RuntimeError: cuDNN error: CUDNN_STATUS_NOT_INITIALIZED

报错：RuntimeError: cuDNN error: CUDNN_STATUS_NOT_INITIALIZED
原因：

accelerate launch --num_processes 3 --multi_gpu --mixed_precision "fp16" \
  tutorial_train.py \
  --pretrained_model_name_or_path="/data0/JM/code/AniPortrait/pretrained_model/stable-diffusion-v1-5/" \
  --image_encoder_path="/data0/JM/code/IP-Adapter/pretrain_model/models--h94--IP-Adapter/models/image_encoder" \
  --data_json_file="/data0/JM/code/lang-segment-anything/process_data/lama_data/chair/data.json" \
  --mixed_precision="fp16" \
  --resolution=512 \
  --train_batch_size=2 \
  --dataloader_num_workers=4 \
  --learning_rate=1e-04 \
  --weight_decay=0.01 \
  --output_dir="/data0/JM/code/IP-Adapter/exp" \
  --save_steps=10000

2号卡用不了，改成–num_processes 2 即可