mirror of
https://github.com/NVIDIA/TensorRT-LLM.git
synced 2026-01-14 06:27:45 +08:00
552 lines
47 KiB
HTML
552 lines
47 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>Best Practices for Tuning the Performance of TensorRT-LLM — 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=19f00094" />
|
||
|
||
|
||
<!--[if lt IE 9]>
|
||
<script src="../_static/js/html5shiv.min.js"></script>
|
||
<![endif]-->
|
||
|
||
<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/js/theme.js"></script>
|
||
<link rel="index" title="Index" href="../genindex.html" />
|
||
<link rel="search" title="Search" href="../search.html" />
|
||
<link rel="next" title="Performance Analysis" href="perf-analysis.html" />
|
||
<link rel="prev" title="Overview" href="perf-overview.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="../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>
|
||
</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>
|
||
<li class="toctree-l1"><a class="reference internal" href="../advanced/gpt-attention.html">Multi-Head, Multi-Query, and Group-Query Attention</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="../advanced/gpt-runtime.html">C++ GPT Runtime</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="../advanced/graph-rewriting.html">Graph Rewriting Module</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="../advanced/batch-manager.html">The Batch Manager in TensorRT-LLM</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="../advanced/inference-request.html">Inference Request</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="../advanced/inference-request.html#responses">Responses</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="../advanced/lora.html">Run gpt-2b + LoRA using GptManager / cpp runtime</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="../advanced/expert-parallelism.html">Expert Parallelism in TensorRT-LLM</a></li>
|
||
</ul>
|
||
<p class="caption" role="heading"><span class="caption-text">Performance</span></p>
|
||
<ul class="current">
|
||
<li class="toctree-l1"><a class="reference internal" href="perf-overview.html">Overview</a></li>
|
||
<li class="toctree-l1 current"><a class="current reference internal" href="#">Best Practices for Tuning the Performance of TensorRT-LLM</a><ul>
|
||
<li class="toctree-l2"><a class="reference internal" href="#how-to-measure-performance">How To Measure Performance?</a></li>
|
||
<li class="toctree-l2"><a class="reference internal" href="#build-options-to-optimize-the-performance-of-tensorrt-llm-models">Build Options to Optimize the Performance of TensorRT-LLM Models</a><ul>
|
||
<li class="toctree-l3"><a class="reference internal" href="#max-batch-size-max-seq-len-and-max-num-tokens"><code class="docutils literal notranslate"><span class="pre">max_batch_size</span></code>, <code class="docutils literal notranslate"><span class="pre">max_seq_len</span></code> and <code class="docutils literal notranslate"><span class="pre">max_num_tokens</span></code></a><ul>
|
||
<li class="toctree-l4"><a class="reference internal" href="#max-batch-size"><code class="docutils literal notranslate"><span class="pre">max_batch_size</span></code></a></li>
|
||
<li class="toctree-l4"><a class="reference internal" href="#max-seq-len"><code class="docutils literal notranslate"><span class="pre">max_seq_len</span></code></a></li>
|
||
<li class="toctree-l4"><a class="reference internal" href="#max-num-tokens"><code class="docutils literal notranslate"><span class="pre">max_num_tokens</span></code></a></li>
|
||
</ul>
|
||
</li>
|
||
<li class="toctree-l3"><a class="reference internal" href="#multiple-profiles">Multiple profiles</a></li>
|
||
<li class="toctree-l3"><a class="reference internal" href="#gpt-attention-plugin-and-context-fused-multi-head-attention">GPT Attention Plugin and Context Fused Multi-Head Attention</a><ul>
|
||
<li class="toctree-l4"><a class="reference internal" href="#fp8-context-fused-multi-head-attention">FP8 Context Fused Multi-Head Attention</a></li>
|
||
</ul>
|
||
</li>
|
||
<li class="toctree-l3"><a class="reference internal" href="#remove-input-padding">Remove Input Padding</a></li>
|
||
<li class="toctree-l3"><a class="reference internal" href="#paged-kv-cache">Paged KV Cache</a></li>
|
||
<li class="toctree-l3"><a class="reference internal" href="#in-flight-sequence-batching">In-flight Sequence Batching</a></li>
|
||
<li class="toctree-l3"><a class="reference internal" href="#multi-block-mode">Multi-Block Mode</a></li>
|
||
<li class="toctree-l3"><a class="reference internal" href="#custom-allreduce-plugin">Custom AllReduce Plugin</a></li>
|
||
<li class="toctree-l3"><a class="reference internal" href="#embedding-parallelism-embedding-sharing-and-look-up-plugin">Embedding Parallelism, Embedding Sharing, and Look-Up Plugin</a></li>
|
||
