This section addresses how to enhance the efficiency of computing clusters when users lack expertise in framework-level optimization.

Since key Point data within Transformer models is already fixed during the algorithm process, attempting improvements at the model layer demands significant manpower and time investment.

Professor Liu suggests that the most effective and practical solution is to boost performance from the storage side.

By employing high-performance software algorithms to improve storage system efficiency—particularly for Checkpoint data writes—organizations can significantly reduce operational costs while achieving better cluster utilization.