Vijay Janapa Reddi

It is shown that a misaligned model can perform inference engine fingerprinting to determine the specific engine (e.g., vLLM, SGLang) which executes the model, and several ways that inference engines could be changed to make fingerprinting attacks more difficult are discussed.

Zero-shot Vision-Language Models can now guide chip floorplanning, beating specialized ML methods by up to 32% without any fine-tuning.

Generative AI in systems design faces the same five challenges—no matter the layer—and surprisingly, the same five design principles keep emerging as solutions.