diffusers/v0.19.2/en/optimization/fp16.html

<meta charset="utf-8" /><meta http-equiv="content-security-policy" content=""><meta name="hf:doc:metadata" content="{&quot;local&quot;:&quot;memory-and-speed&quot;,&quot;sections&quot;:[{&quot;local&quot;:null,&quot;sections&quot;:[{&quot;local&quot;:&quot;use-tf32-instead-of-fp32-on-ampere-and-later-cuda-devices&quot;,&quot;title&quot;:&quot;Use tf32 instead of fp32 (on Ampere and later CUDA devices)&quot;}],&quot;title&quot;:null},{&quot;local&quot;:&quot;half-precision-weights&quot;,&quot;title&quot;:&quot;Half precision weights&quot;},{&quot;local&quot;:&quot;sliced-attention-for-additional-memory-savings&quot;,&quot;title&quot;:&quot;Sliced attention for additional memory savings&quot;},{&quot;local&quot;:&quot;sliced-vae-decode-for-larger-batches&quot;,&quot;title&quot;:&quot;Sliced VAE decode for larger batches&quot;},{&quot;local&quot;:&quot;tiled-vae-decode-and-encode-for-large-images&quot;,&quot;title&quot;:&quot;Tiled VAE decode and encode for large images&quot;},{&quot;local&quot;:&quot;offloading-to-cpu-with-accelerate-for-memory-savings&quot;,&quot;title&quot;:&quot;Offloading to CPU with accelerate for memory savings&quot;},{&quot;local&quot;:&quot;model-offloading-for-fast-inference-and-memory-savings&quot;,&quot;title&quot;:&quot;Model offloading for fast inference and memory savings&quot;},{&quot;local&quot;:&quot;using-channels-last-memory-format&quot;,&quot;title&quot;:&quot;Using Channels Last memory format&quot;},{&quot;local&quot;:&quot;tracing&quot;,&quot;title&quot;:&quot;Tracing&quot;},{&quot;local&quot;:&quot;memory-efficient-attention&quot;,&quot;title&quot;:&quot;Memory Efficient Attention&quot;}],&quot;title&quot;:&quot;Memory and speed&quot;}" data-svelte="svelte-1phssyn">
	<link rel="modulepreload" href="/docs/diffusers/v0.19.2/en/_app/assets/pages/__layout.svelte-hf-doc-builder.css">
	<link rel="modulepreload" href="/docs/diffusers/v0.19.2/en/_app/start-hf-doc-builder.js">
	<link rel="modulepreload" href="/docs/diffusers/v0.19.2/en/_app/chunks/vendor-hf-doc-builder.js">
	<link rel="modulepreload" href="/docs/diffusers/v0.19.2/en/_app/chunks/paths-hf-doc-builder.js">
	<link rel="modulepreload" href="/docs/diffusers/v0.19.2/en/_app/pages/__layout.svelte-hf-doc-builder.js">
	<link rel="modulepreload" href="/docs/diffusers/v0.19.2/en/_app/pages/optimization/fp16.mdx-hf-doc-builder.js">
	<link rel="modulepreload" href="/docs/diffusers/v0.19.2/en/_app/chunks/Tip-hf-doc-builder.js">
	<link rel="modulepreload" href="/docs/diffusers/v0.19.2/en/_app/chunks/IconCopyLink-hf-doc-builder.js">
	<link rel="modulepreload" href="/docs/diffusers/v0.19.2/en/_app/chunks/CodeBlock-hf-doc-builder.js"> 






<h1 class="relative group"><a id="memory-and-speed" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#memory-and-speed"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a>
	<span>Memory and speed
	</span></h1>

