[Training utils] add kohya conversion dict. (#7435)

* add kohya conversion dict.

* update readme

* typo

* add filename

examples/dreambooth/README_sdxl.md

@@ -259,9 +259,9 @@ The authors found that by using DoRA, both the learning capacity and training st
 > This is also aligned with some of the quantitative analysis shown in the paper. 
 
 **Usage**
-1. To use DoRA you need to install `peft` from main: 
+1. To use DoRA you need to upgrade the installation of `peft`: 
 ```bash
-pip install git+https://github.com/huggingface/peft.git
+pip install-U peft
 ```
 2. Enable DoRA training by adding this flag
 ```bash
@@ -269,3 +269,7 @@ pip install git+https://github.com/huggingface/peft.git
 ```
 **Inference** 
 The inference is the same as if you train a regular LoRA 🤗
+
+## Format compatibility
+
+You can pass `--output_kohya_format` to additionally generate a state dictionary which should be compatible with other platforms and tools such as Automatic 1111, Comfy, Kohya, etc. The `output_dir` will contain a file named "pytorch_lora_weights_kohya.safetensors".
\ No newline at end of file

examples/dreambooth/train_dreambooth_lora_sdxl.py

@@ -41,6 +41,7 @@ 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
+from safetensors.torch import load_file, save_file
 from torch.utils.data import Dataset
 from torchvision import transforms
 from torchvision.transforms.functional import crop
@@ -62,7 +63,9 @@ from diffusers.optimization import get_scheduler
 from diffusers.training_utils import _set_state_dict_into_text_encoder, cast_training_params, compute_snr
 from diffusers.utils import (
     check_min_version,
+    convert_all_state_dict_to_peft,
     convert_state_dict_to_diffusers,
+    convert_state_dict_to_kohya,
     convert_unet_state_dict_to_peft,
     is_wandb_available,
 )
@@ -396,6 +399,11 @@ def parse_args(input_args=None):
         default="lora-dreambooth-model",
         help="The output directory where the model predictions and checkpoints will be written.",
     )
+    parser.add_argument(
+        "--output_kohya_format",
+        action="store_true",
+        help="Flag to additionally generate final state dict in the Kohya format so that it becomes compatible with A111, Comfy, Kohya, etc.",
+    )
     parser.add_argument("--seed", type=int, default=None, help="A seed for reproducible training.")
     parser.add_argument(
         "--resolution",
@@ -1890,6 +1898,11 @@ def main(args):
             text_encoder_lora_layers=text_encoder_lora_layers,
             text_encoder_2_lora_layers=text_encoder_2_lora_layers,
         )
+        if args.output_kohya_format:
+            lora_state_dict = load_file(f"{args.output_dir}/pytorch_lora_weights.safetensors")
+            peft_state_dict = convert_all_state_dict_to_peft(lora_state_dict)
+            kohya_state_dict = convert_state_dict_to_kohya(peft_state_dict)
+            save_file(kohya_state_dict, f"{args.output_dir}/pytorch_lora_weights_kohya.safetensors")
 
         # Final inference
         # Load previous pipeline