The issue asks about using Wrap CTC on multiple GPUs. To address this, first examine the current CTC implementation in the repository (likely in the 'lstm ctc ocr' codebase). Investigate TensorFlow's distributed training capabilities (e.g., tf.distribute.MirroredStrategy) and consider how to adapt the existing CTC loss computation for multi GPU. Additionally, search for any existing discussions or related pull requests in the repository for context.