<li class="toctree-l3"><a class="reference internal" href="#horizontal-fusion-in-gated-mlp">Horizontal Fusion in Gated-MLP</a></li>
|
||
<li class="toctree-l3"><a class="reference internal" href="#gemm-swiglu-fusion-in-gated-mlp">GEMM + SwiGLU Fusion in Gated-MLP</a></li>
|
||
<li class="toctree-l3"><a class="reference internal" href="#gemm-plugin">GEMM Plugin</a><ul>
|
||
<li class="toctree-l4"><a class="reference internal" href="#fp8-gemm-plugin-for-small-batch-size-performance-optimization">FP8 GEMM Plugin for Small Batch Size Performance Optimization</a></li>
|
||
</ul>
|
||
</li>
|
||
<li class="toctree-l3"><a class="reference internal" href="#bert-attention-plugin-and-context-fused-multi-head-attention">BERT Attention Plugin and Context Fused Multi-Head Attention</a></li>
|
||
</ul>
|
||
</li>
|
||
<li class="toctree-l2"><a class="reference internal" href="#runtime-options-to-optimize-the-performance-of-tensorrt-llm-models">Runtime Options to Optimize the Performance of TensorRT-LLM Models</a><ul>
|
||
<li class="toctree-l3"><a class="reference internal" href="#gpt-model-type">GPT Model Type</a></li>
|
||
<li class="toctree-l3"><a class="reference internal" href="#max-tokens-in-paged-kv-cache-and-kv-cache-free-gpu-memory-fraction">Max Tokens in Paged KV Cache and KV Cache Free GPU Memory Fraction</a></li>
|
||
<li class="toctree-l3"><a class="reference internal" href="#batch-scheduler-policy">Batch Scheduler Policy</a></li>
|
||
<li class="toctree-l3"><a class="reference internal" href="#tensorrt-overlap">TensorRT Overlap</a></li>
|
||
<li class="toctree-l3"><a class="reference internal" href="#maximum-attention-window-size">Maximum Attention Window Size</a></li>
|
||
<li class="toctree-l3"><a class="reference internal" href="#chunked-context">Chunked Context</a></li>
|
||
</ul>
|
||
</li>
|
||
</ul>
|
||
</li>
|
||
<li class="toctree-l1"><a class="reference internal" href="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">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">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">Best Practices for Tuning the Performance of TensorRT-LLM</li>
|
||
<li class="wy-breadcrumbs-aside">
|
||
<a href="../_sources/performance/perf-best-practices.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="best-practices-for-tuning-the-performance-of-tensorrt-llm">
|
||
<span id="perf-best-practice"></span><h1>Best Practices for Tuning the Performance of TensorRT-LLM<a class="headerlink" href="#best-practices-for-tuning-the-performance-of-tensorrt-llm" title="Link to this heading"></a></h1>
|
||
<p>This document provides some best practices for tuning the performance of TensorRT-LLM.</p>
|
||
<section id="how-to-measure-performance">
|
||
<h2>How To Measure Performance?<a class="headerlink" href="#how-to-measure-performance" title="Link to this heading"></a></h2>
|
||
<p>TensorRT-LLM can be benchmarked using the included
|
||
<span class="xref myst">C++</span>
|
||
and
|
||
<span class="xref myst">Python</span> tools. However, it is <em>strongly</em>
|
||
recommended to use the C++ benchmarking tool. For detailed performance data and
|
||
the steps to reproduce those results, see
|
||
this <a class="reference internal" href="perf-overview.html"><span class="std std-doc">Document</span></a>.
|
||
The <a class="reference external" href="https://github.com/triton-inference-server/tensorrtllm_backend">TensorRT-LLM backend</a>
|
||
can also be used to measure the performance of TensorRT-LLM for online serving.</p>
|
||
</section>
|
||
<section id="build-options-to-optimize-the-performance-of-tensorrt-llm-models">
|
||
<h2>Build Options to Optimize the Performance of TensorRT-LLM Models<a class="headerlink" href="#build-options-to-optimize-the-performance-of-tensorrt-llm-models" title="Link to this heading"></a></h2>
|
||
<p>This part summarizes how to build engines to enhance the performance of the
|
||
runtime and, for some of them, decrease the engine build time.</p>
|
||
<p><em><strong>Note that some of those features and how to enable them may change in the future.</strong></em></p>
|
||
<section id="max-batch-size-max-seq-len-and-max-num-tokens">
|
||
<h3><code class="docutils literal notranslate"><span class="pre">max_batch_size</span></code>, <code class="docutils literal notranslate"><span class="pre">max_seq_len</span></code> and <code class="docutils literal notranslate"><span class="pre">max_num_tokens</span></code><a class="headerlink" href="#max-batch-size-max-seq-len-and-max-num-tokens" title="Link to this heading"></a></h3>
|
||
<p align="center">
|
||
<img src="https://github.com/NVIDIA/TensorRT-LLM/blob/rel/docs/source/media/max_bs_toks_len.svg?raw=true" alt="Explain `max_batch_size`, `max_seq_len` and `max_num_tokens`" width="30%" height="auto">
|
||
</p>
|
||
