transformers/pr_40546/en/deepspeed_alst.html

<meta charset="utf-8" /><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;Ulysses sequence parallelism&quot;,&quot;local&quot;:&quot;ulysses-sequence-parallelism&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Configure&quot;,&quot;local&quot;:&quot;configure&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Combining with data parallelism&quot;,&quot;local&quot;:&quot;combining-with-data-parallelism&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Next steps&quot;,&quot;local&quot;:&quot;next-steps&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2}],&quot;depth&quot;:1}"/>
		<link href="/docs/transformers/main/en/_app/immutable/entry/start.DyW3J4IQ.js" rel="modulepreload">
		<link href="/docs/transformers/main/en/_app/immutable/chunks/DYRUnuSC.js" rel="modulepreload">
		<link href="/docs/transformers/main/en/_app/immutable/chunks/v-ay6ctG.js" rel="modulepreload">
		<link href="/docs/transformers/main/en/_app/immutable/entry/app.D9CUDH73.js" rel="modulepreload">
		<link href="/docs/transformers/main/en/_app/immutable/chunks/BIdhRtf1.js" rel="modulepreload">
		<link href="/docs/transformers/main/en/_app/immutable/chunks/Cmr6hduW.js" rel="modulepreload">
		<link href="/docs/transformers/main/en/_app/immutable/chunks/BRRD8tqb.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.CLEsIFU8.js" rel="modulepreload">
		<link href="/docs/transformers/main/en/_app/immutable/chunks/QxajwaoF.js" rel="modulepreload">
		<link href="/docs/transformers/main/en/_app/immutable/nodes/6.CHrtO2tr.js" rel="modulepreload">
		<link href="/docs/transformers/main/en/_app/immutable/chunks/CxENESZC.js" rel="modulepreload">
		<link href="/docs/transformers/main/en/_app/immutable/chunks/BJV2BSXD.js" rel="modulepreload">
		<!--1ubkjvt--><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;Ulysses sequence parallelism&quot;,&quot;local&quot;:&quot;ulysses-sequence-parallelism&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Configure&quot;,&quot;local&quot;:&quot;configure&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Combining with data parallelism&quot;,&quot;local&quot;:&quot;combining-with-data-parallelism&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Next steps&quot;,&quot;local&quot;:&quot;next-steps&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2}],&quot;depth&quot;: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="ulysses-sequence-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="#ulysses-sequence-parallelism"><span><svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!----></span></a> <span>Ulysses sequence parallelism</span></h1><!--]--><!----> <p>Ulysses sequence parallelism (SP) trains on very long sequences by splitting them across multiple GPUs. To compute attention correctly, an all-to-all collective swaps the sharding dimension from sequence to attention heads. Each GPU then has the full sequence and computes attention locally over a subset of heads. A second all-to-all returns to the sequence-sharded layout so the rest of the forward pass continues locally on each chunk.</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 "><!---->                        GPU 0                       GPU 1
                   ┌───────────────┐          ┌───────────────┐
  forward          │ tokens 0..N/2 │          │ tokens N/2..N │  ← each GPU holds half the sequence
  (seq-sharded)    │  all H heads  │          │  all H heads  │
                   └───────┬───────┘          └───────┬───────┘
                           └───────── all-to-all ──────┘
                   ┌───────────────┐          ┌───────────────┐
  attention        │ all N tokens  │          │ all N tokens  │  ← now each GPU has the full sequence
  (head-sharded)   │ heads 0..H/2  │          │ heads H/2..H  │  ← but only half the heads
                   └───────┬───────┘          └───────┬───────┘
                           └───────── all-to-all ──────┘
                   ┌───────────────┐          ┌───────────────┐
  forward          │ tokens 0..N/2 │          │ tokens N/2..N │  ← back to seq-sharded
  (seq-sharded)    │  all H heads  │          │  all H heads  │
                   └───────────────┘          └───────────────┘<!----></pre></div><!----> <blockquote class="note"><p>This guide covers the Ulysses sequence parallelism component of <a href="https://www.deepspeed.ai/tutorials/ulysses-alst-sequence-parallelism/" rel="nofollow">ALST</a> (Arctic Long Sequence Training). The full ALST system also includes TiledMLP and activation checkpoint offloading, which aren’t available in Transformers. See the <a href="https://www.deepspeed.ai/tutorials/ulysses-alst-sequence-parallelism/" rel="nofollow">DeepSpeed ALST tutorial</a> for the complete system.</p></blockquote> <!--[1--><h2 class="relative group"><a id="configure" 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"><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>Configure</span></h2><!--]--><!----> <p>Sequence parallelism requires Accelerate v1.12.0 and at least 2 GPUs. Configure sequence parallelism in Accelerate’s <code>ParallelismConfig</code> and pass it to <a href="/docs/transformers/pr_40546/en/main_classes/trainer#transformers.TrainingArguments.parallelism_config">parallelism_config</a> or an <a href="./accelerate#accelerate-config-file">Accelerate config file</a>.</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">parallelism_config</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">Accelerate config file</div><!--]--></div> <div class="language-select"><!--[0--><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-py "><!----><span class="hljs-keyword">from</span> accelerate.utils <span class="hljs-keyword">import</span> ParallelismConfig, DeepSpeedSequenceParallelConfig

parallelism_config = ParallelismConfig(
    sp_backend=<span class="hljs-string">&quot;deepspeed&quot;</span>,
    sp_size=<span class="hljs-number">4</span>,
    dp_replicate_size=<span class="hljs-number">1</span>,
    sp_handler=DeepSpeedSequenceParallelConfig(
        sp_seq_length_is_variable=<span class="hljs-literal">True</span>,
        sp_attn_implementation=<span class="hljs-string">&quot;flash_attention_2&quot;</span>,
    ),
)

