Buckets:

hf-doc-build
/

doc-dev

Files

xet

hf-doc-build/doc-dev / kernels /pr_675 /en /builder /writing-kernels.html

HuggingFaceDocBuilder

about 1 month ago

download

raw

93.2 kB

	<meta charset="utf-8" /><meta name="hf:doc:metadata" content="{"title":"Write kernels","local":"write-kernels","sections":[{"title":"Introduction","local":"introduction","sections":[],"depth":2},{"title":"Setting up environment","local":"setting-up-environment","sections":[{"title":"Quick install","local":"quick-install","sections":[],"depth":3},{"title":"Cloud environment","local":"cloud-environment","sections":[],"depth":3}],"depth":2},{"title":"Starting a new kernel","local":"starting-a-new-kernel","sections":[],"depth":2},{"title":"Kernel project layout","local":"kernel-project-layout","sections":[],"depth":2},{"title":"build.toml","local":"buildtoml","sections":[{"title":"general","local":"general","sections":[],"depth":3},{"title":"general.hub","local":"generalhub","sections":[],"depth":3},{"title":"general.cuda","local":"generalcuda","sections":[],"depth":3},{"title":"Framework sections","local":"framework-sections","sections":[],"depth":3},{"title":"torch","local":"torch","sections":[],"depth":3},{"title":"tvm-ffi","local":"tvm-ffi","sections":[],"depth":3},{"title":"torch-noarch","local":"torch-noarch","sections":[],"depth":3},{"title":"kernel.&lt;name&gt;","local":"kernelltnamegt","sections":[{"title":"cuda","local":"cuda","sections":[],"depth":4},{"title":"rocm","local":"rocm","sections":[],"depth":4},{"title":"xpu","local":"xpu","sections":[],"depth":4}],"depth":3},{"title":"cpu","local":"cpu","sections":[],"depth":3}],"depth":2},{"title":"Torch bindings","local":"torch-bindings","sections":[{"title":"Defining bindings","local":"defining-bindings","sections":[],"depth":3}],"depth":2},{"title":"Using kernel functions from Python","local":"using-kernel-functions-from-python","sections":[],"depth":2},{"title":"Registering Torch operators","local":"registering-torch-operators","sections":[],"depth":2},{"title":"Kernel tests","local":"kernel-tests","sections":[{"title":"Use get_kernel in tests","local":"use-getkernel-in-tests","sections":[],"depth":3},{"title":"Mark CI tests","local":"mark-ci-tests","sections":[],"depth":3}],"depth":2},{"title":"Kernel docs","local":"kernel-docs","sections":[],"depth":2}],"depth":1}"/>
	<link href="/docs/kernels/pr_675/en/_app/immutable/entry/start.BbgcxPYy.js" rel="modulepreload">
	<link href="/docs/kernels/pr_675/en/_app/immutable/chunks/C03Cu03y.js" rel="modulepreload">
	<link href="/docs/kernels/pr_675/en/_app/immutable/chunks/DfHjNWj2.js" rel="modulepreload">
	<link href="/docs/kernels/pr_675/en/_app/immutable/entry/app.Dtj_mlUF.js" rel="modulepreload">
	<link href="/docs/kernels/pr_675/en/_app/immutable/chunks/Ba1drWuH.js" rel="modulepreload">
	<link href="/docs/kernels/pr_675/en/_app/immutable/chunks/DTgriGRG.js" rel="modulepreload">
	<link href="/docs/kernels/pr_675/en/_app/immutable/chunks/DsnmJJEf.js" rel="modulepreload">
	<link href="/docs/kernels/pr_675/en/_app/immutable/chunks/PGLvW8OU.js" rel="modulepreload">
	<link href="/docs/kernels/pr_675/en/_app/immutable/nodes/0.DEUbFP7a.js" rel="modulepreload">
	<link href="/docs/kernels/pr_675/en/_app/immutable/chunks/CFswWuIO.js" rel="modulepreload">
	<link href="/docs/kernels/pr_675/en/_app/immutable/nodes/16.BcRqqQdP.js" rel="modulepreload">
	<link href="/docs/kernels/pr_675/en/_app/immutable/chunks/CyvF58-O.js" rel="modulepreload">
	<!--brj1rc--><meta name="hf:doc:metadata" content="{"title":"Write kernels","local":"write-kernels","sections":[{"title":"Introduction","local":"introduction","sections":[],"depth":2},{"title":"Setting up environment","local":"setting-up-environment","sections":[{"title":"Quick install","local":"quick-install","sections":[],"depth":3},{"title":"Cloud environment","local":"cloud-environment","sections":[],"depth":3}],"depth":2},{"title":"Starting a new kernel","local":"starting-a-new-kernel","sections":[],"depth":2},{"title":"Kernel project layout","local":"kernel-project-layout","sections":[],"depth":2},{"title":"build.toml","local":"buildtoml","sections":[{"title":"general","local":"general","sections":[],"depth":3},{"title":"general.hub","local":"generalhub","sections":[],"depth":3},{"title":"general.cuda","local":"generalcuda","sections":[],"depth":3},{"title":"Framework sections","local":"framework-sections","sections":[],"depth":3},{"title":"torch","local":"torch","sections":[],"depth":3},{"title":"tvm-ffi","local":"tvm-ffi","sections":[],"depth":3},{"title":"torch-noarch","local":"torch-noarch","sections":[],"depth":3},{"title":"kernel.&lt;name&gt;","local":"kernelltnamegt","sections":[{"title":"cuda","local":"cuda","sections":[],"depth":4},{"title":"rocm","local":"rocm","sections":[],"depth":4},{"title":"xpu","local":"xpu","sections":[],"depth":4}],"depth":3},{"title":"cpu","local":"cpu","sections":[],"depth":3}],"depth":2},{"title":"Torch bindings","local":"torch-bindings","sections":[{"title":"Defining bindings","local":"defining-bindings","sections":[],"depth":3}],"depth":2},{"title":"Using kernel functions from Python","local":"using-kernel-functions-from-python","sections":[],"depth":2},{"title":"Registering Torch operators","local":"registering-torch-operators","sections":[],"depth":2},{"title":"Kernel tests","local":"kernel-tests","sections":[{"title":"Use get_kernel in tests","local":"use-getkernel-in-tests","sections":[],"depth":3},{"title":"Mark CI tests","local":"mark-ci-tests","sections":[],"depth":3}],"depth":2},{"title":"Kernel docs","local":"kernel-docs","sections":[],"depth":2}],"depth":1}"/><!---->
