Resolving OOM Issues in train_control.py on 4 A100 GPUs (80GB Each)

[Beijia11]

My settings are as follows. Is there any way to reduce OOM (Out of Memory) issues?
CUDA_VISIBLE_DEVICES=0,1,2,3 accelerate launch --main_process_ip=10.140.37.162 --main_process_port=12345 --num_machines=1 --num_processes=4 --machine_rank=0 scripts/train_control.py
--pretrained_model_name_or_path=$MODEL_NAME
--train_data_dir=$DATASET_NAME
--train_data_meta=$DATASET_META_NAME
--image_sample_size=1024
--video_sample_size=256
--token_sample_size=512
--video_sample_stride=3
--video_sample_n_frames=49
--train_batch_size=1
--video_repeat=1
--gradient_accumulation_steps=4
--dataloader_num_workers=1
--num_train_epochs=100
--checkpointing_steps=50
--learning_rate=2e-05
--lr_scheduler="constant_with_warmup"
--lr_warmup_steps=50
--seed=43
--output_dir="output_dir"
--gradient_checkpointing
--mixed_precision="bf16"
--adam_weight_decay=3e-2
--adam_epsilon=1e-10
--vae_mini_batch=1
--max_grad_norm=0.05
--random_hw_adapt
--training_with_video_token_length
--enable_bucket
--trainable_modules "."