Soodoshll focused on stabilizing long-sequence processing in the NVIDIA-NeMo/RL repository, addressing critical memory management and numerical stability issues in DeepScaler. By optimizing memory usage and explicitly deleting intermediate tensors after use, Soodoshll reduced out-of-memory risks for sequence lengths up to 24,000 tokens. The work involved casting logits to float32 before log-probability calculations, improving both memory efficiency and numerical reliability. Using Python and YAML, Soodoshll’s targeted bug fix enhanced production robustness for large language model inference. This contribution demonstrated depth in deep learning and model optimization, laying a foundation for future scalability improvements in long-context workloads within the repository.