	<link href="/docs/kernels/pr_675/en/_app/immutable/assets/0.tn0RQdqM.css" rel="modulepreload"> <!--[--><!--[0--><!--[--><!--[0--><!--[--><p></p> <div class="items-center shrink-0 min-w-[100px] max-sm:min-w-[50px] justify-end ml-auto flex" style="float: right; margin-left: 10px; display: inline-flex; position: relative; z-index: 10;"><div class="inline-flex rounded-md max-sm:rounded-sm"><button class="inline-flex items-center gap-1 h-7 max-sm:h-7 px-2 max-sm:px-1.5 text-sm font-medium text-gray-800 border border-r-0 rounded-l-md max-sm:rounded-l-sm border-gray-200 bg-white hover:shadow-inner dark:border-gray-850 dark:bg-gray-950 dark:text-gray-200 dark:hover:bg-gray-800" aria-live="polite"><span class="inline-flex items-center justify-center rounded-md p-0.5 max-sm:p-0 hover:text-gray-800 dark:hover:text-gray-200"><svg class="sm:size-3.5 size-3" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg><!----></span> <span>Copy page</span></button> <button class="inline-flex items-center justify-center w-6 max-sm:w-5 h-7 max-sm:h-7 disabled:pointer-events-none text-sm text-gray-500 hover:text-gray-700 dark:hover:text-white rounded-r-md max-sm:rounded-r-sm border border-l transition border-gray-200 bg-white hover:shadow-inner dark:border-gray-850 dark:bg-gray-950 dark:text-gray-200 dark:hover:bg-gray-800" aria-haspopup="menu" aria-expanded="false" aria-label="Open copy menu"><svg class="transition-transform text-gray-400 overflow-visible sm:size-3.5 size-3 rotate-0" width="1em" height="1em" viewBox="0 0 12 7" fill="none" xmlns="http://www.w3.org/2000/svg"><path d="M1 1L6 6L11 1" stroke="currentColor"></path></svg><!----></button></div> <!--[-1--><!--]--></div><!----> <!--[0--><h1 class="relative group"><a id="write-kernels" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#write-kernels"><span><svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!----></span></a> <span>Write kernels</span></h1><!--]--><!----> <!--[1--><h2 class="relative group"><a id="introduction" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#introduction"><span><svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!----></span></a> <span>Introduction</span></h2><!--]--><!----> <p>The Kernel Hub allows Python libraries and applications to load compute
	kernels directly from the <a href="https://hf.co/" rel="nofollow">Hub</a>. To support this kind
	of dynamic loading, Hub kernels differ from traditional Python kernel
	packages in that they are made to be:</p> <ul><li>Portable: a kernel can be loaded from paths outside <code>PYTHONPATH</code>.</li> <li>Unique: multiple versions of the same kernel can be loaded in the
	same Python process.</li> <li>Compatible: kernels must support all recent versions of Python and
	the different PyTorch build configurations (various CUDA versions
	and C++ ABIs). Furthermore, older C library versions must be supported.</li></ul> <p><code>kernel-builder</code> is a set of tools that can build conforming kernels. It
	takes care of:</p> <ul><li>Building kernels for all supported PyTorch configurations (C++98/11 and
	different CUDA versions).</li> <li>Compatibility with old glibc and libstdc++ versions, so that kernels also
	work on older Linux distributions.</li> <li>Registering Torch ops, such that multiple versions the same kernel can be
	loaded without namespace conflicts.</li></ul> <p><code>kernel-builder</code> builds are configured through a <code>build.toml</code> file. <code>build.toml</code> is a simple format that does not require intricate knowledge
	of CMake or setuptools.</p> <p>This page describes the directory layout of a kernel-builder project, the
	format of the <code>build.toml</code> file, and some additional Python glue that <code>kernel-builder</code> provides. We will use a <a href="https://github.com/huggingface/kernels/tree/main/examples/kernels/relu" rel="nofollow">simple ReLU kernel</a> as the running example. After reading this page, you may also want to have
	a look at the more realistic <a href="https://github.com/huggingface/kernels/tree/main/examples/kernels/relu-backprop-compile" rel="nofollow">ReLU kernel with backprop and <code>torch.compile</code></a> support.</p> <blockquote class="tip"><p>We maintain a set of conforming kernels in the <a href="https://github.com/huggingface/kernels-community" rel="nofollow">kernels-community repository</a>.
	We try to keep these kernels synced with upstream as much as possible.</p></blockquote> <!--[1--><h2 class="relative group"><a id="setting-up-environment" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#setting-up-environment"><span><svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!----></span></a> <span>Setting up environment</span></h2><!--]--><!----> <!--[2--><h3 class="relative group"><a id="quick-install" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#quick-install"><span><svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!----></span></a> <span>Quick install</span></h3><!--]--><!----> <p>The fastest way to get started is to run the install script. This
	installs <a href="https://docs.determinate.systems/determinate-nix/" rel="nofollow">Determinate Nix</a> and <code>kernel-builder</code> in a single command:</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg><!----> <div class=" absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0 "><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent;"></div> Copied</div><!----></button><!----></div> <pre class="language-bash "><!---->curl -fsSL https://raw.githubusercontent.com/huggingface/kernels/main/install.sh \| bash<!----></pre></div><!----> <p>This will:</p> <ol><li>Install Determinate Nix (if not already installed).</li> <li>Configure the Hugging Face binary cache (to avoid building dependencies from
	source).</li> <li>Install <code>kernel-builder</code> via <code>nix profile install</code>.</li></ol> <p>To update <code>kernel-builder</code> later:</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg><!----> <div class=" absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0 "><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent;"></div> Copied</div><!----></button><!----></div> <pre class="language-bash "><!---->nix profile upgrade --all<!----></pre></div><!----> <p>For a step-by-step breakdown of what the script does, see <a href="nix">Using the kernel builder with Nix</a>.</p> <!--[2--><h3 class="relative group"><a id="cloud-environment" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#cloud-environment"><span><svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!----></span></a> <span>Cloud environment</span></h3><!--]--><!----> <p>In the <a href="https://github.com/huggingface/kernels/tree/main/terraform" rel="nofollow"><code>terraform</code></a> directory, we provide an