<p>Regarding the impacts of those three arguments to the GPU memory usage, please refer to <a class="reference internal" href="../reference/memory.html"><span class="std std-doc">memory.md</span></a></p>
|
||
<section id="max-batch-size">
|
||
<h4><code class="docutils literal notranslate"><span class="pre">max_batch_size</span></code><a class="headerlink" href="#max-batch-size" title="Link to this heading"></a></h4>
|
||
<p><code class="docutils literal notranslate"><span class="pre">max_batch_size</span></code> defines the maximum number of requests that the engine can handle.</p>
|
||
<p>It controls the maximum number of requests that can be scheduled at runtime.</p>
|
||
<p>Set high enough <code class="docutils literal notranslate"><span class="pre">max_batch_size</span></code> when building the engine so that it does not become the bottleneck of the throughput, and use runtime <code class="docutils literal notranslate"><span class="pre">max_batch_size</span></code> to tune it without re-building the engine if you want to get better user throughput or lower latency.</p>
|
||
</section>
|
||
<section id="max-seq-len">
|
||
<h4><code class="docutils literal notranslate"><span class="pre">max_seq_len</span></code><a class="headerlink" href="#max-seq-len" title="Link to this heading"></a></h4>
|
||
<p><code class="docutils literal notranslate"><span class="pre">max_seq_len</span></code> defines the maximum sequence length of single request</p>
|
||
<p>Starting from TensorRT-LLM v0.11, when <code class="docutils literal notranslate"><span class="pre">--remove_input_padding</span></code> and <code class="docutils literal notranslate"><span class="pre">--context_fmha</span></code> are enabled, <code class="docutils literal notranslate"><span class="pre">max_seq_len</span></code> can replace <code class="docutils literal notranslate"><span class="pre">max_input_len</span></code> and <code class="docutils literal notranslate"><span class="pre">max_output_len</span></code>, and is set to <code class="docutils literal notranslate"><span class="pre">max_position_embeddings</span></code> by default.</p>
|
||
<p>Use default <code class="docutils literal notranslate"><span class="pre">max_seq_len</span></code> (which is <code class="docutils literal notranslate"><span class="pre">max_position_embeddings</span></code>), no need to tune it unless you are very sure what max sequence lengths would be on your workloads. If the GPU memory is so limited that it cannot make sure even one request to reach <code class="docutils literal notranslate"><span class="pre">max_seq_len</span></code>, you’ll need to reduce it.</p>
|
||
</section>
|
||
<section id="max-num-tokens">
|
||
<h4><code class="docutils literal notranslate"><span class="pre">max_num_tokens</span></code><a class="headerlink" href="#max-num-tokens" title="Link to this heading"></a></h4>
|
||
<p><code class="docutils literal notranslate"><span class="pre">max_num_tokens</span></code> defines the maximum number of batched input tokens after padding is removed in each batch.</p>
|
||
<p><code class="docutils literal notranslate"><span class="pre">max_num_tokens</span></code> is set to 8192 by default starting from v0.11, you can tune it using the runtime <code class="docutils literal notranslate"><span class="pre">max_num_tokens</span></code> without re-buliding the engine. It is recommended to tune <code class="docutils literal notranslate"><span class="pre">--max_num_tokens</span></code> for better performance.</p>
|
||
<p>The maximum number of tokens equals will not take effects when input padding is
|
||
not removed. When input padding is removed (see <a class="reference internal" href="#remove-input-padding"><span class="xref myst">Remove Input
|
||
Padding</span></a>), the tokens from different sequences are
|
||
packed together and the maximum number of the tokens can be set to a different
|
||
(lower) value, which by default to be 8192.</p>
|
||
<p>There are two aspects that must be considered. Firstly, some input sequences
|
||
will be shorter than the maximum input length. Secondly, when in-flight
|
||
sequence batching is enabled, requests in context phase will be executed with
|
||
requests in generation phase. Those latter requests produce a lot fewer tokens
|
||
than <code class="docutils literal notranslate"><span class="pre">max_input_len</span></code> (at most, <code class="docutils literal notranslate"><span class="pre">beam_width</span></code> tokens).</p>
|
||
<p>Using a more realistic value for <code class="docutils literal notranslate"><span class="pre">max_num_tokens</span></code> allows TensorRT-LLM to
|
||
allocate more memory to store the KV cache and execute more requests together.
|
||
It leads to an increased efficiency.</p>
|
||
<p>Increasing <code class="docutils literal notranslate"><span class="pre">max_num_tokens</span></code> appropriately will be beneficial to performance.
