Pradas developed a Philox-based random number generation context for HPU devices in the pytorch/pytorch repository, focusing on distributed tensor (Dtensor) scenarios. Using Python and CUDA, Pradas implemented device-specific RNG context management and an offset-based RNG tracker to improve randomness control and reproducibility in distributed training workflows. The solution enhanced integration of random operations with CUDA, ensuring reliable and scalable RNG behavior for HPUs in multi-device environments. This work addressed the need for robust random number generation in distributed computing, demonstrating depth in backend development and distributed systems while strengthening the RNG subsystem’s compatibility for advanced training scenarios.