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":[{"title":"Shutting down the manager","local":"shutting-down-the-manager","sections":[],"depth":3},{"title":"Adding requests","local":"adding-requests","sections":[],"depth":3},{"title":"Per-request sampling parameters","local":"per-request-sampling-parameters","sections":[],"depth":3},{"title":"Retrieving results","local":"retrieving-results","sections":[],"depth":3},{"title":"Streaming","local":"streaming","sections":[],"depth":3}],"depth":2},{"title":"ContinuousBatchingConfig","local":"continuousbatchingconfig","sections":[{"title":"Batch and scheduling limits","local":"batch-and-scheduling-limits","sections":[],"depth":3},{"title":"Log probabilities","local":"log-probabilities","sections":[],"depth":3},{"title":"CUDA graphs","local":"cuda-graphs","sections":[],"depth":3},{"title":"Async batching","local":"async-batching","sections":[],"depth":3},{"title":"Compilation","local":"compilation","sections":[],"depth":3},{"title":"Decode fast path","local":"decode-fast-path","sections":[],"depth":3},{"title":"CPU offloading","local":"cpu-offloading","sections":[],"depth":3},{"title":"Tensor parallel timeout","local":"tensor-parallel-timeout","sections":[],"depth":3},{"title":"Prefix caching","local":"prefix-caching","sections":[],"depth":3}],"depth":2},{"title":"Paged attention","local":"paged-attention","sections":[],"depth":2},{"title":"Tensor parallelism","local":"tensor-parallelism","sections":[],"depth":2},{"title":"Sliding window attention","local":"sliding-window-attention","sections":[],"depth":2},{"title":"Next steps","local":"next-steps","sections":[],"depth":2}],"depth":1}"> | |
| <link href="/docs/transformers/pr_43838/en/_app/immutable/assets/0.e3b0c442.css" rel="modulepreload"> | |
| <link rel="modulepreload" href="/docs/transformers/pr_43838/en/_app/immutable/entry/start.1bcc8411.js"> | |
| <link rel="modulepreload" href="/docs/transformers/pr_43838/en/_app/immutable/chunks/scheduler.31fdf58d.js"> | |
| <link rel="modulepreload" href="/docs/transformers/pr_43838/en/_app/immutable/chunks/singletons.49522dd1.js"> | |
| <link rel="modulepreload" href="/docs/transformers/pr_43838/en/_app/immutable/chunks/index.252883d5.js"> | |
| <link rel="modulepreload" href="/docs/transformers/pr_43838/en/_app/immutable/chunks/paths.91497448.js"> | |
| <link rel="modulepreload" href="/docs/transformers/pr_43838/en/_app/immutable/entry/app.9d434a07.js"> | |
| <link rel="modulepreload" href="/docs/transformers/pr_43838/en/_app/immutable/chunks/preload-helper.f02a9c49.js"> | |
| <link rel="modulepreload" href="/docs/transformers/pr_43838/en/_app/immutable/chunks/index.2f76fdf0.js"> | |
| <link rel="modulepreload" href="/docs/transformers/pr_43838/en/_app/immutable/nodes/0.b44e4c87.js"> | |
| <link rel="modulepreload" href="/docs/transformers/pr_43838/en/_app/immutable/chunks/each.e59479a4.js"> | |
| <link rel="modulepreload" href="/docs/transformers/pr_43838/en/_app/immutable/nodes/35.29836eb3.js"> | |
| <link rel="modulepreload" href="/docs/transformers/pr_43838/en/_app/immutable/chunks/CopyLLMTxtMenu.f658c69c.js"> | |
| <link rel="modulepreload" href="/docs/transformers/pr_43838/en/_app/immutable/chunks/MermaidChart.svelte_svelte_type_style_lang.0d9e7a9c.js"> | |
| <link rel="modulepreload" href="/docs/transformers/pr_43838/en/_app/immutable/chunks/IconCopy.ac192424.js"> | |
| <link rel="modulepreload" href="/docs/transformers/pr_43838/en/_app/immutable/chunks/CodeBlock.e52df5d6.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":[{"title":"Shutting down the manager","local":"shutting-down-the-manager","sections":[],"depth":3},{"title":"Adding requests","local":"adding-requests","sections":[],"depth":3},{"title":"Per-request sampling parameters","local":"per-request-sampling-parameters","sections":[],"depth":3},{"title":"Retrieving results","local":"retrieving-results","sections":[],"depth":3},{"title":"Streaming","local":"streaming","sections":[],"depth":3}],"depth":2},{"title":"ContinuousBatchingConfig","local":"continuousbatchingconfig","sections":[{"title":"Batch and scheduling limits","local":"batch-and-scheduling-limits","sections":[],"depth":3},{"title":"Log probabilities","local":"log-probabilities","sections":[],"depth":3},{"title":"CUDA graphs","local":"cuda-graphs","sections":[],"depth":3},{"title":"Async batching","local":"async-batching","sections":[],"depth":3},{"title":"Compilation","local":"compilation","sections":[],"depth":3},{"title":"Decode fast path","local":"decode-fast-path","sections":[],"depth":3},{"title":"CPU offloading","local":"cpu-offloading","sections":[],"depth":3},{"title":"Tensor parallel timeout","local":"tensor-parallel-timeout","sections":[],"depth":3},{"title":"Prefix caching","local":"prefix-caching","sections":[],"depth":3}],"depth":2},{"title":"Paged attention","local":"paged-attention","sections":[],"depth":2},{"title":"Tensor parallelism","local":"tensor-parallelism","sections":[],"depth":2},{"title":"Sliding window attention","local":"sliding-window-attention","sections":[],"depth":2},{"title":"Next steps","local":"next-steps","sections":[],"depth":2}],"depth":1}"><!