[LoRA] don't break offloading for incompatible lora ckpts. (huggingface#5085)

* don't break offloading for incompatible lora ckpts.

* debugging

* better condition.

* fix

* fix

* fix

* fix

---------

Co-authored-by: Patrick von Platen <[email protected]>

Author: sayakpaul

src/diffusers/loaders.py

@@ -13,7 +13,6 @@
 # limitations under the License.
 import os
 import re
-import warnings
 from collections import defaultdict
 from contextlib import nullcontext
 from io import BytesIO
@@ -33,7 +32,6 @@
     _get_model_file,
     deprecate,
     is_accelerate_available,
-    is_accelerate_version,
     is_omegaconf_available,
     is_transformers_available,
     logging,
@@ -308,6 +306,9 @@ def load_attn_procs(self, pretrained_model_name_or_path_or_dict: Union[str, Dict
         # This value has the same meaning as the `--network_alpha` option in the kohya-ss trainer script.
         # See https://github.com/darkstorm2150/sd-scripts/blob/main/docs/train_network_README-en.md#execute-learning
         network_alphas = kwargs.pop("network_alphas", None)
+
+        _pipeline = kwargs.pop("_pipeline", None)
+
         is_network_alphas_none = network_alphas is None
 
         allow_pickle = False
@@ -461,6 +462,7 @@ def load_attn_procs(self, pretrained_model_name_or_path_or_dict: Union[str, Dict
                     load_model_dict_into_meta(lora, value_dict, device=device, dtype=dtype)
                 else:
                     lora.load_state_dict(value_dict)
+
         elif is_custom_diffusion:
             attn_processors = {}
             custom_diffusion_grouped_dict = defaultdict(dict)
@@ -490,19 +492,44 @@ def load_attn_procs(self, pretrained_model_name_or_path_or_dict: Union[str, Dict
                         cross_attention_dim=cross_attention_dim,
                     )
                     attn_processors[key].load_state_dict(value_dict)
-
-            self.set_attn_processor(attn_processors)
         else:
             raise ValueError(
                 f"{model_file} does not seem to be in the correct format expected by LoRA or Custom Diffusion training."
             )
 
+        # <Unsafe code
+        # We can be sure that the following works as it just sets attention processors, lora layers and puts all in the same dtype
+        # Now we remove any existing hooks to
+        is_model_cpu_offload = False
+        is_sequential_cpu_offload = False
+        if _pipeline is not None:
+            for _, component in _pipeline.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_lora_weights()`, the previous hooks will be first removed. Then the LoRA parameters will be loaded and the hooks will be applied again."
+                        )
+                        remove_hook_from_module(component, recurse=is_sequential_cpu_offload)
+
+        # only custom diffusion needs to set attn processors
+        if is_custom_diffusion:
+            self.set_attn_processor(attn_processors)
+
         # set lora layers
         for target_module, lora_layer in lora_layers_list:
             target_module.set_lora_layer(lora_layer)
 
         self.to(dtype=self.dtype, device=self.device)
 
+        # Offload back.
+        if is_model_cpu_offload:
+            _pipeline.enable_model_cpu_offload()
+        elif is_sequential_cpu_offload:
+            _pipeline.enable_sequential_cpu_offload()
+        # Unsafe code />
+
     def convert_state_dict_legacy_attn_format(self, state_dict, network_alphas):
         is_new_lora_format = all(
             key.startswith(self.unet_name) or key.startswith(self.text_encoder_name) for key in state_dict.keys()
@@ -1072,41 +1099,31 @@ def load_lora_weights(self, pretrained_model_name_or_path_or_dict: Union[str, Di
             kwargs (`dict`, *optional*):
                 See [`~loaders.LoraLoaderMixin.lora_state_dict`].
         """
-        # Remove any existing hooks.
-        is_model_cpu_offload = False
-        is_sequential_cpu_offload = False
-        recurive = 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_lora_weights()`, the previous hooks will be first removed. Then the LoRA parameters will be loaded and the hooks will be applied again."
-                    )
-                    recurive = is_sequential_cpu_offload
-                    remove_hook_from_module(component, recurse=recurive)
+        # First, ensure that the checkpoint is a compatible one and can be successfully loaded.
+        state_dict, network_alphas = self.lora_state_dict(pretrained_model_name_or_path_or_dict, **kwargs)
+
+        is_correct_format = all("lora" in key for key in state_dict.keys())
+        if not is_correct_format:
+            raise ValueError("Invalid LoRA checkpoint.")
 
