Hugging Face's logo

Buckets:
hf-doc-build
/
doc-dev

Files
xet
hf-doc-build/doc-dev / diffusers /pr_11739 /en /optimization /para_attn.html
rtrm's picture
rtrm
download
raw
36.9 kB
<meta charset="utf-8" /><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;ParaAttention&quot;,&quot;local&quot;:&quot;paraattention&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;First Block Cache&quot;,&quot;local&quot;:&quot;first-block-cache&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;fp8 quantization&quot;,&quot;local&quot;:&quot;fp8-quantization&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Context Parallelism&quot;,&quot;local&quot;:&quot;context-parallelism&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Benchmarks&quot;,&quot;local&quot;:&quot;benchmarks&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2}],&quot;depth&quot;:1}">
 <link href="/docs/diffusers/pr_11739/en/_app/immutable/assets/0.e3b0c442.css" rel="modulepreload">
 <link rel="modulepreload" href="/docs/diffusers/pr_11739/en/_app/immutable/entry/start.56da5b2b.js">
 <link rel="modulepreload" href="/docs/diffusers/pr_11739/en/_app/immutable/chunks/scheduler.53228c21.js">
 <link rel="modulepreload" href="/docs/diffusers/pr_11739/en/_app/immutable/chunks/singletons.70d778ba.js">
 <link rel="modulepreload" href="/docs/diffusers/pr_11739/en/_app/immutable/chunks/index.e93d0901.js">
 <link rel="modulepreload" href="/docs/diffusers/pr_11739/en/_app/immutable/chunks/paths.a0989064.js">
 <link rel="modulepreload" href="/docs/diffusers/pr_11739/en/_app/immutable/entry/app.794fd245.js">
 <link rel="modulepreload" href="/docs/diffusers/pr_11739/en/_app/immutable/chunks/preload-helper.b5c24ab3.js">
 <link rel="modulepreload" href="/docs/diffusers/pr_11739/en/_app/immutable/chunks/index.100fac89.js">
 <link rel="modulepreload" href="/docs/diffusers/pr_11739/en/_app/immutable/nodes/0.d68fe2c3.js">
 <link rel="modulepreload" href="/docs/diffusers/pr_11739/en/_app/immutable/chunks/each.e59479a4.js">
 <link rel="modulepreload" href="/docs/diffusers/pr_11739/en/_app/immutable/nodes/287.50f85794.js">
 <link rel="modulepreload" href="/docs/diffusers/pr_11739/en/_app/immutable/chunks/CopyLLMTxtMenu.ed0e3681.js">
 <link rel="modulepreload" href="/docs/diffusers/pr_11739/en/_app/immutable/chunks/globals.7f7f1b26.js">
 <link rel="modulepreload" href="/docs/diffusers/pr_11739/en/_app/immutable/chunks/IconCopy.38cf8f56.js">
 <link rel="modulepreload" href="/docs/diffusers/pr_11739/en/_app/immutable/chunks/MermaidChart.svelte_svelte_type_style_lang.dd42f483.js">
 <link rel="modulepreload" href="/docs/diffusers/pr_11739/en/_app/immutable/chunks/CodeBlock.d30a6509.js">
 <link rel="modulepreload" href="/docs/diffusers/pr_11739/en/_app/immutable/chunks/HfOption.fad27e59.js"><!-- HEAD_svelte-u9bgzb_START --><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;ParaAttention&quot;,&quot;local&quot;:&quot;paraattention&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;First Block Cache&quot;,&quot;local&quot;:&quot;first-block-cache&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;fp8 quantization&quot;,&quot;local&quot;:&quot;fp8-quantization&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Context Parallelism&quot;,&quot;local&quot;:&quot;context-parallelism&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Benchmarks&quot;,&quot;local&quot;:&quot;benchmarks&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 max-sm:gap-0.5 h-6 max-sm:h-5 px-2 max-sm:px-1.5 text-[11px] max-sm:text-[9px] 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"><svg class="w-3 h-3 max-sm:w-2.5 max-sm:h-2.5" 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-6 max-sm:h-5 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 w-3 h-3 max-sm:w-2.5 max-sm:h-2.5 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="paraattention" 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="#paraattention"><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>ParaAttention</span></h1> <div class="flex justify-center" data-svelte-h="svelte-1p4slnk"><img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/para-attn/flux-performance.png"></div> <div class="flex justify-center" data-svelte-h="svelte-1kqq4mt"><img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/para-attn/hunyuan-video-performance.png"></div> <p data-svelte-h="svelte-uizfax">Large image and video generation models, such as <a href="https://huggingface.co/black-forest-labs/FLUX.1-dev" rel="nofollow">FLUX.1-dev</a> and <a href="https://huggingface.co/tencent/HunyuanVideo" rel="nofollow">HunyuanVideo</a>, can be an inference challenge for real-time applications and deployment because of their size.</p> <p data-svelte-h="svelte-1ufw59x"><a href="https://github.com/chengzeyi/ParaAttention" rel="nofollow">ParaAttention</a> is a library that implements <strong>context parallelism</strong> and <strong>first block cache</strong>, and can be combined with other techniques (torch.compile, fp8 dynamic quantization), to accelerate inference.</p> <p data-svelte-h="svelte-y6ktum">This guide will show you how to apply ParaAttention to FLUX.1-dev and HunyuanVideo on NVIDIA L20 GPUs.