-- HEAD_svelte-u9bgzb_END --> <p></p> <div class="items-center shrink-0 min-w-[100px] max-sm:min-w-[50px] justify-end ml-auto flex" style="float: right; margin-left: 10px; display: inline-flex; position: relative; z-index: 10;"><div class="inline-flex rounded-md max-sm:rounded-sm"><button class="inline-flex items-center gap-1 h-7 max-sm:h-7 px-2 max-sm:px-1.5 text-sm font-medium text-gray-800 border border-r-0 rounded-l-md max-sm:rounded-l-sm border-gray-200 bg-white hover:shadow-inner dark:border-gray-850 dark:bg-gray-950 dark:text-gray-200 dark:hover:bg-gray-800" aria-live="polite"><span class="inline-flex items-center justify-center rounded-md p-0.5 max-sm:p-0 hover:text-gray-800 dark:hover:text-gray-200"><svg class="sm:size-3.5 size-3" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg></span> <span>Copy page</span></button> <button class="inline-flex items-center justify-center w-6 max-sm:w-5 h-7 max-sm:h-7 disabled:pointer-events-none text-sm text-gray-500 hover:text-gray-700 dark:hover:text-white rounded-r-md max-sm:rounded-r-sm border border-l transition border-gray-200 bg-white hover:shadow-inner dark:border-gray-850 dark:bg-gray-950 dark:text-gray-200 dark:hover:bg-gray-800" aria-haspopup="menu" aria-expanded="false" aria-label="Open copy menu"><svg class="transition-transform text-gray-400 overflow-visible sm:size-3.5 size-3 rotate-0" width="1em" height="1em" viewBox="0 0 12 7" fill="none" xmlns="http://www.w3.org/2000/svg"><path d="M1 1L6 6L11 1" stroke="currentColor"></path></svg></button></div> </div> <h1 class="relative group"><a id="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-10zdlgk">Continuous batching maximizes GPU utilization by dynamically rescheduling the batch at every generation step. As requests finish, new ones join immediately instead of waiting for the whole batch to complete. The GPU stays full and throughput stays high.</p> <blockquote class="tip" data-svelte-h="svelte-18p8hgd"><p>For production deployments, use <a href="./serve-cli/serving_optims#continuous-batching">transformers serve</a>. It builds on <a href="/docs/transformers/pr_43838/en/main_classes/text_generation#transformers.ContinuousBatchingManager">ContinuousBatchingManager</a> and exposes an OpenAI-compatible HTTP endpoint.</p></blockquote> <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-r7cb4v">Continuous batching is supported through <a href="/docs/transformers/pr_43838/en/main_classes/text_generation#transformers.ContinuousMixin.generate_batch">generate_batch()</a>. Pass a list of tokenized prompts and get back results for all of them when they’re done. <code>generate_batch</code> handles scheduling internally and blocks until all requests are complete.</p> <p data-svelte-h="svelte-1dubnml">For serving and streaming use cases, use <a href="#continuousbatchingmanager">ContinuousBatchingManager</a> directly to manage requests.</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class="language-py "><!-- HTML_TAG_START --><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> ContinuousBatchingConfig, GenerationConfig | |
| model = AutoModelForCausalLM.from_pretrained( | |
| <span class="hljs-string">"Qwen/Qwen3-4B"</span>, | |
| attn_implementation=<span class="hljs-string">"flash_attention_2"</span>, | |
| device_map=<span class="hljs-string">"cuda"</span>, | |
| dtype=torch.bfloat16, | |
| ) | |
| tokenizer = AutoTokenizer.from_pretrained(<span class="hljs-string">"Qwen/Qwen3-4B"</span>) | |
| prompts = [ | |
| <span class="hljs-string">"Whats up?"</span>, | |
| <span class="hljs-string">"Name a cat breed."</span>, | |
| <span class="hljs-string">"Write a detailed history of quantum mechanics."</span>, | |
| ] | |
| inputs = [tokenizer.encode(p) <span class="hljs-keyword">for</span> p <span class="hljs-keyword">in</span> prompts] | |
| generation_config = GenerationConfig( | |
| max_new_tokens=<span class="hljs-number">64</span>, | |
| eos_token_id=tokenizer.eos_token_id, | |
| ) | |
| outputs = model.generate_batch(inputs=inputs, generation_config=generation_config) | |
| <span class="hljs-keyword">for</span> request_id, output <span class="hljs-keyword">in</span> outputs.items(): | |
| 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"[<span class="hljs-subst">{request_id}</span>] <span class="hljs-subst">{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-1qrwaol"><a href="/docs/transformers/pr_43838/en/main_classes/text_generation#transformers.ContinuousBatchingManager">ContinuousBatchingManager</a> runs a background thread and lets you submit requests and retrieve results independently. Every generation step, it checks for finished requests and schedules new ones to join the batch. This is useful for streaming, real-time serving, or submitting requests as they arrive.</p> <p data-svelte-h="svelte-w41x9n">Use <a href="/docs/transformers/pr_43838/en/main_classes/text_generation#transformers.ContinuousMixin.continuous_batching_context_manager">continuous_batching_context_manager()</a> to start and stop the manager safely. The example below contains variable length inputs. As soon as the shortest prompt is complete, it leaves the batch while the longer prompts continue generating. With static batching, you’d have to pad them all to the same length. Continuous batching frees up the completed prompt so you can start processing the next prompt immediately.</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class="language-py "><!-- HTML_TAG_START --><span class="hljs-keyword">with</span> model.continuous_batching_context_manager(generation_config=generation_config) <span class="hljs-keyword">as</span> manager: | |
| manager.add_request( | |
| input_ids=tokenizer.encode(<span class="hljs-string">"Write a detailed history of quantum mechanics."</span>), | |
| request_id=<span class="hljs-string">"long"</span>, | |
| max_new_tokens=<span class="hljs-number">512</span>, | |
| ) | |
| manager.add_request( | |
| input_ids=tokenizer.encode(<span class="hljs-string">"What's up?"</span>), | |
| request_id=<span class="hljs-string">"short_0"</span>, | |
| max_new_tokens=<span class="hljs-number">32</span>, | |
| ) | |
| manager.add_request( | |
| input_ids=tokenizer.encode(<span class="hljs-string">"Name a cat breed."</span>), | |
