wassname/pytorch-lightning
1
0
Fork 0
mirror of https://github.com/wassname/pytorch-lightning.git synced 2026-09-09 11:32:07 +08:00
Code Issues Packages Projects Releases Wiki Activity
Files
gh-pages
pytorch-lightning/Trainer/Distributed training/index.html
T

1207 lines
41 KiB
HTML
Raw Permalink Blame History

<!doctype html>
<html lang="en" class="no-js">
<head>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width,initial-scale=1">
<meta http-equiv="x-ua-compatible" content="ie=edge">
<meta name="description" content="Documentation for PyTorch LightningModule, the researcher version of keras.">
<meta name="lang:clipboard.copy" content="Copy to clipboard">
<meta name="lang:clipboard.copied" content="Copied to clipboard">
<meta name="lang:search.language" content="en">
<meta name="lang:search.pipeline.stopwords" content="True">
<meta name="lang:search.pipeline.trimmer" content="True">
<meta name="lang:search.result.none" content="No matching documents">
<meta name="lang:search.result.one" content="1 matching document">
<meta name="lang:search.result.other" content="# matching documents">
<meta name="lang:search.tokenizer" content="[\s\-]+">
<link rel="shortcut icon" href="../../assets/images/favicon.png">
<meta name="generator" content="mkdocs-1.0.4, mkdocs-material-4.4.0">
<title>Distributed training - PyTorch lightning Documentation</title>
<link rel="stylesheet" href="../../assets/stylesheets/application.0284f74d.css">
<script src="../../assets/javascripts/modernizr.74668098.js"></script>
<link href="https://fonts.gstatic.com" rel="preconnect" crossorigin>
<link rel="stylesheet" href="https://fonts.googleapis.com/css?family=Roboto:300,400,400i,700|Roboto+Mono&display=fallback">
<style>body,input{font-family:"Roboto","Helvetica Neue",Helvetica,Arial,sans-serif}code,kbd,pre{font-family:"Roboto Mono","Courier New",Courier,monospace}</style>
<link rel="stylesheet" href="../../assets/fonts/material-icons.css">
</head>
<body dir="ltr">
<svg class="md-svg">
<defs>
<svg xmlns="http://www.w3.org/2000/svg" width="416" height="448" viewBox="0 0 416 448" id="__github"><path fill="currentColor" d="M160 304q0 10-3.125 20.5t-10.75 19T128 352t-18.125-8.5-10.75-19T96 304t3.125-20.5 10.75-19T128 256t18.125 8.5 10.75 19T160 304zm160 0q0 10-3.125 20.5t-10.75 19T288 352t-18.125-8.5-10.75-19T256 304t3.125-20.5 10.75-19T288 256t18.125 8.5 10.75 19T320 304zm40 0q0-30-17.25-51T296 232q-10.25 0-48.75 5.25Q229.5 240 208 240t-39.25-2.75Q130.75 232 120 232q-29.5 0-46.75 21T56 304q0 22 8 38.375t20.25 25.75 30.5 15 35 7.375 37.25 1.75h42q20.5 0 37.25-1.75t35-7.375 30.5-15 20.25-25.75T360 304zm56-44q0 51.75-15.25 82.75-9.5 19.25-26.375 33.25t-35.25 21.5-42.5 11.875-42.875 5.5T212 416q-19.5 0-35.5-.75t-36.875-3.125-38.125-7.5-34.25-12.875T37 371.5t-21.5-28.75Q0 312 0 260q0-59.25 34-99-6.75-20.5-6.75-42.5 0-29 12.75-54.5 27 0 47.5 9.875t47.25 30.875Q171.5 96 212 96q37 0 70 8 26.25-20.5 46.75-30.25T376 64q12.75 25.5 12.75 54.5 0 21.75-6.75 42 34 40 34 99.5z"/></svg>
</defs>
</svg>
<input class="md-toggle" data-md-toggle="drawer" type="checkbox" id="__drawer" autocomplete="off">
<input class="md-toggle" data-md-toggle="search" type="checkbox" id="__search" autocomplete="off">
<label class="md-overlay" data-md-component="overlay" for="__drawer"></label>
<a href="#choosing-a-backend" tabindex="1" class="md-skip">
Skip to content
</a>
<header class="md-header" data-md-component="header">
<nav class="md-header-nav md-grid">
<div class="md-flex">
<div class="md-flex__cell md-flex__cell--shrink">
<a href="../.." title="PyTorch lightning Documentation" class="md-header-nav__button md-logo">
<i class="md-icon"></i>
</a>
</div>
<div class="md-flex__cell md-flex__cell--shrink">
<label class="md-icon md-icon--menu md-header-nav__button" for="__drawer"></label>
</div>
<div class="md-flex__cell md-flex__cell--stretch">
<div class="md-flex__ellipsis md-header-nav__title" data-md-component="title">
<span class="md-header-nav__topic">
PyTorch lightning Documentation
</span>
<span class="md-header-nav__topic">
Distributed training
</span>
</div>
</div>
<div class="md-flex__cell md-flex__cell--shrink">
<label class="md-icon md-icon--search md-header-nav__button" for="__search"></label>
<div class="md-search" data-md-component="search" role="dialog">
<label class="md-search__overlay" for="__search"></label>
<div class="md-search__inner" role="search">
<form class="md-search__form" name="search">
<input type="text" class="md-search__input" name="query" placeholder="Search" autocapitalize="off" autocorrect="off" autocomplete="off" spellcheck="false" data-md-component="query" data-md-state="active">
<label class="md-icon md-search__icon" for="__search"></label>
