diffusers/pr_13881/zh/_app/immutable/nodes/42.Bcc8xow6.js

import"../chunks/DsnmJJEf.js";import{i as _,h as B,C as k,H as a,a as s,E as W,s as E}from"../chunks/DdZvggmf.js";import{p as R,o as X,s as l,f as N,a as f,b as V,c as Z,n as v}from"../chunks/BbekZcyp.js";const C='{"title":"分布式推理","local":"分布式推理","sections":[{"title":"🤗 Accelerate","local":"-accelerate","sections":[],"depth":2},{"title":"PyTorch Distributed","local":"pytorch-distributed","sections":[],"depth":2},{"title":"模型分片","local":"模型分片","sections":[],"depth":2}],"depth":1}';var A=Z('<meta name="hf:doc:metadata"/>'),F=Z(`<p></p> <!> <!> <p>在分布式设置中，您可以使用 🤗 <a href="https://huggingface.co/docs/accelerate/index" rel="nofollow">Accelerate</a> 或 <a href="https://pytorch.org/tutorials/beginner/dist_overview.html" rel="nofollow">PyTorch Distributed</a> 在多个 GPU 上运行推理，这对于并行生成多个提示非常有用。</p> <p>本指南将向您展示如何使用 🤗 Accelerate 和 PyTorch Distributed 进行分布式推理。</p> <!> <p>🤗 <a href="https://huggingface.co/docs/accelerate/index" rel="nofollow">Accelerate</a> 是一个旨在简化在分布式设置中训练或运行推理的库。它简化了设置分布式环境的过程，让您可以专注于您的 PyTorch 代码。</p> <p>首先，创建一个 Python 文件并初始化一个 <code>accelerate.PartialState</code> 来创建分布式环境；您的设置会自动检测，因此您无需明确定义 <code>rank</code> 或 <code>world_size</code>。将 <code>DiffusionPipeline</code> 移动到 <code>distributed_state.device</code> 以为每个进程分配一个 GPU。</p> <p>现在使用 <code>split_between_processes</code> 实用程序作为上下文管理器，自动在进程数之间分发提示。</p> <!> <p>使用 <code>--num_processes</code> 参数指定要使用的 GPU 数量，并调用 <code>accelerate launch</code> 来运行脚本：</p> <!> <blockquote class="tip"><p>参考这个最小示例 <a href="https://gist.github.com/sayakpaul/cfaebd221820d7b43fae638b4dfa01ba" rel="nofollow">脚本</a> 以在多个 GPU 上运行推理。要了解更多信息，请查看 <a href="https://huggingface.co/docs/accelerate/en/usage_guides/distributed_inference#distributed-inference-with-accelerate" rel="nofollow">使用 🤗 Accelerate 进行分布式推理</a> 指南。</p></blockquote> <!> <p>PyTorch 支持 <a href="https://pytorch.org/docs/stable/generated/torch.nn.parallel.DistributedDataParallel.html" rel="nofollow"><code>DistributedDataParallel</code></a>，它启用了数据
并行性。</p> <p>首先，创建一个 Python 文件并导入 <code>torch.distributed</code> 和 <code>torch.multiprocessing</code> 来设置分布式进程组，并为每个 GPU 上的推理生成进程。您还应该初始化一个 <code>DiffusionPipeline</code>：</p> <!> <p>您需要创建一个函数来运行推理；<a href="https://pytorch.org/docs/stable/distributed.html?highlight=init_process_group#torch.distributed.init_process_group" rel="nofollow"><code>init_process_group</code></a> 处理创建一个分布式环境，指定要使用的后端类型、当前进程的 <code>rank</code> 以及参与进程的数量 <code>world_size</code>。如果您在 2 个 GPU 上并行运行推理，那么 <code>world_size</code> 就是 2。</p> <p>将 <code>DiffusionPipeline</code> 移动到 <code>rank</code>，并使用 <code>get_rank</code> 为每个进程分配一个 GPU，其中每个进程处理不同的提示：</p> <!> <p>要运行分布式推理，调用 <a href="https://pytorch.org/docs/stable/multiprocessing.html#torch.multiprocessing.spawn" rel="nofollow"><code>mp.spawn</code></a> 在 <code>world_size</code> 定义的 GPU 数量上运行 <code>run_inference</code> 函数：</p> <!