Buckets:
| <meta charset="utf-8" /><meta name="hf:doc:metadata" content="{"title":"FSDP2","local":"fsdp2","sections":[{"title":"Sharding strategies","local":"sharding-strategies","sections":[],"depth":2},{"title":"Configure FSDP","local":"configure-fsdp","sections":[],"depth":2},{"title":"Next steps","local":"next-steps","sections":[],"depth":2}],"depth":1}"> | |
| <link href="/docs/transformers/pr_41992/en/_app/immutable/assets/0.e3b0c442.css" rel="modulepreload"> | |
| <link rel="modulepreload" href="/docs/transformers/pr_41992/en/_app/immutable/entry/start.6ac19d43.js"> | |
| <link rel="modulepreload" href="/docs/transformers/pr_41992/en/_app/immutable/chunks/scheduler.0dad8db1.js"> | |
| <link rel="modulepreload" href="/docs/transformers/pr_41992/en/_app/immutable/chunks/singletons.9505f7e3.js"> | |
| <link rel="modulepreload" href="/docs/transformers/pr_41992/en/_app/immutable/chunks/index.595571e0.js"> | |
| <link rel="modulepreload" href="/docs/transformers/pr_41992/en/_app/immutable/chunks/paths.a2014728.js"> | |
| <link rel="modulepreload" href="/docs/transformers/pr_41992/en/_app/immutable/entry/app.42da82bd.js"> | |
| <link rel="modulepreload" href="/docs/transformers/pr_41992/en/_app/immutable/chunks/preload-helper.43417560.js"> | |
| <link rel="modulepreload" href="/docs/transformers/pr_41992/en/_app/immutable/chunks/index.855abf52.js"> | |
| <link rel="modulepreload" href="/docs/transformers/pr_41992/en/_app/immutable/nodes/0.dc15b4bc.js"> | |
| <link rel="modulepreload" href="/docs/transformers/pr_41992/en/_app/immutable/chunks/each.e59479a4.js"> | |
| <link rel="modulepreload" href="/docs/transformers/pr_41992/en/_app/immutable/nodes/51.e0024372.js"> | |
| <link rel="modulepreload" href="/docs/transformers/pr_41992/en/_app/immutable/chunks/CopyLLMTxtMenu.9e1f0f49.js"> | |
| <link rel="modulepreload" href="/docs/transformers/pr_41992/en/_app/immutable/chunks/MermaidChart.svelte_svelte_type_style_lang.bf05d4d8.js"> | |
| <link rel="modulepreload" href="/docs/transformers/pr_41992/en/_app/immutable/chunks/IconCopy.ffdeebb6.js"> | |
| <link rel="modulepreload" href="/docs/transformers/pr_41992/en/_app/immutable/chunks/CodeBlock.b28b91fa.js"> | |
| <link rel="modulepreload" href="/docs/transformers/pr_41992/en/_app/immutable/chunks/HfOption.7b5c323c.js"><!-- HEAD_svelte-u9bgzb_START --><meta name="hf:doc:metadata" content="{"title":"FSDP2","local":"fsdp2","sections":[{"title":"Sharding strategies","local":"sharding-strategies","sections":[],"depth":2},{"title":"Configure FSDP","local":"configure-fsdp","sections":[],"depth":2},{"title":"Next steps","local":"next-steps","sections":[],"depth":2}],"depth":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 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> </div> <h1 class="relative group"><a id="fsdp2" 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="#fsdp2"><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>FSDP2</span></h1> <p data-svelte-h="svelte-1s0sn18"><a href="https://docs.pytorch.org/tutorials/intermediate/FSDP_tutorial.html" rel="nofollow">Fully Sharded Data Parallel (FSDP2)</a> shards the model, gradients, and optimizer states across GPUs. Before computation, each GPU gathers a complete set of parameters from all shards, then frees them afterward. Sharding lets you train models larger than a single GPU’s memory, at the cost of more communication than <a href="./ddp">DDP</a>. Use FSDP when your model or optimizer states don’t fit on a single GPU.</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="language-text "><!-- HTML_TAG_START --> ┌─────────────────┐ | |
| │ training data │ | |
| └────────┬────────┘ | |
| ┌──────────────────┼──────────────────┐ | |
