mirror of
https://github.com/NVIDIA/TensorRT-LLM.git
synced 2026-01-14 06:27:45 +08:00
336 lines
42 KiB
HTML
336 lines
42 KiB
HTML
|
||
|
||
<!DOCTYPE html>
|
||
<html class="writer-html5" lang="en" data-content_root="../">
|
||
<head>
|
||
<meta charset="utf-8" /><meta name="viewport" content="width=device-width, initial-scale=1" />
|
||
|
||
<meta name="viewport" content="width=device-width, initial-scale=1.0" />
|
||
<title>Executor API — tensorrt_llm documentation</title>
|
||
<link rel="stylesheet" type="text/css" href="../_static/pygments.css?v=80d5e7a1" />
|
||
<link rel="stylesheet" type="text/css" href="../_static/css/theme.css?v=e59714d7" />
|
||
<link rel="stylesheet" type="text/css" href="../_static/copybutton.css?v=76b2166b" />
|
||
|
||
|
||
<script src="../_static/jquery.js?v=5d32c60e"></script>
|
||
<script src="../_static/_sphinx_javascript_frameworks_compat.js?v=2cd50e6c"></script>
|
||
<script src="../_static/documentation_options.js?v=5929fcd5"></script>
|
||
<script src="../_static/doctools.js?v=9a2dae69"></script>
|
||
<script src="../_static/sphinx_highlight.js?v=dc90522c"></script>
|
||
<script src="../_static/clipboard.min.js?v=a7894cd8"></script>
|
||
<script src="../_static/copybutton.js?v=65e89d2a"></script>
|
||
<script src="../_static/js/theme.js"></script>
|
||
<link rel="index" title="Index" href="../genindex.html" />
|
||
<link rel="search" title="Search" href="../search.html" />
|
||
<link rel="next" title="Graph Rewriting Module" href="graph-rewriting.html" />
|
||
<link rel="prev" title="C++ GPT Runtime" href="gpt-runtime.html" />
|
||
</head>
|
||
|
||
<body class="wy-body-for-nav">
|
||
<div class="wy-grid-for-nav">
|
||
<nav data-toggle="wy-nav-shift" class="wy-nav-side">
|
||
<div class="wy-side-scroll">
|
||
<div class="wy-side-nav-search" >
|
||
|
||
|
||
|
||
<a href="../index.html" class="icon icon-home">
|
||
tensorrt_llm
|
||
</a>
|
||
<div role="search">
|
||
<form id="rtd-search-form" class="wy-form" action="../search.html" method="get">
|
||
<input type="text" name="q" placeholder="Search docs" aria-label="Search docs" />
|
||
<input type="hidden" name="check_keywords" value="yes" />
|
||
<input type="hidden" name="area" value="default" />
|
||
</form>
|
||
</div>
|
||
</div><div class="wy-menu wy-menu-vertical" data-spy="affix" role="navigation" aria-label="Navigation menu">
|
||
<p class="caption" role="heading"><span class="caption-text">Getting Started</span></p>
|
||
<ul>
|
||
<li class="toctree-l1"><a class="reference internal" href="../overview.html">Overview</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="../quick-start-guide.html">Quick Start Guide</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="../key-features.html">Key Features</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="../release-notes.html">Release Notes</a></li>
|
||
</ul>
|
||
<p class="caption" role="heading"><span class="caption-text">Installation</span></p>
|
||
<ul>
|
||
<li class="toctree-l1"><a class="reference internal" href="../installation/linux.html">Installing on Linux</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="../installation/build-from-source-linux.html">Building from Source Code on Linux</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="../installation/windows.html">Installing on Windows</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="../installation/build-from-source-windows.html">Building from Source Code on Windows</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="../installation/grace-hopper.html">Installing on Grace Hopper</a></li>
|
||
</ul>
|
||
<p class="caption" role="heading"><span class="caption-text">LLM API</span></p>
|
||
<ul>
|
||
<li class="toctree-l1"><a class="reference internal" href="../llm-api/index.html">API Introduction</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="../llm-api/reference.html">API Reference</a></li>
|
||
</ul>
|
||
<p class="caption" role="heading"><span class="caption-text">LLM API Examples</span></p>
|
||
<ul>
|
||
<li class="toctree-l1"><a class="reference internal" href="../llm-api-examples/index.html">LLM Examples Introduction</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="../llm-api-examples/customization.html">Common Customizations</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="../llm-api-examples/llm_api_examples.html">Examples</a></li>
|
||
</ul>
|
||
<p class="caption" role="heading"><span class="caption-text">Model Definition API</span></p>
|
||
<ul>
|
||
<li class="toctree-l1"><a class="reference internal" href="../python-api/tensorrt_llm.layers.html">Layers</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="../python-api/tensorrt_llm.functional.html">Functionals</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="../python-api/tensorrt_llm.models.html">Models</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="../python-api/tensorrt_llm.plugin.html">Plugin</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="../python-api/tensorrt_llm.quantization.html">Quantization</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="../python-api/tensorrt_llm.runtime.html">Runtime</a></li>
|
||
</ul>
|
||
<p class="caption" role="heading"><span class="caption-text">C++ API</span></p>
|
||
<ul>
|
||