| request_id=<span class="hljs-string">"short_1"</span>, | |
| max_new_tokens=<span class="hljs-number">32</span>, | |
| ) | |
| <span class="hljs-keyword">for</span> result <span class="hljs-keyword">in</span> manager: | |
| text = tokenizer.decode(result.generated_tokens, skip_special_tokens=<span class="hljs-literal">True</span>) | |
| <span class="hljs-built_in">print</span>(<span class="hljs-string">f"[<span class="hljs-subst">{result.request_id}</span>] <span class="hljs-subst">{text}</span>"</span>)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1h0ii8j">You could also call <a href="/docs/transformers/pr_43838/en/main_classes/text_generation#transformers.ContinuousMixin.init_continuous_batching">init_continuous_batching()</a> to manage the lifecycle yourself.</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class="language-py "><!-- HTML_TAG_START -->manager = model.init_continuous_batching(generation_config=generation_config) | |
| manager.start() | |
| <span class="hljs-comment"># submit and retrieve requests...</span><!-- HTML_TAG_END --></pre></div> <h3 class="relative group"><a id="shutting-down-the-manager" 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="#shutting-down-the-manager"><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>Shutting down the manager</span></h3> <p data-svelte-h="svelte-1ym91bm">The manager runs a background thread and holds distributed resources. Shutdown happens in two stages so you can choose what to do with in-flight work.</p> <p data-svelte-h="svelte-a6hv2i">Call <a href="/docs/transformers/pr_43838/en/main_classes/text_generation#transformers.ContinuousBatchingManager.stop">stop()</a> to halt the background thread. By default, the manager stops accepting new submissions and waits for queued and active requests to finish before the thread exits.</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class="language-py "><!-- HTML_TAG_START -->manager.stop()<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-8vl1hm">Pass <code>hard_stop=True</code> to abandon pending work immediately. Queued and active requests are failed with a <code>RuntimeError</code> instead of finishing.</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class="language-py "><!-- HTML_TAG_START -->manager.stop(hard_stop=<span class="hljs-literal">True</span>)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1qvd22j">Once <code>stop</code> is called, <a href="/docs/transformers/pr_43838/en/main_classes/text_generation#transformers.ContinuousBatchingManager.add_request">add_request()</a> and <a href="/docs/transformers/pr_43838/en/main_classes/text_generation#transformers.ContinuousBatchingManager.add_requests">add_requests()</a> drop new submissions and log a warning. You can still call <code>start</code> again to run another generation session with the same manager.</p> <p data-svelte-h="svelte-x9uc1v">Call <a href="/docs/transformers/pr_43838/en/main_classes/text_generation#transformers.ContinuousBatchingManager.destroy">destroy()</a> to release distributed resources. <code>destroy</code> stops the manager first if it’s still running, and the manager cannot be restarted afterwards. Use it when you’re done with continuous batching for the lifetime of the process.</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class="language-py "><!-- HTML_TAG_START -->manager.destroy()<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1eoav8o"><a href="/docs/transformers/pr_43838/en/main_classes/text_generation#transformers.ContinuousMixin.continuous_batching_context_manager">continuous_batching_context_manager()</a> handles this process. It calls <code>stop</code> on exit and <code>destroy</code> unless you pass <code>persistent_manager=True</code> to cache the manager on the model for the next session.</p> <h3 class="relative group"><a id="adding-requests" 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="#adding-requests"><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>Adding requests</span></h3> <p data-svelte-h="svelte-hdymtv"><a href="/docs/transformers/pr_43838/en/main_classes/text_generation#transformers.ContinuousBatchingManager.add_request">add_request()</a> submits a single request. Provide a <code>request_id</code> or let the manager generate 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="language-py "><!-- HTML_TAG_START -->manager.add_request(input_ids=input_ids, request_id=<span class="hljs-string">"my_request"</span>)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-mfu4lj"><a href="/docs/transformers/pr_43838/en/main_classes/text_generation#transformers.ContinuousBatchingManager.add_requests">add_requests()</a> submits a batch at once. It sorts inputs automatically to maximize prefix cache hits when block sharing is enabled.</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class="language-py "><!-- HTML_TAG_START -->manager.add_requests(inputs=inputs)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1285gt0">Cancel a request with <a href="/docs/transformers/pr_43838/en/main_classes/text_generation#transformers.ContinuousBatchingManager.cancel_request">cancel_request()</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="language-py "><!-- HTML_TAG_START -->manager.cancel_request(request_id=<span class="hljs-string">"my_request"</span>)<!-- HTML_TAG_END --></pre></div> <h3 class="relative group"><a id="per-request-sampling-parameters" 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="#per-request-sampling-parameters"><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>Per-request sampling parameters</span></h3> <p data-svelte-h="svelte-19b8uc9">Enable <code>per_request_processors</code> to apply <code>temperature</code>, <code>top_k</code>, and <code>top_p</code> independently per request within the same forward pass to allow different sampling parameters for different requests (creative, high-temperature outputs versus precise, low-temperature ones for example).</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class="language-py "><!-- HTML_TAG_START -->cb_config = ContinuousBatchingConfig(per_request_processors=<span class="hljs-literal">True</span>) | |