	example of programatically spinning up an EC2 instance that is ready
	with everything needed for you to start developing and building
	kernels.</p> <p>If you use a different provider, the Terraform bridges should be
	similar and straightforward to modify.</p> <!--[1--><h2 class="relative group"><a id="starting-a-new-kernel" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#starting-a-new-kernel"><span><svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!----></span></a> <span>Starting a new kernel</span></h2><!--]--><!----> <p>The easiest way to start a new kernel is by using the <code>init</code> subcommand
	of <code>kernel-builder</code>. This creates a minimal, compilable kernel:</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg><!----> <div class=" absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0 "><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent;"></div> Copied</div><!----></button><!----></div> <pre class="language-bash "><!---->$ kernel-builder init --name myorg/mykernel
	Initialized `myorg/mykernel` at /home/daniel/git/kernels/examples/kernels/mykernel<!----></pre></div><!----> <p>This creates a kernel named <code>mykernel</code> in the directory <code>mykernel</code>. The
	kernel is configured to upload to the <code>myorg/mykernel</code> Hub
	repository when an upload command is used.</p> <p>By default, the <code>init</code> subcommand creates a CUDA kernel. You can specify
	another backend with the <code>--backends</code> option:</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg><!----> <div class=" absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0 "><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent;"></div> Copied</div><!----></button><!----></div> <pre class="language-bash "><!---->$ kernel-builder init --name myorg/mykernel --backends xpu<!----></pre></div><!----> <p>You can also make a multi-backend kernel by adding all the backends
	that you would like to support as arguments to <code>--backends</code>:</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg><!----> <div class=" absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0 "><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent;"></div> Copied</div><!----></button><!----></div> <pre class="language-bash "><!---->$ kernel-builder init --name myorg/mykernel --backends cuda xpu
	Initialized `myorg/mykernel` at /home/daniel/git/kernels/examples/kernels/mykernel<!----></pre></div><!----> <p>Finally, if you want to create a kernel for all supported backends, you
	can use <code>--backends all</code>.</p> <!--[1--><h2 class="relative group"><a id="kernel-project-layout" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#kernel-project-layout"><span><svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!----></span></a> <span>Kernel project layout</span></h2><!--]--><!----> <p>Kernel projects follow this general directory layout:</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg><!----> <div class=" absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0 "><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent;"></div> Copied</div><!----></button><!----></div> <pre class="language-text "><!---->mykernel
	├── benchmarks
	│ └── benchmark.py
	├── build.toml
	├── CARD.md
	├── example.py
	├── flake.nix
	├── mykernel_cuda
	│ └── mykernel.cu
	├── tests
	│ ├── __init__.py
	│ └── test_mykernel.py
	└── torch-ext
	├── mykernel
	│ └── __init__.py
	├── torch_binding.cpp
	└── torch_binding.h<!----></pre></div><!----> <p>In this example we can find:</p> <ul><li>The build configuration in <code>build.toml</code>.</li> <li>One or more top-level directories containing kernels (<code>mykernel_cuda</code>).</li> <li>The <code>torch-ext</code> directory, which contains: <ul><li><code>torch_binding.h</code>: contains declarations for kernel entry points
	(from <code>kernel_a</code> and <code>kernel_b</code>).</li> <li><code>torch_binding.cpp</code>: registers the entry points as Torch ops.</li> <li><code>torch_ext/mykernel</code>: contains any Python wrapping the kernel needs. At the
	bare minimum, it should contain an <code>__init__.py</code> file.</li></ul></li> <li>Kernel tests in the directory <code>tests</code>.</li> <li>Benchmarks in the directory <code>benchmarks</code>.</li> <li>A kernel card template in <code>CARD.md</code>. This placeholders in the card are filled
	during the kernel build.</li> <li>The Nix flake configuration in <code>flake.nix</code>.</li> <li>An example script that uses the kernel in <code>example.py</code>.</li></ul> <!--[1--><h2 class="relative group"><a id="buildtoml" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#buildtoml"><span><svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!----></span></a> <span>build.toml</span></h2><!--]--><!----> <p><code>build.toml</code> tells <code>kernel-builder</code> what to build and how. It looks as
	follows for the <code>mykernel</code> kernel:</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg><!----> <div class=" absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0 "><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent;"></div> Copied</div><!----></button><!----></div> <pre class="language-toml "><!----><span class="hljs-section">[general]</span>
	<span class="hljs-attr">backends</span> = [
	<span class="hljs-string">"cuda"</span>,
	]
	<span class="hljs-attr">name</span> = <span class="hljs-string">"mykernel"</span>
	<span class="hljs-attr">version</span> = <span class="hljs-number">1</span>
	<span class="hljs-attr">edition</span> = <span class="hljs-number">5</span>

