Hugging Face's logo

Buckets:
hf-doc-build
/
doc-dev

Files
xet
hf-doc-build/doc-dev / trl /pr_3582 /en /reducing_memory_usage.html
rtrm's picture
rtrm
download
raw
35.6 kB
<meta charset="utf-8" /><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;Reducing Memory Usage&quot;,&quot;local&quot;:&quot;reducing-memory-usage&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Truncation&quot;,&quot;local&quot;:&quot;truncation&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Packing&quot;,&quot;local&quot;:&quot;packing&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Liger for reducing peak memory usage&quot;,&quot;local&quot;:&quot;liger-for-reducing-peak-memory-usage&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Padding-free&quot;,&quot;local&quot;:&quot;padding-free&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Activation offloading&quot;,&quot;local&quot;:&quot;activation-offloading&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Disabling model gathering for generation in online methods&quot;,&quot;local&quot;:&quot;disabling-model-gathering-for-generation-in-online-methods&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2}],&quot;depth&quot;:1}">
 <link href="/docs/trl/pr_3582/en/_app/immutable/assets/0.e3b0c442.css" rel="modulepreload">
 <link rel="modulepreload" href="/docs/trl/pr_3582/en/_app/immutable/entry/start.0f0f318c.js">
 <link rel="modulepreload" href="/docs/trl/pr_3582/en/_app/immutable/chunks/scheduler.d627b047.js">
 <link rel="modulepreload" href="/docs/trl/pr_3582/en/_app/immutable/chunks/singletons.affb0d47.js">
 <link rel="modulepreload" href="/docs/trl/pr_3582/en/_app/immutable/chunks/index.a57a1c33.js">
 <link rel="modulepreload" href="/docs/trl/pr_3582/en/_app/immutable/chunks/paths.15dc14db.js">
 <link rel="modulepreload" href="/docs/trl/pr_3582/en/_app/immutable/entry/app.b27a462f.js">
 <link rel="modulepreload" href="/docs/trl/pr_3582/en/_app/immutable/chunks/index.73c51727.js">
 <link rel="modulepreload" href="/docs/trl/pr_3582/en/_app/immutable/nodes/0.8cd8e450.js">
 <link rel="modulepreload" href="/docs/trl/pr_3582/en/_app/immutable/chunks/each.e59479a4.js">
 <link rel="modulepreload" href="/docs/trl/pr_3582/en/_app/immutable/nodes/39.18dab992.js">
 <link rel="modulepreload" href="/docs/trl/pr_3582/en/_app/immutable/chunks/Tip.a82942ec.js">
 <link rel="modulepreload" href="/docs/trl/pr_3582/en/_app/immutable/chunks/CodeBlock.5f78c87f.js">
 <link rel="modulepreload" href="/docs/trl/pr_3582/en/_app/immutable/chunks/getInferenceSnippets.256dfbf1.js">
 <link rel="modulepreload" href="/docs/trl/pr_3582/en/_app/immutable/chunks/HfOption.d6c7dc54.js"><!-- HEAD_svelte-u9bgzb_START --><meta name="hf:doc:metadata" content="{&quot;title&quot;:&quot;Reducing Memory Usage&quot;,&quot;local&quot;:&quot;reducing-memory-usage&quot;,&quot;sections&quot;:[{&quot;title&quot;:&quot;Truncation&quot;,&quot;local&quot;:&quot;truncation&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Packing&quot;,&quot;local&quot;:&quot;packing&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Liger for reducing peak memory usage&quot;,&quot;local&quot;:&quot;liger-for-reducing-peak-memory-usage&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Padding-free&quot;,&quot;local&quot;:&quot;padding-free&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Activation offloading&quot;,&quot;local&quot;:&quot;activation-offloading&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2},{&quot;title&quot;:&quot;Disabling model gathering for generation in online methods&quot;,&quot;local&quot;:&quot;disabling-model-gathering-for-generation-in-online-methods&quot;,&quot;sections&quot;:[],&quot;depth&quot;:2}],&quot;depth&quot;:1}"><!-- HEAD_svelte-u9bgzb_END --> <p></p> <h1 class="relative group"><a id="reducing-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="#reducing-memory-usage"><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>Reducing Memory Usage</span></h1> <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 data-svelte-h="svelte-txx63b">Section under construction. Feel free to contribute!