|
||
When increasing <code class="docutils literal notranslate"><span class="pre">--max_num_tokens</span></code> to some point, GPU utilization will plateau,
|
||
going beyond that saturation point may hurt both first token latency as well as
|
||
total end-to-end latency.</p>
|
||
<p>See also <a class="reference internal" href="../advanced/gpt-attention.html#chunked-context"><span class="std std-ref">chunked context</span></a>.</p>
|
||
</section>
|
||
</section>
|
||
<section id="multiple-profiles">
|
||
<h3>Multiple profiles<a class="headerlink" href="#multiple-profiles" title="Link to this heading"></a></h3>
|
||
<p><code class="docutils literal notranslate"><span class="pre">--multiple_profiles</span></code> enables multiple TensorRT optimization profiles in the
|
||
built engines, it will benefits the performance especially when GEMM plugin is
|
||
disabled, because more optimization profiles help TensorRT have more chances to
|
||
select better kernels.</p>
|
||
<p>However, this feature will increase the engine build time.</p>
|
||
</section>
|
||
<section id="gpt-attention-plugin-and-context-fused-multi-head-attention">
|
||
<h3>GPT Attention Plugin and Context Fused Multi-Head Attention<a class="headerlink" href="#gpt-attention-plugin-and-context-fused-multi-head-attention" title="Link to this heading"></a></h3>
|
||
<p>The GPT attention plugin and fused multi-head attention kernel are enabled by
|
||
default. For the context phase, use the <code class="docutils literal notranslate"><span class="pre">--gpt_attention_plugin</span></code>
|
||
and <code class="docutils literal notranslate"><span class="pre">--context_fmha</span></code> arguments with <code class="docutils literal notranslate"><span class="pre">trtllm-build</span></code> to control.</p>
|
||
<p>The TensorRT-LLM GPT attention plugin uses efficient kernels and enables an
|
||
in-place update of the KV cache. It results in reduced memory consumption as
|
||
well as the removal of unneeded memory copy operations (compared with the
|
||
implementation that uses the <code class="docutils literal notranslate"><span class="pre">concat</span></code> operator to update the KV cache).</p>
|
||
<p>Enabling the fused multi-head attention, during the context phase, will trigger
|
||
a kernel that performs the MHA/MQA/GQA block using a single kernel, for more
|
||
details, see this <a class="reference internal" href="#../advanced/gpt-attention.html#context-phase"><span class="xref myst">Document</span></a>.</p>
|
||
<section id="fp8-context-fused-multi-head-attention">
|
||
<h4>FP8 Context Fused Multi-Head Attention<a class="headerlink" href="#fp8-context-fused-multi-head-attention" title="Link to this heading"></a></h4>
|
||
<p><code class="docutils literal notranslate"><span class="pre">--use_fp8_context_fmha</span></code> enables FP8 Context fused multi-head attention, which
|
||
is recommended to be enabled when fp8 quantization is used to improve the
|
||
performance. Note that only NVIDIA Hopper architecture is supported.</p>
|
||
</section>
|
||
</section>
|
||
<section id="remove-input-padding">
|
||
<h3>Remove Input Padding<a class="headerlink" href="#remove-input-padding" title="Link to this heading"></a></h3>
|
||
<p>The remove input padding feature is enabled by default, the <code class="docutils literal notranslate"><span class="pre">--remove_input_padding</span></code>
|
||
argument in <code class="docutils literal notranslate"><span class="pre">trtllm-build</span></code> is used to control it.</p>
|
||
<p>When input padding is removed, the different tokens are packed together. It
|
||
reduces both the amount of computations and memory consumption. For more details, see
|
||
this <a class="reference internal" href="../advanced/gpt-attention.html#padded-and-packed-tensors"><span class="std std-ref">Document</span></a>.</p>
|
||
</section>
|
||
<section id="paged-kv-cache">
|
||
<h3>Paged KV Cache<a class="headerlink" href="#paged-kv-cache" title="Link to this heading"></a></h3>
|
||
<p>Paged KV cache is enabled by default, the <code class="docutils literal notranslate"><span class="pre">--paged_kv_cache</span></code> argument in
|
||
<code class="docutils literal notranslate"><span class="pre">trtllm-build</span></code> is used to control it.</p>
|
||
<p>The paged KV cache helps manage memory for the KV cache more efficiently (see
|
||
this <a class="reference internal" href="#../advanced/gpt-attention.html#paged-kv-cache"><span class="xref myst">Document</span></a>). It usually leads to an
|
||
increase in the batch size and an improved efficiency.</p>
|
||
</section>
|
||
<section id="in-flight-sequence-batching">
|
||
<h3>In-flight Sequence Batching<a class="headerlink" href="#in-flight-sequence-batching" title="Link to this heading"></a></h3>
|
||
<p>In-flight sequence batching is enabled by default with <code class="docutils literal notranslate"><span class="pre">trtllm-build</span></code>,
|
||
which requires that the GPT attention plugin, input padding removal and paged KV
|
||
cache are all enabled together.</p>
|
||
<p>In-flight sequence batching schedules sequences in context phase together with
|
||
sequences in generation phase to increase efficiency and reduce latency, see
|
||
this <a class="reference internal" href="#../advanced/gpt-attention.html#in-flight-batching"><span class="xref myst">Document</span></a> for more details.</p>
|
||
</section>
|
||
<section id="multi-block-mode">
|
||
<h3>Multi-Block Mode<a class="headerlink" href="#multi-block-mode" title="Link to this heading"></a></h3>
|
||
<p>When the following conditions are met, it is recommended to try the
|
||
