transformers/pr_37537/en/quantization/quark.html

<meta charset="utf-8" /><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;Quark&quot;,&quot;local&quot;:&quot;quark&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Support matrix&quot;,&quot;local&quot;:&quot;support-matrix&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Models on Hugging Face Hub&quot;,&quot;local&quot;:&quot;models-on-hugging-face-hub&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Using Quark models in Transformers&quot;,&quot;local&quot;:&quot;using-quark-models-in-transformers&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2}],&quot;depth&quot;:1}">
		<link href="/docs/transformers/pr_37537/en/_app/immutable/assets/0.e3b0c442.css" rel="modulepreload">
		<link rel="modulepreload" href="/docs/transformers/pr_37537/en/_app/immutable/entry/start.f786cb3e.js">
		<link rel="modulepreload" href="/docs/transformers/pr_37537/en/_app/immutable/chunks/scheduler.18a86fab.js">
		<link rel="modulepreload" href="/docs/transformers/pr_37537/en/_app/immutable/chunks/singletons.9d0600a8.js">
		<link rel="modulepreload" href="/docs/transformers/pr_37537/en/_app/immutable/chunks/index.40ab8126.js">
		<link rel="modulepreload" href="/docs/transformers/pr_37537/en/_app/immutable/chunks/paths.abccb988.js">
		<link rel="modulepreload" href="/docs/transformers/pr_37537/en/_app/immutable/entry/app.f503a203.js">
		<link rel="modulepreload" href="/docs/transformers/pr_37537/en/_app/immutable/chunks/index.98837b22.js">
		<link rel="modulepreload" href="/docs/transformers/pr_37537/en/_app/immutable/nodes/0.cbf75854.js">
		<link rel="modulepreload" href="/docs/transformers/pr_37537/en/_app/immutable/chunks/each.e59479a4.js">
		<link rel="modulepreload" href="/docs/transformers/pr_37537/en/_app/immutable/nodes/464.61def05e.js">
		<link rel="modulepreload" href="/docs/transformers/pr_37537/en/_app/immutable/chunks/CodeBlock.8d0c2e8a.js">
		<link rel="modulepreload" href="/docs/transformers/pr_37537/en/_app/immutable/chunks/Heading.7e7e0c5b.js">
		<link rel="modulepreload" href="/docs/transformers/pr_37537/en/_app/immutable/chunks/index.f01015d9.js"><!-- HEAD_svelte-u9bgzb_START --><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;Quark&quot;,&quot;local&quot;:&quot;quark&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Support matrix&quot;,&quot;local&quot;:&quot;support-matrix&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Models on Hugging Face Hub&quot;,&quot;local&quot;:&quot;models-on-hugging-face-hub&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Using Quark models in Transformers&quot;,&quot;local&quot;:&quot;using-quark-models-in-transformers&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2}],&quot;depth&quot;:1}"><!-- HEAD_svelte-u9bgzb_END -->      <p></p>   <h1 class="relative group"><a id="quark" 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="#quark"><span><svg class="" 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>Quark</span></h1> <p data-svelte-h="svelte-f91k68"><a href="https://quark.docs.amd.com/latest/" rel="nofollow">Quark</a> is a deep learning quantization toolkit designed to be agnostic to specific data types, algorithms, and hardware. Different pre-processing strategies, algorithms and data-types can be combined in Quark.</p> <p data-svelte-h="svelte-1bghkp0">The PyTorch support integrated through 🤗 Transformers primarily targets AMD CPUs and GPUs, and is primarily meant to be used for evaluation purposes. For example, it is possible to use <a href="https://github.com/EleutherAI/lm-evaluation-harness" rel="nofollow">lm-evaluation-harness</a> with 🤗 Transformers backend and evaluate a wide range of models quantized through Quark seamlessly.</p> <p data-svelte-h="svelte-5ndvt6">Users interested in Quark can refer to its <a href="https://quark.docs.amd.com/latest/" rel="nofollow">documentation</a> to get started quantizing models and using them in supported open-source libraries!</p> <p data-svelte-h="svelte-8fh62i">Although Quark has its own checkpoint / <a href="https://huggingface.co/amd/Llama-3.1-8B-Instruct-FP8-KV-Quark-test/blob/main/config.json#L26" rel="nofollow">configuration format</a>, the library also supports producing models with a serialization layout compliant with other quantization/runtime implementations (<a href="https://huggingface.co/docs/transformers/quantization/awq" rel="nofollow">AutoAWQ</a>, <a href="https://huggingface.co/docs/transformers/quantization/finegrained_fp8" rel="nofollow">native fp8 in 🤗 Transformers</a>).</p> <p data-svelte-h="svelte-14oxujq">To be able to load Quark quantized models in Transformers, the library first needs to be installed:</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 class="" 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=""><!-- HTML_TAG_START -->pip install amd-quark<!-- HTML_TAG_END --></pre></div>  <h2 class="relative group"><a id="support-matrix" 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="#support-matrix"><span><svg class="" 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>Support matrix</span></h2> <p data-svelte-h="svelte-1vpj1hi">Models quantized through Quark support a large range of features, that can be combined together. All quantized models independently of their configuration can seamlessly be reloaded through <code>PretrainedModel.from_pretrained</code>.</p> <p data-svelte-h="svelte-3upp27">The table below shows a few features supported by Quark:</p> <table data-svelte-h="svelte-vm9870"><thead><tr><th><strong>Feature</strong></th> <th><strong>Supported subset in Quark</strong></th> <th></th></tr></thead> <tbody><tr><td>Data types</td> <td>int8, int4, int2, bfloat16, float16, fp8_e5m2, fp8_e4m3, fp6_e3m2, fp6_e2m3, fp4, OCP MX, MX6, MX9, bfp16</td> <td></td></tr> <tr><td>Pre-quantization transformation</td> <td>SmoothQuant, QuaRot, SpinQuant, AWQ</td> <td></td></tr> <tr><td>Quantization algorithm</td> <td>GPTQ</td> <td></td></tr> <tr><td>Supported operators</td> <td><code>nn.Linear</code>, <code>nn.Conv2d</code>, <code>nn.ConvTranspose2d</code>, <code>nn.Embedding</code>, <code>nn.EmbeddingBag</code></td> <td></td></tr> <tr><td>Granularity</td> <td>per-tensor, per-channel, per-block, per-layer, per-layer type</td> <td></td></tr> <tr><td>KV cache</td> <td>fp8</td> <td></td></tr> <tr><td>Activation calibration</td> <td>MinMax / Percentile / MSE</td> <td></td></tr> <tr><td>Quantization strategy</td> <td>weight-only, static, dynamic, with or without output quantization</td> <td></td></tr></tbody></table>  <h2 class="relative group"><a id="models-on-hugging-face-hub" 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="#models-on-hugging-face-hub"><span><svg class="" 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>Models on Hugging Face Hub</span></h2> <p data-svelte-h="svelte-4nfu69">Public models using Quark native serialization can be found at <a href="https://huggingface.co/models?other=quark" rel="nofollow">https://huggingface.co/models?other=quark</a>.</p> <p data-svelte-h="svelte-1m8qmmj">Although Quark also supports <a href="https://huggingface.co/models?other=fp8" rel="nofollow">models using <code>quant_method=&quot;fp8&quot;</code></a> and <a href="https://huggingface.co/models?other=awq" rel="nofollow">models using <code>quant_method=&quot;awq&quot;</code></a>, Transformers loads these models rather through <a href="https://huggingface.co/docs/transformers/quantization/awq" rel="nofollow">AutoAWQ</a> or uses the <a href="https://huggingface.co/docs/transformers/quantization/finegrained_fp8" rel="nofollow">native fp8 support in 🤗 Transformers</a>.</p>  <h2 class="relative group"><a id="using-quark-models-in-transformers" 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-quark-models-in-transformers"><span><svg class="" 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 Quark models in Transformers</span></h2> <p data-svelte-h="svelte-sp1dmt">Here is an example of how one can load a Quark model in Transformers:</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 class="" 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=""><!-- HTML_TAG_START --><span class="hljs-keyword">from</span> transformers <span class="hljs-keyword">import</span> AutoModelForCausalLM, AutoTokenizer

