BEGIN:VCALENDAR
VERSION:2.0
PRODID:Linklings LLC
BEGIN:VTIMEZONE
TZID:America/Los_Angeles
X-LIC-LOCATION:America/Los_Angeles
BEGIN:DAYLIGHT
TZOFFSETFROM:-0800
TZOFFSETTO:-0700
TZNAME:PDT
DTSTART:19700308T020000
RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=2SU
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:-0700
TZOFFSETTO:-0800
TZNAME:PST
DTSTART:19701101T020000
RRULE:FREQ=YEARLY;BYMONTH=11;BYDAY=1SU
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
DTSTAMP:20260730T152640Z
LOCATION:
DTSTART;TZID=America/Los_Angeles:20260728T173700
DTEND;TZID=America/Los_Angeles:20260728T173700
UID:dac_DAC 2026_sess317_RESEARCH484@linklings.com
SUMMARY:SmartSwap: Swap-Based Memory Optimization for LLM Training under V
 arying Operator Sequences
DESCRIPTION:Zibo Wang, Yuhang Zhou, Zhibin Wang, and Shipeng Li (Nanjing U
 niversity); Xinjing Huang, Chendong Cai, Bingxu Mu, Yuqing Sun, Zhiheng Hu
 , Bin She, Shu You, and Guanghuan Fang (Huawei); and Rong Gu, Wanchun Dou,
  Guihai Chen, and Chen Tian (Nanjing University)\n\nThe increasing size of
  large language models (LLMs) has led to a surge in memory requirements du
 ring training, often exceeding the capacity of high-bandwidth memory (HBM)
 . Swap-based memory optimization incurs neither accuracy loss nor addition
 al end-to-end overhead when effectively overlapped, thus being an attracti
 ve solution. However, existing swap methods assume consistent operator seq
 uences, which is impractical in Eager Mode, where operator sequences can v
 ary across iterations.\n\nWe propose SmartSwap, which redesigns the end-to
 -end process of swap-based memory optimization and is the first work to co
 nsider varying operator sequences in Eager Mode. SmartSwap (i) introduces 
 a lightweight online profiler to enable continuous profiling for monitorin
 g operator sequences, (ii) generates effective swap policies with limited 
 operator information, and (iii) optimizes the policy execution module for 
 accurate policy application and better performance. Experimental results d
 emonstrate that SmartSwap reduces profiling overhead by 84.25%, enables tr
 aining models up to 4× larger than hardware memory while adapting to chang
 es in operator sequences, improves performance by up to 38.94% compared to
  recomputation or high-degree parallelism.\n\nTrack: EDA7-II. Physical Des
 ign and Verification, Design\n\nSession Chairs: Christian Pilato (Politecn
 ico di Milano); Mohsen Imani (University of California, Irvine); Flavio Po
 nzina (San Diego State University); Amin Farshidi (Cadence Design Systems,
  Inc.); Yonggan Fu (Georgia Institute of Technology, NVIDIA); Dimitrios Ts
 aras (Huawei, Hong Kong University of Science and Technology (HKUST)); Anh
  Nguyen (Fujitsu Limited); Ruiyang Qin (Villanova University); Jiliang Zha
 ng (Hunan University); Meng Li (Peking University); and Rassul Bairamkulov
  (Advanced Micro Devices (AMD))\n\n
END:VEVENT
END:VCALENDAR