Buckets:

hf-doc-build
/

doc-dev

Files
xet
hf-doc-build/doc-dev / transformers /pr_45649 /en /quantization /awq.html
HuggingFaceDocBuilder
about 18 hours ago
download
raw
36.3 kB
<meta charset="utf-8" /><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;AWQ&quot;,&quot;local&quot;:&quot;awq&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Fused modules&quot;,&quot;local&quot;:&quot;fused-modules&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;ExLlamaV2&quot;,&quot;local&quot;:&quot;exllamav2&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Resources&quot;,&quot;local&quot;:&quot;resources&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2}],&quot;depth&quot;:1}"/>
		<link href="/docs/transformers/pr_45649/en/_app/immutable/entry/start.Cf6g_lQr.js" rel="modulepreload">
		<link href="/docs/transformers/pr_45649/en/_app/immutable/chunks/bcIanF0A.js" rel="modulepreload">
		<link href="/docs/transformers/pr_45649/en/_app/immutable/chunks/CBySwR1V.js" rel="modulepreload">
		<link href="/docs/transformers/pr_45649/en/_app/immutable/entry/app.Ckn5M_jH.js" rel="modulepreload">
		<link href="/docs/transformers/pr_45649/en/_app/immutable/chunks/Cdb1BouH.js" rel="modulepreload">
		<link href="/docs/transformers/pr_45649/en/_app/immutable/chunks/dUGfULTE.js" rel="modulepreload">
		<link href="/docs/transformers/pr_45649/en/_app/immutable/chunks/D5-p1Lz5.js" rel="modulepreload">
		<link href="/docs/transformers/pr_45649/en/_app/immutable/chunks/DsnmJJEf.js" rel="modulepreload">
		<link href="/docs/transformers/pr_45649/en/_app/immutable/nodes/0.92KZZ6l3.js" rel="modulepreload">
		<link href="/docs/transformers/pr_45649/en/_app/immutable/chunks/Conj1S5Z.js" rel="modulepreload">
		<link href="/docs/transformers/pr_45649/en/_app/immutable/nodes/694.HkUzJ01x.js" rel="modulepreload">
		<link href="/docs/transformers/pr_45649/en/_app/immutable/chunks/DoQ7MZQh.js" rel="modulepreload">
		<link href="/docs/transformers/pr_45649/en/_app/immutable/chunks/uQdIcFiG.js" rel="modulepreload">
		<!--b6axru--><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;AWQ&quot;,&quot;local&quot;:&quot;awq&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Fused modules&quot;,&quot;local&quot;:&quot;fused-modules&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;ExLlamaV2&quot;,&quot;local&quot;:&quot;exllamav2&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Resources&quot;,&quot;local&quot;:&quot;resources&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2}],&quot;depth&quot;:1}"/><!---->
		<link href="/docs/transformers/pr_45649/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: inline-end; margin-inline-start: 10px; display: inline-flex; position: relative; z-index: 10;"><div dir="ltr" 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 dir="auto">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="awq" 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="#awq"><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><!---->AWQ<!----></span></h1><!--]--><!----> <p><a href="https://hf.co/papers/2306.00978" rel="nofollow">Activation-aware Weight Quantization (AWQ)</a> preserves a small fraction of the weights that are important for LLM performance to compress a model to 4-bits with minimal performance degradation.</p> <p>There are several libraries for quantizing models with the AWQ algorithm, such as <a href="https://github.com/mit-han-lab/llm-awq" rel="nofollow">llm-awq</a>, <a href="https://github.com/casper-hansen/AutoAWQ" rel="nofollow">autoawq</a> or <a href="https://huggingface.co/docs/optimum/main/en/intel/optimization_inc" rel="nofollow">optimum-intel</a>. Transformers supports loading models quantized with the llm-awq and autoawq libraries. This guide will show you how to load models quantized with autoawq, but the process is similar for llm-awq quantized models.</p> <p>Run the command below to install autoawq</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 "><!---->pip install autoawq<!----></pre></div><!----> <blockquote class="warning"><p>AutoAWQ downgrades Transformers to version 4.47.1. If you want to do inference with AutoAWQ, you may need to reinstall your Transformers’ version after installing AutoAWQ.</p></blockquote> <p>Identify an AWQ-quantized model by checking the <code>quant_method</code> key in the models <a href="https://huggingface.co/TheBloke/zephyr-7B-alpha-AWQ/blob/main/config.json" rel="nofollow">config.json</a> file.</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-json "><!----><span class="hljs-punctuation">{</span>
  <span class="hljs-attr">&quot;_name_or_path&quot;</span><span class="hljs-punctuation">:</span> <span class="hljs-string">&quot;/workspace/process/huggingfaceh4_zephyr-7b-alpha/source&quot;</span><span class="hljs-punctuation">,</span>
  <span class="hljs-attr">&quot;architectures&quot;</span><span class="hljs-punctuation">:</span> <span class="hljs-punctuation">[</span>
    <span class="hljs-string">&quot;MistralForCausalLM&quot;</span>
  <span class="hljs-punctuation">]</span><span class="hljs-punctuation">,</span>
  ...
  ...
  ...
  <span class="hljs-attr">&quot;quantization_config&quot;</span><span class="hljs-punctuation">:</span> <span class="hljs-punctuation">{</span>
    <span class="hljs-attr">&quot;quant_method&quot;</span><span class="hljs-punctuation">:</span> <span class="hljs-string">&quot;awq&quot;</span><span class="hljs-punctuation">,</span>
    <span class="hljs-attr">&quot;zero_point&quot;</span><span class="hljs-punctuation">:</span> <span class="hljs-literal"><span class="hljs-keyword">true</span></span><span class="hljs-punctuation">,</span>
    <span class="hljs-attr">&quot;group_size&quot;</span><span class="hljs-punctuation">:</span> <span class="hljs-number">128</span><span class="hljs-punctuation">,</span>
    <span class="hljs-attr">&quot;bits&quot;</span><span class="hljs-punctuation">:</span> <span class="hljs-number">4</span><span class="hljs-punctuation">,</span>
    <span class="hljs-attr">&quot;version&quot;</span><span class="hljs-punctuation">:</span> <span class="hljs-string">&quot;gemm&quot;</span>
  <span class="hljs-punctuation">}</span>
<span class="hljs-punctuation">}</span><!----></pre></div><!----> <p>Load the AWQ-quantized model with <a href="/docs/transformers/pr_45649/en/main_classes/model#transformers.PreTrainedModel.from_pretrained">from_pretrained()</a>. This automatically sets the other weights to fp16 by default for performance reasons. Use the <code>dtype</code> parameter to load these other weights in a different format.</p> <p>If the model is loaded on the CPU, use the <code>device_map</code> parameter to move it to an accelerator.</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> transformers <span class="hljs-keyword">import</span> AutoModelForCausalLM, AutoTokenizer
<span class="hljs-keyword">from</span> accelerate <span class="hljs-keyword">import</span> Accelerator
<span class="hljs-keyword">import</span> torch