	<span class="hljs-section">[general.hub]</span>
	<span class="hljs-attr">repo-id</span> = <span class="hljs-string">"myorg/mykernel"</span>

	<span class="hljs-section">[torch]</span>
	<span class="hljs-attr">src</span> = [
	<span class="hljs-string">"torch-ext/torch_binding.cpp"</span>,
	<span class="hljs-string">"torch-ext/torch_binding.h"</span>,
	]

	<span class="hljs-section">[kernel.mykernel]</span>
	<span class="hljs-attr">backend</span> = <span class="hljs-string">"cuda"</span>
	<span class="hljs-attr">depends</span> = [<span class="hljs-string">"torch"</span>]
	<span class="hljs-attr">src</span> = [<span class="hljs-string">"mykernel_cuda/mykernel.cu"</span>]
	<span class="hljs-comment"># If the kernel is only supported on specific capabilities, set the</span>
	<span class="hljs-comment"># cuda-capabilities option:</span>
	<span class="hljs-comment">#</span>
	<span class="hljs-comment"># cuda-capabilities = [ "9.0", "10.0", "12.0" ]</span><!----></pre></div><!----> <p>The following sections enumerate all supported options for <code>build.toml</code>.</p> <!--[2--><h3 class="relative group"><a id="general" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#general"><span><svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!----></span></a> <span>general</span></h3><!--]--><!----> <ul><li><code>name</code> (required): the name of the kernel. The Python code for a Torch
	extension must be stored in <code>torch-ext/<name></code>.</li> <li><code>version</code> (int): the major version of the kernel.
	The version is written to the kernel’s <code>metadata.json</code> and is used
	by the <code>kernels upload</code> command to upload the kernel to a version
	branch named <code>v<version></code>.</li> <li><code>edition</code> (required): the <code>build.toml</code> format edition. The current
	edition is <code>5</code>. Older <code>build.toml</code> files can be migrated with <code>kernel-builder update-build</code>.</li> <li><code>backends</code> (required): a list of supported backends. Must be one or
	more of <code>cpu</code>, <code>cuda</code>, <code>metal</code>, <code>rocm</code>, or <code>xpu</code>.</li> <li><code>upstream</code>: Git-compatible URL (passable to <code>git clone</code>) of the original
	upstream repository where the kernel source code comes from.</li> <li><code>source</code>: Git-compatible URL (passable to <code>git clone</code>) of the kernel-builder
	formatted source repository. This repository must contain a <code>build.toml</code> and <code>flake.nix</code> so that it can be pulled and built with the kernel builder.</li> <li><code>python-depends</code> (<strong>experimental</strong>): a list of additional Python dependencies
	that the kernel requires. The only supported dependencies are <code>einops</code> and <code>nvidia-cutlass-dsl</code>.</li></ul> <!--[2--><h3 class="relative group"><a id="generalhub" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#generalhub"><span><svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!----></span></a> <span>general.hub</span></h3><!--]--><!----> <ul><li><code>repo-id</code>: the Hub repository to upload the kernel to when the <code>upload</code> or <code>build-and-upload</code> subcommands of <code>kernel-builder</code> are used.</li></ul> <!--[2--><h3 class="relative group"><a id="generalcuda" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#generalcuda"><span><svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!----></span></a> <span>general.cuda</span></h3><!--]--><!----> <ul><li><code>maxver</code>: the maximum CUDA toolkit version (inclusive). This option <em>must not</em> be set under normal circumstances, since it can exclude Torch
	build variants that are <a href="../kernel-requirements">required for compliant kernels</a>.
	This option is provided for kernels that cause compiler errors on
	newer CUDA toolkit versions.</li> <li><code>minver</code>: the minimum required CUDA toolkit version. This option <em>must not</em> be set under normal circumstances, since it can exclude Torch
	build variants that are <a href="../kernel-requirements">required for compliant kernels</a>.
	This option is provided for kernels that require functionality only
	provided by newer CUDA toolkits.</li></ul> <!--[2--><h3 class="relative group"><a id="framework-sections" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#framework-sections"><span><svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!----></span></a> <span>Framework sections</span></h3><!--]--><!----> <p>The framework section specifies framework-specific settings. The name of
	the section depends on the framework that is used. The currently supported
	frameworks are:</p> <ul><li>AOT-compiled Torch kernel (<code>torch</code>).</li> <li>AOT-compiled TVM-FFI kernel (<code>tvm-ffi</code>).</li> <li>JIT-compiled or not-compiled Torch kernel (<code>torch-noarch</code>, experimental).</li></ul> <!--[2--><h3 class="relative group"><a id="torch" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#torch"><span><svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!----></span></a> <span>torch</span></h3><!--]--><!----> <p>This framework section is used for AOT-compiled Torch kernels, and has the
	following options:</p> <ul><li><p><code>src</code> (required): a list of source files and headers.</p></li> <li><p><code>pyext</code> (optional): the list of extensions for Python files. Default: <code>["py", "pyi"]</code>.</p></li> <li><p><code>include</code> (optional): include directories relative to the project root.
	Default: <code>[]</code>.</p></li> <li><p><code>maxver</code> (optional): only build for this Torch version and earlier. Use cautiously, since this option produces
	non-compliant kernels if the version range does not correspond to the <a href="build-variants">required variants</a>.</p></li> <li><p><code>minver</code> (optional): only build for this Torch version and later. Use cautiously, since this option produces
	non-compliant kernels if the version range does not correspond to the <a href="build-variants">required variants</a>.</p></li> <li><p><code>stable-abi</code> (<strong>experimental</strong>): a table mapping backend names to the Torch
	version (e.g. <code>"2.11"</code>) that the backend is built against using the Torch
	stable ABI. This requires that the kernel itself only use <a href="https://docs.pytorch.org/docs/2.12/notes/libtorch_stable_abi.html" rel="nofollow">stable ABI headers</a>.
	For an example, see the <a href="https://github.com/huggingface/kernels/tree/main/examples/kernels/relu-torch-stable-abi" rel="nofollow"><code>relu-torch-stable-abi</code></a> example kernel.</p> <p>Backends that are not listed in the table are built normally (without the
	stable ABI), allowing a kernel to mix stable-ABI and full-ABI backends:</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg><!----> <div class=" absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0 "><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent;"></div> Copied</div><!----></button><!----></div> <pre class="language-toml "><!----><span class="hljs-section">[torch.stable-abi]</span>
