Worked on improving inference reliability and consistency across the PaddlePaddle and PaddleNLP repositories by addressing normalization and quantization behaviors in deep learning models. Refactored fused normalization operations in both static and dynamic modes, ensuring output structures matched and reducing debugging complexity. Added FP8 quantization support for fused bias activation, introducing a helper function to handle FP8 outputs while maintaining compatibility with existing quantization logic. Standardized normalization outputs across multiple Transformer model families, such as LLaMA and Qwen, to harmonize behavior. Utilized Python and C++ with a focus on CUDA, neural network operations, and model optimization throughout the development process.