</p></div> <h2 class="relative group"><a id="truncation" 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="#truncation"><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>Truncation</span></h2> <p data-svelte-h="svelte-1lhmd25">Sequence lengths in the dataset can vary widely. When data is batched, sequences are padded to match the longest one in the batch, which can cause high memory usage, even if most sequences are relatively short.</p> <div class="flex justify-center" data-svelte-h="svelte-4pi4g0"><img src="https://huggingface.co/datasets/trl-lib/documentation-images/resolve/main/why_you_should_truncate.png" alt="Truncation prompt completion" width="600"></div> <p data-svelte-h="svelte-15aiza5">To reduce memory usage, it’s important to truncate sequences to a reasonable length. While TRL trainers truncate sequences by default, you may want to adjust the default truncation length to better align with your specific use case.</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">DPO </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">SFT </div></div> <div class="language-select"><p data-svelte-h="svelte-cly35y">DPO truncation is applied first to the prompt and to the completion via the <code>max_prompt_length</code> and <code>max_completion_length</code> parameters. The <code>max_length</code> parameter is then used to truncate the resulting sequence.</p> <div class="flex justify-center" data-svelte-h="svelte-1f9vb58"><img src="https://huggingface.co/datasets/trl-lib/documentation-images/resolve/main/truncation_prompt_completion.png" alt="Truncation prompt completion" width="600"></div> <p data-svelte-h="svelte-1r0zv8c">To set the truncation parameters, use the following code snippet:</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-keyword">from</span> trl <span class="hljs-keyword">import</span> DPOConfig
training_args = DPOConfig(..., max_prompt_length=..., max_length=...)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-10jd0vs">You can also use the <code>max_completion_length</code> parameter to truncate the completion, though this is less common since the goal is typically to preserve the completion’s full length whenever possible.</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-keyword">from</span> trl <span class="hljs-keyword">import</span> DPOConfig
training_args = DPOConfig(..., max_completion_length=...)<!-- HTML_TAG_END --></pre></div> </div> <h2 class="relative group"><a id="packing" 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="#packing"><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>Packing</span></h2> <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"><p data-svelte-h="svelte-o7a208">This technique applies only to SFT.</p></div> <p data-svelte-h="svelte-6rm19q"><a href="#truncation">Truncation</a> has several drawbacks:</p> <ol data-svelte-h="svelte-17zhqal"><li><strong>Loss of information</strong>: Key data at the end of a sequence may be discarded.</li> <li><strong>Choosing truncation length</strong>: Too short loses data; too long undermines efficiency.</li></ol> <p data-svelte-h="svelte-1016bgl">Packing, introduced in <a href="https://huggingface.co/papers/1910.10683" rel="nofollow">Raffel et al., 2020</a>, addresses these issues by grouping sequences instead of truncating. It concatenates and splits dataset sequences into the desired lengths.</p> <div class="flex justify-center" data-svelte-h="svelte-nl0cdp"><img src="https://huggingface.co/datasets/trl-lib/documentation-images/resolve/main/packing_2.png" alt="Packing" width="600"></div> <p data-svelte-h="svelte-43vy0b">Packing reduces padding by merging several sequences in one row when possible. We use an advanced method to be near-optimal in the way we pack the dataset. To enable packing, use <code>packing=True</code> and in the <a href="/docs/trl/pr_3582/en/sft_trainer#trl.SFTConfig">SFTConfig</a>.</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"><p data-svelte-h="svelte-1m2tmjt">In TRL 0.18 and earlier, packing used a more aggressive method that reduced padding to almost nothing, but had the downside of breaking sequence continuity for a large fraction of the dataset. To revert to this strategy, use <code>packing_strategy=&quot;wrapped&quot;</code> in <code>SFTConfig</code>.</p></div> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-keyword">from</span> trl <span class="hljs-keyword">import</span> SFTConfig
