optimum/pr_2430/en/llm_quantization/usage_guides/quantization.html

<meta charset="utf-8" /><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;Quantization&quot;,&quot;local&quot;:&quot;quantization&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;AutoGPTQ Integration&quot;,&quot;local&quot;:&quot;autogptq-integration&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Requirements&quot;,&quot;local&quot;:&quot;requirements&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;Load and quantize a model&quot;,&quot;local&quot;:&quot;load-and-quantize-a-model&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;Save the model&quot;,&quot;local&quot;:&quot;save-the-model&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;Load quantized weights&quot;,&quot;local&quot;:&quot;load-quantized-weights&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;Exllama kernels for faster inference&quot;,&quot;local&quot;:&quot;exllama-kernels-for-faster-inference&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Fine-tune a quantized model&quot;,&quot;local&quot;:&quot;fine-tune-a-quantized-model&quot;,&quot;sections&quot;:[],&quot;depth&quot;:4}],&quot;depth&quot;:3}],&quot;depth&quot;:2}],&quot;depth&quot;:1}">
		<link href="/docs/optimum/pr_2430/en/_app/immutable/assets/0.e3b0c442.css" rel="modulepreload">
		<link rel="modulepreload" href="/docs/optimum/pr_2430/en/_app/immutable/entry/start.7ca01d50.js">
		<link rel="modulepreload" href="/docs/optimum/pr_2430/en/_app/immutable/chunks/scheduler.0cf4ef2e.js">
		<link rel="modulepreload" href="/docs/optimum/pr_2430/en/_app/immutable/chunks/singletons.ed7b08e2.js">
		<link rel="modulepreload" href="/docs/optimum/pr_2430/en/_app/immutable/chunks/paths.a2bc8d9b.js">
		<link rel="modulepreload" href="/docs/optimum/pr_2430/en/_app/immutable/entry/app.cd433aba.js">
		<link rel="modulepreload" href="/docs/optimum/pr_2430/en/_app/immutable/chunks/preload-helper.4e3372a7.js">
		<link rel="modulepreload" href="/docs/optimum/pr_2430/en/_app/immutable/chunks/index.abc433bf.js">
		<link rel="modulepreload" href="/docs/optimum/pr_2430/en/_app/immutable/nodes/0.40e13f95.js">
		<link rel="modulepreload" href="/docs/optimum/pr_2430/en/_app/immutable/chunks/each.e59479a4.js">
		<link rel="modulepreload" href="/docs/optimum/pr_2430/en/_app/immutable/nodes/8.c80cf74a.js">
		<link rel="modulepreload" href="/docs/optimum/pr_2430/en/_app/immutable/chunks/Tip.15dd1ac8.js">
		<link rel="modulepreload" href="/docs/optimum/pr_2430/en/_app/immutable/chunks/MermaidChart.svelte_svelte_type_style_lang.9537ae3a.js">
		<link rel="modulepreload" href="/docs/optimum/pr_2430/en/_app/immutable/chunks/CodeBlock.12edfa53.js"><!-- HEAD_svelte-u9bgzb_START --><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;Quantization&quot;,&quot;local&quot;:&quot;quantization&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;AutoGPTQ Integration&quot;,&quot;local&quot;:&quot;autogptq-integration&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Requirements&quot;,&quot;local&quot;:&quot;requirements&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;Load and quantize a model&quot;,&quot;local&quot;:&quot;load-and-quantize-a-model&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;Save the model&quot;,&quot;local&quot;:&quot;save-the-model&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;Load quantized weights&quot;,&quot;local&quot;:&quot;load-quantized-weights&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;Exllama kernels for faster inference&quot;,&quot;local&quot;:&quot;exllama-kernels-for-faster-inference&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Fine-tune a quantized model&quot;,&quot;local&quot;:&quot;fine-tune-a-quantized-model&quot;,&quot;sections&quot;:[],&quot;depth&quot;:4}],&quot;depth&quot;:3}],&quot;depth&quot;:2}],&quot;depth&quot;:1}"><!-- HEAD_svelte-u9bgzb_END -->      <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> </div>  <h1 class="relative group"><a id="quantization" 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="#quantization"><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>Quantization</span></h1>  <h2 class="relative group"><a id="autogptq-integration" 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="#autogptq-integration"><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>AutoGPTQ Integration</span></h2> <p data-svelte-h="svelte-ugzxpc">🤗 Optimum collaborated with <a href="https://github.com/PanQiWei/AutoGPTQ" rel="nofollow">AutoGPTQ library</a> to provide a simple API that apply GPTQ quantization on language models. With GPTQ quantization, you can quantize your favorite language model to 8, 4, 3 or even 2 bits. This comes without a big drop of performance and with faster inference speed. This is supported by most GPU hardwares.</p> <p data-svelte-h="svelte-1uwtkai">If you want to quantize 🤗 Transformers models with GPTQ, follow this <a href="https://huggingface.co/docs/transformers/main_classes/quantization" rel="nofollow">documentation</a>.</p> <p data-svelte-h="svelte-1ruo6q2">To learn more about the quantization technique used in GPTQ, please refer to:</p> <ul data-svelte-h="svelte-kq5zxq"><li>the <a href="https://arxiv.org/pdf/2210.17323.pdf" rel="nofollow">GPTQ</a> paper</li> <li>the <a href="https://github.com/PanQiWei/AutoGPTQ" rel="nofollow">AutoGPTQ</a> library used as the backend</li></ul> <p data-svelte-h="svelte-1p4dqk2">Note that the AutoGPTQ library provides more advanced usage (triton backend, fused attention, fused MLP) that are not integrated with Optimum. For now, we leverage only the CUDA kernel for GPTQ.</p>  <h3 class="relative group"><a id="requirements" 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="#requirements"><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>Requirements</span></h3> <p data-svelte-h="svelte-3chw6d">You need to have the following requirements installed to run the code below:</p> <ul data-svelte-h="svelte-bd9wx0"><li><p>AutoGPTQ library:
<code>pip install auto-gptq</code></p></li> <li><p>Optimum library:
<code>pip install --upgrade optimum</code></p></li> <li><p>Install latest <code>transformers</code> library from source:
<code>pip install --upgrade git+https://github.com/huggingface/transformers.git</code></p></li> <li><p>Install latest <code>accelerate</code> library:
<code>pip install --upgrade accelerate</code></p></li></ul>  <h3 class="relative group"><a id="load-and-quantize-a-model" 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="#load-and-quantize-a-model"><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>Load and quantize a model</span></h3> <p data-svelte-h="svelte-qnau83">The <code>GPTQQuantizer</code> class is used to quantize your model. In order to quantize your model, you need to provide a few arguments:</p> <ul data-svelte-h="svelte-1gnh4h2"><li>the number of bits: <code>bits</code></li> <li>the dataset used to calibrate the quantization: <code>dataset</code></li> <li>the model sequence length used to process the dataset: <code>model_seqlen</code></li> <li>the block name to quantize: <code>block_name_to_quantize</code></li></ul> <p data-svelte-h="svelte-1palewv">With 🤗 Transformers integration, you don’t need to pass the <code>block_name_to_quantize</code> and <code>model_seqlen</code> as we can retrieve them. However, for custom model, you need to specify them. Also, make sure that your model is converted to <code>torch.float16</code> before quantization.</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
<span class="hljs-keyword">from</span> optimum.gptq <span class="hljs-keyword">import</span> GPTQQuantizer, load_quantized_model
<span class="hljs-keyword">import</span> torch
model_name = <span class="hljs-string">&quot;facebook/opt-125m&quot;</span>
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.float16)

