variable-length-sequences

Introduces Data-Centric Parallel (DCP), a method for training deep learning models on variable long sequences by dynamically adjusting runtime settings per batch, achieving up to 2.88x speedup on 32 H200 GPUs with only 10 lines of code integration.