training_args = SFTConfig(..., packing=<span class="hljs-literal">True</span>, max_length=<span class="hljs-number">512</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 data-svelte-h="svelte-6zyodn">Packing may cause batch contamination, where adjacent sequences influence one another. This can be problematic for some applications. For more details, see <a href="https://github.com/huggingface/trl/issues/1230" rel="nofollow">#1230</a>.</p></div> <h2 class="relative group"><a id="liger-for-reducing-peak-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="#liger-for-reducing-peak-memory-usage"><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>Liger for reducing peak memory usage</span></h2> <blockquote data-svelte-h="svelte-skwjb4"><p><a href="https://github.com/linkedin/Liger-Kernel" rel="nofollow">Liger Kernel</a> is a collection of Triton kernels designed specifically for LLM training. It can effectively increase multi-GPU training throughput by 20% and reduces memory usage by 60%.</p></blockquote> <p data-svelte-h="svelte-tl3zka">For more information, see <a href="liger_kernel_integration">Liger Kernel Integration</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">DPO </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">GRPO </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">KTO </div></div> <div class="language-select"><p data-svelte-h="svelte-tcc5th">To use Liger for reducing peak memory usage, use the following code snippet:</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-keyword">from</span> trl <span class="hljs-keyword">import</span> DPOConfig
training_args = DPOConfig(..., use_liger_loss=<span class="hljs-literal">True</span>)<!-- HTML_TAG_END --></pre></div> </div> <h2 class="relative group"><a id="padding-free" 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="#padding-free"><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>Padding-free</span></h2> <p data-svelte-h="svelte-igqq2g">Padding-free batching is an alternative approach for reducing memory usage. In this method, a batch is first sampled and then flattened into a single sequence, avoiding padding. Unlike packing, which can result in incomplete sequences by combining parts of different samples, padding-free batching ensures that all sequences remain complete and intact.</p> <div class="flex justify-center" data-svelte-h="svelte-k4w76c"><img src="https://huggingface.co/datasets/trl-lib/documentation-images/resolve/main/padding-free.png" alt="Padding-free batching" width="600"></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 data-svelte-h="svelte-1953gpl">It’s highly recommended to use padding-free batching with <strong>Flash Attention 2</strong>. Otherwise, you may encounter batch contamination issues.</p></div> <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">DPO </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">SFT </div></div> <div class="language-select"><div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-keyword">from</span> trl <span class="hljs-keyword">import</span> DPOConfig
training_args = DPOConfig(..., padding_free=<span class="hljs-literal">True</span>, model_init_kwargs={<span class="hljs-string">&quot;attn_implementation&quot;</span>: <span class="hljs-string">&quot;flash_attention_2&quot;</span>})<!-- HTML_TAG_END --></pre></div> </div> <h2 class="relative group"><a id="activation-offloading" 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="#activation-offloading"><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>Activation offloading</span></h2> <p data-svelte-h="svelte-1w4qqip">Activation offloading is a memory efficiency technique that reduces GPU VRAM usage by temporarily moving activation tensors to CPU RAM during the forward pass and bringing them back only when needed for the backward pass. This significantly reduces peak memory usage at the cost of slightly increased training time.</p> <p data-svelte-h="svelte-yymgek">To enable activation offloading in your SFT training configuration:</p>
&lt;hfoptions&gt;
&lt;hfoption id=&quot;SFT&quot;&gt;
<div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-keyword">from</span> trl <span class="hljs-keyword">import</span> SFTConfig
training_args = SFTConfig(..., activation_offloading=<span class="hljs-literal">True</span>)<!-- HTML_TAG_END --></pre></div>
&lt;/hfoption&gt;
&lt;/hfoptions&gt;