         low_cpu_mem_usage = kwargs.pop("low_cpu_mem_usage", _LOW_CPU_MEM_USAGE_DEFAULT)
 
-        state_dict, network_alphas = self.lora_state_dict(pretrained_model_name_or_path_or_dict, **kwargs)
         self.load_lora_into_unet(
-            state_dict, network_alphas=network_alphas, unet=self.unet, low_cpu_mem_usage=low_cpu_mem_usage
+            state_dict,
+            network_alphas=network_alphas,
+            unet=self.unet,
+            low_cpu_mem_usage=low_cpu_mem_usage,
+            _pipeline=self,
         )
         self.load_lora_into_text_encoder(
             state_dict,
             network_alphas=network_alphas,
             text_encoder=self.text_encoder,
             lora_scale=self.lora_scale,
             low_cpu_mem_usage=low_cpu_mem_usage,
+            _pipeline=self,
         )
 
-        # Offload back.
-        if is_model_cpu_offload:
-            self.enable_model_cpu_offload()
-        elif is_sequential_cpu_offload:
-            self.enable_sequential_cpu_offload()
-
     @classmethod
     def lora_state_dict(
         cls,
@@ -1403,7 +1420,7 @@ def _maybe_map_sgm_blocks_to_diffusers(cls, state_dict, unet_config, delimiter="
         return new_state_dict
 
     @classmethod
-    def load_lora_into_unet(cls, state_dict, network_alphas, unet, low_cpu_mem_usage=None):
+    def load_lora_into_unet(cls, state_dict, network_alphas, unet, low_cpu_mem_usage=None, _pipeline=None):
         """
         This will load the LoRA layers specified in `state_dict` into `unet`.
 
@@ -1445,13 +1462,22 @@ def load_lora_into_unet(cls, state_dict, network_alphas, unet, low_cpu_mem_usage
             # Otherwise, we're dealing with the old format. This means the `state_dict` should only
             # contain the module names of the `unet` as its keys WITHOUT any prefix.
             warn_message = "You have saved the LoRA weights using the old format. To convert the old LoRA weights to the new format, you can first load them in a dictionary and then create a new dictionary like the following: `new_state_dict = {f'unet.{module_name}': params for module_name, params in old_state_dict.items()}`."
-            warnings.warn(warn_message)
+            logger.warn(warn_message)
 
-        unet.load_attn_procs(state_dict, network_alphas=network_alphas, low_cpu_mem_usage=low_cpu_mem_usage)
+        unet.load_attn_procs(
+            state_dict, network_alphas=network_alphas, low_cpu_mem_usage=low_cpu_mem_usage, _pipeline=_pipeline
+        )
 
     @classmethod
     def load_lora_into_text_encoder(
-        cls, state_dict, network_alphas, text_encoder, prefix=None, lora_scale=1.0, low_cpu_mem_usage=None
+        cls,
+        state_dict,
+        network_alphas,
+        text_encoder,
+        prefix=None,
+        lora_scale=1.0,
+        low_cpu_mem_usage=None,
+        _pipeline=None,
     ):
         """
         This will load the LoRA layers specified in `state_dict` into `text_encoder`
@@ -1561,11 +1587,15 @@ def load_lora_into_text_encoder(
                     low_cpu_mem_usage=low_cpu_mem_usage,
                 )
 
-                # set correct dtype & device
-                text_encoder_lora_state_dict = {
-                    k: v.to(device=text_encoder.device, dtype=text_encoder.dtype)
-                    for k, v in text_encoder_lora_state_dict.items()
-                }
+                is_pipeline_offloaded = _pipeline is not None and any(
+                    isinstance(c, torch.nn.Module) and hasattr(c, "_hf_hook") for c in _pipeline.components.values()
+                )
+                if is_pipeline_offloaded and low_cpu_mem_usage:
+                    low_cpu_mem_usage = True
+                    logger.info(
+                        f"Pipeline {_pipeline.__class__} is offloaded. Therefore low cpu mem usage loading is forced."
+                    )
+
                 if low_cpu_mem_usage:
                     device = next(iter(text_encoder_lora_state_dict.values())).device
                     dtype = next(iter(text_encoder_lora_state_dict.values())).dtype
@@ -1581,8 +1611,33 @@ def load_lora_into_text_encoder(
                         f"failed to load text encoder state dict, unexpected keys: {load_state_dict_results.unexpected_keys}"
                     )
 