No optimizations are applied for our baseline benchmark, except for HunyuanVideo to avoid out-of-memory errors.</p> <p data-svelte-h="svelte-12v81em">Our baseline benchmark shows that FLUX.1-dev is able to generate a 1024x1024 resolution image in 28 steps in 26.36 seconds, and HunyuanVideo is able to generate 129 frames at 720p resolution in 30 steps in 3675.71 seconds.</p> <blockquote class="tip" data-svelte-h="svelte-au4ots"><p>For even faster inference with context parallelism, try using NVIDIA A100 or H100 GPUs (if available) with NVLink support, especially when there is a large number of GPUs.</p></blockquote> <h2 class="relative group"><a id="first-block-cache" 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="#first-block-cache"><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>First Block Cache</span></h2> <p data-svelte-h="svelte-zipcqj">Caching the output of the transformers blocks in the model and reusing them in the next inference steps reduces the computation cost and makes inference faster.</p> <p data-svelte-h="svelte-11q05n5">However, it is hard to decide when to reuse the cache to ensure quality generated images or videos. ParaAttention directly uses the <strong>residual difference of the first transformer block output</strong> to approximate the difference among model outputs. When the difference is small enough, the residual difference of previous inference steps is reused. In other words, the denoising step is skipped.</p> <p data-svelte-h="svelte-1ecvggt">This achieves a 2x speedup on FLUX.1-dev and HunyuanVideo inference with very good quality.</p> <figure data-svelte-h="svelte-1czvnic"><img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/para-attn/ada-cache.png" alt="Cache in Diffusion Transformer"> <figcaption>How AdaCache works, First Block Cache is a variant of it</figcaption></figure> <div class="flex space-x-2 items-center my-1.5 mr-8 h-7 !pl-0 -mx-3 md:mx-0"><div class="flex items-center border rounded-lg px-1.5 py-1 leading-none select-none text-smd border-gray-800 bg-black dark:bg-gray-700 text-white">FLUX-1.dev </div><div class="flex items-center border rounded-lg px-1.5 py-1 leading-none select-none text-smd text-gray-500 cursor-pointer opacity-90 hover:text-gray-700 dark:hover:text-gray-200 hover:shadow-sm">HunyuanVideo </div></div> <div class="language-select"><p data-svelte-h="svelte-1gztnit">To apply first block cache on FLUX.1-dev, call <code>apply_cache_on_pipe</code> as shown below. 0.08 is the default residual difference value for FLUX models.</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> time
<span class="hljs-keyword">import</span> torch
<span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> FluxPipeline
pipe = FluxPipeline.from_pretrained(
 <span class="hljs-string">&quot;black-forest-labs/FLUX.1-dev&quot;</span>,
 torch_dtype=torch.bfloat16,
).to(<span class="hljs-string">&quot;cuda&quot;</span>)
<span class="hljs-keyword">from</span> para_attn.first_block_cache.diffusers_adapters <span class="hljs-keyword">import</span> apply_cache_on_pipe
apply_cache_on_pipe(pipe, residual_diff_threshold=<span class="hljs-number">0.08</span>)
<span class="hljs-comment"># Enable memory savings</span>
<span class="hljs-comment"># pipe.enable_model_cpu_offload()</span>
<span class="hljs-comment"># pipe.enable_sequential_cpu_offload()</span>
begin = time.time()
image = pipe(
 <span class="hljs-string">&quot;A cat holding a sign that says hello world&quot;</span>,
 num_inference_steps=<span class="hljs-number">28</span>,
).images[<span class="hljs-number">0</span>]
end = time.time()
<span class="hljs-built_in">print</span>(<span class="hljs-string">f&quot;Time: <span class="hljs-subst">{end - begin:<span class="hljs-number">.2</span>f}</span>s&quot;</span>)
<span class="hljs-built_in">print</span>(<span class="hljs-string">&quot;Saving image to flux.png&quot;</span>)
image.save(<span class="hljs-string">&quot;flux.png&quot;</span>)<!-- HTML_TAG_END --></pre></div> <table data-svelte-h="svelte-oatp7r"><thead><tr><th>Optimizations</th> <th>Original</th> <th>FBCache rdt=0.06</th> <th>FBCache rdt=0.08</th> <th>FBCache rdt=0.10</th> <th>FBCache rdt=0.12</th></tr></thead> <tbody><tr><td>Preview</td> <td><img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/para-attn/flux-original.png" alt="Original"></td> <td><img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/para-attn/flux-fbc-0.06.png" alt="FBCache rdt=0.06"></td> <td><img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/para-attn/flux-fbc-0.08.png" alt="FBCache rdt=0.08"></td> <td><img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/para-attn/flux-fbc-0.10.png" alt="FBCache rdt=0.10"></td> <td><img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/para-attn/flux-fbc-0.12.png" alt="FBCache rdt=0.12"></td></tr> <tr><td>Wall Time (s)</td> <td>26.36</td> <td>21.83</td> <td>17.01</td> <td>16.00</td> <td>13.78</td></tr></tbody></table> <p data-svelte-h="svelte-1igi2g7">First Block Cache reduced the inference speed to 17.01 seconds compared to the baseline, or 1.55x faster, while maintaining nearly zero quality loss.</p> </div> <h2 class="relative group"><a id="fp8-quantization" 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="#fp8-quantization"><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>fp8 quantization</span></h2> <p data-svelte-h="svelte-yi5eo1">fp8 with dynamic quantization further speeds up inference and reduces memory usage. Both the activations and weights must be quantized in order to use the 8-bit <a href="https://www.nvidia.com/en-us/data-center/tensor-cores/" rel="nofollow">NVIDIA Tensor Cores</a>.</p> <p data-svelte-h="svelte-10ndi9a">Use <code>float8_weight_only</code> and <code>float8_dynamic_activation_float8_weight</code> to quantize the text encoder and transformer model.</p> <p data-svelte-h="svelte-1dzea4g">The default quantization method is per tensor quantization, but if your GPU supports row-wise quantization, you can also try it for better accuracy.</p> <p data-svelte-h="svelte-edf8tk">Install <a href="https://github.com/pytorch/ao/tree/main" rel="nofollow">torchao</a> with the command below.</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 -->pip3 install -U torch torchao<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-xjm2cj"><a href="https://pytorch.org/tutorials/intermediate/torch_compile_tutorial.html" rel="nofollow">torch.compile</a> with <code>mode=&quot;max-autotune-no-cudagraphs&quot;</code> or <code>mode=&quot;max-autotune&quot;</code> selects the best kernel for performance. Compilation can take a long time if it’s the first time the model is called, but it is worth it once the model has been compiled.</p> <p data-svelte-h="svelte-1akf52n">This example only quantizes the transformer model, but you can also quantize the text encoder to reduce memory usage even more.</p> <blockquote class="tip" data-svelte-h="svelte-6flf2e"><p>Dynamic quantization can significantly change the distribution of the model output, so you need to change the <code>residual_diff_threshold</code> to a larger value for it to take effect.</p></blockquote> <div class="flex space-x-2 items-center my-1.5 mr-8 h-7 !pl-0 -mx-3 md:mx-0"><div class="flex items-center border rounded-lg px-1.5 py-1 leading-none select-none text-smd border-gray-800 bg-black dark:bg-gray-700 text-white">FLUX-1.dev </div><div class="flex items-center border rounded-lg px-1.5 py-1 leading-none select-none text-smd text-gray-500 cursor-pointer opacity-90 hover:text-gray-700 dark:hover:text-gray-200 hover:shadow-sm">HunyuanVideo </div></div> <div class="language-select"><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> time