<button type="reset" class="md-icon md-search__icon" data-md-component="reset" tabindex="-1">
&#xE5CD;
</button>
</form>
<div class="md-search__output">
<div class="md-search__scrollwrap" data-md-scrollfix>
<div class="md-search-result" data-md-component="result">
<div class="md-search-result__meta">
Type to start searching
</div>
<ol class="md-search-result__list"></ol>
</div>
</div>
</div>
</div>
</div>
</div>
<div class="md-flex__cell md-flex__cell--shrink">
<div class="md-header-nav__source">
<a href="https://github.com/williamFalcon/pytorch-lightning/" title="Go to repository" class="md-source" data-md-source="github">
<div class="md-source__icon">
<svg viewBox="0 0 24 24" width="24" height="24">
<use xlink:href="#__github" width="24" height="24"></use>
</svg>
</div>
<div class="md-source__repository">
williamFalcon/pytorch-lightning
</div>
</a>
</div>
</div>
</div>
</nav>
</header>
<div class="md-container">
<main class="md-main">
<div class="md-main__inner md-grid" data-md-component="container">
<div class="md-sidebar md-sidebar--primary" data-md-component="navigation">
<div class="md-sidebar__scrollwrap">
<div class="md-sidebar__inner">
<nav class="md-nav md-nav--primary" data-md-level="0">
<label class="md-nav__title md-nav__title--site" for="__drawer">
<a href="../.." title="PyTorch lightning Documentation" class="md-nav__button md-logo">
<i class="md-icon"></i>
</a>
PyTorch lightning Documentation
</label>
<div class="md-nav__source">
<a href="https://github.com/williamFalcon/pytorch-lightning/" title="Go to repository" class="md-source" data-md-source="github">
<div class="md-source__icon">
<svg viewBox="0 0 24 24" width="24" height="24">
<use xlink:href="#__github" width="24" height="24"></use>
</svg>
</div>
<div class="md-source__repository">
williamFalcon/pytorch-lightning
</div>
</a>
</div>
<ul class="md-nav__list" data-md-scrollfix>
<li class="md-nav__item">
<a href="../.." title="Home" class="md-nav__link">
Home
</a>
</li>
<li class="md-nav__item md-nav__item--nested">
<input class="md-toggle md-nav__toggle" data-md-toggle="nav-2" type="checkbox" id="nav-2">
<label class="md-nav__link" for="nav-2">
LightningModule
</label>
<nav class="md-nav" data-md-component="collapsible" data-md-level="1">
<label class="md-nav__title" for="nav-2">
LightningModule
</label>
<ul class="md-nav__list" data-md-scrollfix>
<li class="md-nav__item">
<a href="../../LightningModule/RequiredTrainerInterface/" title="Lightning Module interface" class="md-nav__link">
Lightning Module interface
</a>
</li>
<li class="md-nav__item">
<a href="../../LightningModule/methods/" title="Methods" class="md-nav__link">
Methods
</a>
</li>
<li class="md-nav__item">
<a href="../../LightningModule/properties/" title="Properties" class="md-nav__link">
Properties
</a>
</li>
</ul>
</nav>
</li>
<li class="md-nav__item md-nav__item--active md-nav__item--nested">
<input class="md-toggle md-nav__toggle" data-md-toggle="nav-3" type="checkbox" id="nav-3" checked>
<label class="md-nav__link" for="nav-3">
Trainer
</label>
<nav class="md-nav" data-md-component="collapsible" data-md-level="1">
<label class="md-nav__title" for="nav-3">
Trainer
</label>
<ul class="md-nav__list" data-md-scrollfix>
<li class="md-nav__item">
<a href="../" title="Trainer" class="md-nav__link">
Trainer
</a>
</li>
<li class="md-nav__item">
<a href="../Checkpointing/" title="Checkpointing" class="md-nav__link">
Checkpointing
</a>
</li>
<li class="md-nav__item md-nav__item--active">
<input class="md-toggle md-nav__toggle" data-md-toggle="toc" type="checkbox" id="__toc">
<label class="md-nav__link md-nav__link--active" for="__toc">
Distributed training
</label>
<a href="./" title="Distributed training" class="md-nav__link md-nav__link--active">
Distributed training
</a>
<nav class="md-nav md-nav--secondary">
<label class="md-nav__title" for="__toc">Table of contents</label>
<ul class="md-nav__list" data-md-scrollfix>
<li class="md-nav__item">
<a href="#choosing-a-backend" title="Choosing a backend" class="md-nav__link">
Choosing a backend
</a>
<nav class="md-nav">
<ul class="md-nav__list">
<li class="md-nav__item">
<a href="#dataparallel-dp" title="DataParallel (dp)" class="md-nav__link">
DataParallel (dp)
</a>
</li>
<li class="md-nav__item">
<a href="#distributeddataparallel-ddp" title="DistributedDataParallel (ddp)" class="md-nav__link">
DistributedDataParallel (ddp)
</a>
</li>
<li class="md-nav__item">
<a href="#distributeddataparallel-2-ddp2" title="DistributedDataParallel-2 (ddp2)" class="md-nav__link">
DistributedDataParallel-2 (ddp2)
</a>
</li>
</ul>
</nav>
</li>
<li class="md-nav__item">
<a href="#distributed-and-16-bit-precision" title="Distributed and 16-bit precision." class="md-nav__link">
Distributed and 16-bit precision.