| <span class="hljs-comment"># each request gets its own sampling parameters</span> | |
| manager.add_request(input_ids=inputs_a, temperature=<span class="hljs-number">0.9</span>, top_p=<span class="hljs-number">0.95</span>) | |
| manager.add_request(input_ids=inputs_b, temperature=<span class="hljs-number">0.1</span>, top_k=<span class="hljs-number">10</span>)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-161jqxl">Each parameter in <a href="/docs/transformers/pr_43838/en/main_classes/text_generation#transformers.GenerationConfig">GenerationConfig</a> must be a non-default value in order to create the associated logits processor at runtime. For example, set <code>temperature</code> to a value other than <code>None</code> or <code>1</code> to support per-request temperature control. Requests with temperatures of <code>1</code> can still be created afterwards.</p> <h3 class="relative group"><a id="retrieving-results" 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="#retrieving-results"><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>Retrieving results</span></h3> <p data-svelte-h="svelte-hvbpnr">Iterate over the manager to receive results as they arrive.</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class="language-py "><!-- HTML_TAG_START --><span class="hljs-keyword">for</span> result <span class="hljs-keyword">in</span> manager: | |
| <span class="hljs-built_in">print</span>(tokenizer.decode(result.generated_tokens, skip_special_tokens=<span class="hljs-literal">True</span>))<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-19f3cek"><a href="/docs/transformers/pr_43838/en/main_classes/text_generation#transformers.ContinuousBatchingManager.get_result">get_result()</a> fetches the next result from the output queue. Pass a <code>request_id</code> to filter for a specific request. If the next result in the queue doesn’t match, it’s requeued and the method returns <code>None</code>.</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class="language-py "><!-- HTML_TAG_START --><span class="hljs-comment"># next available result</span> | |
| result = manager.get_result() | |
| <span class="hljs-comment"># filter for a specific request</span> | |
| result = manager.get_result(request_id=<span class="hljs-string">"my_request"</span>)<!-- HTML_TAG_END --></pre></div> <h3 class="relative group"><a id="streaming" 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="#streaming"><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>Streaming</span></h3> <p data-svelte-h="svelte-1ln86gy">Set <code>streaming=True</code> on a request, then use <a href="/docs/transformers/pr_43838/en/main_classes/text_generation#transformers.ContinuousBatchingManager.request_id_iter">request_id_iter()</a> to iterate over partial outputs as tokens are generated.</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class="language-py "><!-- HTML_TAG_START --><span class="hljs-keyword">from</span> transformers.generation.continuous_batching <span class="hljs-keyword">import</span> RequestStatus | |
| manager.add_request(input_ids=input_ids, request_id=<span class="hljs-string">"streamed"</span>, streaming=<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">"streamed"</span>): | |
| token = tokenizer.decode(chunk.generated_tokens[-<span class="hljs-number">1</span>:], skip_special_tokens=<span class="hljs-literal">True</span>) | |
| <span class="hljs-built_in">print</span>(token, end=<span class="hljs-string">""</span>, flush=<span class="hljs-literal">True</span>) | |
| <span class="hljs-keyword">if</span> chunk.status == RequestStatus.FINISHED: | |
| <span class="hljs-keyword">break</span><!-- HTML_TAG_END --></pre></div> <h2 class="relative group"><a id="continuousbatchingconfig" 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="#continuousbatchingconfig"><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>ContinuousBatchingConfig</span></h2> <p data-svelte-h="svelte-1vflxwp"><a href="/docs/transformers/pr_43838/en/main_classes/continuous_batching#transformers.ContinuousBatchingConfig">ContinuousBatchingConfig</a> controls the KV cache, scheduling, CUDA graphs, memory usage, and more. Pass it alongside <a href="/docs/transformers/pr_43838/en/main_classes/text_generation#transformers.GenerationConfig">GenerationConfig</a> to customize continuous batching.</p> <p data-svelte-h="svelte-15qe7ty">By default, <code>num_blocks</code> and <code>max_batch_tokens</code> are inferred automatically from available GPU memory. Use the table below to help you pick the appropriate features.