diffusers/pr_13881/zh/_app/immutable/nodes/36.qrWwbSet.js

import"../chunks/DsnmJJEf.js";import{i as w,h as C,C as Z,H as l,a as i,E as b,s as I}from"../chunks/DdZvggmf.js";import{p as x,o as j,s as e,f as _,a as U,b as V,c as g,n as G}from"../chunks/BbekZcyp.js";const v='{"title":"xDiT","local":"xdit","sections":[{"title":"支持的模型","local":"支持的模型","sections":[],"depth":2},{"title":"基准测试","local":"基准测试","sections":[{"title":"Flux.1-schnell","local":"flux1-schnell","sections":[],"depth":3},{"title":"Stable Diffusion 3","local":"stable-diffusion-3","sections":[],"depth":3},{"title":"HunyuanDiT","local":"hunyuandit","sections":[],"depth":3}],"depth":2},{"title":"参考文献","local":"参考文献","sections":[],"depth":2}],"depth":1}';var B=g('<meta name="hf:doc:metadata"/>'),X=g(`<p></p> <!> <!> <p><a href="https://github.com/xdit-project/xDiT" rel="nofollow">xDiT</a> 是一个推理引擎，专为大规模并行部署扩散变换器（DiTs）而设计。xDiT 提供了一套用于扩散模型的高效并行方法，以及 GPU 内核加速。</p> <p>xDiT 支持四种并行方法，包括<a href="https://huggingface.co/papers/2405.07719" rel="nofollow">统一序列并行</a>、<a href="https://huggingface.co/papers/2405.14430" rel="nofollow">PipeFusion</a>、CFG 并行和数据并行。xDiT 中的这四种并行方法可以以混合方式配置，优化通信模式以最适合底层网络硬件。</p> <p>与并行化正交的优化侧重于加速单个 GPU 的性能。除了利用知名的注意力优化库外，我们还利用编译加速技术，如 torch.compile 和 onediff。</p> <p>xDiT 的概述如下所示。</p> <div class="flex justify-center"><img src="https://huggingface.co/datasets/xDiT/documentation-images/resolve/main/methods/xdit_overview.png"/></div> 您可以使用以下命令安装 xDiT： <!> <p>以下是一个使用 xDiT 加速 Diffusers 模型推理的示例。</p> <!> <p>如您所见，我们只需要使用 xDiT 中的 xFuserArgs 来获取配置参数，并将这些参数与来自 Diffusers 库的管道对象一起传递给 xDiTParallel，即可完成对 Diffusers 中特定管道的并行化。</p> <p>xDiT 运行时参数可以在命令行中使用 <code>-h</code> 查看，您可以参考此<a href="https://github.com/xdit-project/xDiT?tab=readme-ov-file#2-usage" rel="nofollow">使用</a>示例以获取更多详细信息。
ils。</p> <p>xDiT 需要使用 torchrun 启动，以支持其多节点、多 GPU 并行能力。例如，以下命令可用于 8-GPU 并行推理：</p> <!> <!> <p>在 xDiT 中支持 Diffusers 模型的一个子集，例如 Flux.1、Stable Diffusion 3 等。最新支持的模型可以在<a href="https://github.com/xdit-project/xDiT?tab=readme-ov-file#-supported-dits" rel="nofollow">这里</a>找到。</p> <!> <p>我们在不同机器上测试了各种模型，以下是一些基准数据。</p> <!> <div class="flex justify-center"><img src="https://huggingface.co/datasets/xDiT/documentation-images/resolve/main/performance/flux/Flux-2k-L40.png"/></div> <div class="flex justify-center"><img src="https://huggingface.co/datasets/xDiT/documentation-images/resolve/main/performance/flux/Flux-2K-A100.png"/></div> <!> <div class="flex justify-center"><img src="https://huggingface.co/datasets/xDiT/documentation-images/resolve/main/performance/sd3/L40-SD3.png"/></div> <div class="flex justify-center"><img src="https://huggingface.co/datasets/xDiT/documentation-images/resolve/main/performance/sd3/A100-SD3.png"/></div> <!> <div class="flex justify-center"><img src="https://huggingface.co/datasets/xDiT/documentation-images/resolve/main/performance/hunuyuandit/L40-HunyuanDiT.png"/></div> <div class="flex justify-center"><img src="https://huggingface.co/datasets/xDiT/documentation-images/resolve/main/performance/hunuyuandit/V100-HunyuanDiT.png"/></div> <div class="flex justify-center"><img src="https://huggingface.co/datasets/xDiT/documentation-images/resolve/main/performance/hunuyuandit/T4-HunyuanDiT.png"/></div> <p>更详细的性能指标可以在我们的 <a href="https://github.com/xdit-project/xDiT?tab=readme-ov-file#perf" rel="nofollow">GitHub 页面</a> 上找到。</p> <!> <p><a href="https://github.com/xdit-project/xDiT" rel="nofollow">xDiT-project</a></p> <p><a href="https://huggingface.co/papers/2405.07719" rel="nofollow">USP: A Unified Sequence Parallelism Approach for Long Context Generative AI</a></p> <p><a href="https://huggingface.co/papers/2405.14430" rel="nofollow">PipeFusion: Displaced Patch Pipeline Parallelism for Inference of Diffusion Transformer Models</a></p> <!> <p></p>`,1);function W(h,T){x(T,!1),j(()=>{new URLSearchParams(window.location.search).get("fw")}),w();var a=X();C("16vwjz4",m=>{var f=B();I(f,"content",v),U(m,f)});var t=e(_(a),2);Z(t,{containerStyle:"float: right; margin-left: 10px; display: inline-flex; position: relative; z-index: 10;"});var n=e(t,2);l(n,{title:"xDiT",local:"xdit",headingTag:"h1"});var s=e(n,12);i(s,{code:"cGlwJTIwaW5zdGFsbCUyMHhmdXNlcg==",highlighted:"pip install xfuser",lang:"bash",wrap:!1});var o=e(s,4);i(o,{code:"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",highlighted:` import torch
 from diffusers import StableDiffusion3Pipeline

