transformers/pr_36356/en/perf_infer_gpu_multi.html

<meta charset="utf-8" /><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;Multi-GPU inference&quot;,&quot;local&quot;:&quot;multi-gpu-inference&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Expected speedups&quot;,&quot;local&quot;:&quot;expected-speedups&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3}],&quot;depth&quot;:1}">
		<link href="/docs/transformers/pr_36356/en/_app/immutable/assets/0.e3b0c442.css" rel="modulepreload">
		<link rel="modulepreload" href="/docs/transformers/pr_36356/en/_app/immutable/entry/start.eb3eb862.js">
		<link rel="modulepreload" href="/docs/transformers/pr_36356/en/_app/immutable/chunks/scheduler.bdbef820.js">
		<link rel="modulepreload" href="/docs/transformers/pr_36356/en/_app/immutable/chunks/singletons.230895fd.js">
		<link rel="modulepreload" href="/docs/transformers/pr_36356/en/_app/immutable/chunks/index.8a885b74.js">
		<link rel="modulepreload" href="/docs/transformers/pr_36356/en/_app/immutable/chunks/paths.98b303d2.js">
		<link rel="modulepreload" href="/docs/transformers/pr_36356/en/_app/immutable/entry/app.4f40fe64.js">
		<link rel="modulepreload" href="/docs/transformers/pr_36356/en/_app/immutable/chunks/index.33f81d56.js">
		<link rel="modulepreload" href="/docs/transformers/pr_36356/en/_app/immutable/nodes/0.c9596ba7.js">
		<link rel="modulepreload" href="/docs/transformers/pr_36356/en/_app/immutable/chunks/each.e59479a4.js">
		<link rel="modulepreload" href="/docs/transformers/pr_36356/en/_app/immutable/nodes/408.3607dd12.js">
		<link rel="modulepreload" href="/docs/transformers/pr_36356/en/_app/immutable/chunks/CodeBlock.362b34a4.js">
		<link rel="modulepreload" href="/docs/transformers/pr_36356/en/_app/immutable/chunks/EditOnGithub.a9246e21.js"><!-- HEAD_svelte-u9bgzb_START --><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;Multi-GPU inference&quot;,&quot;local&quot;:&quot;multi-gpu-inference&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Expected speedups&quot;,&quot;local&quot;:&quot;expected-speedups&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3}],&quot;depth&quot;:1}"><!-- HEAD_svelte-u9bgzb_END -->      <p></p>   <h1 class="relative group"><a id="multi-gpu-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="#multi-gpu-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>Multi-GPU inference</span></h1> <p data-svelte-h="svelte-18sueh2">Built-in Tensor Parallelism (TP) is now available with certain models using PyTorch. Tensor parallelism shards a model onto multiple GPUs, enabling larger model sizes, and parallelizes computations such as matrix multiplication.</p> <p data-svelte-h="svelte-8b4tmf">To enable tensor parallel, pass the argument <code>tp_plan=&quot;auto&quot;</code> to <a href="/docs/transformers/pr_36356/en/model_doc/auto#transformers.AutoModel.from_pretrained">from_pretrained()</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 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">import</span> os
<span class="hljs-keyword">import</span> torch
<span class="hljs-keyword">from</span> transformers <span class="hljs-keyword">import</span> AutoModelForCausalLM, AutoTokenizer

model_id = <span class="hljs-string">&quot;meta-llama/Meta-Llama-3-8B-Instruct&quot;</span>

<span class="hljs-comment"># Initialize distributed</span>
rank = <span class="hljs-built_in">int</span>(os.environ[<span class="hljs-string">&quot;RANK&quot;</span>])
device = torch.device(<span class="hljs-string">f&quot;cuda:<span class="hljs-subst">{rank}</span>&quot;</span>)
torch.cuda.set_device(device)
torch.distributed.init_process_group(<span class="hljs-string">&quot;nccl&quot;</span>, device_id=device)

<span class="hljs-comment"># Retrieve tensor parallel model</span>
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    tp_plan=<span class="hljs-string">&quot;auto&quot;</span>,
)

<span class="hljs-comment"># Prepare input tokens</span>
tokenizer = AutoTokenizer.from_pretrained(model_id)
prompt = <span class="hljs-string">&quot;Can I help&quot;</span>
inputs = tokenizer(prompt, return_tensors=<span class="hljs-string">&quot;pt&quot;</span>).input_ids.to(device)

