Worked on optimizing inference performance in the ROCm/FBGEMM repository by implementing a feature that accelerates KV cache prefill for transformer models. The approach involved developing CUDA kernels and new C++ functions to bypass Rotary Positional Embedding (RoPE) calculations during the KV cache fill process. By introducing nope_qkv_varseq_prefill and nope_qkv_decoding, the solution reduced cache fill overhead and improved throughput for FP32 and FP16 inference scenarios. This work focused on deep learning inference and performance optimization, laying the foundation for RoPE-free inference paths and enabling future tuning. No critical bugs were reported during this period, reflecting stable engineering practices.