Developed end-to-end Mixture-of-Experts (MoE) routing replay support for NPU platforms in the volcengine/verl repository, focusing on deployment reliability and consistent training behavior. The work involved implementing NPU-compatible routing replay with Python, integrating compatibility patches for Megatron 0.12.1, and ensuring robust data alignment from rollout to training. Leveraged MindSpeed on Ascend NPUs to validate integration, while introducing dynamic signature detection and standardized rollout tokens to prevent shape mismatches. Routing metadata was preserved and propagated through the agent loop using safe attribute patterns, enabling deterministic rollout and reducing Python-level overhead during training on complex NPU-based machine learning systems.