$$ \newcommand{\bone}{\mathbf{1}} \newcommand{\bbeta}{\mathbf{\beta}} \newcommand{\bdelta}{\mathbf{\delta}} \newcommand{\bepsilon}{\mathbf{\epsilon}} \newcommand{\blambda}{\mathbf{\lambda}} \newcommand{\bomega}{\mathbf{\omega}} \newcommand{\bpi}{\mathbf{\pi}} \newcommand{\bphi}{\mathbf{\phi}} \newcommand{\bvphi}{\mathbf{\varphi}} \newcommand{\bpsi}{\mathbf{\psi}} \newcommand{\bsigma}{\mathbf{\sigma}} \newcommand{\btheta}{\mathbf{\theta}} \newcommand{\btau}{\mathbf{\tau}} \newcommand{\ba}{\mathbf{a}} \newcommand{\bb}{\mathbf{b}} \newcommand{\bc}{\mathbf{c}} \newcommand{\bd}{\mathbf{d}} \newcommand{\be}{\mathbf{e}} \newcommand{\boldf}{\mathbf{f}} \newcommand{\bg}{\mathbf{g}} \newcommand{\bh}{\mathbf{h}} \newcommand{\bi}{\mathbf{i}} \newcommand{\bj}{\mathbf{j}} \newcommand{\bk}{\mathbf{k}} \newcommand{\bell}{\mathbf{\ell}} \newcommand{\bm}{\mathbf{m}} \newcommand{\bn}{\mathbf{n}} \newcommand{\bo}{\mathbf{o}} \newcommand{\bp}{\mathbf{p}} \newcommand{\bq}{\mathbf{q}} \newcommand{\br}{\mathbf{r}} \newcommand{\bs}{\mathbf{s}} \newcommand{\bt}{\mathbf{t}} \newcommand{\bu}{\mathbf{u}} \newcommand{\bv}{\mathbf{v}} \newcommand{\bw}{\mathbf{w}} \newcommand{\bx}{\mathbf{x}} \newcommand{\by}{\mathbf{y}} \newcommand{\bz}{\mathbf{z}} \newcommand{\bA}{\mathbf{A}} \newcommand{\bB}{\mathbf{B}} \newcommand{\bC}{\mathbf{C}} \newcommand{\bD}{\mathbf{D}} \newcommand{\bE}{\mathbf{E}} \newcommand{\bF}{\mathbf{F}} \newcommand{\bG}{\mathbf{G}} \newcommand{\bH}{\mathbf{H}} \newcommand{\bI}{\mathbf{I}} \newcommand{\bJ}{\mathbf{J}} \newcommand{\bK}{\mathbf{K}} \newcommand{\bL}{\mathbf{L}} \newcommand{\bM}{\mathbf{M}} \newcommand{\bN}{\mathbf{N}} \newcommand{\bP}{\mathbf{P}} \newcommand{\bQ}{\mathbf{Q}} \newcommand{\bR}{\mathbf{R}} \newcommand{\bS}{\mathbf{S}} \newcommand{\bT}{\mathbf{T}} \newcommand{\bU}{\mathbf{U}} \newcommand{\bV}{\mathbf{V}} \newcommand{\bW}{\mathbf{W}} \newcommand{\bX}{\mathbf{X}} \newcommand{\bY}{\mathbf{Y}} \newcommand{\bZ}{\mathbf{Z}} \newcommand{\bsa}{\boldsymbol{a}} \newcommand{\bsb}{\boldsymbol{b}} \newcommand{\bsc}{\boldsymbol{c}} \newcommand{\bsd}{\boldsymbol{d}} \newcommand{\bse}{\boldsymbol{e}} \newcommand{\bsoldf}{\boldsymbol{f}} \newcommand{\bsg}{\boldsymbol{g}} \newcommand{\bsh}{\boldsymbol{h}} \newcommand{\bsi}{\boldsymbol{i}} \newcommand{\bsj}{\boldsymbol{j}} \newcommand{\bsk}{\boldsymbol{k}} \newcommand{\bsell}{\boldsymbol{\ell}} \newcommand{\bsm}{\boldsymbol{m}} \newcommand{\bsn