fix bugs when using deepspeed in sdxl (#7917)

fix bugs when using deepspeed
Co-authored-by: mhh001 <mahonghao1@huawei.com>

examples/text_to_image/train_text_to_image_sdxl.py

@@ -35,7 +35,7 @@ import torch.utils.checkpoint
 import transformers
 from accelerate import Accelerator
 from accelerate.logging import get_logger
-from accelerate.utils import ProjectConfiguration, set_seed
+from accelerate.utils import DistributedType, ProjectConfiguration, set_seed
 from datasets import concatenate_datasets, load_dataset
 from huggingface_hub import create_repo, upload_folder
 from packaging import version
@@ -742,7 +742,8 @@ def main(args):
                     model.save_pretrained(os.path.join(output_dir, "unet"))
 
                     # make sure to pop weight so that corresponding model is not saved again
-                    weights.pop()
+                    if weights:
+                        weights.pop()
 
         def load_model_hook(models, input_dir):
             if args.use_ema:
@@ -914,7 +915,7 @@ def main(args):
         train_dataset_with_vae = train_dataset.map(
             compute_vae_encodings_fn,
             batched=True,
-            batch_size=args.train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps,
+            batch_size=args.train_batch_size,
             new_fingerprint=new_fingerprint_for_vae,
         )
         precomputed_dataset = concatenate_datasets(
@@ -1160,7 +1161,8 @@ def main(args):
                 accelerator.log({"train_loss": train_loss}, step=global_step)
                 train_loss = 0.0
 
-                if accelerator.is_main_process:
+                # DeepSpeed requires saving weights on every device; saving weights only on the main process would cause issues.
+                if accelerator.distributed_type == DistributedType.DEEPSPEED or accelerator.is_main_process:
                     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: