Buckets:

hf-doc-build
/

doc

Files
xet
hf-doc-build/doc / diffusers /v0.43.0 /zh /optimization /tensor_parallelism.html
mishig
about 2 hours ago
download
raw
38.1 kB
<meta charset="utf-8" /><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;张量并行&quot;,&quot;local&quot;:&quot;tensor-parallelism&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;加载分片模型&quot;,&quot;local&quot;:&quot;load-a-sharded-model&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;加载性能&quot;,&quot;local&quot;:&quot;loading-performance&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;局限性&quot;,&quot;local&quot;:&quot;limitations&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;为模型添加张量并行&quot;,&quot;local&quot;:&quot;add-tensor-parallelism-to-a-model&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;列向和行向&quot;,&quot;local&quot;:&quot;colwise-and-rowwise&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;融合投影&quot;,&quot;local&quot;:&quot;fused-projections&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;需要排除的内容&quot;,&quot;local&quot;:&quot;what-to-leave-out&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;约束与验证&quot;,&quot;local&quot;:&quot;constraints-and-verification&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3}],&quot;depth&quot;:2}],&quot;depth&quot;:1}"/>
		<link href="/docs/diffusers/main/en/_app/immutable/entry/start.0S0_iEoG.js" rel="modulepreload">
		<link href="/docs/diffusers/main/en/_app/immutable/chunks/I9eooN77.js" rel="modulepreload">
		<link href="/docs/diffusers/main/en/_app/immutable/chunks/Bupkjr9f.js" rel="modulepreload">
		<link href="/docs/diffusers/main/en/_app/immutable/entry/app.CPdOxCbw.js" rel="modulepreload">
		<link href="/docs/diffusers/main/en/_app/immutable/chunks/CyvZ3FS5.js" rel="modulepreload">
		<link href="/docs/diffusers/main/en/_app/immutable/chunks/l0QnnOED.js" rel="modulepreload">
		<link href="/docs/diffusers/main/en/_app/immutable/chunks/DnVIJfC5.js" rel="modulepreload">
		<link href="/docs/diffusers/main/en/_app/immutable/chunks/DsnmJJEf.js" rel="modulepreload">
		<link href="/docs/diffusers/main/en/_app/immutable/nodes/0.DteTvmQZ.js" rel="modulepreload">
		<link href="/docs/diffusers/main/en/_app/immutable/chunks/CoSrdQax.js" rel="modulepreload">
		<link href="/docs/diffusers/main/en/_app/immutable/nodes/347.DnqzUbRP.js" rel="modulepreload">
		<link href="/docs/diffusers/main/en/_app/immutable/chunks/DSPDl1gz.js" rel="modulepreload">
		<!--17gkrup--><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;Tensor parallelism&quot;,&quot;local&quot;:&quot;tensor-parallelism&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Load a sharded model&quot;,&quot;local&quot;:&quot;load-a-sharded-model&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Loading performance&quot;,&quot;local&quot;:&quot;loading-performance&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Limitations&quot;,&quot;local&quot;:&quot;limitations&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Add tensor parallelism to a model&quot;,&quot;local&quot;:&quot;add-tensor-parallelism-to-a-model&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Colwise and rowwise&quot;,&quot;local&quot;:&quot;colwise-and-rowwise&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;Fused projections&quot;,&quot;local&quot;:&quot;fused-projections&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;What to leave out&quot;,&quot;local&quot;:&quot;what-to-leave-out&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3},{&quot;title&quot;:&quot;Constraints and verification&quot;,&quot;local&quot;:&quot;constraints-and-verification&quot;,&quot;sections&quot;:[],&quot;depth&quot;:3}],&quot;depth&quot;:2}],&quot;depth&quot;:1}"/><!---->
		<link href="/docs/diffusers/main/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="tensor-parallelism" 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="#tensor-parallelism"><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><!---->张量并行<!----></span></h1><!--]--><!----> <p><a href="https://huggingface.co/spaces/nanotron/ultrascale-playbook?section=tensor_parallelism" rel="nofollow">张量并行</a>将模型的权重矩阵分片到多个设备上。每个设备持有每一层的按列（<code>"colwise"</code>）或按行（<code>"rowwise"</code>）切片，并计算部分结果。在层边界处，通过 <code>AllReduce</code> 或 <code>AllGather</code> 将部分结果合并为完整输出。与 <a href="./context_parallelism">上下文并行</a> 不同，它减少了每个设备上的 <em>权重</em> 内存占用，这在模型无法装入单个设备时非常有用。请参阅 <a href="../training/distributed_inference">分布式推理</a> 以将其与其他策略进行比较。</p> <!--[1--><h2 class="relative group"><a id="load-a-sharded-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="#load-a-sharded-model"><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><!---->加载分片模型<!----></span></h2><!--]--><!----> <p>Pass a <a href="/docs/diffusers/v0.43.0/zh/api/parallel#diffusers.TensorParallelConfig">TensorParallelConfig</a> to the <code>parallel_config</code> argument of the model’s <a href="/docs/diffusers/v0.43.0/zh/api/models/overview#diffusers.ModelMixin.from_pretrained">from_pretrained()</a>. <code>tp_degree</code> is the number of devices to shard across and must divide the model’s number of attention heads. The model’s class must define a <code>_tp_plan</code>, which maps module names to a sharding style. To check whether a model supports tensor parallelism, look for this attribute on its class, such as <a href="/docs/diffusers/v0.43.0/zh/api/models/flux2_transformer#diffusers.Flux2Transformer2DModel">Flux2Transformer2DModel</a>.</p> <p>要将计划添加到另一个模型，请参阅 <a href="#add-tensor-parallelism-to-a-model">为模型添加张量并行</a>。</p> <p>以这种方式加载时，检查点在读取过程中会被分片。每个 rank 仅读取每个权重中属于自己的切片，并将其直接放置到其 GPU 上，因此没有任何 rank 会持有完整大小的权重。随着 <code>tp_degree</code> 的增加，每个 rank 的内存占用会下降。</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> 已复制</div><!----></button><!----></div> <pre class="language-py "><!----><span class="hljs-keyword">import</span> os
<span class="hljs-keyword">import</span> torch
<span class="hljs-keyword">from</span> torch <span class="hljs-keyword">import</span> distributed <span class="hljs-keyword">as</span> dist
<span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> DiffusionPipeline, Flux2Transformer2DModel, TensorParallelConfig

<span class="hljs-keyword">def</span> <span class="hljs-title function_">main</span>():
    dist.init_process_group(backend=<span class="hljs-string">&quot;nccl&quot;</span>)
    rank, world_size = dist.get_rank(), dist.get_world_size()
    <span class="hljs-comment"># torchrun sets LOCAL_RANK to this process&#x27;s GPU index on its node</span>
    local_rank = <span class="hljs-built_in">int</span>(os.environ[<span class="hljs-string">&quot;LOCAL_RANK&quot;</span>])
    device = torch.device(<span class="hljs-string">f&quot;cuda:<span class="hljs-subst">{local_rank}</span>&quot;</span>)
    torch.cuda.set_device(device)

    <span class="hljs-comment"># Each rank reads only its own shard of every planned weight, straight onto its own GPU.</span>
    transformer = Flux2Transformer2DModel.from_pretrained(
        <span class="hljs-string">&quot;black-forest-labs/FLUX.2-dev&quot;</span>,
        subfolder=<span class="hljs-string">&quot;transformer&quot;</span>,
        dtype=torch.bfloat16,
        parallel_config=TensorParallelConfig(tp_degree=world_size),
    )

    pipeline = DiffusionPipeline.from_pretrained(
        <span class="hljs-string">&quot;black-forest-labs/FLUX.2-dev&quot;</span>, transformer=transformer, dtype=torch.bfloat16
    )
    <span class="hljs-comment"># The transformer is already on its device; move the remaining components individually. Do not call</span>
    <span class="hljs-comment"># `pipeline.to(device)` — that would move every rank&#x27;s shards onto the same device.</span>
    pipeline.text_encoder.to(device)
    pipeline.vae.to(device)

    generator = torch.Generator().manual_seed(<span class="hljs-number">42</span>)
    image = pipeline(prompt=<span class="hljs-string">&quot;a cat holding a sign that says hello&quot;</span>, generator=generator).images[<span class="hljs-number">0</span>]
    <span class="hljs-keyword">if</span> rank == <span class="hljs-number">0</span>:
        image.save(<span class="hljs-string">&quot;output.png&quot;</span>)
    dist.destroy_process_group()