model_id = <span class="hljs-string">&quot;EmbeddedLLM/Llama-3.1-8B-Instruct-w_fp8_per_channel_sym&quot;</span>
model = AutoModelForCausalLM.from_pretrained(model_id)
model = model.to(<span class="hljs-string">&quot;cuda&quot;</span>)

<span class="hljs-built_in">print</span>(model.model.layers[<span class="hljs-number">0</span>].self_attn.q_proj)
<span class="hljs-comment"># QParamsLinear(</span>
<span class="hljs-comment">#   (weight_quantizer): ScaledRealQuantizer()</span>
<span class="hljs-comment">#   (input_quantizer): ScaledRealQuantizer()</span>
<span class="hljs-comment">#   (output_quantizer): ScaledRealQuantizer()</span>
<span class="hljs-comment"># )</span>

tokenizer = AutoTokenizer.from_pretrained(model_id)
inp = tokenizer(<span class="hljs-string">&quot;Where is a good place to cycle around Tokyo?&quot;</span>, return_tensors=<span class="hljs-string">&quot;pt&quot;</span>)
inp = inp.to(<span class="hljs-string">&quot;cuda&quot;</span>)

res = model.generate(**inp, min_new_tokens=<span class="hljs-number">50</span>, max_new_tokens=<span class="hljs-number">100</span>)

<span class="hljs-built_in">print</span>(tokenizer.batch_decode(res)[<span class="hljs-number">0</span>])
<span class="hljs-comment"># &lt;|begin_of_text|&gt;Where is a good place to cycle around Tokyo? There are several places in Tokyo that are suitable for cycling, depending on your skill level and interests. Here are a few suggestions:</span>
<span class="hljs-comment"># 1. Yoyogi Park: This park is a popular spot for cycling and has a wide, flat path that&#x27;s perfect for beginners. You can also visit the Meiji Shrine, a famous Shinto shrine located in the park.</span>
<span class="hljs-comment"># 2. Imperial Palace East Garden: This beautiful garden has a large, flat path that&#x27;s perfect for cycling. You can also visit the</span><!-- HTML_TAG_END --></pre></div> <a class="!text-gray-400 !no-underline text-sm flex items-center not-prose mt-4" href="https://github.com/huggingface/transformers/blob/main/docs/source/en/quantization/quark.md" target="_blank"><span data-svelte-h="svelte-1kd6by1">&lt;</span> <span data-svelte-h="svelte-x0xyl0">&gt;</span> <span data-svelte-h="svelte-1dajgef"><span class="underline ml-1.5">Update</span> on GitHub</span></a>  <p></p> 
			
			<script>
				{
					__sveltekit_1v5zr0c = {
						assets: "/docs/transformers/pr_37537/en",
						base: "/docs/transformers/pr_37537/en",
						env: {}
					};

					const element = document.currentScript.parentElement;

					const data = [null,null];

					Promise.all([
						import("/docs/transformers/pr_37537/en/_app/immutable/entry/start.f786cb3e.js"),
						import("/docs/transformers/pr_37537/en/_app/immutable/entry/app.f503a203.js")
					]).then(([kit, app]) => {
						kit.start(app, element, {
							node_ids: [0, 464],
							data,
							form: null,
							error: null
						});
					});
				}
			</script>