	<span class="hljs-attr">cuda</span> = <span class="hljs-string">"2.11"</span>
	<span class="hljs-attr">rocm</span> = <span class="hljs-string">"2.9"</span><!----></pre></div><!----> <p>Entries for backends that are not in <code>[general].backends</code> are ignored, so a
	backend can be commented out of <code>[general].backends</code> for testing without
	having to also remove it from the table.</p></li></ul> <!--[2--><h3 class="relative group"><a id="tvm-ffi" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#tvm-ffi"><span><svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!----></span></a> <span>tvm-ffi</span></h3><!--]--><!----> <p>This framework section is used for AOT-compiled TVM-FFI kernels.</p> <ul><li><code>src</code> (required): a list of source files and headers.</li> <li><code>pyext</code> (optional): the list of extensions for Python files. Default: <code>["py", "pyi"]</code>.</li> <li><code>include</code> (optional): include directories relative to the project root.
	Default: <code>[]</code>.</li></ul> <!--[2--><h3 class="relative group"><a id="torch-noarch" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#torch-noarch"><span><svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!----></span></a> <span>torch-noarch</span></h3><!--]--><!----> <p>The <code>torch-noarch</code> section is used for JIT-compiled kernels or kernels that
	do not require any ahead-of-time compilation (e.g. a kernel that packages plain PyTorch
	layers).</p> <p>Normally, it is expected that this type of kernel runs on all CUDA capabilities
	or ROCm architectures. However, for kernels that support only a limited range
	of archs, the <code>cuda-capabilites</code> and <code>rocm-archs</code> options can be used to specify
	the supported archs. These are then exported to <code>metadata.json</code> for consumption
	by e.g. the Hugging Face Hub.</p> <ul><li><code>pyext</code> (optional): the list of extensions for Python files. Default: <code>["py", "pyi"]</code>.</li> <li><code>cuda-capabilities</code> (optional): a list of CUDA compute capabilities the
	kernel supports (e.g. <code>["9.0", "10.0"]</code>).</li> <li><code>rocm-archs</code> (optional): a list of ROCm architectures the kernel supports
	(e.g. <code>["gfx942"]</code>).</li></ul> <!--[2--><h3 class="relative group"><a id="kernelltnamegt" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#kernelltnamegt"><span><svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!----></span></a> <span>kernel.<name></span></h3><!--]--><!----> <p>Specification of a kernel with the name <code><name></code>. Multiple <code>kernel.<name></code> sections can be defined in the same <code>build.toml</code>.
	See for example <a href="https://huggingface.co/kernels-community/quantization/" rel="nofollow"><code>kernels-community/quantization</code></a> for an example with multiple kernel sections.</p> <p>The following options can be set for a kernel:</p> <ul><li><code>backend</code> (required): the compute backend of the kernel. The currently
	supported backends are <code>cpu</code>, <code>cuda</code>, <code>metal</code>, <code>rocm</code>, and <code>xpu</code>. <strong>The <code>cpu</code> backend is currently experimental and might still change.</strong></li> <li><code>depends</code> (required): a list of dependencies. The supported dependencies
	are listed in <a href="https://github.com/huggingface/kernels/blob/main/builder/lib/deps.nix" rel="nofollow"><code>deps.nix</code></a>.</li> <li><code>src</code> (required): a list of source files and headers.</li> <li><code>include</code> (optional): include directories relative to the project root.
	Default: <code>[]</code>.</li></ul> <p>Besides these shared options, the following backend-specific options
	are available:</p> <!--[3--><h4 class="relative group"><a id="cuda" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#cuda"><span><svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!----></span></a> <span>cuda</span></h4><!--]--><!----> <ul><li><code>cuda-capabilities</code> (optional): a list of CUDA capabilities that the
	kernel should be compiled for. When absent, the kernel will be built
	using all capabilities that the builder supports. The effective
	capabilities are the intersection of this list and the capabilities
	supported by the CUDA compiler. It is recommended to leave this option
	unspecified <strong>unless</strong> a kernel requires specific capabilities.</li> <li><code>cuda-flags</code> (optional): additional flags to be passed to <code>nvcc</code>. <strong>Warning</strong>: this option should only be used in exceptional circumstances.
	Custom compile flags can interfere with the build process or break
	compatibility requirements.</li></ul> <!--[3--><h4 class="relative group"><a id="rocm" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#rocm"><span><svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!----></span></a> <span>rocm</span></h4><!--]--><!----> <ul><li><code>rocm-archs</code>: a list of ROCm architectures that the kernel should be
	compiled for.</li></ul> <!--[3--><h4 class="relative group"><a id="xpu" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#xpu"><span><svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!----></span></a> <span>xpu</span></h4><!--]--><!----> <ul><li><code>sycl-flags</code>: a list of additional flags to be passed to the SYCL
	compiler.</li></ul> <!--[2--><h3 class="relative group"><a id="cpu" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#cpu"><span><svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!----></span></a> <span>cpu</span></h3><!--]--><!----> <ul><li><code>cxx-flags</code>: a list of additional flags to be passed to the C++
	compiler.</li></ul> <!--[1--><h2 class="relative group"><a id="torch-bindings" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#torch-bindings"><span><svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!----></span></a> <span>Torch bindings</span></h2><!--]--><!----> <!--[2--><h3 class="relative group"><a id="defining-bindings" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#defining-bindings"><span><svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!----></span></a> <span>Defining bindings</span></h3><!--]--><!----> <p>Torch bindings are defined in C++, kernels commonly use two files:</p> <ul><li><code>torch_binding.h</code> containing function declarations.</li> <li><code>torch_binding.cpp</code> registering the functions as Torch ops.</li></ul> <p>For instance, the <code>mykernel</code> kernel discussed above has the following
	declaration in <code>torch_binding.h</code>:</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg><!----> <div class=" absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0 "><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent;"></div> Copied</div><!----></button><!----></div> <pre class="language-cpp "><!----><span class="hljs-meta">#<span class="hljs-keyword">pragma</span> once</span>