device = Accelerator().device

model = AutoModelForCausalLM.from_pretrained(
  <span class="hljs-string">&quot;TheBloke/zephyr-7B-alpha-AWQ&quot;</span>,
  dtype=torch.float32,
  device_map=device
)<!----></pre></div><!----> <p>Use <code>attn_implementation</code> to enable <a href="../kernel_doc/loading_kernels#attention-kernels">FlashAttention2</a> to further accelerate inference.</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> transformers <span class="hljs-keyword">import</span> AutoModelForCausalLM, AutoTokenizer
<span class="hljs-keyword">from</span> accelerate <span class="hljs-keyword">import</span> Accelerator

model = AutoModelForCausalLM.from_pretrained(
  <span class="hljs-string">&quot;TheBloke/zephyr-7B-alpha-AWQ&quot;</span>,
  attn_implementation=<span class="hljs-string">&quot;flash_attention_2&quot;</span>,
  device_map=Accelerator().device
)<!----></pre></div><!----> <!--[1--><h2 class="relative group"><a id="fused-modules" 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="#fused-modules"><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><!---->Fused modules<!----></span></h2><!--]--><!----> <p>Fused modules offer improved accuracy and performance. They are supported out-of-the-box for AWQ modules for <a href="https://huggingface.co/meta-llama" rel="nofollow">Llama</a> and <a href="https://huggingface.co/mistralai/Mistral-7B-v0.1" rel="nofollow">Mistral</a> architectures, but you can also fuse AWQ modules for unsupported architectures.</p> <blockquote class="warning"><p>Fused modules cannot be combined with other optimization techniques such as FlashAttention2.</p></blockquote> <div class="flex space-x-2 items-center my-1.5 mr-8 h-7 !pl-0 -mx-3 md:mx-0"><!--[--><div class="flex items-center border rounded-lg px-1.5 py-1 leading-none select-none text-smd border-gray-800 bg-black dark:bg-gray-700 text-white">supported architectures</div><div class="flex items-center border rounded-lg px-1.5 py-1 leading-none select-none text-smd text-gray-500 cursor-pointer opacity-90 hover:text-gray-700 dark:hover:text-gray-200 hover:shadow-sm">unsupported architectures</div><!--]--></div> <div class="language-select"><div><p>Create an <a href="/docs/transformers/pr_45649/en/main_classes/quantization#transformers.AwqConfig">AwqConfig</a> and set the parameters <code>fuse_max_seq_len</code> and <code>do_fuse=True</code> to enable fused modules. The <code>fuse_max_seq_len</code> parameter is the total sequence length and it should include the context length and the expected generation length. Set it to a larger value to be safe.</p> <p>The example below fuses the AWQ modules of the <a href="https://huggingface.co/TheBloke/Mistral-7B-OpenOrca-AWQ" rel="nofollow">TheBloke/Mistral-7B-OpenOrca-AWQ</a> model.</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> torch
<span class="hljs-keyword">from</span> transformers <span class="hljs-keyword">import</span> AwqConfig, AutoModelForCausalLM

quantization_config = AwqConfig(
    bits=<span class="hljs-number">4</span>,
    fuse_max_seq_len=<span class="hljs-number">512</span>,
    do_fuse=<span class="hljs-literal">True</span>,
)
model = AutoModelForCausalLM.from_pretrained(
  <span class="hljs-string">&quot;TheBloke/Mistral-7B-OpenOrca-AWQ&quot;</span>,
  quantization_config=quantization_config
).to(<span class="hljs-number">0</span>)<!----></pre></div><!----> <p>The <a href="https://huggingface.co/TheBloke/Mistral-7B-OpenOrca-AWQ" rel="nofollow">TheBloke/Mistral-7B-OpenOrca-AWQ</a> model was benchmarked with <code>batch_size=1</code> with and without fused modules.