training_args = TrainingArguments(
    ...,
    deepspeed=<span class="hljs-string">&quot;path/to/deepspeed_config.json&quot;</span>,
    parallelism_config=parallelism_config,
)<!----></pre></div><!----> <p>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_40546/en/main_classes/trainer#transformers.Trainer">Trainer</a>-based script.</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-shell "><!---->accelerate launch --num_processes 4 train.py \
--output_dir output_dir \
--per_device_train_batch_size 1 \
--gradient_accumulation_steps 1<!----></pre></div><!----><!----><!--]--><!----> <!--[-1--><!--]--><!----><!----></div><!----> <p>The following fields are important for configuring sequence parallelism.</p> <blockquote class="tip"><p>The <a href="/docs/transformers/pr_40546/en/main_classes/trainer#transformers.Trainer">Trainer</a> automatically handles DataLoader sharding, <code>position_ids</code> generation, label shifting, and loss aggregation across SP ranks. If you’re writing a custom training loop, see the Accelerate <a href="https://huggingface.co/docs/accelerate/concept_guides/sequence_parallelism" rel="nofollow">Sequence Parallelism</a> guide instead.</p></blockquote> <ul><li><p><code>sp_backend</code> must be set to <code>"deepspeed"</code> to use Ulysses sequence parallelism.</p></li> <li><p><code>sp_size</code> is the number of GPUs that process a single sequence in parallel. Each SP rank receives a unique data stream from the DataLoader, unlike tensor parallelism where all ranks receive identical data. The effective <code>dp_world_size = world_size / sp_size</code>, so with 4 GPUs and <code>sp_size=4</code>, <code>dp_world_size=1</code> for batch size calculations. Sequences must also be padded to a multiple of <code>sp_size</code>. Set <code>pad_to_multiple_of</code> in your data collator accordingly.</p> <blockquote class="warning"><p>The number of attention heads must be divisible by <code>sp_size</code>. A model with 32 heads supports <code>sp_size</code> of 1, 2, 4, 8, 16, or 32.</p></blockquote> <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-py "><!----><span class="hljs-keyword">from</span> transformers <span class="hljs-keyword">import</span> DataCollatorForLanguageModeling

data_collator = DataCollatorForLanguageModeling(
    tokenizer=tokenizer,
    mlm=<span class="hljs-literal">False</span>,
    pad_to_multiple_of=sp_size,
)<!----></pre></div><!----></li> <li><p><code>sp_seq_length_is_variable</code> controls variable sequence length handling. Set it to <code>True</code> (recommended) for varying lengths between batches. Set it to <code>False</code> when all sequences pad to a fixed length specified by <code>sp_seq_length</code>.</p></li> <li><p><code>sp_attn_implementation</code> sets the attention backend. Supported values are <code>"sdpa"</code>, <code>"flash_attention_2"</code>, or <code>"flash_attention_3"</code>. FlashAttention is recommended, especially when packing multiple samples in a batch. SDPA can attend incorrectly across sample boundaries when samples are packed. Eager attention isn’t supported because its 4D <code>attention_mask</code> is discarded for memory and scaling reasons.</p></li></ul> <!--[1--><h2 class="relative group"><a id="combining-with-data-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="#combining-with-data-parallelism"><span><svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!----></span></a> <span>Combining with data parallelism</span></h2><!--]--><!----> <p>Sequence parallelism and data parallelism use the same GPUs, and SP doesn’t require additional hardware. To run both, set <code>dp_replicate_size</code> or <code>dp_shard_size</code> so that <code>dp_replicate_size × dp_shard_size × sp_size</code> equals your total GPU count.</p> <p>For example, with 8 GPUs and <code>sp_size=4</code>, set <code>dp_replicate_size=2</code> (2 × 1 × 4 = 8).</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-py "><!---->parallelism_config = ParallelismConfig(
    sp_backend=<span class="hljs-string">&quot;deepspeed&quot;</span>,
    sp_size=<span class="hljs-number">4</span>,
    dp_replicate_size=<span class="hljs-number">2</span>,
    sp_handler=DeepSpeedSequenceParallelConfig(
        sp_seq_length_is_variable=<span class="hljs-literal">True</span>,
        sp_attn_implementation=<span class="hljs-string">&quot;flash_attention_2&quot;</span>,
    ),
)<!----></pre></div><!----> <!--[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>The Accelerate <a href="https://huggingface.co/docs/accelerate/concept_guides/sequence_parallelism" rel="nofollow">Sequence Parallelism</a> guide covers the Ulysses implementation in more depth and shows how to write a custom training loop.</li> <li>The <a href="https://www.deepspeed.ai/tutorials/ulysses-alst-sequence-parallelism/" rel="nofollow">DeepSpeed ALST tutorial</a> covers the full ALST system, including TiledMLP and activation checkpoint offloading.</li> <li>The <a href="./perf_train_gpu_many">parallelism methods</a> guide shows how to combine sequence parallelism with other strategies like ZeRO.</li> <li>The <a href="https://huggingface.co/blog/ulysses-sp" rel="nofollow">Ulysses Sequence Parallelism: Training with Million-Token Contexts</a> blog post explains how Ulysses works and how it’s integrated in Accelerate, Trainer, and SFTTrainer.</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/deepspeed_alst.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_uog7ni = {
						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.DyW3J4IQ.js"),
						import("/docs/transformers/main/en/_app/immutable/entry/app.D9CUDH73.js")
					]).then(([kit, app]) => {
						kit.start(app, element, {
							node_ids: [0, 6],
							data: [null,null],
							form: null,
							error: null
						});
					});
				}
			</script>