> <p>完成推理脚本后，使用 <code>--nproc_per_node</code> 参数指定要使用的 GPU 数量，并调用 <code>torchrun</code> 来运行脚本：</p> <!> <blockquote class="tip"><p>您可以在 <code>DiffusionPipeline</code> 中使用 <code>device_map</code> 将其模型级组件分布在多个设备上。请参考 <a href="../tutorials/inference_with_big_models#device-placement">设备放置</a> 指南了解更多信息。</p></blockquote> <!> <p>现代扩散系统，如 <a href="../api/pipelines/flux">Flux</a>，非常大且包含多个模型。例如，<a href="https://hf.co/black-forest-labs/FLUX.1-dev" rel="nofollow">Flux.1-Dev</a> 由两个文本编码器 - <a href="https://hf.co/google/t5-v1_1-xxl" rel="nofollow">T5-XXL</a> 和 <a href="https://hf.co/openai/clip-vit-large-patch14" rel="nofollow">CLIP-L</a> - 一个 <a href="../api/models/flux_transformer">扩散变换器</a>，以及一个 <a href="../api/models/autoencoderkl">VAE</a> 组成。对于如此大的模型，在消费级 GPU 上运行推理可能具有挑战性。</p> <p>模型分片是一种技术，当模型无法容纳在单个 GPU 上时，将模型分布在多个 GPU 上。下面的示例假设有两个 16GB GPU 可用于推理。</p> <p>开始使用文本编码器计算文本嵌入。通过设置 <code>device_map="balanced"</code> 将文本编码器保持在两个GPU上。<code>balanced</code> 策略将模型均匀分布在所有可用GPU上。使用 <code>max_memory</code> 参数为每个GPU上的每个文本编码器分配最大内存量。</p> <blockquote class="tip"><p><strong>仅</strong> 在此步骤加载文本编码器！扩散变换器和VAE在后续步骤中加载以节省内存。</p></blockquote> <!> <p>一旦文本嵌入计算完成，从GPU中移除它们以为扩散变换器腾出空间。</p> <!> <p>接下来加载扩散变换器，它有125亿参数。这次，设置 <code>device_map="auto"</code> 以自动将模型分布在两个16GB GPU上。<code>auto</code> 策略由 <a href="https://hf.co/docs/accelerate/index" rel="nofollow">Accelerate</a> 支持，并作为 <a href="https://hf.co/docs/accelerate/concept_guides/big_model_inference" rel="nofollow">大模型推理</a> 功能的一部分可用。它首先将模型分布在最快的设备（GPU）上，然后在需要时移动到较慢的设备如CPU和硬盘。将模型参数存储在较慢设备上的权衡是推理延迟较慢。</p> <!> <blockquote class="tip"><p>在任何时候，您可以尝试 <code>print(pipeline.hf_device_map)</code> 来查看各种模型如何在设备上分布。这对于跟踪模型的设备放置很有用。您也可以尝试 <code>print(transformer.hf_device_map)</code> 来查看变换器模型如何在设备上分片。</p></blockquote> <p>将变换器模型添加到管道中以进行去噪，但将其他模型级组件如文本编码器和VAE设置为 <code>None</code>，因为您还不需要它们。</p> <!> <p>从内存中移除管道和变换器，因为它们不再需要。</p> <!> <p>最后，使用变分自编码器（VAE）将潜在表示解码为图像。VAE通常足够小，可以在单个GPU上加载。</p> <!> <p>通过选择性加载和卸载在特定阶段所需的模型，并将最大模型分片到多个GPU上，可以在消费级GPU上运行大型模型的推理。</p> <!> <p></p>`,1);function S(g,G){R(G,!1),X(()=>{new URLSearchParams(window.location.search).get("fw")}),_();var e=F();B("el5ux9",b=>{var T=A();E(T,"content",C),f(b,T)});var t=l(N(e),2);k(t,{containerStyle:"float: right; margin-left: 10px; display: inline-flex; position: relative; z-index: 10;"});var n=l(t,2);a(n,{title:"分布式推理",local:"分布式推理",headingTag:"h1"});var o=l(n,6);a(o,{title:"🤗 Accelerate",local:"-accelerate",headingTag:"h2"});var c=l(o,8);s(c,{code:"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",highlighted:`<span class="hljs-keyword">import</span> torch
<span class="hljs-keyword">from</span> accelerate <span class="hljs-keyword">import</span> PartialState
<span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> DiffusionPipeline

