Buckets:
| <meta charset="utf-8" /><meta name="hf:doc:metadata" content="{"title":"GPU memory usage","local":"gpu-memory-usage","sections":[{"title":"Tensors","local":"tensors","sections":[],"depth":2},{"title":"Ops","local":"ops","sections":[],"depth":2},{"title":"Next steps","local":"next-steps","sections":[],"depth":2}],"depth":1}"/> | |
| <link href="/docs/transformers/main/en/_app/immutable/entry/start.BukNISTP.js" rel="modulepreload"> | |
| <link href="/docs/transformers/main/en/_app/immutable/chunks/BPzh_Jza.js" rel="modulepreload"> | |
| <link href="/docs/transformers/main/en/_app/immutable/chunks/BhmhGQv5.js" rel="modulepreload"> | |
| <link href="/docs/transformers/main/en/_app/immutable/entry/app.5sK1nhM5.js" rel="modulepreload"> | |
| <link href="/docs/transformers/main/en/_app/immutable/chunks/qMuWdE-Y.js" rel="modulepreload"> | |
| <link href="/docs/transformers/main/en/_app/immutable/chunks/CfMEBQAN.js" rel="modulepreload"> | |
| <link href="/docs/transformers/main/en/_app/immutable/chunks/B4BfNKKv.js" rel="modulepreload"> | |
| <link href="/docs/transformers/main/en/_app/immutable/chunks/DsnmJJEf.js" rel="modulepreload"> | |
| <link href="/docs/transformers/main/en/_app/immutable/nodes/0.C2Q_m8zG.js" rel="modulepreload"> | |
| <link href="/docs/transformers/main/en/_app/immutable/chunks/B76T2TVW.js" rel="modulepreload"> | |
| <link href="/docs/transformers/main/en/_app/immutable/nodes/626.2LwK1xsU.js" rel="modulepreload"> | |
| <link href="/docs/transformers/main/en/_app/immutable/chunks/vlAHDplH.js" rel="modulepreload"> | |
| <!--71qes8--><meta name="hf:doc:metadata" content="{"title":"GPU memory usage","local":"gpu-memory-usage","sections":[{"title":"Tensors","local":"tensors","sections":[],"depth":2},{"title":"Ops","local":"ops","sections":[],"depth":2},{"title":"Next steps","local":"next-steps","sections":[],"depth":2}],"depth":1}"/><!----> | |
| <link href="/docs/transformers/main/en/_app/immutable/assets/0.tn0RQdqM.css" rel="modulepreload"> <!--[--><!--[0--><!--[--><!--[0--><!--[--><p></p> <div class="items-center shrink-0 min-w-[100px] max-sm:min-w-[50px] justify-end ml-auto flex" style="float: 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 h-7 max-sm:h-7 px-2 max-sm:px-1.5 text-sm font-medium text-gray-800 border border-r-0 rounded-l-md max-sm:rounded-l-sm border-gray-200 bg-white hover:shadow-inner dark:border-gray-850 dark:bg-gray-950 dark:text-gray-200 dark:hover:bg-gray-800" aria-live="polite"><span class="inline-flex items-center justify-center rounded-md p-0.5 max-sm:p-0 hover:text-gray-800 dark:hover:text-gray-200"><svg class="sm:size-3.5 size-3" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg><!----></span> <span>Copy page</span></button> <button class="inline-flex items-center justify-center w-6 max-sm:w-5 h-7 max-sm:h-7 disabled:pointer-events-none text-sm text-gray-500 hover:text-gray-700 dark:hover:text-white rounded-r-md max-sm:rounded-r-sm border border-l transition border-gray-200 bg-white hover:shadow-inner dark:border-gray-850 dark:bg-gray-950 dark:text-gray-200 dark:hover:bg-gray-800" aria-haspopup="menu" aria-expanded="false" aria-label="Open copy menu"><svg class="transition-transform text-gray-400 overflow-visible sm:size-3.5 size-3 rotate-0" width="1em" height="1em" viewBox="0 0 12 7" fill="none" xmlns="http://www.w3.org/2000/svg"><path d="M1 1L6 6L11 1" stroke="currentColor"></path></svg><!----></button></div> <!--[-1--><!--]--></div><!----> <!--[0--><h1 class="relative group"><a id="gpu-memory-usage" 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="#gpu-memory-usage"><span><svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!----></span></a> <span>GPU memory usage</span></h1><!--]--><!----> <p>Training a 4B parameter model in mixed precision on a batch size of 16 requires roughly 85GB of GPU memory. Understanding what occupies GPU memory and the compute operations that occur during training help identify where you can reduce your memory usage.</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg><!----> <div class=" absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0 "><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent;"></div> Copied</div><!----></button><!----></div> <pre class="language-text "><!---->┌─────────────────────────── TENSORS ──────────────────────────────────┐ | |
| │ │ | |
| │ MODEL WEIGHTS ████████████████████████ 6 bytes/param │ | |
| │ (mixed precision) ├── fp16 copy (2B) — forward/backward pass │ | |
| │ └── fp32 copy (4B) — stable weight updates │ | |
| │ │ | |
| │ OPTIMIZER STATES ████████████████████████████████ 8 bytes/param │ | |
| │ (Adam) ├── fp32 momentum (4B) │ | |
| │ └── fp32 variance (4B) │ | |
| │ ↳ quantized Adam (bitsandbytes) → 2 bytes/param │ | |
| │ │ | |
| │ GRADIENTS ████████████████ 4 bytes/param (fp32) │ | |
| │ └── computed in backward pass, per parameter │ | |
| │ │ | |
| │ ACTIVATIONS ████ varies — batch × seq_len × depth × hidden │ | |
| │ (forward cache) └── cached for backward; can OOM even if model │ | |
| │ fits (long seqs / large batches) │ | |
| │ bf16/fp16 if using mixed precision │ | |
| │ │ | |
| │ TEMPORARY TENSORS ▓ short-lived (softmax, matmul scratch) │ | |
| │ └── peak spikes can cause OOM │ | |
| │ │ | |
| │ OTHER OVERHEAD ▒ beam search caches, large embedding tables │ | |