	<span class="hljs-meta">#<span class="hljs-keyword">include</span> <span class="hljs-string"><torch/torch.h></span></span>

	<span class="hljs-function"><span class="hljs-type">void</span> <span class="hljs-title">mykernel</span><span class="hljs-params">(torch::Tensor &out, torch::Tensor <span class="hljs-type">const</span> &input)</span></span>;<!----></pre></div><!----> <p>This function is then registered as a Torch op in <code>torch_binding.cpp</code>:</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg><!----> <div class=" absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0 "><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent;"></div> Copied</div><!----></button><!----></div> <pre class="language-cpp "><!----><span class="hljs-meta">#<span class="hljs-keyword">include</span> <span class="hljs-string"><torch/library.h></span></span>

	<span class="hljs-meta">#<span class="hljs-keyword">include</span> <span class="hljs-string">"registration.h"</span></span>
	<span class="hljs-meta">#<span class="hljs-keyword">include</span> <span class="hljs-string">"torch_binding.h"</span></span>

	<span class="hljs-built_in">TORCH_LIBRARY_EXPAND</span>(TORCH_EXTENSION_NAME, ops) {
	ops.<span class="hljs-built_in">def</span>(<span class="hljs-string">"mykernel(Tensor! out, Tensor input) -> ()"</span>);
	<span class="hljs-meta">#<span class="hljs-keyword">if</span> defined(CUDA_KERNEL) \|\| defined(ROCM_KERNEL)</span>
	ops.<span class="hljs-built_in">impl</span>(<span class="hljs-string">"mykernel"</span>, torch::kCUDA, &mykernel);
	<span class="hljs-meta">#<span class="hljs-keyword">endif</span></span>
	}

	<span class="hljs-built_in">REGISTER_EXTENSION</span>(TORCH_EXTENSION_NAME)<!----></pre></div><!----> <p>This snippet uses macros from <code>registration.h</code> to register the function. <code>registration.h</code> is generated by <code>kernel-builder</code> itself. A function
	is registered through the <code>def</code>/<code>ops</code> methods. <code>ops</code> specifies the
	function signature following the <a href="https://github.com/pytorch/pytorch/blob/main/aten/src/ATen/native/README.md#func" rel="nofollow">function schema</a>. <code>impl</code> associates the function name with the C/C++ function and
	the applicable device.</p> <!--[1--><h2 class="relative group"><a id="using-kernel-functions-from-python" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#using-kernel-functions-from-python"><span><svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!----></span></a> <span>Using kernel functions from Python</span></h2><!--]--><!----> <p>The bindings are typically wrapped in Python code in <code>torch_ext/<name></code>.
	The native code is exposed under the <code>torch.ops</code> namespace. However,
	we add some unique material to the name of the extension to ensure that
	different versions of the same extension can be loaded at the same time.
	As a result, the extension is registered as <code>torch.ops.<name>_<unique_material></code>.</p> <p>To deal with this uniqueness, <code>kernel_builder</code> generates a Python module
	named <code>_ops</code> that contains an alias for the name. This can be used to
	refer to the correct <code>torch.ops</code> module. For example:</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg><!----> <div class=" absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0 "><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent;"></div> Copied</div><!----></button><!----></div> <pre class="language-python "><!----><span class="hljs-keyword">from</span> typing <span class="hljs-keyword">import</span> <span class="hljs-type">Optional</span>

	<span class="hljs-keyword">import</span> torch

	<span class="hljs-keyword">from</span> ._ops <span class="hljs-keyword">import</span> ops


