[Training] fix training resuming problem when using FP16 (SDXL LoRA DreamBooth)

examples/dreambooth/train_dreambooth_lora_sdxl.py

@@ -34,7 +34,7 @@
 from huggingface_hub import create_repo, upload_folder
 from huggingface_hub.utils import insecure_hashlib
 from packaging import version
-from peft import LoraConfig
+from peft import LoraConfig, set_peft_model_state_dict
 from peft.utils import get_peft_model_state_dict
 from PIL import Image
 from PIL.ImageOps import exif_transpose
@@ -56,6 +56,7 @@
 from diffusers.training_utils import compute_snr
 from diffusers.utils import check_min_version, convert_state_dict_to_diffusers, is_wandb_available
 from diffusers.utils.import_utils import is_xformers_available
+from diffusers.utils.state_dict_utils import convert_state_dict_to_peft, convert_unet_state_dict_to_peft
 
 
 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
@@ -996,17 +997,6 @@ def main(args):
         text_encoder_one.add_adapter(text_lora_config)
         text_encoder_two.add_adapter(text_lora_config)
 
-    # Make sure the trainable params are in float32.
-    if args.mixed_precision == "fp16":
-        models = [unet]
-        if args.train_text_encoder:
-            models.extend([text_encoder_one, text_encoder_two])
-        for model in models:
-            for param in model.parameters():
-                # only upcast trainable parameters (LoRA) into fp32
-                if param.requires_grad:
-                    param.data = param.to(torch.float32)
-
     # create custom saving & loading hooks so that `accelerator.save_state(...)` serializes in a nice format
     def save_model_hook(models, weights, output_dir):
         if accelerator.is_main_process:
@@ -1058,17 +1048,52 @@ def load_model_hook(models, input_dir):
                 raise ValueError(f"unexpected save model: {model.__class__}")
 
         lora_state_dict, network_alphas = LoraLoaderMixin.lora_state_dict(input_dir)
-        LoraLoaderMixin.load_lora_into_unet(lora_state_dict, network_alphas=network_alphas, unet=unet_)
 
-        text_encoder_state_dict = {k: v for k, v in lora_state_dict.items() if "text_encoder." in k}
-        LoraLoaderMixin.load_lora_into_text_encoder(
-            text_encoder_state_dict, network_alphas=network_alphas, text_encoder=text_encoder_one_
-        )
+        unet_state_dict = {f'{k.replace("unet.", "")}': v for k, v in lora_state_dict.items() if k.startswith("unet.")}
+        unet_state_dict = convert_unet_state_dict_to_peft(unet_state_dict)
+        incompatible_keys = set_peft_model_state_dict(unet_, unet_state_dict, adapter_name="default")
+        if incompatible_keys is not None:
+            # check only for unexpected keys
+            unexpected_keys = getattr(incompatible_keys, "unexpected_keys", None)
+            if unexpected_keys:
+                logger.warning(
+                    f"Loading adapter weights from state_dict led to unexpected keys not found in the model: "
+                    f" {unexpected_keys}. "
+                )
 
-        text_encoder_2_state_dict = {k: v for k, v in lora_state_dict.items() if "text_encoder_2." in k}
-        LoraLoaderMixin.load_lora_into_text_encoder(
-            text_encoder_2_state_dict, network_alphas=network_alphas, text_encoder=text_encoder_two_
-        )
+        if args.train_text_encoder:
+            # Do we need to call `scale_lora_layers()` here?
+            text_encoder_state_dict = {
+                f'{k.replace("text_encoder.", "")}': v
+                for k, v in lora_state_dict.items()
+                if k.startswith("text_encoder.")
+            }
+            text_encoder_state_dict = convert_state_dict_to_peft(
+                convert_state_dict_to_diffusers(text_encoder_state_dict)
+            )
+            set_peft_model_state_dict(text_encoder_one_, text_encoder_state_dict, adapter_name="default")
+
+            text_encoder_2_state_dict = {
+                f'{k.replace("text_encoder_2.", "")}': v
+                for k, v in lora_state_dict.items()
+                if k.startswith("text_encoder_2.")
+            }
+            text_encoder_2_state_dict = convert_state_dict_to_peft(
+                convert_state_dict_to_diffusers(text_encoder_2_state_dict)
+            )
+            set_peft_model_state_dict(text_encoder_two_, text_encoder_2_state_dict, adapter_name="default")
+
+        # Make sure the trainable params are in float32. This is again needed since the base models
+        # are in `weight_dtype`.
+        if args.mixed_precision == "fp16":
+            models = [unet_]
+            if args.train_text_encoder:
+                models.extend([text_encoder_one_, text_encoder_two_])
+            for model in models:
+                for param in model.parameters():
+                    # only upcast trainable parameters (LoRA) into fp32
+                    if param.requires_grad:
+                        param.data = param.to(torch.float32)
 
     accelerator.register_save_state_pre_hook(save_model_hook)
     accelerator.register_load_state_pre_hook(load_model_hook)
@@ -1083,6 +1108,17 @@ def load_model_hook(models, input_dir):
             args.learning_rate * args.gradient_accumulation_steps * args.train_batch_size * accelerator.num_processes
         )
 
+    # Make sure the trainable params are in float32.
+    if args.mixed_precision == "fp16":
+        models = [unet]
+        if args.train_text_encoder:
+            models.extend([text_encoder_one, text_encoder_two])
+        for model in models:
+            for param in model.parameters():
+                # only upcast trainable parameters (LoRA) into fp32
+                if param.requires_grad:
+                    param.data = param.to(torch.float32)
+
     unet_lora_parameters = list(filter(lambda p: p.requires_grad, unet.parameters()))
 
     if args.train_text_encoder:
@@ -1500,6 +1536,7 @@ def compute_text_embeddings(prompt, text_encoders, tokenizers):
                         else unet_lora_parameters
                     )
                     accelerator.clip_grad_norm_(params_to_clip, args.max_grad_norm)
+
                 optimizer.step()
                 lr_scheduler.step()
                 optimizer.zero_grad()

src/diffusers/loaders/lora.py

@@ -581,7 +581,6 @@ def load_lora_into_text_encoder(
                     lora_config_kwargs = get_peft_kwargs(
                         rank, network_alphas, text_encoder_lora_state_dict, is_unet=False
                     )
-
                     lora_config = LoraConfig(**lora_config_kwargs)
 
                     # adapter_name