kanshan/TensorRT-LLMs
1
0
Fork 0
mirror of https://github.com/NVIDIA/TensorRT-LLM.git synced 2026-01-26 05:32:57 +08:00
Code Issues Actions 1 Packages Projects Releases Wiki Activity
130807da9f
TensorRT-LLMs/performance/perf-benchmarking.html
Iman Tabrizian 130807da9f Update docs (#2732) 
* Update docs

* Update windows install version

Update gh pages (#2741)

update gh pages (#2743)

gh pages update (#2746)

Update gh-pages (#2764)

Update
2025-02-11 02:56:32 +00:00

957 lines
106 KiB
HTML
Raw Blame History

This file contains ambiguous Unicode characters

This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

<!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>TensorRT-LLM Benchmarking &mdash; 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="Performance Tuning Guide" href="performance-tuning-guide/index.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="../key-features.html">Key Features</a></li>
<li class="toctree-l1"><a class="reference internal" href="../torch.html">PyTorch Backend</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>
<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/executor.html">Executor API</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/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>
<li class="toctree-l1"><a class="reference internal" href="../advanced/kv-cache-reuse.html">KV cache reuse</a></li>
<li class="toctree-l1"><a class="reference internal" href="../advanced/speculative-decoding.html">Speculative Sampling</a></li>
<li class="toctree-l1"><a class="reference internal" href="../advanced/disaggregated-service.html">Disaggregated-Service (experimental)</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="#">Benchmarking</a><ul>
<li class="toctree-l2"><a class="reference internal" href="#throughput-benchmarking">Throughput Benchmarking</a><ul>
<li class="toctree-l3"><a class="reference internal" href="#limitations-and-caveats">Limitations and Caveats</a><ul>
<li class="toctree-l4"><a class="reference internal" href="#validated-networks-for-benchmarking">Validated Networks for Benchmarking</a></li>
<li class="toctree-l4"><a class="reference internal" href="#supported-quantization-modes">Supported Quantization Modes</a></li>
</ul>
</li>
<li class="toctree-l3"><a class="reference internal" href="#quickstart">Quickstart</a></li>
<li class="toctree-l3"><a class="reference internal" href="#workflow">Workflow</a><ul>
<li class="toctree-l4"><a class="reference internal" href="#preparing-a-dataset">Preparing a Dataset</a></li>
</ul>
</li>
<li class="toctree-l3"><a class="reference internal" href="#building-a-benchmark-engine">Building a Benchmark Engine</a><ul>
<li class="toctree-l4"><a class="reference internal" href="#default-build-behavior">Default Build Behavior</a></li>
<li class="toctree-l4"><a class="reference internal" href="#other-build-modes">Other Build Modes</a></li>
<li class="toctree-l4"><a class="reference internal" href="#parallelism-mapping-support">Parallelism Mapping Support</a></li>
<li class="toctree-l4"><a class="reference internal" href="#example-of-build-subcommand-output">Example of Build Subcommand Output:</a></li>
</ul>
</li>
<li class="toctree-l3"><a class="reference internal" href="#max-throughput-benchmark">Max Throughput Benchmark</a><ul>
<li class="toctree-l4"><a class="reference internal" href="#how-the-benchmarker-works">How the Benchmarker Works</a></li>
</ul>
</li>
<li class="toctree-l3"><a class="reference internal" href="#running-with-the-pytorch-workflow">Running with the PyTorch Workflow</a><ul>
<li class="toctree-l4"><a class="reference internal" href="#quantization-in-the-pytorch-flow">Quantization in the PyTorch Flow</a></li>
</ul>
</li>
</ul>
</li>
<li class="toctree-l2"><a class="reference internal" href="#low-latency-benchmark">Low Latency Benchmark</a><ul>
<li class="toctree-l3"><a class="reference internal" href="#low-latency-tensorrt-llm-engine-for-llama-3-70b">Low Latency TensorRT-LLM Engine for Llama-3 70B</a><ul>
<li class="toctree-l4"><a class="reference internal" href="#benchmarking-a-non-medusa-low-latency-engine">Benchmarking a non-Medusa Low Latency Engine</a></li>
</ul>
</li>
<li class="toctree-l3"><a class="reference internal" href="#building-a-medusa-low-latency-engine">Building a Medusa Low-Latency Engine</a></li>
</ul>
</li>
<li class="toctree-l2"><a class="reference internal" href="#summary">Summary</a></li>
</ul>
</li>
<li class="toctree-l1"><a class="reference internal" href="performance-tuning-guide/index.html">Performance Tuning Guide</a></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">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">TensorRT-LLM Benchmarking</li>
<li class="wy-breadcrumbs-aside">
<a href="../_sources/performance/perf-benchmarking.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="tensorrt-llm-benchmarking">
<span id="perf-benchmarking"></span><h1>TensorRT-LLM Benchmarking<a class="headerlink" href="#tensorrt-llm-benchmarking" title="Link to this heading"></a></h1>
<div class="admonition important">
<p class="admonition-title">Important</p>
<p>This benchmarking suite is a work in progress.
Expect breaking API changes.</p>
</div>
<p>TensorRT-LLM provides the <code class="docutils literal notranslate"><span class="pre">trtllm-bench</span></code> CLI, a packaged benchmarking utility that aims to make it
easier for users to reproduce our officially published <a class="reference internal" href="perf-overview.html#throughput-measurements"><span class="std std-ref">performance overiew</span></a>. <code class="docutils literal notranslate"><span class="pre">trtllm-bench</span></code> provides the follows:</p>
<ul class="simple">
<li><p>A streamlined way to build tuned engines for benchmarking for a variety of models and platforms.</p></li>
<li><p>An entirely Python workflow for benchmarking.</p></li>
<li><p>Ability to benchmark various flows and features within TensorRT-LLM.</p></li>
</ul>
<p><code class="docutils literal notranslate"><span class="pre">trtllm-bench</span></code> executes all benchmarks using [in-flight batching] – for more information see
the <a class="reference internal" href="../advanced/gpt-attention.html#in-flight-batching"><span class="std std-ref">this section</span></a> that describes the concept
in further detail.</p>
<section id="throughput-benchmarking">
<h2>Throughput Benchmarking<a class="headerlink" href="#throughput-benchmarking" title="Link to this heading"></a></h2>
<section id="limitations-and-caveats">
<h3>Limitations and Caveats<a class="headerlink" href="#limitations-and-caveats" title="Link to this heading"></a></h3>
<section id="validated-networks-for-benchmarking">
<h4>Validated Networks for Benchmarking<a class="headerlink" href="#validated-networks-for-benchmarking" title="Link to this heading"></a></h4>
<p>While <code class="docutils literal notranslate"><span class="pre">trtllm-bench</span></code> should be able to run any network that TensorRT-LLM supports, the following are the list
that have been validated extensively and is the same listing as seen on the
<a class="reference internal" href="perf-overview.html"><span class="std std-doc">Performance Overview</span></a> page.</p>
<ul class="simple">
<li><p><a class="reference external" href="https://huggingface.co/meta-llama/Llama-2-7b-hf">meta-llama/Llama-2-7b-hf</a></p></li>
<li><p><a class="reference external" href="https://huggingface.co/meta-llama/Llama-2-70b-hf">meta-llama/Llama-2-70b-hf</a></p></li>
<li><p><a class="reference external" href="https://huggingface.co/tiiuae/falcon-180B">tiiuae/falcon-180B</a></p></li>
<li><p><a class="reference external" href="https://huggingface.co/EleutherAI/gpt-j-6b">EleutherAI/gpt-j-6b</a></p></li>
<li><p><a class="reference external" href="https://huggingface.co/meta-llama/Meta-Llama-3-8B">meta-llama/Meta-Llama-3-8B</a></p></li>
<li><p><a class="reference external" href="https://huggingface.co/meta-llama/Llama-3.1-8B">meta-llama/Llama-3.1-8B</a></p></li>
<li><p><a class="reference external" href="https://huggingface.co/meta-llama/Meta-Llama-3-70B">meta-llama/Meta-Llama-3-70B</a></p></li>
<li><p><a class="reference external" href="https://huggingface.co/meta-llama/Llama-3.1-70B">meta-llama/Llama-3.1-70B</a></p></li>
<li><p><a class="reference external" href="https://huggingface.co/meta-llama/Llama-3.1-405B">meta-llama/Llama-3.1-405B</a></p></li>
<li><p><a class="reference external" href="https://huggingface.co/mistralai/Mixtral-8x7B-v0.1">mistralai/Mixtral-8x7B-v0.1</a></p></li>
<li><p><a class="reference external" href="https://huggingface.co/mistralai/Mistral-7B-v0.1">mistralai/Mistral-7B-v0.1</a></p></li>
<li><p><a class="reference external" href="https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct">meta-llama/Llama-3.1-8B-Instruct</a></p></li>
<li><p><a class="reference external" href="https://huggingface.co/meta-llama/Llama-3.1-70B-Instruct">meta-llama/Llama-3.1-70B-Instruct</a></p></li>
<li><p><a class="reference external" href="https://huggingface.co/meta-llama/Llama-3.1-405B-Instruct">meta-llama/Llama-3.1-405B-Instruct</a></p></li>
<li><p><a class="reference external" href="https://huggingface.co/mistralai/Mixtral-8x7B-v0.1-Instruct">mistralai/Mixtral-8x7B-v0.1-Instruct</a></p></li>
</ul>
<div class="admonition tip">
<p class="admonition-title">Tip</p>
<p><code class="docutils literal notranslate"><span class="pre">trtllm-bench</span></code> can automatically download the model from Hugging Face Model Hub.
Export your token in the <code class="docutils literal notranslate"><span class="pre">HF_TOKEN</span></code> environment variable.</p>
</div>
</section>
<section id="supported-quantization-modes">
<h4>Supported Quantization Modes<a class="headerlink" href="#supported-quantization-modes" title="Link to this heading"></a></h4>
<p><code class="docutils literal notranslate"><span class="pre">trtllm-bench</span></code> supports the following quantization modes:</p>
<ul class="simple">
<li><p>None (no quantization applied)</p></li>
<li><p><code class="docutils literal notranslate"><span class="pre">FP8</span></code></p></li>
<li><p><code class="docutils literal notranslate"><span class="pre">NVFP4</span></code></p></li>
</ul>
<p>For more information about quantization, refer to <a class="reference internal" href="../reference/precision.html"><span class="std std-doc">Numerical Precision</span></a> and
the <a class="reference internal" href="../reference/precision.html#support-matrix"><span class="std std-ref">support matrix</span></a> of the supported quantization methods for each network.</p>
<div class="admonition tip">
<p class="admonition-title">Tip</p>
<p>Although TensorRT-LLM supports more quantization modes than listed above, <code class="docutils literal notranslate"><span class="pre">trtllm-bench</span></code> currently only configures for
a smaller subset.</p>
</div>
</section>
</section>
<section id="quickstart">
<h3>Quickstart<a class="headerlink" href="#quickstart" title="Link to this heading"></a></h3>
<p>This quick start focuses on running a short max throughput benchmark on
<code class="docutils literal notranslate"><span class="pre">meta-llama/Llama-3.1-8B</span></code> on a synthetic dataset with a uniform distribution of prompts with ISL:OSL
of 128:128.
