Sanghyun Han
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Sanghyun Han has a total of 15 co-authors. Here are the top co-authors by number of shared publications:
- Kyunghan Lee
Seoul National University, Seoul, Republic of Korea
4 shared publications - Gibum Park
Seoul National University, Seoul, Republic of Korea
3 shared publications - Chanjeong Park
Seoul National University, Seoul, Republic of Korea
2 shared publications - Yonghwa Cho
Seoul National University, Seoul, Republic of Korea
2 shared publications - Byunggu Kang
Seoul National University, Seoul, Republic of Korea
1 shared publication - Daewook Hwang
Seoul National University, Seoul, Republic of Korea
1 shared publication
Latest Publications
- 7 October 2026
- Published by Elsevier in Neurocomputing
- Vol. 709
BVSR: BitVector-sum representation for distribution-aware LLM weight quantization
- 7 October 2026
- Published by Elsevier in Neurocomputing
- Vol. 709
BVSR: BitVector-sum representation for distribution-aware LLM weight quantization
- Open Access
- 20 June 2026
- Published by Association for Computing Machinery (ACM)
- p. 580-592
SAIL: Redesigning Collaborative Language Inference with a Single Server-to-Mobile Handoff
- 20 June 2026
- Published by Association for Computing Machinery (ACM)
- p. 580-592
Open AccessSAIL: Redesigning Collaborative Language Inference with a Single Server-to-Mobile Handoff
- Open Access
- 20 June 2026
- Published by Association for Computing Machinery (ACM)
- p. 23-24
Poster: Selective Semantic Cache Transfer for Mobile-Cloud LLM Inference
Direct key-value (KV) cache transfer between large language models (LLMs) is emerging as a richer alternative to token exchange, but the KV cache is orders of magnitude larger than tokens, and mobile-cloud inference systems are not yet prepared for this shift. We present CAIL, a system that reframes semantic cache delivery as a selection problem: rather than compressing or scheduling the entire cache, CAIL identifies which cache positions genuinely help the edge small language model (SLM) using an uncertainty-aware selection policy. Building on recent advances in cross-model cache projection (C2C) and server-assisted edge inference (SAIL), we ask which positions to send rather than how to compress them. Preliminary experiments on the WikiText-103 dataset show that selective transfer at 25% of positions outperforms full-cache transfer in perplexity, closing 28.7% of the SLM-to-LLM gap.
- 20 June 2026
- Published by Association for Computing Machinery (ACM)
- p. 23-24
Open AccessPoster: Selective Semantic Cache Transfer for Mobile-Cloud LLM Inference
Direct key-value (KV) cache transfer between large language models (LLMs) is emerging as a richer alternative to token exchange, but the KV cache is orders of magnitude larger than tokens, and mobile-cloud inference systems are not yet prepared for this shift. We present CAIL, a system that reframes semantic cache delivery as a selection problem: rather than compressing or scheduling the entire cache, CAIL identifies which cache positions genuinely help the edge small language model (SLM) using an uncertainty-aware selection policy. Building on recent advances in cross-model cache projection (C2C) and server-assisted edge inference (SAIL), we ask which positions to send rather than how to compress them. Preliminary experiments on the WikiText-103 dataset show that selective transfer at 25% of positions outperforms full-cache transfer in perplexity, closing 28.7% of the SLM-to-LLM gap.
- Open Access
- 1 June 2026
- Published by Association for Computing Machinery (ACM) in Proceedings of the ACM on Networking
- Vol. 4 (CoNEXT2) , 1-23
PRC: Process-centric Rate Control for Transient Stragglers in Hybrid-Parallel Training
Stragglers remain an open challenge in large model training, forcing other GPUs to wait at synchronization barriers. While severe, persistent stragglers have been actively studied, their transient counterparts are often overlooked due to their short-lived nature. We reveal that the interplay between the transient straggler effects and network contention drastically amplifies the relative delays across workers, significantly slowing down training iterations. To address this, an intuitive approach is to prioritize bandwidth for the straggler, allowing it to catch up to the front-runners. However, while the transport layer is well-positioned to accommodate such responsive bandwidth control, conventional designs lack visibility into process-level progress. To bridge this gap, we propose PRC (Process-centric Rate Control), a new sending-rate control designed to mitigate the transient straggler effects. PRC adjusts NIC sending rates by inferring local GPU process-level information, enabling transient straggler processes to utilize more bandwidth on time. Extensive experiments on both a real-world cluster and large-scale simulations confirm that PRC effectively accelerates transient stragglers, achieving a training speedup of up to 28% compared to using state-of-the-art datacenter congestion control schemes.
- 1 June 2026
- Published by Association for Computing Machinery (ACM) in Proceedings of the ACM on Networking
- Vol. 4 (CoNEXT2) , 1-23
Open AccessPRC: Process-centric Rate Control for Transient Stragglers in Hybrid-Parallel Training
Stragglers remain an open challenge in large model training, forcing other GPUs to wait at synchronization barriers. While severe, persistent stragglers have been actively studied, their transient counterparts are often overlooked due to their short-lived nature. We reveal that the interplay between the transient straggler effects and network contention drastically amplifies the relative delays across workers, significantly slowing down training iterations. To address this, an intuitive approach is to prioritize bandwidth for the straggler, allowing it to catch up to the front-runners. However, while the transport layer is well-positioned to accommodate such responsive bandwidth control, conventional designs lack visibility into process-level progress. To bridge this gap, we propose PRC (Process-centric Rate Control), a new sending-rate control designed to mitigate the transient straggler effects. PRC adjusts NIC sending rates by inferring local GPU process-level information, enabling transient straggler processes to utilize more bandwidth on time. Extensive experiments on both a real-world cluster and large-scale simulations confirm that PRC effectively accelerates transient stragglers, achieving a training speedup of up to 28% compared to using state-of-the-art datacenter congestion control schemes.