<span class="hljs-keyword">import</span> torch
<span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> FluxPipeline
pipe = FluxPipeline.from_pretrained(
 <span class="hljs-string">&quot;black-forest-labs/FLUX.1-dev&quot;</span>,
 torch_dtype=torch.bfloat16,
).to(<span class="hljs-string">&quot;cuda&quot;</span>)
<span class="hljs-keyword">from</span> para_attn.first_block_cache.diffusers_adapters <span class="hljs-keyword">import</span> apply_cache_on_pipe
apply_cache_on_pipe(
 pipe,
 residual_diff_threshold=<span class="hljs-number">0.12</span>, <span class="hljs-comment"># Use a larger value to make the cache take effect</span>
)
<span class="hljs-keyword">from</span> torchao.quantization <span class="hljs-keyword">import</span> quantize_, float8_dynamic_activation_float8_weight, float8_weight_only
quantize_(pipe.text_encoder, float8_weight_only())
quantize_(pipe.transformer, float8_dynamic_activation_float8_weight())
pipe.transformer = torch.<span class="hljs-built_in">compile</span>(
 pipe.transformer, mode=<span class="hljs-string">&quot;max-autotune-no-cudagraphs&quot;</span>,
)
<span class="hljs-comment"># Enable memory savings</span>
<span class="hljs-comment"># pipe.enable_model_cpu_offload()</span>
<span class="hljs-comment"># pipe.enable_sequential_cpu_offload()</span>
<span class="hljs-keyword">for</span> i <span class="hljs-keyword">in</span> <span class="hljs-built_in">range</span>(<span class="hljs-number">2</span>):
 begin = time.time()
 image = pipe(
 <span class="hljs-string">&quot;A cat holding a sign that says hello world&quot;</span>,
 num_inference_steps=<span class="hljs-number">28</span>,
 ).images[<span class="hljs-number">0</span>]
 end = time.time()
 <span class="hljs-keyword">if</span> i == <span class="hljs-number">0</span>:
 <span class="hljs-built_in">print</span>(<span class="hljs-string">f&quot;Warm up time: <span class="hljs-subst">{end - begin:<span class="hljs-number">.2</span>f}</span>s&quot;</span>)
 <span class="hljs-keyword">else</span>:
 <span class="hljs-built_in">print</span>(<span class="hljs-string">f&quot;Time: <span class="hljs-subst">{end - begin:<span class="hljs-number">.2</span>f}</span>s&quot;</span>)
<span class="hljs-built_in">print</span>(<span class="hljs-string">&quot;Saving image to flux.png&quot;</span>)
image.save(<span class="hljs-string">&quot;flux.png&quot;</span>)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-2k6y70">fp8 dynamic quantization and torch.compile reduced the inference speed to 7.56 seconds compared to the baseline, or 3.48x faster.</p> </div> <h2 class="relative group"><a id="context-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="#context-parallelism"><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>Context Parallelism</span></h2> <p data-svelte-h="svelte-vw4vn">Context Parallelism parallelizes inference and scales with multiple GPUs. The ParaAttention compositional design allows you to combine Context Parallelism with First Block Cache and dynamic quantization.</p> <blockquote class="tip" data-svelte-h="svelte-1jqb1qc"><p>Refer to the <a href="https://github.com/chengzeyi/ParaAttention/tree/main" rel="nofollow">ParaAttention</a> repository for detailed instructions and examples of how to scale inference with multiple GPUs.