pipeline = DiffusionPipeline.from_pretrained(
    <span class="hljs-string">&quot;stable-diffusion-v1-5/stable-diffusion-v1-5&quot;</span>, torch_dtype=torch.float16, use_safetensors=<span class="hljs-literal">True</span>
)
distributed_state = PartialState()
pipeline.to(distributed_state.device)

<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 = pipeline(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>)`,lang:"py",wrap:!1});var p=l(c,4);s(p,{code:"YWNjZWxlcmF0ZSUyMGxhdW5jaCUyMHJ1bl9kaXN0cmlidXRlZC5weSUyMC0tbnVtX3Byb2Nlc3NlcyUzRDI=",highlighted:"accelerate launch run_distributed.py --num_processes=2",lang:"bash",wrap:!1});var r=l(p,4);a(r,{title:"PyTorch Distributed",local:"pytorch-distributed",headingTag:"h2"});var d=l(r,6);s(d,{code:"aW1wb3J0JTIwdG9yY2glMEFpbXBvcnQlMjB0b3JjaC5kaXN0cmlidXRlZCUyMGFzJTIwZGlzdCUwQWltcG9ydCUyMHRvcmNoLm11bHRpcHJvY2Vzc2luZyUyMGFzJTIwbXAlMEElMEFmcm9tJTIwZGlmZnVzZXJzJTIwaW1wb3J0JTIwRGlmZnVzaW9uUGlwZWxpbmUlMEElMEFzZCUyMCUzRCUyMERpZmZ1c2lvblBpcGVsaW5lLmZyb21fcHJldHJhaW5lZCglMEElMjAlMjAlMjAlMjAlMjJzdGFibGUtZGlmZnVzaW9uLXYxLTUlMkZzdGFibGUtZGlmZnVzaW9uLXYxLTUlMjIlMkMlMjB0b3JjaF9kdHlwZSUzRHRvcmNoLmZsb2F0MTYlMkMlMjB1c2Vfc2FmZXRlbnNvcnMlM0RUcnVlJTBBKQ==",highlighted:`<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">import</span> torch.multiprocessing <span class="hljs-keyword">as</span> mp

<span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> DiffusionPipeline

sd = DiffusionPipeline.from_pretrained(
    <span class="hljs-string">&quot;stable-diffusion-v1-5/stable-diffusion-v1-5&quot;</span>, torch_dtype=torch.float16, use_safetensors=<span class="hljs-literal">True</span>
)`,lang:"py",wrap:!1});var i=l(d,6);s(i,{code:"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",highlighted:`<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)

    sd.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>

    image = sd(prompt).images[<span class="hljs-number">0</span>]
    image.save(<span class="hljs-string">f&quot;./<span class="hljs-subst">{<span class="hljs-string">&#x27;_&#x27;</span>.join(prompt)}</span>.png&quot;</span>)`,lang:"py",wrap:!1});var M=l(i,4);s(M,{code:"ZGVmJTIwbWFpbigpJTNBJTBBJTIwJTIwJTIwJTIwd29ybGRfc2l6ZSUyMCUzRCUyMDIlMEElMjAlMjAlMjAlMjBtcC5zcGF3bihydW5faW5mZXJlbmNlJTJDJTIwYXJncyUzRCh3b3JsZF9zaXplJTJDKSUyQyUyMG5wcm9jcyUzRHdvcmxkX3NpemUlMkMlMjBqb2luJTNEVHJ1ZSklMEElMEElMEFpZiUyMF9fbmFtZV9fJTIwJTNEJTNEJTIwJTIyX19tYWluX18lMjIlM0ElMEElMjAlMjAlMjAlMjBtYWluKCk=",highlighted:`<span class="hljs-keyword">def</span> <span class="hljs-title function_">main</span>():
    world_size = <span class="hljs-number">2</span>
    mp.spawn(run_inference, args=(world_size,), nprocs=world_size, join=<span class="hljs-literal">True</span>)


<span class="hljs-keyword">if</span> __name__ == <span class="hljs-string">&quot;__main__&quot;</span>:
    main()`,lang:"py",wrap:!1});var h=l(M,4);s(h,{code:"dG9yY2hydW4lMjBydW5fZGlzdHJpYnV0ZWQucHklMjAtLW5wcm9jX3Blcl9ub2RlJTNEMg==",highlighted:"torchrun run_distributed.py --nproc_per_node=2",lang:"bash",wrap:!1});var J=l(h,4);a(J,{title:"模型分片",local:"模型分片",headingTag:"h2"});var y=l(J,10);s(y,{code:"ZnJvbSUyMGRpZmZ1c2VycyUyMGltcG9ydCUyMEZsdXhQaXBlbGluZSUwQWltcG9ydCUyMHRvcmNoJTBBJTBBcHJvbXB0JTIwJTNEJTIwJTIyYSUyMHBob3RvJTIwb2YlMjBhJTIwZG9nJTIwd2l0aCUyMGNhdC1saWtlJTIwbG9vayUyMiUwQSUwQXBpcGVsaW5lJTIwJTNEJTIwRmx1eFBpcGVsaW5lLmZyb21fcHJldHJhaW5lZCglMEElMjAlMjAlMjAlMjAlMjJibGFjay1mb3Jlc3QtbGFicyUyRkZMVVguMS1kZXYlMjIlMkMlMEElMjAlMjAlMjAlMjB0cmFuc2Zvcm1lciUzRE5vbmUlMkMlMEElMjAlMjAlMjAlMjB2YWUlM0ROb25lJTJDJTBBJTIwJTIwJTIwJTIwZGV2aWNlX21hcCUzRCUyMmJhbGFuY2VkJTIyJTJDJTBBJTIwJTIwJTIwJTIwbWF4X21lbW9yeSUzRCU3QjAlM0ElMjAlMjIxNkdCJTIyJTJDJTIwMSUzQSUyMCUyMjE2R0IlMjIlN0QlMkMlMEElMjAlMjAlMjAlMjB0b3JjaF9kdHlwZSUzRHRvcmNoLmJmbG9hdDE2JTBBKSUwQXdpdGglMjB0b3JjaC5ub19ncmFkKCklM0ElMEElMjAlMjAlMjAlMjBwcmludCglMjJFbmNvZGluZyUyMHByb21wdHMuJTIyKSUwQSUyMCUyMCUyMCUyMHByb21wdF9lbWJlZHMlMkMlMjBwb29sZWRfcHJvbXB0X2VtYmVkcyUyQyUyMHRleHRfaWRzJTIwJTNEJTIwcGlwZWxpbmUuZW5jb2RlX3Byb21wdCglMEElMjAlMjAlMjAlMjAlMjAlMjAlMjAlMjBwcm9tcHQlM0Rwcm9tcHQlMkMlMjBwcm9tcHRfMiUzRE5vbmUlMkMlMjBtYXhfc2VxdWVuY2VfbGVuZ3RoJTNENTEyJTBBJTIwJTIwJTIwJTIwKQ==",highlighted:`<span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> FluxPipeline
<span class="hljs-keyword">import</span> torch