<code class="docutils literal notranslate"><span class="pre">--multi_block_mode</span></code> argument with <code class="docutils literal notranslate"><span class="pre">trtllm-build</span></code> and evaluate the impact on
|
||
performance:</p>
|
||
<ol class="arabic simple">
|
||
<li><p><code class="docutils literal notranslate"><span class="pre">input_seq_len</span></code> > 1024 (An empirically derived value that indicates that the
|
||
context length is long enough),</p></li>
|
||
<li><p><code class="docutils literal notranslate"><span class="pre">sequence_count</span></code> * <code class="docutils literal notranslate"><span class="pre">num_head</span></code> < <code class="docutils literal notranslate"><span class="pre">multiprocessor_count</span></code> / 2</p></li>
|
||
</ol>
|
||
<p>Multi-block mode can be beneficial when <code class="docutils literal notranslate"><span class="pre">batch_size</span> <span class="pre">*</span> <span class="pre">num_heads</span></code> is not large
|
||
enough to fully utilize the GPU (the number of CUDA thread blocks is low
|
||
compared to the number of streaming multiprocessors). Hence, the multi-block
|
||
mode is expected to reduce the latency of the multi-head attention kernel in
|
||
the generation phase. However, it requires the context length to be long enough
|
||
for the work performed by each CUDA thread block to remain sufficient for
|
||
efficiency.</p>
|
||
<p>Note that, the <code class="docutils literal notranslate"><span class="pre">--multi_block_mode</span></code> argument works more like a suggestion to the
|
||
runtime, hence it’s possible that multi-block is not used even when
|
||
<code class="docutils literal notranslate"><span class="pre">--multi_block_mode</span></code> argument is specified due to no performance gain, and it’s
|
||
also possible that multi-block is automatically used even when <code class="docutils literal notranslate"><span class="pre">--multi_block_mode</span></code>
|
||
argument is disabled.</p>
|
||
</section>
|
||
<section id="custom-allreduce-plugin">
|
||
<h3>Custom AllReduce Plugin<a class="headerlink" href="#custom-allreduce-plugin" title="Link to this heading"></a></h3>
|
||
<p>The custom AllReduce plugin activates a latency-optimized algorithm for
|
||
the AllReduce operation instead of the native NCCL operator. However, the
|
||
performance benefits may not be seen on PCIE-based systems.</p>
|
||
<p>Starting from v0.10, an automatic fallback mechanism is supported for allreduce plugin, so that native NCCL kernel is used when hardware requirements are not satisfied to get the best performance. Currently, the <code class="docutils literal notranslate"><span class="pre">--use_custom_all_reduce</span></code> argument works more like a suggestion
|
||
to the runtime, and will possibly be removed in the future releases.</p>
|
||
<p>On NVLink-based nodes, it is recommended to enable the custom AllReduce plugin
|
||
by using the <code class="docutils literal notranslate"><span class="pre">--use_custom_all_reduce</span></code> argument with <code class="docutils literal notranslate"><span class="pre">trtllm-build</span></code>. On PCIE-based
|
||
nodes, it is not recommended to enabled that plugin.</p>
|
||
<p>If you use <code class="docutils literal notranslate"><span class="pre">CUDA_VISIBLE_DEVICES</span></code> or <code class="docutils literal notranslate"><span class="pre">NVIDIA_VISIBLE_DEVICES</span></code>, please provide the full device list
|
||
instead of limiting to single device, otherwise custom all reduce will be disabled since its kernels
|
||
rely on P2P access to peer devices, which is not allowed when only a single device is visible.</p>
|
||
<p>In addition, there is an experimental feature called “Reduce Norm Fusion”
|
||
available to extend the custom AllReduce functionality. It can be enabled by
|
||
using the <code class="docutils literal notranslate"><span class="pre">--reduce_fusion</span> <span class="pre">enable</span></code> argument with <code class="docutils literal notranslate"><span class="pre">trtllm-build</span></code> when the
|
||
custom AllReduce is already enabled. This feature aims to fuse the ResidualAdd
|
||
and LayerNorm kernels after AllReduce into a single kernel, resulting in
|
||
improved end-to-end performance. Please note that currently, this feature is
|
||
only supported for the llama model. It is recommended to enable this feature when the batch size is small and the generation phase time is the dominant factor.</p>
|
||
</section>
|
||
<section id="embedding-parallelism-embedding-sharing-and-look-up-plugin">
|
||
<h3>Embedding Parallelism, Embedding Sharing, and Look-Up Plugin<a class="headerlink" href="#embedding-parallelism-embedding-sharing-and-look-up-plugin" title="Link to this heading"></a></h3>
|
||
<p>The embedding parallelism feature enables the sharding of the embedding table
|
||
across multiple GPUs, so that the memory usage could be reduced and the
|
||
throughput improved. The embedding sharing feature enables the sharing of the
|
||
embedding table between <code class="docutils literal notranslate"><span class="pre">look_up</span></code> and <code class="docutils literal notranslate"><span class="pre">lm_head</span></code> layers.</p>
|
||
<p>The look-up plugin implements the embedding sharing feature and is required to
|
||
enable the aforementioned features for now (until TensorRT native layers
|
||
support embedding sharing).</p>
|
||
<p>It is recommended to enable the embedding parallelism and sharing features to
|
||
improve throughput. However, the following conditions have to be satisfied:</p>
|
||
<ol class="arabic simple">
|
||
<li><p>The model shares the embedding table between <code class="docutils literal notranslate"><span class="pre">look_up</span></code> and <code class="docutils literal notranslate"><span class="pre">lm_head</span></code> layers,</p></li>
