Hugging Face's logo

Spaces:
siro1
/
triton-for-llms
Running

App Files Community
triton-for-llms / src /_site /notebooks /single_gpu.html
siro1's picture
siro1 HF Staff
commit from `quarto publish`
cf7a8a2
raw
history blame contribute delete
110 kB
<!DOCTYPE html>
<html xmlns="http://www.w3.org/1999/xhtml" lang="en" xml:lang="en"><head>
<meta charset="utf-8">
<meta name="generator" content="quarto-1.6.40">
<meta name="viewport" content="width=device-width, initial-scale=1.0, user-scalable=yes">
<title>Single GPU Fine-tuning – Open-Source AI Cookbook</title>
<style>
code{white-space: pre-wrap;}
span.smallcaps{font-variant: small-caps;}
div.columns{display: flex; gap: min(4vw, 1.5em);}
div.column{flex: auto; overflow-x: auto;}
div.hanging-indent{margin-left: 1.5em; text-indent: -1.5em;}
ul.task-list{list-style: none;}
ul.task-list li input[type="checkbox"] {
width: 0.8em;
margin: 0 0.8em 0.2em -1em; /* quarto-specific, see https://github.com/quarto-dev/quarto-cli/issues/4556 */
vertical-align: middle;
}
/* CSS for syntax highlighting */
pre > code.sourceCode { white-space: pre; position: relative; }
pre > code.sourceCode > span { line-height: 1.25; }
pre > code.sourceCode > span:empty { height: 1.2em; }
.sourceCode { overflow: visible; }
code.sourceCode > span { color: inherit; text-decoration: inherit; }
div.sourceCode { margin: 1em 0; }
pre.sourceCode { margin: 0; }
@media screen {
div.sourceCode { overflow: auto; }
}
@media print {
pre > code.sourceCode { white-space: pre-wrap; }
pre > code.sourceCode > span { display: inline-block; text-indent: -5em; padding-left: 5em; }
}
pre.numberSource code
{ counter-reset: source-line 0; }
pre.numberSource code > span
{ position: relative; left: -4em; counter-increment: source-line; }
pre.numberSource code > span > a:first-child::before
{ content: counter(source-line);
position: relative; left: -1em; text-align: right; vertical-align: baseline;
border: none; display: inline-block;
-webkit-touch-callout: none; -webkit-user-select: none;
-khtml-user-select: none; -moz-user-select: none;
-ms-user-select: none; user-select: none;
padding: 0 4px; width: 4em;
}
pre.numberSource { margin-left: 3em; padding-left: 4px; }
div.sourceCode
{ }
@media screen {
pre > code.sourceCode > span > a:first-child::before { text-decoration: underline; }
}
</style>
<script src="https://cdnjs.cloudflare.com/ajax/libs/jquery/3.5.1/jquery.min.js" integrity="sha512-bLT0Qm9VnAYZDflyKcBaQ2gg0hSYNQrJ8RilYldYQ1FxQYoCLtUjuuRuZo+fjqhx/qtq/1itJ0C2ejDxltZVFg==" crossorigin="anonymous"></script><script src="../site_libs/quarto-nav/quarto-nav.js"></script>
<script src="../site_libs/quarto-nav/headroom.min.js"></script>
<script src="../site_libs/clipboard/clipboard.min.js"></script>
<script src="../site_libs/quarto-search/autocomplete.umd.js"></script>
<script src="../site_libs/quarto-search/fuse.min.js"></script>
<script src="../site_libs/quarto-search/quarto-search.js"></script>
<meta name="quarto:offset" content="../">
<script src="../site_libs/quarto-html/quarto.js"></script>
<script src="../site_libs/quarto-html/popper.min.js"></script>
<script src="../site_libs/quarto-html/tippy.umd.min.js"></script>
<script src="../site_libs/quarto-html/anchor.min.js"></script>
<link href="../site_libs/quarto-html/tippy.css" rel="stylesheet">
<link href="../site_libs/quarto-html/quarto-syntax-highlighting-549806ee2085284f45b00abea8c6df48.css" rel="stylesheet" id="quarto-text-highlighting-styles">
<script src="../site_libs/bootstrap/bootstrap.min.js"></script>
<link href="../site_libs/bootstrap/bootstrap-icons.css" rel="stylesheet">
<link href="../site_libs/bootstrap/bootstrap-2be10d9e998f81ff6e49e26833438aa5.min.css" rel="stylesheet" append-hash="true" id="quarto-bootstrap" data-mode="light">
<script id="quarto-search-options" type="application/json">{
"location": "sidebar",
"copy-button": false,
"collapse-after": 3,
"panel-placement": "start",
"type": "textbox",
"limit": 50,
"keyboard-shortcut": [
"f",
"/",
"s"
],
"show-item-context": false,
"language": {
"search-no-results-text": "No results",
"search-matching-documents-text": "matching documents",
"search-copy-link-title": "Copy link to search",
"search-hide-matches-text": "Hide additional matches",
"search-more-match-text": "more match in this document",
"search-more-matches-text": "more matches in this document",
"search-clear-button-title": "Clear",
"search-text-placeholder": "",
"search-detached-cancel-button-title": "Cancel",
"search-submit-button-title": "Submit",
"search-label": "Search"
}
}</script>
<script src="https://cdnjs.cloudflare.com/ajax/libs/require.js/2.3.6/require.min.js" integrity="sha512-c3Nl8+7g4LMSTdrm621y7kf9v3SDPnhxLNhcjFJbKECVnmZHTdo+IRO05sNLTH/D3vA6u1X32ehoLC7WFVdheg==" crossorigin="anonymous"></script>
<script type="application/javascript">define('jquery', [],function() {return window.jQuery;})</script>
<link rel="stylesheet" href="../styles.css">
</head>
<body class="nav-sidebar docked">
<div id="quarto-search-results"></div>
<header id="quarto-header" class="headroom fixed-top">
<nav class="quarto-secondary-nav">
<div class="container-fluid d-flex">
<button type="button" class="quarto-btn-toggle btn" data-bs-toggle="collapse" role="button" data-bs-target=".quarto-sidebar-collapse-item" aria-controls="quarto-sidebar" aria-expanded="false" aria-label="Toggle sidebar navigation" onclick="if (window.quartoToggleHeadroom) { window.quartoToggleHeadroom(); }">
<i class="bi bi-layout-text-sidebar-reverse"></i>
</button>
<nav class="quarto-page-breadcrumbs" aria-label="breadcrumb"><ol class="breadcrumb"><li class="breadcrumb-item">Open-Source AI Cookbook</li><li class="breadcrumb-item"><a href="../notebooks/automatic_embedding.html">Additional Techniques</a></li><li class="breadcrumb-item"><a href="../notebooks/single_gpu.html">Single GPU Optimization</a></li></ol></nav>
<a class="flex-grow-1" role="navigation" data-bs-toggle="collapse" data-bs-target=".quarto-sidebar-collapse-item" aria-controls="quarto-sidebar" aria-expanded="false" aria-label="Toggle sidebar navigation" onclick="if (window.quartoToggleHeadroom) { window.quartoToggleHeadroom(); }">
</a>
<button type="button" class="btn quarto-search-button" aria-label="Search" onclick="window.quartoOpenSearch();">
<i class="bi bi-search"></i>
</button>
</div>
</nav>
</header>
<!-- content -->
<div id="quarto-content" class="quarto-container page-columns page-rows-contents page-layout-article">
<!-- sidebar -->
<nav id="quarto-sidebar" class="sidebar collapse collapse-horizontal quarto-sidebar-collapse-item sidebar-navigation docked overflow-auto">
<div class="pt-lg-2 mt-2 text-left sidebar-header">
<div class="sidebar-title mb-0 py-0">
<a href="../">Open-Source AI Cookbook</a>
</div>
</div>
<div class="mt-2 flex-shrink-0 align-items-center">
<div class="sidebar-search">
<div id="quarto-search" class="" title="Search"></div>
</div>
</div>
<div class="sidebar-menu-container">
<ul class="list-unstyled mt-1">
<li class="sidebar-item sidebar-item-section">
<div class="sidebar-item-container">
<a class="sidebar-item-text sidebar-link text-start" data-bs-toggle="collapse" data-bs-target="#quarto-sidebar-section-1" role="navigation" aria-expanded="true">
<span class="menu-text">About</span></a>
<a class="sidebar-item-toggle text-start" data-bs-toggle="collapse" data-bs-target="#quarto-sidebar-section-1" role="navigation" aria-expanded="true" aria-label="Toggle section">
<i class="bi bi-chevron-right ms-2"></i>
</a>
</div>
<ul id="quarto-sidebar-section-1" class="collapse list-unstyled sidebar-section depth1 show">
<li class="sidebar-item">
<div class="sidebar-item-container">
<a href="../index.html" class="sidebar-item-text sidebar-link">
<span class="menu-text">About Quarto</span></a>
</div>
</li>
</ul>
</li>
<li class="sidebar-item sidebar-item-section">
<div class="sidebar-item-container">
<a class="sidebar-item-text sidebar-link text-start" data-bs-toggle="collapse" data-bs-target="#quarto-sidebar-section-2" role="navigation" aria-expanded="true">
<span class="menu-text">Open-Source AI Cookbook</span></a>
<a class="sidebar-item-toggle text-start" data-bs-toggle="collapse" data-bs-target="#quarto-sidebar-section-2" role="navigation" aria-expanded="true" aria-label="Toggle section">
<i class="bi bi-chevron-right ms-2"></i>
</a>
</div>
<ul id="quarto-sidebar-section-2" class="collapse list-unstyled sidebar-section depth1 show">
<li class="sidebar-item sidebar-item-section">
<div class="sidebar-item-container">
<a class="sidebar-item-text sidebar-link text-start" data-bs-toggle="collapse" data-bs-target="#quarto-sidebar-section-3" role="navigation" aria-expanded="true">
<span class="menu-text">RAG Techniques</span></a>
<a class="sidebar-item-toggle text-start" data-bs-toggle="collapse" data-bs-target="#quarto-sidebar-section-3" role="navigation" aria-expanded="true" aria-label="Toggle section">
<i class="bi bi-chevron-right ms-2"></i>
</a>
</div>
<ul id="quarto-sidebar-section-3" class="collapse list-unstyled sidebar-section depth2 show">
<li class="sidebar-item">
<div class="sidebar-item-container">
<a href="../notebooks/rag_zephyr_langchain.html" class="sidebar-item-text sidebar-link">
<span class="menu-text">RAG Zephyr &amp; LangChain</span></a>
</div>
</li>
<li class="sidebar-item">
<div class="sidebar-item-container">
<a href="../notebooks/advanced_rag.html" class="sidebar-item-text sidebar-link">
<span class="menu-text">Advanced RAG</span></a>
</div>
</li>
<li class="sidebar-item">
<div class="sidebar-item-container">
<a href="../notebooks/rag_evaluation.html" class="sidebar-item-text sidebar-link">
<span class="menu-text">RAG Evaluation</span></a>
</div>
</li>
</ul>
</li>
<li class="sidebar-item sidebar-item-section">
<div class="sidebar-item-container">
<a class="sidebar-item-text sidebar-link text-start" data-bs-toggle="collapse" data-bs-target="#quarto-sidebar-section-4" role="navigation" aria-expanded="true">
<span class="menu-text">Additional Techniques</span></a>
<a class="sidebar-item-toggle text-start" data-bs-toggle="collapse" data-bs-target="#quarto-sidebar-section-4" role="navigation" aria-expanded="true" aria-label="Toggle section">
<i class="bi bi-chevron-right ms-2"></i>
</a>
</div>
<ul id="quarto-sidebar-section-4" class="collapse list-unstyled sidebar-section depth2 show">
<li class="sidebar-item">
<div class="sidebar-item-container">
<a href="../notebooks/automatic_embedding.html" class="sidebar-item-text sidebar-link">
<span class="menu-text">Automatic Embedding</span></a>
</div>
</li>
<li class="sidebar-item">
<div class="sidebar-item-container">
<a href="../notebooks/faiss.html" class="sidebar-item-text sidebar-link">
<span class="menu-text">FAISS for Efficient Search</span></a>
</div>
</li>
<li class="sidebar-item">
<div class="sidebar-item-container">
<a href="../notebooks/single_gpu.html" class="sidebar-item-text sidebar-link active">
<span class="menu-text">Single GPU Optimization</span></a>
</div>
</li>
</ul>
</li>
</ul>
</li>
</ul>
</div>
</nav>
<div id="quarto-sidebar-glass" class="quarto-sidebar-collapse-item" data-bs-toggle="collapse" data-bs-target=".quarto-sidebar-collapse-item"></div>
<!-- margin-sidebar -->
<div id="quarto-margin-sidebar" class="sidebar margin-sidebar">
<nav id="TOC" role="doc-toc" class="toc-active">
<h2 id="toc-title">On this page</h2>
<ul>
<li><a href="#fine-tuning-a-code-llm-on-custom-code-on-a-single-gpu" id="toc-fine-tuning-a-code-llm-on-custom-code-on-a-single-gpu" class="nav-link active" data-scroll-target="#fine-tuning-a-code-llm-on-custom-code-on-a-single-gpu">Fine-tuning a Code LLM on Custom Code on a single GPU</a>
<ul class="collapse">
<li><a href="#dataset" id="toc-dataset" class="nav-link" data-scroll-target="#dataset">Dataset</a></li>
<li><a href="#model" id="toc-model" class="nav-link" data-scroll-target="#model">Model</a></li>
<li><a href="#prepare-the-data" id="toc-prepare-the-data" class="nav-link" data-scroll-target="#prepare-the-data">Prepare the data</a></li>
<li><a href="#prepare-the-model" id="toc-prepare-the-model" class="nav-link" data-scroll-target="#prepare-the-model">Prepare the model</a></li>
<li><a href="#train-the-model" id="toc-train-the-model" class="nav-link" data-scroll-target="#train-the-model">Train the model</a></li>
<li><a href="#inference" id="toc-inference" class="nav-link" data-scroll-target="#inference">Inference</a></li>
</ul></li>
</ul>
</nav>
</div>
<!-- main -->
<main class="content" id="quarto-document-content">
<header id="title-block-header" class="quarto-title-block default"><nav class="quarto-page-breadcrumbs quarto-title-breadcrumbs d-none d-lg-block" aria-label="breadcrumb"><ol class="breadcrumb"><li class="breadcrumb-item">Open-Source AI Cookbook</li><li class="breadcrumb-item"><a href="../notebooks/automatic_embedding.html">Additional Techniques</a></li><li class="breadcrumb-item"><a href="../notebooks/single_gpu.html">Single GPU Optimization</a></li></ol></nav>
<div class="quarto-title">