<span class="hljs-keyword">if</span> __name__ == <span class="hljs-string">&quot;__main__&quot;</span>:
    main()<!----></pre></div><!----> <p>将脚本保存为 <code>tensor_parallel_flux2.py</code>，并使用 <code>torchrun</code> 启动它。</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> 已复制</div><!----></button><!----></div> <pre class="language-shell "><!---->torchrun --nproc-per-node 4 tensor_parallel_flux2.py<!----></pre></div><!----> <p>该脚本将 <code>tp_degree</code> 设置为 <code>world_size</code>，因此 <code>--nproc-per-node 4</code> 会将 transformer 分片到 4 个 GPU 上。</p> <p>仅对 transformer 进行分片。每个 rank 都持有文本编码器的完整副本，因此每个 GPU 需要足够的内存来容纳文本编码器及其所分配的 transformer 部分。</p> <!--[1--><h2 class="relative group"><a id="loading-performance" 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="#loading-performance"><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><!---->加载性能<!----></span></h2><!--]--><!----> <p>使用 <code>parallel_config</code> 进行加载比加载完整模型后再调用 <code>enable_parallelism()</code> 速度更快，且每个 rank 占用的 CPU 内存更少。下表展示了 <a href="https://huggingface.co/black-forest-labs/FLUX.2-dev" rel="nofollow">black-forest-labs/FLUX.2-dev</a>（仅 transformer，32B 参数，bf16）在 4x A10G（23GB）上的结果。</p> <table><thead><tr><th>tp_degree</th><th>方法</th><th>加载时间</th><th>每个 rank 的峰值 CPU</th></tr></thead><tbody><tr><td>4</td><td><code>from_pretrained(parallel_config=...)</code></td><td>12.5s</td><td>6.8GB</td></tr><tr><td>4</td><td><code>from_pretrained</code> + <code>enable_parallelism</code></td><td>30.4s</td><td>64.1GB</td></tr></tbody></table> <!--[1--><h2 class="relative group"><a id="limitations" 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="#limitations"><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><!---->局限性<!----></span></h2><!--]--><!----> <p>张量并行与某些加载选项和功能不兼容，并且无法保存分片模型。</p> <ul><li>使用张量并行的 <code>parallel_config</code> 加载时，与 <code>device_map</code>、<code>quantization_config</code>、<code>low_cpu_mem_usage=False</code>、<code>use_flashpack=True</code> 或非 safetensors 权重不兼容。上述任一情况都会引发错误，而不是静默加载完整的检查点。</li> <li>将张量并行与量化、卸载或 LoRA 适配器结合使用时，无论模型如何分片，都会引发错误。</li> <li>要对已在内存中的模型进行分片，请使用相同的配置调用 <code>enable_parallelism()</code>。它会先加载完整模型，然后再进行分片，因此每个 rank 都需要容纳完整检查点的内存。使用 <code>parallel_config</code> 进行加载可以避免这一开销。</li> <li>保存张量并行模型时，<a href="/docs/diffusers/v0.43.0/zh/api/models/overview#diffusers.ModelMixin.save_pretrained">save_pretrained()</a> 会引发错误。请在分片之前保存模型。</li></ul> <!--[1--><h2 class="relative group"><a id="add-tensor-parallelism-to-a-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="#add-tensor-parallelism-to-a-model"><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><!---->为模型添加张量并行<!----></span></h2><!--]--><!----> <p>本节面向希望为 Diffusers 模型添加张量并行支持的贡献者。</p> <p>张量并行仅适用于定义了 <code>_tp_plan</code> 的模型，这是一个将模块名称映射到分片样式的类属性。编写该属性通常意味着将每个 <em>扩展</em>隐藏维度的投影与将其 <em>收缩</em>回原状的投影配对。</p> <p>每个键最多可以包含一个 <code>*</code>，且 <code>*</code> 之前的部分必须解析为一个 <a href="https://pytorch.org/docs/stable/generated/torch.nn.ModuleList.html" rel="nofollow"><code>nn.ModuleList</code></a>，因此一个条目可以覆盖所有块。不包含 <code>*</code> 的键应用于模型本身。路径是相对于模型的。</p> <!--[2--><h3 class="relative group"><a id="colwise-and-rowwise" 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="#colwise-and-rowwise"><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><!---->列向和行向<!----></span></h3><!--]--><!