To run the benchmark from start to finish, run the following commands:</p>
<div class="highlight-shell notranslate"><div class="highlight"><pre><span></span>python<span class="w"> </span>benchmarks/cpp/prepare_dataset.py<span class="w"> </span>--stdout<span class="w"> </span>--tokenizer<span class="w"> </span>meta-llama/Llama-3.1-8B<span class="w"> </span>token-norm-dist<span class="w"> </span>--input-mean<span class="w"> </span><span class="m">128</span><span class="w"> </span>--output-mean<span class="w"> </span><span class="m">128</span><span class="w"> </span>--input-stdev<span class="w"> </span><span class="m">0</span><span class="w"> </span>--output-stdev<span class="w"> </span><span class="m">0</span><span class="w"> </span>--num-requests<span class="w"> </span><span class="m">3000</span><span class="w"> </span>&gt;<span class="w"> </span>/tmp/synthetic_128_128.txt
trtllm-bench<span class="w"> </span>--model<span class="w"> </span>meta-llama/Llama-3.1-8B<span class="w"> </span>build<span class="w"> </span>--dataset<span class="w"> </span>/tmp/synthetic_128_128.txt<span class="w"> </span>--quantization<span class="w"> </span>FP8
trtllm-bench<span class="w"> </span>--model<span class="w"> </span>meta-llama/Llama-3.1-8B<span class="w"> </span>throughput<span class="w"> </span>--dataset<span class="w"> </span>/tmp/synthetic_128_128.txt<span class="w"> </span>--engine_dir<span class="w"> </span>/tmp/meta-llama/Llama-3.1-8B/tp_1_pp_1
</pre></div>
</div>
<p>After the benchmark completes, <code class="docutils literal notranslate"><span class="pre">trtllm-bench</span></code> prints a summary with summary metrics.</p>
<div class="highlight-shell notranslate"><div class="highlight"><pre><span></span><span class="o">===========================================================</span>
<span class="o">=</span><span class="w"> </span>ENGINE<span class="w"> </span><span class="nv">DETAILS</span>
<span class="o">===========================================================</span>
Model:<span class="w"> </span>meta-llama/Llama-3.1-8B
Engine<span class="w"> </span>Directory:<span class="w"> </span>/tmp/meta-llama/Llama-3.1-8B/tp_1_pp_1
TensorRT-LLM<span class="w"> </span>Version:<span class="w"> </span><span class="m">0</span>.17.0
Dtype:<span class="w"> </span>bfloat16
KV<span class="w"> </span>Cache<span class="w"> </span>Dtype:<span class="w"> </span>FP8
Quantization:<span class="w"> </span>FP8
Max<span class="w"> </span>Input<span class="w"> </span>Length:<span class="w"> </span><span class="m">256</span>
Max<span class="w"> </span>Sequence<span class="w"> </span>Length:<span class="w"> </span><span class="nv">256</span>
<span class="o">===========================================================</span>
<span class="o">=</span><span class="w"> </span>WORLD<span class="w"> </span>+<span class="w"> </span>RUNTIME<span class="w"> </span><span class="nv">INFORMATION</span>
<span class="o">===========================================================</span>
TP<span class="w"> </span>Size:<span class="w"> </span><span class="m">1</span>
PP<span class="w"> </span>Size:<span class="w"> </span><span class="m">1</span>
Max<span class="w"> </span>Runtime<span class="w"> </span>Batch<span class="w"> </span>Size:<span class="w"> </span><span class="m">4096</span>
Max<span class="w"> </span>Runtime<span class="w"> </span>Tokens:<span class="w"> </span><span class="m">8192</span>
Scheduling<span class="w"> </span>Policy:<span class="w"> </span>Guaranteed<span class="w"> </span>No<span class="w"> </span>Evict
KV<span class="w"> </span>Memory<span class="w"> </span>Percentage:<span class="w"> </span><span class="m">90</span>.00%
Issue<span class="w"> </span>Rate<span class="w"> </span><span class="o">(</span>req/sec<span class="o">)</span>:<span class="w"> </span><span class="m">5</span>.0689E+14
<span class="o">===========================================================</span>
<span class="o">=</span><span class="w"> </span>PERFORMANCE<span class="w"> </span><span class="nv">OVERVIEW</span>
<span class="o">===========================================================</span>
Number<span class="w"> </span>of<span class="w"> </span>requests:<span class="w"> </span><span class="m">3000</span>
Average<span class="w"> </span>Input<span class="w"> </span>Length<span class="w"> </span><span class="o">(</span>tokens<span class="o">)</span>:<span class="w"> </span><span class="m">128</span>.0000
Average<span class="w"> </span>Output<span class="w"> </span>Length<span class="w"> </span><span class="o">(</span>tokens<span class="o">)</span>:<span class="w"> </span><span class="m">128</span>.0000
Token<span class="w"> </span>Throughput<span class="w"> </span><span class="o">(</span>tokens/sec<span class="o">)</span>:<span class="w"> </span><span class="m">28390</span>.4265
Request<span class="w"> </span>Throughput<span class="w"> </span><span class="o">(</span>req/sec<span class="o">)</span>:<span class="w"> </span><span class="m">221</span>.8002
Total<span class="w"> </span>Latency<span class="w"> </span><span class="o">(</span>ms<span class="o">)</span>:<span class="w"> </span><span class="m">13525</span>.6862
<span class="o">===========================================================</span>
</pre></div>
</div>
</section>
<section id="workflow">
<h3>Workflow<a class="headerlink" href="#workflow" title="Link to this heading"></a></h3>
<p>The workflow for <code class="docutils literal notranslate"><span class="pre">trtllm-bench</span></code> is composed of the following steps:</p>
<ol class="arabic simple">
<li><p>Prepare a dataset to drive the inflight batching benchmark.</p></li>
<li><p>Build a benchmark engine using <code class="docutils literal notranslate"><span class="pre">trtllm-bench</span> <span class="pre">build</span></code> subcommand (not required for <a class="reference internal" href="#running-with-the-pytorch-workflow">PyTorch flow</a>).</p></li>
<li><p>Run the max throughput benchmark using the <code class="docutils literal notranslate"><span class="pre">trtllm-bench</span> <span class="pre">throughput</span></code> subcommand or low latency benchmark using the <code class="docutils literal notranslate"><span class="pre">trtllm-bench</span> <span class="pre">latency</span></code> subcommand.</p></li>
</ol>
<section id="preparing-a-dataset">
<h4>Preparing a Dataset<a class="headerlink" href="#preparing-a-dataset" title="Link to this heading"></a></h4>
<p>The throughput benchmark utilizes a fixed JSON schema to specify requests. The schema is defined as follows:</p>
<table class="docutils align-default">
<thead>
<tr class="row-odd"><th class="head text-left"><p>Key</p></th>
<th class="head text-center"><p>Required</p></th>
<th class="head text-center"><p>Type</p></th>
<th class="head text-left"><p>Description</p></th>
</tr>
</thead>
<tbody>
<tr class="row-even"><td class="text-left"><p><code class="docutils literal notranslate"><span class="pre">task_id</span></code></p></td>
<td class="text-center"><p>Y</p></td>
<td class="text-center"><p>String</p></td>
<td class="text-left"><p>Unique identifier for the request.</p></td>
</tr>
<tr class="row-odd"><td class="text-left"><p><code class="docutils literal notranslate"><span class="pre">prompt</span></code></p></td>
<td class="text-center"><p>N*</p></td>
<td class="text-center"><p>String</p></td>
<td class="text-left"><p>Input text for a generation request.</p></td>
</tr>
<tr class="row-even"><td class="text-left"><p><code class="docutils literal notranslate"><span class="pre">input_ids</span></code></p></td>
<td class="text-center"><p>Y*</p></td>
<td class="text-center"><p>List[Integer]</p></td>
<td class="text-left"><p>List of logits that make up the request prompt.</p></td>
</tr>
<tr class="row-odd"><td class="text-left"><p><code class="docutils literal notranslate"><span class="pre">output_tokens</span></code></p></td>
<td class="text-center"><p>Y</p></td>
<td class="text-center"><p>Integer</p></td>
<td class="text-left"><p>Number of generated tokens for this request.</p></td>
</tr>
</tbody>
</table>
<div class="admonition tip">
<p class="admonition-title">Tip</p>
<p>* Specifying <code class="docutils literal notranslate"><span class="pre">prompt</span></code> or <code class="docutils literal notranslate"><span class="pre">input_ids</span></code> is required. However, you can not have both prompts and logits (<code class="docutils literal notranslate"><span class="pre">input_ids</span></code>)
defined at the same time. If you specify <code class="docutils literal notranslate"><span class="pre">input_ids</span></code>, the <code class="docutils literal notranslate"><span class="pre">prompt</span></code> entry is ignored for request generation.</p>
</div>
<p>Refer to the following examples of valid entries for the benchmark:</p>
<ul>
<li><p>Entries with a human-readable prompt and no logits.</p>
<div class="highlight-json notranslate"><div class="highlight"><pre><span></span><span class="p">{</span><span class="nt">&quot;task_id&quot;</span><span class="p">:</span><span class="w"> </span><span class="mi">1</span><span class="p">,</span><span class="w"> </span><span class="nt">&quot;prompt&quot;</span><span class="p">:</span><span class="w"> </span><span class="s2">&quot;Generate an infinite response to the following: This is the song that never ends, it goes on and on my friend.&quot;</span><span class="p">,</span><span class="w"> </span><span class="nt">&quot;output_tokens&quot;</span><span class="p">:</span><span class="w"> </span><span class="mi">1000</span><span class="p">}</span>
<span class="p">{</span><span class="nt">&quot;task_id&quot;</span><span class="p">:</span><span class="w"> </span><span class="mi">2</span><span class="p">,</span><span class="w"> </span><span class="nt">&quot;prompt&quot;</span><span class="p">:</span><span class="w"> </span><span class="s2">&quot;Generate an infinite response to the following: Na, na, na, na&quot;</span><span class="p">,</span><span class="w"> </span><span class="nt">&quot;output_tokens&quot;</span><span class="p">:</span><span class="w"> </span><span class="mi">1000</span><span class="p">}</span>
</pre></div>
</div>
</li>
<li><p>Entries which contain logits.</p>
<div class="highlight-json notranslate"><div class="highlight"><pre><span></span><span class="p">{</span><span class="nt">&quot;task_id&quot;</span><span class="p">:</span><span class="mi">0</span><span class="p">,</span><span class="nt">&quot;input_ids&quot;</span><span class="p">:[</span><span class="mi">863</span><span class="p">,</span><span class="mi">22056</span><span class="p">,</span><span class="mi">25603</span><span class="p">,</span><span class="mi">11943</span><span class="p">,</span><span class="mi">8932</span><span class="p">,</span><span class="mi">13195</span><span class="p">,</span><span class="mi">3132</span><span class="p">,</span><span class="mi">25032</span><span class="p">,</span><span class="mi">21747</span><span class="p">,</span><span class="mi">22213</span><span class="p">],</span><span class="nt">&quot;output_tokens&quot;</span><span class="p">:</span><span class="mi">128</span><span class="p">}</span>
<span class="p">{</span><span class="nt">&quot;task_id&quot;</span><span class="p">:</span><span class="mi">1</span><span class="p">,</span><span class="nt">&quot;input_ids&quot;</span><span class="p">:[</span><span class="mi">14480</span><span class="p">,</span><span class="mi">13598</span><span class="p">,</span><span class="mi">15585</span><span class="p">,</span><span class="mi">6591</span><span class="p">,</span><span class="mi">1252</span><span class="p">,</span><span class="mi">8259</span><span class="p">,</span><span class="mi">30990</span><span class="p">,</span><span class="mi">26778</span><span class="p">,</span><span class="mi">7063</span><span class="p">,</span><span class="mi">30065</span><span class="p">,</span><span class="mi">21764</span><span class="p">,</span><span class="mi">11023</span><span class="p">,</span><span class="mi">1418</span><span class="p">],</span><span class="nt">&quot;output_tokens&quot;</span><span class="p">:</span><span class="mi">128</span><span class="p">}</span>
</pre></div>
</div>
</li>
</ul>
<div class="admonition tip">
<p class="admonition-title">Tip</p>
<p>Specify each entry on one line.
