Buckets:
| <meta charset="utf-8" /><meta name="hf:doc:metadata" content="{"title":"Continuous batching","local":"continuous-batching","sections":[{"title":"generate_batch","local":"generatebatch","sections":[],"depth":2},{"title":"ContinuousBatchingManager","local":"continuousbatchingmanager","sections":[],"depth":2},{"title":"PagedAttention","local":"pagedattention","sections":[],"depth":2},{"title":"Sliding window attention","local":"sliding-window-attention","sections":[],"depth":2},{"title":"How it works","local":"how-it-works","sections":[],"depth":2},{"title":"Resources","local":"resources","sections":[],"depth":2}],"depth":1}"> | |
| <link href="/docs/transformers/pr_26617/en/_app/immutable/assets/0.e3b0c442.css" rel="modulepreload"> | |
| <link rel="modulepreload" href="/docs/transformers/pr_26617/en/_app/immutable/entry/start.b5ae2c21.js"> | |
| <link rel="modulepreload" href="/docs/transformers/pr_26617/en/_app/immutable/chunks/scheduler.31fdf58d.js"> | |
| <link rel="modulepreload" href="/docs/transformers/pr_26617/en/_app/immutable/chunks/singletons.512cdb48.js"> | |
| <link rel="modulepreload" href="/docs/transformers/pr_26617/en/_app/immutable/chunks/index.252883d5.js"> | |
| <link rel="modulepreload" href="/docs/transformers/pr_26617/en/_app/immutable/chunks/paths.81255c3b.js"> | |
| <link rel="modulepreload" href="/docs/transformers/pr_26617/en/_app/immutable/entry/app.9acf2c3e.js"> | |
| <link rel="modulepreload" href="/docs/transformers/pr_26617/en/_app/immutable/chunks/preload-helper.bb442aeb.js"> | |
| <link rel="modulepreload" href="/docs/transformers/pr_26617/en/_app/immutable/chunks/index.2f76fdf0.js"> | |
| <link rel="modulepreload" href="/docs/transformers/pr_26617/en/_app/immutable/nodes/0.da6b3909.js"> | |
| <link rel="modulepreload" href="/docs/transformers/pr_26617/en/_app/immutable/chunks/each.e59479a4.js"> | |
| <link rel="modulepreload" href="/docs/transformers/pr_26617/en/_app/immutable/nodes/30.3025937f.js"> | |
| <link rel="modulepreload" href="/docs/transformers/pr_26617/en/_app/immutable/chunks/CopyLLMTxtMenu.a69e059a.js"> | |
| <link rel="modulepreload" href="/docs/transformers/pr_26617/en/_app/immutable/chunks/MermaidChart.svelte_svelte_type_style_lang.e4c7f916.js"> | |
| <link rel="modulepreload" href="/docs/transformers/pr_26617/en/_app/immutable/chunks/IconCopy.ac192424.js"> | |
| <link rel="modulepreload" href="/docs/transformers/pr_26617/en/_app/immutable/chunks/CodeBlock.ab12f8e1.js"><!-- HEAD_svelte-u9bgzb_START --><meta name="hf:doc:metadata" content="{"title":"Continuous batching","local":"continuous-batching","sections":[{"title":"generate_batch","local":"generatebatch","sections":[],"depth":2},{"title":"ContinuousBatchingManager","local":"continuousbatchingmanager","sections":[],"depth":2},{"title":"PagedAttention","local":"pagedattention","sections":[],"depth":2},{"title":"Sliding window attention","local":"sliding-window-attention","sections":[],"depth":2},{"title":"How it works","local":"how-it-works","sections":[],"depth":2},{"title":"Resources","local":"resources","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="continuous-batching" 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="#continuous-batching"><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>Continuous batching</span></h1> <p data-svelte-h="svelte-18mo4lc">Continuous batching maximizes GPU utilization. It increases throughput and reduces latency by using dynamic scheduling to rearrange the batch at each step. The system removes completed requests and adds new requests immediately to prevent GPU idling. Chunked prefill prevents expensive prefill work from stalling the batch while still allowing new requests still join.</p> <p data-svelte-h="svelte-18szo6u">Continuous batching works with <a href="./serving">transformers serve</a>, a server for deploying local models, and <a href="/docs/transformers/pr_26617/en/main_classes/text_generation#transformers.ContinuousMixin.generate_batch">generate_batch()</a>.</p> <h2 class="relative group"><a id="generatebatch" 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="#generatebatch"><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>generate_batch</span></h2> <p data-svelte-h="svelte-sxofjh">The <a href="/docs/transformers/pr_26617/en/main_classes/text_generation#transformers.ContinuousMixin.generate_batch">generate_batch()</a> method works with all autoregressive text models. It accepts a list of tokenized inputs and a <a href="/docs/transformers/pr_26617/en/main_classes/text_generation#transformers.GenerationConfig">GenerationConfig</a> to configure generation 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=""><!-- HTML_TAG_START --><span class="hljs-keyword">import</span> datasets | |