</a>
</li>
<li class="md-nav__item">
<a href="#cuda-flags" title="CUDA flags" class="md-nav__link">
CUDA flags
</a>
</li>
<li class="md-nav__item">
<a href="#16-bit-mixed-precision" title="16-bit mixed precision" class="md-nav__link">
16-bit mixed precision
</a>
</li>
<li class="md-nav__item">
<a href="#single-gpu" title="Single-gpu" class="md-nav__link">
Single-gpu
</a>
</li>
<li class="md-nav__item">
<a href="#multi-gpu" title="multi-gpu" class="md-nav__link">
multi-gpu
</a>
</li>
<li class="md-nav__item">
<a href="#multi-node" title="Multi-node" class="md-nav__link">
Multi-node
</a>
</li>
<li class="md-nav__item">
<a href="#auto-slurm-job-submission" title="Auto-slurm-job-submission" class="md-nav__link">
Auto-slurm-job-submission
</a>
</li>
<li class="md-nav__item">
<a href="#self-balancing-architecture" title="Self-balancing architecture" class="md-nav__link">
Self-balancing architecture
</a>
</li>
</ul>
</nav>
</li>
<li class="md-nav__item">
<a href="../Logging/" title="Logging" class="md-nav__link">
Logging
</a>
</li>
<li class="md-nav__item">
<a href="../SLURM Managed Cluster/" title="SLURM Managed Cluster" class="md-nav__link">
SLURM Managed Cluster
</a>
</li>
<li class="md-nav__item">
<a href="../Testing loop/" title="Testing loop" class="md-nav__link">
Testing loop
</a>
</li>
<li class="md-nav__item">
<a href="../Training Loop/" title="Training Loop" class="md-nav__link">
Training Loop
</a>
</li>
<li class="md-nav__item">
<a href="../Validation loop/" title="Validation loop" class="md-nav__link">
Validation loop
</a>
</li>
<li class="md-nav__item">
<a href="../debugging/" title="Debugging" class="md-nav__link">
Debugging
</a>
</li>
<li class="md-nav__item">
<a href="../hooks/" title="Hooks" class="md-nav__link">
Hooks
</a>
</li>
</ul>
</nav>
</li>
<li class="md-nav__item md-nav__item--nested">
<input class="md-toggle md-nav__toggle" data-md-toggle="nav-4" type="checkbox" id="nav-4">
<label class="md-nav__link" for="nav-4">
Examples
</label>
<nav class="md-nav" data-md-component="collapsible" data-md-level="1">
<label class="md-nav__title" for="nav-4">
Examples
</label>
<ul class="md-nav__list" data-md-scrollfix>
<li class="md-nav__item">
<a href="../../examples/Examples/" title="Examples" class="md-nav__link">
Examples
</a>
</li>
</ul>
</nav>
</li>
</ul>
</nav>
</div>
</div>
</div>
<div class="md-sidebar md-sidebar--secondary" data-md-component="toc">
<div class="md-sidebar__scrollwrap">
<div class="md-sidebar__inner">
<nav class="md-nav md-nav--secondary">
<label class="md-nav__title" for="__toc">Table of contents</label>
<ul class="md-nav__list" data-md-scrollfix>
<li class="md-nav__item">
<a href="#choosing-a-backend" title="Choosing a backend" class="md-nav__link">
Choosing a backend
</a>
<nav class="md-nav">
<ul class="md-nav__list">
<li class="md-nav__item">
<a href="#dataparallel-dp" title="DataParallel (dp)" class="md-nav__link">
DataParallel (dp)
</a>
</li>
<li class="md-nav__item">
<a href="#distributeddataparallel-ddp" title="DistributedDataParallel (ddp)" class="md-nav__link">
DistributedDataParallel (ddp)
</a>
</li>
<li class="md-nav__item">
<a href="#distributeddataparallel-2-ddp2" title="DistributedDataParallel-2 (ddp2)" class="md-nav__link">
DistributedDataParallel-2 (ddp2)
</a>
</li>
</ul>
</nav>
</li>
<li class="md-nav__item">
<a href="#distributed-and-16-bit-precision" title="Distributed and 16-bit precision." class="md-nav__link">
Distributed and 16-bit precision.