<li class="toctree-l1"><a class="reference internal" href="../_cpp_gen/executor.html">Executor</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="../_cpp_gen/runtime.html">Runtime</a></li>
|
||
</ul>
|
||
<p class="caption" role="heading"><span class="caption-text">Command-Line Reference</span></p>
|
||
<ul>
|
||
<li class="toctree-l1"><a class="reference internal" href="../commands/trtllm-build.html">trtllm-build</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="../commands/trtllm-serve.html">trtllm-serve</a></li>
|
||
</ul>
|
||
<p class="caption" role="heading"><span class="caption-text">Architecture</span></p>
|
||
<ul>
|
||
<li class="toctree-l1"><a class="reference internal" href="../architecture/overview.html">TensorRT-LLM Architecture</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="../architecture/core-concepts.html">Model Definition</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="../architecture/core-concepts.html#compilation">Compilation</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="../architecture/core-concepts.html#runtime">Runtime</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="../architecture/core-concepts.html#multi-gpu-and-multi-node-support">Multi-GPU and Multi-Node Support</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="../architecture/checkpoint.html">TensorRT-LLM Checkpoint</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="../architecture/workflow.html">TensorRT-LLM Build Workflow</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="../architecture/add-model.html">Adding a Model</a></li>
|
||
</ul>
|
||
<p class="caption" role="heading"><span class="caption-text">Advanced</span></p>
|
||
<ul class="current">
|
||
<li class="toctree-l1"><a class="reference internal" href="gpt-attention.html">Multi-Head, Multi-Query, and Group-Query Attention</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="gpt-runtime.html">C++ GPT Runtime</a></li>
|
||
<li class="toctree-l1 current"><a class="current reference internal" href="#">Executor API</a><ul>
|
||
<li class="toctree-l2"><a class="reference internal" href="#api">API</a><ul>
|
||
<li class="toctree-l3"><a class="reference internal" href="#the-executor-class">The Executor Class</a></li>
|
||
<li class="toctree-l3"><a class="reference internal" href="#the-request-class">The Request Class</a></li>
|
||
<li class="toctree-l3"><a class="reference internal" href="#the-response-class">The Response Class</a></li>
|
||
<li class="toctree-l3"><a class="reference internal" href="#the-result-class">The Result Class</a></li>
|
||
<li class="toctree-l3"><a class="reference internal" href="#sending-requests-with-different-beam-widths">Sending Requests with Different Beam Widths</a></li>
|
||
<li class="toctree-l3"><a class="reference internal" href="#controlling-output-with-logits-post-processor">Controlling output with Logits Post-Processor</a></li>
|
||
<li class="toctree-l3"><a class="reference internal" href="#structured-output-with-guided-decoding">Structured output with guided decoding</a></li>
|
||
</ul>
|
||
</li>
|
||
<li class="toctree-l2"><a class="reference internal" href="#c-executor-api-example">C++ Executor API Example</a></li>
|
||
<li class="toctree-l2"><a class="reference internal" href="#python-bindings-for-the-executor-api">Python Bindings for the Executor API</a></li>
|
||
<li class="toctree-l2"><a class="reference internal" href="#in-flight-batching-with-the-triton-inference-server">In-flight Batching with the Triton Inference Server</a></li>
|
||
</ul>
|
||
</li>
|
||
<li class="toctree-l1"><a class="reference internal" href="graph-rewriting.html">Graph Rewriting Module</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="inference-request.html">Inference Request</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="inference-request.html#responses">Responses</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="lora.html">Run gpt-2b + LoRA using GptManager / cpp runtime</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="expert-parallelism.html">Expert Parallelism in TensorRT-LLM</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="kv-cache-reuse.html">KV cache reuse</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="speculative-decoding.html">Speculative Sampling</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="disaggregated-service.html">Disaggregated-Service (experimental)</a></li>
|
||
</ul>
|
||
<p class="caption" role="heading"><span class="caption-text">Performance</span></p>
|
||
<ul>
|
||
<li class="toctree-l1"><a class="reference internal" href="../performance/perf-overview.html">Overview</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="../performance/perf-benchmarking.html">Benchmarking</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="../performance/perf-best-practices.html">Best Practices</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="../performance/perf-analysis.html">Performance Analysis</a></li>
|
||
</ul>
|
||
<p class="caption" role="heading"><span class="caption-text">Reference</span></p>
|
||
<ul>
|
||
<li class="toctree-l1"><a class="reference internal" href="../reference/troubleshooting.html">Troubleshooting</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="../reference/support-matrix.html">Support Matrix</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="../reference/precision.html">Numerical Precision</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="../reference/memory.html">Memory Usage of TensorRT-LLM</a></li>