----> <table><thead><tr><th>风格</th><th>分片</th><th>每个 rank</th><th>用于</th></tr></thead><tbody><tr><td><code>"colwise"</code></td><td>输出特征（<code>weight</code> 的第 0 维）</td><td>计算输出的一个切片</td><td><code>to_q</code>、<code>to_k</code>、<code>to_v</code>、FFN 输入投影</td></tr><tr><td><code>"rowwise"</code></td><td>输入特征（<code>weight</code> 维度 1）</td><td>计算部分和</td><td><code>to_out.0</code>，FFN 输出投影</td></tr><tr><td><code>"rowwise_split_input"</code></td><td>输入特征（<code>weight</code> 维度 1）</td><td>从其复制输入的切片中计算部分和</td><td>一个独立的投影层，其输入是复制的，例如读取时间步嵌入的 AdaLN 调制投影</td></tr></tbody></table> <p>请务必按顺序将 <code>"colwise"</code> 与 <code>"rowwise"</code> 配对使用。<code>"colwise"</code> 投影会使其输出保持分片状态，紧随其后的 <code>"rowwise"</code> 投影会直接消费该分片，并在块边界处执行一次 <code>AllReduce</code> 来重构结果。若以其他方式对该组合进行分片，则会在中间强制触发一次 gather 操作，从而导致通信量大幅增加。</p> <p>在注意力机制中，按列划分的 Q、K 和 V 投影为每个 rank 分配一组头，因此 <code>tp_degree</code> 必须能整除头的数量。编码器流的重复部分（<code>add_q_proj</code>、<code>to_add_out</code>、<code>ff_context</code>）遵循与其图像流对应部分相同的模式。</p> <!--[2--><h3 class="relative group"><a id="fused-projections" 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-projections"><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><!---->融合投影<!----></span></h3><!--]--><!----> <p>当一个 <code>Linear</code> 沿分片维度打包了多个逻辑张量时，普通的 <code>"colwise"</code>/<code>"rowwise"</code> 会直接跨拼接进行切片，导致各部分错位。请改用 <code>PackedColwiseParallel</code>/<code>PackedRowwiseParallel</code>，它们会独立地对每个打包块进行分片。</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> 已复制</div><!----></button><!----></div> <pre class="language-py "><!----><span class="hljs-comment"># in src/diffusers/models/transformers/your_model.py</span>
<span class="hljs-keyword">from</span> ...hooks.tensor_parallel <span class="hljs-keyword">import</span> PackedColwiseParallel, PackedRowwiseParallel<!----></pre></div><!----> <p>传入 <code>blocks</code>，这是一个成比例整数列表，其总和必须能整除打包维度。例如，对于等分的 SwiGLU 门控和上投影，使用 <code>[1, 1]</code>；对于融合的 Q、K、V、门控和上投影且 <code>mlp_ratio=3</code>，使用 <code>[1, 1, 1, 3, 3]</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 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> 已复制</div><!----></button><!----></div> <pre class="language-py "><!----><span class="hljs-string">&quot;transformer_blocks.*.ff.linear_in&quot;</span>: PackedColwiseParallel([<span class="hljs-number">1</span>, <span class="hljs-number">1</span>]),<!----></pre></div><!----> <p>当块大小仅能从配置中获知时，请省略 <code>blocks</code>，并改为在 <code>__init__</code> 期间将绝对大小存储在 <code>Linear</code> 上，作为 <code>_tp_packed_col_blocks</code> 或 <code>_tp_packed_row_blocks</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 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> 已复制</div><!----></button><!----></div> <pre class="language-py "><!----><span class="hljs-comment"># in the attention module&#x27;s __init__</span>
<span class="hljs-variable language_">self</span>.to_out._tp_packed_row_blocks = [<span class="hljs-variable language_">self</span>.inner_dim, <span class="hljs-variable language_">self</span>.mlp_hidden_dim]<!----></pre></div><!----> <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> 已复制</div><!----></button><!----></div> <pre class="language-py "><!----><span class="hljs-comment"># in the model&#x27;s _tp_plan</span>