</a>
</li>
<li class="md-nav__item">
<a href="#cuda-flags" title="CUDA flags" class="md-nav__link">
CUDA flags
</a>
</li>
<li class="md-nav__item">
<a href="#16-bit-mixed-precision" title="16-bit mixed precision" class="md-nav__link">
16-bit mixed precision
</a>
</li>
<li class="md-nav__item">
<a href="#single-gpu" title="Single-gpu" class="md-nav__link">
Single-gpu
</a>
</li>
<li class="md-nav__item">
<a href="#multi-gpu" title="multi-gpu" class="md-nav__link">
multi-gpu
</a>
</li>
<li class="md-nav__item">
<a href="#multi-node" title="Multi-node" class="md-nav__link">
Multi-node
</a>
</li>
<li class="md-nav__item">
<a href="#auto-slurm-job-submission" title="Auto-slurm-job-submission" class="md-nav__link">
Auto-slurm-job-submission
</a>
</li>
<li class="md-nav__item">
<a href="#self-balancing-architecture" title="Self-balancing architecture" class="md-nav__link">
Self-balancing architecture
</a>
</li>
</ul>
</nav>
</div>
</div>
</div>
<div class="md-content">
<article class="md-content__inner md-typeset">
<a href="https://github.com/williamFalcon/pytorch-lightning/edit/master/docs/Trainer/Distributed training.md" title="Edit this page" class="md-icon md-content__icon">&#xE3C9;</a>
<h1>Distributed training</h1>
<p>Lightning makes multi-gpu training and 16 bit training trivial.</p>
<p><em>Note:</em> <br />
None of the flags below require changing anything about your lightningModel definition. </p>
<hr />
<h4 id="choosing-a-backend">Choosing a backend</h4>
<p>Lightning supports two backends. DataParallel and DistributedDataParallel. Both can be used for single-node multi-GPU training.
For multi-node training you must use DistributedDataParallel. </p>
<h5 id="dataparallel-dp">DataParallel (dp)</h5>
<p>Splits a batch across multiple GPUs on the same node. Cannot be used for multi-node training. </p>
<h5 id="distributeddataparallel-ddp">DistributedDataParallel (ddp)</h5>
<p>Trains a copy of the model on each GPU and only syncs gradients. If used with DistributedSampler, each GPU trains
on a subset of the full dataset. </p>
<h5 id="distributeddataparallel-2-ddp2">DistributedDataParallel-2 (ddp2)</h5>
<p>Works like DDP, except each node trains a single copy of the model using ALL GPUs on that node.
Very useful when dealing with negative samples, etc...</p>
<p>You can toggle between each mode by setting this flag.</p>
<table class="codehilitetable"><tr><td class="linenos"><div class="linenodiv"><pre> 1
2
3
4
5
6
7
8
9
10
11</pre></div></td><td class="code"><div class="codehilite"><pre><span></span><span class="c1"># DEFAULT (when using single GPU or no GPUs)</span>
<span class="n">trainer</span> <span class="o">=</span> <span class="n">Trainer</span><span class="p">(</span><span class="n">distributed_backend</span><span class="o">=</span><span class="bp">None</span><span class="p">)</span>
<span class="c1"># Change to DataParallel (gpus &gt; 1)</span>
<span class="n">trainer</span> <span class="o">=</span> <span class="n">Trainer</span><span class="p">(</span><span class="n">distributed_backend</span><span class="o">=</span><span class="s1">&#39;dp&#39;</span><span class="p">)</span>
<span class="c1"># change to distributed data parallel (gpus &gt; 1)</span>
<span class="n">trainer</span> <span class="o">=</span> <span class="n">Trainer</span><span class="p">(</span><span class="n">distributed_backend</span><span class="o">=</span><span class="s1">&#39;ddp&#39;</span><span class="p">)</span>
<span class="c1"># change to distributed data parallel (gpus &gt; 1)</span>
<span class="n">trainer</span> <span class="o">=</span> <span class="n">Trainer</span><span class="p">(</span><span class="n">distributed_backend</span><span class="o">=</span><span class="s1">&#39;ddp2&#39;</span><span class="p">)</span>
</pre></div>
</td></tr></table>
<p>If you request multiple nodes, the back-end will auto-switch to ddp.