}{\boldsymbol{n}} \newcommand{\bso}{\boldsymbol{o}} \newcommand{\bsp}{\boldsymbol{p}} \newcommand{\bsq}{\boldsymbol{q}} \newcommand{\bsr}{\boldsymbol{r}} \newcommand{\bss}{\boldsymbol{s}} \newcommand{\bst}{\boldsymbol{t}} \newcommand{\bsu}{\boldsymbol{u}} \newcommand{\bsv}{\boldsymbol{v}} \newcommand{\bsw}{\boldsymbol{w}} \newcommand{\bsx}{\boldsymbol{x}} \newcommand{\bsy}{\boldsymbol{y}} \newcommand{\bsz}{\boldsymbol{z}} \newcommand{\bsA}{\boldsymbol{A}} \newcommand{\bsB}{\boldsymbol{B}} \newcommand{\bsC}{\boldsymbol{C}} \newcommand{\bsD}{\boldsymbol{D}} \newcommand{\bsE}{\boldsymbol{E}} \newcommand{\bsF}{\boldsymbol{F}} \newcommand{\bsG}{\boldsymbol{G}} \newcommand{\bsH}{\boldsymbol{H}} \newcommand{\bsI}{\boldsymbol{I}} \newcommand{\bsJ}{\boldsymbol{J}} \newcommand{\bsK}{\boldsymbol{K}} \newcommand{\bsL}{\boldsymbol{L}} \newcommand{\bsM}{\boldsymbol{M}} \newcommand{\bsN}{\boldsymbol{N}} \newcommand{\bsP}{\boldsymbol{P}} \newcommand{\bsQ}{\boldsymbol{Q}} \newcommand{\bsR}{\boldsymbol{R}} \newcommand{\bsS}{\boldsymbol{S}} \newcommand{\bsT}{\boldsymbol{T}} \newcommand{\bsU}{\boldsymbol{U}} \newcommand{\bsV}{\boldsymbol{V}} \newcommand{\bsW}{\boldsymbol{W}} \newcommand{\bsX}{\boldsymbol{X}} \newcommand{\bsY}{\boldsymbol{Y}} \newcommand{\bsZ}{\boldsymbol{Z}} \newcommand{\calA}{\mathcal{A}} \newcommand{\calB}{\mathcal{B}} \newcommand{\calC}{\mathcal{C}} \newcommand{\calD}{\mathcal{D}} \newcommand{\calE}{\mathcal{E}} \newcommand{\calF}{\mathcal{F}} \newcommand{\calG}{\mathcal{G}} \newcommand{\calH}{\mathcal{H}} \newcommand{\calI}{\mathcal{I}} \newcommand{\calJ}{\mathcal{J}} \newcommand{\calK}{\mathcal{K}} \newcommand{\calL}{\mathcal{L}} \newcommand{\calM}{\mathcal{M}} \newcommand{\calN}{\mathcal{N}} \newcommand{\calO}{\mathcal{O}} \newcommand{\calP}{\mathcal{P}} \newcommand{\calQ}{\mathcal{Q}} \newcommand{\calR}{\mathcal{R}} \newcommand{\calS}{\mathcal{S}} \newcommand{\calT}{\mathcal{T}} \newcommand{\calU}{\mathcal{U}} \newcommand{\calV}{\mathcal{V}} \newcommand{\calW}{\mathcal{W}} \newcommand{\calX}{\mathcal{X}} \newcommand{\calY}{\mathcal{Y}} \newcommand{\calZ}{\mathcal{Z}} \newcommand{\R}{\mathbb{R}} \newcommand{\C}{\mathbb{C}} \newcommand{\N}{\mathbb{N}} \newcommand{\Z}{\mathbb{Z}} \newcommand{\F}{\mathbb{F}} \newcommand{\Q}{\mathbb{Q}} \DeclareMathOperator*{\argmax}{arg\,max} \DeclareMathOperator*{\argmin}{arg\,min} \newcommand{\nnz}[1]{\mbox{nnz}(#1)} \newcommand{\dotprod}[2]{\langle #1, #2 \rangle} \newcommand{\ignore}[1]{} \let\Pr\relax \DeclareMathOperator*{\Pr}{\mathbf{Pr}} \newcommand{\E}{\mathbb{E}} \DeclareMathOperator*{\Ex}{\mathbf{E}} \DeclareMathOperator*{\Var}{\mathbf{Var}} \DeclareMathOperator*{\Cov}{\mathbf{Cov}} \DeclareMathOperator*{\stddev}{\mathbf{stddev}} \DeclareMathOperator*{\avg}{avg} \DeclareMathOperator{\poly}{poly} \DeclareMathOperator{\polylog}{polylog} \DeclareMathOperator{\size}{size} \DeclareMathOperator{\sgn}{sgn} \DeclareMathOperator{\dist}{dist} \DeclareMathOperator{\vol}{vol} \DeclareMathOperator{\spn}{span} \DeclareMathOperator{\supp}{supp} \DeclareMathOperator{\tr}{tr} \DeclareMathOperator{\Tr}{Tr} \DeclareMathOperator{\codim}{codim} \DeclareMathOperator{\diag}{diag} \newcommand{\PTIME}{\mathsf{P}} \newcommand{\LOGSPACE}{\mathsf{L}} \newcommand{\ZPP}{\mathsf{ZPP}} \newcommand{\RP}{\mathsf{RP}} \newcommand{\BPP}{\mathsf{BPP}} \newcommand{\P}{\mathsf{P}} \newcommand{\NP}{\mathsf{NP}} \newcommand{\TC}{\mathsf{TC}} \newcommand{\AC}{\mathsf{AC}} \newcommand{\SC}{\mathsf{SC}} \newcommand{\SZK}{\mathsf{SZK}} \newcommand{\AM}{\mathsf{AM}} \newcommand{\IP}{\mathsf{IP}} \newcommand{\PSPACE}{\mathsf{PSPACE}} \newcommand{\EXP}{\mathsf{EXP}} \newcommand{\MIP}{\mathsf{MIP}} \newcommand{\NEXP}{\mathsf{NEXP}} \newcommand{\BQP}{\mathsf{BQP}} \newcommand{\distP}{\mathsf{dist\textbf{P}}} \newcommand{\distNP}{\mathsf{dist\textbf{NP}}} \newcommand{\eps}{\epsilon} \newcommand{\lam}{\lambda} \newcommand{\dleta}{\delta} \newcommand{\simga}{\sigma} \newcommand{\vphi}{\varphi} \newcommand{\la}{\langle} \newcommand{\ra}{\rangle} \newcommand{\wt}[1]{\widetilde{#1}} \newcommand{\wh}[1]{\widehat{#1}} \newcommand{\ol}[1]{\overline{#1}} \newcommand{\ul}[1]{\underline{#1}} \newcommand{\ot}{\otimes} \newcommand{\zo}{\{0,1\}} \newcommand{\co}{:} %\newcommand{\co}{\colon} \newcommand{\bdry}{\partial} \newcommand{\grad}{\nabla} \newcommand{\transp}{^\intercal} \newcommand{\inv}{^{-1}} \newcommand{\symmdiff}{\triangle} \newcommand{\symdiff}{\symmdiff} \newcommand{\half}{\tfrac{1}{2}} \newcommand{\mathbbm}{\Bbb} \newcommand{\bbone}{\mathbbm 1} \newcommand{\Id}{\bbone} \newcommand{\SAT}{\mathsf{SAT}} \newcommand{\bcalG}{\boldsymbol{\calG}} \newcommand{\calbG}{\bcalG} \newcommand{\bcalX}{\boldsymbol{\calX}} \newcommand{\calbX}{\bcalX} \newcommand{\bcalY}{\boldsymbol{\calY}} \newcommand{\calbY}{\bcalY} \newcommand{\bcalZ}{\boldsymbol{\calZ}} \newcommand{\calbZ}{\bcalZ} $$