prompt = <span class="hljs-string">&quot;a photo of a dog with cat-like look&quot;</span>

pipeline = FluxPipeline.from_pretrained(
    <span class="hljs-string">&quot;black-forest-labs/FLUX.1-dev&quot;</span>,
    transformer=<span class="hljs-literal">None</span>,
    vae=<span class="hljs-literal">None</span>,
    device_map=<span class="hljs-string">&quot;balanced&quot;</span>,
    max_memory={<span class="hljs-number">0</span>: <span class="hljs-string">&quot;16GB&quot;</span>, <span class="hljs-number">1</span>: <span class="hljs-string">&quot;16GB&quot;</span>},
    torch_dtype=torch.bfloat16
)
<span class="hljs-keyword">with</span> torch.no_grad():
    <span class="hljs-built_in">print</span>(<span class="hljs-string">&quot;Encoding prompts.&quot;</span>)
    prompt_embeds, pooled_prompt_embeds, text_ids = pipeline.encode_prompt(
        prompt=prompt, prompt_2=<span class="hljs-literal">None</span>, max_sequence_length=<span class="hljs-number">512</span>
    )`,lang:"py",wrap:!1});var w=l(y,4);s(w,{code:"aW1wb3J0JTIwZ2MlMjAlMEElMEFkZWYlMjBmbHVzaCgpJTNBJTBBJTIwJTIwJTIwJTIwZ2MuY29sbGVjdCgpJTBBJTIwJTIwJTIwJTIwdG9yY2guY3VkYS5lbXB0eV9jYWNoZSgpJTBBJTIwJTIwJTIwJTIwdG9yY2guY3VkYS5yZXNldF9tYXhfbWVtb3J5X2FsbG9jYXRlZCgpJTBBJTIwJTIwJTIwJTIwdG9yY2guY3VkYS5yZXNldF9wZWFrX21lbW9yeV9zdGF0cygpJTBBJTBBZGVsJTIwcGlwZWxpbmUudGV4dF9lbmNvZGVyJTBBZGVsJTIwcGlwZWxpbmUudGV4dF9lbmNvZGVyXzIlMEFkZWwlMjBwaXBlbGluZS50b2tlbml6ZXIlMEFkZWwlMjBwaXBlbGluZS50b2tlbml6ZXJfMiUwQWRlbCUyMHBpcGVsaW5lJTBBJTBBZmx1c2goKQ==",highlighted:`<span class="hljs-keyword">import</span> gc 

<span class="hljs-keyword">def</span> <span class="hljs-title function_">flush</span>():
    gc.collect()
    torch.cuda.empty_cache()
    torch.cuda.reset_max_memory_allocated()
    torch.cuda.reset_peak_memory_stats()

<span class="hljs-keyword">del</span> pipeline.text_encoder
<span class="hljs-keyword">del</span> pipeline.text_encoder_2
<span class="hljs-keyword">del</span> pipeline.tokenizer
<span class="hljs-keyword">del</span> pipeline.tokenizer_2
<span class="hljs-keyword">del</span> pipeline

flush()`,lang:"py",wrap:!1});var m=l(w,4);s(m,{code:"ZnJvbSUyMGRpZmZ1c2VycyUyMGltcG9ydCUyMEF1dG9Nb2RlbCUwQWltcG9ydCUyMHRvcmNoJTIwJTBBJTBBdHJhbnNmb3JtZXIlMjAlM0QlMjBBdXRvTW9kZWwuZnJvbV9wcmV0cmFpbmVkKCUwQSUyMCUyMCUyMCUyMCUyMmJsYWNrLWZvcmVzdC1sYWJzJTJGRkxVWC4xLWRldiUyMiUyQyUyMCUwQSUyMCUyMCUyMCUyMHN1YmZvbGRlciUzRCUyMnRyYW5zZm9ybWVyJTIyJTJDJTBBJTIwJTIwJTIwJTIwZGV2aWNlX21hcCUzRCUyMmF1dG8lMjIlMkMlMEElMjAlMjAlMjAlMjB0b3JjaF9kdHlwZSUzRHRvcmNoLmJmbG9hdDE2JTBBKQ==",highlighted:`<span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> AutoModel
<span class="hljs-keyword">import</span> torch 