	<span class="hljs-keyword">def</span> <span class="hljs-title function_">mykernel</span>(<span class="hljs-params">x: torch.Tensor, out: <span class="hljs-type">Optional</span>[torch.Tensor] = <span class="hljs-literal">None</span></span>) -> torch.Tensor:
	<span class="hljs-keyword">if</span> out <span class="hljs-keyword">is</span> <span class="hljs-literal">None</span>:
	out = torch.empty_like(x)
	ops.mykernel(out, x)
	<span class="hljs-keyword">return</span> out<!----></pre></div><!----> <!--[1--><h2 class="relative group"><a id="registering-torch-operators" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#registering-torch-operators"><span><svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!----></span></a> <span>Registering Torch operators</span></h2><!--]--><!----> <p>You may want to register Torch ops from your kernel’s Python code or <a href="https://docs.pytorch.org/tutorials/advanced/custom_ops_landing_page.html" rel="nofollow">register</a> fake ops for <code>torch.compile</code> support. It is important to register
	such ops in the namespace that kernel-builder makes for your kernel
	build. This is required for compliant kernels to ensure that multiple
	versions of the same kernel can be loaded at the same time without
	namespace conflicts.</p> <p>You can use the <code>add_op_namespace_prefix</code> to prefix an op name with the
	correct prefix. So for instance, replace</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg><!----> <div class=" absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0 "><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent;"></div> Copied</div><!----></button><!----></div> <pre class="language-python "><!----><span class="hljs-meta">@torch.library.register_fake(<span class="hljs-params"><span class="hljs-string">"relu::relu_fwd"</span></span>)</span>
	<span class="hljs-keyword">def</span> <span class="hljs-title function_">relu_fwd_fake</span>(<span class="hljs-params"><span class="hljs-built_in">input</span>: torch.Tensor</span>) -> torch.Tensor:
	<span class="hljs-keyword">return</span> torch.empty_like(<span class="hljs-built_in">input</span>)<!----></pre></div><!----> <p>by</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg><!----> <div class=" absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0 "><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent;"></div> Copied</div><!----></button><!----></div> <pre class="language-python "><!----><span class="hljs-keyword">from</span> ._ops <span class="hljs-keyword">import</span> add_op_namespace_prefix

	<span class="hljs-meta">@torch.library.register_fake(<span class="hljs-params">add_op_namespace_prefix(<span class="hljs-params"><span class="hljs-string">"relu_fwd"</span></span>)</span>)</span>
	<span class="hljs-keyword">def</span> <span class="hljs-title function_">relu_fwd_fake</span>(<span class="hljs-params"><span class="hljs-built_in">input</span>: torch.Tensor</span>) -> torch.Tensor:
	<span class="hljs-keyword">return</span> torch.empty_like(<span class="hljs-built_in">input</span>)<!----></pre></div><!----> <p>As mentioned in the above, the <code>_ops</code> module is generated by kernel-builder.</p> <p>kernel-builder uses a hook to reject incorrect usage of Torch op registration
	functions. However, it can only catch direct use of certain <code>torch.library</code> decorators. For instance, the hook would not reject the following decorator,
	so it should be seen as a last-resort check if human review failed:</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg><!----> <div class=" absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0 "><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent;"></div> Copied</div><!----></button><!----></div> <pre class="language-python "><!----><span class="hljs-meta">@some_indirection_for_register_fake(<span class="hljs-params"><span class="hljs-string">"relu::relu_fwd"</span></span>)</span>
	<span class="hljs-keyword">def</span> <span class="hljs-title function_">relu_fwd_fake</span>(<span class="hljs-params"><span class="hljs-built_in">input</span>: torch.Tensor</span>) -> torch.Tensor:
	<span class="hljs-keyword">return</span> torch.empty_like(<span class="hljs-built_in">input</span>)<!----></pre></div><!----> <blockquote class="warning"><p>To facilitate static analysis of ops by kernel-builder, <code>add_op_namespace_prefix</code> should not be rewrapped. Furthermore, no fallbacks should be added when importing <code>add_op_namnespace_prefix</code>, since such fallbacks can mask issues (e.g. incorrect import paths), resulting in non-unique op names. Below is an example of this antipattern:</p></blockquote> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg><!----> <div class=" absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0 "><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent;"></div> Copied</div><!----></button><!----></div> <pre class="language-py "><!----><span class="hljs-keyword">try</span>:
	<span class="hljs-keyword">from</span> ._ops <span class="hljs-keyword">import</span> add_op_namespace_prefix <span class="hljs-keyword">as</span> _generated_add_op_namespace_prefix
	<span class="hljs-keyword">except</span> ImportError:
	<span class="hljs-keyword">def</span> <span class="hljs-title function_">_generated_add_op_namespace_prefix</span>(<span class="hljs-params">name: <span class="hljs-built_in">str</span></span>) -> <span class="hljs-built_in">str</span>:
	<span class="hljs-keyword">return</span> name <span class="hljs-keyword">if</span> <span class="hljs-string">"::"</span> <span class="hljs-keyword">in</span> name <span class="hljs-keyword">else</span> <span class="hljs-string">f"my_kernel::<span class="hljs-subst">{name}</span>"</span>