 <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 data-svelte-h="svelte-cna7dj">When using activation offloading with models that use Liger kernels, you must disable Liger cross entropy due to compatibility issues. The issue occurs specifically with <code>use_liger_kernel=True</code> because Liger cross entropy performs in-place operations which conflict with activation offloading. The default setting (<code>use_liger_kernel=False</code>) works:</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-comment"># When using activation offloading with a model that uses Liger kernels:</span>
<span class="hljs-keyword">from</span> trl <span class="hljs-keyword">import</span> SFTConfig
training_args = SFTConfig(
 activation_offloading=<span class="hljs-literal">True</span>,
 use_liger_kernel=<span class="hljs-literal">False</span>, <span class="hljs-comment"># Disable Liger cross entropy</span>
 <span class="hljs-comment"># Other parameters...</span>
)<!-- HTML_TAG_END --></pre></div></div> <p data-svelte-h="svelte-12o8vk8">Under the hood, activation offloading implements PyTorch’s <a href="https://pytorch.org/tutorials/intermediate/autograd_saved_tensors_hooks_tutorial.html#hooks-for-autograd-saved-tensors" rel="nofollow"><code>saved_tensors_hooks</code></a> to intercept activations during the forward pass. It intelligently manages which tensors to offload based on size and context, avoiding offloading output tensors which would be inefficient. For performance optimization, it can optionally use CUDA streams to overlap computation with CPU-GPU transfers.</p> <h2 class="relative group"><a id="disabling-model-gathering-for-generation-in-online-methods" 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="#disabling-model-gathering-for-generation-in-online-methods"><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>Disabling model gathering for generation in online methods</span></h2> <p data-svelte-h="svelte-1dxf8zd">When using DeepSpeed ZeRO-3, model weights are sharded across multiple GPUs. Online methods involve generating completions from the model as part of the training process. During this step, the model weights are temporarily gathered on a single GPU for generation. For very large models, this gathering can lead to out-of-memory (OOM) errors, as described in this issue: <a href="https://github.com/huggingface/trl/issues/2250#issue-2598304204" rel="nofollow">#2250</a>.</p> <p data-svelte-h="svelte-gun0x9">If you encounter this issue, you can disable the gathering of model weights for generation by setting the following parameter:</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">GRPO </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">Online DPO </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">PPO </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">RLOO </div></div> <div class="language-select"><div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-keyword">from</span> trl <span class="hljs-keyword">import</span> GRPOConfig
training_args = GRPOConfig(..., ds3_gather_for_generation=<span class="hljs-literal">False</span>)<!-- HTML_TAG_END --></pre></div> </div> <p data-svelte-h="svelte-14mh10t">This adjustment prevents model weights from being gathered, avoiding OOM errors, but it may result in slower generation speeds.</p> <a class="!text-gray-400 !no-underline text-sm flex items-center not-prose mt-4" href="https://github.com/huggingface/trl/blob/main/docs/source/reducing_memory_usage.md" target="_blank"><span data-svelte-h="svelte-1kd6by1">&lt;</span> <span data-svelte-h="svelte-x0xyl0">&gt;</span> <span data-svelte-h="svelte-1dajgef"><span class="underline ml-1.5">Update</span> on GitHub</span></a> <p></p>
<script>
 {
 __sveltekit_4tczb2 = {
 assets: "/docs/trl/pr_3582/en",
 base: "/docs/trl/pr_3582/en",
 env: {}
 };
 const element = document.currentScript.parentElement;
 const data = [null,null];
 Promise.all([
 import("/docs/trl/pr_3582/en/_app/immutable/entry/start.0f0f318c.js"),
 import("/docs/trl/pr_3582/en/_app/immutable/entry/app.b27a462f.js")
 ]).then(([kit, app]) => {
 kit.start(app, element, {
 node_ids: [0, 39],
 data,
 form: null,
 error: null
 });
 });
 }
 </script>

Xet Storage Details

Size:
35.6 kB
·
Xet hash:
de1dc9ddfdf61f0db10d2f499229ecc5c7d982b9cf426ad3f8cc15473b949c68

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