Add --lora_alpha and metadata handling to train_dreambooth_lora_sd3.py

examples/dreambooth/test_dreambooth_lora_sd3.py

@@ -13,13 +13,15 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+import json
 import logging
 import os
 import sys
 import tempfile
 
 import safetensors
 
+from diffusers.loaders.lora_base import LORA_ADAPTER_METADATA_KEY
 
 sys.path.append("..")
 from test_examples_utils import ExamplesTestsAccelerate, run_command  # noqa: E402
@@ -207,6 +209,46 @@ def test_dreambooth_lora_layer(self):
             starts_with_transformer = all("attn.to_k" in key for key in lora_state_dict.keys())
             self.assertTrue(starts_with_transformer)
 
+    def test_dreambooth_lora_sd3_with_metadata(self):
+        lora_alpha = 8
+        rank = 4
+        with tempfile.TemporaryDirectory() as tmpdir:
+            test_args = f"""
+            {self.script_path}
+            --pretrained_model_name_or_path={self.pretrained_model_name_or_path}
+            --instance_data_dir={self.instance_data_dir}
+            --output_dir={tmpdir}
+            --resolution=32
+            --train_batch_size=1
+            --gradient_accumulation_steps=1
+            --max_train_steps=4
+            --lora_alpha={lora_alpha}
+            --rank={rank}
+            --checkpointing_steps=2
+            --max_sequence_length 166
+            """.split()
+
+            test_args.extend(["--instance_prompt", ""])
+            run_command(self._launch_args + test_args)
+
+            state_dict_file = os.path.join(tmpdir, "pytorch_lora_weights.safetensors")
+            self.assertTrue(os.path.isfile(state_dict_file))
+
+            # Check if the metadata was properly serialized.
+            with safetensors.torch.safe_open(state_dict_file, framework="pt", device="cpu") as f:
+                metadata = f.metadata() or {}
+
+            metadata.pop("format", None)
+            raw = metadata.get(LORA_ADAPTER_METADATA_KEY)
+            if raw:
+                raw = json.loads(raw)
+
+            loaded_lora_alpha = raw["transformer.lora_alpha"]
+            self.assertTrue(loaded_lora_alpha == lora_alpha)
+            loaded_lora_rank = raw["transformer.r"]
+            self.assertTrue(loaded_lora_rank == rank)
+
+
     def test_dreambooth_lora_sd3_checkpointing_checkpoints_total_limit(self):
         with tempfile.TemporaryDirectory() as tmpdir:
             test_args = f"""

examples/dreambooth/train_dreambooth_lora_sd3.py

@@ -53,6 +53,7 @@
 )
 from diffusers.optimization import get_scheduler
 from diffusers.training_utils import (
+    _collate_lora_metadata,
     _set_state_dict_into_text_encoder,
     cast_training_params,
     compute_density_for_timestep_sampling,
@@ -321,6 +322,12 @@ def parse_args(input_args=None):
         required=False,
         help="A folder containing the training data of class images.",
     )
+    parser.add_argument(
+        "--lora_alpha",
+        type=int,
+        default=4,
+        help="LoRA alpha to be used for additional scaling.",
+    )
     parser.add_argument(
         "--instance_prompt",
         type=str,
@@ -1266,7 +1273,7 @@ def main(args):
     # now we will add new LoRA weights to the attention layers
     transformer_lora_config = LoraConfig(
         r=args.rank,
-        lora_alpha=args.rank,
+        lora_alpha=args.lora_alpha,
         lora_dropout=args.lora_dropout,
         init_lora_weights="gaussian",
         target_modules=target_modules,
@@ -1295,13 +1302,15 @@ def save_model_hook(models, weights, output_dir):
             transformer_lora_layers_to_save = None
             text_encoder_one_lora_layers_to_save = None
             text_encoder_two_lora_layers_to_save = None
+            modules_to_save = {}
 
             for model in models:
                 if isinstance(unwrap_model(model), type(unwrap_model(transformer))):
                     model = unwrap_model(model)
                     if args.upcast_before_saving:
                         model = model.to(torch.float32)
                     transformer_lora_layers_to_save = get_peft_model_state_dict(model)
+                    modules_to_save["transformer"] = model
                 elif args.train_text_encoder and isinstance(
                     unwrap_model(model), type(unwrap_model(text_encoder_one))
                 ):  # or text_encoder_two
@@ -1324,6 +1333,7 @@ def save_model_hook(models, weights, output_dir):
                 transformer_lora_layers=transformer_lora_layers_to_save,
                 text_encoder_lora_layers=text_encoder_one_lora_layers_to_save,
                 text_encoder_2_lora_layers=text_encoder_two_lora_layers_to_save,
+                **_collate_lora_metadata(modules_to_save),
             )
 
     def load_model_hook(models, input_dir):
@@ -1925,10 +1935,12 @@ def get_sigmas(timesteps, n_dim=4, dtype=torch.float32):
     accelerator.wait_for_everyone()
     if accelerator.is_main_process:
         transformer = unwrap_model(transformer)
+        modules_to_save = {}
         if args.upcast_before_saving:
             transformer.to(torch.float32)
         else:
             transformer = transformer.to(weight_dtype)
+        modules_to_save["transformer"] = transformer
         transformer_lora_layers = get_peft_model_state_dict(transformer)
 
         if args.train_text_encoder:
@@ -1945,6 +1957,7 @@ def get_sigmas(timesteps, n_dim=4, dtype=torch.float32):
             transformer_lora_layers=transformer_lora_layers,
             text_encoder_lora_layers=text_encoder_lora_layers,
             text_encoder_2_lora_layers=text_encoder_2_lora_layers,
+            **_collate_lora_metadata(modules_to_save),
         )
 
         # Final inference