<h1 class="title">Single GPU Fine-tuning</h1>
</div>
<div class="quarto-title-meta">
</div>
</header>
<section id="fine-tuning-a-code-llm-on-custom-code-on-a-single-gpu" class="level1">
<h1>Fine-tuning a Code LLM on Custom Code on a single GPU</h1>
<p><em>Authored by: <a href="https://github.com/MKhalusova">Maria Khalusova</a></em></p>
<p>Publicly available code LLMs such as Codex, StarCoder, and Code Llama are great at generating code that adheres to general programming principles and syntax, but they may not align with an organization’s internal conventions, or be aware of proprietary libraries.</p>
<p>In this notebook, we’ll see show how you can fine-tune a code LLM on private code bases to enhance its contextual awareness and improve a model’s usefulness to your organization’s needs. Since the code LLMs are quite large, fine-tuning them in a traditional manner can be resource-draining. Worry not! We will show how you can optimize fine-tuning to fit on a single GPU.</p>
<section id="dataset" class="level2">
<h2 class="anchored" data-anchor-id="dataset">Dataset</h2>
<p>For this example, we picked the top 10 Hugging Face public repositories on GitHub. We have excluded non-code files from the data, such as images, audio files, presentations, and so on. For Jupyter notebooks, we’ve kept only cells containing code. The resulting code is stored as a dataset that you can find on the Hugging Face Hub under <a href="https://huggingface.co/datasets/smangrul/hf-stack-v1"><code>smangrul/hf-stack-v1</code></a>. It contains repo id, file path, and file content.</p>
</section>
<section id="model" class="level2">
<h2 class="anchored" data-anchor-id="model">Model</h2>
<p>We’ll finetune <a href="https://huggingface.co/bigcode/starcoderbase-1b"><code>bigcode/starcoderbase-1b</code></a>, which is a 1B parameter model trained on 80+ programming languages. This is a gated model, so if you plan to run this notebook with this exact model, you’ll need to gain access to it on the model’s page. Log in to your Hugging Face account to do so:</p>
<div id="cell-1" class="cell">
<div class="sourceCode cell-code" id="cb1"><pre class="sourceCode python code-with-copy"><code class="sourceCode python"><span id="cb1-1"><a href="#cb1-1" aria-hidden="true" tabindex="-1"></a><span class="im">from</span> huggingface_hub <span class="im">import</span> notebook_login</span>
<span id="cb1-2"><a href="#cb1-2" aria-hidden="true" tabindex="-1"></a></span>
<span id="cb1-3"><a href="#cb1-3" aria-hidden="true" tabindex="-1"></a>notebook_login()</span></code><button title="Copy to Clipboard" class="code-copy-button"><i class="bi"></i></button></pre></div>
</div>
<p>To get started, let’s install all the necessary libraries. As you can see, in addition to <code>transformers</code> and <code>datasets</code>, we’ll be using <code>peft</code>, <code>bitsandbytes</code>, and <code>flash-attn</code> to optimize the training.</p>
<p>By employing parameter-efficient training techniques, we can run this notebook on a single A100 High-RAM GPU.</p>
<div id="cell-3" class="cell">
<div class="sourceCode cell-code" id="cb2"><pre class="sourceCode python code-with-copy"><code class="sourceCode python"><span id="cb2-1"><a href="#cb2-1" aria-hidden="true" tabindex="-1"></a><span class="op">!</span>pip install <span class="op">-</span>q transformers datasets peft bitsandbytes flash<span class="op">-</span>attn</span></code><button title="Copy to Clipboard" class="code-copy-button"><i class="bi"></i></button></pre></div>
</div>
<p>Let’s define some variables now. Feel free to play with these.</p>
<div id="cell-5" class="cell">
<div class="sourceCode cell-code" id="cb3"><pre class="sourceCode python code-with-copy"><code class="sourceCode python"><span id="cb3-1"><a href="#cb3-1" aria-hidden="true" tabindex="-1"></a>MODEL<span class="op">=</span><span class="st">"bigcode/starcoderbase-1b"</span> <span class="co"># Model checkpoint on the Hugging Face Hub</span></span>
<span id="cb3-2"><a href="#cb3-2" aria-hidden="true" tabindex="-1"></a>DATASET<span class="op">=</span><span class="st">"smangrul/hf-stack-v1"</span> <span class="co"># Dataset on the Hugging Face Hub</span></span>
<span id="cb3-3"><a href="#cb3-3" aria-hidden="true" tabindex="-1"></a>DATA_COLUMN<span class="op">=</span><span class="st">"content"</span> <span class="co"># Column name containing the code content</span></span>
<span id="cb3-4"><a href="#cb3-4" aria-hidden="true" tabindex="-1"></a></span>
<span id="cb3-5"><a href="#cb3-5" aria-hidden="true" tabindex="-1"></a>SEQ_LENGTH<span class="op">=</span><span class="dv">2048</span> <span class="co"># Sequence length</span></span>
<span id="cb3-6"><a href="#cb3-6" aria-hidden="true" tabindex="-1"></a></span>
<span id="cb3-7"><a href="#cb3-7" aria-hidden="true" tabindex="-1"></a><span class="co"># Training arguments</span></span>
<span id="cb3-8"><a href="#cb3-8" aria-hidden="true" tabindex="-1"></a>MAX_STEPS<span class="op">=</span><span class="dv">2000</span> <span class="co"># max_steps</span></span>
<span id="cb3-9"><a href="#cb3-9" aria-hidden="true" tabindex="-1"></a>BATCH_SIZE<span class="op">=</span><span class="dv">16</span> <span class="co"># batch_size</span></span>
<span id="cb3-10"><a href="#cb3-10" aria-hidden="true" tabindex="-1"></a>GR_ACC_STEPS<span class="op">=</span><span class="dv">1</span> <span class="co"># gradient_accumulation_steps</span></span>
<span id="cb3-11"><a href="#cb3-11" aria-hidden="true" tabindex="-1"></a>LR<span class="op">=</span><span class="fl">5e-4</span> <span class="co"># learning_rate</span></span>
<span id="cb3-12"><a href="#cb3-12" aria-hidden="true" tabindex="-1"></a>LR_SCHEDULER_TYPE<span class="op">=</span><span class="st">"cosine"</span> <span class="co"># lr_scheduler_type</span></span>
<span id="cb3-13"><a href="#cb3-13" aria-hidden="true" tabindex="-1"></a>WEIGHT_DECAY<span class="op">=</span><span class="fl">0.01</span> <span class="co"># weight_decay</span></span>
<span id="cb3-14"><a href="#cb3-14" aria-hidden="true" tabindex="-1"></a>NUM_WARMUP_STEPS<span class="op">=</span><span class="dv">30</span> <span class="co"># num_warmup_steps</span></span>
<span id="cb3-15"><a href="#cb3-15" aria-hidden="true" tabindex="-1"></a>EVAL_FREQ<span class="op">=</span><span class="dv">100</span> <span class="co"># eval_freq</span></span>
<span id="cb3-16"><a href="#cb3-16" aria-hidden="true" tabindex="-1"></a>SAVE_FREQ<span class="op">=</span><span class="dv">100</span> <span class="co"># save_freq</span></span>
<span id="cb3-17"><a href="#cb3-17" aria-hidden="true" tabindex="-1"></a>LOG_FREQ<span class="op">=</span><span class="dv">25</span> <span class="co"># log_freq</span></span>
<span id="cb3-18"><a href="#cb3-18" aria-hidden="true" tabindex="-1"></a>OUTPUT_DIR<span class="op">=</span><span class="st">"peft-starcoder-lora-a100"</span> <span class="co"># output_dir</span></span>
<span id="cb3-19"><a href="#cb3-19" aria-hidden="true" tabindex="-1"></a>BF16<span class="op">=</span><span class="va">True</span> <span class="co"># bf16</span></span>
<span id="cb3-20"><a href="#cb3-20" aria-hidden="true" tabindex="-1"></a>FP16<span class="op">=</span><span class="va">False</span> <span class="co"># no_fp16</span></span>
<span id="cb3-21"><a href="#cb3-21" aria-hidden="true" tabindex="-1"></a></span>
<span id="cb3-22"><a href="#cb3-22" aria-hidden="true" tabindex="-1"></a><span class="co"># FIM trasformations arguments</span></span>
<span id="cb3-23"><a href="#cb3-23" aria-hidden="true" tabindex="-1"></a>FIM_RATE<span class="op">=</span><span class="fl">0.5</span> <span class="co"># fim_rate</span></span>
<span id="cb3-24"><a href="#cb3-24" aria-hidden="true" tabindex="-1"></a>FIM_SPM_RATE<span class="op">=</span><span class="fl">0.5</span> <span class="co"># fim_spm_rate</span></span>
<span id="cb3-25"><a href="#cb3-25" aria-hidden="true" tabindex="-1"></a></span>
<span id="cb3-26"><a href="#cb3-26" aria-hidden="true" tabindex="-1"></a><span class="co"># LORA</span></span>
<span id="cb3-27"><a href="#cb3-27" aria-hidden="true" tabindex="-1"></a>LORA_R<span class="op">=</span><span class="dv">8</span> <span class="co"># lora_r</span></span>
<span id="cb3-28"><a href="#cb3-28" aria-hidden="true" tabindex="-1"></a>LORA_ALPHA<span class="op">=</span><span class="dv">32</span> <span class="co"># lora_alpha</span></span>
<span id="cb3-29"><a href="#cb3-29" aria-hidden="true" tabindex="-1"></a>LORA_DROPOUT<span class="op">=</span><span class="fl">0.0</span> <span class="co"># lora_dropout</span></span>
<span id="cb3-30"><a href="#cb3-30" aria-hidden="true" tabindex="-1"></a>LORA_TARGET_MODULES<span class="op">=</span><span class="st">"c_proj,c_attn,q_attn,c_fc,c_proj"</span> <span class="co"># lora_target_modules</span></span>
<span id="cb3-31"><a href="#cb3-31" aria-hidden="true" tabindex="-1"></a></span>
<span id="cb3-32"><a href="#cb3-32" aria-hidden="true" tabindex="-1"></a><span class="co"># bitsandbytes config</span></span>
<span id="cb3-33"><a href="#cb3-33" aria-hidden="true" tabindex="-1"></a>USE_NESTED_QUANT<span class="op">=</span><span class="va">True</span> <span class="co"># use_nested_quant</span></span>
<span id="cb3-34"><a href="#cb3-34" aria-hidden="true" tabindex="-1"></a>BNB_4BIT_COMPUTE_DTYPE<span class="op">=</span><span class="st">"bfloat16"</span><span class="co"># bnb_4bit_compute_dtype</span></span>
<span id="cb3-35"><a href="#cb3-35" aria-hidden="true" tabindex="-1"></a></span>
<span id="cb3-36"><a href="#cb3-36" aria-hidden="true" tabindex="-1"></a>SEED<span class="op">=</span><span class="dv">0</span></span></code><button title="Copy to Clipboard" class="code-copy-button"><i class="bi"></i></button></pre></div>
</div>
<div id="cell-6" class="cell">
<div class="sourceCode cell-code" id="cb4"><pre class="sourceCode python code-with-copy"><code class="sourceCode python"><span id="cb4-1"><a href="#cb4-1" aria-hidden="true" tabindex="-1"></a><span class="im">from</span> transformers <span class="im">import</span> (</span>
<span id="cb4-2"><a href="#cb4-2" aria-hidden="true" tabindex="-1"></a> AutoModelForCausalLM,</span>
<span id="cb4-3"><a href="#cb4-3" aria-hidden="true" tabindex="-1"></a> AutoTokenizer,</span>
<span id="cb4-4"><a href="#cb4-4" aria-hidden="true" tabindex="-1"></a> Trainer,</span>
<span id="cb4-5"><a href="#cb4-5" aria-hidden="true" tabindex="-1"></a> TrainingArguments,</span>
<span id="cb4-6"><a href="#cb4-6" aria-hidden="true" tabindex="-1"></a> logging,</span>
<span id="cb4-7"><a href="#cb4-7" aria-hidden="true" tabindex="-1"></a> set_seed,</span>
<span id="cb4-8"><a href="#cb4-8" aria-hidden="true" tabindex="-1"></a> BitsAndBytesConfig,</span>
<span id="cb4-9"><a href="#cb4-9" aria-hidden="true" tabindex="-1"></a>)</span>
<span id="cb4-10"><a href="#cb4-10" aria-hidden="true" tabindex="-1"></a></span>
<span id="cb4-11"><a href="#cb4-11" aria-hidden="true" tabindex="-1"></a>set_seed(SEED)</span></code><button title="Copy to Clipboard" class="code-copy-button"><i class="bi"></i></button></pre></div>
</div>
</section>
<section id="prepare-the-data" class="level2">
<h2 class="anchored" data-anchor-id="prepare-the-data">Prepare the data</h2>
<p>Begin by loading the data. As the dataset is likely to be quite large, make sure to enable the streaming mode. Streaming allows us to load the data progressively as we iterate over the dataset instead of downloading the whole dataset at once.</p>
<p>We’ll reserve the first 4000 examples as the validation set, and everything else will be the training data.</p>
<div id="cell-9" class="cell">
<div class="sourceCode cell-code" id="cb5"><pre class="sourceCode python code-with-copy"><code class="sourceCode python"><span id="cb5-1"><a href="#cb5-1" aria-hidden="true" tabindex="-1"></a><span class="im">from</span> datasets <span class="im">import</span> load_dataset</span>
<span id="cb5-2"><a href="#cb5-2" aria-hidden="true" tabindex="-1"></a><span class="im">import</span> torch</span>
<span id="cb5-3"><a href="#cb5-3" aria-hidden="true" tabindex="-1"></a><span class="im">from</span> tqdm <span class="im">import</span> tqdm</span>
<span id="cb5-4"><a href="#cb5-4" aria-hidden="true" tabindex="-1"></a></span>
<span id="cb5-5"><a href="#cb5-5" aria-hidden="true" tabindex="-1"></a></span>
<span id="cb5-6"><a href="#cb5-6" aria-hidden="true" tabindex="-1"></a>dataset <span class="op">=</span> load_dataset(</span>
<span id="cb5-7"><a href="#cb5-7" aria-hidden="true" tabindex="-1"></a> DATASET,</span>
<span id="cb5-8"><a href="#cb5-8" aria-hidden="true" tabindex="-1"></a> data_dir<span class="op">=</span><span class="st">"data"</span>,</span>
<span id="cb5-9"><a href="#cb5-9" aria-hidden="true" tabindex="-1"></a> split<span class="op">=</span><span class="st">"train"</span>,</span>
<span id="cb5-10"><a href="#cb5-10" aria-hidden="true" tabindex="-1"></a> streaming<span class="op">=</span><span class="va">True</span>,</span>
<span id="cb5-11"><a href="#cb5-11" aria-hidden="true" tabindex="-1"></a>)</span>