	<span class="hljs-keyword">def</span> <span class="hljs-title function_">add_op_namespace_prefix</span>(<span class="hljs-params">name: <span class="hljs-built_in">str</span></span>) -> <span class="hljs-built_in">str</span>:
	<span class="hljs-keyword">return</span> _generated_add_op_namespace_prefix(name)<!----></pre></div><!----> <!--[1--><h2 class="relative group"><a id="kernel-tests" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#kernel-tests"><span><svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!----></span></a> <span>Kernel tests</span></h2><!--]--><!----> <!--[2--><h3 class="relative group"><a id="use-getkernel-in-tests" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#use-getkernel-in-tests"><span><svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!----></span></a> <span>Use get_kernel in tests</span></h3><!--]--><!----> <p>Kernel tests are stored in the <code>tests</code> directory. Tests must not use direct
	imports, but instead use <code>get_kernel</code> to test the kernel as it will be used.
	For example:</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg><!----> <div class=" absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0 "><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent;"></div> Copied</div><!----></button><!----></div> <pre class="language-python "><!----><span class="hljs-keyword">import</span> kernels
	<span class="hljs-keyword">import</span> torch
	<span class="hljs-keyword">import</span> torch.nn.functional <span class="hljs-keyword">as</span> F

	relu = kernels.get_kernel(<span class="hljs-string">"kernels-community/relu"</span>, version=<span class="hljs-number">1</span>)

	<span class="hljs-keyword">def</span> <span class="hljs-title function_">test_relu</span>():
	x = torch.randn(<span class="hljs-number">1024</span>, <span class="hljs-number">1024</span>, dtype=torch.float32, device=torch.device(<span class="hljs-string">"cuda"</span>))
	y = relu.relu(x, torch.empty_like(x))
	y_ref = F.relu(x)
	torch.testing.assert_close(y_ref, y)<!----></pre></div><!----> <p>Development shells (<code>kernel-builder devshell</code>/<code>kernel-builder testshell</code>)
	will set the <code>LOCAL_KERNELS</code> variable to ensure that the kernel will be
	loaded from the development environment.</p> <!--[2--><h3 class="relative group"><a id="mark-ci-tests" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#mark-ci-tests"><span><svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!----></span></a> <span>Mark CI tests</span></h3><!--]--><!----> <p>Since running all kernel tests in CI may be prohibitively expensive, the <code>pyproject.toml</code> generated by the builder adds support for the special <code>kernels_ci</code> PyTest marker that can be used as follows:</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg><!----> <div class=" absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0 "><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent;"></div> Copied</div><!----></button><!----></div> <pre class="language-python "><!----><span class="hljs-keyword">import</span> pytest

	<span class="hljs-meta">@pytest.mark.kernels_ci</span>
	<span class="hljs-keyword">def</span> <span class="hljs-title function_">test_mykernel</span>():
	...<!----></pre></div><!----> <p>We recommend that you to pick tests that together would catch most error
	cases while running within 60 seconds.</p> <p>You can run the tests (e.g. in CI) using:</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg><!----> <div class=" absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0 "><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent;"></div> Copied</div><!----></button><!----></div> <pre class="language-bash "><!---->$ nix run .#ci-test<!----></pre></div><!----> <p>If the kernel supports multiple backends, it will run the test for the
	first supported backend that was found, obeying the following order: CUDA,
	ROCm, XPU, Metal, CPU. If you would like to the tests for a specific build
	variant, you can use <code>nix run .#ciTests.<variant></code>. For instance:</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg><!----> <div class=" absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0 "><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent;"></div> Copied</div><!----></button><!----></div> <pre class="language-bash "><!---->$ nix run .#ciTests.torch210-cxx11-cpu-x86_64-linux<!----></pre></div><!----> <p>When running the tests on a non-NixOS systems, make sure that <a href="https://danieldk.eu/Software/Nix/Nix-CUDA-on-non-NixOS-systems#solutions" rel="nofollow">the CUDA driver library can be found</a>.</p> <!--[1--><h2 class="relative group"><a id="kernel-docs" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#kernel-docs"><span><svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!----></span></a> <span>Kernel docs</span></h2><!--]--><!----> <p>We provide a utility to generate a system card for a given kernel, utilizing
	information from its <code>build.toml</code> and metadata. This system card provides a
	reasonable starting point and is meant to be edited afterward by the kernel
	developer.</p> <p>The template card is generated as a part of <code>kernel-builder init</code> command and is serialized in the root directory of the kernel.</p> <p>The card will be filled automatically by the builder when using the <code>build-and-upload</code> or <code>build-and-copy</code> command. It will be serialized
	to the <code>build</code> sub-directory inside the main kernel directory. It
	will be uploaded as <code>README.md</code> to the Hub.</p> <a class="!text-gray-400 !no-underline text-sm flex items-center not-prose mt-4" href="https://github.com/huggingface/kernels/blob/main/docs/source/builder/writing-kernels.md" target="_blank"><svg class="mr-1" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M31,16l-7,7l-1.41-1.41L28.17,16l-5.58-5.59L24,9l7,7z"></path><path d="M1,16l7-7l1.41,1.41L3.83,16l5.58,5.59L8,23l-7-7z"></path><path d="M12.419,25.484L17.639,6.552l1.932,0.518L14.351,26.002z"></path></svg><!----> <span><span class="underline">Update</span> on GitHub</span></a><!----> <p></p><!--]--><!----><!--]--><!--]--><!--]--> <!--[-1--><!--]--><!--]-->

	<script>
	{
	__sveltekit_f0hrjw = {
	base: "/docs/kernels/pr_675/en",
	assets: "/docs/kernels/pr_675/en"
	};

	const element = document.currentScript.parentElement;

	Promise.all([
	import("/docs/kernels/pr_675/en/_app/immutable/entry/start.BbgcxPYy.js"),
	import("/docs/kernels/pr_675/en/_app/immutable/entry/app.Dtj_mlUF.js")
	]).then(([kit, app]) => {
	kit.start(app, element, {
	node_ids: [0, 16],
	data: [null,null],
	form: null,
	error: null
	});
	});
	}
	</script>

Xet Storage Details

Size:: 93.2 kB
Xet hash:: c6a65a6630b64efeb95c666930f4965d72f42ebb48aec05d95b197a64d861475

Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.