<p>We present some techniques and ideas to optimize 🤗 Diffusers <em>inference</em> for memory or speed. As a general rule, we recommend the use of <a href="https://github.com/facebookresearch/xformers" rel="nofollow">xFormers</a> for memory efficient attention, please see the recommended <a href="xformers">installation instructions</a>.</p>
<p>We’ll discuss how the following settings impact performance and memory.</p>
<table><thead><tr><th></th>
<th>Latency</th>
<th>Speedup</th></tr></thead>
<tbody><tr><td>original</td>
<td>9.50s</td>
<td>x1</td></tr>
<tr><td>fp16</td>
<td>3.61s</td>
<td>x2.63</td></tr>
<tr><td>channels last</td>
<td>3.30s</td>
<td>x2.88</td></tr>
<tr><td>traced UNet</td>
<td>3.21s</td>
<td>x2.96</td></tr>
<tr><td>memory efficient attention</td>
<td>2.63s</td>
<td>x3.61</td></tr></tbody></table>
<em>obtained on NVIDIA TITAN RTX by generating a single image of size 512x512 from
  the prompt &quot;a photo of an astronaut riding a horse on mars&quot; with 50 DDIM
  steps.
</em>
<h3 class="relative group"><a id="use-tf32-instead-of-fp32-on-ampere-and-later-cuda-devices" 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="#use-tf32-instead-of-fp32-on-ampere-and-later-cuda-devices"><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>Use tf32 instead of fp32 (on Ampere and later CUDA devices)
	</span></h3>

<p>On Ampere and later CUDA devices matrix multiplications and convolutions can use the TensorFloat32 (TF32) mode for faster but slightly less accurate computations. By default PyTorch enables TF32 mode for convolutions but not matrix multiplications, and unless a network requires full float32 precision we recommend enabling this setting for matrix multiplications, too. It can significantly speed up computations with typically negligible loss of numerical accuracy. You can read more about it <a href="https://huggingface.co/docs/transformers/v4.18.0/en/performance#tf32" rel="nofollow">here</a>. All you need to do is to add this before your inference:</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><!-- HTML_TAG_START --><span class="hljs-keyword">import</span> torch

torch.backends.cuda.matmul.allow_tf32 = <span class="hljs-literal">True</span><!-- HTML_TAG_END --></pre></div>
<h2 class="relative group"><a id="half-precision-weights" 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="#half-precision-weights"><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>Half precision weights
	</span></h2>

<p>To save more GPU memory and get more speed, you can load and run the model weights directly in half precision. This involves loading the float16 version of the weights, which was saved to a branch named <code>fp16</code>, and telling PyTorch to use the <code>float16</code> type when loading them:</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><!-- HTML_TAG_START --><span class="hljs-keyword">import</span> torch
<span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> DiffusionPipeline

pipe = DiffusionPipeline.from_pretrained(
    <span class="hljs-string">&quot;runwayml/stable-diffusion-v1-5&quot;</span>,
    torch_dtype=torch.float16,
)
pipe = pipe.to(<span class="hljs-string">&quot;cuda&quot;</span>)

prompt = <span class="hljs-string">&quot;a photo of an astronaut riding a horse on mars&quot;</span>
image = pipe(prompt).images[<span class="hljs-number">0</span>]<!-- HTML_TAG_END --></pre></div>


<div class="course-tip course-tip-orange bg-gradient-to-br dark:bg-gradient-to-r before:border-orange-500 dark:before:border-orange-800 from-orange-50 dark:from-gray-900 to-white dark:to-gray-950 border border-orange-50 text-orange-700 dark:text-gray-400"><p>It is strongly discouraged to make use of <a href="https://pytorch.org/docs/stable/amp.html#torch.autocast" rel="nofollow"><code>torch.autocast</code></a> in any of the pipelines as it can lead to black images and is always slower than using pure
float16 precision.</p></div>
<h2 class="relative group"><a id="sliced-attention-for-additional-memory-savings" 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="#sliced-attention-for-additional-memory-savings"><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>Sliced attention for additional memory savings
	</span></h2>

<p>For even additional memory savings, you can use a sliced version of attention that performs the computation in steps instead of all at once.</p>


