train with Wan2.2_I2V_A14B will encounter the Wan2.1-T2V-1.3B, although the Wan2.2 local ckpts paths are given.

pretty curious.

#!/bin/bash
# set -ex

accelerate launch \
  --main_process_ip $MASTER_ADDR \
  --main_process_port $MASTER_PORT \
  --machine_rank $RANK \
  --num_machines $WORLD_SIZE \
  --num_processes $WORLD_SIZE \
  --multi_gpu \
  --mixed_precision bf16 \
  --num_cpu_threads_per_process 1 \
  examples/wanvideo/model_training/train.py \
  --dataset_base_path /home/work/data/ \
  --dataset_metadata_path /home/work/data/metadata.csv \
  --height 544 \
  --width 960 \
  --num_frames 37 \
  --dataset_repeat 100 \
  --model_paths '[
      [
          "/home/work/DiffSynth-Studio/Wan-AI/Wan2.2-I2V-A14B/high_noise_model/diffusion_pytorch_model-00001-of-00006.safetensors",
          "/home/work/DiffSynth-Studio/Wan-AI/Wan2.2-I2V-A14B/high_noise_model/diffusion_pytorch_model-00002-of-00006.safetensors",
          "/home/work/DiffSynth-Studio/Wan-AI/Wan2.2-I2V-A14B/high_noise_model/diffusion_pytorch_model-00003-of-00006.safetensors",
          "/home/work/DiffSynth-Studio/Wan-AI/Wan2.2-I2V-A14B/high_noise_model/diffusion_pytorch_model-00004-of-00006.safetensors",
          "/home/work/DiffSynth-Studio/Wan-AI/Wan2.2-I2V-A14B/high_noise_model/diffusion_pytorch_model-00005-of-00006.safetensors",
          "/home/work/DiffSynth-Studio/Wan-AI/Wan2.2-I2V-A14B/high_noise_model/diffusion_pytorch_model-00006-of-00006.safetensors"
      ],
      "/home/work/DiffSynth-Studio/Wan-AI/Wan2.2-I2V-A14B/models_t5_umt5-xxl-enc-bf16.pth",
      "/home/work/DiffSynth-Studio/Wan-AI/Wan2.2-I2V-A14B/Wan2.1_VAE.pth"
  ]' \
  --learning_rate 1e-4 \
  --num_epochs 5 \
  --remove_prefix_in_ckpt "pipe.dit." \
  --output_path "/home/work/data_OUTPUT/Wan2.2-I2V-A14B_high_noise_lora" \
  --lora_base_model "dit" \
  --lora_target_modules "q,k,v,o,ffn.0,ffn.2" \
  --lora_rank 32 \
  --extra_inputs "input_image" \
  --max_timestep_boundary 0.358 \
  --min_timestep_boundary 0

Start with examples/wanvideo/model_training/train.py and reproduce the provided accelerate command with the Wan2.2-I2V-A14B paths. Trace how --model_paths is parsed and which checkpoint is selected; done means training uses the supplied Wan2.2 model rather than Wan2.1-T2V-1.3B.