accelerate launch --use_deepspeed --num_processes 32 \

--num_machines ${WORLD_SIZE} \

--machine_rank ${RANK} \

--main_process_ip ${MASTER_ADDR} \

--main_process_port ${MASTER_PORT} \

--mixed_precision "bf16" \

--deepspeed_multinode_launcher standard \

--zero_stage 2 --offload_param_device none --offload_optimizer_device none --gradient_accumulation_steps 4 --zero3_init_flag false \

finetune_sdxl.py \

--pretrained_model_name_or_path="stabilityai/stable-diffusion-xl-base-1.0" \

--pretrained_easy_ref_path="{outputs/single_ref_finetuning/checkpoint-250000/pytorch_model/mp_rank_00_model_states.pt}" \

--image_encoder_path="Qwen/Qwen2-VL-2B-Instruct" \

--data_json_file="{data.json}" \

--data_root_path="{image_path}" \

--mixed_precision="bf16" \

--resolution=1024 \

--train_batch_size=1 \

--dataloader_num_workers=2 \

--learning_rate=2e-05 \

--weight_decay=0.0 \

--lr_warmup_steps=0 \

--t_drop_rate=0.05 \

--truncate_rate=0.4 \

--num_train_epochs=100 \

--gradient_accumulation_steps=4 \

--num_reference_tokens=64 \

--multi_ref_finetuning \

--unfreeze_mllm \

--use_lora \

--lora_rank=128 \

--output_dir="{outputs/multi_ref_finetuning}" \

--save_steps=20000