<div class="course-tip  bg-gradient-to-br dark:bg-gradient-to-r before:border-green-500 dark:before:border-green-800 from-green-50 dark:from-gray-900 to-white dark:to-gray-950 border border-green-50 text-green-700 dark:text-gray-400">Attention slicing is useful even if a batch size of just 1 is used - as long
  as the model uses more than one attention head. If there is more than one
  attention head the *QK^T* attention matrix can be computed sequentially for
  each head which can save a significant amount of memory.
</div>
<p>To perform the attention computation sequentially over each head, you only need to invoke <a href="/docs/diffusers/v0.19.2/en/api/pipelines/stable_diffusion/inpaint#diffusers.StableDiffusionInpaintPipeline.enable_attention_slicing">enable_attention_slicing()</a> in your pipeline before inference, like here:</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><!-- HTML_TAG_START --><span class="hljs-keyword">import</span> torch
<span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> DiffusionPipeline

pipe = DiffusionPipeline.from_pretrained(
    <span class="hljs-string">&quot;runwayml/stable-diffusion-v1-5&quot;</span>,
    torch_dtype=torch.float16,
)
pipe = pipe.to(<span class="hljs-string">&quot;cuda&quot;</span>)

prompt = <span class="hljs-string">&quot;a photo of an astronaut riding a horse on mars&quot;</span>
pipe.enable_attention_slicing()
image = pipe(prompt).images[<span class="hljs-number">0</span>]<!-- HTML_TAG_END --></pre></div>
<p>There’s a small performance penalty of about 10% slower inference times, but this method allows you to use Stable Diffusion in as little as 3.2 GB of VRAM!</p>
<h2 class="relative group"><a id="sliced-vae-decode-for-larger-batches" 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="#sliced-vae-decode-for-larger-batches"><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>Sliced VAE decode for larger batches
	</span></h2>

<p>To decode large batches of images with limited VRAM, or to enable batches with 32 images or more, you can use sliced VAE decode that decodes the batch latents one image at a time.</p>
<p>You likely want to couple this with <a href="/docs/diffusers/v0.19.2/en/api/pipelines/stable_diffusion/inpaint#diffusers.StableDiffusionInpaintPipeline.enable_attention_slicing">enable_attention_slicing()</a> or <a href="/docs/diffusers/v0.19.2/en/api/pipelines/stable_diffusion/inpaint#diffusers.StableDiffusionInpaintPipeline.enable_xformers_memory_efficient_attention">enable_xformers_memory_efficient_attention()</a> to further minimize memory use.</p>
<p>To perform the VAE decode one image at a time, invoke <a href="/docs/diffusers/v0.19.2/en/api/pipelines/stable_diffusion/text2img#diffusers.StableDiffusionPipeline.enable_vae_slicing">enable_vae_slicing()</a> in your pipeline before inference. For example:</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><!-- HTML_TAG_START --><span class="hljs-keyword">import</span> torch
<span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> StableDiffusionPipeline

pipe = StableDiffusionPipeline.from_pretrained(
    <span class="hljs-string">&quot;runwayml/stable-diffusion-v1-5&quot;</span>,
    torch_dtype=torch.float16,
)
pipe = pipe.to(<span class="hljs-string">&quot;cuda&quot;</span>)

prompt = <span class="hljs-string">&quot;a photo of an astronaut riding a horse on mars&quot;</span>
pipe.enable_vae_slicing()
images = pipe([prompt] * <span class="hljs-number">32</span>).images<!-- HTML_TAG_END --></pre></div>
<p>You may see a small performance boost in VAE decode on multi-image batches. There should be no performance impact on single-image batches.</p>
<h2 class="relative group"><a id="tiled-vae-decode-and-encode-for-large-images" 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="#tiled-vae-decode-and-encode-for-large-images"><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>Tiled VAE decode and encode for large images
	</span></h2>

<p>Tiled VAE processing makes it possible to work with large images on limited VRAM. For example, generating 4k images in 8GB of VRAM. Tiled VAE decoder splits the image into overlapping tiles, decodes the tiles, and blends the outputs to make the final image.</p>
<p>You want to couple this with <a href="/docs/diffusers/v0.19.2/en/api/pipelines/stable_diffusion/inpaint#diffusers.StableDiffusionInpaintPipeline.enable_attention_slicing">enable_attention_slicing()</a> or <a href="/docs/diffusers/v0.19.2/en/api/pipelines/stable_diffusion/inpaint#diffusers.StableDiffusionInpaintPipeline.enable_xformers_memory_efficient_attention">enable_xformers_memory_efficient_attention()</a> to further minimize memory use.</p>
<p>To use tiled VAE processing, invoke <a href="/docs/diffusers/v0.19.2/en/api/pipelines/stable_diffusion/text2img#diffusers.StableDiffusionPipeline.enable_vae_tiling">enable_vae_tiling()</a> in your pipeline before inference. For example:</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><!-- HTML_TAG_START --><span class="hljs-keyword">import</span> torch
<span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> StableDiffusionPipeline, UniPCMultistepScheduler

pipe = StableDiffusionPipeline.from_pretrained(
    <span class="hljs-string">&quot;runwayml/stable-diffusion-v1-5&quot;</span>,
    torch_dtype=torch.float16,
)
pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config)
pipe = pipe.to(<span class="hljs-string">&quot;cuda&quot;</span>)
prompt = <span class="hljs-string">&quot;a beautiful landscape photograph&quot;</span>
pipe.enable_vae_tiling()
pipe.enable_xformers_memory_efficient_attention()

image = pipe([prompt], width=<span class="hljs-number">3840</span>, height=<span class="hljs-number">2224</span>, num_inference_steps=<span class="hljs-number">20</span>).images[<span class="hljs-number">0</span>]<!-- HTML_TAG_END --></pre></div>
<p>The output image will have some tile-to-tile tone variation from the tiles having separate decoders, but you shouldn’t see sharp seams between the tiles. The tiling is turned off for images that are 512x512 or smaller.</p>
<a name="sequential_offloading"></a>
<h2 class="relative group"><a id="offloading-to-cpu-with-accelerate-for-memory-savings" 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="#offloading-to-cpu-with-accelerate-for-memory-savings"><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>Offloading to CPU with accelerate for memory savings
	</span></h2>

<p>For additional memory savings, you can offload the weights to CPU and only load them to GPU when performing the forward pass.</p>
<p>To perform CPU offloading, all you have to do is invoke <a href="/docs/diffusers/v0.19.2/en/api/pipelines/stable_diffusion/latent_upscale#diffusers.StableDiffusionLatentUpscalePipeline.enable_sequential_cpu_offload">enable_sequential_cpu_offload()</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><!-- HTML_TAG_START --><span class="hljs-keyword">import</span> torch
<span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> StableDiffusionPipeline

pipe = StableDiffusionPipeline.from_pretrained(
    <span class="hljs-string">&quot;runwayml/stable-diffusion-v1-5&quot;</span>,
    torch_dtype=torch.float16,
)

prompt = <span class="hljs-string">&quot;a photo of an astronaut riding a horse on mars&quot;</span>
pipe.enable_sequential_cpu_offload()
image = pipe(prompt).images[<span class="hljs-number">0</span>]<!-- HTML_TAG_END --></pre></div>
<p>And you can get the memory consumption to &lt; 3GB.</p>
<p>Note that this method works at the submodule level, not on whole models. This is the best way to minimize memory consumption, but inference is much slower due to the iterative nature of the process. The UNet component of the pipeline runs several times (as many as <code>num_inference_steps</code>); each time, the different submodules of the UNet are sequentially onloaded and then offloaded as they are needed, so the number of memory transfers is large.</p>


<div class="course-tip  bg-gradient-to-br dark:bg-gradient-to-r before:border-green-500 dark:before:border-green-800 from-green-50 dark:from-gray-900 to-white dark:to-gray-950 border border-green-50 text-green-700 dark:text-gray-400">Consider using <a href="#model_offloading">model offloading</a> as another point in the optimization space: it will be much faster, but memory savings won&#39;t be as large.
</div>
<p>It is also possible to chain offloading with attention slicing for minimal memory consumption (&lt; 2GB).</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><!-- HTML_TAG_START --><span class="hljs-keyword">import</span> torch
<span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> StableDiffusionPipeline

pipe = StableDiffusionPipeline.from_pretrained(
    <span class="hljs-string">&quot;runwayml/stable-diffusion-v1-5&quot;</span>,
    torch_dtype=torch.float16,
)

prompt = <span class="hljs-string">&quot;a photo of an astronaut riding a horse on mars&quot;</span>
pipe.enable_sequential_cpu_offload()
pipe.enable_attention_slicing(<span class="hljs-number">1</span>)

