[training] add ds support to lora hidream (huggingface#11737)

* [training] add ds support to lora hidream

* Apply style fixes

---------

Co-authored-by: J石页 <[email protected]>
Co-authored-by: Sayak Paul <[email protected]>
Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>

Author: 4 people

examples/dreambooth/train_dreambooth_lora_hidream.py

@@ -29,7 +29,7 @@
 import numpy as np
 import torch
 import transformers
-from accelerate import Accelerator
+from accelerate import Accelerator, DistributedType
 from accelerate.logging import get_logger
 from accelerate.utils import DistributedDataParallelKwargs, ProjectConfiguration, set_seed
 from huggingface_hub import create_repo, upload_folder
@@ -1181,13 +1181,15 @@ def save_model_hook(models, weights, output_dir):
             transformer_lora_layers_to_save = None
 
             for model in models:
-                if isinstance(model, type(unwrap_model(transformer))):
+                if isinstance(unwrap_model(model), type(unwrap_model(transformer))):
+                    model = unwrap_model(model)
                     transformer_lora_layers_to_save = get_peft_model_state_dict(model)
                 else:
                     raise ValueError(f"unexpected save model: {model.__class__}")
 
                 # make sure to pop weight so that corresponding model is not saved again
-                weights.pop()
+                if weights:
+                    weights.pop()
 
             HiDreamImagePipeline.save_lora_weights(
                 output_dir,
@@ -1197,13 +1199,20 @@ def save_model_hook(models, weights, output_dir):
     def load_model_hook(models, input_dir):
         transformer_ = None
 
-        while len(models) > 0:
-            model = models.pop()
+        if not accelerator.distributed_type == DistributedType.DEEPSPEED:
+            while len(models) > 0:
+                model = models.pop()
 
-            if isinstance(model, type(unwrap_model(transformer))):
-                transformer_ = model
-            else:
-                raise ValueError(f"unexpected save model: {model.__class__}")
+                if isinstance(unwrap_model(model), type(unwrap_model(transformer))):
+                    model = unwrap_model(model)
+                    transformer_ = model
+                else:
+                    raise ValueError(f"unexpected save model: {model.__class__}")
+        else:
+            transformer_ = HiDreamImageTransformer2DModel.from_pretrained(
+                args.pretrained_model_name_or_path, subfolder="transformer"
+            )
+            transformer_.add_adapter(transformer_lora_config)
 
         lora_state_dict = HiDreamImagePipeline.lora_state_dict(input_dir)
 
@@ -1655,7 +1664,7 @@ def get_sigmas(timesteps, n_dim=4, dtype=torch.float32):
                 progress_bar.update(1)
                 global_step += 1
 
-                if accelerator.is_main_process:
+                if accelerator.is_main_process or accelerator.distributed_type == DistributedType.DEEPSPEED:
                     if global_step % args.checkpointing_steps == 0:
                         # _before_ saving state, check if this save would set us over the `checkpoints_total_limit`
                         if args.checkpoints_total_limit is not None: