Buckets:

hf-doc-build/doc-dev / accelerate /pr_4021 /en /usage_guides /distributed_inference.html
download
raw
40 kB
<meta charset="utf-8" /><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;Distributed inference&quot;,&quot;local&quot;:&quot;distributed-inference&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Sending chunks of a batch automatically to each loaded model&quot;,&quot;local&quot;:&quot;sending-chunks-of-a-batch-automatically-to-each-loaded-model&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Memory-efficient pipeline parallelism (experimental)&quot;,&quot;local&quot;:&quot;memory-efficient-pipeline-parallelism-experimental&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2}],&quot;depth&quot;:1}">
<link href="/docs/accelerate/pr_4021/en/_app/immutable/assets/0.e3b0c442.css" rel="modulepreload">
<link rel="modulepreload" href="/docs/accelerate/pr_4021/en/_app/immutable/entry/start.8a49e72b.js">
<link rel="modulepreload" href="/docs/accelerate/pr_4021/en/_app/immutable/chunks/scheduler.b9285784.js">
<link rel="modulepreload" href="/docs/accelerate/pr_4021/en/_app/immutable/chunks/singletons.7547c222.js">
<link rel="modulepreload" href="/docs/accelerate/pr_4021/en/_app/immutable/chunks/index.6d423e5c.js">
<link rel="modulepreload" href="/docs/accelerate/pr_4021/en/_app/immutable/chunks/paths.d42c9205.js">
<link rel="modulepreload" href="/docs/accelerate/pr_4021/en/_app/immutable/entry/app.1df4d18e.js">
<link rel="modulepreload" href="/docs/accelerate/pr_4021/en/_app/immutable/chunks/preload-helper.b0bd19d1.js">
<link rel="modulepreload" href="/docs/accelerate/pr_4021/en/_app/immutable/chunks/index.26bc89a1.js">
<link rel="modulepreload" href="/docs/accelerate/pr_4021/en/_app/immutable/nodes/0.0e7c56e8.js">
<link rel="modulepreload" href="/docs/accelerate/pr_4021/en/_app/immutable/chunks/each.e59479a4.js">
<link rel="modulepreload" href="/docs/accelerate/pr_4021/en/_app/immutable/nodes/44.22ccf593.js">
<link rel="modulepreload" href="/docs/accelerate/pr_4021/en/_app/immutable/chunks/Tip.e4eba3d6.js">
<link rel="modulepreload" href="/docs/accelerate/pr_4021/en/_app/immutable/chunks/MermaidChart.svelte_svelte_type_style_lang.7a0ae628.js">
<link rel="modulepreload" href="/docs/accelerate/pr_4021/en/_app/immutable/chunks/CodeBlock.844ff9c3.js"><!-- HEAD_svelte-u9bgzb_START --><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;Distributed inference&quot;,&quot;local&quot;:&quot;distributed-inference&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Sending chunks of a batch automatically to each loaded model&quot;,&quot;local&quot;:&quot;sending-chunks-of-a-batch-automatically-to-each-loaded-model&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Memory-efficient pipeline parallelism (experimental)&quot;,&quot;local&quot;:&quot;memory-efficient-pipeline-parallelism-experimental&quot;,&quot;sections&quot;:[],&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="distributed-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="#distributed-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>Distributed inference</span></h1> <p data-svelte-h="svelte-pzys4w">Distributed inference can fall into three brackets:</p> <ol data-svelte-h="svelte-auhvyh"><li>Loading an entire model onto each GPU and sending chunks of a batch through each GPU’s model copy at a time</li> <li>Loading parts of a model onto each GPU and processing a single input at one time</li> <li>Loading parts of a model onto each GPU and using what is called scheduled Pipeline Parallelism to combine the two prior techniques.</li></ol> <p data-svelte-h="svelte-105ktyk">We’re going to go through the first and the last bracket, showcasing how to do each as they are more realistic scenarios.</p> <h2 class="relative group"><a id="sending-chunks-of-a-batch-automatically-to-each-loaded-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="#sending-chunks-of-a-batch-automatically-to-each-loaded-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>Sending chunks of a batch automatically to each loaded model</span></h2> <p data-svelte-h="svelte-1ojc10y">This is the most memory-intensive solution, as it requires each GPU to keep a full copy of the model in memory at a given time.</p> <p data-svelte-h="svelte-1qi053h">Normally when doing this, users send the model to a specific device to load it from the CPU, and then move each prompt to a different device.</p> <p data-svelte-h="svelte-rhvic4">A basic pipeline using the <code>diffusers</code> library might look something like so:</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> torch
<span class="hljs-keyword">import</span> torch.distributed <span class="hljs-keyword">as</span> dist
<span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> DiffusionPipeline
pipe = DiffusionPipeline.from_pretrained(<span class="hljs-string">&quot;runwayml/stable-diffusion-v1-5&quot;</span>, torch_dtype=torch.float16)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1dzspg1">Followed then by performing inference based on the specific prompt:</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">def</span> <span class="hljs-title function_">run_inference</span>(<span class="hljs-params">rank, world_size</span>):
dist.init_process_group(<span class="hljs-string">&quot;nccl&quot;</span>, rank=rank, world_size=world_size)
pipe.to(rank)
<span class="hljs-keyword">if</span> torch.distributed.get_rank() == <span class="hljs-number">0</span>:
prompt = <span class="hljs-string">&quot;a dog&quot;</span>
<span class="hljs-keyword">elif</span> torch.distributed.get_rank() == <span class="hljs-number">1</span>:
prompt = <span class="hljs-string">&quot;a cat&quot;</span>
result = pipe(prompt).images[<span class="hljs-number">0</span>]
result.save(<span class="hljs-string">f&quot;result_<span class="hljs-subst">{rank}</span>.png&quot;</span>)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1ov43u6">One will notice how we have to check the rank to know what prompt to send, which can be a bit tedious.</p> <p data-svelte-h="svelte-1n5a8vc">A user might then also think that with Accelerate, using the <code>Accelerator</code> to prepare a dataloader for such a task might also be
a simple way to manage this. (To learn more, check out the relevant section in the <a href="../quicktour#distributed-evaluation">Quick Tour</a>)</p> <p data-svelte-h="svelte-12rt8sz">Can it manage it? Yes. Does it add unneeded extra code however: also yes.</p> <p data-svelte-h="svelte-gw34xi">With Accelerate, we can simplify this process by using the <a href="/docs/accelerate/pr_4021/en/package_reference/accelerator#accelerate.Accelerator.split_between_processes">Accelerator.split_between_processes()</a> context manager (which also exists in <code>PartialState</code> and <code>AcceleratorState</code>).
This function will automatically split whatever data you pass to it (be it a prompt, a set of tensors, a dictionary of the prior data, etc.) across all the processes (with a potential
to be padded) for you to use right away.</p> <p data-svelte-h="svelte-8uk323">Let’s rewrite the above example using this context manager:</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> torch
<span class="hljs-keyword">from</span> accelerate <span class="hljs-keyword">import</span> PartialState <span class="hljs-comment"># Can also be Accelerator or AcceleratorState</span>
<span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> DiffusionPipeline
pipe = DiffusionPipeline.from_pretrained(<span class="hljs-string">&quot;runwayml/stable-diffusion-v1-5&quot;</span>, torch_dtype=torch.float16)
distributed_state = PartialState()
pipe.to(distributed_state.device)
<span class="hljs-comment"># Assume two processes</span>
<span class="hljs-keyword">with</span> distributed_state.split_between_processes([<span class="hljs-string">&quot;a dog&quot;</span>, <span class="hljs-string">&quot;a cat&quot;</span>]) <span class="hljs-keyword">as</span> prompt:
result = pipe(prompt).images[<span class="hljs-number">0</span>]
result.save(<span class="hljs-string">f&quot;result_<span class="hljs-subst">{distributed_state.process_index}</span>.png&quot;</span>)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-k1tgc3">And then to launch the code, we can use the Accelerate:</p> <p data-svelte-h="svelte-glszdf">If you have generated a config file to be used using <code>accelerate 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 -->accelerate launch distributed_inference.py<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1beq8se">If you have a specific config file you want to use:</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 -->accelerate launch --config_file my_config.json distributed_inference.py<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-52p69u">Or if don’t want to make any config files and launch on two GPUs:</p> <blockquote data-svelte-h="svelte-ix7ij8"><p>Note: You will get some warnings about values being guessed based on your system. To remove these you can do <code>accelerate config default</code> or go through <code>accelerate config</code> to create a config file.</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 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 -->accelerate launch --num_processes 2 distributed_inference.py<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-ui4crb">We’ve now reduced the boilerplate code needed to split this data to a few lines of code quite easily.</p> <p data-svelte-h="svelte-dyqt5a">But what if we have an odd distribution of prompts to GPUs? For example, what if we have 3 prompts, but only 2 GPUs?</p> <p data-svelte-h="svelte-1o56krx">Under the context manager, the first GPU would receive the first two prompts and the second GPU the third, ensuring that