|
||
<li><p>Both look_up plugin and gemm plugin are enabled,</p></li>
|
||
<li><p>The sharding dimension of the embedding lookup table is set correctly.</p></li>
|
||
</ol>
|
||
<p>To enable the features, use the <code class="docutils literal notranslate"><span class="pre">--use_parallel_embedding</span></code>,
|
||
<code class="docutils literal notranslate"><span class="pre">--use_embedding_sharing</span></code>, <code class="docutils literal notranslate"><span class="pre">--use_lookup_plugin</span></code>, <code class="docutils literal notranslate"><span class="pre">--use_gemm_plugin</span></code>
|
||
arguments, and set correct dimension to <code class="docutils literal notranslate"><span class="pre">--embedding_sharding_dim</span></code> argument
|
||
with <code class="docutils literal notranslate"><span class="pre">trtllm-build</span></code>. See those
|
||
<a class="reference internal" href="#../../../examples/gpt#embedding-parallelism-and-sharing"><span class="xref myst">Examples</span></a>
|
||
for details.</p>
|
||
</section>
|
||
<section id="horizontal-fusion-in-gated-mlp">
|
||
<h3>Horizontal Fusion in Gated-MLP<a class="headerlink" href="#horizontal-fusion-in-gated-mlp" title="Link to this heading"></a></h3>
|
||
<p>Horizontal fusion in Gated-MLP combines two Matmul operations into a single one
|
||
followed by a separate SwiGLU kernel. However, for FP8 PTQ, the
|
||
downside is slight reduction of accuracy because one of the quantization scaling
|
||
factors are discarded.</p>
|
||
<p>If both model and batch sizes are large, it is recommended to enable the feature
|
||
by using the <code class="docutils literal notranslate"><span class="pre">--use_fused_mlp</span></code> argument with <code class="docutils literal notranslate"><span class="pre">trtllm-build</span></code>. When the workload
|
||
is very small, or if you’re using FP8 PTQ and the accuracy after enabling it
|
||
does not satisfy your requirement, it is not recommended to enable that feature.</p>
|
||
</section>
|
||
<section id="gemm-swiglu-fusion-in-gated-mlp">
|
||
<h3>GEMM + SwiGLU Fusion in Gated-MLP<a class="headerlink" href="#gemm-swiglu-fusion-in-gated-mlp" title="Link to this heading"></a></h3>
|
||
<p>GEMM + SwiGLU fusion in Gated-MLP combines two Matmul operations and one SwiGLU
|
||
operation into a single kernel. It only supports FP8 on Hopper now. For FP8 PTQ,
|
||
the downside is slight reduction of accuracy because one of the quantization
|
||
scaling factors are discarded.</p>
|
||
<p>If model is large and you are running it on Hopper with FP8 precision, it is
|
||
recommended to enable the feature by using the <code class="docutils literal notranslate"><span class="pre">--use_fused_mlp</span> <span class="pre">--gemm_swiglu_plugin</span> <span class="pre">fp8</span></code>
|
||
argument with <code class="docutils literal notranslate"><span class="pre">trtllm-build</span></code>. When the workload is very small, or the accuracy
|
||
after enabling it does not satisfy your requirement, it is not recommended to
|
||
enable that feature.</p>
|
||
</section>
|
||
<section id="gemm-plugin">
|
||
<h3>GEMM Plugin<a class="headerlink" href="#gemm-plugin" title="Link to this heading"></a></h3>
|
||
<p>The GEMM plugin utilizes NVIDIA cuBLASLt to perform GEMM operations. On FP16 and
|
||
BF16, it’s recommended to be enabled for better performance and smaller GPU
|
||
memory usage. On FP8, it’s recommended to be disabled.</p>
|
||
<section id="fp8-gemm-plugin-for-small-batch-size-performance-optimization">
|
||
<h4>FP8 GEMM Plugin for Small Batch Size Performance Optimization<a class="headerlink" href="#fp8-gemm-plugin-for-small-batch-size-performance-optimization" title="Link to this heading"></a></h4>
|
||
<p>FP8 gemm plugin is an experimental feature aimed to improve performance in
|
||
small-batch-size cases(e.g. BS<=4) and can be enabled by <code class="docutils literal notranslate"><span class="pre">--gemm_plugin</span> <span class="pre">fp8</span></code>
|
||
when building FP8 models. Although inputs with larger batch size can be correctly
|
||
inferenced, the performance may decrease as batch size grows. Therefore, this
|
||
feature is only recommended for latency reduction in small-batch-size scenarios
|
||
currently.</p>
|
||
</section>
|
||
</section>
|
||
<section id="bert-attention-plugin-and-context-fused-multi-head-attention">
|
||
<h3>BERT Attention Plugin and Context Fused Multi-Head Attention<a class="headerlink" href="#bert-attention-plugin-and-context-fused-multi-head-attention" title="Link to this heading"></a></h3>
|
||
<p>BERT attention plugin and context fused multi-head attention are both
|
||
recommended for the BERT model. They are enabled by default using the
|
||
<code class="docutils literal notranslate"><span class="pre">--bert_attention_plugin</span></code> and <code class="docutils literal notranslate"><span class="pre">--context_fmha</span></code> arguments with
|
||
<code class="docutils literal notranslate"><span class="pre">trtllm-build</span></code>.</p>
|
||
</section>
|
||
</section>
|
||
<section id="runtime-options-to-optimize-the-performance-of-tensorrt-llm-models">
|
||
<h2>Runtime Options to Optimize the Performance of TensorRT-LLM Models<a class="headerlink" href="#runtime-options-to-optimize-the-performance-of-tensorrt-llm-models" title="Link to this heading"></a></h2>
|
||
<p>This part summarizes the runtime configuration knobs that can be tweaked to
|
||
enhance the performance of already built engines. Note that currently the
|
||
configurations can be modified using the
|
||