| └──────────────────────────────────────────────────────────────────────┘<!----></pre></div><!----> <p>GPU memory holds two categories of things: stored tensors and the ops that process them.</p> <!--[1--><h2 class="relative group"><a id="tensors" 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="#tensors"><span><svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!----></span></a> <span>Tensors</span></h2><!--]--><!----> <p>Training requires and produces many tensor types that need to be stored.</p> <ul><li><p>Model weights are stored on the GPU. In mixed precision training, two copies of the weights are required - one in fp16 for the forward/backward pass and one in fp32 as the “main copy” for stable weight updates. This equates to 6 bytes per parameter.</p></li> <li><p>Optimizer states such as Adam store two extra tensors per parameter, the momentum and variance, which are both in fp32. That’s an additional 8 bytes per parameter.</p> <p>You could use a different optimizer like a quantized version of Adam from <a href="https://huggingface.co/docs/bitsandbytes/main/en/index" rel="nofollow">bitsandbytes</a> to compress it to 2 bytes per parameter.</p></li> <li><p>Gradient tensors are computed for each parameter in the backward pass. This is kept in fp32 so that’s 4 bytes per parameter.</p></li> <li><p>Forward activations are computed in the forward pass and cached for the backward pass to compute the gradients. These activations vary in size with batch size, sequence length, model depth, and hidden size. This is why batch size or sequence length can exhaust GPU memory even if the model itself fits.</p></li> <li><p>Temporary tensors are created by ops like softmax and matrix multiplications and released after each op. If the peak of a single op is very intensive, it can create a temporary spike that causes you to run out of memory.</p></li> <li><p>Some features add their own overhead. For example, beam search maintains multiple outputs, and embedding tables for large vocabularies can be very large.</p></li></ul> <!--[1--><h2 class="relative group"><a id="ops" 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="#ops"><span><svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!----></span></a> <span>Ops</span></h2><!--]--><!----> <p>There are three main types of training ops.</p> <ul><li><p>Matrix multiplications (matmuls) are the main op type, covering linear layers, QKV projections, attention output projections, and FFN layers. The main memory hogs are the attention score matrices which grow with the square of the sequence length. This is why long sequences are so expensive and different <a href="./attention_interface">attention backends</a> exist to avoid materializing the full matrix in memory.</p></li> <li><p>Reduction ops, like softmax and layer norm, read a full tensor, compute a statistic across a dimension, and then read the tensor again to apply it. This requires accessing memory multiple times per operation.</p></li> <li><p>Element-wise ops, like activations and dropout, apply a function to each element independently and their memory usage is proportional to tensor size.</p></li></ul> <!--[1--><h2 class="relative group"><a id="next-steps" 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="#next-steps"><span><svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!----></span></a> <span>Next steps</span></h2><!--]--><!----> <ul><li>See the <a href="./grad_accumulation">Gradient accumulation</a> guide to learn how to simulate training on a larger effective batch size without running out of GPU memory.</li> <li>See the <a href="./grad_checkpointing">Gradient checkpointing</a> guide to learn how to reduce memory usage by only storing some intermediate activations.</li> <li>See the <a href="./mixed_precision_training">Mixed precision training</a> guide to learn how to use lower precision data types to reduce memory and speed up training.</li> <li>See the <a href="./kernels">Kernels</a> guide to learn how to speed up training with custom fused kernels.</li></ul> <a class="!text-gray-400 !no-underline text-sm flex items-center not-prose mt-4" href="https://github.com/huggingface/transformers/blob/main/docs/source/en/model_memory_anatomy.md" target="_blank"><svg class="mr-1" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M31,16l-7,7l-1.41-1.41L28.17,16l-5.58-5.59L24,9l7,7z"></path><path d="M1,16l7-7l1.41,1.41L3.83,16l5.58,5.59L8,23l-7-7z"></path><path d="M12.419,25.484L17.639,6.552l1.932,0.518L14.351,26.002z"></path></svg><!----> <span><span class="underline">Update</span> on GitHub</span></a><!----> <p></p><!--]--><!----><!--]--><!--]--><!--]--> <!--[-1--><!--]--><!--]--> | |
| <script> | |
| { | |
| __sveltekit_hvxp82 = { | |
| base: "/docs/transformers/main/en", | |
| assets: "/docs/transformers/main/en" | |
| }; | |
| const element = document.currentScript.parentElement; | |
| Promise.all([ | |
| import("/docs/transformers/main/en/_app/immutable/entry/start.BukNISTP.js"), | |
| import("/docs/transformers/main/en/_app/immutable/entry/app.5sK1nhM5.js") | |
| ]).then(([kit, app]) => { | |
| kit.start(app, element, { | |
| node_ids: [0, 626], | |
| data: [null,null], | |
| form: null, | |
| error: null | |
| }); | |
| }); | |
| } | |
| </script> | |
Xet Storage Details
- Size:
- 17 kB
- Xet hash:
- 91a3fbfc67d313f906ab472c2a2cf7b280e1a8caf0f0a56d30deccbae5e8f0d7
·
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.