<span class="hljs-string">&quot;single_transformer_blocks.*.attn.to_out&quot;</span>: PackedRowwiseParallel(),<!----></pre></div><!----> <p><a href="https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/transformers/transformer_flux2.py" rel="nofollow"><code>transformer_flux2.py</code></a> 中的计划同时使用了这两种形式。双流 SwiGLU 对直接传递 <code>blocks</code>。融合的单流 QKV 和 MLP 投影将其尺寸存储在 <code>Flux2ParallelSelfAttention.__init__</code> 中的 <code>Linear</code> 上，因为它们依赖于 <code>mlp_ratio</code> 和 <code>mlp_mult_factor</code>。</p> <!--[2--><h3 class="relative group"><a id="what-to-leave-out" 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="#what-to-leave-out"><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><!---->需要排除的内容<!----></span></h3><!--]--><!----> <p>未包含在计划中的模块会在每个 rank 上复制。请排除归一化层、patch 和文本嵌入，以及最终的 <code>norm_out</code> 和 <code>proj_out</code>。这些层很小，对它们进行分片节省的内存很少，却会增加通信开销。AdaLN 调制（<code>img_mod</code> 和 <code>txt_mod</code>）通常也足够小，可以排除。当它很大时，例如 MiniMax-H3 的 <code>adaln_proj.linear</code>，请使用 <code>"rowwise_split_input"</code> 对其进行分片。</p> <!--[2--><h3 class="relative group"><a id="constraints-and-verification" 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="#constraints-and-verification"><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><!---->约束与验证<!----></span></h3><!--]--><!----> <ul><li><code>tp_degree</code> 必须能整除 <code>config.num_attention_heads</code>。</li> <li>每个打包块都必须<em>单独</em>能被<code>tp_degree</code>整除，而不仅仅是它们的总和。</li></ul> <p><a href="/docs/diffusers/v0.43.0/zh/api/models/overview#diffusers.ModelMixin.from_pretrained">from_pretrained()</a> 和 <code>enable_parallelism()</code> 会在加载或分片任何权重之前同时检查这两项。</p> <p>通过数值验证而非目视检查来验证新方案：在单个设备上使用固定种子生成，然后在张量并行下再次生成，并比较输出。错误放置的 <code>"colwise"</code>/<code>"rowwise"</code> 通常仍能运行，并生成看似合理但实际错误的图像。</p> <blockquote class="tip"><p>从一个类似架构的现有计划开始。</p> <ul><li><a href="https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/transformers/transformer_qwenimage.py" rel="nofollow"><code>transformer_qwenimage.py</code></a> 没有融合投影，因此每个条目都是普通的 <code>"colwise"</code> 或 <code>"rowwise"</code>。</li> <li><a href="https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/transformers/transformer_flux.py" rel="nofollow"><code>transformer_flux.py</code></a> 添加了一个打包的行投影。</li> <li><a href="https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/transformers/transformer_flux2.py" rel="nofollow"><code>transformer_flux2.py</code></a> 同时使用了两种打包样式。</li></ul></blockquote> <a class="!text-gray-400 !no-underline text-sm flex items-center not-prose mt-4" href="https://github.com/huggingface/diffusers/blob/main/docs/source/en/optimization/tensor_parallelism.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_1qsdbwf = {
						base: "/docs/diffusers/main/en",
						assets: "/docs/diffusers/main/en"
					};
					const element = document.currentScript.parentElement;
					Promise.all([
						import("/docs/diffusers/main/en/_app/immutable/entry/start.0S0_iEoG.js"),
						import("/docs/diffusers/main/en/_app/immutable/entry/app.CPdOxCbw.js")
					]).then(([kit, app]) => {
						kit.start(app, element, {
							node_ids: [0, 347],
							data: [null,null],
							form: null,
							error: null
						});
					});
				}
			</script>
		
Xet Storage Details

Size:: 38.1 kB
Xet hash:: bdb7313c73f424b9965c291d2eec428cfb671ada1d1483731e8c338d5aa4e7db
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.