<span class="hljs-comment"># Distributed run</span>
outputs = model(inputs)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-ey9nzh">You can use <code>torchrun</code> to launch the above script with multiple processes, each mapping to a GPU:</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 -->torchrun --nproc-per-<span class="hljs-keyword">node</span> <span class="hljs-title">4</span> demo.py<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-188qg3a">PyTorch tensor parallel is currently supported for the following models:</p> <ul data-svelte-h="svelte-1ghi9eu"><li><a href="https://huggingface.co/docs/transformers/model_doc/llama#transformers.LlamaModel" rel="nofollow">Llama</a></li> <li><a href="https://huggingface.co/docs/transformers/en/model_doc/gemma" rel="nofollow">Gemma</a>, <a href="https://huggingface.co/docs/transformers/en/model_doc/gemma2" rel="nofollow">Gemma2</a></li> <li><a href="https://huggingface.co/docs/transformers/en/model_doc/granite" rel="nofollow">Granite</a></li> <li><a href="https://huggingface.co/docs/transformers/en/model_doc/mistral" rel="nofollow">Mistral</a></li> <li><a href="https://huggingface.co/docs/transformers/en/model_doc/qwen2" rel="nofollow">Qwen2</a>, <a href="https://huggingface.co/docs/transformers/en/model_doc/qwen2_moe" rel="nofollow">Qwen2MoE</a>, <a href="https://huggingface.co/docs/transformers/v4.48.0/en/model_doc/qwen2_vl" rel="nofollow">Qwen2-VL</a></li> <li><a href="https://huggingface.co/docs/transformers/en/model_doc/starcoder2" rel="nofollow">Starcoder2</a></li> <li><a href="https://huggingface.co/docs/transformers/en/model_doc/cohere" rel="nofollow">Cohere</a>, <a href="https://huggingface.co/docs/transformers/en/model_doc/cohere2" rel="nofollow">Cohere2</a></li> <li><a href="https://huggingface.co/docs/transformers/en/model_doc/glm" rel="nofollow">GLM</a></li> <li><a href="https://huggingface.co/docs/transformers/en/model_doc/mixtral" rel="nofollow">Mixtral</a></li> <li><a href="https://huggingface.co/docs/transformers/en/model_doc/olmo" rel="nofollow">OLMo</a>, <a href="https://huggingface.co/docs/transformers/en/model_doc/olmo2" rel="nofollow">OLMo2</a></li> <li><a href="https://huggingface.co/docs/transformers/en/model_doc/phi" rel="nofollow">Phi</a>, <a href="https://huggingface.co/docs/transformers/en/model_doc/phi3" rel="nofollow">Phi-3</a></li></ul> <p data-svelte-h="svelte-1s0fi8f">You can request to add tensor parallel support for another model by opening a GitHub Issue or Pull Request.</p>  <h3 class="relative group"><a id="expected-speedups" 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="#expected-speedups"><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>Expected speedups</span></h3> <p data-svelte-h="svelte-yctpd2">You can benefit from considerable speedups for inference, especially for inputs with large batch size or long sequences.</p> <p data-svelte-h="svelte-192kbjl">For a single forward pass on <a href="https://huggingface.co/docs/transformers/model_doc/llama#transformers.LlamaModel" rel="nofollow">Llama</a> with a sequence length of 512 and various batch sizes, the expected speedup is as follows:</p> <div style="text-align: center" data-svelte-h="svelte-lg3kfi"><img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/Meta-Llama-3-8B-Instruct%2C%20seqlen%20%3D%20512%2C%20python%2C%20w_%20compile.png"></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/perf_infer_gpu_multi.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_gc6ivo = {
						assets: "/docs/transformers/pr_36356/en",
						base: "/docs/transformers/pr_36356/en",
						env: {}
					};

					const element = document.currentScript.parentElement;

					const data = [null,null];

					Promise.all([
						import("/docs/transformers/pr_36356/en/_app/immutable/entry/start.eb3eb862.js"),
						import("/docs/transformers/pr_36356/en/_app/immutable/entry/app.4f40fe64.js")
					]).then(([kit, app]) => {
						kit.start(app, element, {
							node_ids: [0, 408],
							data,
							form: null,
							error: null
						});
					});
				}
			</script>