To simplify passing the data, a complete JSON entry is on each line so that the benchmarker
can simply read a line and assume a complete entry. When creating a dataset, be sure that a complete
JSON entry is on every line.</p>
</div>
<p>In order to prepare a synthetic dataset, you can use the provided script in the <code class="docutils literal notranslate"><span class="pre">benchmarks/cpp</span></code>
directory. For example, to generate a synthetic dataset of 1000 requests with a uniform ISL/OSL of
128/128 for <a class="reference external" href="https://huggingface.co/meta-llama/Llama-3.1-8B">meta-llama/Llama-3.1-8B</a>, run:</p>
<div class="highlight-shell notranslate"><div class="highlight"><pre><span></span>benchmarks/cpp/prepare_dataset.py<span class="w"> </span>--stdout<span class="w"> </span>--tokenizer<span class="w"> </span>meta-llama/Llama-3.1-8B<span class="w"> </span>token-norm-dist<span class="w"> </span>--input-mean<span class="w"> </span><span class="m">128</span><span class="w"> </span>--output-mean<span class="w"> </span><span class="m">128</span><span class="w"> </span>--input-stdev<span class="w"> </span><span class="m">0</span><span class="w"> </span>--output-stdev<span class="w"> </span><span class="m">0</span><span class="w"> </span>--num-requests<span class="w"> </span><span class="m">1000</span><span class="w"> </span>&gt;<span class="w"> </span>/tmp/synthetic_128_128.txt
</pre></div>
</div>
</section>
</section>
<section id="building-a-benchmark-engine">
<h3>Building a Benchmark Engine<a class="headerlink" href="#building-a-benchmark-engine" title="Link to this heading"></a></h3>
<section id="default-build-behavior">
<h4>Default Build Behavior<a class="headerlink" href="#default-build-behavior" title="Link to this heading"></a></h4>
<p>The <code class="docutils literal notranslate"><span class="pre">trtllm-bench</span></code> CLI tool provides the <code class="docutils literal notranslate"><span class="pre">build</span></code> subcommand to build the TRT-LLM engines for max throughput benchmark.
To build an engine for benchmarking, you can specify the dataset generated with <code class="docutils literal notranslate"><span class="pre">prepare_dataset.py</span></code> through <code class="docutils literal notranslate"><span class="pre">--dataset</span></code> option.
By default, <code class="docutils literal notranslate"><span class="pre">trtllm-bench</span></code>’s tuning heuristic uses the high-level statistics of the dataset (average ISL/OSL, max sequence length)
to optimize engine build settings. The following command builds an FP8 quantized engine optimized using the dataset’s ISL/OSL.</p>
<div class="highlight-shell notranslate"><div class="highlight"><pre><span></span>trtllm-bench<span class="w"> </span>--model<span class="w"> </span>meta-llama/Llama-3.1-8B<span class="w"> </span>build<span class="w"> </span>--quantization<span class="w"> </span>FP8<span class="w"> </span>--dataset<span class="w"> </span>/tmp/synthetic_128_128.txt
</pre></div>
</div>
</section>
<section id="other-build-modes">
<h4>Other Build Modes<a class="headerlink" href="#other-build-modes" title="Link to this heading"></a></h4>
<p>The build subcommand also provides other ways to build the engine where users have larger control over the tuning values.</p>
<ul class="simple">
<li><p>Build engine with self-defined tuning values:
You specify the tuning values to build the engine with by setting <code class="docutils literal notranslate"><span class="pre">--max_batch_size</span></code> and <code class="docutils literal notranslate"><span class="pre">--max_num_tokens</span></code> directly.
<code class="docutils literal notranslate"><span class="pre">max_batch_size</span></code> and <code class="docutils literal notranslate"><span class="pre">max_num_tokens</span></code> control the maximum number of requests and tokens that can be scheduled in each iteration.
If no value is specified, the default <code class="docutils literal notranslate"><span class="pre">max_batch_size</span></code> and <code class="docutils literal notranslate"><span class="pre">max_num_tokens</span></code> values of <code class="docutils literal notranslate"><span class="pre">2048</span></code> and <code class="docutils literal notranslate"><span class="pre">8192</span></code> are used.
The following command builds an FP8 quantized engine by specifying the engine tuning values.</p></li>
</ul>
<div class="highlight-shell notranslate"><div class="highlight"><pre><span></span>trtllm-bench<span class="w"> </span>--model<span class="w"> </span>meta-llama/Llama-3.1-8B<span class="w"> </span>build<span class="w"> </span>--quantization<span class="w"> </span>FP8<span class="w"> </span>--max_seq_len<span class="w"> </span><span class="m">4096</span><span class="w"> </span>--max_batch_size<span class="w"> </span><span class="m">1024</span><span class="w"> </span>--max_num_tokens<span class="w"> </span><span class="m">2048</span>
</pre></div>
</div>
<ul class="simple">
<li><p>[Experimental] Build engine with target ISL/OSL for optimization:
In this experimental mode, you can provide hints to <code class="docutils literal notranslate"><span class="pre">trtllm-bench</span></code>’s tuning heuristic to optimize the engine on specific ISL and OSL targets.
Generally, the target ISL and OSL aligns with the average ISL and OSL of the dataset, but you can experiment with different values to optimize the engine using this mode.
The following command builds an FP8 quantized engine and optmizes for ISL:OSL targets of 128:128.</p></li>
</ul>
<div class="highlight-shell notranslate"><div class="highlight"><pre><span></span>trtllm-bench<span class="w"> </span>--model<span class="w"> </span>meta-llama/Llama-3.1-8B<span class="w"> </span>build<span class="w"> </span>--quantization<span class="w"> </span>FP8<span class="w"> </span>--max_seq_len<span class="w"> </span><span class="m">4096</span><span class="w"> </span>--target_isl<span class="w"> </span><span class="m">128</span><span class="w"> </span>--target_osl<span class="w"> </span><span class="m">128</span>
</pre></div>
</div>
</section>
<section id="parallelism-mapping-support">
<h4>Parallelism Mapping Support<a class="headerlink" href="#parallelism-mapping-support" title="Link to this heading"></a></h4>
<p>The <code class="docutils literal notranslate"><span class="pre">trtllm-bench</span> <span class="pre">build</span></code> subcommand supports combinations of tensor-parallel (TP) and pipeline-parallel (PP) mappings as long as the world size (<code class="docutils literal notranslate"><span class="pre">tp_size</span> <span class="pre">x</span> <span class="pre">pp_size</span></code>) <code class="docutils literal notranslate"><span class="pre">&lt;=</span></code> <code class="docutils literal notranslate"><span class="pre">8</span></code>. The parallelism mapping in build subcommad is controlled by <code class="docutils literal notranslate"><span class="pre">--tp_size</span></code> and <code class="docutils literal notranslate"><span class="pre">--pp_size</span></code> options. The following command builds an engine with TP2-PP2 mapping.</p>
<div class="highlight-shell notranslate"><div class="highlight"><pre><span></span>trtllm-bench<span class="w"> </span>--model<span class="w"> </span>meta-llama/Llama-3.1-8B<span class="w"> </span>build<span class="w"> </span>--quantization<span class="w"> </span>FP8<span class="w"> </span>--dataset<span class="w"> </span>/tmp/synthetic_128_128.txt<span class="w"> </span>--tp_size<span class="w"> </span><span class="m">2</span><span class="w"> </span>--pp_size<span class="w"> </span><span class="m">2</span>
</pre></div>
</div>
</section>
<section id="example-of-build-subcommand-output">
<h4>Example of Build Subcommand Output:<a class="headerlink" href="#example-of-build-subcommand-output" title="Link to this heading"></a></h4>
<p>The output of the <code class="docutils literal notranslate"><span class="pre">build</span></code> subcommand looks similar to the snippet below (for <code class="docutils literal notranslate"><span class="pre">meta-llama/Llama-3.1-8B</span></code>):</p>
<div class="highlight-shell notranslate"><div class="highlight"><pre><span></span>user@387b12598a9e:/scratch/code/trt-llm/tekit_2025$<span class="w"> </span>trtllm-bench<span class="w"> </span>--model<span class="w"> </span>meta-llama/Llama-3.1-8B<span class="w"> </span>build<span class="w"> </span>--dataset<span class="w"> </span>/tmp/synthetic_128_128.txt<span class="w"> </span>--quantization<span class="w"> </span>FP8
<span class="o">[</span>TensorRT-LLM<span class="o">]</span><span class="w"> </span>TensorRT-LLM<span class="w"> </span>version:<span class="w"> </span><span class="m">0</span>.17.0
<span class="o">[</span><span class="m">01</span>/18/2025-00:55:14<span class="o">]</span><span class="w"> </span><span class="o">[</span>TRT-LLM<span class="o">]</span><span class="w"> </span><span class="o">[</span>I<span class="o">]</span><span class="w"> </span>Found<span class="w"> </span>dataset.
<span class="o">[</span><span class="m">01</span>/18/2025-00:55:14<span class="o">]</span><span class="w"> </span><span class="o">[</span>TRT-LLM<span class="o">]</span><span class="w"> </span><span class="o">[</span>I<span class="o">]</span>
<span class="o">===========================================================</span>
<span class="o">=</span><span class="w"> </span>DATASET<span class="w"> </span><span class="nv">DETAILS</span>
<span class="o">===========================================================</span>
Max<span class="w"> </span>Input<span class="w"> </span>Sequence<span class="w"> </span>Length:<span class="w"> </span><span class="m">128</span>
Max<span class="w"> </span>Output<span class="w"> </span>Sequence<span class="w"> </span>Length:<span class="w"> </span><span class="m">128</span>
Max<span class="w"> </span>Sequence<span class="w"> </span>Length:<span class="w"> </span><span class="m">256</span>
Target<span class="w"> </span><span class="o">(</span>Average<span class="o">)</span><span class="w"> </span>Input<span class="w"> </span>Sequence<span class="w"> </span>Length:<span class="w"> </span><span class="m">128</span>
Target<span class="w"> </span><span class="o">(</span>Average<span class="o">)</span><span class="w"> </span>Output<span class="w"> </span>Sequence<span class="w"> </span>Length:<span class="w"> </span><span class="m">128</span>
Number<span class="w"> </span>of<span class="w"> </span>Sequences:<span class="w"> </span><span class="nv">3000</span>
<span class="o">===========================================================</span>
<span class="o">[</span><span class="m">01</span>/18/2025-00:55:14<span class="o">]</span><span class="w"> </span><span class="o">[</span>TRT-LLM<span class="o">]</span><span class="w"> </span><span class="o">[</span>I<span class="o">]</span><span class="w"> </span>Max<span class="w"> </span>batch<span class="w"> </span>size<span class="w"> </span>and<span class="w"> </span>max<span class="w"> </span>num<span class="w"> </span>tokens<span class="w"> </span>are<span class="w"> </span>not<span class="w"> </span>provided,<span class="w"> </span>use<span class="w"> </span>tuning<span class="w"> </span>heuristics<span class="w"> </span>or<span class="w"> </span>pre-defined<span class="w"> </span>setting<span class="w"> </span>from<span class="w"> </span>trtllm-bench.