| <span class="hljs-keyword">import</span> torch | |
| <span class="hljs-keyword">from</span> transformers <span class="hljs-keyword">import</span> AutoModelForCausalLM, AutoTokenizer | |
| <span class="hljs-keyword">from</span> transformers.generation <span class="hljs-keyword">import</span> GenerationConfig | |
| model = AutoModelForCausalLM.from_pretrained( | |
| <span class="hljs-string">"Qwen/Qwen3-4B-Instruct-2507"</span>, | |
| attn_implementation=<span class="hljs-string">"sdpa_paged"</span>, | |
| device_map=<span class="hljs-string">"cuda"</span>, | |
| dtype=torch.bfloat16, | |
| ) | |
| tokenizer = AutoTokenizer.from_pretrained(<span class="hljs-string">"Qwen/Qwen3-4B-Instruct-2507"</span>, padding_side=<span class="hljs-string">"left"</span>) | |
| dataset = datasets.load_dataset(<span class="hljs-string">"openai/gsm8k"</span>, <span class="hljs-string">"socratic"</span>, split=<span class="hljs-string">"test"</span>) | |
| dataset = dataset.select(<span class="hljs-built_in">range</span>(args.samples)) | |
| tokenized_datasets = dataset.<span class="hljs-built_in">map</span>(<span class="hljs-keyword">lambda</span> x: tokenizer(x[<span class="hljs-string">"question"</span>]), batched=<span class="hljs-literal">True</span>) | |
| simple_batch_inputs = [item[<span class="hljs-string">"input_ids"</span>] <span class="hljs-keyword">for</span> item <span class="hljs-keyword">in</span> tokenized_datasets] | |
| generation_config = GenerationConfig( | |
| max_new_tokens=<span class="hljs-number">32</span>, | |
| use_cuda_graph=<span class="hljs-literal">False</span>, | |
| eos_token_id=tokenizer.eos_token_id, | |
| pad_token_id=tokenizer.pad_token_id, | |
| do_sample=<span class="hljs-literal">False</span>, | |
| max_batch_tokens=<span class="hljs-number">512</span>, | |
| ) | |
| batch_outputs = model.generate_batch( | |
| inputs=simple_batch_inputs, | |
| generation_config=generation_config, | |
| ) | |
| <span class="hljs-keyword">for</span> request_id, output <span class="hljs-keyword">in</span> batch_outputs.items(): | |
| generated_text = tokenizer.decode(output.generated_tokens, skip_special_tokens=<span class="hljs-literal">True</span>) | |
| <span class="hljs-built_in">print</span>(<span class="hljs-string">f"Request <span class="hljs-subst">{request_id}</span> output: <span class="hljs-subst">{generated_text}</span>"</span>)<!-- HTML_TAG_END --></pre></div> <h2 class="relative group"><a id="continuousbatchingmanager" 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="#continuousbatchingmanager"><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>ContinuousBatchingManager</span></h2> <p data-svelte-h="svelte-ucf4zx">The <a href="/docs/transformers/pr_26617/en/main_classes/text_generation#transformers.ContinuousBatchingManager">ContinuousBatchingManager</a> orchestrates the background thread by pulling requests from the queue and filling the GPU to capacity. Every iteration checks for finished requests and schedules new ones to join the batch. Use this manager to customize request scheduling.</p> <p data-svelte-h="svelte-jnngf7">Call <a href="/docs/transformers/pr_26617/en/main_classes/text_generation#transformers.ContinuousMixin.init_continuous_batching">init_continuous_batching()</a> to initialize the manager with a <a href="/docs/transformers/pr_26617/en/main_classes/text_generation#transformers.GenerationConfig">GenerationConfig</a> and <a href="/docs/transformers/pr_26617/en/main_classes/text_generation#transformers.ContinuousBatchingManager.start">start()</a> the background thread.</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> transformers.generation.continuous_batching <span class="hljs-keyword">import</span> RequestStatus | |
| manager = model.init_continuous_batching(generation_config=generation_config) | |
| manager.start()<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-187htnb">Use <a href="/docs/transformers/pr_26617/en/main_classes/text_generation#transformers.ContinuousBatchingManager.add_request">add_request()</a> to asynchronously submit individual requests. Provide a specific request id or the manager wgenerates one automatically.</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">for</span> i, input_ids <span class="hljs-keyword">in</span> <span class="hljs-built_in">enumerate</span>(simple_batch_inputs): | |
