python -m training.visualize_multilayer_features \
    --sae_ckpt  training/multilayer_sae_ckpt/last.ckpt \
    --model_name llava-hf/llava-1.5-7b-hf \
    --data_dir  COCO-Dataset/val \
    --output_dir training/outputs_features \
    --device_id 0 \
    --top_features 20 \
    --max_new_tokens 64 \
    --batch_size 8

    # [--feature_ids 0 1 2]   # explicit feature IDs; otherwise ranks top-N \\
    # [--top_features 20] \\
    # [--max_new_tokens 64] \\
    # [--batch_size 2] \\
    # [--top_hooks 5]          # how many most-active hooks to show per feature