quantizer = GPTQQuantizer(bits=<span class="hljs-number">4</span>, dataset=<span class="hljs-string">&quot;c4&quot;</span>, block_name_to_quantize = <span class="hljs-string">&quot;model.decoder.layers&quot;</span>, model_seqlen = <span class="hljs-number">2048</span>)
quantized_model = quantizer.quantize_model(model, tokenizer)<!-- HTML_TAG_END --></pre></div> <blockquote class="warning">GPTQ quantization only works for text model for now. Furthermore, the quantization process can take a lot of time depending on one&#39;s hardware (175B model = 4 gpu hours using NVIDIA A100). Please check on the Hugging Face Hub if there is not already a GPTQ quantized version of the model you would like to quantize.</blockquote>  <h3 class="relative group"><a id="save-the-model" 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="#save-the-model"><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>Save the model</span></h3> <p data-svelte-h="svelte-o4ss2s">To save your model, use the save method from <code>GPTQQuantizer</code> class. It will create a folder with your model state dict along with the quantization config.</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 -->save_folder = <span class="hljs-string">&quot;/path/to/save_folder/&quot;</span>
quantizer.save(model,save_folder)<!-- HTML_TAG_END --></pre></div>  <h3 class="relative group"><a id="load-quantized-weights" 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="#load-quantized-weights"><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>Load quantized weights</span></h3> <p data-svelte-h="svelte-1x9uzwt">You can load your quantized weights by using the <code>load_quantized_model()</code> function.
Through the Accelerate library, it is possible to load a model faster with a lower memory usage. The model needs to be initialized using empty weights, with weights loaded as a next step.</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> accelerate <span class="hljs-keyword">import</span> init_empty_weights
<span class="hljs-keyword">with</span> init_empty_weights():
    empty_model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.float16)
empty_model.tie_weights()
quantized_model = load_quantized_model(empty_model, save_folder=save_folder, device_map=<span class="hljs-string">&quot;auto&quot;</span>)<!-- HTML_TAG_END --></pre></div>  <h3 class="relative group"><a id="exllama-kernels-for-faster-inference" 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="#exllama-kernels-for-faster-inference"><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>Exllama kernels for faster inference</span></h3> <p data-svelte-h="svelte-12tew79">With the release of exllamav2 kernels, you can get faster inference speed compared to exllama kernels for 4-bit model. It is activated by default: <code>disable_exllamav2=False</code> in <code>load_quantized_model()</code>. In order to use these kernels, you need to have the entire model on gpus.</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> optimum.gptq <span class="hljs-keyword">import</span> GPTQQuantizer, load_quantized_model
<span class="hljs-keyword">import</span> torch