</p> <table data-svelte-h="svelte-1extjki"><thead><tr><th>Feature</th> <th>Memory</th> <th>Throughput</th> <th>Latency</th></tr></thead> <tbody><tr><td><code>max_memory_percent</code> / <code>block_size</code></td> <td>✓ controls KV budget</td> <td></td> <td></td></tr> <tr><td><code>scheduler</code></td> <td></td> <td>✓ scheduling policy</td> <td>✓ TTFT</td></tr> <tr><td>CUDA graphs</td> <td>↑ graph storage</td> <td>✓ less dispatch overhead</td> <td>✓</td></tr> <tr><td>Async batching</td> <td>↑ ~2× I/O buffers</td> <td>✓ overlaps CPU/GPU</td> <td></td></tr> <tr><td>Compilation</td> <td>↑ warmup-time only</td> <td>✓ faster forward passes</td> <td>✓</td></tr> <tr><td>Decode fast path</td> <td>↑ block table per request</td> <td>✓ faster decode-only steps</td> <td>✓</td></tr> <tr><td>CPU offloading</td> <td>↑ pinned CPU memory</td> <td>✓ skips some re-prefills</td> <td></td></tr> <tr><td>Prefix caching</td> <td>↓ shared KV blocks</td> <td>✓ skips redundant prefill</td> <td>✓ TTFT</td></tr> <tr><td>Paged attention</td> <td>↓ no fragmentation</td> <td>✓ dynamic batch membership</td> <td></td></tr> <tr><td>Sliding window</td> <td>↓ bounded KV per layer</td> <td></td> <td></td></tr> <tr><td>Per-request processors</td> <td></td> <td>✓ mixed sampling params per batch</td> <td></td></tr> <tr><td><code>max_requests_per_batch</code></td> <td>↓ caps logits buffer</td> <td>✓ bounds batch size</td> <td></td></tr> <tr><td><code>safety_margin</code></td> <td></td> <td></td> <td>✓ protects decode latency</td></tr></tbody></table> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class="language-py "><!-- HTML_TAG_START --><span class="hljs-keyword">from</span> transformers.generation <span class="hljs-keyword">import</span> ContinuousBatchingConfig | |
| cb_config = ContinuousBatchingConfig( | |
| max_memory_percent=<span class="hljs-number">0.8</span>, <span class="hljs-comment"># fraction of free GPU memory to use for the KV cache</span> | |
| block_size=<span class="hljs-number">256</span>, <span class="hljs-comment"># KV cache block size in tokens</span> | |
| scheduler_type=<span class="hljs-string">"fifo"</span>, <span class="hljs-comment"># "fifo" or "prefill_first"</span> | |
| ) | |
| outputs = model.generate_batch( | |
| inputs=inputs, | |
| generation_config=generation_config, | |
| continuous_batching_config=cb_config, | |
| )<!-- HTML_TAG_END --></pre></div> <h3 class="relative group"><a id="batch-and-scheduling-limits" 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="#batch-and-scheduling-limits"><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>Batch and scheduling limits</span></h3> <p data-svelte-h="svelte-yrzifv"><code>max_requests_per_batch</code> caps how many requests run in a single forward pass. The model produces a logits tensor sized by the number of batch tokens and the vocabulary size, and it’s cast to <code>fp32</code> for sampling, which doubles the footprint. Each request only needs one next-token prediction, so a smaller cap keeps this tensor smaller and avoids out-of-memory errors on prefill-heavy batches with large vocabularies. By default it’s set to the number of prompts submitted, with a fallback of 1024, then capped to fit <code>max_batch_tokens</code> and <code>num_blocks</code>.</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class="language-py "><!-- HTML_TAG_START -->cb_config = ContinuousBatchingConfig(max_requests_per_batch=<span class="hljs-number">256</span>)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1y0irce"><code>safety_margin</code> reserves a fraction of the KV cache for active requests. When free blocks fall below <code>safety_margin * num_blocks</code>, the scheduler stops admitting new prefills and only continues decoding the requests already in progress. This prioritizes finishing active work over starting new requests, which protects decode latency and delays <a href="./continuous_batching_architecture#offloading">offloading</a>. The value must fall between <code>0</code> and <code>1</code>, where <code>0</code> disables the margin.</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class="language-py "><!-- HTML_TAG_START -->cb_config = ContinuousBatchingConfig(safety_margin=<span class="hljs-number">0.15</span>)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-2wku14">The default depends on the scheduler. For the FIFO scheduler, it’s <code>0.15</code> and for <code>prefill_first</code>, it’s <code>0.0</code>. See <a href="./continuous_batching_architecture#admission">Admission</a> for how the margin interacts with scheduling.</p> <h3 class="relative group"><a id="log-probabilities" 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="#log-probabilities"><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>Log probabilities</span></h3> <p data-svelte-h="svelte-1s0fn88"><a href="/docs/transformers/pr_43838/en/main_classes/continuous_batching#transformers.ContinuousBatchingConfig">ContinuousBatchingConfig</a> returns each generated token’s log probability when <code>return_logprobs=True</code>. This is useful for RL where logprobs are an input to some of the training loops.</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class="language-py "><!-- HTML_TAG_START -->cb_config = ContinuousBatchingConfig(return_logprobs=<span class="hljs-literal">True</span>) | |
| <span class="hljs-comment"># generate_batch()</span> | |
| <span class="hljs-keyword">for</span> request_id, output <span class="hljs-keyword">in</span> outputs.items(): | |
| <span class="hljs-keyword">for</span> token_id, log_prob <span class="hljs-keyword">in</span> <span class="hljs-built_in">zip</span>(output.generated_tokens, output.logprobs): | |
| token = tokenizer.decode([token_id]) | |
