cuAether Assistant: Using LLMs to rewrite UDFs for GPU acceleration

The RAPIDS Accelerator for Apache Spark enables GPU-accelerated query execution for data processing workloads. However, a common performance bottleneck for vectorized query engines are User-Defined Functions (UDFs). Since UDFs contain opaque custom code, the engine must fall back to row-at-a-time CPU execution and incur the overhead of GPU-to-CPU data transfer.We present cuAether Assistant, an LLM-powered developer tool that converts Spark UDFs into GPU-compatible implementations: either native SQL expressions or custom RapidsUDFs leveraging cuDF columnar APIs. The tool provides an end-to-end workflow: 1) generating unit tests, 2) converting the UDF with RAG-augmented reasoning, and 3) benchmarking the result, with verification through iterative feedback loops.By transforming UDFs into forms that can be accelerated on GPUs, cuAether Assistant enables full query acceleration for the RAPIDS Accelerator for Apache Spark, unlocking significant UDF speedups (20x+).