image = pipe(prompt).images[<span class="hljs-number">0</span>]<!-- HTML_TAG_END --></pre></div>
<p><strong>Note</strong>: When using <code>enable_sequential_cpu_offload()</code>, it is important to <strong>not</strong> move the pipeline to CUDA beforehand or else the gain in memory consumption will only be minimal. See <a href="https://github.com/huggingface/diffusers/issues/1934" rel="nofollow">this issue</a> for more information.</p>
<p><strong>Note</strong>: <code>enable_sequential_cpu_offload()</code> is a stateful operation that installs hooks on the models.</p>
<a name="model_offloading"></a>
<h2 class="relative group"><a id="model-offloading-for-fast-inference-and-memory-savings" 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="#model-offloading-for-fast-inference-and-memory-savings"><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>Model offloading for fast inference and memory savings
	</span></h2>

<p><a href="#sequential_offloading">Sequential CPU offloading</a>, as discussed in the previous section, preserves a lot of memory but makes inference slower, because submodules are moved to GPU as needed, and immediately returned to CPU when a new module runs.</p>
<p>Full-model offloading is an alternative that moves whole models to the GPU, instead of handling each model’s constituent <em>modules</em>. This results in a negligible impact on inference time (compared with moving the pipeline to <code>cuda</code>), while still providing some memory savings.</p>
<p>In this scenario, only one of the main components of the pipeline (typically: text encoder, unet and vae)
will be in the GPU while the others wait in the CPU. Components like the UNet that run for multiple iterations will stay on GPU until they are no longer needed.</p>
<p>This feature can be enabled by invoking <code>enable_model_cpu_offload()</code> on the pipeline, as shown 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><!-- HTML_TAG_START --><span class="hljs-keyword">import</span> torch
<span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> StableDiffusionPipeline

pipe = StableDiffusionPipeline.from_pretrained(
    <span class="hljs-string">&quot;runwayml/stable-diffusion-v1-5&quot;</span>,  
    torch_dtype=torch.float16,
)

prompt = <span class="hljs-string">&quot;a photo of an astronaut riding a horse on mars&quot;</span>
pipe.enable_model_cpu_offload()
image = pipe(prompt).images[<span class="hljs-number">0</span>]<!-- HTML_TAG_END --></pre></div>
<p>This is also compatible with attention slicing for additional memory savings.</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><!-- HTML_TAG_START --><span class="hljs-keyword">import</span> torch
<span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> StableDiffusionPipeline

pipe = StableDiffusionPipeline.from_pretrained(
    <span class="hljs-string">&quot;runwayml/stable-diffusion-v1-5&quot;</span>,
    torch_dtype=torch.float16,
)

prompt = <span class="hljs-string">&quot;a photo of an astronaut riding a horse on mars&quot;</span>
pipe.enable_model_cpu_offload()
pipe.enable_attention_slicing(<span class="hljs-number">1</span>)

image = pipe(prompt).images[<span class="hljs-number">0</span>]<!-- HTML_TAG_END --></pre></div>


<div class="course-tip  bg-gradient-to-br dark:bg-gradient-to-r before:border-green-500 dark:before:border-green-800 from-green-50 dark:from-gray-900 to-white dark:to-gray-950 border border-green-50 text-green-700 dark:text-gray-400">This feature requires `accelerate` version 0.17.0 or larger.
</div>
<p><strong>Note</strong>: <code>enable_model_cpu_offload()</code> is a stateful operation that installs hooks on the models and state on the pipeline. In order to properly offload
models after they are called, it is required that the entire pipeline is run and models are called in the order the pipeline expects them to be. Exercise caution
if models are re-used outside the context of the pipeline after hooks have been installed. See <a href="https://huggingface.co/docs/accelerate/v0.18.0/en/package_reference/big_modeling#accelerate.hooks.remove_hook_from_module" rel="nofollow">accelerate</a>
for further docs on removing hooks.</p>
<h2 class="relative group"><a id="using-channels-last-memory-format" 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="#using-channels-last-memory-format"><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>Using Channels Last memory format
	</span></h2>