| <span class="hljs-built_in">print</span>(<span class="hljs-string">f"<span class="hljs-subst">{token}</span> | logprob: <span class="hljs-subst">{log_prob}</span>"</span>)<!-- HTML_TAG_END --></pre></div> <h3 class="relative group"><a id="cuda-graphs" 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="#cuda-graphs"><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>CUDA graphs</span></h3> <p data-svelte-h="svelte-1u6umkv">CUDA graphs eliminate CPU dispatch overhead by recording the GPU execution graph once and replaying it for batches with matching shapes. Enable them explicitly with <code>use_cuda_graph=True</code>.</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class="language-py "><!-- HTML_TAG_START -->cb_config = ContinuousBatchingConfig(use_cuda_graph=<span class="hljs-literal">True</span>)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1sub3ar">When active, the manager pads query and KV lengths to fixed intervals so shapes repeat and graphs reuse. Smaller values of <code>q_padding_interval_size</code> and <code>kv_padding_interval_size</code> reduce wasted compute on padding, but this means there are more unique shapes the graph has to record and store which costs more memory.</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class="language-py "><!-- HTML_TAG_START -->cb_config = ContinuousBatchingConfig( | |
| use_cuda_graph=<span class="hljs-literal">True</span>, | |
| q_padding_interval_size=<span class="hljs-number">64</span>, | |
| kv_padding_interval_size=<span class="hljs-number">16384</span>, | |
| )<!-- HTML_TAG_END --></pre></div> <h3 class="relative group"><a id="async-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="#async-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>Async batching</span></h3> <p data-svelte-h="svelte-gptoqa">Async batching overlaps CPU scheduling of the next batch with GPU computation of the current one. It requires CUDA graphs and roughly doubles the VRAM used for input tensors.</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class="language-py "><!-- HTML_TAG_START -->cb_config = ContinuousBatchingConfig( | |
| use_cuda_graph=<span class="hljs-literal">True</span>, | |
| use_async_batching=<span class="hljs-literal">True</span>, | |
| )<!-- HTML_TAG_END --></pre></div> <h3 class="relative group"><a id="compilation" 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="#compilation"><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>Compilation</span></h3> <p data-svelte-h="svelte-1mbblsk"><code>default_compile_level</code> applies <code>torch.compile</code> to the model’s forward passes. Compilation trades a one-time warmup cost for faster forward passes during generation. Higher levels run more aggressive optimization, which improves throughput but lengthens warmup. Keep the level low for tests and benchmark iteration, and raise it for long-running serving workloads where the warmup cost is paid back over many requests.</p> <p data-svelte-h="svelte-15aazza">The level ranges from <code>0</code> to <code>3</code>. Level <code>0</code> is the default and skips compilation entirely.</p> <table data-svelte-h="svelte-mz6pz4"><thead><tr><th>Level</th> <th><code>mode</code></th> <th><code>dynamic</code></th> <th>Trade-off</th></tr></thead> <tbody><tr><td>0</td> <td>—</td> <td>—</td> <td>No compilation (default), fastest startup</td></tr> <tr><td>1</td> <td><code>default</code></td> <td><code>True</code></td> <td>Modest speedup, short warmup</td></tr> <tr><td>2</td> <td><code>max-autotune-no-cudagraphs</code></td> <td><code>True</code></td> <td>More speedup, longer warmup</td></tr> <tr><td>3</td> <td><code>max-autotune-no-cudagraphs</code></td> <td><code>False</code></td> <td>Best throughput, longest warmup</td></tr></tbody></table> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class="language-py "><!-- HTML_TAG_START -->cb_config = ContinuousBatchingConfig(default_compile_level=<span class="hljs-number">1</span>)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-29yyuy">The level supplies a default <a href="/docs/transformers/pr_43838/en/internal/generation_utils#transformers.CompileConfig">CompileConfig</a> to the varlen and decode execution paths. It only applies to a path that has no explicit config, so <code>varlen_compile_config</code> and <code>decode_compile_config</code> take precedence when set. Under FlashAttention, the varlen path skips compilation because <code>max_seqlen_k</code> triggers frequent recompilation, so the level affects only the decode path in that case.</p> <h3 class="relative group"><a id="decode-fast-path" 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="#decode-fast-path"><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>Decode fast path</span></h3> <p data-svelte-h="svelte-1wdklbx">When a batch contains only decode requests (one query token per sequence), the manager can dispatch to the <code>flash_attn_with_kvcache</code> kernel instead of the variable-length kernel. This is faster than the varlen path because the kernel reads and writes the paged KV cache in-place through a block table rather than going through a manual update. See <a href="./paged_attention">Paged attention</a> for kernel-level details.</p> <p data-svelte-h="svelte-iel068">The fast path is sized by <code>max_blocks_per_request</code>, which dimensions the per-request block table. By default this is auto-inferred. If <code>max_prompt_length</code> and <code>max_generated_length</code> are set on the manager, the block table is sized to fit the maximum sequence length. Otherwise, a fallback default (32 blocks per request) is used.</p> <p data-svelte-h="svelte-1u3ptky">Set <code>max_blocks_per_request</code> to a specific value to size the block table explicitly. This is useful when you know the maximum sequence length per request and want to bound the block table memory cost.</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class="language-py "><!-- HTML_TAG_START -->cb_config = ContinuousBatchingConfig(max_blocks_per_request=<span class="hljs-number">64</span>)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-607upp">Set <code>max_blocks_per_request=0</code> to disable the fast path and force every batch through the varlen kernel. This recovers the pre-default behavior and is useful when the fast path is unavailable for your attention implementation (the manager also disables it automatically when the underlying kernel can’t be used).