We recommend you use DistributedDataparallel even for single-node multi-GPU training. It is MUCH faster than DP but <em>may</em>
have configuration issues depending on your cluster.</p>
<p>For a deeper understanding of what lightning is doing, feel free to read <a href="https://medium.com/@_willfalcon/9-tips-for-training-lightning-fast-neural-networks-in-pytorch-8e63a502f565">this guide</a>. </p>
<hr />
<h4 id="distributed-and-16-bit-precision">Distributed and 16-bit precision.</h4>
<p>Due to an issue with apex and DistributedDataParallel (PyTorch and NVIDIA issue), Lightning does
not allow 16-bit and DP training. We tried to get this to work, but it's an issue on their end. </p>
<p>Below are the possible configurations we support. </p>
<table>
<thead>
<tr>
<th>1 GPU</th>
<th>1+ GPUs</th>
<th>DP</th>
<th>DDP</th>
<th>16-bit</th>
<th>command</th>
</tr>
</thead>
<tbody>
<tr>
<td>Y</td>
<td></td>
<td></td>
<td></td>
<td></td>
<td><code>Trainer(gpus=1)</code></td>
</tr>
<tr>
<td>Y</td>
<td></td>
<td></td>
<td></td>
<td>Y</td>
<td><code>Trainer(gpus=1, use_amp=True)</code></td>
</tr>
<tr>
<td></td>
<td>Y</td>
<td>Y</td>
<td></td>
<td></td>
<td><code>Trainer(gpus=k, distributed_backend='dp')</code></td>
</tr>
<tr>
<td></td>
<td>Y</td>
<td></td>
<td>Y</td>
<td></td>
<td><code>Trainer(gpus=k, distributed_backend='ddp')</code></td>
</tr>
<tr>
<td></td>
<td>Y</td>
<td></td>
<td>Y</td>
<td>Y</td>
<td><code>Trainer(gpus=k, distributed_backend='ddp', use_amp=True)</code></td>
</tr>
</tbody>
</table>
<p>You also have the option of specifying which GPUs to use by passing a list: </p>
<table class="codehilitetable"><tr><td class="linenos"><div class="linenodiv"><pre> 1
2
3
4
5
6
7
8
9
10
11
12
13
14
15</pre></div></td><td class="code"><div class="codehilite"><pre><span></span><span class="c1"># DEFAULT (int) specifies how many GPUs to use.</span>
<span class="n">Trainer</span><span class="p">(</span><span class="n">gpus</span><span class="o">=</span><span class="n">k</span><span class="p">)</span>
<span class="c1"># Above is equivalent to </span>
<span class="n">Trainer</span><span class="p">(</span><span class="n">gpus</span><span class="o">=</span><span class="nb">list</span><span class="p">(</span><span class="nb">range</span><span class="p">(</span><span class="n">k</span><span class="p">)))</span>
<span class="c1"># You specify which GPUs (don&#39;t use if running on cluster) </span>
<span class="n">Trainer</span><span class="p">(</span><span class="n">gpus</span><span class="o">=</span><span class="p">[</span><span class="mi">0</span><span class="p">,</span> <span class="mi">1</span><span class="p">])</span>
<span class="c1"># can also be a string</span>
<span class="n">Trainer</span><span class="p">(</span><span class="n">gpus</span><span class="o">=</span><span class="s1">&#39;0, 1&#39;</span><span class="p">)</span>
<span class="c1"># can also be -1 or &#39;-1&#39;, this uses all available GPUs</span>
<span class="c1"># this is equivalent to list(range(torch.cuda.available_devices()))</span>
<span class="n">Trainer</span><span class="p">(</span><span class="n">gpus</span><span class="o">=-</span><span class="mi">1</span><span class="p">)</span>
</pre></div>
</td></tr></table>
<hr />
<h4 id="cuda-flags">CUDA flags</h4>
<p>CUDA flags make certain GPUs visible to your script.
Lightning sets these for you automatically, there's NO NEED to do this yourself.</p>
<table class="codehilitetable"><tr><td class="linenos"><div class="linenodiv"><pre>1
2
3</pre></div></td><td class="code"><div class="codehilite"><pre><span></span><span class="c1"># lightning will set according to what you give the trainer</span>
<span class="c1"># os.environ[&quot;CUDA_DEVICE_ORDER&quot;] = &quot;PCI_BUS_ID&quot;</span>
<span class="c1"># os.environ[&quot;CUDA_VISIBLE_DEVICES&quot;] = &quot;0&quot;</span>
</pre></div>
</td></tr></table>
<p>However, when using a cluster, Lightning will NOT set these flags (and you should not either).
SLURM will set these for you. </p>
<hr />
<h4 id="16-bit-mixed-precision">16-bit mixed precision</h4>
<p>16 bit precision can cut your memory footprint by half. If using volta architecture GPUs it can give a dramatic training speed-up as well. <br />
First, install apex (if install fails, look <a href="https://github.com/NVIDIA/apex">here</a>):</p>
<table class="codehilitetable"><tr><td class="linenos"><div class="linenodiv"><pre> 1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18</pre></div></td><td class="code"><div class="codehilite"><pre><span></span>$ git clone https://github.com/NVIDIA/apex
$ <span class="nb">cd</span> apex
<span class="c1"># ------------------------</span>
<span class="c1"># OPTIONAL: on your cluster you might need to load cuda 10 or 9</span>
<span class="c1"># depending on how you installed PyTorch</span>
<span class="c1"># see available modules</span>
module avail
<span class="c1"># load correct cuda before install</span>
module load cuda-10.0
<span class="c1"># ------------------------</span>
<span class="c1"># make sure you&#39;ve loaded a cuda version &gt; 4.0 and &lt; 7.0</span>
module load gcc-6.1.0
$ pip install -v --no-cache-dir --global-option<span class="o">=</span><span class="s2">&quot;--cpp_ext&quot;</span> --global-option<span class="o">=</span><span class="s2">&quot;--cuda_ext&quot;</span> ./
</pre></div>
</td></tr></table>
<p>then set this use_amp to True.</p>
<table class="codehilitetable"><tr><td class="linenos"><div class="linenodiv"><pre>1
2</pre></div></td><td class="code"><div class="codehilite"><pre><span></span><span class="c1"># DEFAULT</span>
<span class="n">trainer</span> <span class="o">=</span> <span class="n">Trainer</span><span class="p">(</span><span class="n">amp_level</span><span class="o">=</span><span class="s1">&#39;O2&#39;</span><span class="p">,</span> <span class="n">use_amp</span><span class="o">=</span><span class="bp">False</span><span class="p">)</span>
</pre></div>
</td></tr></table>
<hr />
<h4 id="single-gpu">Single-gpu</h4>
<p>Make sure you're on a GPU machine. </p>
<table class="codehilitetable"><tr><td class="linenos"><div class="linenodiv"><pre>1
2</pre></div></td><td class="code"><div class="codehilite"><pre><span></span><span class="c1"># DEFAULT</span>
<span class="n">trainer</span> <span class="o">=</span> <span class="n">Trainer</span><span class="p">(</span><span class="n">gpus</span><span class="o">=</span><span class="mi">1</span><span class="p">)</span>
</pre></div>
</td></tr></table>
<hr />
<h4 id="multi-gpu">multi-gpu</h4>
<p>Make sure you're on a GPU machine. You can set as many GPUs as you want.
