finalf0's picture
Add vision_batch_size to avoid cuda OOM
74e6636