<span id="cb5-12"><a href="#cb5-12" aria-hidden="true" tabindex="-1"></a></span>
<span id="cb5-13"><a href="#cb5-13" aria-hidden="true" tabindex="-1"></a>valid_data <span class="op">=</span> dataset.take(<span class="dv">4000</span>)</span>
<span id="cb5-14"><a href="#cb5-14" aria-hidden="true" tabindex="-1"></a>train_data <span class="op">=</span> dataset.skip(<span class="dv">4000</span>)</span>
<span id="cb5-15"><a href="#cb5-15" aria-hidden="true" tabindex="-1"></a>train_data <span class="op">=</span> train_data.shuffle(buffer_size<span class="op">=</span><span class="dv">5000</span>, seed<span class="op">=</span>SEED)</span></code><button title="Copy to Clipboard" class="code-copy-button"><i class="bi"></i></button></pre></div>
</div>
<p>At this step, the dataset still contains raw data with code of arbitraty length. For training, we need inputs of fixed length. Let’s create an Iterable dataset that would return constant-length chunks of tokens from a stream of text files.</p>
<p>First, let’s estimate the average number of characters per token in the dataset, which will help us later estimate the number of tokens in the text buffer later. By default, we’ll only take 400 examples (<code>nb_examples</code>) from the dataset. Using only a subset of the entire dataset will reduce computational cost while still providing a reasonable estimate of the overall character-to-token ratio.</p>
<div id="cell-11" class="cell" data-outputid="cabf7fd0-a922-4371-cbc6-60ee99ef7469">
<div class="sourceCode cell-code" id="cb6"><pre class="sourceCode python code-with-copy"><code class="sourceCode python"><span id="cb6-1"><a href="#cb6-1" aria-hidden="true" tabindex="-1"></a>tokenizer <span class="op">=</span> AutoTokenizer.from_pretrained(MODEL, trust_remote_code<span class="op">=</span><span class="va">True</span>)</span>
<span id="cb6-2"><a href="#cb6-2" aria-hidden="true" tabindex="-1"></a></span>
<span id="cb6-3"><a href="#cb6-3" aria-hidden="true" tabindex="-1"></a><span class="kw">def</span> chars_token_ratio(dataset, tokenizer, data_column, nb_examples<span class="op">=</span><span class="dv">400</span>):</span>
<span id="cb6-4"><a href="#cb6-4" aria-hidden="true" tabindex="-1"></a> <span class="co">"""</span></span>
<span id="cb6-5"><a href="#cb6-5" aria-hidden="true" tabindex="-1"></a><span class="co"> Estimate the average number of characters per token in the dataset.</span></span>
<span id="cb6-6"><a href="#cb6-6" aria-hidden="true" tabindex="-1"></a><span class="co"> """</span></span>
<span id="cb6-7"><a href="#cb6-7" aria-hidden="true" tabindex="-1"></a></span>
<span id="cb6-8"><a href="#cb6-8" aria-hidden="true" tabindex="-1"></a> total_characters, total_tokens <span class="op">=</span> <span class="dv">0</span>, <span class="dv">0</span></span>
<span id="cb6-9"><a href="#cb6-9" aria-hidden="true" tabindex="-1"></a> <span class="cf">for</span> _, example <span class="kw">in</span> tqdm(<span class="bu">zip</span>(<span class="bu">range</span>(nb_examples), <span class="bu">iter</span>(dataset)), total<span class="op">=</span>nb_examples):</span>
<span id="cb6-10"><a href="#cb6-10" aria-hidden="true" tabindex="-1"></a> total_characters <span class="op">+=</span> <span class="bu">len</span>(example[data_column])</span>
<span id="cb6-11"><a href="#cb6-11" aria-hidden="true" tabindex="-1"></a> total_tokens <span class="op">+=</span> <span class="bu">len</span>(tokenizer(example[data_column]).tokens())</span>
<span id="cb6-12"><a href="#cb6-12" aria-hidden="true" tabindex="-1"></a></span>
<span id="cb6-13"><a href="#cb6-13" aria-hidden="true" tabindex="-1"></a> <span class="cf">return</span> total_characters <span class="op">/</span> total_tokens</span>
<span id="cb6-14"><a href="#cb6-14" aria-hidden="true" tabindex="-1"></a></span>
<span id="cb6-15"><a href="#cb6-15" aria-hidden="true" tabindex="-1"></a></span>
<span id="cb6-16"><a href="#cb6-16" aria-hidden="true" tabindex="-1"></a>chars_per_token <span class="op">=</span> chars_token_ratio(train_data, tokenizer, DATA_COLUMN)</span>
<span id="cb6-17"><a href="#cb6-17" aria-hidden="true" tabindex="-1"></a><span class="bu">print</span>(<span class="ss">f"The character to token ratio of the dataset is: </span><span class="sc">{</span>chars_per_token<span class="sc">:.2f}</span><span class="ss">"</span>)</span></code><button title="Copy to Clipboard" class="code-copy-button"><i class="bi"></i></button></pre></div>
<div class="cell-output cell-output-stderr">
<pre><code>100%|██████████| 400/400 [00:10&lt;00:00, 39.87it/s] </code></pre>
</div>
<div class="cell-output cell-output-stdout">
<pre><code>The character to token ratio of the dataset is: 2.43</code></pre>
</div>
<div class="cell-output cell-output-stderr">
<pre><code></code></pre>
</div>
</div>
<p>The character-to-token ratio can also be used as an indicator of the quality of text tokenization. For instance, a character-to-token ratio of 1.0 would mean that each character is represented with a token, which is not very meaningful. This would indicate poor tokenization. In standard English text, one token is typically equivalent to approximately four characters, meaning the character-to-token ratio is around 4.0. We can expect a lower ratio in the code dataset, but generally speaking, a number between 2.0 and 3.5 can be considered good enough.</p>
<p><strong>Optional FIM transformations</strong></p>
<p>Autoregressive language models typically generate sequences from left to right. By applying the FIM transformations, the model can also learn to infill text. Check out <a href="https://arxiv.org/pdf/2207.14255.pdf">“Efficient Training of Language Models to Fill in the Middle” paper</a> to learn more about the technique. We’ll define the FIM transformations here and will use them when creating the Iterable Dataset. However, if you want to omit transformations, feel free to set <code>fim_rate</code> to 0.</p>
<div id="cell-14" class="cell">
<div class="sourceCode cell-code" id="cb10"><pre class="sourceCode python code-with-copy"><code class="sourceCode python"><span id="cb10-1"><a href="#cb10-1" aria-hidden="true" tabindex="-1"></a><span class="im">import</span> functools</span>
<span id="cb10-2"><a href="#cb10-2" aria-hidden="true" tabindex="-1"></a><span class="im">import</span> numpy <span class="im">as</span> np</span>
<span id="cb10-3"><a href="#cb10-3" aria-hidden="true" tabindex="-1"></a></span>
<span id="cb10-4"><a href="#cb10-4" aria-hidden="true" tabindex="-1"></a></span>
<span id="cb10-5"><a href="#cb10-5" aria-hidden="true" tabindex="-1"></a><span class="co"># Helper function to get token ids of the special tokens for prefix, suffix and middle for FIM transformations.</span></span>
<span id="cb10-6"><a href="#cb10-6" aria-hidden="true" tabindex="-1"></a><span class="at">@functools.lru_cache</span>(maxsize<span class="op">=</span><span class="va">None</span>)</span>
<span id="cb10-7"><a href="#cb10-7" aria-hidden="true" tabindex="-1"></a><span class="kw">def</span> get_fim_token_ids(tokenizer):</span>
<span id="cb10-8"><a href="#cb10-8" aria-hidden="true" tabindex="-1"></a> <span class="cf">try</span>:</span>
<span id="cb10-9"><a href="#cb10-9" aria-hidden="true" tabindex="-1"></a> FIM_PREFIX, FIM_MIDDLE, FIM_SUFFIX, FIM_PAD <span class="op">=</span> tokenizer.special_tokens_map[<span class="st">"additional_special_tokens"</span>][<span class="dv">1</span>:<span class="dv">5</span>]</span>
<span id="cb10-10"><a href="#cb10-10" aria-hidden="true" tabindex="-1"></a> suffix_tok_id, prefix_tok_id, middle_tok_id, pad_tok_id <span class="op">=</span> (</span>
<span id="cb10-11"><a href="#cb10-11" aria-hidden="true" tabindex="-1"></a> tokenizer.vocab[tok] <span class="cf">for</span> tok <span class="kw">in</span> [FIM_SUFFIX, FIM_PREFIX, FIM_MIDDLE, FIM_PAD]</span>
<span id="cb10-12"><a href="#cb10-12" aria-hidden="true" tabindex="-1"></a> )</span>
<span id="cb10-13"><a href="#cb10-13" aria-hidden="true" tabindex="-1"></a> <span class="cf">except</span> <span class="pp">KeyError</span>:</span>
<span id="cb10-14"><a href="#cb10-14" aria-hidden="true" tabindex="-1"></a> suffix_tok_id, prefix_tok_id, middle_tok_id, pad_tok_id <span class="op">=</span> <span class="va">None</span>, <span class="va">None</span>, <span class="va">None</span>, <span class="va">None</span></span>
<span id="cb10-15"><a href="#cb10-15" aria-hidden="true" tabindex="-1"></a> <span class="cf">return</span> suffix_tok_id, prefix_tok_id, middle_tok_id, pad_tok_id</span>
<span id="cb10-16"><a href="#cb10-16" aria-hidden="true" tabindex="-1"></a></span>
<span id="cb10-17"><a href="#cb10-17" aria-hidden="true" tabindex="-1"></a></span>
<span id="cb10-18"><a href="#cb10-18" aria-hidden="true" tabindex="-1"></a><span class="co">## Adapted from https://github.com/bigcode-project/Megatron-LM/blob/6c4bf908df8fd86b4977f54bf5b8bd4b521003d1/megatron/data/gpt_dataset.py</span></span>
<span id="cb10-19"><a href="#cb10-19" aria-hidden="true" tabindex="-1"></a><span class="kw">def</span> permute(</span>
<span id="cb10-20"><a href="#cb10-20" aria-hidden="true" tabindex="-1"></a> sample,</span>
<span id="cb10-21"><a href="#cb10-21" aria-hidden="true" tabindex="-1"></a> np_rng,</span>
<span id="cb10-22"><a href="#cb10-22" aria-hidden="true" tabindex="-1"></a> suffix_tok_id,</span>
<span id="cb10-23"><a href="#cb10-23" aria-hidden="true" tabindex="-1"></a> prefix_tok_id,</span>
<span id="cb10-24"><a href="#cb10-24" aria-hidden="true" tabindex="-1"></a> middle_tok_id,</span>
<span id="cb10-25"><a href="#cb10-25" aria-hidden="true" tabindex="-1"></a> pad_tok_id,</span>
<span id="cb10-26"><a href="#cb10-26" aria-hidden="true" tabindex="-1"></a> fim_rate<span class="op">=</span><span class="fl">0.5</span>,</span>
<span id="cb10-27"><a href="#cb10-27" aria-hidden="true" tabindex="-1"></a> fim_spm_rate<span class="op">=</span><span class="fl">0.5</span>,</span>
<span id="cb10-28"><a href="#cb10-28" aria-hidden="true" tabindex="-1"></a> truncate_or_pad<span class="op">=</span><span class="va">False</span>,</span>
<span id="cb10-29"><a href="#cb10-29" aria-hidden="true" tabindex="-1"></a>):</span>
<span id="cb10-30"><a href="#cb10-30" aria-hidden="true" tabindex="-1"></a> <span class="co">"""</span></span>
<span id="cb10-31"><a href="#cb10-31" aria-hidden="true" tabindex="-1"></a><span class="co"> Take in a sample (list of tokens) and perform a FIM transformation on it with a probability of fim_rate, using two FIM modes:</span></span>
<span id="cb10-32"><a href="#cb10-32" aria-hidden="true" tabindex="-1"></a><span class="co"> PSM and SPM (with a probability of fim_spm_rate).</span></span>
<span id="cb10-33"><a href="#cb10-33" aria-hidden="true" tabindex="-1"></a><span class="co"> """</span></span>
<span id="cb10-34"><a href="#cb10-34" aria-hidden="true" tabindex="-1"></a></span>
<span id="cb10-35"><a href="#cb10-35" aria-hidden="true" tabindex="-1"></a> <span class="co"># The if condition will trigger with the probability of fim_rate</span></span>
<span id="cb10-36"><a href="#cb10-36" aria-hidden="true" tabindex="-1"></a> <span class="co"># This means FIM transformations will apply to samples with a probability of fim_rate</span></span>
<span id="cb10-37"><a href="#cb10-37" aria-hidden="true" tabindex="-1"></a> <span class="cf">if</span> np_rng.binomial(<span class="dv">1</span>, fim_rate):</span>
<span id="cb10-38"><a href="#cb10-38" aria-hidden="true" tabindex="-1"></a></span>
<span id="cb10-39"><a href="#cb10-39" aria-hidden="true" tabindex="-1"></a> <span class="co"># Split the sample into prefix, middle, and suffix, based on randomly generated indices stored in the boundaries list.</span></span>
<span id="cb10-40"><a href="#cb10-40" aria-hidden="true" tabindex="-1"></a> boundaries <span class="op">=</span> <span class="bu">list</span>(np_rng.randint(low<span class="op">=</span><span class="dv">0</span>, high<span class="op">=</span><span class="bu">len</span>(sample) <span class="op">+</span> <span class="dv">1</span>, size<span class="op">=</span><span class="dv">2</span>))</span>
<span id="cb10-41"><a href="#cb10-41" aria-hidden="true" tabindex="-1"></a> boundaries.sort()</span>
<span id="cb10-42"><a href="#cb10-42" aria-hidden="true" tabindex="-1"></a></span>
<span id="cb10-43"><a href="#cb10-43" aria-hidden="true" tabindex="-1"></a> prefix <span class="op">=</span> np.array(sample[: boundaries[<span class="dv">0</span>]], dtype<span class="op">=</span>np.int64)</span>
<span id="cb10-44"><a href="#cb10-44" aria-hidden="true" tabindex="-1"></a> middle <span class="op">=</span> np.array(sample[boundaries[<span class="dv">0</span>] : boundaries[<span class="dv">1</span>]], dtype<span class="op">=</span>np.int64)</span>
<span id="cb10-45"><a href="#cb10-45" aria-hidden="true" tabindex="-1"></a> suffix <span class="op">=</span> np.array(sample[boundaries[<span class="dv">1</span>] :], dtype<span class="op">=</span>np.int64)</span>
