During October 2025, Andexpouni developed robust cross-device pixel value dtype handling for the bytedance/Dolphin repository, focusing on seamless execution across CPU and CUDA environments. Leveraging Python and CUDA, Andexpouni implemented explicit per-device path management in the DOLPHIN class, ensuring CPU operations use float32 while GPU operations use float16, with safe conversions between them. This approach addressed dtype mismatch and bias-type errors, reducing runtime failures and stabilizing multi-device workflows. The work included targeted refactoring for compatibility and maintainability, particularly in demo components, and laid a solid foundation for future performance optimizations in data processing and machine learning pipelines.