</p> <figcaption class="text-center text-gray-500 text-lg">Unfused module</figcaption> <table><thead><tr><th align="right">Batch Size</th><th align="right">Prefill Length</th><th align="right">Decode Length</th><th align="right">Prefill tokens/s</th><th align="right">Decode tokens/s</th><th align="left">Memory (VRAM)</th></tr></thead><tbody><tr><td align="right">1</td><td align="right">32</td><td align="right">32</td><td align="right">60.0984</td><td align="right">38.4537</td><td align="left">4.50 GB (5.68%)</td></tr><tr><td align="right">1</td><td align="right">64</td><td align="right">64</td><td align="right">1333.67</td><td align="right">31.6604</td><td align="left">4.50 GB (5.68%)</td></tr><tr><td align="right">1</td><td align="right">128</td><td align="right">128</td><td align="right">2434.06</td><td align="right">31.6272</td><td align="left">4.50 GB (5.68%)</td></tr><tr><td align="right">1</td><td align="right">256</td><td align="right">256</td><td align="right">3072.26</td><td align="right">38.1731</td><td align="left">4.50 GB (5.68%)</td></tr><tr><td align="right">1</td><td align="right">512</td><td align="right">512</td><td align="right">3184.74</td><td align="right">31.6819</td><td align="left">4.59 GB (5.80%)</td></tr><tr><td align="right">1</td><td align="right">1024</td><td align="right">1024</td><td align="right">3148.18</td><td align="right">36.8031</td><td align="left">4.81 GB (6.07%)</td></tr><tr><td align="right">1</td><td align="right">2048</td><td align="right">2048</td><td align="right">2927.33</td><td align="right">35.2676</td><td align="left">5.73 GB (7.23%)</td></tr></tbody></table> <figcaption class="text-center text-gray-500 text-lg">Fused module</figcaption> <table><thead><tr><th align="right">Batch Size</th><th align="right">Prefill Length</th><th align="right">Decode Length</th><th align="right">Prefill tokens/s</th><th align="right">Decode tokens/s</th><th align="left">Memory (VRAM)</th></tr></thead><tbody><tr><td align="right">1</td><td align="right">32</td><td align="right">32</td><td align="right">81.4899</td><td align="right">80.2569</td><td align="left">4.00 GB (5.05%)</td></tr><tr><td align="right">1</td><td align="right">64</td><td align="right">64</td><td align="right">1756.1</td><td align="right">106.26</td><td align="left">4.00 GB (5.05%)</td></tr><tr><td align="right">1</td><td align="right">128</td><td align="right">128</td><td align="right">2479.32</td><td align="right">105.631</td><td align="left">4.00 GB (5.06%)</td></tr><tr><td align="right">1</td><td align="right">256</td><td align="right">256</td><td align="right">1813.6</td><td align="right">85.7485</td><td align="left">4.01 GB (5.06%)</td></tr><tr><td align="right">1</td><td align="right">512</td><td align="right">512</td><td align="right">2848.9</td><td align="right">97.701</td><td align="left">4.11 GB (5.19%)</td></tr><tr><td align="right">1</td><td align="right">1024</td><td align="right">1024</td><td align="right">3044.35</td><td align="right">87.7323</td><td align="left">4.41 GB (5.57%)</td></tr><tr><td align="right">1</td><td align="right">2048</td><td align="right">2048</td><td align="right">2715.11</td><td align="right">89.4709</td><td align="left">5.57 GB (7.04%)</td></tr></tbody></table> <p>The speed and throughput of fused and unfused modules were also tested with the <a href="https://github.com/huggingface/optimum-benchmark" rel="nofollow">optimum-benchmark</a> library.</p> <div class="flex gap-4"><div><img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/quantization/fused_forward_memory_plot.png" alt="generate throughput per batch size"/> <figcaption class="mt-2 text-center text-sm text-gray-500">forward peak memory/batch size</figcaption></div> <div><img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/quantization/fused_generate_throughput_plot.png" alt="forward latency per batch size"/> <figcaption class="mt-2 text-center text-sm text-gray-500">generate throughput/batch size</figcaption></div></div><!----></div><!----> <div hidden=""><p>For architectures that don’t support fused modules, create an <a href="/docs/transformers/pr_45649/en/main_classes/quantization#transformers.AwqConfig">AwqConfig</a> and define a custom fusing mapping in <code>modules_to_fuse</code> to determine which modules need to be fused.</p> <p>The example below fuses the AWQ modules of the <a href="https://huggingface.co/TheBloke/Yi-34B-AWQ" rel="nofollow">TheBloke/Yi-34B-AWQ</a> model.</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> torch
<span class="hljs-keyword">from</span> transformers <span class="hljs-keyword">import</span> AwqConfig, AutoModelForCausalLM