<a class="reference internal" href="#../advanced/batch-manager.html#the-batch-manager-api"><span class="xref myst">Batch Manager API</span></a>
|
||
as well as the
|
||
<a class="reference external" href="https://github.com/triton-inference-server/tensorrtllm_backend">TensorRT-LLM backend</a>.</p>
|
||
<section id="gpt-model-type">
|
||
<h3>GPT Model Type<a class="headerlink" href="#gpt-model-type" title="Link to this heading"></a></h3>
|
||
<p>The GPT model type can be set to <code class="docutils literal notranslate"><span class="pre">V1</span></code>, <code class="docutils literal notranslate"><span class="pre">inflight_batching</span></code> and
|
||
<code class="docutils literal notranslate"><span class="pre">inflight_fused_batching</span></code>. It is recommended to use <code class="docutils literal notranslate"><span class="pre">inflight_fused_batching</span></code>
|
||
to increase throughput and reduce latency.</p>
|
||
</section>
|
||
<section id="max-tokens-in-paged-kv-cache-and-kv-cache-free-gpu-memory-fraction">
|
||
<h3>Max Tokens in Paged KV Cache and KV Cache Free GPU Memory Fraction<a class="headerlink" href="#max-tokens-in-paged-kv-cache-and-kv-cache-free-gpu-memory-fraction" title="Link to this heading"></a></h3>
|
||
<p>The <code class="docutils literal notranslate"><span class="pre">max_tokens_in_paged_kv_cache</span></code> and <code class="docutils literal notranslate"><span class="pre">kv_cache_free_gpu_mem_fraction</span></code>
|
||
parameters can be used to control the maximum number of tokens handled by the
|
||
KV cache manager. Setting them properly helps better control the amount of
|
||
available memory for the KV cache manager during inference. Keeping in mind
|
||
that increasing the amount of memory available to the KV cache manager tends to
|
||
translate to a higher achievable throughput.</p>
|
||
<p>The <code class="docutils literal notranslate"><span class="pre">max_tokens_in_paged_kv_cache</span></code> flag directly sets the maximum number of
|
||
tokens in the KV cache manager. When left unset, that value will be computed
|
||
based on the <code class="docutils literal notranslate"><span class="pre">kv_cache_free_gpu_mem_fraction</span></code> setting.</p>
|
||
<p>The <code class="docutils literal notranslate"><span class="pre">kv_cache_free_gpu_mem_fraction</span></code> is a floating-point number between <code class="docutils literal notranslate"><span class="pre">0.0</span></code>
|
||
and <code class="docutils literal notranslate"><span class="pre">1.0</span></code> that indicates the maximum fraction of GPU memory (after loading the
|
||
model) that will be used for the KV cache. The default value is <code class="docutils literal notranslate"><span class="pre">0.90</span></code> and
|
||
means that 90% of the free GPU memory will be used to save tokens in the KV
|
||
cache. Based on that value, TensorRT-LLM can determine the maximum number of
|
||
tokens in the KV cache manager.</p>
|
||
<p>When both parameters are set, the maximum number of tokens in the KV cache
|
||
manager will be set to the smaller value between <code class="docutils literal notranslate"><span class="pre">max_tokens_in_paged_kv_cache</span></code>
|
||
and the value computed from the amount of memory available for the KV cache.</p>
|
||
<p>Unless users clearly know the maximum number of tokens in the KV cache needed
|
||
by the model, it is recommended to leave <code class="docutils literal notranslate"><span class="pre">max_tokens_in_paged_kv_cache</span></code> unset.
|
||
For <code class="docutils literal notranslate"><span class="pre">kv_cache_free_gpu_mem_fraction</span></code>, if no other programs are executed on the
|
||
same GPU, it is recommended to test with a as high value as <code class="docutils literal notranslate"><span class="pre">0.95</span></code> to target a
|
||
high throughput. Note that the <code class="docutils literal notranslate"><span class="pre">kv_cache_free_gpu_mem_fraction</span></code> parameter
|
||
cannot be set to <code class="docutils literal notranslate"><span class="pre">1.0</span></code> because some amount of memory has to be reserved for
|
||
inputs and outputs.</p>
|
||
</section>
|
||
<section id="batch-scheduler-policy">
|
||
<h3>Batch Scheduler Policy<a class="headerlink" href="#batch-scheduler-policy" title="Link to this heading"></a></h3>
|
||
<p>There currently are two batch scheduler policies: <code class="docutils literal notranslate"><span class="pre">MAX_UTILIZATION</span></code> and
|
||
<code class="docutils literal notranslate"><span class="pre">GUARANTEED_NO_EVICT</span></code>.</p>
|
||
<p>As explained in the <a class="reference internal" href="#../advanced/batch-manager.html#gptmanager-design"><span class="xref myst">GPT Manager Design</span></a>
|
||
section, the scheduling policy can be set to <code class="docutils literal notranslate"><span class="pre">MAX_UTILIZATION</span></code> to pack as many
|
||
requests as possible at each iteration of the forward loop, when in-flight
|
||
sequence batching is enabled. It maximizes the utilization of the GPUs by
|
||
aggressively scheduling requests at the risk of having to pause requests if the
|
||
KV cache size limit is reached.</p>
|
||
<p>For a more conservative approach with respect to the KV cache limitations in
|
||
terms of memory allocation, <code class="docutils literal notranslate"><span class="pre">CapacitySchedulerPolicy</span></code> should be set to
|
||
<code class="docutils literal notranslate"><span class="pre">GUARANTEED_NO_EVICT</span></code> to guarantee that a started request is never paused.</p>
|
||
<p>If the goal is to maximizes the throughput, users should try <code class="docutils literal notranslate"><span class="pre">MAX_UTILIZATION</span></code>.
