Worked on the pytorch/pytorch repository to deliver a Philox-based random number generation (RNG) context for HPU devices in distributed tensor (Dtensor) scenarios. The implementation introduced device-specific RNG context management and an offset-based RNG tracker, improving control over randomness and ensuring reproducibility in distributed training workflows. By enhancing integration of random operations with CUDA, the work enabled more reliable and scalable training on HPUs within distributed computing environments. Leveraging Python and CUDA, the solution strengthened the RNG subsystem’s compatibility for HPUs, addressing the challenges of randomness management and reproducibility in backend development for large-scale machine learning systems.