I am currently a part of LDOS at UT-Austin. I graduated from UW-Madison, where I was Advised by Dimitris Papailiopoulos and Shivaram Venkataraman . My research interests are primarily in Systems for Machine Learning, especially around distributed training and inference of ML workloads. During my PhD I have been very fortunate to intern with Bilge Acun at FAIR, Amar Phanishayee at Microsoft Research and Yucheng Low at Apple.

During my time in Madison, when I was not being a grad student, you will find me racing keelboats on Lake Mendota or alpine skiing in the winters. I also doubled up as a sailing instructor at the UW-Madison’s Hoofers Sailing club. Since moving to Austin, I have been racing keelboats on Lake Travis.

Teaching

CS 395T, Principles of Learned Systems

Service

Reviewer: ICML ‘23, ICLR ‘23, Neurips ‘22, Neurips ‘21
ERC: MLSys ‘22, Usenix ATC ‘23

Publications

• CHAI: Clustered Head Attention for Efficient LLM Inference.
  S Agarwal, B Acun, B Hosmer, M Elhoushi, Y Lee, S Venkataraman, D Papailiopoulos, C Wu. ICML’ 24
  [paper]

• LayerSkip: Enabling Early Exit Inference and Self-Speculative Decoding
  M Elhoushi, A Shrivastava, D Liskovich, B Hosmer, B Wasti, L Lai, A Mahmoud, B Acun, S Agarwal, A Roman, A Aly, B Chen, C Wu. ACL’ 24
  [paper]

• Decoding Speculative Decoding.
  M Yan, S Agarwal, S Venkataraman.
  [paper]

• Blox: A Modular Toolkit for Deep Learning Schedulers.
  S Agarwal, A Phanishayee, S Venkataraman. Eurosys’24.
  [paper] [source]

• Bagpipe: Accelerating deep recommendation model training.
  S Agarwal, C Yan, Z Zhang, S Venkataraman. SOSP’23.
  [paper][source]

• Cuttlefish: Low-rank Model Training without All The Tuning.
  H Wang, S Agarwal, Y Tanaka, E Xing, D Papailiopoulos. MLSys’23.
  [paper]

• Pufferfish: Communication-efficient models at no extra cost.
  H Wang, S Agarwal, D Papailiopoulos. MLSys’22.
  [paper]

• On the utility of Gradient compression
  S Agarwal, H Wang, S Venkataraman, D Papailiopoulos. MLSys’22.
  [paper][source]

• Adaptive Gradient Communication via Critical Learning Regime Identification.
  S Agarwal, H Wang, K Lee, S Venkataraman, D Papailiopoulos. MLSys’21.
  [paper][source]

• AutoFreeze: Automatically Freezing Model Blocks to Accelerate Fine-tuning.
  Y Liu, S Agarwal, S Venkataraman.
  [paper]

• Attack of the tails: Yes, you really can backdoor federated learning.
  H Wang, K Sreenivasan, S Rajput, H Vishwakarma, S Agarwal, J Sohn, K Lee, D Papailiopoulos. Neurips’21
  [paper]