<span class="o">[</span><span class="m">01</span>/18/2025-00:55:14<span class="o">]</span><span class="w"> </span><span class="o">[</span>TRT-LLM<span class="o">]</span><span class="w"> </span><span class="o">[</span>I<span class="o">]</span><span class="w"> </span>Estimated<span class="w"> </span>total<span class="w"> </span>available<span class="w"> </span>memory<span class="w"> </span><span class="k">for</span><span class="w"> </span>KV<span class="w"> </span>cache:<span class="w"> </span><span class="m">132</span>.37<span class="w"> </span>GB
<span class="o">[</span><span class="m">01</span>/18/2025-00:55:14<span class="o">]</span><span class="w"> </span><span class="o">[</span>TRT-LLM<span class="o">]</span><span class="w"> </span><span class="o">[</span>I<span class="o">]</span><span class="w"> </span>Estimated<span class="w"> </span>total<span class="w"> </span>KV<span class="w"> </span>cache<span class="w"> </span>memory:<span class="w"> </span><span class="m">125</span>.75<span class="w"> </span>GB
<span class="o">[</span><span class="m">01</span>/18/2025-00:55:14<span class="o">]</span><span class="w"> </span><span class="o">[</span>TRT-LLM<span class="o">]</span><span class="w"> </span><span class="o">[</span>I<span class="o">]</span><span class="w"> </span>Estimated<span class="w"> </span>max<span class="w"> </span>number<span class="w"> </span>of<span class="w"> </span>requests<span class="w"> </span><span class="k">in</span><span class="w"> </span>KV<span class="w"> </span>cache<span class="w"> </span>memory:<span class="w"> </span><span class="m">8048</span>.16
<span class="o">[</span><span class="m">01</span>/18/2025-00:55:14<span class="o">]</span><span class="w"> </span><span class="o">[</span>TRT-LLM<span class="o">]</span><span class="w"> </span><span class="o">[</span>I<span class="o">]</span><span class="w"> </span>Estimated<span class="w"> </span>max<span class="w"> </span>batch<span class="w"> </span>size<span class="w"> </span><span class="o">(</span>after<span class="w"> </span>fine-tune<span class="o">)</span>:<span class="w"> </span><span class="m">4096</span>
<span class="o">[</span><span class="m">01</span>/18/2025-00:55:14<span class="o">]</span><span class="w"> </span><span class="o">[</span>TRT-LLM<span class="o">]</span><span class="w"> </span><span class="o">[</span>I<span class="o">]</span><span class="w"> </span>Estimated<span class="w"> </span>max<span class="w"> </span>num<span class="w"> </span>tokens<span class="w"> </span><span class="o">(</span>after<span class="w"> </span>fine-tune<span class="o">)</span>:<span class="w"> </span><span class="m">8192</span>
<span class="o">[</span><span class="m">01</span>/18/2025-00:55:14<span class="o">]</span><span class="w"> </span><span class="o">[</span>TRT-LLM<span class="o">]</span><span class="w"> </span><span class="o">[</span>I<span class="o">]</span><span class="w"> </span>Set<span class="w"> </span>dtype<span class="w"> </span>to<span class="w"> </span>bfloat16.
<span class="o">[</span><span class="m">01</span>/18/2025-00:55:14<span class="o">]</span><span class="w"> </span><span class="o">[</span>TRT-LLM<span class="o">]</span><span class="w"> </span><span class="o">[</span>I<span class="o">]</span><span class="w"> </span>Set<span class="w"> </span>multiple_profiles<span class="w"> </span>to<span class="w"> </span>True.
<span class="o">[</span><span class="m">01</span>/18/2025-00:55:14<span class="o">]</span><span class="w"> </span><span class="o">[</span>TRT-LLM<span class="o">]</span><span class="w"> </span><span class="o">[</span>I<span class="o">]</span><span class="w"> </span>Set<span class="w"> </span>use_paged_context_fmha<span class="w"> </span>to<span class="w"> </span>True.
<span class="o">[</span><span class="m">01</span>/18/2025-00:55:14<span class="o">]</span><span class="w"> </span><span class="o">[</span>TRT-LLM<span class="o">]</span><span class="w"> </span><span class="o">[</span>I<span class="o">]</span><span class="w"> </span>Set<span class="w"> </span>use_fp8_context_fmha<span class="w"> </span>to<span class="w"> </span>True.
<span class="o">[</span><span class="m">01</span>/18/2025-00:55:14<span class="o">]</span><span class="w"> </span><span class="o">[</span>TRT-LLM<span class="o">]</span><span class="w"> </span><span class="o">[</span>I<span class="o">]</span>
<span class="o">===========================================================</span>
<span class="o">=</span><span class="w"> </span>ENGINE<span class="w"> </span>BUILD<span class="w"> </span><span class="nv">INFO</span>
<span class="o">===========================================================</span>
Model<span class="w"> </span>Name:<span class="w"> </span>meta-llama/Llama-3.1-8B
Model<span class="w"> </span>Path:<span class="w"> </span>None
Workspace<span class="w"> </span>Directory:<span class="w"> </span>/tmp
Engine<span class="w"> </span>Directory:<span class="w"> </span>/tmp/meta-llama/Llama-3.1-8B/tp_1_pp_1
<span class="o">===========================================================</span>
<span class="o">=</span><span class="w"> </span>ENGINE<span class="w"> </span>CONFIGURATION<span class="w"> </span><span class="nv">DETAILS</span>
<span class="o">===========================================================</span>
Max<span class="w"> </span>Sequence<span class="w"> </span>Length:<span class="w"> </span><span class="m">256</span>
Max<span class="w"> </span>Batch<span class="w"> </span>Size:<span class="w"> </span><span class="m">4096</span>
Max<span class="w"> </span>Num<span class="w"> </span>Tokens:<span class="w"> </span><span class="m">8192</span>
Quantization:<span class="w"> </span>FP8
KV<span class="w"> </span>Cache<span class="w"> </span>Dtype:<span class="w"> </span><span class="nv">FP8</span>
<span class="o">===========================================================</span>
Loading<span class="w"> </span>Model:<span class="w"> </span><span class="o">[</span><span class="m">1</span>/3<span class="o">]</span><span class="w"> </span>Downloading<span class="w"> </span>HF<span class="w"> </span>model
Downloaded<span class="w"> </span>model<span class="w"> </span>to<span class="w"> </span>/data/models--meta-llama--Llama-3.1-8B/snapshots/d04e592bb4f6aa9cfee91e2e20afa771667e1d4b
Time:<span class="w"> </span><span class="m">0</span>.321s
Loading<span class="w"> </span>Model:<span class="w"> </span><span class="o">[</span><span class="m">2</span>/3<span class="o">]</span><span class="w"> </span>Loading<span class="w"> </span>HF<span class="w"> </span>model<span class="w"> </span>to<span class="w"> </span>memory
Loading<span class="w"> </span>checkpoint<span class="w"> </span>shards:<span class="w"> </span><span class="m">100</span>%<span class="p">|</span>█████████████████████████████████████████████████████████████████████████████████████████████████████<span class="p">|</span><span class="w"> </span><span class="m">4</span>/4<span class="w"> </span><span class="o">[</span><span class="m">00</span>:59&lt;<span class="m">00</span>:00,<span class="w"> </span><span class="m">14</span>.79s/it<span class="o">]</span>
Generating<span class="w"> </span>train<span class="w"> </span>split:<span class="w"> </span><span class="m">100</span>%<span class="p">|</span>████████████████████████████████████████████████████████████████████████████████████<span class="p">|</span><span class="w"> </span><span class="m">287113</span>/287113<span class="w"> </span><span class="o">[</span><span class="m">00</span>:06&lt;<span class="m">00</span>:00,<span class="w"> </span><span class="m">41375</span>.57<span class="w"> </span>examples/s<span class="o">]</span>
Generating<span class="w"> </span>validation<span class="w"> </span>split:<span class="w"> </span><span class="m">100</span>%<span class="p">|</span>█████████████████████████████████████████████████████████████████████████████████<span class="p">|</span><span class="w"> </span><span class="m">13368</span>/13368<span class="w"> </span><span class="o">[</span><span class="m">00</span>:00&lt;<span class="m">00</span>:00,<span class="w"> </span><span class="m">41020</span>.63<span class="w"> </span>examples/s<span class="o">]</span>
Generating<span class="w"> </span><span class="nb">test</span><span class="w"> </span>split:<span class="w"> </span><span class="m">100</span>%<span class="p">|</span>███████████████████████████████████████████████████████████████████████████████████████<span class="p">|</span><span class="w"> </span><span class="m">11490</span>/11490<span class="w"> </span><span class="o">[</span><span class="m">00</span>:00&lt;<span class="m">00</span>:00,<span class="w"> </span><span class="m">41607</span>.11<span class="w"> </span>examples/s<span class="o">]</span>
Inserted<span class="w"> </span><span class="m">675</span><span class="w"> </span>quantizers
/usr/local/lib/python3.12/dist-packages/modelopt/torch/quantization/model_quant.py:71:<span class="w"> </span>DeprecationWarning:<span class="w"> </span>forward_loop<span class="w"> </span>should<span class="w"> </span>take<span class="w"> </span>model<span class="w"> </span>as<span class="w"> </span>argument,<span class="w"> </span>but<span class="w"> </span>got<span class="w"> </span>forward_loop<span class="w"> </span>without<span class="w"> </span>any<span class="w"> </span>arguments.<span class="w"> </span>This<span class="w"> </span>usage<span class="w"> </span>will<span class="w"> </span>be<span class="w"> </span>deprecated<span class="w"> </span><span class="k">in</span><span class="w"> </span>future<span class="w"> </span>versions.
<span class="w"> </span>warnings.warn<span class="o">(</span>
Disable<span class="w"> </span>lm_head<span class="w"> </span>quantization<span class="w"> </span><span class="k">for</span><span class="w"> </span>TRT-LLM<span class="w"> </span><span class="nb">export</span><span class="w"> </span>due<span class="w"> </span>to<span class="w"> </span>deployment<span class="w"> </span>limitations.
current<span class="w"> </span>rank:<span class="w"> </span><span class="m">0</span>,<span class="w"> </span>tp<span class="w"> </span>rank:<span class="w"> </span><span class="m">0</span>,<span class="w"> </span>pp<span class="w"> </span>rank:<span class="w"> </span><span class="m">0</span>
Time:<span class="w"> </span><span class="m">122</span>.568s
Loading<span class="w"> </span>Model:<span class="w"> </span><span class="o">[</span><span class="m">3</span>/3<span class="o">]</span><span class="w"> </span>Building<span class="w"> </span>TRT-LLM<span class="w"> </span>engine
/usr/local/lib/python3.12/dist-packages/tensorrt/__init__.py:85:<span class="w"> </span>DeprecationWarning:<span class="w"> </span>Context<span class="w"> </span>managers<span class="w"> </span><span class="k">for</span><span class="w"> </span>TensorRT<span class="w"> </span>types<span class="w"> </span>are<span class="w"> </span>deprecated.<span class="w"> </span>Memory<span class="w"> </span>will<span class="w"> </span>be<span class="w"> </span>freed<span class="w"> </span>automatically<span class="w"> </span>when<span class="w"> </span>the<span class="w"> </span>reference<span class="w"> </span>count<span class="w"> </span>reaches<span class="w"> </span><span class="m">0</span>.
<span class="w"> </span>warnings.warn<span class="o">(</span>
Time:<span class="w"> </span><span class="m">53</span>.820s
Loading<span class="w"> </span>model<span class="w"> </span><span class="k">done</span>.
Total<span class="w"> </span>latency:<span class="w"> </span><span class="m">176</span>.709s
&lt;snip<span class="w"> </span>verbose<span class="w"> </span>logging&gt;
<span class="o">===========================================================</span>
ENGINE<span class="w"> </span>SAVED:<span class="w"> </span>/tmp/meta-llama/Llama-3.1-8B/tp_1_pp_1
<span class="o">===========================================================</span>
</pre></div>
</div>
<p>The engine in this case will be written to <code class="docutils literal notranslate"><span class="pre">/tmp/meta-llama/Llama-3.1-8B/tp_1_pp_1</span></code> (the end of the log).</p>
</section>
</section>
<section id="max-throughput-benchmark">
<h3>Max Throughput Benchmark<a class="headerlink" href="#max-throughput-benchmark" title="Link to this heading"></a></h3>
<p>The <code class="docutils literal notranslate"><span class="pre">trtllm-bench</span></code> command line tool provides a max throughput benchmark that is accessible via the
<code class="docutils literal notranslate"><span class="pre">throughput</span></code> subcommand. This benchmark tests a TensorRT-LLM engine or PyTorch backend under maximum load to provide an
upper bound throughput number.</p>
<section id="how-the-benchmarker-works">
<h4>How the Benchmarker Works<a class="headerlink" href="#how-the-benchmarker-works" title="Link to this heading"></a></h4>
<p>The benchmarker reads a data file where a single line contains
a complete JSON request entry as specified in <a class="reference internal" href="#preparing-a-dataset"><span class="std std-ref">Preparing a Dataset</span></a>.
