Worked on the modelscope/ms-swift repository to enhance the stability of multi-GPU training on AMD ROCm platforms. Focused on resolving an inter-process communication issue in asynchronous mode, the developer implemented a targeted fix to ensure correct GPU visibility and prevent device ordinal errors during distributed training and rollout communications. Leveraging expertise in Python, concurrency, and GPU programming, the solution improved cross-GPU coordination and reduced failure modes related to IPC in ROCm environments. Additional regression checks and improved logging were introduced to aid future debugging, supporting more reliable model training pipelines and smoother production deployments on AMD-based systems.