Worked on stabilizing distributed benchmarking in the AMD-AGI/Primus repository by addressing a critical crash in the RCCL all-to-all benchmark. The solution involved deferring the use of torch.distributed.group.WORLD until after the init_process_group() call, ensuring proper initialization and preventing execution failures. This change improved the reliability of distributed computing workflows and reduced flakiness in continuous integration environments. The work focused on debugging and resolving a complex issue in Python code, leveraging expertise in benchmarking and distributed systems. No new features were added during this period, but the targeted bug fix enhanced the stability and maintainability of the benchmarking infrastructure.