<span id="cb10-46"><a href="#cb10-46" aria-hidden="true" tabindex="-1"></a></span>
<span id="cb10-47"><a href="#cb10-47" aria-hidden="true" tabindex="-1"></a> <span class="cf">if</span> truncate_or_pad:</span>
<span id="cb10-48"><a href="#cb10-48" aria-hidden="true" tabindex="-1"></a> <span class="co"># calculate the new total length of the sample, taking into account tokens indicating prefix, middle, and suffix</span></span>
<span id="cb10-49"><a href="#cb10-49" aria-hidden="true" tabindex="-1"></a> new_length <span class="op">=</span> suffix.shape[<span class="dv">0</span>] <span class="op">+</span> prefix.shape[<span class="dv">0</span>] <span class="op">+</span> middle.shape[<span class="dv">0</span>] <span class="op">+</span> <span class="dv">3</span></span>
<span id="cb10-50"><a href="#cb10-50" aria-hidden="true" tabindex="-1"></a> diff <span class="op">=</span> new_length <span class="op">-</span> <span class="bu">len</span>(sample)</span>
<span id="cb10-51"><a href="#cb10-51" aria-hidden="true" tabindex="-1"></a></span>
<span id="cb10-52"><a href="#cb10-52" aria-hidden="true" tabindex="-1"></a> <span class="co"># trancate or pad if there's a difference in length between the new length and the original</span></span>
<span id="cb10-53"><a href="#cb10-53" aria-hidden="true" tabindex="-1"></a> <span class="cf">if</span> diff <span class="op">&gt;</span> <span class="dv">0</span>:</span>
<span id="cb10-54"><a href="#cb10-54" aria-hidden="true" tabindex="-1"></a> <span class="cf">if</span> suffix.shape[<span class="dv">0</span>] <span class="op">&lt;=</span> diff:</span>
<span id="cb10-55"><a href="#cb10-55" aria-hidden="true" tabindex="-1"></a> <span class="cf">return</span> sample, np_rng</span>
<span id="cb10-56"><a href="#cb10-56" aria-hidden="true" tabindex="-1"></a> suffix <span class="op">=</span> suffix[: suffix.shape[<span class="dv">0</span>] <span class="op">-</span> diff]</span>
<span id="cb10-57"><a href="#cb10-57" aria-hidden="true" tabindex="-1"></a> <span class="cf">elif</span> diff <span class="op">&lt;</span> <span class="dv">0</span>:</span>
<span id="cb10-58"><a href="#cb10-58" aria-hidden="true" tabindex="-1"></a> suffix <span class="op">=</span> np.concatenate([suffix, np.full((<span class="op">-</span><span class="dv">1</span> <span class="op">*</span> diff), pad_tok_id)])</span>
<span id="cb10-59"><a href="#cb10-59" aria-hidden="true" tabindex="-1"></a></span>
<span id="cb10-60"><a href="#cb10-60" aria-hidden="true" tabindex="-1"></a> <span class="co"># With the probability of fim_spm_rateapply SPM variant of FIM transformations</span></span>
<span id="cb10-61"><a href="#cb10-61" aria-hidden="true" tabindex="-1"></a> <span class="co"># SPM: suffix, prefix, middle</span></span>
<span id="cb10-62"><a href="#cb10-62" aria-hidden="true" tabindex="-1"></a> <span class="cf">if</span> np_rng.binomial(<span class="dv">1</span>, fim_spm_rate):</span>
<span id="cb10-63"><a href="#cb10-63" aria-hidden="true" tabindex="-1"></a> new_sample <span class="op">=</span> np.concatenate(</span>
<span id="cb10-64"><a href="#cb10-64" aria-hidden="true" tabindex="-1"></a> [</span>
<span id="cb10-65"><a href="#cb10-65" aria-hidden="true" tabindex="-1"></a> [prefix_tok_id, suffix_tok_id],</span>
<span id="cb10-66"><a href="#cb10-66" aria-hidden="true" tabindex="-1"></a> suffix,</span>
<span id="cb10-67"><a href="#cb10-67" aria-hidden="true" tabindex="-1"></a> [middle_tok_id],</span>
<span id="cb10-68"><a href="#cb10-68" aria-hidden="true" tabindex="-1"></a> prefix,</span>
<span id="cb10-69"><a href="#cb10-69" aria-hidden="true" tabindex="-1"></a> middle,</span>
<span id="cb10-70"><a href="#cb10-70" aria-hidden="true" tabindex="-1"></a> ]</span>
<span id="cb10-71"><a href="#cb10-71" aria-hidden="true" tabindex="-1"></a> )</span>
<span id="cb10-72"><a href="#cb10-72" aria-hidden="true" tabindex="-1"></a> <span class="co"># Otherwise, apply the PSM variant of FIM transformations</span></span>
<span id="cb10-73"><a href="#cb10-73" aria-hidden="true" tabindex="-1"></a> <span class="co"># PSM: prefix, suffix, middle</span></span>
<span id="cb10-74"><a href="#cb10-74" aria-hidden="true" tabindex="-1"></a> <span class="cf">else</span>:</span>
<span id="cb10-75"><a href="#cb10-75" aria-hidden="true" tabindex="-1"></a></span>
<span id="cb10-76"><a href="#cb10-76" aria-hidden="true" tabindex="-1"></a> new_sample <span class="op">=</span> np.concatenate(</span>
<span id="cb10-77"><a href="#cb10-77" aria-hidden="true" tabindex="-1"></a> [</span>
<span id="cb10-78"><a href="#cb10-78" aria-hidden="true" tabindex="-1"></a> [prefix_tok_id],</span>
<span id="cb10-79"><a href="#cb10-79" aria-hidden="true" tabindex="-1"></a> prefix,</span>
<span id="cb10-80"><a href="#cb10-80" aria-hidden="true" tabindex="-1"></a> [suffix_tok_id],</span>
<span id="cb10-81"><a href="#cb10-81" aria-hidden="true" tabindex="-1"></a> suffix,</span>
<span id="cb10-82"><a href="#cb10-82" aria-hidden="true" tabindex="-1"></a> [middle_tok_id],</span>
<span id="cb10-83"><a href="#cb10-83" aria-hidden="true" tabindex="-1"></a> middle,</span>
<span id="cb10-84"><a href="#cb10-84" aria-hidden="true" tabindex="-1"></a> ]</span>
<span id="cb10-85"><a href="#cb10-85" aria-hidden="true" tabindex="-1"></a> )</span>
<span id="cb10-86"><a href="#cb10-86" aria-hidden="true" tabindex="-1"></a> <span class="cf">else</span>:</span>
<span id="cb10-87"><a href="#cb10-87" aria-hidden="true" tabindex="-1"></a> <span class="co"># don't apply FIM transformations</span></span>
<span id="cb10-88"><a href="#cb10-88" aria-hidden="true" tabindex="-1"></a> new_sample <span class="op">=</span> sample</span>
<span id="cb10-89"><a href="#cb10-89" aria-hidden="true" tabindex="-1"></a></span>
<span id="cb10-90"><a href="#cb10-90" aria-hidden="true" tabindex="-1"></a> <span class="cf">return</span> <span class="bu">list</span>(new_sample), np_rng</span></code><button title="Copy to Clipboard" class="code-copy-button"><i class="bi"></i></button></pre></div>
</div>
<p>Let’s define the <code>ConstantLengthDataset</code>, an Iterable dataset that will return constant-length chunks of tokens. To do so, we’ll read a buffer of text from the original dataset until we hit the size limits and then apply tokenizer to convert the raw text into tokenized inputs. Optionally, we’ll perform FIM transformations on some sequences (the proportion of sequences affected is controlled by <code>fim_rate</code>).</p>
<p>Once defined, we can create instances of the <code>ConstantLengthDataset</code> from both training and validation data.</p>
<div id="cell-16" class="cell">
<div class="sourceCode cell-code" id="cb11"><pre class="sourceCode python code-with-copy"><code class="sourceCode python"><span id="cb11-1"><a href="#cb11-1" aria-hidden="true" tabindex="-1"></a><span class="im">from</span> torch.utils.data <span class="im">import</span> IterableDataset</span>
<span id="cb11-2"><a href="#cb11-2" aria-hidden="true" tabindex="-1"></a><span class="im">from</span> torch.utils.data.dataloader <span class="im">import</span> DataLoader</span>
<span id="cb11-3"><a href="#cb11-3" aria-hidden="true" tabindex="-1"></a><span class="im">import</span> random</span>
<span id="cb11-4"><a href="#cb11-4" aria-hidden="true" tabindex="-1"></a></span>
<span id="cb11-5"><a href="#cb11-5" aria-hidden="true" tabindex="-1"></a><span class="co"># Create an Iterable dataset that returns constant-length chunks of tokens from a stream of text files.</span></span>
<span id="cb11-6"><a href="#cb11-6" aria-hidden="true" tabindex="-1"></a></span>
<span id="cb11-7"><a href="#cb11-7" aria-hidden="true" tabindex="-1"></a><span class="kw">class</span> ConstantLengthDataset(IterableDataset):</span>
<span id="cb11-8"><a href="#cb11-8" aria-hidden="true" tabindex="-1"></a> <span class="co">"""</span></span>
<span id="cb11-9"><a href="#cb11-9" aria-hidden="true" tabindex="-1"></a><span class="co"> Iterable dataset that returns constant length chunks of tokens from stream of text files.</span></span>
<span id="cb11-10"><a href="#cb11-10" aria-hidden="true" tabindex="-1"></a><span class="co"> Args:</span></span>
<span id="cb11-11"><a href="#cb11-11" aria-hidden="true" tabindex="-1"></a><span class="co"> tokenizer (Tokenizer): The processor used for proccessing the data.</span></span>
<span id="cb11-12"><a href="#cb11-12" aria-hidden="true" tabindex="-1"></a><span class="co"> dataset (dataset.Dataset): Dataset with text files.</span></span>
<span id="cb11-13"><a href="#cb11-13" aria-hidden="true" tabindex="-1"></a><span class="co"> infinite (bool): If True the iterator is reset after dataset reaches end else stops.</span></span>
<span id="cb11-14"><a href="#cb11-14" aria-hidden="true" tabindex="-1"></a><span class="co"> seq_length (int): Length of token sequences to return.</span></span>
<span id="cb11-15"><a href="#cb11-15" aria-hidden="true" tabindex="-1"></a><span class="co"> num_of_sequences (int): Number of token sequences to keep in buffer.</span></span>
<span id="cb11-16"><a href="#cb11-16" aria-hidden="true" tabindex="-1"></a><span class="co"> chars_per_token (int): Number of characters per token used to estimate number of tokens in text buffer.</span></span>
<span id="cb11-17"><a href="#cb11-17" aria-hidden="true" tabindex="-1"></a><span class="co"> fim_rate (float): Rate (0.0 to 1.0) that sample will be permuted with FIM.</span></span>
<span id="cb11-18"><a href="#cb11-18" aria-hidden="true" tabindex="-1"></a><span class="co"> fim_spm_rate (float): Rate (0.0 to 1.0) of FIM permuations that will use SPM.</span></span>
<span id="cb11-19"><a href="#cb11-19" aria-hidden="true" tabindex="-1"></a><span class="co"> seed (int): Seed for random number generator.</span></span>
<span id="cb11-20"><a href="#cb11-20" aria-hidden="true" tabindex="-1"></a><span class="co"> """</span></span>
<span id="cb11-21"><a href="#cb11-21" aria-hidden="true" tabindex="-1"></a></span>
<span id="cb11-22"><a href="#cb11-22" aria-hidden="true" tabindex="-1"></a> <span class="kw">def</span> <span class="fu">__init__</span>(</span>
<span id="cb11-23"><a href="#cb11-23" aria-hidden="true" tabindex="-1"></a> <span class="va">self</span>,</span>
<span id="cb11-24"><a href="#cb11-24" aria-hidden="true" tabindex="-1"></a> tokenizer,</span>
<span id="cb11-25"><a href="#cb11-25" aria-hidden="true" tabindex="-1"></a> dataset,</span>
<span id="cb11-26"><a href="#cb11-26" aria-hidden="true" tabindex="-1"></a> infinite<span class="op">=</span><span class="va">False</span>,</span>
<span id="cb11-27"><a href="#cb11-27" aria-hidden="true" tabindex="-1"></a> seq_length<span class="op">=</span><span class="dv">1024</span>,</span>
<span id="cb11-28"><a href="#cb11-28" aria-hidden="true" tabindex="-1"></a> num_of_sequences<span class="op">=</span><span class="dv">1024</span>,</span>
<span id="cb11-29"><a href="#cb11-29" aria-hidden="true" tabindex="-1"></a> chars_per_token<span class="op">=</span><span class="fl">3.6</span>,</span>
<span id="cb11-30"><a href="#cb11-30" aria-hidden="true" tabindex="-1"></a> content_field<span class="op">=</span><span class="st">"content"</span>,</span>
<span id="cb11-31"><a href="#cb11-31" aria-hidden="true" tabindex="-1"></a> fim_rate<span class="op">=</span><span class="fl">0.5</span>,</span>
<span id="cb11-32"><a href="#cb11-32" aria-hidden="true" tabindex="-1"></a> fim_spm_rate<span class="op">=</span><span class="fl">0.5</span>,</span>
<span id="cb11-33"><a href="#cb11-33" aria-hidden="true" tabindex="-1"></a> seed<span class="op">=</span><span class="dv">0</span>,</span>
<span id="cb11-34"><a href="#cb11-34" aria-hidden="true" tabindex="-1"></a> ):</span>
<span id="cb11-35"><a href="#cb11-35" aria-hidden="true" tabindex="-1"></a> <span class="va">self</span>.tokenizer <span class="op">=</span> tokenizer</span>
<span id="cb11-36"><a href="#cb11-36" aria-hidden="true" tabindex="-1"></a> <span class="va">self</span>.concat_token_id <span class="op">=</span> tokenizer.eos_token_id</span>
<span id="cb11-37"><a href="#cb11-37" aria-hidden="true" tabindex="-1"></a> <span class="va">self</span>.dataset <span class="op">=</span> dataset</span>
<span id="cb11-38"><a href="#cb11-38" aria-hidden="true" tabindex="-1"></a> <span class="va">self</span>.seq_length <span class="op">=</span> seq_length</span>
<span id="cb11-39"><a href="#cb11-39" aria-hidden="true" tabindex="-1"></a> <span class="va">self</span>.infinite <span class="op">=</span> infinite</span>
<span id="cb11-40"><a href="#cb11-40" aria-hidden="true" tabindex="-1"></a> <span class="va">self</span>.current_size <span class="op">=</span> <span class="dv">0</span></span>
<span id="cb11-41"><a href="#cb11-41" aria-hidden="true" tabindex="-1"></a> <span class="va">self</span>.max_buffer_size <span class="op">=</span> seq_length <span class="op">*</span> chars_per_token <span class="op">*</span> num_of_sequences</span>
<span id="cb11-42"><a href="#cb11-42" aria-hidden="true" tabindex="-1"></a> <span class="va">self</span>.content_field <span class="op">=</span> content_field</span>