| request_id = manager.add_request(input_ids=input_ids, request_id=<span class="hljs-string">f"request_<span class="hljs-subst">{i}</span>"</span>)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-mjd39x">Retrieve <em>all</em> results as they arrive with <a href="/docs/transformers/pr_26617/en/main_classes/text_generation#transformers.ContinuousBatchingManager.get_result">get_result()</a>.</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START --><span class="hljs-keyword">for</span> request_id, request <span class="hljs-keyword">in</span> manager.get_result(): | |
| generated_text = tokenizer.decode(request.generated_tokens, skip_special_tokens=<span class="hljs-literal">True</span>) | |
| <span class="hljs-built_in">print</span>(<span class="hljs-string">f"Request <span class="hljs-subst">{request_id}</span> output: <span class="hljs-subst">{generated_text}</span>"</span>)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-mf7hjh">Use the <code>request_id</code> of a specific request to get its results. This is a blocking operation that waits until the result is ready.</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 -->result = manager.get_result(request_id=<span class="hljs-string">"request_5"</span>)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-soon5q">Stream partial results for a specific request with <a href="/docs/transformers/pr_26617/en/main_classes/text_generation#transformers.ContinuousBatchingManager.request_id_iter">request_id_iter()</a>.</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class=""><!-- HTML_TAG_START -->manager.add_request( | |
| input_ids=input_ids, | |
| request_id=<span class="hljs-string">"streaming_request"</span>, | |
| stream=<span class="hljs-literal">True</span>, | |
| ) | |
| <span class="hljs-keyword">for</span> chunk <span class="hljs-keyword">in</span> manager.request_id_iter(request_id=<span class="hljs-string">"streaming_request"</span>): | |
| generated_text = tokenizer.decode(chunk.generated_tokens, skip_special_tokens=<span class="hljs-literal">True</span>) | |
| <span class="hljs-built_in">print</span>(generated_text) | |
| <span class="hljs-comment"># <span class="hljs-doctag">FIXME:</span> stop iteration in `request_id_iter` when finished instead of doing it externally</span> | |
| <span class="hljs-keyword">if</span> chunk.status == RequestStatus.FINISHED: | |
| <span class="hljs-keyword">break</span><!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1bte4zg">Call <a href="/docs/transformers/pr_26617/en/main_classes/text_generation#transformers.ContinuousBatchingManager.stop">stop()</a> to terminate the manager.</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 -->manager.stop()<!-- HTML_TAG_END --></pre></div> <h2 class="relative group"><a id="pagedattention" 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="#pagedattention"><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>PagedAttention</span></h2> <p data-svelte-h="svelte-9f8pns">PagedAttention breaks large key-value caches into smaller, non-contiguous fixed-size pages to avoid GPU memory fragmentation and support variable-length requests. Transformers automatically enables PagedAttention when using continuous batching.</p> <p data-svelte-h="svelte-1yu4u9x">You could explicitly enable PagedAttention when instantiating a model rather than waiting for <a href="/docs/transformers/pr_26617/en/main_classes/text_generation#transformers.ContinuousMixin.generate_batch">generate_batch()</a> to dynamically enable it.</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">import</span> torch | |
| <span class="hljs-keyword">from</span> transformers <span class="hljs-keyword">import</span> AutoModelForCausalLM | |
| model = AutoModelForCausalLM.from_pretrained( | |
| <span class="hljs-string">"Qwen/Qwen3-4B-Instruct-2507"</span>, | |
| attn_implementation=<span class="hljs-string">"paged|flash_attention_2"</span>, | |
| device_map=<span class="hljs-string">"cuda"</span>, | |
| torch_dtype=torch.bfloat16 | |