all prompts are split and no overhead is needed.</p> <p data-svelte-h="svelte-1y9mak3"><em>However</em>, what if we then wanted to do something with the results of <em>all the GPUs</em>? (Say gather them all and perform some kind of post processing)
You can pass in <code>apply_padding=True</code> to ensure that the lists of prompts are padded to the same length, with extra data being taken
from the last sample. This way all GPUs will have the same number of prompts, and you can then gather the results.</p> <blockquote class="tip"><p data-svelte-h="svelte-rtowy2">This is only needed when trying to perform an action such as gathering the results, where the data on each device
needs to be the same length. Basic inference does not require this.</p></blockquote> <p data-svelte-h="svelte-4vay6o">For instance:</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> torch
<span class="hljs-keyword">from</span> accelerate <span class="hljs-keyword">import</span> PartialState <span class="hljs-comment"># Can also be Accelerator or AcceleratorState</span>
<span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> DiffusionPipeline
pipe = DiffusionPipeline.from_pretrained(<span class="hljs-string">&quot;runwayml/stable-diffusion-v1-5&quot;</span>, torch_dtype=torch.float16)
distributed_state = PartialState()
pipe.to(distributed_state.device)
<span class="hljs-comment"># Assume two processes</span>
<span class="hljs-keyword">with</span> distributed_state.split_between_processes([<span class="hljs-string">&quot;a dog&quot;</span>, <span class="hljs-string">&quot;a cat&quot;</span>, <span class="hljs-string">&quot;a chicken&quot;</span>], apply_padding=<span class="hljs-literal">True</span>) <span class="hljs-keyword">as</span> prompt:
result = pipe(prompt).images<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-42h08l">On the first GPU, the prompts will be <code>[&quot;a dog&quot;, &quot;a cat&quot;]</code>, and on the second GPU it will be <code>[&quot;a chicken&quot;, &quot;a chicken&quot;]</code>.
Make sure to drop the final sample, as it will be a duplicate of the previous one.</p> <p data-svelte-h="svelte-13go2fo">You can find more complex examples <a href="https://github.com/huggingface/accelerate/tree/main/examples/inference/distributed" rel="nofollow">here</a> such as how to use it with LLMs.</p> <h2 class="relative group"><a id="memory-efficient-pipeline-parallelism-experimental" 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="#memory-efficient-pipeline-parallelism-experimental"><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>Memory-efficient pipeline parallelism (experimental)</span></h2> <p data-svelte-h="svelte-ynzvto">This next part will discuss using <em>pipeline parallelism</em>. This is an <strong>experimental</strong> API that utilizes <a href="https://pytorch.org/docs/stable/distributed.pipelining.html#" rel="nofollow">torch.distributed.pipelining</a> as a native solution.</p> <p data-svelte-h="svelte-1l3sfxh">The general idea with pipeline parallelism is: say you have 4 GPUs and a model big enough it can be <em>split</em> on four GPUs using <code>device_map=&quot;auto&quot;</code>. With this method you can send in 4 inputs at a time (for example here, any amount works) and each model chunk will work on an input, then receive the next input once the prior chunk finished, making it <em>much</em> more efficient <strong>and faster</strong> than the method described earlier. Here’s a visual taken from the PyTorch repository:</p> <p data-svelte-h="svelte-l2zgqj"><img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/accelerate/pipeline_parallel.png" alt="Pipeline parallelism example"></p> <p data-svelte-h="svelte-heqe2k">To illustrate how you can use this with Accelerate, we have created an <a href="https://github.com/huggingface/accelerate/tree/main/examples/inference" rel="nofollow">example zoo</a> showcasing a number of different models and situations. In this tutorial, we’ll show this method for GPT2 across two GPUs.</p> <p data-svelte-h="svelte-1pr52sj">Before you proceed, please make sure you have the latest PyTorch version installed by running the following:</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 torch<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1mviugk">Start by creating the model on the CPU:</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 -->from transformers <span class="hljs-keyword">import</span> GPT2ForSequenceClassification, GPT2Config
config = <span class="hljs-built_in">GPT2Config</span>()
model = <span class="hljs-built_in">GPT2ForSequenceClassification</span>(config)
model.<span class="hljs-built_in">eval</span>()<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-15xaqzv">Next you’ll need to create some example inputs to use. These help <code>torch.distributed.pipelining</code> trace the model.</p> <blockquote class="warning">However you make this example will determine the relative batch size that will be used/passed
