[LoRA] make LoRAs trained with `peft` loadable when `peft` isn't installed (#6306)

* spit diffusers-native format from the get go.

* rejig the peft_to_diffusers mapping.

examples/dreambooth/train_dreambooth_lora_sdxl.py

@@ -54,7 +54,7 @@ from diffusers import (
 from diffusers.loaders import LoraLoaderMixin
 from diffusers.optimization import get_scheduler
 from diffusers.training_utils import compute_snr
-from diffusers.utils import check_min_version, is_wandb_available
+from diffusers.utils import check_min_version, convert_state_dict_to_diffusers, is_wandb_available
 from diffusers.utils.import_utils import is_xformers_available
 
 
@@ -1019,11 +1019,15 @@ def main(args):
 
             for model in models:
                 if isinstance(model, type(accelerator.unwrap_model(unet))):
-                    unet_lora_layers_to_save = get_peft_model_state_dict(model)
+                    unet_lora_layers_to_save = convert_state_dict_to_diffusers(get_peft_model_state_dict(model))
                 elif isinstance(model, type(accelerator.unwrap_model(text_encoder_one))):
-                    text_encoder_one_lora_layers_to_save = get_peft_model_state_dict(model)
+                    text_encoder_one_lora_layers_to_save = convert_state_dict_to_diffusers(
+                        get_peft_model_state_dict(model)
+                    )
                 elif isinstance(model, type(accelerator.unwrap_model(text_encoder_two))):
-                    text_encoder_two_lora_layers_to_save = get_peft_model_state_dict(model)
+                    text_encoder_two_lora_layers_to_save = convert_state_dict_to_diffusers(
+                        get_peft_model_state_dict(model)
+                    )
                 else:
                     raise ValueError(f"unexpected save model: {model.__class__}")
 
@@ -1615,13 +1619,17 @@ def main(args):
     if accelerator.is_main_process:
         unet = accelerator.unwrap_model(unet)
         unet = unet.to(torch.float32)
-        unet_lora_layers = get_peft_model_state_dict(unet)
+        unet_lora_layers = convert_state_dict_to_diffusers(get_peft_model_state_dict(unet))
 
         if args.train_text_encoder:
             text_encoder_one = accelerator.unwrap_model(text_encoder_one)
-            text_encoder_lora_layers = get_peft_model_state_dict(text_encoder_one.to(torch.float32))
+            text_encoder_lora_layers = convert_state_dict_to_diffusers(
+                get_peft_model_state_dict(text_encoder_one.to(torch.float32))
+            )
             text_encoder_two = accelerator.unwrap_model(text_encoder_two)
-            text_encoder_2_lora_layers = get_peft_model_state_dict(text_encoder_two.to(torch.float32))
+            text_encoder_2_lora_layers = convert_state_dict_to_diffusers(
+                get_peft_model_state_dict(text_encoder_two.to(torch.float32))
+            )
         else:
             text_encoder_lora_layers = None
             text_encoder_2_lora_layers = None

src/diffusers/utils/state_dict_utils.py

@@ -79,6 +79,14 @@ PEFT_TO_DIFFUSERS = {
     ".v_proj.lora_A": ".v_proj.lora_linear_layer.down",
     ".out_proj.lora_B": ".out_proj.lora_linear_layer.up",
     ".out_proj.lora_A": ".out_proj.lora_linear_layer.down",
+    "to_k.lora_A": "to_k.lora.down",
+    "to_k.lora_B": "to_k.lora.up",
+    "to_q.lora_A": "to_q.lora.down",
+    "to_q.lora_B": "to_q.lora.up",
+    "to_v.lora_A": "to_v.lora.down",
+    "to_v.lora_B": "to_v.lora.up",
+    "to_out.0.lora_A": "to_out.0.lora.down",
+    "to_out.0.lora_B": "to_out.0.lora.up",
 }
 
 DIFFUSERS_OLD_TO_DIFFUSERS = {