Work

LORA: A Latency-Oriented Recurrent Architecture for GPT Model on Multi-FPGA Platform with Communication Optimization