<p>Channels last memory format is an alternative way of ordering NCHW tensors in memory preserving dimensions ordering. Channels last tensors ordered in such a way that channels become the densest dimension (aka storing images pixel-per-pixel). Since not all operators currently support channels last format it may result in a worst performance, so it’s better to try it and see if it works for your model.</p>
<p>For example, in order to set the UNet model in our pipeline to use channels last format, we can use the following:</p>

	<div class="code-block relative"><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5   text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg>
	<div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div>
	Copied</div></button></div>
	<pre><!-- HTML_TAG_START --><span class="hljs-built_in">print</span>(pipe.unet.conv_out.state_dict()[<span class="hljs-string">&quot;weight&quot;</span>].stride())  <span class="hljs-comment"># (2880, 9, 3, 1)</span>
pipe.unet.to(memory_format=torch.channels_last)  <span class="hljs-comment"># in-place operation</span>
<span class="hljs-built_in">print</span>(
    pipe.unet.conv_out.state_dict()[<span class="hljs-string">&quot;weight&quot;</span>].stride()
)  <span class="hljs-comment"># (2880, 1, 960, 320) having a stride of 1 for the 2nd dimension proves that it works</span><!-- HTML_TAG_END --></pre></div>
<h2 class="relative group"><a id="tracing" 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="#tracing"><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>Tracing
	</span></h2>

<p>Tracing runs an example input tensor through your model, and captures the operations that are invoked as that input makes its way through the model’s layers so that an executable or <code>ScriptFunction</code> is returned that will be optimized using just-in-time compilation.</p>
<p>To trace our UNet model, we can use the following:</p>

	<div class="code-block relative"><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5   text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg>
	<div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div>
	Copied</div></button></div>
	<pre><!-- 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> StableDiffusionPipeline
<span class="hljs-keyword">import</span> functools

<span class="hljs-comment"># torch disable grad</span>
torch.set_grad_enabled(<span class="hljs-literal">False</span>)

<span class="hljs-comment"># set variables</span>
n_experiments = <span class="hljs-number">2</span>
unet_runs_per_experiment = <span class="hljs-number">50</span>


<span class="hljs-comment"># load inputs</span>
<span class="hljs-keyword">def</span> <span class="hljs-title function_">generate_inputs</span>():
    sample = torch.randn(<span class="hljs-number">2</span>, <span class="hljs-number">4</span>, <span class="hljs-number">64</span>, <span class="hljs-number">64</span>).half().cuda()
    timestep = torch.rand(<span class="hljs-number">1</span>).half().cuda() * <span class="hljs-number">999</span>
    encoder_hidden_states = torch.randn(<span class="hljs-number">2</span>, <span class="hljs-number">77</span>, <span class="hljs-number">768</span>).half().cuda()
    <span class="hljs-keyword">return</span> sample, timestep, encoder_hidden_states


pipe = StableDiffusionPipeline.from_pretrained(
    <span class="hljs-string">&quot;runwayml/stable-diffusion-v1-5&quot;</span>,
    torch_dtype=torch.float16,
).to(<span class="hljs-string">&quot;cuda&quot;</span>)
unet = pipe.unet
unet.<span class="hljs-built_in">eval</span>()
unet.to(memory_format=torch.channels_last)  <span class="hljs-comment"># use channels_last memory format</span>
unet.forward = functools.partial(unet.forward, return_dict=<span class="hljs-literal">False</span>)  <span class="hljs-comment"># set return_dict=False as default</span>

<span class="hljs-comment"># warmup</span>
<span class="hljs-keyword">for</span> _ <span class="hljs-keyword">in</span> <span class="hljs-built_in">range</span>(<span class="hljs-number">3</span>):
    <span class="hljs-keyword">with</span> torch.inference_mode():
        inputs = generate_inputs()
        orig_output = unet(*inputs)

<span class="hljs-comment"># trace</span>
<span class="hljs-built_in">print</span>(<span class="hljs-string">&quot;tracing..&quot;</span>)
unet_traced = torch.jit.trace(unet, inputs)
unet_traced.<span class="hljs-built_in">eval</span>()
<span class="hljs-built_in">print</span>(<span class="hljs-string">&quot;done tracing&quot;</span>)


