export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True

accelerate launch --num_cpu_threads_per_process 2 --mixed_precision bf16 ../src/musubi_tuner/fpack_train_network.py \
    --dit /mnt/workspace/checkpoints/lllyasviel/FramePackI2V_HY/diffusion_pytorch_model-00001-of-00003.safetensors \
    --vae /mnt/workspace/checkpoints/hunyuanvideo-community/HunyuanVideo/vae/diffusion_pytorch_model.safetensors \
    --text_encoder1 /mnt/workspace/checkpoints/hunyuanvideo-community/HunyuanVideo/text_encoder/model-00001-of-00004.safetensors \
    --text_encoder2 /mnt/workspace/checkpoints/hunyuanvideo-community/HunyuanVideo/text_encoder_2/model.safetensors \
    --image_encoder /mnt/workspace/checkpoints/lllyasviel/flux_redux_bfl/image_encoder/model.safetensors \
    --dataset_config /mnt/workspace/ysh/Code/Efficient_Model/2_code/musubi-tuner/src/dataset_config.toml \
    --mixed_precision bf16 \
    --optimizer_type adamw8bit --learning_rate 2e-3 --gradient_checkpointing \
    --timestep_sampling shift --weighting_scheme none --discrete_flow_shift 3.0 \
    --max_data_loader_n_workers 8 --persistent_data_loader_workers \
    --network_module networks.lora_framepack --network_dim 128 \
    --max_train_epochs 100 --save_every_n_epochs 10 --seed 42 \
    --output_dir /mnt/workspace/ysh/Code/Efficient_Model/2_code/musubi-tuner/demo_output \
    --output_name name-of-lora \
    --flash_attn