<span id="cb11-43"><a href="#cb11-43" aria-hidden="true" tabindex="-1"></a> <span class="va">self</span>.fim_rate <span class="op">=</span> fim_rate</span>
<span id="cb11-44"><a href="#cb11-44" aria-hidden="true" tabindex="-1"></a> <span class="va">self</span>.fim_spm_rate <span class="op">=</span> fim_spm_rate</span>
<span id="cb11-45"><a href="#cb11-45" aria-hidden="true" tabindex="-1"></a> <span class="va">self</span>.seed <span class="op">=</span> seed</span>
<span id="cb11-46"><a href="#cb11-46" aria-hidden="true" tabindex="-1"></a></span>
<span id="cb11-47"><a href="#cb11-47" aria-hidden="true" tabindex="-1"></a> (</span>
<span id="cb11-48"><a href="#cb11-48" aria-hidden="true" tabindex="-1"></a> <span class="va">self</span>.suffix_tok_id,</span>
<span id="cb11-49"><a href="#cb11-49" aria-hidden="true" tabindex="-1"></a> <span class="va">self</span>.prefix_tok_id,</span>
<span id="cb11-50"><a href="#cb11-50" aria-hidden="true" tabindex="-1"></a> <span class="va">self</span>.middle_tok_id,</span>
<span id="cb11-51"><a href="#cb11-51" aria-hidden="true" tabindex="-1"></a> <span class="va">self</span>.pad_tok_id,</span>
<span id="cb11-52"><a href="#cb11-52" aria-hidden="true" tabindex="-1"></a> ) <span class="op">=</span> get_fim_token_ids(<span class="va">self</span>.tokenizer)</span>
<span id="cb11-53"><a href="#cb11-53" aria-hidden="true" tabindex="-1"></a> <span class="cf">if</span> <span class="kw">not</span> <span class="va">self</span>.suffix_tok_id <span class="kw">and</span> <span class="va">self</span>.fim_rate <span class="op">&gt;</span> <span class="dv">0</span>:</span>
<span id="cb11-54"><a href="#cb11-54" aria-hidden="true" tabindex="-1"></a> <span class="bu">print</span>(<span class="st">"FIM is not supported by tokenizer, disabling FIM"</span>)</span>
<span id="cb11-55"><a href="#cb11-55" aria-hidden="true" tabindex="-1"></a> <span class="va">self</span>.fim_rate <span class="op">=</span> <span class="dv">0</span></span>
<span id="cb11-56"><a href="#cb11-56" aria-hidden="true" tabindex="-1"></a></span>
<span id="cb11-57"><a href="#cb11-57" aria-hidden="true" tabindex="-1"></a> <span class="kw">def</span> <span class="fu">__iter__</span>(<span class="va">self</span>):</span>
<span id="cb11-58"><a href="#cb11-58" aria-hidden="true" tabindex="-1"></a> iterator <span class="op">=</span> <span class="bu">iter</span>(<span class="va">self</span>.dataset)</span>
<span id="cb11-59"><a href="#cb11-59" aria-hidden="true" tabindex="-1"></a> more_examples <span class="op">=</span> <span class="va">True</span></span>
<span id="cb11-60"><a href="#cb11-60" aria-hidden="true" tabindex="-1"></a> np_rng <span class="op">=</span> np.random.RandomState(seed<span class="op">=</span><span class="va">self</span>.seed)</span>
<span id="cb11-61"><a href="#cb11-61" aria-hidden="true" tabindex="-1"></a> <span class="cf">while</span> more_examples:</span>
<span id="cb11-62"><a href="#cb11-62" aria-hidden="true" tabindex="-1"></a> <span class="bu">buffer</span>, buffer_len <span class="op">=</span> [], <span class="dv">0</span></span>
<span id="cb11-63"><a href="#cb11-63" aria-hidden="true" tabindex="-1"></a> <span class="cf">while</span> <span class="va">True</span>:</span>
<span id="cb11-64"><a href="#cb11-64" aria-hidden="true" tabindex="-1"></a> <span class="cf">if</span> buffer_len <span class="op">&gt;=</span> <span class="va">self</span>.max_buffer_size:</span>
<span id="cb11-65"><a href="#cb11-65" aria-hidden="true" tabindex="-1"></a> <span class="cf">break</span></span>
<span id="cb11-66"><a href="#cb11-66" aria-hidden="true" tabindex="-1"></a> <span class="cf">try</span>:</span>
<span id="cb11-67"><a href="#cb11-67" aria-hidden="true" tabindex="-1"></a> <span class="bu">buffer</span>.append(<span class="bu">next</span>(iterator)[<span class="va">self</span>.content_field])</span>
<span id="cb11-68"><a href="#cb11-68" aria-hidden="true" tabindex="-1"></a> buffer_len <span class="op">+=</span> <span class="bu">len</span>(<span class="bu">buffer</span>[<span class="op">-</span><span class="dv">1</span>])</span>
<span id="cb11-69"><a href="#cb11-69" aria-hidden="true" tabindex="-1"></a> <span class="cf">except</span> <span class="pp">StopIteration</span>:</span>
<span id="cb11-70"><a href="#cb11-70" aria-hidden="true" tabindex="-1"></a> <span class="cf">if</span> <span class="va">self</span>.infinite:</span>
<span id="cb11-71"><a href="#cb11-71" aria-hidden="true" tabindex="-1"></a> iterator <span class="op">=</span> <span class="bu">iter</span>(<span class="va">self</span>.dataset)</span>
<span id="cb11-72"><a href="#cb11-72" aria-hidden="true" tabindex="-1"></a> <span class="cf">else</span>:</span>
<span id="cb11-73"><a href="#cb11-73" aria-hidden="true" tabindex="-1"></a> more_examples <span class="op">=</span> <span class="va">False</span></span>
<span id="cb11-74"><a href="#cb11-74" aria-hidden="true" tabindex="-1"></a> <span class="cf">break</span></span>
<span id="cb11-75"><a href="#cb11-75" aria-hidden="true" tabindex="-1"></a> tokenized_inputs <span class="op">=</span> <span class="va">self</span>.tokenizer(<span class="bu">buffer</span>, truncation<span class="op">=</span><span class="va">False</span>)[<span class="st">"input_ids"</span>]</span>
<span id="cb11-76"><a href="#cb11-76" aria-hidden="true" tabindex="-1"></a> all_token_ids <span class="op">=</span> []</span>
<span id="cb11-77"><a href="#cb11-77" aria-hidden="true" tabindex="-1"></a></span>
<span id="cb11-78"><a href="#cb11-78" aria-hidden="true" tabindex="-1"></a> <span class="cf">for</span> tokenized_input <span class="kw">in</span> tokenized_inputs:</span>
<span id="cb11-79"><a href="#cb11-79" aria-hidden="true" tabindex="-1"></a> <span class="co"># optionally do FIM permutations</span></span>
<span id="cb11-80"><a href="#cb11-80" aria-hidden="true" tabindex="-1"></a> <span class="cf">if</span> <span class="va">self</span>.fim_rate <span class="op">&gt;</span> <span class="dv">0</span>:</span>
<span id="cb11-81"><a href="#cb11-81" aria-hidden="true" tabindex="-1"></a> tokenized_input, np_rng <span class="op">=</span> permute(</span>
<span id="cb11-82"><a href="#cb11-82" aria-hidden="true" tabindex="-1"></a> tokenized_input,</span>
<span id="cb11-83"><a href="#cb11-83" aria-hidden="true" tabindex="-1"></a> np_rng,</span>
<span id="cb11-84"><a href="#cb11-84" aria-hidden="true" tabindex="-1"></a> <span class="va">self</span>.suffix_tok_id,</span>
<span id="cb11-85"><a href="#cb11-85" aria-hidden="true" tabindex="-1"></a> <span class="va">self</span>.prefix_tok_id,</span>
<span id="cb11-86"><a href="#cb11-86" aria-hidden="true" tabindex="-1"></a> <span class="va">self</span>.middle_tok_id,</span>
<span id="cb11-87"><a href="#cb11-87" aria-hidden="true" tabindex="-1"></a> <span class="va">self</span>.pad_tok_id,</span>
<span id="cb11-88"><a href="#cb11-88" aria-hidden="true" tabindex="-1"></a> fim_rate<span class="op">=</span><span class="va">self</span>.fim_rate,</span>
<span id="cb11-89"><a href="#cb11-89" aria-hidden="true" tabindex="-1"></a> fim_spm_rate<span class="op">=</span><span class="va">self</span>.fim_spm_rate,</span>
<span id="cb11-90"><a href="#cb11-90" aria-hidden="true" tabindex="-1"></a> truncate_or_pad<span class="op">=</span><span class="va">False</span>,</span>
<span id="cb11-91"><a href="#cb11-91" aria-hidden="true" tabindex="-1"></a> )</span>
<span id="cb11-92"><a href="#cb11-92" aria-hidden="true" tabindex="-1"></a></span>
<span id="cb11-93"><a href="#cb11-93" aria-hidden="true" tabindex="-1"></a> all_token_ids.extend(tokenized_input <span class="op">+</span> [<span class="va">self</span>.concat_token_id])</span>
<span id="cb11-94"><a href="#cb11-94" aria-hidden="true" tabindex="-1"></a> examples <span class="op">=</span> []</span>
<span id="cb11-95"><a href="#cb11-95" aria-hidden="true" tabindex="-1"></a> <span class="cf">for</span> i <span class="kw">in</span> <span class="bu">range</span>(<span class="dv">0</span>, <span class="bu">len</span>(all_token_ids), <span class="va">self</span>.seq_length):</span>
<span id="cb11-96"><a href="#cb11-96" aria-hidden="true" tabindex="-1"></a> input_ids <span class="op">=</span> all_token_ids[i : i <span class="op">+</span> <span class="va">self</span>.seq_length]</span>
<span id="cb11-97"><a href="#cb11-97" aria-hidden="true" tabindex="-1"></a> <span class="cf">if</span> <span class="bu">len</span>(input_ids) <span class="op">==</span> <span class="va">self</span>.seq_length:</span>
<span id="cb11-98"><a href="#cb11-98" aria-hidden="true" tabindex="-1"></a> examples.append(input_ids)</span>
<span id="cb11-99"><a href="#cb11-99" aria-hidden="true" tabindex="-1"></a> random.shuffle(examples)</span>
<span id="cb11-100"><a href="#cb11-100" aria-hidden="true" tabindex="-1"></a> <span class="cf">for</span> example <span class="kw">in</span> examples:</span>
<span id="cb11-101"><a href="#cb11-101" aria-hidden="true" tabindex="-1"></a> <span class="va">self</span>.current_size <span class="op">+=</span> <span class="dv">1</span></span>
<span id="cb11-102"><a href="#cb11-102" aria-hidden="true" tabindex="-1"></a> <span class="cf">yield</span> {</span>
<span id="cb11-103"><a href="#cb11-103" aria-hidden="true" tabindex="-1"></a> <span class="st">"input_ids"</span>: torch.LongTensor(example),</span>
<span id="cb11-104"><a href="#cb11-104" aria-hidden="true" tabindex="-1"></a> <span class="st">"labels"</span>: torch.LongTensor(example),</span>
<span id="cb11-105"><a href="#cb11-105" aria-hidden="true" tabindex="-1"></a> }</span>
<span id="cb11-106"><a href="#cb11-106" aria-hidden="true" tabindex="-1"></a></span>
<span id="cb11-107"><a href="#cb11-107" aria-hidden="true" tabindex="-1"></a></span>
<span id="cb11-108"><a href="#cb11-108" aria-hidden="true" tabindex="-1"></a>train_dataset <span class="op">=</span> ConstantLengthDataset(</span>
<span id="cb11-109"><a href="#cb11-109" aria-hidden="true" tabindex="-1"></a> tokenizer,</span>
<span id="cb11-110"><a href="#cb11-110" aria-hidden="true" tabindex="-1"></a> train_data,</span>
<span id="cb11-111"><a href="#cb11-111" aria-hidden="true" tabindex="-1"></a> infinite<span class="op">=</span><span class="va">True</span>,</span>
<span id="cb11-112"><a href="#cb11-112" aria-hidden="true" tabindex="-1"></a> seq_length<span class="op">=</span>SEQ_LENGTH,</span>
<span id="cb11-113"><a href="#cb11-113" aria-hidden="true" tabindex="-1"></a> chars_per_token<span class="op">=</span>chars_per_token,</span>
<span id="cb11-114"><a href="#cb11-114" aria-hidden="true" tabindex="-1"></a> content_field<span class="op">=</span>DATA_COLUMN,</span>
<span id="cb11-115"><a href="#cb11-115" aria-hidden="true" tabindex="-1"></a> fim_rate<span class="op">=</span>FIM_RATE,</span>
<span id="cb11-116"><a href="#cb11-116" aria-hidden="true" tabindex="-1"></a> fim_spm_rate<span class="op">=</span>FIM_SPM_RATE,</span>
<span id="cb11-117"><a href="#cb11-117" aria-hidden="true" tabindex="-1"></a> seed<span class="op">=</span>SEED,</span>
<span id="cb11-118"><a href="#cb11-118" aria-hidden="true" tabindex="-1"></a>)</span>
<span id="cb11-119"><a href="#cb11-119" aria-hidden="true" tabindex="-1"></a>eval_dataset <span class="op">=</span> ConstantLengthDataset(</span>
<span id="cb11-120"><a href="#cb11-120" aria-hidden="true" tabindex="-1"></a> tokenizer,</span>
<span id="cb11-121"><a href="#cb11-121" aria-hidden="true" tabindex="-1"></a> valid_data,</span>
<span id="cb11-122"><a href="#cb11-122" aria-hidden="true" tabindex="-1"></a> infinite<span class="op">=</span><span class="va">False</span>,</span>
<span id="cb11-123"><a href="#cb11-123" aria-hidden="true" tabindex="-1"></a> seq_length<span class="op">=</span>SEQ_LENGTH,</span>
<span id="cb11-124"><a href="#cb11-124" aria-hidden="true" tabindex="-1"></a> chars_per_token<span class="op">=</span>chars_per_token,</span>
<span id="cb11-125"><a href="#cb11-125" aria-hidden="true" tabindex="-1"></a> content_field<span class="op">=</span>DATA_COLUMN,</span>
<span id="cb11-126"><a href="#cb11-126" aria-hidden="true" tabindex="-1"></a> fim_rate<span class="op">=</span>FIM_RATE,</span>
<span id="cb11-127"><a href="#cb11-127" aria-hidden="true" tabindex="-1"></a> fim_spm_rate<span class="op">=</span>FIM_SPM_RATE,</span>
<span id="cb11-128"><a href="#cb11-128" aria-hidden="true" tabindex="-1"></a> seed<span class="op">=</span>SEED,</span>
<span id="cb11-129"><a href="#cb11-129" aria-hidden="true" tabindex="-1"></a>)</span></code><button title="Copy to Clipboard" class="code-copy-button"><i class="bi"></i></button></pre></div>
</div>
</section>
<section id="prepare-the-model" class="level2">
<h2 class="anchored" data-anchor-id="prepare-the-model">Prepare the model</h2>
<p>Now that the data is prepared, it’s time to load the model! We’re going to load the quantized version of the model.</p>