In this setting, the model will run on all 8 GPUs at once using DataParallel under the hood.</p>
<table class="codehilitetable"><tr><td class="linenos"><div class="linenodiv"><pre>1
2
3
4
5</pre></div></td><td class="code"><div class="codehilite"><pre><span></span><span class="c1"># to use DataParallel</span>
<span class="n">trainer</span> <span class="o">=</span> <span class="n">Trainer</span><span class="p">(</span><span class="n">gpus</span><span class="o">=</span><span class="mi">8</span><span class="p">,</span> <span class="n">distributed_backend</span><span class="o">=</span><span class="s1">&#39;dp&#39;</span><span class="p">)</span>
<span class="c1"># RECOMMENDED use DistributedDataParallel</span>
<span class="n">trainer</span> <span class="o">=</span> <span class="n">Trainer</span><span class="p">(</span><span class="n">gpus</span><span class="o">=</span><span class="mi">8</span><span class="p">,</span> <span class="n">distributed_backend</span><span class="o">=</span><span class="s1">&#39;ddp&#39;</span><span class="p">)</span>
</pre></div>
</td></tr></table>
<hr />
<h4 id="multi-node">Multi-node</h4>
<p>Multi-node training is easily done by specifying these flags. </p>
<table class="codehilitetable"><tr><td class="linenos"><div class="linenodiv"><pre>1
2</pre></div></td><td class="code"><div class="codehilite"><pre><span></span><span class="c1"># train on 12*8 GPUs</span>
<span class="n">trainer</span> <span class="o">=</span> <span class="n">Trainer</span><span class="p">(</span><span class="n">gpus</span><span class="o">=</span><span class="mi">8</span><span class="p">,</span> <span class="n">nb_gpu_nodes</span><span class="o">=</span><span class="mi">12</span><span class="p">,</span> <span class="n">distributed_backend</span><span class="o">=</span><span class="s1">&#39;ddp&#39;</span><span class="p">)</span>
</pre></div>
</td></tr></table>
<p>You must configure your job submission script correctly for the trainer to work. Here is an example
script for the above trainer configuration. </p>
<table class="codehilitetable"><tr><td class="linenos"><div class="linenodiv"><pre> 1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33</pre></div></td><td class="code"><div class="codehilite"><pre><span></span><span class="ch">#!/bin/bash -l</span>
<span class="c1"># SLURM SUBMIT SCRIPT</span>
<span class="c1">#SBATCH --nodes=12</span>
<span class="c1">#SBATCH --gres=gpu:8</span>
<span class="c1">#SBATCH --ntasks-per-node=8</span>
<span class="c1">#SBATCH --mem=0</span>
<span class="c1">#SBATCH --time=0-02:00:00</span>
<span class="c1"># activate conda env</span>
conda activate my_env
<span class="c1"># -------------------------</span>
<span class="c1"># OPTIONAL</span>
<span class="c1"># -------------------------</span>
<span class="c1"># debugging flags (optional)</span>
<span class="c1"># export NCCL_DEBUG=INFO</span>
<span class="c1"># export PYTHONFAULTHANDLER=1</span>
<span class="c1"># PyTorch comes with prebuilt NCCL support... but if you have issues with it</span>
<span class="c1"># you might need to load the latest version from your modules</span>
<span class="c1"># module load NCCL/2.4.7-1-cuda.10.0</span>
<span class="c1"># on your cluster you might need these:</span>
<span class="c1"># set the network interface</span>
<span class="c1"># export NCCL_SOCKET_IFNAME=^docker0,lo</span>
<span class="c1"># -------------------------</span>
<span class="c1"># random port between 12k and 20k</span>
<span class="nb">export</span> <span class="nv">MASTER_PORT</span><span class="o">=</span><span class="k">$((</span><span class="m">12000</span> <span class="o">+</span> RANDOM <span class="o">%</span> <span class="m">20000</span><span class="k">))</span>
<span class="c1"># run script from above</span>
python my_main_file.py
</pre></div>
</td></tr></table>
<p><strong>NOTE:</strong> When running in DDP mode, any errors in your code will show up as an NCCL issue.