|
||
</ul>
|
||
<p class="caption" role="heading"><span class="caption-text">Blogs</span></p>
|
||
<ul>
|
||
<li class="toctree-l1"><a class="reference internal" href="../blogs/H100vsA100.html">H100 has 4.6x A100 Performance in TensorRT-LLM, achieving 10,000 tok/s at 100ms to first token</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="../blogs/H200launch.html">H200 achieves nearly 12,000 tokens/sec on Llama2-13B with TensorRT-LLM</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="../blogs/Falcon180B-H200.html">Falcon-180B on a single H200 GPU with INT4 AWQ, and 6.7x faster Llama-70B over A100</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="../blogs/quantization-in-TRT-LLM.html">Speed up inference with SOTA quantization techniques in TRT-LLM</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="../blogs/XQA-kernel.html">New XQA-kernel provides 2.4x more Llama-70B throughput within the same latency budget</a></li>
|
||
</ul>
|
||
|
||
</div>
|
||
</div>
|
||
</nav>
|
||
|
||
<section data-toggle="wy-nav-shift" class="wy-nav-content-wrap"><nav class="wy-nav-top" aria-label="Mobile navigation menu" >
|
||
<i data-toggle="wy-nav-top" class="fa fa-bars"></i>
|
||
<a href="../index.html">tensorrt_llm</a>
|
||
</nav>
|
||
|
||
<div class="wy-nav-content">
|
||
<div class="rst-content">
|
||
<div role="navigation" aria-label="Page navigation">
|
||
<ul class="wy-breadcrumbs">
|
||
<li><a href="../index.html" class="icon icon-home" aria-label="Home"></a></li>
|
||
<li class="breadcrumb-item active">Executor API</li>
|
||
<li class="wy-breadcrumbs-aside">
|
||
<a href="../_sources/advanced/executor.md.txt" rel="nofollow"> View page source</a>
|
||
</li>
|
||
</ul>
|
||
<hr/>
|
||
</div>
|
||
<div role="main" class="document" itemscope="itemscope" itemtype="http://schema.org/Article">
|
||
<div itemprop="articleBody">
|
||
|
||
<section id="executor-api">
|
||
<span id="executor"></span><h1>Executor API<a class="headerlink" href="#executor-api" title="Link to this heading"></a></h1>
|
||
<p>TensorRT-LLM includes a high-level C++ API called the Executor API which allows you to execute requests
|
||
asynchronously, with in-flight batching, and without the need to define callbacks.</p>
|
||
<p>A software component (referred to as “the client” in the text that follows) can interact
|
||
with the executor using the API defined in the <a class="reference external" href="https://github.com/NVIDIA/TensorRT-LLM/tree/user/kaiyu/update_disagg_doc/cpp/include/tensorrt_llm/executor/executor.h"><code class="docutils literal notranslate"><span class="pre">executor.h</span></code></a> file.
|
||
For details about the API, refer to the <span class="xref std std-ref">_cpp_gen/executor.rst</span>.</p>
|
||
<p>The following sections provide an overview of the main classes defined in the Executor API.</p>
|
||
<section id="api">
|
||
<h2>API<a class="headerlink" href="#api" title="Link to this heading"></a></h2>
|
||
<section id="the-executor-class">
|
||
<h3>The Executor Class<a class="headerlink" href="#the-executor-class" title="Link to this heading"></a></h3>
|
||
<p>The <code class="docutils literal notranslate"><span class="pre">Executor</span></code> class is responsible for receiving requests from the client, and providing responses for those requests. The executor is constructed by providing a path to a directory containing the TensorRT-LLM engine or buffers containing the engine and the model JSON configuration. The client can create requests and enqueue those requests for execution using the <code class="docutils literal notranslate"><span class="pre">enqueueRequest</span></code> or <code class="docutils literal notranslate"><span class="pre">enqueueRequests</span></code> methods of the <code class="docutils literal notranslate"><span class="pre">Executor</span></code> class. Enqueued requests will be scheduled for execution by the executor, and multiple independent requests can be batched together at every iteration of the main execution loop (a process often referred to as continuous batching or iteration-level batching). Responses for a particular request can be awaited for by calling the <code class="docutils literal notranslate"><span class="pre">awaitResponses</span></code> method, and by providing the request id. Alternatively, responses for any requests can be awaited for by omitting to provide the request id when calling <code class="docutils literal notranslate"><span class="pre">awaitResponses</span></code>. The <code class="docutils literal notranslate"><span class="pre">Executor</span></code> class also allows to cancel requests using the <code class="docutils literal notranslate"><span class="pre">cancelRequest</span></code> method and to obtain per-iteration and per-request statistics using the <code class="docutils literal notranslate"><span class="pre">getLatestIterationStats</span></code>.</p>
|
||
</section>
|
||
<section id="the-request-class">
|
||
<h3>The Request Class<a class="headerlink" href="#the-request-class" title="Link to this heading"></a></h3>
|
||
