Danying Ge Presented at Euro-Par 2026
Published:
From August 24–28, 2026, the 32nd International European Conference on Parallel and Distributed Computing (Euro-Par 2026) was held in Pisa, Italy. Danying Ge, a doctoral student in the group, attended the conference and gave a presentation titled ApproxRAG.
About the Conference
Euro-Par is Europe’s flagship academic conference in the field of parallel and distributed computing. Held over a full week, it covers the field’s complete research spectrum — from fundamental algorithms to end-to-end applications, from low-power edge devices to supercomputing and cloud-edge collaboration, and from architectural breakthroughs and compilation innovations to emerging programming models, performance engineering, reproducibility, and sustainability — providing scholars worldwide with an ample platform for academic exchange.
Presentation Content
Danying Ge presented the research work ApproxRAG, an inference acceleration technique for retrieval-augmented generation (RAG). To address the high inference latency and heavy disk I/O overhead that existing RAG techniques face when processing large volumes of external documents, this work proposes a method that jointly optimizes data movement among disk, host memory, and device memory. The presentation elaborated on three core techniques of the framework:
Hybrid compression strategy: reducing the disk storage overhead of the key-value cache (KV Cache).
Semantic-similarity-based KV Cache reuse mechanism: reducing unnecessary disk accesses.
Cache eviction strategy: building an eviction policy that combines access frequency, temporal locality, and retrieval score to improve memory resource utilization.
Experimental Results
The experiments were based on the LLaMA 2-7B and LLaMA 2-13B models and validated using a Wikipedia knowledge base along with several question-answering benchmarks. The results show that ApproxRAG achieves up to 3.66× inference speedup while maintaining text generation quality, demonstrating its potential for efficient large-model inference and storage optimization.
Highlights
