Over a two-month period, contributed to the ROCm/jax and jax-ml/jax repositories by developing advanced attention mechanisms for deep learning workloads. Focused on enabling dynamic masking in splash attention, the work introduced conditional mask processing for JAX arrays and helper utilities to support flexible attention workflows. Further enhancements included sharded dynamic masks, allowing scalable distributed computation and relaxing head dimension constraints for broader applicability. Implemented using Python and JAX, these features improved memory and compute efficiency, expanded test coverage, and increased the scalability and flexibility of attention mechanisms for both training and inference in distributed GPU and TPU environments.