The process that the benchmarker is as follows:</p>
<ol class="arabic simple">
<li><p>Iterate over all input requests. If <code class="docutils literal notranslate"><span class="pre">logits</span></code> is specified, construct the request using the specified
list of logits. Otherwise, tokenize the <code class="docutils literal notranslate"><span class="pre">prompt</span></code> with as specified by <code class="docutils literal notranslate"><span class="pre">--model</span> <span class="pre">$HF_MODEL_NAME</span></code>.</p></li>
<li><p>Submit the dataset to the TensorRT-LLM <code class="docutils literal notranslate"><span class="pre">Executor</span></code> API as fast as possible (offline mode).</p></li>
<li><p>Wait for all requests to return, compute statistics, and then report results.</p></li>
</ol>
<p>To run the benchmarker, run the following commands with the <a class="reference internal" href="#building-a-benchmark-engine">engine</a> and
<a class="reference internal" href="#preparing-a-dataset">dataset</a> generated from previous steps:</p>
<div class="highlight-shell notranslate"><div class="highlight"><pre><span></span>trtllm-bench<span class="w"> </span>--model<span class="w"> </span>meta-llama/Llama-3.1-8B<span class="w"> </span>throughput<span class="w"> </span>--dataset<span class="w"> </span>/tmp/synthetic_128_128.txt<span class="w"> </span>--engine_dir<span class="w"> </span>/tmp/meta-llama/Llama-3.1-8B/tp_1_pp_1
<span class="o">[</span>TensorRT-LLM<span class="o">]</span><span class="w"> </span>TensorRT-LLM<span class="w"> </span>version:<span class="w"> </span><span class="m">0</span>.17.0
<span class="o">[</span><span class="m">01</span>/18/2025-01:01:13<span class="o">]</span><span class="w"> </span><span class="o">[</span>TRT-LLM<span class="o">]</span><span class="w"> </span><span class="o">[</span>I<span class="o">]</span><span class="w"> </span>Preparing<span class="w"> </span>to<span class="w"> </span>run<span class="w"> </span>throughput<span class="w"> </span>benchmark...
<span class="o">[</span><span class="m">01</span>/18/2025-01:01:13<span class="o">]</span><span class="w"> </span><span class="o">[</span>TRT-LLM<span class="o">]</span><span class="w"> </span><span class="o">[</span>I<span class="o">]</span><span class="w"> </span>Setting<span class="w"> </span>up<span class="w"> </span>throughput<span class="w"> </span>benchmark.
&lt;snip<span class="w"> </span>verbose<span class="w"> </span>logging&gt;
<span class="o">[</span><span class="m">01</span>/18/2025-01:01:26<span class="o">]</span><span class="w"> </span><span class="o">[</span>TRT-LLM<span class="o">]</span><span class="w"> </span><span class="o">[</span>I<span class="o">]</span><span class="w"> </span>Setting<span class="w"> </span>up<span class="w"> </span><span class="k">for</span><span class="w"> </span>warmup...
<span class="o">[</span><span class="m">01</span>/18/2025-01:01:26<span class="o">]</span><span class="w"> </span><span class="o">[</span>TRT-LLM<span class="o">]</span><span class="w"> </span><span class="o">[</span>I<span class="o">]</span><span class="w"> </span>Running<span class="w"> </span>warmup.
<span class="o">[</span><span class="m">01</span>/18/2025-01:01:26<span class="o">]</span><span class="w"> </span><span class="o">[</span>TRT-LLM<span class="o">]</span><span class="w"> </span><span class="o">[</span>I<span class="o">]</span><span class="w"> </span>Starting<span class="w"> </span>benchmarking<span class="w"> </span>async<span class="w"> </span>task.
<span class="o">[</span><span class="m">01</span>/18/2025-01:01:26<span class="o">]</span><span class="w"> </span><span class="o">[</span>TRT-LLM<span class="o">]</span><span class="w"> </span><span class="o">[</span>I<span class="o">]</span><span class="w"> </span>Starting<span class="w"> </span>benchmark...
<span class="o">[</span><span class="m">01</span>/18/2025-01:01:26<span class="o">]</span><span class="w"> </span><span class="o">[</span>TRT-LLM<span class="o">]</span><span class="w"> </span><span class="o">[</span>I<span class="o">]</span><span class="w"> </span>Request<span class="w"> </span>submission<span class="w"> </span>complete.<span class="w"> </span><span class="o">[</span><span class="nv">count</span><span class="o">=</span><span class="m">2</span>,<span class="w"> </span><span class="nv">time</span><span class="o">=</span><span class="m">0</span>.0000s,<span class="w"> </span><span class="nv">rate</span><span class="o">=</span><span class="m">121847</span>.20<span class="w"> </span>req/s<span class="o">]</span>
<span class="o">[</span><span class="m">01</span>/18/2025-01:01:28<span class="o">]</span><span class="w"> </span><span class="o">[</span>TRT-LLM<span class="o">]</span><span class="w"> </span><span class="o">[</span>I<span class="o">]</span><span class="w"> </span>Benchmark<span class="w"> </span>complete.
<span class="o">[</span><span class="m">01</span>/18/2025-01:01:28<span class="o">]</span><span class="w"> </span><span class="o">[</span>TRT-LLM<span class="o">]</span><span class="w"> </span><span class="o">[</span>I<span class="o">]</span><span class="w"> </span>Stopping<span class="w"> </span>LLM<span class="w"> </span>backend.
<span class="o">[</span><span class="m">01</span>/18/2025-01:01:28<span class="o">]</span><span class="w"> </span><span class="o">[</span>TRT-LLM<span class="o">]</span><span class="w"> </span><span class="o">[</span>I<span class="o">]</span><span class="w"> </span>Cancelling<span class="w"> </span>all<span class="w"> </span><span class="m">0</span><span class="w"> </span>tasks<span class="w"> </span>to<span class="w"> </span>complete.
<span class="o">[</span><span class="m">01</span>/18/2025-01:01:28<span class="o">]</span><span class="w"> </span><span class="o">[</span>TRT-LLM<span class="o">]</span><span class="w"> </span><span class="o">[</span>I<span class="o">]</span><span class="w"> </span>All<span class="w"> </span>tasks<span class="w"> </span>cancelled.
<span class="o">[</span><span class="m">01</span>/18/2025-01:01:28<span class="o">]</span><span class="w"> </span><span class="o">[</span>TRT-LLM<span class="o">]</span><span class="w"> </span><span class="o">[</span>I<span class="o">]</span><span class="w"> </span>LLM<span class="w"> </span>Backend<span class="w"> </span>stopped.
<span class="o">[</span><span class="m">01</span>/18/2025-01:01:28<span class="o">]</span><span class="w"> </span><span class="o">[</span>TRT-LLM<span class="o">]</span><span class="w"> </span><span class="o">[</span>I<span class="o">]</span><span class="w"> </span>Warmup<span class="w"> </span><span class="k">done</span>.
<span class="o">[</span><span class="m">01</span>/18/2025-01:01:28<span class="o">]</span><span class="w"> </span><span class="o">[</span>TRT-LLM<span class="o">]</span><span class="w"> </span><span class="o">[</span>I<span class="o">]</span><span class="w"> </span>Starting<span class="w"> </span>benchmarking<span class="w"> </span>async<span class="w"> </span>task.
<span class="o">[</span><span class="m">01</span>/18/2025-01:01:28<span class="o">]</span><span class="w"> </span><span class="o">[</span>TRT-LLM<span class="o">]</span><span class="w"> </span><span class="o">[</span>I<span class="o">]</span><span class="w"> </span>Starting<span class="w"> </span>benchmark...
<span class="o">[</span><span class="m">01</span>/18/2025-01:01:28<span class="o">]</span><span class="w"> </span><span class="o">[</span>TRT-LLM<span class="o">]</span><span class="w"> </span><span class="o">[</span>I<span class="o">]</span><span class="w"> </span>Request<span class="w"> </span>submission<span class="w"> </span>complete.<span class="w"> </span><span class="o">[</span><span class="nv">count</span><span class="o">=</span><span class="m">3000</span>,<span class="w"> </span><span class="nv">time</span><span class="o">=</span><span class="m">0</span>.0012s,<span class="w"> </span><span class="nv">rate</span><span class="o">=</span><span class="m">2590780</span>.97<span class="w"> </span>req/s<span class="o">]</span>
<span class="o">[</span><span class="m">01</span>/18/2025-01:01:42<span class="o">]</span><span class="w"> </span><span class="o">[</span>TRT-LLM<span class="o">]</span><span class="w"> </span><span class="o">[</span>I<span class="o">]</span><span class="w"> </span>Benchmark<span class="w"> </span>complete.
<span class="o">[</span><span class="m">01</span>/18/2025-01:01:42<span class="o">]</span><span class="w"> </span><span class="o">[</span>TRT-LLM<span class="o">]</span><span class="w"> </span><span class="o">[</span>I<span class="o">]</span><span class="w"> </span>Stopping<span class="w"> </span>LLM<span class="w"> </span>backend.
<span class="o">[</span><span class="m">01</span>/18/2025-01:01:42<span class="o">]</span><span class="w"> </span><span class="o">[</span>TRT-LLM<span class="o">]</span><span class="w"> </span><span class="o">[</span>I<span class="o">]</span><span class="w"> </span>Cancelling<span class="w"> </span>all<span class="w"> </span><span class="m">0</span><span class="w"> </span>tasks<span class="w"> </span>to<span class="w"> </span>complete.
<span class="o">[</span><span class="m">01</span>/18/2025-01:01:42<span class="o">]</span><span class="w"> </span><span class="o">[</span>TRT-LLM<span class="o">]</span><span class="w"> </span><span class="o">[</span>I<span class="o">]</span><span class="w"> </span>All<span class="w"> </span>tasks<span class="w"> </span>cancelled.
<span class="o">[</span><span class="m">01</span>/18/2025-01:01:42<span class="o">]</span><span class="w"> </span><span class="o">[</span>TRT-LLM<span class="o">]</span><span class="w"> </span><span class="o">[</span>I<span class="o">]</span><span class="w"> </span>LLM<span class="w"> </span>Backend<span class="w"> </span>stopped.