| )<!-- HTML_TAG_END --></pre></div> <h2 class="relative group"><a id="sliding-window-attention" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#sliding-window-attention"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>Sliding window attention</span></h2> <p data-svelte-h="svelte-1ale9j1">Sliding window attention limits the backward context of a token to save compute. Generation cost stays proportional to window size. This reduces compute per step and simplifies continuous batching.</p> <p data-svelte-h="svelte-1kalxhd">Transformers models like Mistral and Gemma 2 natively support sliding window attention. Manually enable it in the model config if the architecture supports it. This helps with fine-tuning or running custom experiments.</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> transformers <span class="hljs-keyword">import</span> AutoConfig | |
| config = AutoConfig.from_pretrained(<span class="hljs-string">"google/gemma-2-2b"</span>) | |
| config.sliding_window = <span class="hljs-number">4096</span> | |
| model = AutoModelForCausalLM.from_pretrained( | |
| <span class="hljs-string">"google/gemma-2-2b"</span>, | |
| config=config, | |
| attn_implementation=<span class="hljs-string">"paged|flash_attention_2"</span>, | |
| device_map=<span class="hljs-string">"cuda"</span>, | |
| dtype=torch.bfloat16, | |
| )<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-hjnj3n">Usage remains the same with <a href="/docs/transformers/pr_26617/en/main_classes/text_generation#transformers.ContinuousMixin.generate_batch">generate_batch()</a>.</p> <h2 class="relative group"><a id="how-it-works" 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="#how-it-works"><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>How it works</span></h2> <p data-svelte-h="svelte-1okgvli">The <a href="/docs/transformers/pr_26617/en/main_classes/text_generation#transformers.ContinuousMixin">ContinuousMixin</a> class serves as the main interface for continuous batching through <a href="/docs/transformers/pr_26617/en/main_classes/text_generation#transformers.ContinuousMixin.generate_batch">generate_batch()</a>. This method internally creates a <a href="/docs/transformers/pr_26617/en/main_classes/text_generation#transformers.ContinuousBatchingManager">ContinuousBatchingManager</a>.</p> <p data-svelte-h="svelte-18tqh0h"><a href="/docs/transformers/pr_26617/en/main_classes/text_generation#transformers.ContinuousBatchingManager">ContinuousBatchingManager</a> manages requests by creating a background thread for the generation loop and adding requests to the queue. The manager is thread-safe, allowing asynchronous request additions while the model generates.</p> <p data-svelte-h="svelte-1rat3qz">The <code>Scheduler</code> selects requests for processing at each step based on the token budget. <code>FIFOScheduler</code> is the default scheduler. It prioritizes decoding requests over prefilling requests and assigns them to specific memory blocks. <code>PrefillFirstScheduler</code> prioritizes prefill requests instead.</p> <p data-svelte-h="svelte-gzetn4"><a href="/docs/transformers/pr_26617/en/main_classes/text_generation#transformers.ContinuousBatchingManager">ContinuousBatchingManager</a> runs the model forward pass for the scheduled requests. It then collects and returns the results.</p> <h2 class="relative group"><a id="resources" 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="#resources"><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>Resources</span></h2> <p data-svelte-h="svelte-1vghojf">The <a href="https://huggingface.co/blog/continuous_batching" rel="nofollow">Continuous batching</a> blog post explains KV caching, chunked prefill, and ragged batching with dynamic scheduling in more detail.</p> <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/continuous_batching.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_1x0t0ja = { | |
| assets: "/docs/transformers/pr_26617/en", | |
| base: "/docs/transformers/pr_26617/en", | |
| env: {} | |
| }; | |
| const element = document.currentScript.parentElement; | |
| const data = [null,null]; | |
| Promise.all([ | |
| import("/docs/transformers/pr_26617/en/_app/immutable/entry/start.b5ae2c21.js"), | |
| import("/docs/transformers/pr_26617/en/_app/immutable/entry/app.9acf2c3e.js") | |
| ]).then(([kit, app]) => { | |
| kit.start(app, element, { | |
| node_ids: [0, 30], | |
| data, | |
| form: null, | |
| error: null | |
| }); | |
| }); | |
| } | |
| </script> | |
Xet Storage Details
- Size:
- 37.4 kB
- Xet hash:
- ebf70b27c02015c6556953b4944694a4e012ab8d0fc3c119d9b5c426334daf75
·
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.