<span class="hljs-comment"># warmup and optimize graph</span>
<span class="hljs-keyword">for</span> _ <span class="hljs-keyword">in</span> <span class="hljs-built_in">range</span>(<span class="hljs-number">5</span>):
    <span class="hljs-keyword">with</span> torch.inference_mode():
        inputs = generate_inputs()
        orig_output = unet_traced(*inputs)


<span class="hljs-comment"># benchmarking</span>
<span class="hljs-keyword">with</span> torch.inference_mode():
    <span class="hljs-keyword">for</span> _ <span class="hljs-keyword">in</span> <span class="hljs-built_in">range</span>(n_experiments):
        torch.cuda.synchronize()
        start_time = time.time()
        <span class="hljs-keyword">for</span> _ <span class="hljs-keyword">in</span> <span class="hljs-built_in">range</span>(unet_runs_per_experiment):
            orig_output = unet_traced(*inputs)
        torch.cuda.synchronize()
        <span class="hljs-built_in">print</span>(<span class="hljs-string">f&quot;unet traced inference took <span class="hljs-subst">{time.time() - start_time:<span class="hljs-number">.2</span>f}</span> seconds&quot;</span>)
    <span class="hljs-keyword">for</span> _ <span class="hljs-keyword">in</span> <span class="hljs-built_in">range</span>(n_experiments):
        torch.cuda.synchronize()
        start_time = time.time()
        <span class="hljs-keyword">for</span> _ <span class="hljs-keyword">in</span> <span class="hljs-built_in">range</span>(unet_runs_per_experiment):
            orig_output = unet(*inputs)
        torch.cuda.synchronize()
        <span class="hljs-built_in">print</span>(<span class="hljs-string">f&quot;unet inference took <span class="hljs-subst">{time.time() - start_time:<span class="hljs-number">.2</span>f}</span> seconds&quot;</span>)

<span class="hljs-comment"># save the model</span>
unet_traced.save(<span class="hljs-string">&quot;unet_traced.pt&quot;</span>)<!-- HTML_TAG_END --></pre></div>
<p>Then we can replace the <code>unet</code> attribute of the pipeline with the traced model like the following</p>

	<div class="code-block relative"><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5   text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg>
	<div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div>
	Copied</div></button></div>
	<pre><!-- HTML_TAG_START --><span class="hljs-keyword">from</span> diffusers <span class="hljs-keyword">import</span> StableDiffusionPipeline
<span class="hljs-keyword">import</span> torch
<span class="hljs-keyword">from</span> dataclasses <span class="hljs-keyword">import</span> dataclass


<span class="hljs-meta">@dataclass</span>
<span class="hljs-keyword">class</span> <span class="hljs-title class_">UNet2DConditionOutput</span>:
    sample: torch.FloatTensor


pipe = StableDiffusionPipeline.from_pretrained(
    <span class="hljs-string">&quot;runwayml/stable-diffusion-v1-5&quot;</span>,
    torch_dtype=torch.float16,
).to(<span class="hljs-string">&quot;cuda&quot;</span>)

<span class="hljs-comment"># use jitted unet</span>
unet_traced = torch.jit.load(<span class="hljs-string">&quot;unet_traced.pt&quot;</span>)


<span class="hljs-comment"># del pipe.unet</span>
<span class="hljs-keyword">class</span> <span class="hljs-title class_">TracedUNet</span>(torch.nn.Module):
    <span class="hljs-keyword">def</span> <span class="hljs-title function_">__init__</span>(<span class="hljs-params">self</span>):
        <span class="hljs-built_in">super</span>().__init__()
        self.in_channels = pipe.unet.in_channels
        self.device = pipe.unet.device

    <span class="hljs-keyword">def</span> <span class="hljs-title function_">forward</span>(<span class="hljs-params">self, latent_model_input, t, encoder_hidden_states</span>):
        sample = unet_traced(latent_model_input, t, encoder_hidden_states)[<span class="hljs-number">0</span>]
        <span class="hljs-keyword">return</span> UNet2DConditionOutput(sample=sample)


pipe.unet = TracedUNet()