| │ shard 0 │ shard 1 │ shard 2 | |
| ▼ ▼ ▼ | |
| ┌─────────────┐ ┌─────────────┐ ┌─────────────┐ | |
| │ param │ │ param │ │ param │ | |
| │ shard 0 │ │ shard 1 │ │ shard 2 │ | |
| │ GPU 0 │ │ GPU 1 │ │ GPU 2 │ | |
| └──────┬──────┘ └──────┬──────┘ └──────┬──────┘ | |
| │ │ │ | |
| └──────── all-gather (params) ────────┘ | |
| │ | |
| full params on each GPU | |
| │ | |
| ┌──────────────────┼──────────────────┐ | |
| ▼ ▼ ▼ | |
| forward forward forward | |
| │ │ │ | |
| └───── reduce-scatter (grads) ────────┘ | |
| │ | |
| ┌──────────────────┼──────────────────┐ | |
| ▼ ▼ ▼ | |
| grad shard 0 grad shard 1 grad shard 2 | |
| optim shard 0 optim shard 1 optim shard 2 | |
| step step step<!-- HTML_TAG_END --></pre></div> <h2 class="relative group"><a id="sharding-strategies" 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="#sharding-strategies"><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>Sharding strategies</span></h2> <p data-svelte-h="svelte-18bxh6b">FSDP2 controls sharding with <code>~TrainingArguments.fsdp_config</code>. Set <code>fsdp=True</code> to enable FSDP, and set <code>reshard_after_forward</code> in the FSDP config to choose the memory and throughput tradeoff.</p> <table data-svelte-h="svelte-munzxc"><thead><tr><th><code>reshard_after_forward</code></th> <th>behavior</th></tr></thead> <tbody><tr><td><code>true</code></td> <td>reshard parameters after the forward pass to save more memory</td></tr> <tr><td><code>false</code></td> <td>keep parameters gathered between forward and backward to avoid the re-all-gather, at the cost of higher peak memory</td></tr></tbody></table> <p data-svelte-h="svelte-ta2q7r"><code>auto_wrap_policy</code> controls how modules are wrapped into FSDP units. It defaults to <code>"TRANSFORMER_BASED_WRAP"</code>, which wraps the model’s transformer layers. Without wrapping (<code>"NO_WRAP"</code>), the entire model is one FSDP unit and you lose the memory benefit of sharding.</p> <h2 class="relative group"><a id="configure-fsdp" 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="#configure-fsdp"><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>Configure FSDP</span></h2> <p data-svelte-h="svelte-9gvwnw">These fields control how FSDP2 wraps, shards, and loads the model. <code>reshard_after_forward</code> and <code>auto_wrap_policy</code> are covered in <a href="#sharding-strategies">Sharding strategies</a>.</p> <ul data-svelte-h="svelte-16mbje"><li><p><code>cpu_offload</code> offloads parameters and gradients to CPU when they aren’t in use to save GPU memory.</p></li> <li><p><code>transformer_layer_cls_to_wrap</code> defines the transformer layer to wrap into an FSDP unit when <code>auto_wrap_policy</code> is <code>"TRANSFORMER_BASED_WRAP"</code>. Each unit manages its own gather and scatter ops. Only the current unit’s parameters are gathered during the forward pass. The previous units’ parameters are released to save memory.</p> <p>Wrapping only the top-level model yields no GPU memory savings. Wrapping every individual <code>Linear</code> layer makes inter-unit communication very expensive. Leave this field empty and FSDP reads the value from the model definition.</p></li> <li><p><code>min_num_params</code> sets the minimum number of parameters per module for size-based wrapping. It is only used when <code>auto_wrap_policy</code> is <code>"SIZE_BASED_WRAP"</code>.</p></li> <li><p><code>state_dict_type</code> controls the checkpoint format. Defaults to <code>"FULL_STATE_DICT"</code> for a single Transformers-compatible checkpoint. Use <code>"SHARDED_STATE_DICT"</code> for one checkpoint file per rank, which is faster for large models. Sharded checkpoints only load back into FSDP, so save a <code>"FULL_STATE_DICT"</code> for the final checkpoint you want to share or load outside FSDP.