<p>The <code class="docutils literal notranslate"><span class="pre">Request</span></code> class is used to define properties of the request, such as the input token ids and the maximum number of tokens to generate. The <code class="docutils literal notranslate"><span class="pre">streaming</span></code> parameter can be used to indicate if the request should generate a response for each new generated tokens (<code class="docutils literal notranslate"><span class="pre">streaming</span> <span class="pre">=</span> <span class="pre">true</span></code>) or only after all tokens have been generated (<code class="docutils literal notranslate"><span class="pre">streaming</span> <span class="pre">=</span> <span class="pre">false</span></code>). Other mandatory parameters of the request include the sampling configuration (defined by the <code class="docutils literal notranslate"><span class="pre">SamplingConfig</span></code> class) which contains parameters controlling the decoding process and the output configuration (defined by the <code class="docutils literal notranslate"><span class="pre">OutputConfig</span></code> class) which controls what information should be included in the <code class="docutils literal notranslate"><span class="pre">Result</span></code> for a particular response.</p>
|
||
<p>Optional parameters can also be provided when constructing a request such as a list of bad words, a list of stop words, a client id, or configurations objects for prompt tuning, LoRA, or speculative decoding, or a number of sequences to generate for example.</p>
|
||
</section>
|
||
<section id="the-response-class">
|
||
<h3>The Response Class<a class="headerlink" href="#the-response-class" title="Link to this heading"></a></h3>
|
||
<p>The <code class="docutils literal notranslate"><span class="pre">awaitResponses</span></code> method of the <code class="docutils literal notranslate"><span class="pre">Executor</span></code> class returns a vector of responses. Each response contains the request id associated with this response, and also contains either an error or a <code class="docutils literal notranslate"><span class="pre">Result</span></code>. Check if the response has an error by using the <code class="docutils literal notranslate"><span class="pre">hasError</span></code> method before trying to obtain the <code class="docutils literal notranslate"><span class="pre">Result</span></code> associated with this response using the <code class="docutils literal notranslate"><span class="pre">getResult</span></code> method.</p>
|
||
</section>
|
||
<section id="the-result-class">
|
||
<h3>The Result Class<a class="headerlink" href="#the-result-class" title="Link to this heading"></a></h3>
|
||
<p>The <code class="docutils literal notranslate"><span class="pre">Result</span></code> class holds the result for a given request. It contains a Boolean parameter called <code class="docutils literal notranslate"><span class="pre">isFinal</span></code> that indicates if this is the last <code class="docutils literal notranslate"><span class="pre">Result</span></code> that will be returned for the given request id. It also contains the generated tokens. If the request is configured with <code class="docutils literal notranslate"><span class="pre">streaming</span> <span class="pre">=</span> <span class="pre">false</span></code> and <code class="docutils literal notranslate"><span class="pre">numReturnSequences</span> <span class="pre">=</span> <span class="pre">1</span></code>, a single response will be returned, the <code class="docutils literal notranslate"><span class="pre">isFinal</span></code> Boolean will be set to <code class="docutils literal notranslate"><span class="pre">true</span></code> and all generated tokens will be included in the <code class="docutils literal notranslate"><span class="pre">outputTokenIds</span></code>. If <code class="docutils literal notranslate"><span class="pre">streaming</span> <span class="pre">=</span> <span class="pre">true</span></code> and <code class="docutils literal notranslate"><span class="pre">numReturnSequences</span> <span class="pre">=</span> <span class="pre">1</span></code> is used, a <code class="docutils literal notranslate"><span class="pre">Result</span></code> will include one or more tokens (depending on the request <code class="docutils literal notranslate"><span class="pre">returnAllGeneratedTokens</span></code> parameter) except the last result and the <code class="docutils literal notranslate"><span class="pre">isFinal</span></code> flag will be set to <code class="docutils literal notranslate"><span class="pre">true</span></code> for the last result associated with this request.</p>
|
||
<p>The request <code class="docutils literal notranslate"><span class="pre">numReturnSequences</span></code> parameter controls the number of output sequences to generate for each prompt. When this option is used, the Executor will return at least <code class="docutils literal notranslate"><span class="pre">numReturnSequences</span></code> responses for each request, each containing one Result. In beam search (<code class="docutils literal notranslate"><span class="pre">beamWidth</span> <span class="pre">></span> <span class="pre">1</span></code>), the number of beams to be returned will be limited by <code class="docutils literal notranslate"><span class="pre">numReturnSequences</span></code> and the <code class="docutils literal notranslate"><span class="pre">sequenceIndex</span></code> attribute