peft/pr_3111/en/package_reference/lora_variant_kasa.html

<meta charset="utf-8" /><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;KaSA&quot;,&quot;local&quot;:&quot;kasa&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Caveats&quot;,&quot;local&quot;:&quot;caveats&quot;,&quot;sections&quot;:[],&quot;depth&quot;:4}],&quot;depth&quot;:3}"/>
		<link href="/docs/peft/main/en/_app/immutable/entry/start.H0B-zMV-.js" rel="modulepreload">
		<link href="/docs/peft/main/en/_app/immutable/chunks/B6nluxQH.js" rel="modulepreload">
		<link href="/docs/peft/main/en/_app/immutable/chunks/CCQPhADn.js" rel="modulepreload">
		<link href="/docs/peft/main/en/_app/immutable/entry/app.DFygnjAi.js" rel="modulepreload">
		<link href="/docs/peft/main/en/_app/immutable/chunks/DW4-mrp-.js" rel="modulepreload">
		<link href="/docs/peft/main/en/_app/immutable/chunks/RUwc_6JX.js" rel="modulepreload">
		<link href="/docs/peft/main/en/_app/immutable/chunks/DsnmJJEf.js" rel="modulepreload">
		<link href="/docs/peft/main/en/_app/immutable/chunks/DsPOG0MZ.js" rel="modulepreload">
		<link href="/docs/peft/main/en/_app/immutable/nodes/0.C09delyQ.js" rel="modulepreload">
		<link href="/docs/peft/main/en/_app/immutable/chunks/C5swDPgD.js" rel="modulepreload">
		<link href="/docs/peft/main/en/_app/immutable/nodes/50.DFcBSNnv.js" rel="modulepreload">
		<link href="/docs/peft/main/en/_app/immutable/chunks/C96YZLYZ.js" rel="modulepreload">
		<!--zwhu6p--><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;KaSA&quot;,&quot;local&quot;:&quot;kasa&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Caveats&quot;,&quot;local&quot;:&quot;caveats&quot;,&quot;sections&quot;:[],&quot;depth&quot;:4}],&quot;depth&quot;:3}"/><!---->
		<link href="/docs/peft/main/en/_app/immutable/assets/0.tn0RQdqM.css" rel="modulepreload"> <!--[--><!--[0--><!--[--><!--[0--><!--[--><p></p> <!--[2--><h3 class="relative group"><a id="kasa" 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="#kasa"><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>KaSA</span></h3><!--]--><!----> <blockquote class="note"><p>This is a variant of LoRA and therefore everything that is possible with LoRA is valid for this method except otherwise stated on this page.</p></blockquote> <p><a href="https://huggingface.co/papers/2412.06071" rel="nofollow">KaSA</a> (Knowledge-aware Singular-value Adaptation) is a LoRA variant that uses the singular value decomposition of the base weight to filter out task-irrelevant knowledge and parametrizes the update with learnable singular values. It changes vanilla LoRA in two ways:</p> <ol><li><strong>Knowledge-based SVD truncation of the frozen base weight.</strong> At initialization, the base weight <code>W</code> is SVD-factored and its <code>r</code> smallest (“noisy”/long-tail) singular components are discarded, leaving the rank-<code>(k - r)</code> approximation as the new frozen base (<code>k = min(in_features, out_features)</code>). The trainable branch then re-learns in the discarded residual subspace.</li> <li><strong>Knowledge-aware singular-value adaptation.</strong> The trainable update is parametrized in SVD form with a learnable diagonal of singular values inserted between the LoRA factors: <code>ΔW = scaling * B @ diag(ΔΣ) @ A</code>, where <code>ΔΣ</code> (<code>lora_diag</code>) is a learnable <code>r</code>-vector and the only new parameter per layer.</li></ol> <p>In PEFT, KaSA is configured as a LoRA variant through the <code>kasa_config</code> argument on <a href="/docs/peft/pr_3111/en/package_reference/lora#peft.LoraConfig">LoraConfig</a>:</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-py "><!----><span class="hljs-keyword">from</span> peft <span class="hljs-keyword">import</span> KasaConfig, LoraConfig

config = LoraConfig(
    target_modules=[<span class="hljs-string">&quot;q_proj&quot;</span>, <span class="hljs-string">&quot;v_proj&quot;</span>],
    kasa_config=KasaConfig(beta=<span class="hljs-number">1e-4</span>, gamma=<span class="hljs-number">1e-3</span>),
)<!----></pre></div><!----> <p>The paper additionally trains with two auxiliary regularizers: an L2 penalty on the learnable singular values (weighted by <code>beta</code>) and an orthogonal regularization on the adapter factors (weighted by <code>gamma</code>), which softly enforces the semi-orthogonality assumed by the SVD parametrization. These cannot be injected automatically by PEFT, so during training you must add them to the task loss by calling <code>LoraModel._get_kasa_loss()</code> on the underlying <code>LoraModel</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-py "><!---->task_loss = ...  <span class="hljs-comment"># standard loss returned by your model</span>
kasa_loss = model._get_kasa_loss()  <span class="hljs-comment"># 0.0 if KaSA is not used</span>
total_loss = task_loss + kasa_loss<!----></pre></div><!----> <p>For detailed usage, see <a href="https://github.com/huggingface/peft/tree/main/examples/kasa_finetuning" rel="nofollow">these instructions</a>.</p> <!--[3--><h4 class="relative group"><a id="caveats" 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="#caveats"><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>Caveats</span></h4><!--]--><!----> <ul><li>KaSA is currently supported on standard LoRA linear layers only, and not with <code>fan_in_fan_out=True</code> layers (e.g. transformers <code>Conv1D</code>).</li> <li>KaSA adapters cannot be combined with non-KaSA adapters on the same model, since the base-weight truncation would change the base weights under the other adapters’ feet. Multiple KaSA adapters are allowed.</li> <li><code>convert_to_lora</code> is not supported: the KaSA update depends on <code>lora_diag</code> and on the truncated base weight, neither of which is representable in a vanilla LoRA adapter.</li> <li>The SVD truncation of the base weight is <strong>destructive</strong>: adding a KaSA adapter permanently changes the layer’s frozen weight. Disabling or unloading the adapter does not restore the original base weight, and <code>merge</code> followed by <code>unmerge</code> round-trips to the truncated weight, not the original one. This is inherent to the method. Keep the original checkpoint if you need to recover the unmodified base model.</li> <li>Loading a trained KaSA adapter with <code>PeftModel.from_pretrained</code> re-applies the same truncation to the freshly loaded base weight, so saving and reloading is consistent.</li></ul> <a class="!text-gray-400 !no-underline text-sm flex items-center not-prose mt-4" href="https://github.com/huggingface/peft/blob/main/docs/source/package_reference/lora_variant_kasa.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_1jptnp0 = {
						base: "/docs/peft/main/en",
						assets: "/docs/peft/main/en"
					};

					const element = document.currentScript.parentElement;

					Promise.all([
						import("/docs/peft/main/en/_app/immutable/entry/start.H0B-zMV-.js"),
						import("/docs/peft/main/en/_app/immutable/entry/app.DFygnjAi.js")
					]).then(([kit, app]) => {
						kit.start(app, element, {
							node_ids: [0, 50],
							data: [null,null],
							form: null,
							error: null
						});
					});
				}
			</script>