During April 2026, contributed to the jeejeelee/vllm repository by implementing ROCm backend support for dflash, enabling both causal and non-causal attention decoding. This work involved updating the attention mechanism to include a causal flag in the metadata and optimizing the decoding logic to efficiently handle varying query lengths, which improved performance and flexibility across different datasets. The development process emphasized backend engineering and deep learning techniques, utilizing Python and PyTorch. All changes were traceable through signed commits, supporting robust code review and future ROCm deployment. No major bugs were addressed during this period, with focus placed on feature delivery.