|
||
However, they need to keep in mind that it may have a negative impact on
|
||
latency if requests have to be paused.</p>
|
||
</section>
|
||
<section id="tensorrt-overlap">
|
||
<h3>TensorRT Overlap<a class="headerlink" href="#tensorrt-overlap" title="Link to this heading"></a></h3>
|
||
<p><em><strong>Note that this option is now deprecated and only available with the GptManager API.</strong></em></p>
|
||
<p>This option allowed to partition available requests into 2
|
||
micro-batches that could be run concurrently and thereby allowed TensorRT-LLM to hide
|
||
some exposed CPU runtime. However, optimization work has been done to reduce this
|
||
exposed CPU runtime and it has been found that the concurrent execution
|
||
of micro-batches did not provide additional benefits in terms of throughput,
|
||
and in most cases, was hurting latency.</p>
|
||
</section>
|
||
<section id="maximum-attention-window-size">
|
||
<h3>Maximum Attention Window Size<a class="headerlink" href="#maximum-attention-window-size" title="Link to this heading"></a></h3>
|
||
<p>The <code class="docutils literal notranslate"><span class="pre">max_attention_window_size</span></code> flag sets the maximum number of tokens that are
|
||
attended to in order to generate one token when using techniques like sliding window
|
||
attention. See this
|
||
<a class="reference internal" href="../advanced/gpt-attention.html#sliding-window-attention-cyclic-rolling-buffer-kv-cache"><span class="std std-ref">Document</span></a>
|
||
for more details. It defaults to the maximum sequence length
|
||
(<code class="docutils literal notranslate"><span class="pre">max_input_length</span> <span class="pre">+</span> <span class="pre">max_output_length</span></code> when building the engine), which means
|
||
that the feature is disabled by default.</p>
|
||
<p>When set to a smaller value than <code class="docutils literal notranslate"><span class="pre">max_input_length</span> <span class="pre">+</span> <span class="pre">max_output_length</span></code> (during
|
||
engine build), only the KV cache of the last <code class="docutils literal notranslate"><span class="pre">max_attention_window_size</span></code> tokens
|
||
will be stored. If the input sequence length at runtime exceeds the
|
||
<code class="docutils literal notranslate"><span class="pre">max_attention_window_size</span></code> value, the accuracy may start dropping, but the
|
||
runtime performance will be better (due to the reduction in terms of
|
||
computations and GPU memory allocation). Users can modify that value to
|
||
increase runtime performance at the expense of reduced accuracy.</p>
|
||
</section>
|
||
<section id="chunked-context">
|
||
<h3>Chunked Context<a class="headerlink" href="#chunked-context" title="Link to this heading"></a></h3>
|
||
<p>Turning on context chunking by specifying <code class="docutils literal notranslate"><span class="pre">enable_chunked_context</span></code> in
|
||
<code class="docutils literal notranslate"><span class="pre">TrtGptModelOptionalParams</span></code> will increase the chance of batch processing between
|
||
the context and the generation phase, thereby balancing the calculation amount
|
||
of each iteration and increasing throughput. When this function is turned on,
|
||
different performance can be obtained by adjusting <code class="docutils literal notranslate"><span class="pre">max_num_tokens</span></code>. Usually
|
||
its recommended value is <code class="docutils literal notranslate"><span class="pre">N</span> <span class="pre">*</span> <span class="pre">tokens_per_block</span></code>, and <code class="docutils literal notranslate"><span class="pre">N</span></code> is an integer that is
|
||
recommended to start from <code class="docutils literal notranslate"><span class="pre">1</span></code> and increase until the best performance is achieved.</p>
|
||
</section>
|
||
</section>
|
||
</section>
|
||
|
||
|
||
</div>
|
||
</div>
|
||
<footer><div class="rst-footer-buttons" role="navigation" aria-label="Footer">
|
||
<a href="perf-overview.html" class="btn btn-neutral float-left" title="Overview" accesskey="p" rel="prev"><span class="fa fa-arrow-circle-left" aria-hidden="true"></span> Previous</a>
|
||
<a href="perf-analysis.html" class="btn btn-neutral float-right" title="Performance Analysis" 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 0x7f8e35b9d990>
|
||
|
||
<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> |