Skip to main content

Deep Algorithms and Systems Lab (DASLab)

Computer Science

Image of Dan Alistarh

Dan Alistarh

Professor


1 Career

  • Since 2022 Professor, Institute of Science and Technology Austria (ISTA); tenured since 2021
  • 2017 – 2022 Assistant Professor, Institute of Science and Technology Austria (ISTA)
  • 2023 Visiting Professor, Massachusetts Institute of Technology (MIT), Cambridge, USA
  • 2018 – 2023 Research Lead, Neural Magic Inc., Boston, USA (consulting since 2024)
  • 2016 – 2017 Visiting Professor / Senior Researcher (“Ambizione Fellow”), ETH Zurich, Switzerland
  • 2014 – 2016 Researcher, Microsoft Research, Cambridge, UK
  • 2014 – 2016 Morgan Fellow, Downing College, University of Cambridge, UK
  • 2012 – 2013 Postdoc, Massachusetts Institute of Technology, Cambridge, USA
  • 2012 PhD – “Deterministic versus Randomized Implementations of Concurrent Data Structures,” EPFL, School of Computer Science and Communication Systems, Lausanne, Switzerland

1.1 Selected Distinctions

  • 2026 NVIDIA Faculty Award for impact on GPU technology (1 of 10 recipients worldwide)
  • 2024 Lower Austria Research Prize (Wissenschaftspreis des Landes Niederösterreich) for work on model compression
  • 2024 Local Organizer, International Conference on Machine Learning (ICML 2024), Vienna
  • 2024 Key researcher, Austrian Cluster of Excellence on Bilateral Artificial Intelligence (BILAI)
  • 2023 Outstanding Paper Award, ACM Symposium on Parallelism in Algorithms and Architectures (SPAA 2023)
  • 2021 Best Paper Award, International Symposium on Distributed Computing (DISC 2021)

1.2 Selected Grants

  • 2024 – 2029 FWF Cluster of Excellence “Bilateral Artificial Intelligence” (BILAI), Co-PI (Project Co-Investigator). Awarded overall: €33,000,000 (FWF-funded portion: €19,760,512); awarded to RG: €612,318.
  • 2018 – 2023 ERC Starting Grant (No. 805223) “ScaleML: Scalable Distributed Machine Learning,” Sole PI. Awarded: €1,500,000. Two follow-up Proof-of-Concept grants awarded (2024 and 2026).
  • 2024 – 2025 NVIDIA Research Grant on post-training compression of large language models, Sole PI. Awarded: €100,000.

1.3 Tech Transfer (Patents, Spin-Offs)

  • 2018 – 2023 Research Lead, Neural Magic Inc. (Boston, USA) — spin-off from MIT building inference software for compressed deep learning models. Acquired by Red Hat in 2024 (consulting role since). Major contributor to the open-source DeepSparse engine and to vLLM, the most widely used open-source runtime for large language models.
  • GPTQ post-training quantization technology (developed at ISTA) is integrated in production software stacks of NVIDIA, HuggingFace, IBM, and Red Hat. Models compressed using GPTQ have been downloaded more than 30 million times from the HuggingFace Hub.
  • MARLIN mixed-precision GPU inference kernels (developed at ISTA) have been adopted as a default inference path in the two most popular open-source LLM serving systems, vLLM and SGLang.

1.4 Selected Publications

  1. E. Frantar, S. Ashkboos, T. Hoefler, and D. Alistarh, “GPTQ: Accurate post-training quantization for generative pre-trained transformers,” in International Conference on Learning Representations (ICLR), 2023. doi: 10.48550/arXiv.2210.17323. [Online]. Available: https://openreview.net/forum?id=tcbBPnfwxS.
  2. E. Frantar and D. Alistarh, “SparseGPT: Massive language models can be accurately pruned in one-shot,” in Proceedings of the 40th International Conference on Machine Learning (ICML), vol. 202, pp. 10323–10337, 2023. [Online]. Available: https://proceedings.mlr.press/v202/frantar23a.html.
  3. V. Egiazarian, A. Panferov, D. Kuznedelev, E. Frantar, A. Babenko, and D. Alistarh, “Extreme compression of large language models via additive quantization,” in Proceedings of the 41st International Conference on Machine Learning (ICML), vol. 235, pp. 12284–12303, 2024. [Online]. Available: https://proceedings.mlr.press/v235/egiazarian24a.html.
  4. E. Frantar, C. Riquelme, N. Houlsby, D. Alistarh, and U. Evci, “Scaling laws for sparsely-connected foundation models,” in International Conference on Learning Representations (ICLR), 2024. [Online]. Available: https://openreview.net/forum?id=i9K2ZWkYIP.
  5. E. Frantar, R. L. Castro, J. Chen, T. Hoefler, and D. Alistarh, “MARLIN: Mixed-precision auto-regressive parallel inference on large language models,” in Proceedings of the 30th ACM SIGPLAN Symposium on Principles and Practice of Parallel Programming (PPoPP), pp. 239–251, 2025. doi: 10.1145/3710848.3710871.

