When I trained the Wan-VACE-14B model with LoRA, an OOM occurred.

An OOM error occurred when I used the official LoRA fine-tuning code to train VACE-14B. My code script is as follows:

accelerate launch examples/wanvideo/model_training/train.py \
  --dataset_base_path xxx(my datasets root path) \
  --dataset_metadata_path xxx(my datasets path) \
  --data_file_keys "video,vace_video,vace_reference_image" \
  --height 480 \
  --width 832 \
  --num_frames 81\
  --model_id_with_origin_paths "Wan-AI/Wan2.1-VACE-14B:diffusion_pytorch_model*.safetensors,Wan-AI/Wan2.1-VACE-14B:models_t5_umt5-xxl-enc-bf16.pth,Wan-AI/Wan2.1-VACE-14B:Wan2.1_VAE.pth" \
  --learning_rate 1e-4 \
  --num_epochs 200 \
  --remove_prefix_in_ckpt "pipe.vace." \
  --output_path "./models/train/Wan2.1-VACE-14B_lora" \
  --lora_base_model "vace" \
  --lora_target_modules "q,k,v,o,ffn.0,ffn.2" \
  --lora_rank 16\
  --extra_inputs "vace_video,vace_reference_image" \
  --use_gradient_checkpointing_offload

When I was testing, I found that num_frames could only reach a maximum of 25 under 8 A100 frames. An OOM error will occur if it exceeds 25. How can I solve this problem? Otherwise, only the first 25 frames of the video can be input. If frame extraction training is conducted, the problem of camera shake will occur. （我在测试的时候发现，num_frames在8张A100下面最多只能到25。超过25就会发生OOM错误。我该如何解决这个问题？否则只能输入视频的前25帧，如果抽帧训练会出现镜头抖动的问题。）

Start with examples/wanvideo/model_training/train.py and the provided accelerate launch command, then compare memory behavior as num_frames increases beyond 25 with 8 A100 GPUs. Done means identifying a reproducible cause or configuration change that permits longer VACE-14B LoRA training without OOM, while preserving the stated video inputs.