---
title: 'ZeRO: Memory Optimizations Toward Training Trillion Parameter Models'
license: CC-BY-4.0
license_url: https://creativecommons.org/licenses/by/4.0/
license_note: 'paper is arXiv non-exclusive distribution license (verified on arXiv abs page 2026-11-12); this pointer note is our own words and is CC BY 4.0; we link to the paper rather than reproduce it'
source: https://iinventedlight.com/library/notes/1910-02054-zero.md
source_license: 'arXiv non-exclusive distribution license'
attribution: 'Pointer by iinventedlight (AI-assisted). Source: Samyam Rajbhandari, Jeff Rasley, Olatunji Ruwase, Yuxiong He, ZeRO: Memory Optimizations Toward Training Trillion Parameter Models (arXiv non-exclusive distribution license). This note links to the paper rather than reproducing it.'
author: 'iinventedlight (AI-assisted)'
date: 2026-11-12
status: published
approved_by: "Cole Withington, 2026-11-12 14:44 PT, Librarian chat widget t51s316: \"Yes, publish all 100 second-list arXiv notes\""
checked_on: 2026-11-12
as_of: 2026-11-12
confidence: medium
untested: 'Pointer note only. Paper not reproduced. Abs page and license checked 2026-11-12.'
category: arxiv
topic_area: 'ai-research: efficiency'
tags: arxiv, efficiency, distributed-training, optimizer
arxiv_id: 1910.02054
alias: 1910.02054-zero
type: pointer
format: pointer
sources:
  - id: S1
    title: 'ZeRO: Memory Optimizations Toward Training Trillion Parameter Models'
    publisher: arXiv
    authors: 'Samyam Rajbhandari, Jeff Rasley, Olatunji Ruwase, Yuxiong He'
    url: https://arxiv.org/abs/1910.02054
    version_read: 'v3 (13 May 2020)'
    accessed: 2026-11-12
    license: 'arXiv non-exclusive distribution license'
---
# ZeRO: Memory Optimizations Toward Training Trillion Parameter Models

## Summary
[ZeRO: Memory Optimizations Toward Training Trillion Parameter Models](https://arxiv.org/abs/1910.02054) (2019) introduced ZeRO, a set of memory optimizations that partition optimizer state, gradients, and parameters across data-parallel devices so that much larger models can be trained without a full copy of every state on every GPU.

What it introduced: a staged redundancy-elimination design for data-parallel training. Early stages shard optimizer state and gradients; later stages can also shard parameters, trading communication for a much smaller per-device memory footprint.

What it led to: DeepSpeed ZeRO and a default large-model training stack, often used together with model-parallel methods such as Megatron. See also the Megatron-LM pointer in this batch, and [2305-13245-gqa](https://iinventedlight.com/library/notes/2305-13245-gqa/) for a later inference-side memory theme.

Use it when you want the original sharded-optimizer design. Where to read it: https://arxiv.org/abs/1910.02054 (v3 (13 May 2020)).

The paper is posted on arXiv under the arXiv non-exclusive distribution license. This note links to it rather than reproducing it. The note itself is CC BY 4.0.

License: this note is CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Pointer by iinventedlight (AI-assisted). Paper license: arXiv non-exclusive distribution; this note links to it and does not reproduce it.

---

Agents: [/agents.md](/agents.md) | [/.well-known/agent-library.json](/.well-known/agent-library.json).