Set the <code>NCCL_DEBUG=INFO</code> flag to see the ACTUAL error.</p>
<p>Finally, make sure to add a distributed sampler to your dataset. The distributed sampler copies a
portion of your dataset onto each GPU. (World_size = gpus_per_node * nb_nodes). </p>
<table class="codehilitetable"><tr><td class="linenos"><div class="linenodiv"><pre>1
2
3
4
5
6
7
8</pre></div></td><td class="code"><div class="codehilite"><pre><span></span><span class="c1"># ie: this:</span>
<span class="n">dataset</span> <span class="o">=</span> <span class="n">myDataset</span><span class="p">()</span>
<span class="n">dataloader</span> <span class="o">=</span> <span class="n">Dataloader</span><span class="p">(</span><span class="n">dataset</span><span class="p">)</span>
<span class="c1"># becomes:</span>
<span class="n">dataset</span> <span class="o">=</span> <span class="n">myDataset</span><span class="p">()</span>
<span class="n">dist_sampler</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">utils</span><span class="o">.</span><span class="n">data</span><span class="o">.</span><span class="n">distributed</span><span class="o">.</span><span class="n">DistributedSampler</span><span class="p">(</span><span class="n">dataset</span><span class="p">)</span>
<span class="n">dataloader</span> <span class="o">=</span> <span class="n">Dataloader</span><span class="p">(</span><span class="n">dataset</span><span class="p">,</span> <span class="n">sampler</span><span class="o">=</span><span class="n">dist_sampler</span><span class="p">)</span>
</pre></div>
</td></tr></table>
<h4 id="auto-slurm-job-submission">Auto-slurm-job-submission</h4>
<p>Instead of manually building SLURM scripts, you can use the <a href="https://williamfalcon.github.io/test-tube/hpc/SlurmCluster/">SlurmCluster object</a> to
do this for you. The SlurmCluster can also run a grid search if you pass in a <a href="https://williamfalcon.github.io/test-tube/hyperparameter_optimization/HyperOptArgumentParser/">HyperOptArgumentParser</a>.</p>
<p>Here is an example where you run a grid search of 9 combinations of hyperparams.
<a href="https://github.com/williamFalcon/pytorch-lightning/tree/master/pl_examples/new_project_templates/multi_node_examples">The full examples are here</a>.</p>
<table class="codehilitetable"><tr><td class="linenos"><div class="linenodiv"><pre> 1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41</pre></div></td><td class="code"><div class="codehilite"><pre><span></span><span class="c1"># grid search 3 values of learning rate and 3 values of number of layers for your net</span>
<span class="c1"># this generates 9 experiments (lr=1e-3, layers=16), (lr=1e-3, layers=32), (lr=1e-3, layers=64), ... (lr=1e-1, layers=64)</span>
<span class="n">parser</span> <span class="o">=</span> <span class="n">HyperOptArgumentParser</span><span class="p">(</span><span class="n">strategy</span><span class="o">=</span><span class="s1">&#39;grid_search&#39;</span><span class="p">,</span> <span class="n">add_help</span><span class="o">=</span><span class="bp">False</span><span class="p">)</span>
<span class="n">parser</span><span class="o">.</span><span class="n">opt_list</span><span class="p">(</span><span class="s1">&#39;--learning_rate&#39;</span><span class="p">,</span> <span class="n">default</span><span class="o">=</span><span class="mf">0.001</span><span class="p">,</span> <span class="nb">type</span><span class="o">=</span><span class="nb">float</span><span class="p">,</span> <span class="n">options</span><span class="o">=</span><span class="p">[</span><span class="mf">1e-3</span><span class="p">,</span> <span class="mf">1e-2</span><span class="p">,</span> <span class="mf">1e-1</span><span class="p">],</span> <span class="n">tunable</span><span class="o">=</span><span class="bp">True</span><span class="p">)</span>
<span class="n">parser</span><span class="o">.</span><span class="n">opt_list</span><span class="p">(</span><span class="s1">&#39;--layers&#39;</span><span class="p">,</span> <span class="n">default</span><span class="o">=</span><span class="mi">1</span><span class="p">,</span> <span class="nb">type</span><span class="o">=</span><span class="nb">float</span><span class="p">,</span> <span class="n">options</span><span class="o">=</span><span class="p">[</span><span class="mi">16</span><span class="p">,</span> <span class="mi">32</span><span class="p">,</span> <span class="mi">64</span><span class="p">],</span> <span class="n">tunable</span><span class="o">=</span><span class="bp">True</span><span class="p">)</span>
<span class="n">hyperparams</span> <span class="o">=</span> <span class="n">parser</span><span class="o">.</span><span class="n">parse_args</span><span class="p">()</span>
<span class="c1"># Slurm cluster submits 9 jobs, each with a set of hyperparams</span>
<span class="n">cluster</span> <span class="o">=</span> <span class="n">SlurmCluster</span><span class="p">(</span>