</p></li> <li><p><code>cpu_ram_efficient_loading</code> loads the checkpoint from disk on rank 0 only. Other GPUs initialize an empty model and receive the weights by broadcast, avoiding multiple processes loading a large model into CPU RAM.</p></li> <li><p><code>activation_checkpointing</code> recomputes activations during the backward pass instead of storing them. Use this instead of <a href="./grad_checkpointing">gradient checkpointing</a> in <a href="/docs/transformers/pr_41992/en/main_classes/trainer#transformers.TrainingArguments">TrainingArguments</a>. Setting both raises an error.</p></li></ul> <p data-svelte-h="svelte-1yrydyo">Configure FSDP training with either an <a href="./accelerate#accelerate-config-file">Accelerate config file</a> or an FSDP config file passed to <code>fsdp_config</code>.</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">Accelerate config file </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">FSDP config file </div></div> <div class="language-select"><p data-svelte-h="svelte-1t08a7v">Run the <a href="https://huggingface.co/docs/accelerate/en/package_reference/cli#accelerate-config" rel="nofollow">accelerate config</a> command and answer questions about your hardware and training setup. This creates a <code>default_config.yaml</code> file in your cache.</p> <p data-svelte-h="svelte-uoiphc">Run <a href="https://huggingface.co/docs/accelerate/en/package_reference/cli#accelerate-launch" rel="nofollow">accelerate launch</a> with a <a href="/docs/transformers/pr_41992/en/main_classes/trainer#transformers.Trainer">Trainer</a>-based script. The <code>fsdp_config</code> is unnecessary because the Accelerate config file covers the same settings.</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="language-cli "><!-- HTML_TAG_START -->accelerate <span class="hljs-built_in">launch</span> train.py<!-- HTML_TAG_END --></pre></div> </div> <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 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>Next steps</span></h2> <ul data-svelte-h="svelte-1pb2j8y"><li>See <a href="./ddp">DDP</a> for data-parallel training when your model fits on one GPU.</li> <li>See <a href="./deepspeed">DeepSpeed</a> for ZeRO optimization and NVMe offloading.</li> <li>For FSDP on TPUs with PyTorch/XLA, set <code>xla</code>, <code>xla_fsdp_settings</code>, and <code>xla_fsdp_grad_ckpt</code> in <code>~TrainingArguments.fsdp_config</code>.</li> <li>Read the <a href="https://nanotron-ultrascale-playbook.static.hf.space/index.html#zero-3:_adding_parameter_partitioning_(fsdp)" rel="nofollow">FSDP chapter</a> from The Ultra-Scale Playbook for more information about how FSDP works.</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/fsdp.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_1cvdjju = { | |
| assets: "/docs/transformers/pr_41992/en", | |
| base: "/docs/transformers/pr_41992/en", | |
| env: {} | |
| }; | |
| const element = document.currentScript.parentElement; | |
| const data = [null,null]; | |
| Promise.all([ | |
| import("/docs/transformers/pr_41992/en/_app/immutable/entry/start.6ac19d43.js"), | |
| import("/docs/transformers/pr_41992/en/_app/immutable/entry/app.42da82bd.js") | |
| ]).then(([kit, app]) => { | |
| kit.start(app, element, { | |
| node_ids: [0, 51], | |
| data, | |
| form: null, | |
| error: null | |
| }); | |
| }); | |
| } | |
| </script> | |
Xet Storage Details
- Size:
- 21.1 kB
- Xet hash:
- e201c50c5759f6ebbb74342694e7a29dd09f9f6e2403ffcecafea3c9aea852d1
·
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.