of the <code class="docutils literal notranslate"><span class="pre">Result</span></code> class will always be zero. Otherwise, in sampling (<code class="docutils literal notranslate"><span class="pre">beamWidth</span> <span class="pre">=</span> <span class="pre">1</span></code>), the <code class="docutils literal notranslate"><span class="pre">sequenceIndex</span></code> attribute indicates the index of the generated sequence in the result (<code class="docutils literal notranslate"><span class="pre">0</span> <span class="pre"><=</span> <span class="pre">sequenceIndex</span> <span class="pre"><</span> <span class="pre">numReturnSequences</span></code>). It contains a Boolean parameter called <code class="docutils literal notranslate"><span class="pre">isSequenceFinal</span></code> that indicates if this is the last result for the sequence and also contains a Boolean parameter <code class="docutils literal notranslate"><span class="pre">isFinal</span></code> that indicates when all sequences for the request have been generated. When <code class="docutils literal notranslate"><span class="pre">numReturnSequences</span> <span class="pre">=</span> <span class="pre">1</span></code>, <code class="docutils literal notranslate"><span class="pre">isFinal</span></code> is identical to <code class="docutils literal notranslate"><span class="pre">isSequenceFinal</span></code>.</p>
|
||
<p>Here is an example that shows how a subset of 3 responses might look like for <code class="docutils literal notranslate"><span class="pre">numReturnSequences</span> <span class="pre">=</span> <span class="pre">3</span></code>:</p>
|
||
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="n">Response</span> <span class="mi">1</span><span class="p">:</span> <span class="n">requestId</span> <span class="o">=</span> <span class="mi">1</span><span class="p">,</span> <span class="n">Result</span> <span class="k">with</span> <span class="n">sequenceIndex</span> <span class="o">=</span> <span class="mi">0</span><span class="p">,</span> <span class="n">isSequenceFinal</span> <span class="o">=</span> <span class="n">false</span><span class="p">,</span> <span class="n">isFinal</span> <span class="o">=</span> <span class="n">false</span>
|
||
<span class="n">Response</span> <span class="mi">2</span><span class="p">:</span> <span class="n">requestId</span> <span class="o">=</span> <span class="mi">1</span><span class="p">,</span> <span class="n">Result</span> <span class="k">with</span> <span class="n">sequenceIndex</span> <span class="o">=</span> <span class="mi">1</span><span class="p">,</span> <span class="n">isSequenceFinal</span> <span class="o">=</span> <span class="n">true</span><span class="p">,</span> <span class="n">isFinal</span> <span class="o">=</span> <span class="n">false</span>
|
||
<span class="n">Response</span> <span class="mi">3</span><span class="p">:</span> <span class="n">requestId</span> <span class="o">=</span> <span class="mi">1</span><span class="p">,</span> <span class="n">Result</span> <span class="k">with</span> <span class="n">sequenceIndex</span> <span class="o">=</span> <span class="mi">2</span><span class="p">,</span> <span class="n">isSequenceFinal</span> <span class="o">=</span> <span class="n">false</span><span class="p">,</span> <span class="n">isFinal</span> <span class="o">=</span> <span class="n">false</span>
|
||
</pre></div>
|
||
</div>
|
||
<p>In this example, each response contains one result for different sequences. The <code class="docutils literal notranslate"><span class="pre">isSequenceFinal</span></code> flag of the second Result is set to true, indicating that it is the last result for <code class="docutils literal notranslate"><span class="pre">sequenceIndex</span> <span class="pre">=</span> <span class="pre">1</span></code>, however, the isFinal flag of each Response is set to false because sequences 0 and 2 are not completed.</p>
|
||
</section>
|
||
<section id="sending-requests-with-different-beam-widths">
|
||
<h3>Sending Requests with Different Beam Widths<a class="headerlink" href="#sending-requests-with-different-beam-widths" title="Link to this heading"></a></h3>
|
||
<p>The executor can process requests with different beam widths if the following conditions are met:</p>
|
||
<ul class="simple">
|
||
<li><p>The model was built with a <code class="docutils literal notranslate"><span class="pre">max_beam_width</span> <span class="pre">></span> <span class="pre">1</span></code>.</p></li>
|
||
<li><p>The executor is configured with a <code class="docutils literal notranslate"><span class="pre">maxBeamWidth</span> <span class="pre">></span> <span class="pre">1</span></code> (the configured <code class="docutils literal notranslate"><span class="pre">maxBeamWidth</span></code> must be less than or equal to the model’s <code class="docutils literal notranslate"><span class="pre">max_beam_width</span></code>).</p></li>
|
||
<li><p>The requested beam widths are less than or equal to the configured <code class="docutils literal notranslate"><span class="pre">maxBeamWidth</span></code>.</p></li>
|
||