<span class="n">hyperparam_optimizer</span><span class="o">=</span><span class="n">hyperparams</span><span class="p">,</span>
<span class="n">log_path</span><span class="o">=</span><span class="s1">&#39;/some/path/to/save&#39;</span><span class="p">,</span>
<span class="p">)</span>
<span class="c1"># OPTIONAL FLAGS WHICH MAY BE CLUSTER DEPENDENT</span>
<span class="c1"># which interface your nodes use for communication</span>
<span class="n">cluster</span><span class="o">.</span><span class="n">add_command</span><span class="p">(</span><span class="s1">&#39;export NCCL_SOCKET_IFNAME=^docker0,lo&#39;</span><span class="p">)</span>
<span class="c1"># see output of the NCCL connection process</span>
<span class="c1"># NCCL is how the nodes talk to each other</span>
<span class="n">cluster</span><span class="o">.</span><span class="n">add_command</span><span class="p">(</span><span class="s1">&#39;export NCCL_DEBUG=INFO&#39;</span><span class="p">)</span>
<span class="c1"># setting a master port here is a good idea.</span>
<span class="n">cluster</span><span class="o">.</span><span class="n">add_command</span><span class="p">(</span><span class="s1">&#39;export MASTER_PORT=</span><span class="si">%r</span><span class="s1">&#39;</span> <span class="o">%</span> <span class="n">PORT</span><span class="p">)</span>
<span class="c1"># ************** DON&#39;T FORGET THIS ***************</span>
<span class="c1"># MUST load the latest NCCL version</span>
<span class="n">cluster</span><span class="o">.</span><span class="n">load_modules</span><span class="p">([</span><span class="s1">&#39;NCCL/2.4.7-1-cuda.10.0&#39;</span><span class="p">])</span>
<span class="c1"># configure cluster</span>
<span class="n">cluster</span><span class="o">.</span><span class="n">per_experiment_nb_nodes</span> <span class="o">=</span> <span class="mi">12</span>
<span class="n">cluster</span><span class="o">.</span><span class="n">per_experiment_nb_gpus</span> <span class="o">=</span> <span class="mi">8</span>
<span class="n">cluster</span><span class="o">.</span><span class="n">add_slurm_cmd</span><span class="p">(</span><span class="n">cmd</span><span class="o">=</span><span class="s1">&#39;ntasks-per-node&#39;</span><span class="p">,</span> <span class="n">value</span><span class="o">=</span><span class="mi">8</span><span class="p">,</span> <span class="n">comment</span><span class="o">=</span><span class="s1">&#39;1 task per gpu&#39;</span><span class="p">)</span>
<span class="c1"># submit a script with 9 combinations of hyper params</span>
<span class="c1"># (lr=1e-3, layers=16), (lr=1e-3, layers=32), (lr=1e-3, layers=64), ... (lr=1e-1, layers=64)</span>
<span class="n">cluster</span><span class="o">.</span><span class="n">optimize_parallel_cluster_gpu</span><span class="p">(</span>
<span class="n">main</span><span class="p">,</span>
<span class="n">nb_trials</span><span class="o">=</span><span class="mi">9</span><span class="p">,</span> <span class="c1"># how many permutations of the grid search to run</span>
<span class="n">job_name</span><span class="o">=</span><span class="s1">&#39;name_for_squeue&#39;</span>
<span class="p">)</span>
</pre></div>
</td></tr></table>
<p>The other option is that you generate scripts on your own via a bash command or use another library...</p>
<hr />
<h4 id="self-balancing-architecture">Self-balancing architecture</h4>
<p>Here lightning distributes parts of your module across available GPUs to optimize for speed and memory. </p>
<p>COMING SOON.</p>
</article>
</div>
</div>
</main>
<footer class="md-footer">
<div class="md-footer-nav">
<nav class="md-footer-nav__inner md-grid">
<a href="../Checkpointing/" title="Checkpointing" class="md-flex md-footer-nav__link md-footer-nav__link--prev" rel="prev">
<div class="md-flex__cell md-flex__cell--shrink">
<i class="md-icon md-icon--arrow-back md-footer-nav__button"></i>
</div>
<div class="md-flex__cell md-flex__cell--stretch md-footer-nav__title">
<span class="md-flex__ellipsis">
<span class="md-footer-nav__direction">
Previous
</span>
Checkpointing
</span>
</div>
</a>
<a href="../Logging/" title="Logging" class="md-flex md-footer-nav__link md-footer-nav__link--next" rel="next">
<div class="md-flex__cell md-flex__cell--stretch md-footer-nav__title">
<span class="md-flex__ellipsis">
<span class="md-footer-nav__direction">
Next
</span>
Logging
</span>
</div>
<div class="md-flex__cell md-flex__cell--shrink">
<i class="md-icon md-icon--arrow-forward md-footer-nav__button"></i>
</div>
</a>
</nav>
</div>
<div class="md-footer-meta md-typeset">
<div class="md-footer-meta__inner md-grid">
<div class="md-footer-copyright">
powered by
<a href="https://www.mkdocs.org">MkDocs</a>
and
<a href="https://squidfunk.github.io/mkdocs-material/">
Material for MkDocs</a>
</div>
</div>
</div>
</footer>
</div>
<script src="../../assets/javascripts/application.245445c6.js"></script>
<script>app.initialize({version:"1.0.4",url:{base:"../.."}})</script>
</body>
</html>
Reference in New Issue View Git Blame Copy Permalink