 from xfuser import xFuserArgs, xDiTParallel
 from xfuser.config import FlexibleArgumentParser
 from xfuser.core.distributed import get_world_group

 def main():
<span class="hljs-addition">+    parser = FlexibleArgumentParser(description=&quot;xFuser Arguments&quot;)</span>
<span class="hljs-addition">+    args = xFuserArgs.add_cli_args(parser).parse_args()</span>
<span class="hljs-addition">+    engine_args = xFuserArgs.from_cli_args(args)</span>
<span class="hljs-addition">+    engine_config, input_config = engine_args.create_config()</span>

     local_rank = get_world_group().local_rank
     pipe = StableDiffusion3Pipeline.from_pretrained(
         pretrained_model_name_or_path=engine_config.model_config.model,
         torch_dtype=torch.float16,
     ).to(f&quot;cuda:{local_rank}&quot;)
    
# 在这里对管道进行任何操作

<span class="hljs-addition">+    pipe = xDiTParallel(pipe, engine_config, input_config)</span>

     pipe(
         height=input_config.height,
         width=input_config.height,
         prompt=input_config.prompt,
         num_inference_steps=input_config.num_inference_steps,
         output_type=input_config.output_type,
         generator=torch.Generator(device=&quot;cuda&quot;).manual_seed(input_config.seed),
     )

<span class="hljs-addition">+    if input_config.output_type == &quot;pil&quot;:</span>
<span class="hljs-addition">+        pipe.save(&quot;results&quot;, &quot;stable_diffusion_3&quot;)</span>

if __name__ == &quot;__main__&quot;:
    main()`,lang:"diff",wrap:!1});var c=e(o,8);i(c,{code:"dG9yY2hydW4lMjAtLW5wcm9jX3Blcl9ub2RlJTNEOCUyMC4lMkZpbmZlcmVuY2UucHklMjAtLW1vZGVsJTIwbW9kZWxzJTJGRkxVWC4xLWRldiUyMC0tZGF0YV9wYXJhbGxlbF9kZWdyZWUlMjAyJTIwLS11bHlzc2VzX2RlZ3JlZSUyMDIlMjAtLXJpbmdfZGVncmVlJTIwMiUyMC0tcHJvbXB0JTIwJTIyQSUyMHNub3d5JTIwbW91bnRhaW4lMjIlMjAlMjJBJTIwc21hbGwlMjBkb2clMjIlMjAtLW51bV9pbmZlcmVuY2Vfc3RlcHMlMjA1MA==",highlighted:'torchrun --nproc_per_node=8 ./inference.py --model models/FLUX.1-dev --data_parallel_degree 2 --ulysses_degree 2 --ring_degree 2 --prompt <span class="hljs-string">&quot;A snowy mountain&quot;</span> <span class="hljs-string">&quot;A small dog&quot;</span> --num_inference_steps 50',lang:"bash",wrap:!1});var r=e(c,2);l(r,{title:"支持的模型",local:"支持的模型",headingTag:"h2"});var p=e(r,4);l(p,{title:"基准测试",local:"基准测试",headingTag:"h2"});var d=e(p,4);l(d,{title:"Flux.1-schnell",local:"flux1-schnell",headingTag:"h3"});var y=e(d,6);l(y,{title:"Stable Diffusion 3",local:"stable-diffusion-3",headingTag:"h3"});var u=e(y,6);l(u,{title:"HunyuanDiT",local:"hunyuandit",headingTag:"h3"});var M=e(u,10);l(M,{title:"参考文献",local:"参考文献",headingTag:"h2"});var J=e(M,8);b(J,{source:"https://github.com/huggingface/diffusers/blob/main/docs/source/zh/optimization/xdit.md"}),G(2),U(h,a),V()}export{W as component};