<li><p>For requests with two different beam widths, <code class="docutils literal notranslate"><span class="pre">x</span></code> and <code class="docutils literal notranslate"><span class="pre">y</span></code>, requests with beam width <code class="docutils literal notranslate"><span class="pre">y</span></code> are not enqueued until all responses for requests with beam width <code class="docutils literal notranslate"><span class="pre">x</span></code> have been awaited.</p></li>
|
||
</ul>
|
||
<p>The request queue of the executor must be empty to allow it to reconfigure itself for a new beam width. This reconfiguration will happen automatically when requests with a new beam width are enqueued. If requests with different beam widths are enqueued at the same time, the executor will encounter an error and terminate all requests prematurely.</p>
|
||
</section>
|
||
<section id="controlling-output-with-logits-post-processor">
|
||
<h3>Controlling output with Logits Post-Processor<a class="headerlink" href="#controlling-output-with-logits-post-processor" title="Link to this heading"></a></h3>
|
||
<p>Optionally, you can alter the logits produced by the network by providing an instance of <code class="docutils literal notranslate"><span class="pre">Executor::LogitsPostProcessorConfig</span></code>. For instance, this feature can be used to generate JSON formatted output. <a class="reference internal" href="../_cpp_gen/executor.html#_CPPv4N12tensorrt_llm8executor25LogitsPostProcessorConfigE" title="tensorrt_llm::executor::LogitsPostProcessorConfig"><code class="xref cpp cpp-class docutils literal notranslate"><span class="pre">Executor::LogitsPostProcessorConfig</span></code></a> specifies a map of named callbacks in the following form</p>
|
||
<div class="highlight-cpp notranslate"><div class="highlight"><pre><span></span><span class="n">std</span><span class="o">::</span><span class="n">unordered_map</span><span class="o"><</span><span class="n">std</span><span class="o">::</span><span class="n">string</span><span class="p">,</span><span class="w"> </span><span class="n">function</span><span class="o"><</span><span class="n">Tensor</span><span class="p">(</span><span class="n">IdType</span><span class="p">,</span><span class="w"> </span><span class="n">Tensor</span><span class="o">&</span><span class="p">,</span><span class="w"> </span><span class="n">BeamTokens</span><span class="w"> </span><span class="k">const</span><span class="o">&</span><span class="p">,</span><span class="w"> </span><span class="n">StreamPtr</span><span class="w"> </span><span class="k">const</span><span class="o">&</span><span class="p">,</span><span class="w"> </span><span class="n">std</span><span class="o">::</span><span class="n">optional</span><span class="o"><</span><span class="n">IdType</span><span class="o">></span><span class="p">)</span><span class="o">>></span>
|
||
</pre></div>
|
||
</div>
|
||
<p>The map key is the name associated with that logits post-processing callback. Each request can then specify the name of the logits post-processor to use for that particular request, if any.</p>
|
||
<p>The first argument to the callback is the request id, second is the logits tensor, third are the tokens produced by the request so far, fourth is the operation stream used by the logits tensor, and last one is an optional client id. The callback returns a modified tensor of logits. Multiple requests can share same client id and callback can use different logic based on client id.</p>
|
||
<p>You must use the stream to access the logits tensor. For example, to perform an addition with a bias tensor, the addition operation is enqueued on that stream. Alternatively, you can call <code class="docutils literal notranslate"><span class="pre">stream->synchronize()</span></code>, however, that will slow down the entire execution pipeline.</p>
|
||
<p>The executor also includes a <a class="reference internal" href="../_cpp_gen/executor.html#_CPPv4N12tensorrt_llm8executor26LogitsPostProcessorBatchedE" title="tensorrt_llm::executor::LogitsPostProcessorBatched"><code class="xref cpp cpp-class docutils literal notranslate"><span class="pre">LogitsPostProcessorBatched</span></code></a> method that enables altering logits of multiple requests in a batch. The batched method allows further optimizations and reduces callback overheads.</p>
|
||
<div class="highlight-cpp notranslate"><div class="highlight"><pre><span></span><span class="n">std</span><span class="o">::</span><span class="n">function</span><span class="o"><</span><span class="kt">void</span><span class="p">(</span><span class="n">std</span><span class="o">::</span><span class="n">vector</span><span class="o"><</span><span class="n">IdType</span><span class="o">></span><span class="w"> </span><span class="k">const</span><span class="o">&</span><span class="p">,</span><span class="w"> </span><span class="n">std</span><span class="o">::</span><span class="n">vector</span><span class="o"><</span><span class="n">Tensor</span><span class="o">>&</span><span class="p">,</span><span class="w"> </span><span class="n">std</span><span class="o">::</span><span class="n">vector</span><span class="o"><</span><span class="n">std</span><span class="o">::</span><span class="n">reference_wrapper</span><span class="o"><</span><span class="n">BeamTokens</span><span class="w"> </span><span class="k">const</span><span class="o">>></span><span class="w"> </span><span class="k">const</span><span class="o">&</span><span class="p">,</span><span class="w"> </span><span class="n">StreamPtr</span><span class="w"> </span><span class="k">const</span><span class="o">&</span><span class="p">,</span><span class="w"> </span><span class="n">std</span><span class="o">::</span><span class="n">vector</span><span class="o"><</span><span class="n">std</span><span class="o">::</span><span class="n">optional</span><span class="o"><</span><span class="n">IdType</span><span class="o">>></span><span class="w"> </span><span class="k">const</span><span class="o">&</span><span class="p">)</span><span class="o">></span>