<span class="o">[</span><span class="m">01</span>/18/2025-01:01:42<span class="o">]</span><span class="w"> </span><span class="o">[</span>TRT-LLM<span class="o">]</span><span class="w"> </span><span class="o">[</span>I<span class="o">]</span>
<span class="o">===========================================================</span>
<span class="o">=</span><span class="w"> </span>ENGINE<span class="w"> </span><span class="nv">DETAILS</span>
<span class="o">===========================================================</span>
Model:<span class="w"> </span>meta-llama/Llama-3.1-8B
Engine<span class="w"> </span>Directory:<span class="w"> </span>/tmp/meta-llama/Llama-3.1-8B/tp_1_pp_1
TensorRT-LLM<span class="w"> </span>Version:<span class="w"> </span><span class="m">0</span>.17.0
Dtype:<span class="w"> </span>bfloat16
KV<span class="w"> </span>Cache<span class="w"> </span>Dtype:<span class="w"> </span>FP8
Quantization:<span class="w"> </span>FP8
Max<span class="w"> </span>Input<span class="w"> </span>Length:<span class="w"> </span><span class="m">256</span>
Max<span class="w"> </span>Sequence<span class="w"> </span>Length:<span class="w"> </span><span class="nv">256</span>
<span class="o">===========================================================</span>
<span class="o">=</span><span class="w"> </span>WORLD<span class="w"> </span>+<span class="w"> </span>RUNTIME<span class="w"> </span><span class="nv">INFORMATION</span>
<span class="o">===========================================================</span>
TP<span class="w"> </span>Size:<span class="w"> </span><span class="m">1</span>
PP<span class="w"> </span>Size:<span class="w"> </span><span class="m">1</span>
Max<span class="w"> </span>Runtime<span class="w"> </span>Batch<span class="w"> </span>Size:<span class="w"> </span><span class="m">4096</span>
Max<span class="w"> </span>Runtime<span class="w"> </span>Tokens:<span class="w"> </span><span class="m">8192</span>
Scheduling<span class="w"> </span>Policy:<span class="w"> </span>Guaranteed<span class="w"> </span>No<span class="w"> </span>Evict
KV<span class="w"> </span>Memory<span class="w"> </span>Percentage:<span class="w"> </span><span class="m">90</span>.00%
Issue<span class="w"> </span>Rate<span class="w"> </span><span class="o">(</span>req/sec<span class="o">)</span>:<span class="w"> </span><span class="m">5</span>.0689E+14
<span class="o">===========================================================</span>
<span class="o">=</span><span class="w"> </span>PERFORMANCE<span class="w"> </span><span class="nv">OVERVIEW</span>
<span class="o">===========================================================</span>
Number<span class="w"> </span>of<span class="w"> </span>requests:<span class="w"> </span><span class="m">3000</span>
Average<span class="w"> </span>Input<span class="w"> </span>Length<span class="w"> </span><span class="o">(</span>tokens<span class="o">)</span>:<span class="w"> </span><span class="m">128</span>.0000
Average<span class="w"> </span>Output<span class="w"> </span>Length<span class="w"> </span><span class="o">(</span>tokens<span class="o">)</span>:<span class="w"> </span><span class="m">128</span>.0000
Token<span class="w"> </span>Throughput<span class="w"> </span><span class="o">(</span>tokens/sec<span class="o">)</span>:<span class="w"> </span><span class="m">28390</span>.4265
Request<span class="w"> </span>Throughput<span class="w"> </span><span class="o">(</span>req/sec<span class="o">)</span>:<span class="w"> </span><span class="m">221</span>.8002
Total<span class="w"> </span>Latency<span class="w"> </span><span class="o">(</span>ms<span class="o">)</span>:<span class="w"> </span><span class="m">13525</span>.6862
<span class="o">===========================================================</span>
<span class="o">[</span><span class="m">01</span>/18/2025-01:01:42<span class="o">]</span><span class="w"> </span><span class="o">[</span>TRT-LLM<span class="o">]</span><span class="w"> </span><span class="o">[</span>I<span class="o">]</span><span class="w"> </span>Thread<span class="w"> </span>proxy_dispatch_result_thread<span class="w"> </span>stopped.
<span class="o">[</span>TensorRT-LLM<span class="o">][</span>INFO<span class="o">]</span><span class="w"> </span>Refreshed<span class="w"> </span>the<span class="w"> </span>MPI<span class="w"> </span><span class="nb">local</span><span class="w"> </span>session
</pre></div>
</div>
</section>
</section>
<section id="running-with-the-pytorch-workflow">
<h3>Running with the PyTorch Workflow<a class="headerlink" href="#running-with-the-pytorch-workflow" title="Link to this heading"></a></h3>
<p>To benchmark the PyTorch backend (<code class="docutils literal notranslate"><span class="pre">tensorrt_llm._torch</span></code>), use the following command with <a class="reference internal" href="#preparing-a-dataset">dataset</a> generated from previous steps. With the PyTorch flow, you will not need to
run <code class="docutils literal notranslate"><span class="pre">trtllm-bench</span> <span class="pre">build</span></code>; the <code class="docutils literal notranslate"><span class="pre">throughput</span></code> benchmark initializes the backend by tuning against the
dataset provided via <code class="docutils literal notranslate"><span class="pre">--dataset</span></code> (or the other build mode settings described <a class="reference internal" href="#other-build-modes">above</a>).
Note that CUDA graph is enabled by default. You can add additional pytorch config with
<code class="docutils literal notranslate"><span class="pre">--extra_llm_api_options</span></code> followed by the path to a YAML file. For more details, please refer to the
help text by running the command with <code class="docutils literal notranslate"><span class="pre">--help</span></code>.</p>
<div class="admonition tip">
<p class="admonition-title">Tip</p>
<p>The command below specifies the <code class="docutils literal notranslate"><span class="pre">--model_path</span></code> option. The model path is optional and used only when you want to run a locally
stored checkpoint. When using <code class="docutils literal notranslate"><span class="pre">--model_path</span></code>, the <code class="docutils literal notranslate"><span class="pre">--model</span></code> is still required for reporting reasons and in order to look up parameters
for build heuristics.</p>
</div>
<div class="highlight-shell notranslate"><div class="highlight"><pre><span></span>trtllm-bench<span class="w"> </span>--model<span class="w"> </span>meta-llama/Llama-3.1-8B<span class="w"> </span>--model_path<span class="w"> </span>/Ckpt/Path/To/Llama-3.1-8B<span class="w"> </span>throughput<span class="w"> </span>--dataset<span class="w"> </span>/tmp/synthetic_128_128.txt<span class="w"> </span>--backend<span class="w"> </span>pytorch
<span class="c1"># Example output</span>
&lt;snip<span class="w"> </span>verbose<span class="w"> </span>logging&gt;
<span class="o">===========================================================</span>
<span class="o">=</span><span class="w"> </span>PyTorch<span class="w"> </span><span class="nv">backend</span>
<span class="o">===========================================================</span>
Model:<span class="w"> </span>meta-llama/Llama-3.1-8B
Model<span class="w"> </span>Path:<span class="w"> </span>/Ckpt/Path/To/Llama-3.1-8B
TensorRT-LLM<span class="w"> </span>Version:<span class="w"> </span><span class="m">0</span>.17.0
Dtype:<span class="w"> </span>bfloat16
KV<span class="w"> </span>Cache<span class="w"> </span>Dtype:<span class="w"> </span>None
Quantization:<span class="w"> </span><span class="nv">FP8</span>
<span class="o">===========================================================</span>
<span class="o">=</span><span class="w"> </span>WORLD<span class="w"> </span>+<span class="w"> </span>RUNTIME<span class="w"> </span><span class="nv">INFORMATION</span>
<span class="o">===========================================================</span>
TP<span class="w"> </span>Size:<span class="w"> </span><span class="m">1</span>
PP<span class="w"> </span>Size:<span class="w"> </span><span class="m">1</span>
Max<span class="w"> </span>Runtime<span class="w"> </span>Batch<span class="w"> </span>Size:<span class="w"> </span><span class="m">2048</span>
Max<span class="w"> </span>Runtime<span class="w"> </span>Tokens:<span class="w"> </span><span class="m">4096</span>
Scheduling<span class="w"> </span>Policy:<span class="w"> </span>Guaranteed<span class="w"> </span>No<span class="w"> </span>Evict
KV<span class="w"> </span>Memory<span class="w"> </span>Percentage:<span class="w"> </span><span class="m">90</span>.00%
Issue<span class="w"> </span>Rate<span class="w"> </span><span class="o">(</span>req/sec<span class="o">)</span>:<span class="w"> </span><span class="m">7</span>.6753E+14
<span class="o">===========================================================</span>
<span class="o">=</span><span class="w"> </span>PERFORMANCE<span class="w"> </span><span class="nv">OVERVIEW</span>
<span class="o">===========================================================</span>
Number<span class="w"> </span>of<span class="w"> </span>requests:<span class="w"> </span><span class="m">3000</span>
Average<span class="w"> </span>Input<span class="w"> </span>Length<span class="w"> </span><span class="o">(</span>tokens<span class="o">)</span>:<span class="w"> </span><span class="m">128</span>.0000
Average<span class="w"> </span>Output<span class="w"> </span>Length<span class="w"> </span><span class="o">(</span>tokens<span class="o">)</span>:<span class="w"> </span><span class="m">128</span>.0000
Token<span class="w"> </span>Throughput<span class="w"> </span><span class="o">(</span>tokens/sec<span class="o">)</span>:<span class="w"> </span><span class="m">20685</span>.5510
Request<span class="w"> </span>Throughput<span class="w"> </span><span class="o">(</span>req/sec<span class="o">)</span>:<span class="w"> </span><span class="m">161</span>.6059
Total<span class="w"> </span>Latency<span class="w"> </span><span class="o">(</span>ms<span class="o">)</span>:<span class="w"> </span><span class="m">18563</span>.6825
</pre></div>
</div>
<section id="quantization-in-the-pytorch-flow">
<h4>Quantization in the PyTorch Flow<a class="headerlink" href="#quantization-in-the-pytorch-flow" title="Link to this heading"></a></h4>
<p>In order to run a quantized run with <code class="docutils literal notranslate"><span class="pre">trtllm-bench</span></code> utilizing the PyTorch flow, you will need to use a pre-quantized
To run a quantized benchmark with <code class="docutils literal notranslate"><span class="pre">trtllm-bench</span></code> utilizing the PyTorch flow, you will need to use a pre-quantized
checkpoint. For the Llama-3.1 models, TensorRT-LLM provides the following checkpoints via HuggingFace:</p>
<ul class="simple">
<li><p><a class="reference external" href="https://huggingface.co/nvidia/Llama-3.1-8B-Instruct-FP8"><code class="docutils literal notranslate"><span class="pre">nvidia/Llama-3.1-8B-Instruct-FP8</span></code></a></p></li>
<li><p><a class="reference external" href="https://huggingface.co/nvidia/Llama-3.1-70B-Instruct-FP8"><code class="docutils literal notranslate"><span class="pre">nvidia/Llama-3.1-70B-Instruct-FP8</span></code></a></p></li>
<li><p><a class="reference external" href="https://huggingface.co/nvidia/Llama-3.1-405B-Instruct-FP8"><code class="docutils literal notranslate"><span class="pre">nvidia/Llama-3.1-405B-Instruct-FP8</span></code></a></p></li>
</ul>
<p><code class="docutils literal notranslate"><span class="pre">trtllm-bench</span></code> utilizes the <code class="docutils literal notranslate"><span class="pre">hf_quant_config.json</span></code> file present in the pre-quantized checkpoints above. The configuration
file is present in checkpoints quantized with <a class="reference external" href="https://github.com/NVIDIA/TensorRT-Model-Optimizer">TensorRT Model Optimizer</a>
and describes the compute and KV cache quantization that checkpoint was compiled with. For example, from the checkpoints
above:</p>
<div class="highlight-json notranslate"><div class="highlight"><pre><span></span><span class="p">{</span>
<span class="w"> </span><span class="nt">&quot;producer&quot;</span><span class="p">:</span><span class="w"> </span><span class="p">{</span>