quantization_config = AwqConfig(
    bits=<span class="hljs-number">4</span>,
    fuse_max_seq_len=<span class="hljs-number">512</span>,
    modules_to_fuse={
        <span class="hljs-string">&quot;attention&quot;</span>: [<span class="hljs-string">&quot;q_proj&quot;</span>, <span class="hljs-string">&quot;k_proj&quot;</span>, <span class="hljs-string">&quot;v_proj&quot;</span>, <span class="hljs-string">&quot;o_proj&quot;</span>],
        <span class="hljs-string">&quot;layernorm&quot;</span>: [<span class="hljs-string">&quot;ln1&quot;</span>, <span class="hljs-string">&quot;ln2&quot;</span>, <span class="hljs-string">&quot;norm&quot;</span>],
        <span class="hljs-string">&quot;mlp&quot;</span>: [<span class="hljs-string">&quot;gate_proj&quot;</span>, <span class="hljs-string">&quot;up_proj&quot;</span>, <span class="hljs-string">&quot;down_proj&quot;</span>],
        <span class="hljs-string">&quot;use_alibi&quot;</span>: <span class="hljs-literal">False</span>,
        <span class="hljs-string">&quot;num_attention_heads&quot;</span>: <span class="hljs-number">56</span>,
        <span class="hljs-string">&quot;num_key_value_heads&quot;</span>: <span class="hljs-number">8</span>,
        <span class="hljs-string">&quot;hidden_size&quot;</span>: <span class="hljs-number">7168</span>
    }
)