|
||
</pre></div>
|
||
</div>
|
||
<p>A single batched callback can be specified in <code class="docutils literal notranslate"><span class="pre">LogitsPostProcessorConfig</span></code>. Each request can opt to apply this callback by specifying the name of the logits post-processor as <code class="docutils literal notranslate"><span class="pre">Request::kBatchedPostProcessorName</span></code>.</p>
|
||
<p>Note: Neither callback variant is supported with the <code class="docutils literal notranslate"><span class="pre">STATIC</span></code> batching type for the moment.</p>
|
||
<p>In a multi-GPU run, the callback is invoked on all ranks in the first tensor-parallel group, by default. To ensure correct execution, replicate the client-side state that is accessed by the callback on these ranks. If replication is expensive or infeasible, use <code class="docutils literal notranslate"><span class="pre">LogitsPostProcessorConfig::setReplicate(false)</span></code> to invoke the callback only on rank 0. The executor broadcasts the sampled tokens internally to ensure correct execution.</p>
|
||
</section>
|
||
<section id="structured-output-with-guided-decoding">
|
||
<h3>Structured output with guided decoding<a class="headerlink" href="#structured-output-with-guided-decoding" title="Link to this heading"></a></h3>
|
||
<p>Guided decoding controls the generation outputs to be amenable to pre-defined structured formats, e.g., JSON or XML. Currently, guided decoding is supported with the <a class="reference external" href="https://github.com/mlc-ai/xgrammar">XGrammar</a> backend.</p>
|
||
<p>To enable guided decoding, a valid instance of <code class="docutils literal notranslate"><span class="pre">GuidedDecodingConfig</span></code> must be provided when constructing <code class="docutils literal notranslate"><span class="pre">Executor</span></code>. <code class="docutils literal notranslate"><span class="pre">GuidedDecodingConfig</span></code> should be constructed with some tokenizer information, including <code class="docutils literal notranslate"><span class="pre">encodedVocab</span></code>, <code class="docutils literal notranslate"><span class="pre">tokenizerStr</span></code> (optional) and <code class="docutils literal notranslate"><span class="pre">stopTokenIds</span></code> (optional). Given a Hugging Face tokenizer, these can be extracted by:</p>
|
||
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="n">encoded_vocab</span> <span class="o">=</span> <span class="n">tokenizer</span><span class="o">.</span><span class="n">get_vocab</span><span class="p">()</span>
|
||
<span class="n">encoded_vocab</span> <span class="o">=</span> <span class="p">[</span><span class="n">token</span> <span class="k">for</span> <span class="n">token</span><span class="p">,</span> <span class="n">_</span> <span class="ow">in</span> <span class="nb">sorted</span><span class="p">(</span><span class="n">encoded_vocab</span><span class="o">.</span><span class="n">items</span><span class="p">(),</span> <span class="n">key</span><span class="o">=</span><span class="k">lambda</span> <span class="n">x</span><span class="p">:</span> <span class="n">x</span><span class="p">[</span><span class="mi">1</span><span class="p">])]</span>
|
||
<span class="n">tokenizer_str</span> <span class="o">=</span> <span class="n">tokenizer</span><span class="o">.</span><span class="n">backend_tokenizer</span><span class="o">.</span><span class="n">to_str</span><span class="p">()</span>
|
||
<span class="n">stop_token_ids</span> <span class="o">=</span> <span class="p">[</span><span class="n">tokenizer</span><span class="o">.</span><span class="n">eos_token_id</span><span class="p">]</span>
|
||
</pre></div>
|
||
</div>
|
||
<p>Refer to <a class="reference external" href="https://github.com/NVIDIA/TensorRT-LLM/tree/user/kaiyu/update_disagg_doc/tensorrt_llm/llmapi/tokenizer.py"><code class="docutils literal notranslate"><span class="pre">tensorrt_llm/llmapi/tokenizer.py</span></code></a> for more details. You may dump these materials to disk, and reload them to C++ runtime for use.</p>
|
||
<p>Each request can be optionally specified with a <code class="docutils literal notranslate"><span class="pre">GuidedDecodingParams</span></code>, which defines the desired structured format. Currently, it supports four types:</p>
|
||
<ul class="simple">
|
||
<li><p><code class="docutils literal notranslate"><span class="pre">GuidedDecodingParams::GuideType::kJSON</span></code>: The generated text is amenable to JSON format;</p></li>
|
||
<li><p><code class="docutils literal notranslate"><span class="pre">GuidedDecodingParams::GuideType::kJSON_SCHEMA</span></code>: The generated text is amenable to JSON format with additional restrictions;</p></li>
|
||
