Peking University and DeepSeek Open-Source DSpark, Delivering Major Leap in LLM Inference Efficiency

Peking University and DeepSeek have launched DSpark, a new framework designed to enhance the efficiency of large language model inference. DSpark achieves a text generation acceleration of 60-85% under real server loads, with productivity improvements of up to 661% under strict time constraints. DSpark employs a dual optimization mechanism, utilizing a semi-automated architecture that combines a parallel backend with lightweight sequential units. This design allows it to outperform traditional models, as evidenced by DSpark's two Transformer layers surpassing the performance of five layers in conventional models.