<span class="w"> </span><span class="nt">&quot;name&quot;</span><span class="p">:</span><span class="w"> </span><span class="s2">&quot;modelopt&quot;</span><span class="p">,</span>
<span class="w"> </span><span class="nt">&quot;version&quot;</span><span class="p">:</span><span class="w"> </span><span class="s2">&quot;0.23.0rc1&quot;</span>
<span class="w"> </span><span class="p">},</span>
<span class="w"> </span><span class="nt">&quot;quantization&quot;</span><span class="p">:</span><span class="w"> </span><span class="p">{</span>
<span class="w"> </span><span class="nt">&quot;quant_algo&quot;</span><span class="p">:</span><span class="w"> </span><span class="s2">&quot;FP8&quot;</span><span class="p">,</span>
<span class="w"> </span><span class="nt">&quot;kv_cache_quant_algo&quot;</span><span class="p">:</span><span class="w"> </span><span class="kc">null</span>
<span class="w"> </span><span class="p">}</span>
</pre></div>
</div>
<p>The checkpoints above are quantized to run with a compute precision of <code class="docutils literal notranslate"><span class="pre">FP8</span></code> and default to no KV cache quantization (full
<code class="docutils literal notranslate"><span class="pre">FP16</span></code> cache). When running <code class="docutils literal notranslate"><span class="pre">trtllm-bench</span> <span class="pre">throughput</span></code>. The benchmark will select a KV cache quantization that is best suited
for the compute precision in the checkpoint automatically if <code class="docutils literal notranslate"><span class="pre">kv_cache_quant_algo</span></code> is specified as <code class="docutils literal notranslate"><span class="pre">null</span></code>, otherwise it will
be forced to match the specified non-null KV cache quantization. The following are the mappings that <code class="docutils literal notranslate"><span class="pre">trtllm-bench</span></code> will
follow when a checkpoint does not specify a KV cache quantization algorithm:</p>
<table class="docutils align-default">
<thead>
<tr class="row-odd"><th class="head"><p>Checkpoint Compute Quant</p></th>
<th class="head"><p>Checkpoint KV Cache Quant</p></th>
<th class="head"><p><code class="docutils literal notranslate"><span class="pre">trtllm-bench</span></code></p></th>
<th class="head"><p>Note</p></th>
</tr>
</thead>
<tbody>
<tr class="row-even"><td><p><code class="docutils literal notranslate"><span class="pre">null</span></code></p></td>
<td><p><code class="docutils literal notranslate"><span class="pre">null</span></code></p></td>
<td><p><code class="docutils literal notranslate"><span class="pre">null</span></code></p></td>
<td><p>In this case, a quantization config doesn’t exist.</p></td>
</tr>
<tr class="row-odd"><td><p><code class="docutils literal notranslate"><span class="pre">FP8</span></code></p></td>
<td><p><code class="docutils literal notranslate"><span class="pre">FP8</span></code></p></td>
<td><p><code class="docutils literal notranslate"><span class="pre">FP8</span></code></p></td>
<td><p>Matches the checkpoint</p></td>
</tr>
<tr class="row-even"><td><p><code class="docutils literal notranslate"><span class="pre">FP8</span></code></p></td>
<td><p><code class="docutils literal notranslate"><span class="pre">null</span></code></p></td>
<td><p><code class="docutils literal notranslate"><span class="pre">FP8</span></code></p></td>
<td><p>Set to <code class="docutils literal notranslate"><span class="pre">FP8</span></code> via benchmark</p></td>
</tr>
<tr class="row-odd"><td><p><code class="docutils literal notranslate"><span class="pre">NVFP4</span></code></p></td>
<td><p><code class="docutils literal notranslate"><span class="pre">null</span></code></p></td>
<td><p><code class="docutils literal notranslate"><span class="pre">FP8</span></code></p></td>
<td><p>Set to <code class="docutils literal notranslate"><span class="pre">FP8</span></code> via benchmark</p></td>
</tr>
</tbody>
</table>
<p>If you would like to force the KV cache quantizaton, you can specify the following in the YAML file to force the precision
when the checkpoint precision is <code class="docutils literal notranslate"><span class="pre">null</span></code>:</p>
<div class="highlight-yaml notranslate"><div class="highlight"><pre><span></span><span class="nt">pytorch_backend_config</span><span class="p">:</span>
<span class="w"> </span><span class="nt">kv_cache_dtype</span><span class="p">:</span><span class="w"> </span><span class="s">&quot;fp8&quot;</span>
</pre></div>
</div>
<div class="admonition tip">
<p class="admonition-title">Tip</p>
<p>The two valid values for <code class="docutils literal notranslate"><span class="pre">kv_cache_dtype</span></code> are <code class="docutils literal notranslate"><span class="pre">auto</span></code> and <code class="docutils literal notranslate"><span class="pre">fp8</span></code>.</p>
</div>
</section>
</section>
</section>
<section id="low-latency-benchmark">
<h2>Low Latency Benchmark<a class="headerlink" href="#low-latency-benchmark" title="Link to this heading"></a></h2>
<p>The low latency benchmark follows a similar workflow to the <a class="reference internal" href="#max-throughput-benchmark">throughput benchmark</a>
but requires building the engine separately from <code class="docutils literal notranslate"><span class="pre">trtllm-bench</span></code>. Low latency benchmarks has the following modes:</p>
<ul class="simple">
<li><p>A single-request low-latency engine</p></li>
<li><p>A Medusa-enabled speculative-decoding engine</p></li>
</ul>
<section id="low-latency-tensorrt-llm-engine-for-llama-3-70b">
<h3>Low Latency TensorRT-LLM Engine for Llama-3 70B<a class="headerlink" href="#low-latency-tensorrt-llm-engine-for-llama-3-70b" title="Link to this heading"></a></h3>
<p>To build a low-latency engine for the latency benchmark, run the following quantize and build commands.
The <code class="docutils literal notranslate"><span class="pre">$checkpoint_dir</span></code> is the path to the <a class="reference external" href="https://huggingface.co/meta-llama/Meta-Llama-3-70B">meta-llama/Meta-Llama-3-70B</a> Hugging Face checkpoint in your cache or downloaded to a specific location with the <a class="reference external" href="https://huggingface.co/docs/huggingface_hub/en/guides/cli">huggingface-cli</a>.
To prepare a dataset, follow the same process as specified in <a class="reference internal" href="#preparing-a-dataset"><span class="std std-ref">Preparing a Dataset</span></a>.</p>
<section id="benchmarking-a-non-medusa-low-latency-engine">
<h4>Benchmarking a non-Medusa Low Latency Engine<a class="headerlink" href="#benchmarking-a-non-medusa-low-latency-engine" title="Link to this heading"></a></h4>
<p>To quantize the checkpoint:</p>
<div class="highlight-shell notranslate"><div class="highlight"><pre><span></span><span class="nb">cd</span><span class="w"> </span>tensorrt_llm/examples/llama
python<span class="w"> </span>../quantization/quantize.py<span class="w"> </span><span class="se">\</span>
<span class="w"> </span>--model_dir<span class="w"> </span><span class="nv">$checkpoint_dir</span><span class="w"> </span><span class="se">\</span>
<span class="w"> </span>--dtype<span class="w"> </span>bfloat16<span class="w"> </span><span class="se">\</span>
<span class="w"> </span>--qformat<span class="w"> </span>fp8<span class="w"> </span><span class="se">\</span>
<span class="w"> </span>--kv_cache_dtype<span class="w"> </span>fp8<span class="w"> </span><span class="se">\</span>
<span class="w"> </span>--output_dir<span class="w"> </span>/tmp/meta-llama/Meta-Llama-3-70B/checkpoint<span class="w"> </span><span class="se">\</span>
<span class="w"> </span>--calib_size<span class="w"> </span><span class="m">512</span><span class="w"> </span><span class="se">\</span>
<span class="w"> </span>--tp_size<span class="w"> </span><span class="nv">$tp_size</span>
</pre></div>
</div>
<p>then build,</p>
<div class="highlight-shell notranslate"><div class="highlight"><pre><span></span>trtllm-build<span class="w"> </span><span class="se">\</span>
<span class="w"> </span>--checkpoint_dir<span class="w"> </span>/tmp/meta-llama/Meta-Llama-3-70B/checkpoint<span class="w"> </span><span class="se">\</span>
<span class="w"> </span>--use_fused_mlp<span class="w"> </span><span class="nb">enable</span><span class="w"> </span><span class="se">\</span>
<span class="w"> </span>--gpt_attention_plugin<span class="w"> </span>bfloat16<span class="w"> </span><span class="se">\</span>
<span class="w"> </span>--output_dir<span class="w"> </span>/tmp/meta-llama/Meta-Llama-3-70B/engine<span class="w"> </span><span class="se">\</span>
<span class="w"> </span>--max_batch_size<span class="w"> </span><span class="m">1</span><span class="w"> </span><span class="se">\</span>
<span class="w"> </span>--max_seq_len<span class="w"> </span><span class="k">$((</span><span class="nv">$isl</span><span class="o">+</span><span class="nv">$osl</span><span class="k">))</span><span class="w"> </span><span class="se">\</span>
<span class="w"> </span>--reduce_fusion<span class="w"> </span><span class="nb">enable</span><span class="w"> </span><span class="se">\</span>
<span class="w"> </span>--gemm_plugin<span class="w"> </span>fp8<span class="w"> </span><span class="se">\</span>
<span class="w"> </span>--workers<span class="w"> </span><span class="nv">$tp_size</span><span class="w"> </span><span class="se">\</span>
<span class="w"> </span>--use_fp8_context_fmha<span class="w"> </span><span class="nb">enable</span><span class="w"> </span><span class="se">\</span>
<span class="w"> </span>--max_num_tokens<span class="w"> </span><span class="nv">$isl</span><span class="w"> </span><span class="se">\</span>
<span class="w"> </span>--use_paged_context_fmha<span class="w"> </span>disable<span class="w"> </span><span class="se">\</span>
<span class="w"> </span>--multiple_profiles<span class="w"> </span><span class="nb">enable</span>
</pre></div>
</div>
<p>After the engine is built, run the low-latency benchmark:</p>
<div class="highlight-shell notranslate"><div class="highlight"><pre><span></span>env<span class="w"> </span><span class="nv">TRTLLM_ENABLE_MMHA_MULTI_BLOCK_DEBUG</span><span class="o">=</span><span class="m">1</span><span class="w"> </span><span class="se">\</span>
<span class="w"> </span><span class="nv">TRTLLM_MMHA_KERNEL_BLOCK_SIZE</span><span class="o">=</span><span class="m">256</span><span class="w"> </span><span class="se">\</span>
<span class="w"> </span><span class="nv">TRTLLM_MMHA_BLOCKS_PER_SEQUENCE</span><span class="o">=</span><span class="m">32</span><span class="w"> </span><span class="se">\</span>
<span class="w"> </span><span class="nv">FORCE_MULTI_BLOCK_MODE</span><span class="o">=</span>ON<span class="w"> </span><span class="se">\</span>
<span class="w"> </span><span class="nv">TRTLLM_ENABLE_PDL</span><span class="o">=</span><span class="m">1</span><span class="w"> </span><span class="se">\</span>
<span class="w"> </span>trtllm-bench<span class="w"> </span>--model<span class="w"> </span>meta-llama/Meta-Llama-3-70B<span class="w"> </span><span class="se">\</span>
<span class="w"> </span>latency<span class="w"> </span><span class="se">\</span>
<span class="w"> </span>--dataset<span class="w"> </span><span class="nv">$DATASET_PATH</span><span class="w"> </span><span class="se">\</span>
<span class="w"> </span>--engine_dir<span class="w"> </span>/tmp/meta-llama/Meta-Llama-3-70B/engine
</pre></div>
</div>
</section>
</section>
<section id="building-a-medusa-low-latency-engine">
<h3>Building a Medusa Low-Latency Engine<a class="headerlink" href="#building-a-medusa-low-latency-engine" title="Link to this heading"></a></h3>
<p>To build a Medusa-enabled engine requires checkpoints that contain Medusa heads.
NVIDIA provides TensorRT-LLM checkpoints on the <a class="reference external" href="https://huggingface.co/nvidia">NVIDIA</a> page on Hugging Face.
The checkpoints are pre-quantized and can be directly built after downloading them with the
<a class="reference external" href="https://huggingface.co/docs/huggingface_hub/en/guides/cli">huggingface-cli</a>.
After you download the checkpoints, run the following command. Make sure to
specify the <code class="docutils literal notranslate"><span class="pre">$tp_size</span></code> supported by your Medusa checkpoint and the path to its stored location <code class="docutils literal notranslate"><span class="pre">$checkpoint_dir</span></code>.