+                # <Unsafe code
+                # We can be sure that the following works as all we do is change the dtype and device of the text encoder
+                # Now we remove any existing hooks to
+                is_model_cpu_offload = False
+                is_sequential_cpu_offload = False
+                if _pipeline is not None:
+                    for _, component in _pipeline.components.items():
+                        if isinstance(component, torch.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_lora_weights()`, the previous hooks will be first removed. Then the LoRA parameters will be loaded and the hooks will be applied again."
+                                )
+                                remove_hook_from_module(component, recurse=is_sequential_cpu_offload)
+
                 text_encoder.to(device=text_encoder.device, dtype=text_encoder.dtype)
 
+                # Offload back.
+                if is_model_cpu_offload:
+                    _pipeline.enable_model_cpu_offload()
+                elif is_sequential_cpu_offload:
+                    _pipeline.enable_sequential_cpu_offload()
+                # Unsafe code />
+
     @property
     def lora_scale(self) -> float:
         # property function that returns the lora scale which can be set at run time by the pipeline.
@@ -2652,31 +2707,17 @@ def load_lora_weights(self, pretrained_model_name_or_path_or_dict: Union[str, Di
         # it here explicitly to be able to tell that it's coming from an SDXL
         # pipeline.
 
-        # Remove any existing hooks.
-        if is_accelerate_available() and is_accelerate_version(">=", "0.17.0.dev0"):
-            from accelerate.hooks import AlignDevicesHook, CpuOffload, remove_hook_from_module
-        else:
-            raise ImportError("Offloading requires `accelerate v0.17.0` or higher.")
-
-        is_model_cpu_offload = False
-        is_sequential_cpu_offload = False
-        for _, component in self.components.items():
-            if isinstance(component, torch.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_lora_weights()`, the previous hooks will be first removed. Then the LoRA parameters will be loaded and the hooks will be applied again."
-                    )
-                    remove_hook_from_module(component, recurse=is_sequential_cpu_offload)
-
+        # First, ensure that the checkpoint is a compatible one and can be successfully loaded.
         state_dict, network_alphas = self.lora_state_dict(
             pretrained_model_name_or_path_or_dict,
             unet_config=self.unet.config,
             **kwargs,
         )
-        self.load_lora_into_unet(state_dict, network_alphas=network_alphas, unet=self.unet)
+        is_correct_format = all("lora" in key for key in state_dict.keys())
+        if not is_correct_format:
+            raise ValueError("Invalid LoRA checkpoint.")
 
+        self.load_lora_into_unet(state_dict, network_alphas=network_alphas, unet=self.unet, _pipeline=self)
         text_encoder_state_dict = {k: v for k, v in state_dict.items() if "text_encoder." in k}
         if len(text_encoder_state_dict) > 0:
             self.load_lora_into_text_encoder(
@@ -2685,6 +2726,7 @@ def load_lora_weights(self, pretrained_model_name_or_path_or_dict: Union[str, Di
                 text_encoder=self.text_encoder,
                 prefix="text_encoder",
                 lora_scale=self.lora_scale,
+                _pipeline=self,
             )
 
         text_encoder_2_state_dict = {k: v for k, v in state_dict.items() if "text_encoder_2." in k}
@@ -2695,14 +2737,9 @@ def load_lora_weights(self, pretrained_model_name_or_path_or_dict: Union[str, Di
                 text_encoder=self.text_encoder_2,
                 prefix="text_encoder_2",
                 lora_scale=self.lora_scale,
+                _pipeline=self,
             )
 
-        # Offload back.
-        if is_model_cpu_offload:
-            self.enable_model_cpu_offload()
-        elif is_sequential_cpu_offload:
-            self.enable_sequential_cpu_offload()
-
     @classmethod
     def save_lora_weights(
         self,