<span class="hljs-keyword">with</span> torch.inference_mode():
    image = pipe([prompt] * <span class="hljs-number">1</span>, num_inference_steps=<span class="hljs-number">50</span>).images[<span class="hljs-number">0</span>]<!-- HTML_TAG_END --></pre></div>
<h2 class="relative group"><a id="memory-efficient-attention" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#memory-efficient-attention"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a>
	<span>Memory Efficient Attention
	</span></h2>

<p>Recent work on optimizing the bandwitdh in the attention block has generated huge speed ups and gains in GPU memory usage. The most recent being Flash Attention from @tridao: <a href="https://github.com/HazyResearch/flash-attention" rel="nofollow">code</a>, <a href="https://arxiv.org/pdf/2205.14135.pdf" rel="nofollow">paper</a>.</p>
<p>Here are the speedups we obtain on a few Nvidia GPUs when running the inference at 512x512 with a batch size of 1 (one prompt):</p>
<table><thead><tr><th>GPU</th>
<th>Base Attention FP16</th>
<th>Memory Efficient Attention FP16</th></tr></thead>
<tbody><tr><td>NVIDIA Tesla T4</td>
<td>3.5it/s</td>
<td>5.5it/s</td></tr>
<tr><td>NVIDIA 3060 RTX</td>
<td>4.6it/s</td>
<td>7.8it/s</td></tr>
<tr><td>NVIDIA A10G</td>
<td>8.88it/s</td>
<td>15.6it/s</td></tr>
<tr><td>NVIDIA RTX A6000</td>
<td>11.7it/s</td>
<td>21.09it/s</td></tr>
<tr><td>NVIDIA TITAN RTX</td>
<td>12.51it/s</td>
<td>18.22it/s</td></tr>
<tr><td>A100-SXM4-40GB</td>
<td>18.6it/s</td>
<td>29.it/s</td></tr>
<tr><td>A100-SXM-80GB</td>
<td>18.7it/s</td>
<td>29.5it/s</td></tr></tbody></table>
<p>To leverage it just make sure you have:</p>


<div class="course-tip course-tip-orange bg-gradient-to-br dark:bg-gradient-to-r before:border-orange-500 dark:before:border-orange-800 from-orange-50 dark:from-gray-900 to-white dark:to-gray-950 border border-orange-50 text-orange-700 dark:text-gray-400"><p>If you have PyTorch 2.0 installed, you shouldn’t use xFormers!</p></div>
<ul><li>PyTorch &gt; 1.12</li>
<li>Cuda available</li>
<li><a href="xformers">Installed the xformers library</a>.</li></ul>

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

pipe = DiffusionPipeline.from_pretrained(
    <span class="hljs-string">&quot;runwayml/stable-diffusion-v1-5&quot;</span>,
    torch_dtype=torch.float16,
).to(<span class="hljs-string">&quot;cuda&quot;</span>)

pipe.enable_xformers_memory_efficient_attention()

<span class="hljs-keyword">with</span> torch.inference_mode():
    sample = pipe(<span class="hljs-string">&quot;a small cat&quot;</span>)

<span class="hljs-comment"># optional: You can disable it via</span>
<span class="hljs-comment"># pipe.disable_xformers_memory_efficient_attention()</span><!-- HTML_TAG_END --></pre></div>


		<script type="module" data-hydrate="1b04r2h">
		import { start } from "/docs/diffusers/v0.19.2/en/_app/start-hf-doc-builder.js";
		start({
			target: document.querySelector('[data-hydrate="1b04r2h"]').parentNode,
			paths: {"base":"/docs/diffusers/v0.19.2/en","assets":"/docs/diffusers/v0.19.2/en"},
			session: {},
			route: false,
			spa: false,
			trailing_slash: "never",
			hydrate: {
				status: 200,
				error: null,
				nodes: [
					import("/docs/diffusers/v0.19.2/en/_app/pages/__layout.svelte-hf-doc-builder.js"),
						import("/docs/diffusers/v0.19.2/en/_app/pages/optimization/fp16.mdx-hf-doc-builder.js")
				],
				params: {}
			}
		});
	</script>