Additionally, <code class="docutils literal notranslate"><span class="pre">$max_seq_len</span></code> should be set to the model’s maximum position embedding.</p>
<p>Using Llama-3.1 70B as an example, for a tensor parallel 8 and bfloat16 dtype:</p>
<div class="highlight-shell notranslate"><div class="highlight"><pre><span></span><span class="nv">tp_size</span><span class="o">=</span><span class="m">8</span>
<span class="nv">max_seq_len</span><span class="o">=</span><span class="m">131072</span>
trtllm-build<span class="w"> </span>--checkpoint_dir<span class="w"> </span><span class="nv">$checkpoint_dir</span><span class="w"> </span><span class="se">\</span>
<span class="w"> </span>--speculative_decoding_mode<span class="w"> </span>medusa<span class="w"> </span><span class="se">\</span>
<span class="w"> </span>--max_batch_size<span class="w"> </span><span class="m">1</span><span class="w"> </span><span class="se">\</span>
<span class="w"> </span>--gpt_attention_plugin<span class="w"> </span>bfloat16<span class="w"> </span><span class="se">\</span>
<span class="w"> </span>--max_seq_len<span class="w"> </span><span class="nv">$max_seq_len</span><span class="w"> </span><span class="se">\</span>
<span class="w"> </span>--output_dir<span class="w"> </span>/tmp/meta-llama/Meta-Llama-3.1-70B/medusa/engine<span class="w"> </span><span class="se">\</span>
<span class="w"> </span>--use_fused_mlp<span class="w"> </span><span class="nb">enable</span><span class="w"> </span><span class="se">\</span>
<span class="w"> </span>--paged_kv_cache<span class="w"> </span><span class="nb">enable</span><span class="w"> </span><span class="se">\</span>
<span class="w"> </span>--use_paged_context_fmha<span class="w"> </span>disable<span class="w"> </span><span class="se">\</span>
<span class="w"> </span>--multiple_profiles<span class="w"> </span><span class="nb">enable</span><span class="w"> </span><span class="se">\</span>
<span class="w"> </span>--reduce_fusion<span class="w"> </span><span class="nb">enable</span><span class="w"> </span><span class="se">\</span>
<span class="w"> </span>--use_fp8_context_fmha<span class="w"> </span><span class="nb">enable</span><span class="w"> </span><span class="se">\</span>
<span class="w"> </span>--workers<span class="w"> </span><span class="nv">$tp_size</span><span class="w"> </span><span class="se">\</span>
<span class="w"> </span>--low_latency_gemm_plugin<span class="w"> </span>fp8
</pre></div>
</div>
<p>After the engine is built, you need to define the Medusa choices.
The choices are specified with a YAML file like the following example (<code class="docutils literal notranslate"><span class="pre">medusa.yaml</span></code>):</p>
<div class="highlight-yaml notranslate"><div class="highlight"><pre><span></span><span class="p p-Indicator">-</span><span class="w"> </span><span class="p p-Indicator">[</span><span class="nv">0</span><span class="p p-Indicator">]</span>
<span class="p p-Indicator">-</span><span class="w"> </span><span class="p p-Indicator">[</span><span class="nv">0</span><span class="p p-Indicator">,</span><span class="w"> </span><span class="nv">0</span><span class="p p-Indicator">]</span>
<span class="p p-Indicator">-</span><span class="w"> </span><span class="p p-Indicator">[</span><span class="nv">1</span><span class="p p-Indicator">]</span>
<span class="p p-Indicator">-</span><span class="w"> </span><span class="p p-Indicator">[</span><span class="nv">0</span><span class="p p-Indicator">,</span><span class="w"> </span><span class="nv">1</span><span class="p p-Indicator">]</span>
<span class="p p-Indicator">-</span><span class="w"> </span><span class="p p-Indicator">[</span><span class="nv">2</span><span class="p p-Indicator">]</span>
<span class="p p-Indicator">-</span><span class="w"> </span><span class="p p-Indicator">[</span><span class="nv">0</span><span class="p p-Indicator">,</span><span class="w"> </span><span class="nv">0</span><span class="p p-Indicator">,</span><span class="w"> </span><span class="nv">0</span><span class="p p-Indicator">]</span>
<span class="p p-Indicator">-</span><span class="w"> </span><span class="p p-Indicator">[</span><span class="nv">1</span><span class="p p-Indicator">,</span><span class="w"> </span><span class="nv">0</span><span class="p p-Indicator">]</span>
<span class="p p-Indicator">-</span><span class="w"> </span><span class="p p-Indicator">[</span><span class="nv">0</span><span class="p p-Indicator">,</span><span class="w"> </span><span class="nv">2</span><span class="p p-Indicator">]</span>
<span class="p p-Indicator">-</span><span class="w"> </span><span class="p p-Indicator">[</span><span class="nv">3</span><span class="p p-Indicator">]</span>
<span class="p p-Indicator">-</span><span class="w"> </span><span class="p p-Indicator">[</span><span class="nv">0</span><span class="p p-Indicator">,</span><span class="w"> </span><span class="nv">3</span><span class="p p-Indicator">]</span>
<span class="p p-Indicator">-</span><span class="w"> </span><span class="p p-Indicator">[</span><span class="nv">4</span><span class="p p-Indicator">]</span>
<span class="p p-Indicator">-</span><span class="w"> </span><span class="p p-Indicator">[</span><span class="nv">0</span><span class="p p-Indicator">,</span><span class="w"> </span><span class="nv">4</span><span class="p p-Indicator">]</span>
<span class="p p-Indicator">-</span><span class="w"> </span><span class="p p-Indicator">[</span><span class="nv">2</span><span class="p p-Indicator">,</span><span class="w"> </span><span class="nv">0</span><span class="p p-Indicator">]</span>
<span class="p p-Indicator">-</span><span class="w"> </span><span class="p p-Indicator">[</span><span class="nv">0</span><span class="p p-Indicator">,</span><span class="w"> </span><span class="nv">5</span><span class="p p-Indicator">]</span>
<span class="p p-Indicator">-</span><span class="w"> </span><span class="p p-Indicator">[</span><span class="nv">0</span><span class="p p-Indicator">,</span><span class="w"> </span><span class="nv">0</span><span class="p p-Indicator">,</span><span class="w"> </span><span class="nv">1</span><span class="p p-Indicator">]</span>
</pre></div>
</div>
<p>To run the Medusa-enabled engine, run the following command:</p>
<div class="highlight-shell notranslate"><div class="highlight"><pre><span></span>env<span class="w"> </span><span class="nv">TRTLLM_ENABLE_PDL</span><span class="o">=</span><span class="m">1</span><span class="w"> </span><span class="se">\</span>
<span class="w"> </span><span class="nv">UB_ONESHOT</span><span class="o">=</span><span class="m">1</span><span class="w"> </span><span class="se">\</span>
<span class="w"> </span><span class="nv">UB_TP_SIZE</span><span class="o">=</span><span class="nv">$tp_size</span><span class="w"> </span><span class="se">\</span>
<span class="w"> </span><span class="nv">TRTLLM_ENABLE_PDL</span><span class="o">=</span><span class="m">1</span><span class="w"> </span><span class="se">\</span>
<span class="w"> </span><span class="nv">TRTLLM_PDL_OVERLAP_RATIO</span><span class="o">=</span><span class="m">0</span>.15<span class="w"> </span><span class="se">\</span>
<span class="w"> </span><span class="nv">TRTLLM_PREFETCH_RATIO</span><span class="o">=</span>-1<span class="w"> </span><span class="se">\</span>
<span class="w"> </span>trtllm-bench<span class="w"> </span>--model<span class="w"> </span>meta-llama/Meta-Llama-3-70B<span class="w"> </span><span class="se">\</span>
<span class="w"> </span>latency<span class="w"> </span><span class="se">\</span>
<span class="w"> </span>--dataset<span class="w"> </span><span class="nv">$DATASET_PATH</span><span class="w"> </span><span class="se">\</span>
<span class="w"> </span>--engine_dir<span class="w"> </span>/tmp/meta-llama/Meta-Llama-3-70B/medusa/engine<span class="w"> </span><span class="se">\</span>
<span class="w"> </span>--medusa_choices<span class="w"> </span>medusa.yml
</pre></div>
</div>
</section>
</section>
<section id="summary">
<h2>Summary<a class="headerlink" href="#summary" title="Link to this heading"></a></h2>
<p>The following table summarizes the commands needed for running benchmarks:</p>
<table class="docutils align-default">
<thead>
<tr class="row-odd"><th class="head"><p>Scenario</p></th>
<th class="head"><p>Phase</p></th>
<th class="head"><p>Command</p></th>
</tr>
</thead>
<tbody>
<tr class="row-even"><td><p>Dataset</p></td>
<td><p>Preparation</p></td>
<td><p><code class="docutils literal notranslate"><span class="pre">python</span> <span class="pre">benchmarks/cpp/prepare_dataset.py</span> <span class="pre">--stdout</span> <span class="pre">--tokenizer</span> <span class="pre">$HF_MODEL</span> <span class="pre">token-norm-dist</span> <span class="pre">--input-mean</span> <span class="pre">$ISL</span> <span class="pre">--output-mean</span> <span class="pre">$OSL</span> <span class="pre">--input-stdev</span> <span class="pre">0</span> <span class="pre">--output-stdev</span> <span class="pre">0</span> <span class="pre">--num-requests</span> <span class="pre">$NUM_REQUESTS</span> <span class="pre">&gt;</span> <span class="pre">$DATASET_PATH</span></code></p></td>
</tr>
<tr class="row-odd"><td><p>Throughput</p></td>
<td><p>Build</p></td>
<td><p><code class="docutils literal notranslate"><span class="pre">trtllm-bench</span> <span class="pre">--model</span> <span class="pre">$HF_MODEL</span> <span class="pre">build</span> <span class="pre">--dataset</span> <span class="pre">$DATASET_PATH</span></code></p></td>
</tr>
<tr class="row-even"><td><p>Throughput</p></td>
<td><p>Benchmark</p></td>
<td><p><code class="docutils literal notranslate"><span class="pre">trtllm-bench</span> <span class="pre">--model</span> <span class="pre">$HF_MODEL</span> <span class="pre">throughput</span> <span class="pre">--dataset</span> <span class="pre">$DATASET_PATH</span> <span class="pre">--engine_dir</span> <span class="pre">$ENGINE_DIR</span></code></p></td>
</tr>
<tr class="row-odd"><td><p>Latency</p></td>
<td><p>Build</p></td>
<td><p>See <a class="reference internal" href="#low-latency-tensorrt-llm-engine-for-llama-3-70b">section about building low latency engines</a></p></td>
</tr>
<tr class="row-even"><td><p>Non-Medusa Latency</p></td>
<td><p>Benchmark</p></td>
<td><p><code class="docutils literal notranslate"><span class="pre">trtllm-bench</span> <span class="pre">--model</span> <span class="pre">$HF_MODEL</span> <span class="pre">latency</span> <span class="pre">--dataset</span> <span class="pre">$DATASET_PATH</span> <span class="pre">--engine_dir</span> <span class="pre">$ENGINE_DIR</span></code></p></td>
</tr>
<tr class="row-odd"><td><p>Medusa Latency</p></td>
<td><p>Benchmark</p></td>
<td><p><code class="docutils literal notranslate"><span class="pre">trtllm-bench</span> <span class="pre">--model</span> <span class="pre">$HF_MODEL</span> <span class="pre">latency</span> <span class="pre">--dataset</span> <span class="pre">$DATASET_PATH</span> <span class="pre">--engine_dir</span> <span class="pre">$ENGINE_DIR</span> <span class="pre">--medusa_choices</span> <span class="pre">$MEDUSA_CHOICES</span></code></p></td>
</tr>
</tbody>
</table>
<p>where,</p>
<dl class="simple myst">
<dt><code class="docutils literal notranslate"><span class="pre">$HF_MODEL</span></code></dt><dd><p>The Hugging Face name of a model.</p>
</dd>
<dt><code class="docutils literal notranslate"><span class="pre">$NUM_REQUESTS</span></code></dt><dd><p>The number of requests to generate.</p>
</dd>
<dt><code class="docutils literal notranslate"><span class="pre">$DATASET_PATH</span></code></dt><dd><p>The path where the dataset was written when preparing the dataset.</p>
</dd>
<dt><code class="docutils literal notranslate"><span class="pre">$ENGINE_DIR</span></code></dt><dd><p>The engine directory as printed by <code class="docutils literal notranslate"><span class="pre">trtllm-bench</span> <span class="pre">build</span></code>.</p>
</dd>
<dt><code class="docutils literal notranslate"><span class="pre">$MEDUSA_CHOICES</span></code></dt><dd><p>A YAML config representing the Medusa tree for the benchmark.</p>
</dd>
</dl>
</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="performance-tuning-guide/index.html" class="btn btn-neutral float-right" title="Performance Tuning Guide" 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 0x7fa255098b30>
<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>
Reference in New Issue View Git Blame Copy Permalink