---
_id: '14461'
abstract:
- lang: eng
  text: 'Communication-reduction techniques are a popular way to improve scalability
    in data-parallel training of deep neural networks (DNNs). The recent emergence
    of large language models such as GPT has created the need for new approaches to
    exploit data-parallelism. Among these, fully-sharded data parallel (FSDP) training
    is highly popular, yet it still encounters scalability bottlenecks. One reason
    is that applying compression techniques to FSDP is challenging: as the vast majority
    of the communication involves the model’s weights, direct compression alters convergence
    and leads to accuracy loss. We present QSDP, a variant of FSDP which supports
    both gradient and weight quantization with theoretical guarantees, is simple to
    implement and has essentially no overheads. To derive QSDP we prove that a natural
    modification of SGD achieves convergence even when we only maintain quantized
    weights, and thus the domain over which we train consists of quantized points
    and is, therefore, highly non-convex. We validate this approach by training GPT-family
    models with up to 1.3 billion parameters on a multi-node cluster. Experiments
    show that QSDP preserves model accuracy, while completely removing the communication
    bottlenecks of FSDP, providing end-to-end speedups of up to 2.2x.'
acknowledged_ssus:
- _id: ScienComp
acknowledgement: The authors gratefully acknowledge funding from the European Research
  Council (ERC) under the European Union’s Horizon 2020 research and innovation programme
  (grant agreement No 805223 ScaleML), as well as experimental support from the IST
  Austria IT department, in particular Stefano Elefante, Andrei Hornoiu, and Alois
  Schloegl. AV acknowledges the support of the French Agence Nationale de la Recherche
  (ANR), under grant ANR-21-CE48-0016 (project COMCOPT), the support of Fondation
  Hadamard with a PRMO grant, and the support of CNRS with a CoopIntEER IEA grant
  (project ALFRED).
alternative_title:
- PMLR
article_processing_charge: No
arxiv: 1
author:
- first_name: Ilia
  full_name: Markov, Ilia
  id: D0CF4148-C985-11E9-8066-0BDEE5697425
  last_name: Markov
- first_name: Adrian
  full_name: Vladu, Adrian
  last_name: Vladu
- first_name: Qi
  full_name: Guo, Qi
  last_name: Guo
- first_name: Dan-Adrian
  full_name: Alistarh, Dan-Adrian
  id: 4A899BFC-F248-11E8-B48F-1D18A9856A87
  last_name: Alistarh
  orcid: 0000-0003-3650-940X
citation:
  ama: 'Markov I, Vladu A, Guo Q, Alistarh D-A. Quantized distributed training of
    large models with convergence guarantees. In: <i>Proceedings of the 40th International
    Conference on Machine Learning</i>. Vol 202. ML Research Press; 2023:24020-24044.'
  apa: 'Markov, I., Vladu, A., Guo, Q., &#38; Alistarh, D.-A. (2023). Quantized distributed
    training of large models with convergence guarantees. In <i>Proceedings of the
    40th International Conference on Machine Learning</i> (Vol. 202, pp. 24020–24044).
    Honolulu, Hawaii, HI, United States: ML Research Press.'
  chicago: Markov, Ilia, Adrian Vladu, Qi Guo, and Dan-Adrian Alistarh. “Quantized
    Distributed Training of Large Models with Convergence Guarantees.” In <i>Proceedings
    of the 40th International Conference on Machine Learning</i>, 202:24020–44. ML
    Research Press, 2023.
  ieee: I. Markov, A. Vladu, Q. Guo, and D.-A. Alistarh, “Quantized distributed training
    of large models with convergence guarantees,” in <i>Proceedings of the 40th International
    Conference on Machine Learning</i>, Honolulu, Hawaii, HI, United States, 2023,
    vol. 202, pp. 24020–24044.
  ista: 'Markov I, Vladu A, Guo Q, Alistarh D-A. 2023. Quantized distributed training
    of large models with convergence guarantees. Proceedings of the 40th International
    Conference on Machine Learning. ICML: International Conference on Machine Learning,
    PMLR, vol. 202, 24020–24044.'
  mla: Markov, Ilia, et al. “Quantized Distributed Training of Large Models with Convergence
    Guarantees.” <i>Proceedings of the 40th International Conference on Machine Learning</i>,
    vol. 202, ML Research Press, 2023, pp. 24020–44.
  short: I. Markov, A. Vladu, Q. Guo, D.-A. Alistarh, in:, Proceedings of the 40th
    International Conference on Machine Learning, ML Research Press, 2023, pp. 24020–24044.
conference:
  end_date: 2023-07-29
  location: Honolulu, Hawaii, HI, United States
  name: 'ICML: International Conference on Machine Learning'
  start_date: 2023-07-23
corr_author: '1'
date_created: 2023-10-29T23:01:17Z
date_published: 2023-07-30T00:00:00Z
date_updated: 2026-04-07T13:00:54Z
day: '30'
department:
- _id: DaAl
ec_funded: 1
external_id:
  arxiv:
  - '2302.02390'
intvolume: '       202'
language:
- iso: eng
main_file_link:
- open_access: '1'
  url: https://doi.org/10.48550/arXiv.2302.02390
month: '07'
oa: 1
oa_version: Preprint
page: 24020-24044
project:
- _id: 268A44D6-B435-11E9-9278-68D0E5697425
  call_identifier: H2020
  grant_number: '805223'
  name: Elastic Coordination for Scalable Machine Learning
publication: Proceedings of the 40th International Conference on Machine Learning
publication_identifier:
  eissn:
  - 2640-3498
publication_status: published
publisher: ML Research Press
quality_controlled: '1'
related_material:
  record:
  - id: '17490'
    relation: dissertation_contains
    status: public
scopus_import: '1'
status: public
title: Quantized distributed training of large models with convergence guarantees
type: conference
user_id: 2DF688A6-F248-11E8-B48F-1D18A9856A87
volume: 202
year: '2023'
...