PD-Swap: Efficient LLM Inference on Edge FPGAs via Dynamic Partial Reconfiguration

This research paper introduces PD-Swap, a novel approach for optimizing Large Language Model (LLM) inference on edge FPGAs. The technique focuses on dynamic partial reconfiguration to improve efficiency.

• •PD-Swap aims to enhance LLM inference performance on edge devices.
• •The core methodology involves swapping prefill and decode logic using dynamic partial reconfiguration.
• •The work targets improving efficiency for resource-constrained edge FPGA platforms.