<li><p><code class="docutils literal notranslate"><span class="pre">GuidedDecodingParams::GuideType::kREGEX</span></code>: The generated text is amenable to regular expression;</p></li>
|
||
<li><p><code class="docutils literal notranslate"><span class="pre">GuidedDecodingParams::GuideType::kEBNF_GRAMMAR</span></code>: The generated text is amenable to the extended Backus-Naur form (EBNF) grammar.</p></li>
|
||
</ul>
|
||
<p>The latter three types should be used with the schema/regex/grammar provided to <code class="docutils literal notranslate"><span class="pre">GuidedDecodingParams</span></code>.</p>
|
||
</section>
|
||
</section>
|
||
<section id="c-executor-api-example">
|
||
<h2>C++ Executor API Example<a class="headerlink" href="#c-executor-api-example" title="Link to this heading"></a></h2>
|
||
<p>Two C++ examples are provided that shows how to use the Executor API and can be found in the <a class="reference external" href="https://github.com/NVIDIA/TensorRT-LLM/tree/user/kaiyu/update_disagg_doc/examples/cpp/executor/"><code class="docutils literal notranslate"><span class="pre">examples/cpp/executor</span></code></a> folder.</p>
|
||
</section>
|
||
<section id="python-bindings-for-the-executor-api">
|
||
<h2>Python Bindings for the Executor API<a class="headerlink" href="#python-bindings-for-the-executor-api" title="Link to this heading"></a></h2>
|
||
<p>Python bindings for the Executor API are also available to use the Executor API from Python. The Python bindings are defined in <a class="reference external" href="https://github.com/NVIDIA/TensorRT-LLM/tree/user/kaiyu/update_disagg_doc/cpp/tensorrt_llm/pybind/executor/bindings.cpp">bindings.cpp</a> and once built, are available in package <code class="docutils literal notranslate"><span class="pre">tensorrt_llm.bindings.executor</span></code>. Running <code class="docutils literal notranslate"><span class="pre">'help('tensorrt_llm.bindings.executor')</span></code> in a Python interpreter will provide an overview of the classes available.</p>
|
||
<p>In addition, three Python examples are provided to demonstrate how to use the Python bindings to the Executor API for single and multi-GPU models. They can be found in <a class="reference external" href="https://github.com/NVIDIA/TensorRT-LLM/tree/user/kaiyu/update_disagg_doc/examples/bindings"><code class="docutils literal notranslate"><span class="pre">examples/bindings</span></code></a>.</p>
|
||
</section>
|
||
<section id="in-flight-batching-with-the-triton-inference-server">
|
||
<h2>In-flight Batching with the Triton Inference Server<a class="headerlink" href="#in-flight-batching-with-the-triton-inference-server" title="Link to this heading"></a></h2>
|
||
<p>A Triton Inference Server C++ <a class="reference external" href="https://github.com/triton-inference-server/tensorrtllm_backend">backend</a> is provided with TensorRT-LLM that
|
||
includes the mechanisms needed to serve models using in-flight batching. That
|
||
backend is also a good starting example of how to implement in-flight batching using
|
||
the TensorRT-LLM C++ Executor API.</p>
|
||
</section>
|
||
</section>
|
||
|
||
|
||
</div>
|
||
</div>
|
||
<footer><div class="rst-footer-buttons" role="navigation" aria-label="Footer">
|
||
<a href="gpt-runtime.html" class="btn btn-neutral float-left" title="C++ GPT Runtime" accesskey="p" rel="prev"><span class="fa fa-arrow-circle-left" aria-hidden="true"></span> Previous</a>
|
||
<a href="graph-rewriting.html" class="btn btn-neutral float-right" title="Graph Rewriting Module" accesskey="n" rel="next">Next <span class="fa fa-arrow-circle-right" aria-hidden="true"></span></a>
|
||
</div>
|
||
|
||
<hr/>
|
||
|
||
<div role="contentinfo">
|
||
<jinja2.runtime.BlockReference object at 0x7da747e36000>
|
||
|
||
<div class="footer">
|
||
<p>
|
||
Copyright © 2024 NVIDIA Corporation
|
||
</p>
|
||
<p>
|
||
<a class="Link" href="https://www.nvidia.com/en-us/about-nvidia/privacy-policy/" target="_blank" rel="noopener"
|
||
data-cms-ai="0">Privacy Policy</a> |
|
||
<a class="Link" href="https://www.nvidia.com/en-us/about-nvidia/privacy-center/" target="_blank" rel="noopener"
|
||
data-cms-ai="0">Manage My Privacy</a> |
|
||
<a class="Link" href="https://www.nvidia.com/en-us/preferences/start/" target="_blank" rel="noopener"
|
||
data-cms-ai="0">Do Not Sell or Share My Data</a> |
|
||
<a class="Link" href="https://www.nvidia.com/en-us/about-nvidia/terms-of-service/" target="_blank"
|
||
rel="noopener" data-cms-ai="0">Terms of Service</a> |
|
||
<a class="Link" href="https://www.nvidia.com/en-us/about-nvidia/accessibility/" target="_blank" rel="noopener"
|
||
data-cms-ai="0">Accessibility</a> |
|
||
<a class="Link" href="https://www.nvidia.com/en-us/about-nvidia/company-policies/" target="_blank"
|
||
rel="noopener" data-cms-ai="0">Corporate Policies</a> |
|
||
<a class="Link" href="https://www.nvidia.com/en-us/product-security/" target="_blank" rel="noopener"
|
||
data-cms-ai="0">Product Security</a> |
|
||
<a class="Link" href="https://www.nvidia.com/en-us/contact/" target="_blank" rel="noopener"
|
||
data-cms-ai="0">Contact</a>
|
||
</p>
|
||
</div>
|
||
|
||
|
||
</div>
|
||
|
||
|
||
|
||
</footer>
|
||
</div>
|
||
</div>
|
||
</section>
|
||
</div>
|
||
<script>
|
||
jQuery(function () {
|
||
SphinxRtdTheme.Navigation.enable(true);
|
||
});
|
||
</script>
|
||
|
||
</body>
|
||
</html> |