train_14B_single_node.sh

# We have tested this code on 5-second multi-shot data with a single GPU. 
# If you want train 14B model on longer video data, you need to implement sequence parallel on our code.

pkill python
source /m2v_intern/wangqinghe/miniconda3/bin/activate wan_multishot

# resume
RESUME_CKPT=None # "checkpoints/download_models/14B/MultiShotMaster_model_14B.safetensors"

# output dir
TIMESTAMP=$(date +%Y%m%d_%H%M)
OUTPUT_DIR="models/train/14B/${TIMESTAMP}-T2V-1.3B_full"

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 accelerate launch --num_processes=8 --multi_gpu train.py \
  --dataset_metadata_path "toy_cases/train_multishot_5s.csv" \
  --height 480 \
  --width 832 \
  --dataset_repeat 1 \
  --learning_rate 1e-5 \
  --num_epochs 100 \
  --remove_prefix_in_ckpt "pipe.dit." \
  --output_path "${OUTPUT_DIR}" \
  --trainable_models "dit" \
  --save_steps 500 \
  --wan_version "t2v-14B" \
  --dataset_num_workers 8 \
  --resume_ckpt "$RESUME_CKPT" \
  --load_path_json "checkpoints/model_configs/model_path_14B.json"