model = AutoModelForCausalLM.from_pretrained(
  <span class="hljs-string">&quot;TheBloke/Yi-34B-AWQ&quot;</span>,
  quantization_config=quantization_config
).to(<span class="hljs-number">0</span>)<!----></pre></div><!----> <p>The parameter <code>modules_to_fuse</code> should include the following keys.</p> <ul><li><p><code>"attention"</code>: The names of the attention layers to fuse in the following order: query, key, value and output projection layer. If you don’t want to fuse these layers, pass an empty list.</p></li> <li><p><code>"layernorm"</code>: The names of all the LayerNorm layers you want to replace with a custom fused LayerNorm. If you don’t want to fuse these layers, pass an empty list.</p></li> <li><p><code>"mlp"</code>: The names of the MLP layers you want to fuse into a single MLP layer in the order: (gate (dense, layer, post-attention) / up / down layers).</p></li> <li><p><code>"use_alibi"</code>: If your model uses ALiBi positional embedding.</p></li> <li><p><code>"num_attention_heads"</code>: The number of attention heads.</p></li> <li><p><code>"num_key_value_heads"</code>: The number of key value heads that should be used to implement Grouped Query Attention (GQA).</p> <table><thead><tr><th>parameter value</th><th>attention</th></tr></thead><tbody><tr><td><code>num_key_value_heads=num_attention_heads</code></td><td>Multi-Head Attention</td></tr><tr><td><code>num_key_value_heads=1</code></td><td>Multi-Query Attention</td></tr><tr><td><code>num_key_value_heads=...</code></td><td>Grouped Query Attention</td></tr></tbody></table></li> <li><p><code>"hidden_size"</code>: The dimension of the hidden representations.</p></li></ul><!----></div><!----><!----></div><!----> <!--[1--><h2 class="relative group"><a id="exllamav2" 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="#exllamav2"><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><!---->ExLlamaV2<!----></span></h2><!--]--><!----> <p><a href="https://github.com/turboderp/exllamav2" rel="nofollow">ExLlamaV2</a> kernels support faster prefill and decoding. Run the command below to install the latest version of autoawq with ExLlamaV2 support.</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 "><!---->pip install git+https://github.com/casper-hansen/AutoAWQ.git<!----></pre></div><!----> <p>Set <code>version="exllama"</code> in <a href="/docs/transformers/pr_45649/en/main_classes/quantization#transformers.AwqConfig">AwqConfig</a> to enable ExLlamaV2 kernels.</p> <blockquote class="tip"><p>ExLlamaV2 is supported on AMD GPUs.</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">import</span> torch
<span class="hljs-keyword">from</span> transformers <span class="hljs-keyword">import</span> AutoModelForCausalLM, AutoTokenizer, AwqConfig

quantization_config = AwqConfig(version=<span class="hljs-string">&quot;exllama&quot;</span>)

model = AutoModelForCausalLM.from_pretrained(
    <span class="hljs-string">&quot;TheBloke/Mistral-7B-Instruct-v0.1-AWQ&quot;</span>,
    quantization_config=quantization_config,
    device_map=<span class="hljs-string">&quot;auto&quot;</span>,
)<!----></pre></div><!----> <!--[1--><h2 class="relative group"><a id="resources" 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="#resources"><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><!---->Resources<!----></span></h2><!--]--><!----> <p>Run the AWQ demo <a href="https://colab.research.google.com/drive/1HzZH89yAXJaZgwJDhQj9LqSBux932BvY#scrollTo=Wwsg6nCwoThm" rel="nofollow">notebook</a> for more examples of how to quantize a model, push a quantized model to the Hub, and more.</p> <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/awq.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_j5svul = {
						base: "/docs/transformers/pr_45649/en",
						assets: "/docs/transformers/pr_45649/en"
					};
					const element = document.currentScript.parentElement;
					Promise.all([
						import("/docs/transformers/pr_45649/en/_app/immutable/entry/start.Cf6g_lQr.js"),
						import("/docs/transformers/pr_45649/en/_app/immutable/entry/app.Ckn5M_jH.js")
					]).then(([kit, app]) => {
						kit.start(app, element, {
							node_ids: [0, 694],
							data: [null,null],
							form: null,
							error: null
						});
					});
				}
			</script>
		
Xet Storage Details

Size:: 36.3 kB
Xet hash:: e69f52e7b0a0e3d61dbecbddd31b01d728d73273b9d33eb84e0a2b6bae67b7e0
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.