2023

  1. DiLoCo: Distributed Low-Communication Training of Language Models

    Arthur Douillard, Qixuan Feng, Andrei A. Rusu, and 6 more authors
    Nov 2023
    arXiv Bib PDF

    Paper Abstract

    Large language models (LLM) have become a critical component in many applications of machine learning. However, standard approaches to training LLM require a large number of tightly interconnected accelerators, with devices exchanging gradients and other intermediate states at each optimization step. While it is difficult to build and maintain a single computing cluster hosting many accelerators, it might be easier to find several computing clusters each hosting a smaller number of devices. In this work, we propose a distributed optimization algorithm, Distributed Low-Communication (DiLoCo), that enables training of language models on islands of devices that are poorly connected. The approach is a variant of federated averaging, where the number of inner steps is large, the inner optimizer is AdamW, and the outer optimizer is Nesterov momentum. On the widely used C4 dataset, we show that DiLoCo on 8 workers performs as well as fully synchronous optimization while communicating 500 times less. DiLoCo exhibits great robustness to the data distribution of each worker. It is also robust to resources becoming unavailable over time, and vice versa, it can seamlessly leverage resources that become available during training.

Three Important Things

1. Challenges with Distributed Training

Training distributed models is challenging: current methods requires all nodes to be online and networked by high-bandwidth interconnects. Attempting to shard model weights, activations, and optimizer states in a way that can still keep the GPUs busy requires careful analysis and software engineering to get right. As the size of our clusters grow, the mean time to failure for any node is also decreasing, which can interrupt training.

The paper introduces Distributed Low-Communication (DiLoCo) training that tries to address these issues. This is an algorithm that requires low communication, is resilient against device failures, and supports heterogeneous devices.

2. DiLoCo

The main idea behind DiLoCo is actually quite simple.

There’s 2 optimization stages going on, dubbed the inner and outer optimizers.

In the inner optimization, each worker takes a sample of data from its shard of data, and optimizes its parameters independently for \(H\) steps. Only after these \(H\) steps does it share its updated parameters with all other workers, hence helping to hide communication overhead. In the paper, they used AdamW for the inner optimizer.

The outer optimizer then takes the average of the changes of all the inner gradients, and uses this to update the overall parameters for the next step. They used Nesterov momentum for this as they found that it empirically gave the best convergence results.

3. Resiliency to Compute and Communication Changes

They did some interesting ablations to show that this technique is resistant to both compute and communication changes.

Compute Changes

They tried various settings of how the total number of nodes in the cluster change over time (i.e constant, increasing over time, decreasing, halving, etc). It’s interesting to note that for strategies that use the same final total compute (i.e doubling/halving, or ramping up/down) that their final perplexity scores were very similar.

In general, schedules that used more compute overall had lower final perplexity.

Communication Changes

On the network side of things, they tested how well training will proceed if each worker (out of 8 total) had a 10%, 30%, and 50% chance of failing to communicate their gradients. This will result in the isolated workers continuing to use their old gradients for the next step of inner optimization training. Surprisingly, even though the perplexity was spiky it still converged to nearly the synchronous case:

Most Glaring Deficiency

Perplexity is an ok metric to measure how training is improving over time, but may not necessarily correlate to downstream task performance.

Results would be stronger if there were downstream evals as well to ensure that there were no side effects to having such large inner optimization steps that cannot be glanced from just perplexity alone.

Conclusions for Future Work

Many traditional practices of model training (i.e we must update all weights after each backward pass!) are probably not strictly necessary for good convergence and generalization of training. Techniques like DiLoCo show that we can trade-off some of these for improvements in communication overhead and cluster resiliency with little impact on final model performance. This could pave way for more efficient and cheaper training of large models in the future.

Written May 5, 2025