</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class="language-py "><!-- HTML_TAG_START -->cb_config = ContinuousBatchingConfig(max_blocks_per_request=<span class="hljs-number">0</span>)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1an69pb">The fast path relies on the <code>flash_attn_with_kvcache</code> kernel, which is available for two device and attention implementation combinations.</p> <table data-svelte-h="svelte-1pl37zg"><thead><tr><th>Device</th> <th><code>attn_implementation</code></th></tr></thead> <tbody><tr><td>CUDA</td> <td><code>flash_attention_2</code> or <code>flash_attention_3</code></td></tr> <tr><td>XPU</td> <td><a href="https://huggingface.co/kernels-community/flash-attn2" rel="nofollow">flash_attention_2</a></td></tr></tbody></table> <p data-svelte-h="svelte-16aftzg">For any other combination, or when the kernel can’t be imported, the manager falls back to the varlen path. It logs a warning only when you set <code>max_blocks_per_request</code> explicitly.</p> <h3 class="relative group"><a id="cpu-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="#cpu-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>CPU offloading</span></h3> <p data-svelte-h="svelte-makf1p">CPU offloading copies evicted KV cache blocks to a pre-allocated pinned CPU buffer when the GPU KV cache is full. After cache space becomes available, the manager copies the blocks back to the GPU and resumes the request without recomputing its prompt and generated tokens.</p> <p data-svelte-h="svelte-ca3xz6">Set <code>cpu_offload_space</code> to the CPU swap space in GiB. The default value, <code>0.0</code>, disables CPU offloading.</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class="language-py "><!-- HTML_TAG_START -->cb_config = ContinuousBatchingConfig(cpu_offload_space=<span class="hljs-number">8.0</span>)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1g3agp1">By default, <code>cpu_offload_space_safety_threshold=0.8</code> limits the requested space to 80% of available system RAM when <code>psutil</code> is installed. Set <code>cpu_offload_space=None</code> to size the swap pool from the safety threshold.</p> <h3 class="relative group"><a id="tensor-parallel-timeout" 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="#tensor-parallel-timeout"><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>Tensor parallel timeout</span></h3> <p data-svelte-h="svelte-whmacg">Under tensor parallelism, the manager creates a CPU communication group to coordinate request submissions, cancellations, and shutdown across ranks. <code>cpu_group_timeout</code> limits how long a collective on this group can block before the process crashes. If one rank stalls, the timeout prevents the others from waiting forever.</p> <p data-svelte-h="svelte-gabf09">Set a longer timeout for workloads that issue infrequent collectives, or pass <code>None</code> to disable 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="language-py "><!-- HTML_TAG_START -->cb_config = ContinuousBatchingConfig(cpu_group_timeout=<span class="hljs-number">600.0</span>)<!-- HTML_TAG_END --></pre></div> <h3 class="relative group"><a id="prefix-caching" 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="#prefix-caching"><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>Prefix caching</span></h3> <p data-svelte-h="svelte-1uubcg5">When multiple requests share a common prefix, like a system prompt, the manager reuses their KV cache blocks instead of recomputing them. This is enabled by default and requires all model layers to use full attention (it’s automatically disabled for sliding window models).</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class="language-py "><!-- HTML_TAG_START -->cb_config = ContinuousBatchingConfig( | |
| allow_block_sharing=<span class="hljs-literal">True</span>, <span class="hljs-comment"># default</span> | |
| )<!-- HTML_TAG_END --></pre></div> <h2 class="relative group"><a id="paged-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="#paged-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>Paged attention</span></h2> <p data-svelte-h="svelte-s10ikk">Continuous batching requires a paged attention backend. Set <code>attn_implementation</code> when loading the model. If you load a model with a non-paged backend (<code>"flash_attention_2"</code>), the <code>"paged|"</code> prefix is added automatically when continuous batching starts.</p> <table data-svelte-h="svelte-1mdsu6u"><thead><tr><th>Backend</th> <th><code>attn_implementation</code></th> <th>Requirements</th></tr></thead> <tbody><tr><td>FlashAttention</td> <td><code>“paged|flash_attention_2”</code></td> <td><code>flash-attn</code> package</td></tr> <tr><td>SDPA (PyTorch native)</td> <td><code>“paged|sdpa”</code></td> <td>None</td></tr> <tr><td>Eager</td> <td><code>“paged|eager”</code></td> <td>None</td></tr></tbody></table> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class="language-py "><!-- HTML_TAG_START -->model = AutoModelForCausalLM.from_pretrained( | |
| <span class="hljs-string">"Qwen/Qwen3-4B"</span>, | |
