key_norm.scale"
+        )
+
+    for lora_key in ["lora_A", "lora_B"]:
+        converted_state_dict[f"proj_out.{lora_key}.weight"] = original_state_dict.pop(
+            f"final_layer.linear.{lora_key}.weight"
+        )
+        if f"final_layer.linear.{lora_key}.bias" in original_state_dict_keys:
+            converted_state_dict[f"proj_out.{lora_key}.bias"] = original_state_dict.pop(
+                f"final_layer.linear.{lora_key}.bias"
+            )
+
+        converted_state_dict[f"norm_out.linear.{lora_key}.weight"] = swap_scale_shift(
+            original_state_dict.pop(f"final_layer.adaLN_modulation.1.{lora_key}.weight")
+        )
+        if f"final_layer.adaLN_modulation.1.{lora_key}.bias" in original_state_dict_keys:
+            converted_state_dict[f"norm_out.linear.{lora_key}.bias"] = swap_scale_shift(
+                original_state_dict.pop(f"final_layer.adaLN_modulation.1.{lora_key}.bias")
+            )
+
+    if len(original_state_dict) > 0:
+        raise ValueError(f"`original_state_dict` should be empty at this point but has {original_state_dict.keys()=}.")
+
+    for key in list(converted_state_dict.keys()):
+        converted_state_dict[f"transformer.{key}"] = converted_state_dict.pop(key)
+
+    return converted_state_dict