<span class="hljs-keyword">from</span> accelerate <span class="hljs-keyword">import</span> init_empty_weights
<span class="hljs-keyword">with</span> init_empty_weights():
    empty_model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.float16)
empty_model.tie_weights()
quantized_model = load_quantized_model(empty_model, save_folder=save_folder, device_map=<span class="hljs-string">&quot;auto&quot;</span>)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-m8dfkf">If you wish to use exllama kernels, you will have to change the version by setting <code>exllama_config</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 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> optimum.gptq <span class="hljs-keyword">import</span> GPTQQuantizer, load_quantized_model
<span class="hljs-keyword">import</span> torch

<span class="hljs-keyword">from</span> accelerate <span class="hljs-keyword">import</span> init_empty_weights
<span class="hljs-keyword">with</span> init_empty_weights():
    empty_model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.float16)
empty_model.tie_weights()
quantized_model = load_quantized_model(empty_model, save_folder=save_folder, device_map=<span class="hljs-string">&quot;auto&quot;</span>, exllama_config = {<span class="hljs-string">&quot;version&quot;</span>:<span class="hljs-number">1</span>})<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1wjtw1l">Note that only 4-bit models are supported with exllama/exllamav2 kernels for now. Furthermore, it is recommended to disable exllama/exllamav2 kernels when you are finetuning your model with peft.</p> <p data-svelte-h="svelte-fpw9ll">You can find the benchmark of these kernels <a href="https://github.com/huggingface/optimum/tree/main/tests/benchmark#gptq-benchmark" rel="nofollow">here</a></p>  <h4 class="relative group"><a id="fine-tune-a-quantized-model" 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="#fine-tune-a-quantized-model"><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>Fine-tune a quantized model</span></h4> <p data-svelte-h="svelte-1rghhrt">With the official support of adapters in the Hugging Face ecosystem, you can fine-tune models that have been quantized with GPTQ.
Please have a look at <a href="https://github.com/huggingface/peft" rel="nofollow"><code>peft</code></a> library for more details.</p>  <p></p> 
			
			<script>
				{
					__sveltekit_mx90ez = {
						assets: "/docs/optimum/pr_2430/en",
						base: "/docs/optimum/pr_2430/en",
						env: {}
					};

					const element = document.currentScript.parentElement;

					const data = [null,null];

					Promise.all([
						import("/docs/optimum/pr_2430/en/_app/immutable/entry/start.7ca01d50.js"),
						import("/docs/optimum/pr_2430/en/_app/immutable/entry/app.cd433aba.js")
					]).then(([kit, app]) => {
						kit.start(app, element, {
							node_ids: [0, 8],
							data,
							form: null,
							error: null
						});
					});
				}
			</script>