| 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> <h2 class="relative group"><a id="tensor-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="#tensor-parallelism"><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>Tensor parallelism</span></h2> <p data-svelte-h="svelte-4j3hf3">For models too large to fit on a single GPU, shard the weights across devices with tensor parallelism. Load the model with <code>tp_plan="auto"</code> and continuous batching reads the tensor parallel size from the model to size the paged KV cache per shard. See <a href="./tensor_parallelism">Tensor parallelism</a> for the list of supported architectures and how sharding 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="language-py "><!-- HTML_TAG_START --><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> ContinuousBatchingConfig, GenerationConfig | |
| model = AutoModelForCausalLM.from_pretrained( | |
| <span class="hljs-string">"Qwen/Qwen3-32B"</span>, | |
| attn_implementation=<span class="hljs-string">"paged|flash_attention_2"</span>, | |
| tp_plan=<span class="hljs-string">"auto"</span>, | |
| ) | |
| tokenizer = AutoTokenizer.from_pretrained(<span class="hljs-string">"Qwen/Qwen3-32B"</span>) | |
| inputs = [tokenizer.encode(p) <span class="hljs-keyword">for</span> p <span class="hljs-keyword">in</span> [<span class="hljs-string">"Whats up?"</span>, <span class="hljs-string">"Name a cat breed."</span>]] | |
| generation_config = GenerationConfig(max_new_tokens=<span class="hljs-number">64</span>, eos_token_id=tokenizer.eos_token_id) | |
| outputs = model.generate_batch(inputs=inputs, generation_config=generation_config)<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-1evgct3">Launch the script with <code>torchrun</code>, setting <code>--nproc-per-node</code> to the number of GPUs you want to shard across.</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class="language-shell "><!-- HTML_TAG_START -->torchrun --nproc-per-node 4 cb_tp.py<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-mx63zt">The tensor parallel size must divide the model’s <code>num_key_value_heads</code> (check the model config). The paged cache raises an error at startup otherwise, so choose an appropriate <code>--nproc-per-node</code>.</p> <blockquote class="warning" data-svelte-h="svelte-1mefv2x"><p>Don’t set <code>device_map</code> with <code>tp_plan</code>. The two conflict because <code>device_map</code> places whole modules on specific GPUs, while <code>tp_plan</code> shards those same parameters across all GPUs.</p></blockquote> <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-leb0w7">Models with sliding window attention (Mistral, Gemma 2) work with continuous batching. To manually configure a sliding window for fine-tuning or custom experiments, set it in the model config before loading.</p> <div class="code-block relative "><div class="absolute top-2.5 right-4"><button class="inline-flex items-center relative text-sm focus:text-green-500 cursor-pointer focus:outline-none transition duration-200 ease-in-out opacity-0 mx-0.5 text-gray-600 " title="code excerpt" type="button"><svg class="" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" fill="currentColor" focusable="false" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 32 32"><path d="M28,10V28H10V10H28m0-2H10a2,2,0,0,0-2,2V28a2,2,0,0,0,2,2H28a2,2,0,0,0,2-2V10a2,2,0,0,0-2-2Z" transform="translate(0)"></path><path d="M4,18H2V4A2,2,0,0,1,4,2H18V4H4Z" transform="translate(0)"></path><rect fill="none" width="32" height="32"></rect></svg> <div class="absolute pointer-events-none transition-opacity bg-black text-white py-1 px-2 leading-tight rounded font-normal shadow left-1/2 top-full transform -translate-x-1/2 translate-y-2 opacity-0"><div class="absolute bottom-full left-1/2 transform -translate-x-1/2 w-0 h-0 border-black border-4 border-t-0" style="border-left-color: transparent; border-right-color: transparent; "></div> Copied</div></button></div> <pre class="language-py "><!-- HTML_TAG_START --><span class="hljs-keyword">from</span> transformers <span class="hljs-keyword">import</span> AutoConfig, AutoModelForCausalLM | |
| 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|sdpa"</span>, | |
| device_map=<span class="hljs-string">"cuda"</span>, | |
| dtype=torch.bfloat16, | |
| )<!-- HTML_TAG_END --></pre></div> <p data-svelte-h="svelte-lwztss">Prefix caching is disabled automatically when sliding window attention is active.</p> <h2 class="relative group"><a id="next-steps" class="header-link block pr-1.5 text-lg no-hover:hidden with-hover:absolute with-hover:p-1.5 with-hover:opacity-0 with-hover:group-hover:opacity-100 with-hover:right-full" href="#next-steps"><span><svg class="" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg></span></a> <span>Next steps</span></h2> <ul data-svelte-h="svelte-1lttzwr"><li>The <a href="https://huggingface.co/blog/continuous_batching" rel="nofollow">Continuous batching blog post</a> covers KV caching, chunked prefill, and dynamic scheduling with performance benchmark numbers.</li> <li>For a deeper look at how the continuous batching system works, see the <a href="./continuous_batching_architecture">Continuous batching architecture</a> doc.</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/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_1o1ef6h = { | |
| assets: "/docs/transformers/pr_43838/en", | |
| base: "/docs/transformers/pr_43838/en", | |
| env: {} | |
| }; | |
| const element = document.currentScript.parentElement; | |
| const data = [null,null]; | |
| Promise.all([ | |
| import("/docs/transformers/pr_43838/en/_app/immutable/entry/start.1bcc8411.js"), | |
| import("/docs/transformers/pr_43838/en/_app/immutable/entry/app.9d434a07.js") | |
| ]).then(([kit, app]) => { | |
| kit.start(app, element, { | |
| node_ids: [0, 35], | |
| data, | |
| form: null, | |
| error: null | |
| }); | |
| }); | |
| } | |
| </script> | |
Xet Storage Details
- Size:
- 103 kB
- Xet hash:
- b912e3832d966af2133a8eb62e738fbdecd3788a49255151802912fbd460de81
·
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.