accelerate launch --num_processes 8 --multi_gpu --mixed_precision "fp16" --main_process_port 29500 \
  -m src.train.distil_sdxl \
  --pretrained_model_name_or_path="stabilityai/stable-diffusion-xl-base-1.0" \
  --mixed_precision="fp16" \
  --resolution=1024 \
  --train_batch_size=2 \
  --gradient_accumulation_steps=4 \
  --dataloader_num_workers=6 \
  --learning_rate=1e-04 \
  --weight_decay=0. \
  --output_dir="ckpt/linfusion_sdxl" \
  --save_steps=10000 \
  --mid_dim_scale=16