</p></blockquote> <p data-svelte-h="svelte-1vj39hr">If the inference process needs to be persistent and serviceable, it is suggested to use <a href="https://pytorch.org/docs/stable/multiprocessing.html" rel="nofollow">torch.multiprocessing</a> to write your own inference processor. This can eliminate the overhead of launching the process and loading and recompiling the model.</p> <div class="flex space-x-2 items-center my-1.5 mr-8 h-7 !pl-0 -mx-3 md:mx-0"><div class="flex items-center border rounded-lg px-1.5 py-1 leading-none select-none text-smd border-gray-800 bg-black dark:bg-gray-700 text-white">FLUX-1.dev </div><div class="flex items-center border rounded-lg px-1.5 py-1 leading-none select-none text-smd text-gray-500 cursor-pointer opacity-90 hover:text-gray-700 dark:hover:text-gray-200 hover:shadow-sm">HunyuanVideo </div></div> <div class="language-select"><p data-svelte-h="svelte-l7izr3">The code sample below combines First Block Cache, fp8 dynamic quantization, torch.compile, and Context Parallelism for the fastest inference speed.</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> time
<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> FluxPipeline
dist.init_process_group()
torch.cuda.set_device(dist.get_rank())
pipe = FluxPipeline.from_pretrained(
 <span class="hljs-string">&quot;black-forest-labs/FLUX.1-dev&quot;</span>,
 torch_dtype=torch.bfloat16,
).to(<span class="hljs-string">&quot;cuda&quot;</span>)
<span class="hljs-keyword">from</span> para_attn.context_parallel <span class="hljs-keyword">import</span> init_context_parallel_mesh
<span class="hljs-keyword">from</span> para_attn.context_parallel.diffusers_adapters <span class="hljs-keyword">import</span> parallelize_pipe
<span class="hljs-keyword">from</span> para_attn.parallel_vae.diffusers_adapters <span class="hljs-keyword">import</span> parallelize_vae
mesh = init_context_parallel_mesh(
 pipe.device.<span class="hljs-built_in">type</span>,
 max_ring_dim_size=<span class="hljs-number">2</span>,
)
parallelize_pipe(
 pipe,
 mesh=mesh,
)
parallelize_vae(pipe.vae, mesh=mesh._flatten())
<span class="hljs-keyword">from</span> para_attn.first_block_cache.diffusers_adapters <span class="hljs-keyword">import</span> apply_cache_on_pipe
apply_cache_on_pipe(
 pipe,
 residual_diff_threshold=<span class="hljs-number">0.12</span>, <span class="hljs-comment"># Use a larger value to make the cache take effect</span>
)
<span class="hljs-keyword">from</span> torchao.quantization <span class="hljs-keyword">import</span> quantize_, float8_dynamic_activation_float8_weight, float8_weight_only
quantize_(pipe.text_encoder, float8_weight_only())
quantize_(pipe.transformer, float8_dynamic_activation_float8_weight())
torch._inductor.config.reorder_for_compute_comm_overlap = <span class="hljs-literal">True</span>
pipe.transformer = torch.<span class="hljs-built_in">compile</span>(
 pipe.transformer, mode=<span class="hljs-string">&quot;max-autotune-no-cudagraphs&quot;</span>,
)
<span class="hljs-comment"># Enable memory savings</span>
<span class="hljs-comment"># pipe.enable_model_cpu_offload(gpu_id=dist.get_rank())</span>
<span class="hljs-comment"># pipe.enable_sequential_cpu_offload(gpu_id=dist.get_rank())</span>
<span class="hljs-keyword">for</span> i <span class="hljs-keyword">in</span> <span class="hljs-built_in">range</span>(<span class="hljs-number">2</span>):
 begin = time.time()
 image = pipe(
 <span class="hljs-string">&quot;A cat holding a sign that says hello world&quot;</span>,
 num_inference_steps=<span class="hljs-number">28</span>,
 output_type=<span class="hljs-string">&quot;pil&quot;</span> <span class="hljs-keyword">if</span> dist.get_rank() == <span class="hljs-number">0</span> <span class="hljs-keyword">else</span> <span class="hljs-string">&quot;pt&quot;</span>,
 ).images[<span class="hljs-number">0</span>]
 end = time.time()