transformer = AutoModel.from_pretrained(
    <span class="hljs-string">&quot;black-forest-labs/FLUX.1-dev&quot;</span>, 
    subfolder=<span class="hljs-string">&quot;transformer&quot;</span>,
    device_map=<span class="hljs-string">&quot;auto&quot;</span>,
    torch_dtype=torch.bfloat16
)`,lang:"py",wrap:!1});var j=l(m,6);s(j,{code:"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",highlighted:`pipeline = FluxPipeline.from_pretrained(
    <span class="hljs-string">&quot;black-forest-labs/FLUX.1-dev&quot;</span>,
    text_encoder=<span class="hljs-literal">None</span>,
    text_encoder_2=<span class="hljs-literal">None</span>,
    tokenizer=<span class="hljs-literal">None</span>,
    tokenizer_2=<span class="hljs-literal">None</span>,
    vae=<span class="hljs-literal">None</span>,
    transformer=transformer,
    torch_dtype=torch.bfloat16
)

<span class="hljs-built_in">print</span>(<span class="hljs-string">&quot;Running denoising.&quot;</span>)
height, width = <span class="hljs-number">768</span>, <span class="hljs-number">1360</span>
latents = pipeline(
   
     
prompt_embeds=prompt_embeds,
pooled_prompt_embeds=pooled_prompt_embeds,
num_inference_steps=<span class="hljs-number">50</span>,
guidance_scale=<span class="hljs-number">3.5</span>,
height=height,
width=width,
output_type=<span class="hljs-string">&quot;latent&quot;</span>,
).images`,lang:"py",wrap:!1});var u=l(j,4);s(u,{code:"ZGVsJTIwcGlwZWxpbmUudHJhbnNmb3JtZXIlMEFkZWwlMjBwaXBlbGluZSUwQSUwQWZsdXNoKCk=",highlighted:`<span class="hljs-keyword">del</span> pipeline.transformer
<span class="hljs-keyword">del</span> pipeline

flush()`,lang:"py",wrap:!1});var U=l(u,4);s(U,{code:"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",highlighted:`<span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> AutoencoderKL
<span class="hljs-keyword">from</span> diffusers.image_processor <span class="hljs-keyword">import</span> VaeImageProcessor
<span class="hljs-keyword">import</span> torch 

vae = AutoencoderKL.from_pretrained(ckpt_id, subfolder=<span class="hljs-string">&quot;vae&quot;</span>, torch_dtype=torch.bfloat16).to(<span class="hljs-string">&quot;cuda&quot;</span>)
vae_scale_factor = <span class="hljs-number">2</span> ** (<span class="hljs-built_in">len</span>(vae.config.block_out_channels) - <span class="hljs-number">1</span>)
image_processor = VaeImageProcessor(vae_scale_factor=vae_scale_factor)

<span class="hljs-keyword">with</span> torch.no_grad():
    <span class="hljs-built_in">print</span>(<span class="hljs-string">&quot;运行解码中。&quot;</span>)
    latents = FluxPipeline._unpack_latents(latents, height, width, vae_scale_factor)
    latents = (latents / vae.config.scaling_factor) + vae.config.shift_factor

    image = vae.decode(latents, return_dict=<span class="hljs-literal">False</span>)[<span class="hljs-number">0</span>]
    image = image_processor.postprocess(image, output_type=<span class="hljs-string">&quot;pil&quot;</span>)
    image[<span class="hljs-number">0</span>].save(<span class="hljs-string">&quot;split_transformer.png&quot;</span>)`,lang:"py",wrap:!1});var I=l(U,4);W(I,{source:"https://github.com/huggingface/diffusers/blob/main/docs/source/zh/training/distributed_inference.md"}),v(2),f(g,e),V()}export{S as component};