2 Group Members

  • Jiale Chen, PhD Student since 2024
  • Ionut-Vlad Modoranu, PhD Student since 2022
  • Mahdi Nikdan, PhD Student since 2022
  • Andrei Panferov, PhD Student since 2025
  • Sergei Pankratov, PhD Student since 2025
  • Soroush Tabesh, PhD Student since 2025
  • Alexandra Volkova, PhD Student since 2025
  • Vage Egiazarian, Postdoc since 2025
  • Erik Schultheis, Postdoc since 2025
  • Eldar Kurtic, Research Technician (Machine Learning) since 2022
  • Matin Ansaripour, Scientific Intern
  • Maximilian Kleinegger, Scientific Intern

3 Research

Artificial Intelligence has made remarkable progress over the past decade, with breakthroughs across applications from computer vision to large language models (LLMs). Yet the sustainability of this pace is in question: the compute required to train and deploy state-of-the-art models has been doubling every few months, while single-GPU compute has grown less than tenfold over the same period. This widening gap risks centralizing AI progress, leaving smaller institutions—and entire countries—unable to fully participate.

The Alistarh group, also known as the Deep Algorithms and Systems Lab (DASLab), works to remove these barriers by developing training and inference algorithms that are dramatically more efficient than conventional ones. We design new algorithms for learning over compressed (e.g., sparse, quantized, or vector-coded) representations, and pair them with high-performance system implementations that turn compression gains into real-world speed and energy savings. Our work on post-training compression of large language models, pioneered through methods such as GPTQ and SparseGPT, was the first to enable fast, accurate inference for GPT-scale models, and is now deployed in widely used software from NVIDIA, HuggingFace, and Red Hat (formerly Neural Magic), with compressed open-source models downloaded over 30 million times. Looking ahead, we aim to extend these ideas to enable accurate training directly over highly compact representations.


3.1 Current Scientific Projects/Directions

  • Efficient inference for large language models via post-training compression (quantization, sparsity, vector and additive quantization)
  • Compact learning: training neural networks directly over compressed (quantized, sparse, structured-matrix) representations
  • Memory- and communication-efficient distributed training of foundation models
  • High-performance GPU kernels for low-precision and sparse deep learning
  • Adaptive concurrent data structures and the foundations of distributed computation

3.2 Research in Focus

GPTQ: Accurate post-training quantization for large language models

The advent of large language models (LLMs) has created a fundamental tension. State-of-the-art models with hundreds of billions of parameters offer remarkable capabilities, but their computational and memory demands place them out of reach for most institutions and individuals. Compressing such models after training, without retraining, would democratize their use—but doing so accurately at GPT scale was widely believed to be infeasible.

In this work, we developed GPTQ, the first method to perform accurate post-training quantization on language models with billions of parameters. Building on classical second-order information from the loss surface, GPTQ quantizes pretrained weights one column at a time while compensating for the error introduced in the remaining weights. The result was the first demonstration that LLMs as large as 175 billion parameters could be compressed to 4 bits per weight—and even 3 bits—with negligible accuracy loss, in a matter of GPU-hours rather than weeks of retraining.

The 4-bit GPTQ format has become a standard for LLM deployment: it is implemented in production software from NVIDIA, HuggingFace, IBM, and Red Hat (formerly Neural Magic), and underpins the popular open-source vLLM inference engine. More than 30 million GPTQ-compressed models have been downloaded from the HuggingFace Hub, enabling researchers, students, and small companies to run state-of-the-art LLMs on commodity hardware. The paper has accumulated more than 2,000 citations.



theme sidebar-arrow-up
Back to Top