through the model at a given time, so make sure to remember how many items there are!</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 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 -->input = torch.randint(
<span class="hljs-attribute">low</span>=0,
<span class="hljs-attribute">high</span>=config.vocab_size,
size=(2, 1024), # bs x seq_len
<span class="hljs-attribute">device</span>=<span class="hljs-string">&quot;cpu&quot;</span>,
<span class="hljs-attribute">dtype</span>=torch.int64,
<span class="hljs-attribute">requires_grad</span>=<span class="hljs-literal">False</span>,
)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-11mjb3e">Next we need to actually perform the tracing and get the model ready. To do so, use the <a href="/docs/accelerate/pr_4021/en/package_reference/inference#accelerate.prepare_pippy">inference.prepare_pippy()</a> function and it will fully wrap the model for pipeline parallelism automatically:</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.inference <span class="hljs-keyword">import</span> prepare_pippy
example_inputs = {&quot;input_ids&quot;: <span class="hljs-keyword">input</span>}
model = prepare_pippy(model, example_args=(<span class="hljs-keyword">input</span>,))<!-- HTML_TAG_END --></pre></div> <blockquote class="tip"><p data-svelte-h="svelte-123k9fd">There are a variety of parameters you can pass through to <code>prepare_pippy</code>:</p> <ul data-svelte-h="svelte-1bax0io"><li><p><code>split_points</code> lets you determine what layers to split the model at. By default we use wherever <code>device_map=&quot;auto&quot; declares, such as </code>fc<code>or</code>conv1`.</p></li> <li><p><code>num_chunks</code> determines how the batch will be split and sent to the model itself (so <code>num_chunks=1</code> with four split points/four GPUs will have a naive MP where a single input gets passed between the four layer split points)</p></li></ul></blockquote> <p data-svelte-h="svelte-t45mca">From here, all that’s left is to actually perform the distributed inference!</p> <blockquote class="warning"><p data-svelte-h="svelte-1cjitly">When passing inputs, we highly recommend to pass them in as a tuple of arguments. Using <code>kwargs</code> is supported, however, this approach is experimental.</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 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-variable">args</span> = <span class="hljs-variable">some_more_arguments</span>
<span class="hljs-variable">with</span> <span class="hljs-variable">torch.no_grad</span>():
<span class="hljs-variable">output</span> = <span class="hljs-function"><span class="hljs-title">model</span>(*<span class="hljs-variable">args</span>)</span><!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1f6tfb4">When finished all the data will be on the last process only:</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> PartialState
<span class="hljs-keyword">if</span> PartialState().is_last_process:
<span class="hljs-built_in">print</span>(output)<!-- HTML_TAG_END --></pre></div> <blockquote class="tip"><p data-svelte-h="svelte-fqz35p">If you pass in <code>gather_output=True</code> to <a href="/docs/accelerate/pr_4021/en/package_reference/inference#accelerate.prepare_pippy">inference.prepare_pippy()</a>, the output will be sent
across to all the GPUs afterwards without needing the <code>is_last_process</code> check. This is
<code>False</code> by default as it incurs a communication call.</p></blockquote> <p data-svelte-h="svelte-bz2a3m">And that’s it! To explore more, please check out the inference examples in the <a href="https://github.com/huggingface/accelerate/tree/main/examples/inference/pippy" rel="nofollow">Accelerate repo</a> and our <a href="../package_reference/inference">documentation</a> as we work to improving this integration.</p> <a class="!text-gray-400 !no-underline text-sm flex items-center not-prose mt-4" href="https://github.com/huggingface/accelerate/blob/main/docs/source/usage_guides/distributed_inference.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 data-svelte-h="svelte-zjs2n5"><span class="underline">Update</span> on GitHub</span></a> <p></p>
<script>
{
__sveltekit_1q7nz6m = {
assets: "/docs/accelerate/pr_4021/en",
base: "/docs/accelerate/pr_4021/en",
env: {}
};
const element = document.currentScript.parentElement;
const data = [null,null];
Promise.all([
import("/docs/accelerate/pr_4021/en/_app/immutable/entry/start.8a49e72b.js"),
import("/docs/accelerate/pr_4021/en/_app/immutable/entry/app.1df4d18e.js")
]).then(([kit, app]) => {
kit.start(app, element, {
node_ids: [0, 44],
data,
form: null,
error: null
});
});
}
</script>

Xet Storage Details

Size:
40 kB
·
Xet hash:
ebd060607811ae285c78bd87077ace23959c59eccf055096d782fc03afe565fe

Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.