<p>This will allow us to reduce memory usage, as quantization represents data with fewer bits. We’ll use the <code>bitsandbytes</code> library to quantize the model, as it has a nice integration with <code>transformers</code>. All we need to do is define a <code>bitsandbytes</code> config, and then use it when loading the model.</p>
<p>There are different variants of 4bit quantization, but generally, we recommend using NF4 quantization for better performance (<code>bnb_4bit_quant_type="nf4"</code>).</p>
<p>The <code>bnb_4bit_use_double_quant</code> option adds a second quantization after the first one to save an additional 0.4 bits per parameter.</p>
<p>To learn more about quantization, check out the <a href="https://huggingface.co/blog/4bit-transformers-bitsandbytes">“Making LLMs even more accessible with bitsandbytes, 4-bit quantization and QLoRA” blog post</a>.</p>
<p>Once defined, pass the config to the <code>from_pretrained</code> method to load the quantized version of the model.</p>
<div id="cell-19" class="cell">
<div class="sourceCode cell-code" id="cb12"><pre class="sourceCode python code-with-copy"><code class="sourceCode python"><span id="cb12-1"><a href="#cb12-1" aria-hidden="true" tabindex="-1"></a><span class="im">from</span> peft <span class="im">import</span> LoraConfig, get_peft_model, prepare_model_for_kbit_training</span>
<span id="cb12-2"><a href="#cb12-2" aria-hidden="true" tabindex="-1"></a><span class="im">from</span> peft.tuners.lora <span class="im">import</span> LoraLayer</span>
<span id="cb12-3"><a href="#cb12-3" aria-hidden="true" tabindex="-1"></a></span>
<span id="cb12-4"><a href="#cb12-4" aria-hidden="true" tabindex="-1"></a>load_in_8bit <span class="op">=</span> <span class="va">False</span></span>
<span id="cb12-5"><a href="#cb12-5" aria-hidden="true" tabindex="-1"></a></span>
<span id="cb12-6"><a href="#cb12-6" aria-hidden="true" tabindex="-1"></a><span class="co"># 4-bit quantization</span></span>
<span id="cb12-7"><a href="#cb12-7" aria-hidden="true" tabindex="-1"></a>compute_dtype <span class="op">=</span> <span class="bu">getattr</span>(torch, BNB_4BIT_COMPUTE_DTYPE)</span>
<span id="cb12-8"><a href="#cb12-8" aria-hidden="true" tabindex="-1"></a></span>
<span id="cb12-9"><a href="#cb12-9" aria-hidden="true" tabindex="-1"></a>bnb_config <span class="op">=</span> BitsAndBytesConfig(</span>
<span id="cb12-10"><a href="#cb12-10" aria-hidden="true" tabindex="-1"></a> load_in_4bit<span class="op">=</span><span class="va">True</span>,</span>
<span id="cb12-11"><a href="#cb12-11" aria-hidden="true" tabindex="-1"></a> bnb_4bit_quant_type<span class="op">=</span><span class="st">"nf4"</span>,</span>
<span id="cb12-12"><a href="#cb12-12" aria-hidden="true" tabindex="-1"></a> bnb_4bit_compute_dtype<span class="op">=</span>compute_dtype,</span>
<span id="cb12-13"><a href="#cb12-13" aria-hidden="true" tabindex="-1"></a> bnb_4bit_use_double_quant<span class="op">=</span>USE_NESTED_QUANT,</span>
<span id="cb12-14"><a href="#cb12-14" aria-hidden="true" tabindex="-1"></a>)</span>
<span id="cb12-15"><a href="#cb12-15" aria-hidden="true" tabindex="-1"></a></span>
<span id="cb12-16"><a href="#cb12-16" aria-hidden="true" tabindex="-1"></a>device_map <span class="op">=</span> {<span class="st">""</span>: <span class="dv">0</span>}</span>
<span id="cb12-17"><a href="#cb12-17" aria-hidden="true" tabindex="-1"></a></span>
<span id="cb12-18"><a href="#cb12-18" aria-hidden="true" tabindex="-1"></a>model <span class="op">=</span> AutoModelForCausalLM.from_pretrained(</span>
<span id="cb12-19"><a href="#cb12-19" aria-hidden="true" tabindex="-1"></a> MODEL,</span>
<span id="cb12-20"><a href="#cb12-20" aria-hidden="true" tabindex="-1"></a> load_in_8bit<span class="op">=</span>load_in_8bit,</span>
<span id="cb12-21"><a href="#cb12-21" aria-hidden="true" tabindex="-1"></a> quantization_config<span class="op">=</span>bnb_config,</span>
<span id="cb12-22"><a href="#cb12-22" aria-hidden="true" tabindex="-1"></a> device_map<span class="op">=</span>device_map,</span>
<span id="cb12-23"><a href="#cb12-23" aria-hidden="true" tabindex="-1"></a> use_cache<span class="op">=</span><span class="va">False</span>, <span class="co"># We will be using gradient checkpointing</span></span>
<span id="cb12-24"><a href="#cb12-24" aria-hidden="true" tabindex="-1"></a> trust_remote_code<span class="op">=</span><span class="va">True</span>,</span>
<span id="cb12-25"><a href="#cb12-25" aria-hidden="true" tabindex="-1"></a> use_flash_attention_2<span class="op">=</span><span class="va">True</span>,</span>
<span id="cb12-26"><a href="#cb12-26" aria-hidden="true" tabindex="-1"></a>)</span></code><button title="Copy to Clipboard" class="code-copy-button"><i class="bi"></i></button></pre></div>
</div>
<p>When using a quantized model for training, you need to call the <code>prepare_model_for_kbit_training()</code> function to preprocess the quantized model for training.</p>
<div id="cell-21" class="cell">
<div class="sourceCode cell-code" id="cb13"><pre class="sourceCode python code-with-copy"><code class="sourceCode python"><span id="cb13-1"><a href="#cb13-1" aria-hidden="true" tabindex="-1"></a>model <span class="op">=</span> prepare_model_for_kbit_training(model)</span></code><button title="Copy to Clipboard" class="code-copy-button"><i class="bi"></i></button></pre></div>
</div>
<p>Now that the quantized model is ready, we can set up a LoRA configuration. LoRA makes fine-tuning more efficient by drastically reducing the number of trainable parameters.</p>
<p>To train a model using LoRA technique, we need to wrap the base model as a <code>PeftModel</code>. This involves definign LoRA configuration with <code>LoraConfig</code>, and wrapping the original model with <code>get_peft_model()</code> using the <code>LoraConfig</code>.</p>
<p>To learn more about LoRA and its parameters, refer to <a href="https://huggingface.co/docs/peft/conceptual_guides/lora">PEFT documentation</a>.</p>
<div id="cell-23" class="cell" data-outputid="63328c2b-e693-49b1-ce0a-3ca8722f852a">
<div class="sourceCode cell-code" id="cb14"><pre class="sourceCode python code-with-copy"><code class="sourceCode python"><span id="cb14-1"><a href="#cb14-1" aria-hidden="true" tabindex="-1"></a><span class="co"># Set up lora</span></span>
<span id="cb14-2"><a href="#cb14-2" aria-hidden="true" tabindex="-1"></a>peft_config <span class="op">=</span> LoraConfig(</span>
<span id="cb14-3"><a href="#cb14-3" aria-hidden="true" tabindex="-1"></a> lora_alpha<span class="op">=</span>LORA_ALPHA,</span>
<span id="cb14-4"><a href="#cb14-4" aria-hidden="true" tabindex="-1"></a> lora_dropout<span class="op">=</span>LORA_DROPOUT,</span>
<span id="cb14-5"><a href="#cb14-5" aria-hidden="true" tabindex="-1"></a> r<span class="op">=</span>LORA_R,</span>
<span id="cb14-6"><a href="#cb14-6" aria-hidden="true" tabindex="-1"></a> bias<span class="op">=</span><span class="st">"none"</span>,</span>
<span id="cb14-7"><a href="#cb14-7" aria-hidden="true" tabindex="-1"></a> task_type<span class="op">=</span><span class="st">"CAUSAL_LM"</span>,</span>
<span id="cb14-8"><a href="#cb14-8" aria-hidden="true" tabindex="-1"></a> target_modules<span class="op">=</span>LORA_TARGET_MODULES.split(<span class="st">","</span>),</span>
<span id="cb14-9"><a href="#cb14-9" aria-hidden="true" tabindex="-1"></a>)</span>
<span id="cb14-10"><a href="#cb14-10" aria-hidden="true" tabindex="-1"></a></span>
<span id="cb14-11"><a href="#cb14-11" aria-hidden="true" tabindex="-1"></a>model <span class="op">=</span> get_peft_model(model, peft_config)</span>
<span id="cb14-12"><a href="#cb14-12" aria-hidden="true" tabindex="-1"></a>model.print_trainable_parameters()</span></code><button title="Copy to Clipboard" class="code-copy-button"><i class="bi"></i></button></pre></div>
<div class="cell-output cell-output-stdout">
<pre><code>trainable params: 5,554,176 || all params: 1,142,761,472 || trainable%: 0.4860310866343243</code></pre>
</div>
</div>
<p>As you can see, by applying LoRA technique we will now need to train less than 1% of the parameters.</p>
</section>
<section id="train-the-model" class="level2">
<h2 class="anchored" data-anchor-id="train-the-model">Train the model</h2>
<p>Now that we have prepared the data, and optimized the model, we are ready to bring everything together to start the training.</p>
<p>To instantiate a <code>Trainer</code>, you need to define the training configuration. The most important is the <code>TrainingArguments</code>, which is a class that contains all the attributes to configure the training.</p>
<p>These are similar to any other kind of model training you may run, so we won’t go into detail here.</p>
<div id="cell-27" class="cell">
<div class="sourceCode cell-code" id="cb16"><pre class="sourceCode python code-with-copy"><code class="sourceCode python"><span id="cb16-1"><a href="#cb16-1" aria-hidden="true" tabindex="-1"></a>train_data.start_iteration <span class="op">=</span> <span class="dv">0</span></span>
<span id="cb16-2"><a href="#cb16-2" aria-hidden="true" tabindex="-1"></a></span>
<span id="cb16-3"><a href="#cb16-3" aria-hidden="true" tabindex="-1"></a></span>
<span id="cb16-4"><a href="#cb16-4" aria-hidden="true" tabindex="-1"></a>training_args <span class="op">=</span> TrainingArguments(</span>
<span id="cb16-5"><a href="#cb16-5" aria-hidden="true" tabindex="-1"></a> output_dir<span class="op">=</span><span class="ss">f"Your_HF_username/</span><span class="sc">{</span>OUTPUT_DIR<span class="sc">}</span><span class="ss">"</span>,</span>
<span id="cb16-6"><a href="#cb16-6" aria-hidden="true" tabindex="-1"></a> dataloader_drop_last<span class="op">=</span><span class="va">True</span>,</span>
<span id="cb16-7"><a href="#cb16-7" aria-hidden="true" tabindex="-1"></a> evaluation_strategy<span class="op">=</span><span class="st">"steps"</span>,</span>
<span id="cb16-8"><a href="#cb16-8" aria-hidden="true" tabindex="-1"></a> save_strategy<span class="op">=</span><span class="st">"steps"</span>,</span>
<span id="cb16-9"><a href="#cb16-9" aria-hidden="true" tabindex="-1"></a> max_steps<span class="op">=</span>MAX_STEPS,</span>
<span id="cb16-10"><a href="#cb16-10" aria-hidden="true" tabindex="-1"></a> eval_steps<span class="op">=</span>EVAL_FREQ,</span>
<span id="cb16-11"><a href="#cb16-11" aria-hidden="true" tabindex="-1"></a> save_steps<span class="op">=</span>SAVE_FREQ,</span>
<span id="cb16-12"><a href="#cb16-12" aria-hidden="true" tabindex="-1"></a> logging_steps<span class="op">=</span>LOG_FREQ,</span>
<span id="cb16-13"><a href="#cb16-13" aria-hidden="true" tabindex="-1"></a> per_device_train_batch_size<span class="op">=</span>BATCH_SIZE,</span>
<span id="cb16-14"><a href="#cb16-14" aria-hidden="true" tabindex="-1"></a> per_device_eval_batch_size<span class="op">=</span>BATCH_SIZE,</span>
<span id="cb16-15"><a href="#cb16-15" aria-hidden="true" tabindex="-1"></a> learning_rate<span class="op">=</span>LR,</span>
<span id="cb16-16"><a href="#cb16-16" aria-hidden="true" tabindex="-1"></a> lr_scheduler_type<span class="op">=</span>LR_SCHEDULER_TYPE,</span>
<span id="cb16-17"><a href="#cb16-17" aria-hidden="true" tabindex="-1"></a> warmup_steps<span class="op">=</span>NUM_WARMUP_STEPS,</span>
<span id="cb16-18"><a href="#cb16-18" aria-hidden="true" tabindex="-1"></a> gradient_accumulation_steps<span class="op">=</span>GR_ACC_STEPS,</span>
<span id="cb16-19"><a href="#cb16-19" aria-hidden="true" tabindex="-1"></a> gradient_checkpointing<span class="op">=</span><span class="va">True</span>,</span>
<span id="cb16-20"><a href="#cb16-20" aria-hidden="true" tabindex="-1"></a> fp16<span class="op">=</span>FP16,</span>
<span id="cb16-21"><a href="#cb16-21" aria-hidden="true" tabindex="-1"></a> bf16<span class="op">=</span>BF16,</span>
<span id="cb16-22"><a href="#cb16-22" aria-hidden="true" tabindex="-1"></a> weight_decay<span class="op">=</span>WEIGHT_DECAY,</span>
<span id="cb16-23"><a href="#cb16-23" aria-hidden="true" tabindex="-1"></a> push_to_hub<span class="op">=</span><span class="va">True</span>,</span>
<span id="cb16-24"><a href="#cb16-24" aria-hidden="true" tabindex="-1"></a> include_tokens_per_second<span class="op">=</span><span class="va">True</span>,</span>
<span id="cb16-25"><a href="#cb16-25" aria-hidden="true" tabindex="-1"></a>)</span></code><button title="Copy to Clipboard" class="code-copy-button"><i class="bi"></i></button></pre></div>
</div>
<p>As a final step, instantiate the <code>Trainer</code> and call the <code>train</code> method.</p>
<div id="cell-29" class="cell" data-outputid="61a5bdb2-b7d0-4aed-8290-4bf20c2ccd38">
<div class="sourceCode cell-code" id="cb17"><pre class="sourceCode python code-with-copy"><code class="sourceCode python"><span id="cb17-1"><a href="#cb17-1" aria-hidden="true" tabindex="-1"></a>trainer <span class="op">=</span> Trainer(</span>
<span id="cb17-2"><a href="#cb17-2" aria-hidden="true" tabindex="-1"></a> model<span class="op">=</span>model, args<span class="op">=</span>training_args, train_dataset<span class="op">=</span>train_dataset, eval_dataset<span class="op">=</span>eval_dataset</span>
