February 2025 monthly summary for PaddlePaddle/Paddle and PaddleNLP focused on XPU performance and data-type support. Key features delivered include BFLOAT16 support for XPU set_value_grad and set_value_with_scalar_grad kernels, and XPU-optimized DeepseekV2 with fused operations, RMS normalization improvements, rotary position embeddings optimizations, and refactored z-loss calculations in MoE gates for better numerical stability and hardware utilization. Major bugs fixed: none explicitly reported this month; the primary value came from feature work that also enhances stability and compatibility on XPU. Overall impact: accelerated training and inference on XPU devices, expanded data-type coverage, and improved hardware utilization for Paddle and PaddleNLP workloads. Technologies/skills demonstrated: XPU kernel development and integration, BFLOAT16 data path, fused operations, RMS normalization, rotary position embeddings, and MoE gate stabilization techniques.