 <span class="hljs-keyword">if</span> dist.get_rank() == <span class="hljs-number">0</span>:
 <span class="hljs-keyword">if</span> i == <span class="hljs-number">0</span>:
 <span class="hljs-built_in">print</span>(<span class="hljs-string">f&quot;Warm up time: <span class="hljs-subst">{end - begin:<span class="hljs-number">.2</span>f}</span>s&quot;</span>)
 <span class="hljs-keyword">else</span>:
 <span class="hljs-built_in">print</span>(<span class="hljs-string">f&quot;Time: <span class="hljs-subst">{end - begin:<span class="hljs-number">.2</span>f}</span>s&quot;</span>)
<span class="hljs-keyword">if</span> dist.get_rank() == <span class="hljs-number">0</span>:
 <span class="hljs-built_in">print</span>(<span class="hljs-string">&quot;Saving image to flux.png&quot;</span>)
 image.save(<span class="hljs-string">&quot;flux.png&quot;</span>)
dist.destroy_process_group()<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1gsxpvn">Save to <code>run_flux.py</code> and launch it with <a href="https://pytorch.org/docs/stable/elastic/run.html" rel="nofollow">torchrun</a>.</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-comment"># Use --nproc_per_node to specify the number of GPUs</span>
torchrun --nproc_per_node=2 run_flux.py<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-11xwpjh">Inference speed is reduced to 8.20 seconds compared to the baseline, or 3.21x faster, with 2 NVIDIA L20 GPUs. On 4 L20s, inference speed is 3.90 seconds, or 6.75x faster.</p> </div> <h2 class="relative group"><a id="benchmarks" 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="#benchmarks"><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>Benchmarks</span></h2> <div class="flex space-x-2 items-center my-1.5 mr-8 h-7 !pl-0 -mx-3 md:mx-0"><div class="flex items-center border rounded-lg px-1.5 py-1 leading-none select-none text-smd border-gray-800 bg-black dark:bg-gray-700 text-white">FLUX-1.dev </div><div class="flex items-center border rounded-lg px-1.5 py-1 leading-none select-none text-smd text-gray-500 cursor-pointer opacity-90 hover:text-gray-700 dark:hover:text-gray-200 hover:shadow-sm">HunyuanVideo </div></div> <div class="language-select"><table data-svelte-h="svelte-1eiq4i6"><thead><tr><th>GPU Type</th> <th>Number of GPUs</th> <th>Optimizations</th> <th>Wall Time (s)</th> <th>Speedup</th></tr></thead> <tbody><tr><td>NVIDIA L20</td> <td>1</td> <td>Baseline</td> <td>26.36</td> <td>1.00x</td></tr> <tr><td>NVIDIA L20</td> <td>1</td> <td>FBCache (rdt=0.08)</td> <td>17.01</td> <td>1.55x</td></tr> <tr><td>NVIDIA L20</td> <td>1</td> <td>FP8 DQ</td> <td>13.40</td> <td>1.96x</td></tr> <tr><td>NVIDIA L20</td> <td>1</td> <td>FBCache (rdt=0.12) + FP8 DQ</td> <td>7.56</td> <td>3.48x</td></tr> <tr><td>NVIDIA L20</td> <td>2</td> <td>FBCache (rdt=0.12) + FP8 DQ + CP</td> <td>4.92</td> <td>5.35x</td></tr> <tr><td>NVIDIA L20</td> <td>4</td> <td>FBCache (rdt=0.12) + FP8 DQ + CP</td> <td>3.90</td> <td>6.75x</td></tr></tbody></table> </div> <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/para_attn.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_kxvna0 = {
 assets: "/docs/diffusers/pr_11739/en",
 base: "/docs/diffusers/pr_11739/en",
 env: {}
 };
 const element = document.currentScript.parentElement;
 const data = [null,null];
 Promise.all([
 import("/docs/diffusers/pr_11739/en/_app/immutable/entry/start.56da5b2b.js"),
 import("/docs/diffusers/pr_11739/en/_app/immutable/entry/app.794fd245.js")
 ]).then(([kit, app]) => {
 kit.start(app, element, {
 node_ids: [0, 287],
 data,
 form: null,
 error: null
 });
 });
 }
 </script>

Xet Storage Details

Size:
36.9 kB
·
Xet hash:
e6a4bf19d2b896f4652f2ffdf2b2b8481c0754adc9a79e8ea181d23001586d79

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