<span id="cb17-3"><a href="#cb17-3" aria-hidden="true" tabindex="-1"></a>)</span>
<span id="cb17-4"><a href="#cb17-4" aria-hidden="true" tabindex="-1"></a></span>
<span id="cb17-5"><a href="#cb17-5" aria-hidden="true" tabindex="-1"></a><span class="bu">print</span>(<span class="st">"Training..."</span>)</span>
<span id="cb17-6"><a href="#cb17-6" aria-hidden="true" tabindex="-1"></a>trainer.train()</span></code><button title="Copy to Clipboard" class="code-copy-button"><i class="bi"></i></button></pre></div>
<div class="cell-output cell-output-stdout">
<pre><code>Training...</code></pre>
</div>
<div class="cell-output cell-output-display">
<div>
<progress value="2000" max="2000" style="width:300px; height:20px; vertical-align: middle;"></progress>
[2000/2000 4:16:10, Epoch 1/9223372036854775807]
</div>
<table class="dataframe caption-top table table-sm table-striped small" data-quarto-postprocess="true" data-border="1">
<thead>
<tr class="header">
<th data-quarto-table-cell-role="th">Step</th>
<th data-quarto-table-cell-role="th">Training Loss</th>
<th data-quarto-table-cell-role="th">Validation Loss</th>
</tr>
</thead>
<tbody>
<tr class="odd">
<td>100</td>
<td>5.524600</td>
<td>7.456872</td>
</tr>
<tr class="even">
<td>200</td>
<td>5.617800</td>
<td>7.262190</td>
</tr>
<tr class="odd">
<td>300</td>
<td>5.129100</td>
<td>6.410039</td>
</tr>
<tr class="even">
<td>400</td>
<td>5.052200</td>
<td>6.306774</td>
</tr>
<tr class="odd">
<td>500</td>
<td>5.202900</td>
<td>6.117062</td>
</tr>
<tr class="even">
<td>600</td>
<td>4.654100</td>
<td>6.018349</td>
</tr>
<tr class="odd">
<td>700</td>
<td>5.100200</td>
<td>6.000355</td>
</tr>
<tr class="even">
<td>800</td>
<td>5.049800</td>
<td>5.889457</td>
</tr>
<tr class="odd">
<td>900</td>
<td>4.541200</td>
<td>5.813823</td>
</tr>
<tr class="even">
<td>1000</td>
<td>5.000700</td>
<td>5.834208</td>
</tr>
<tr class="odd">
<td>1100</td>
<td>5.026500</td>
<td>5.781939</td>
</tr>
<tr class="even">
<td>1200</td>
<td>4.411800</td>
<td>5.720596</td>
</tr>
<tr class="odd">
<td>1300</td>
<td>4.782500</td>
<td>5.736376</td>
</tr>
<tr class="even">
<td>1400</td>
<td>4.980200</td>
<td>5.712276</td>
</tr>
<tr class="odd">
<td>1500</td>
<td>4.368700</td>
<td>5.689637</td>
</tr>
<tr class="even">
<td>1600</td>
<td>4.884700</td>
<td>5.675920</td>
</tr>
<tr class="odd">
<td>1700</td>
<td>4.914400</td>
<td>5.662421</td>
</tr>
<tr class="even">
<td>1800</td>
<td>4.248700</td>
<td>5.660122</td>
</tr>
<tr class="odd">
<td>1900</td>
<td>4.798400</td>
<td>5.664026</td>
</tr>
<tr class="even">
<td>2000</td>
<td>4.704200</td>
<td>5.655665</td>
</tr>
</tbody>
</table>
<p>
</p></div>
<div class="cell-output cell-output-display" data-execution_count="19">
<pre><code>TrainOutput(global_step=2000, training_loss=4.885598585128784, metrics={'train_runtime': 15380.3075, 'train_samples_per_second': 2.081, 'train_steps_per_second': 0.13, 'train_tokens_per_second': 4261.033, 'total_flos': 4.0317260660736e+17, 'train_loss': 4.885598585128784, 'epoch': 1.0})</code></pre>
</div>
</div>
<p>Finally, you can push the fine-tuned model to your Hub repository to share with your team.</p>
<div id="cell-31" class="cell">
<div class="sourceCode cell-code" id="cb20"><pre class="sourceCode python code-with-copy"><code class="sourceCode python"><span id="cb20-1"><a href="#cb20-1" aria-hidden="true" tabindex="-1"></a>trainer.push_to_hub()</span></code><button title="Copy to Clipboard" class="code-copy-button"><i class="bi"></i></button></pre></div>
</div>
</section>
<section id="inference" class="level2">
<h2 class="anchored" data-anchor-id="inference">Inference</h2>
<p>Once the model is uploaded to Hub, we can use it for inference. To do so we first initialize the original base model and its tokenizer. Next, we need to merge the fine-duned weights with the base model.</p>
<div id="cell-33" class="cell">
<div class="sourceCode cell-code" id="cb21"><pre class="sourceCode python code-with-copy"><code class="sourceCode python"><span id="cb21-1"><a href="#cb21-1" aria-hidden="true" tabindex="-1"></a><span class="im">from</span> peft <span class="im">import</span> PeftModel</span>
<span id="cb21-2"><a href="#cb21-2" aria-hidden="true" tabindex="-1"></a><span class="im">import</span> torch</span>
<span id="cb21-3"><a href="#cb21-3" aria-hidden="true" tabindex="-1"></a></span>
<span id="cb21-4"><a href="#cb21-4" aria-hidden="true" tabindex="-1"></a><span class="co"># load the original model first</span></span>
<span id="cb21-5"><a href="#cb21-5" aria-hidden="true" tabindex="-1"></a>tokenizer <span class="op">=</span> AutoTokenizer.from_pretrained(MODEL, trust_remote_code<span class="op">=</span><span class="va">True</span>)</span>
<span id="cb21-6"><a href="#cb21-6" aria-hidden="true" tabindex="-1"></a>base_model <span class="op">=</span> AutoModelForCausalLM.from_pretrained(</span>
<span id="cb21-7"><a href="#cb21-7" aria-hidden="true" tabindex="-1"></a> MODEL,</span>
<span id="cb21-8"><a href="#cb21-8" aria-hidden="true" tabindex="-1"></a> quantization_config<span class="op">=</span><span class="va">None</span>,</span>
<span id="cb21-9"><a href="#cb21-9" aria-hidden="true" tabindex="-1"></a> device_map<span class="op">=</span><span class="va">None</span>,</span>
<span id="cb21-10"><a href="#cb21-10" aria-hidden="true" tabindex="-1"></a> trust_remote_code<span class="op">=</span><span class="va">True</span>,</span>
<span id="cb21-11"><a href="#cb21-11" aria-hidden="true" tabindex="-1"></a> torch_dtype<span class="op">=</span>torch.bfloat16,</span>
<span id="cb21-12"><a href="#cb21-12" aria-hidden="true" tabindex="-1"></a>).cuda()</span>
<span id="cb21-13"><a href="#cb21-13" aria-hidden="true" tabindex="-1"></a></span>
<span id="cb21-14"><a href="#cb21-14" aria-hidden="true" tabindex="-1"></a><span class="co"># merge fine-tuned weights with the base model</span></span>
<span id="cb21-15"><a href="#cb21-15" aria-hidden="true" tabindex="-1"></a>peft_model_id <span class="op">=</span> <span class="ss">f"Your_HF_username/</span><span class="sc">{</span>OUTPUT_DIR<span class="sc">}</span><span class="ss">"</span></span>
<span id="cb21-16"><a href="#cb21-16" aria-hidden="true" tabindex="-1"></a>model <span class="op">=</span> PeftModel.from_pretrained(base_model, peft_model_id)</span>
<span id="cb21-17"><a href="#cb21-17" aria-hidden="true" tabindex="-1"></a>model.merge_and_unload()</span></code><button title="Copy to Clipboard" class="code-copy-button"><i class="bi"></i></button></pre></div>
</div>
<p>Now we can use the merged model for inference. For convenience, we’ll define a <code>get_code_completion</code> - feel free to experiment with text generation parameters!</p>
<div id="cell-35" class="cell">
<div class="sourceCode cell-code" id="cb22"><pre class="sourceCode python code-with-copy"><code class="sourceCode python"><span id="cb22-1"><a href="#cb22-1" aria-hidden="true" tabindex="-1"></a><span class="kw">def</span> get_code_completion(prefix, suffix):</span>
<span id="cb22-2"><a href="#cb22-2" aria-hidden="true" tabindex="-1"></a> text <span class="op">=</span> prompt <span class="op">=</span> <span class="ss">f"""&lt;fim_prefix&gt;</span><span class="sc">{</span>prefix<span class="sc">}</span><span class="ss">&lt;fim_suffix&gt;</span><span class="sc">{</span>suffix<span class="sc">}</span><span class="ss">&lt;fim_middle&gt;"""</span></span>
<span id="cb22-3"><a href="#cb22-3" aria-hidden="true" tabindex="-1"></a> model.<span class="bu">eval</span>()</span>
<span id="cb22-4"><a href="#cb22-4" aria-hidden="true" tabindex="-1"></a> outputs <span class="op">=</span> model.generate(</span>
<span id="cb22-5"><a href="#cb22-5" aria-hidden="true" tabindex="-1"></a> input_ids<span class="op">=</span>tokenizer(text, return_tensors<span class="op">=</span><span class="st">"pt"</span>).input_ids.cuda(),</span>
<span id="cb22-6"><a href="#cb22-6" aria-hidden="true" tabindex="-1"></a> max_new_tokens<span class="op">=</span><span class="dv">128</span>,</span>
<span id="cb22-7"><a href="#cb22-7" aria-hidden="true" tabindex="-1"></a> temperature<span class="op">=</span><span class="fl">0.2</span>,</span>
<span id="cb22-8"><a href="#cb22-8" aria-hidden="true" tabindex="-1"></a> top_k<span class="op">=</span><span class="dv">50</span>,</span>
<span id="cb22-9"><a href="#cb22-9" aria-hidden="true" tabindex="-1"></a> top_p<span class="op">=</span><span class="fl">0.95</span>,</span>
<span id="cb22-10"><a href="#cb22-10" aria-hidden="true" tabindex="-1"></a> do_sample<span class="op">=</span><span class="va">True</span>,</span>
<span id="cb22-11"><a href="#cb22-11" aria-hidden="true" tabindex="-1"></a> repetition_penalty<span class="op">=</span><span class="fl">1.0</span>,</span>
<span id="cb22-12"><a href="#cb22-12" aria-hidden="true" tabindex="-1"></a> )</span>
<span id="cb22-13"><a href="#cb22-13" aria-hidden="true" tabindex="-1"></a> <span class="cf">return</span> tokenizer.batch_decode(outputs, skip_special_tokens<span class="op">=</span><span class="va">True</span>)[<span class="dv">0</span>]</span></code><button title="Copy to Clipboard" class="code-copy-button"><i class="bi"></i></button></pre></div>
</div>
<p>Now all we need to do to get code completion is call the <code>get_code_complete</code> function and pass the first few lines that we want to be completed as a prefix, and an empty string as a suffix.</p>
<div id="cell-37" class="cell" data-outputid="41c411ad-b7dc-4277-f975-c173888234bb">
<div class="sourceCode cell-code" id="cb23"><pre class="sourceCode python code-with-copy"><code class="sourceCode python"><span id="cb23-1"><a href="#cb23-1" aria-hidden="true" tabindex="-1"></a>prefix <span class="op">=</span> <span class="st">"""from peft import LoraConfig, TaskType, get_peft_model</span></span>
<span id="cb23-2"><a href="#cb23-2" aria-hidden="true" tabindex="-1"></a><span class="st">from transformers import AutoModelForCausalLM</span></span>
<span id="cb23-3"><a href="#cb23-3" aria-hidden="true" tabindex="-1"></a><span class="st">peft_config = LoraConfig(</span></span>
<span id="cb23-4"><a href="#cb23-4" aria-hidden="true" tabindex="-1"></a><span class="st">"""</span></span>
<span id="cb23-5"><a href="#cb23-5" aria-hidden="true" tabindex="-1"></a>suffix <span class="op">=</span><span class="st">""""""</span></span>
<span id="cb23-6"><a href="#cb23-6" aria-hidden="true" tabindex="-1"></a></span>
<span id="cb23-7"><a href="#cb23-7" aria-hidden="true" tabindex="-1"></a><span class="bu">print</span>(get_code_completion(prefix, suffix))</span></code><button title="Copy to Clipboard" class="code-copy-button"><i class="bi"></i></button></pre></div>
<div class="cell-output cell-output-stdout">
<pre><code>from peft import LoraConfig, TaskType, get_peft_model
from transformers import AutoModelForCausalLM
peft_config = LoraConfig(
task_type=TaskType.CAUSAL_LM,
r=8,
lora_alpha=32,
target_modules=["q_proj", "v_proj"],
lora_dropout=0.1,
bias="none",
modules_to_save=["q_proj", "v_proj"],
inference_mode=False,
)
model = AutoModelForCausalLM.from_pretrained("gpt2")
model = get_peft_model(model, peft_config)
model.print_trainable_parameters()</code></pre>
</div>
</div>
<p>As someone who has just used the PEFT library earlier in this notebook, you can see that the generated result for creating a <code>LoraConfig</code> is rather good!</p>
<p>If you go back to the cell where we instantiate the model for inference, and comment out the lines where we merge the fine-tuned weights, you can see what the original model would’ve generated for the exact same prefix:</p>
<div id="cell-39" class="cell" data-outputid="c6d597a2-01da-4d25-a32f-3a551212c5b4">
<div class="sourceCode cell-code" id="cb25"><pre class="sourceCode python code-with-copy"><code class="sourceCode python"><span id="cb25-1"><a href="#cb25-1" aria-hidden="true" tabindex="-1"></a>prefix <span class="op">=</span> <span class="st">"""from peft import LoraConfig, TaskType, get_peft_model</span></span>
<span id="cb25-2"><a href="#cb25-2" aria-hidden="true" tabindex="-1"></a><span class="st">from transformers import AutoModelForCausalLM</span></span>
<span id="cb25-3"><a href="#cb25-3" aria-hidden="true" tabindex="-1"></a><span class="st">peft_config = LoraConfig(</span></span>
<span id="cb25-4"><a href="#cb25-4" aria-hidden="true" tabindex="-1"></a><span class="st">"""</span></span>
<span id="cb25-5"><a href="#cb25-5" aria-hidden="true" tabindex="-1"></a>suffix <span class="op">=</span><span class="st">""""""</span></span>
<span id="cb25-6"><a href="#cb25-6" aria-hidden="true" tabindex="-1"></a></span>
<span id="cb25-7"><a href="#cb25-7" aria-hidden="true" tabindex="-1"></a><span class="bu">print</span>(get_code_completion(prefix, suffix))</span></code><button title="Copy to Clipboard" class="code-copy-button"><i class="bi"></i></button></pre></div>
<div class="cell-output cell-output-stdout">
<pre><code>from peft import LoraConfig, TaskType, get_peft_model
from transformers import AutoModelForCausalLM
peft_config = LoraConfig(
model_name_or_path="facebook/wav2vec2-base-960h",
num_labels=1,
num_features=1,
num_hidden_layers=1,
num_attention_heads=1,
num_hidden_layers_per_attention_head=1,
num_attention_heads_per_hidden_layer=1,
hidden_size=1024,
hidden_dropout_prob=0.1,
hidden_act="gelu",
hidden_act_dropout_prob=0.1,
hidden</code></pre>
</div>
</div>
<p>While it is Python syntax, you can see that the original model has no understanding of what a <code>LoraConfig</code> should be doing.</p>
<p>To learn how this kind of fine-tuning compares to full fine-tuning, and how to use a model like this as your copilot in VS Code via Inference Endpoints, or locally, check out the <a href="https://huggingface.co/blog/personal-copilot">“Personal Copilot: Train Your Own Coding Assistant” blog post</a>. This notebook complements the original blog post.</p>
</section>
</section>
</main> <!-- /main -->
<script id="quarto-html-after-body" type="application/javascript">
window.document.addEventListener("DOMContentLoaded", function (event) {
const toggleBodyColorMode = (bsSheetEl) => {
const mode = bsSheetEl.getAttribute("data-mode");
const bodyEl = window.document.querySelector("body");
if (mode === "dark") {
bodyEl.classList.add("quarto-dark");
bodyEl.classList.remove("quarto-light");
} else {
bodyEl.classList.add("quarto-light");
bodyEl.classList.remove("quarto-dark");
}
}
const toggleBodyColorPrimary = () => {
const bsSheetEl = window.document.querySelector("link#quarto-bootstrap");
if (bsSheetEl) {
toggleBodyColorMode(bsSheetEl);
}
}
toggleBodyColorPrimary();
const icon = "";
const anchorJS = new window.AnchorJS();
anchorJS.options = {
placement: 'right',
icon: icon
};
anchorJS.add('.anchored');
const isCodeAnnotation = (el) => {
for (const clz of el.classList) {
if (clz.startsWith('code-annotation-')) {
return true;
}
}
return false;
}
const onCopySuccess = function(e) {
// button target
const button = e.trigger;
// don't keep focus
button.blur();
// flash "checked"
button.classList.add('code-copy-button-checked');
var currentTitle = button.getAttribute("title");
button.setAttribute("title", "Copied!");
let tooltip;
if (window.bootstrap) {
button.setAttribute("data-bs-toggle", "tooltip");
button.setAttribute("data-bs-placement", "left");
button.setAttribute("data-bs-title", "Copied!");
tooltip = new bootstrap.Tooltip(button,
{ trigger: "manual",
customClass: "code-copy-button-tooltip",
offset: [0, -8]});
tooltip.show();
}
setTimeout(function() {
if (tooltip) {
tooltip.hide();
button.removeAttribute("data-bs-title");
button.removeAttribute("data-bs-toggle");
button.removeAttribute("data-bs-placement");
}
button.setAttribute("title", currentTitle);
button.classList.remove('code-copy-button-checked');
}, 1000);
// clear code selection
e.clearSelection();
}
const getTextToCopy = function(trigger) {
const codeEl = trigger.previousElementSibling.cloneNode(true);
for (const childEl of codeEl.children) {
if (isCodeAnnotation(childEl)) {
childEl.remove();
}
}
return codeEl.innerText;
}
const clipboard = new window.ClipboardJS('.code-copy-button:not([data-in-quarto-modal])', {
text: getTextToCopy
});
clipboard.on('success', onCopySuccess);
if (window.document.getElementById('quarto-embedded-source-code-modal')) {
const clipboardModal = new window.ClipboardJS('.code-copy-button[data-in-quarto-modal]', {
text: getTextToCopy,
container: window.document.getElementById('quarto-embedded-source-code-modal')
});
clipboardModal.on('success', onCopySuccess);
}
var localhostRegex = new RegExp(/^(?:http|https):\/\/localhost\:?[0-9]*\//);
var mailtoRegex = new RegExp(/^mailto:/);
var filterRegex = new RegExp('/' + window.location.host + '/');
var isInternal = (href) => {
return filterRegex.test(href) || localhostRegex.test(href) || mailtoRegex.test(href);
}
// Inspect non-navigation links and adorn them if external
var links = window.document.querySelectorAll('a[href]:not(.nav-link):not(.navbar-brand):not(.toc-action):not(.sidebar-link):not(.sidebar-item-toggle):not(.pagination-link):not(.no-external):not([aria-hidden]):not(.dropdown-item):not(.quarto-navigation-tool):not(.about-link)');
for (var i=0; i<links.length; i++) {
const link = links[i];
if (!isInternal(link.href)) {
// undo the damage that might have been done by quarto-nav.js in the case of
// links that we want to consider external
if (link.dataset.originalHref !== undefined) {
link.href = link.dataset.originalHref;
}
}
}
function tippyHover(el, contentFn, onTriggerFn, onUntriggerFn) {
const config = {
allowHTML: true,
maxWidth: 500,
delay: 100,
arrow: false,
appendTo: function(el) {
return el.parentElement;
},
interactive: true,
interactiveBorder: 10,
theme: 'quarto',
placement: 'bottom-start',
};
if (contentFn) {
config.content = contentFn;
}
if (onTriggerFn) {
config.onTrigger = onTriggerFn;
}
if (onUntriggerFn) {
config.onUntrigger = onUntriggerFn;
}
window.tippy(el, config);
}
const noterefs = window.document.querySelectorAll('a[role="doc-noteref"]');
for (var i=0; i<noterefs.length; i++) {
const ref = noterefs[i];
tippyHover(ref, function() {
// use id or data attribute instead here
let href = ref.getAttribute('data-footnote-href') || ref.getAttribute('href');
try { href = new URL(href).hash; } catch {}
const id = href.replace(/^#\/?/, "");
const note = window.document.getElementById(id);
if (note) {
return note.innerHTML;
} else {
return "";
}
});
}
const xrefs = window.document.querySelectorAll('a.quarto-xref');
const processXRef = (id, note) => {
// Strip column container classes
const stripColumnClz = (el) => {
el.classList.remove("page-full", "page-columns");
if (el.children) {
for (const child of el.children) {
stripColumnClz(child);
}
}
}
stripColumnClz(note)
if (id === null || id.startsWith('sec-')) {
// Special case sections, only their first couple elements
const container = document.createElement("div");
if (note.children && note.children.length > 2) {
container.appendChild(note.children[0].cloneNode(true));
for (let i = 1; i < note.children.length; i++) {
const child = note.children[i];
if (child.tagName === "P" && child.innerText === "") {
continue;
} else {
container.appendChild(child.cloneNode(true));
break;
}
}
if (window.Quarto?.typesetMath) {
window.Quarto.typesetMath(container);
}
return container.innerHTML
} else {
if (window.Quarto?.typesetMath) {
window.Quarto.typesetMath(note);
}
return note.innerHTML;
}
} else {
// Remove any anchor links if they are present
const anchorLink = note.querySelector('a.anchorjs-link');
if (anchorLink) {
anchorLink.remove();
}
if (window.Quarto?.typesetMath) {
window.Quarto.typesetMath(note);
}
if (note.classList.contains("callout")) {
return note.outerHTML;
} else {
return note.innerHTML;
}
}
}
for (var i=0; i<xrefs.length; i++) {
const xref = xrefs[i];
tippyHover(xref, undefined, function(instance) {
instance.disable();
let url = xref.getAttribute('href');
let hash = undefined;
if (url.startsWith('#')) {
hash = url;
} else {
try { hash = new URL(url).hash; } catch {}
}
if (hash) {
const id = hash.replace(/^#\/?/, "");
const note = window.document.getElementById(id);
if (note !== null) {
try {
const html = processXRef(id, note.cloneNode(true));
instance.setContent(html);
} finally {
instance.enable();
instance.show();
}
} else {
// See if we can fetch this
fetch(url.split('#')[0])
.then(res => res.text())
.then(html => {
const parser = new DOMParser();
const htmlDoc = parser.parseFromString(html, "text/html");
const note = htmlDoc.getElementById(id);
if (note !== null) {
const html = processXRef(id, note);
instance.setContent(html);
}
}).finally(() => {
instance.enable();
instance.show();
});
}
} else {
// See if we can fetch a full url (with no hash to target)
// This is a special case and we should probably do some content thinning / targeting
fetch(url)
.then(res => res.text())
.then(html => {
const parser = new DOMParser();
const htmlDoc = parser.parseFromString(html, "text/html");
const note = htmlDoc.querySelector('main.content');
if (note !== null) {
// This should only happen for chapter cross references
// (since there is no id in the URL)
// remove the first header
if (note.children.length > 0 && note.children[0].tagName === "HEADER") {
note.children[0].remove();
}
const html = processXRef(null, note);
instance.setContent(html);
}
}).finally(() => {
instance.enable();
instance.show();
});
}
}, function(instance) {
});
}
let selectedAnnoteEl;
const selectorForAnnotation = ( cell, annotation) => {
let cellAttr = 'data-code-cell="' + cell + '"';
let lineAttr = 'data-code-annotation="' + annotation + '"';
const selector = 'span[' + cellAttr + '][' + lineAttr + ']';
return selector;
}
const selectCodeLines = (annoteEl) => {
const doc = window.document;
const targetCell = annoteEl.getAttribute("data-target-cell");
const targetAnnotation = annoteEl.getAttribute("data-target-annotation");
const annoteSpan = window.document.querySelector(selectorForAnnotation(targetCell, targetAnnotation));
const lines = annoteSpan.getAttribute("data-code-lines").split(",");
const lineIds = lines.map((line) => {
return targetCell + "-" + line;
})
let top = null;
let height = null;
let parent = null;
if (lineIds.length > 0) {
//compute the position of the single el (top and bottom and make a div)
const el = window.document.getElementById(lineIds[0]);
top = el.offsetTop;
height = el.offsetHeight;
parent = el.parentElement.parentElement;
if (lineIds.length > 1) {
const lastEl = window.document.getElementById(lineIds[lineIds.length - 1]);
const bottom = lastEl.offsetTop + lastEl.offsetHeight;
height = bottom - top;
}
if (top !== null && height !== null && parent !== null) {
// cook up a div (if necessary) and position it
let div = window.document.getElementById("code-annotation-line-highlight");
if (div === null) {
div = window.document.createElement("div");
div.setAttribute("id", "code-annotation-line-highlight");
div.style.position = 'absolute';
parent.appendChild(div);
}
div.style.top = top - 2 + "px";
div.style.height = height + 4 + "px";
div.style.left = 0;
let gutterDiv = window.document.getElementById("code-annotation-line-highlight-gutter");
if (gutterDiv === null) {
gutterDiv = window.document.createElement("div");
gutterDiv.setAttribute("id", "code-annotation-line-highlight-gutter");
gutterDiv.style.position = 'absolute';
const codeCell = window.document.getElementById(targetCell);
const gutter = codeCell.querySelector('.code-annotation-gutter');
gutter.appendChild(gutterDiv);
}
gutterDiv.style.top = top - 2 + "px";
gutterDiv.style.height = height + 4 + "px";
}
selectedAnnoteEl = annoteEl;
}
};
const unselectCodeLines = () => {
const elementsIds = ["code-annotation-line-highlight", "code-annotation-line-highlight-gutter"];
elementsIds.forEach((elId) => {
const div = window.document.getElementById(elId);
if (div) {
div.remove();
}
});
selectedAnnoteEl = undefined;
};
// Handle positioning of the toggle
window.addEventListener(
"resize",
throttle(() => {
elRect = undefined;
if (selectedAnnoteEl) {
selectCodeLines(selectedAnnoteEl);
}
}, 10)
);
function throttle(fn, ms) {
let throttle = false;
let timer;
return (...args) => {
if(!throttle) { // first call gets through
fn.apply(this, args);
throttle = true;
} else { // all the others get throttled
if(timer) clearTimeout(timer); // cancel #2
timer = setTimeout(() => {
fn.apply(this, args);
timer = throttle = false;
}, ms);
}
};
}
// Attach click handler to the DT
const annoteDls = window.document.querySelectorAll('dt[data-target-cell]');
for (const annoteDlNode of annoteDls) {
annoteDlNode.addEventListener('click', (event) => {
const clickedEl = event.target;
if (clickedEl !== selectedAnnoteEl) {
unselectCodeLines();
const activeEl = window.document.querySelector('dt[data-target-cell].code-annotation-active');
if (activeEl) {
activeEl.classList.remove('code-annotation-active');
}
selectCodeLines(clickedEl);
clickedEl.classList.add('code-annotation-active');
} else {
// Unselect the line
unselectCodeLines();
clickedEl.classList.remove('code-annotation-active');
}
});
}
const findCites = (el) => {
const parentEl = el.parentElement;
if (parentEl) {
const cites = parentEl.dataset.cites;
if (cites) {
return {
el,
cites: cites.split(' ')
};
} else {
return findCites(el.parentElement)
}
} else {
return undefined;
}
};
var bibliorefs = window.document.querySelectorAll('a[role="doc-biblioref"]');
for (var i=0; i<bibliorefs.length; i++) {
const ref = bibliorefs[i];
const citeInfo = findCites(ref);
if (citeInfo) {
tippyHover(citeInfo.el, function() {
var popup = window.document.createElement('div');
citeInfo.cites.forEach(function(cite) {
var citeDiv = window.document.createElement('div');
citeDiv.classList.add('hanging-indent');
citeDiv.classList.add('csl-entry');
var biblioDiv = window.document.getElementById('ref-' + cite);
if (biblioDiv) {
citeDiv.innerHTML = biblioDiv.innerHTML;
}
popup.appendChild(citeDiv);
});
return popup.innerHTML;
});
}
}
});
</script>
</div> <!-- /content -->
</body></html>