mirror of
https://github.com/NVIDIA/TensorRT-LLM.git
synced 2026-01-14 06:27:45 +08:00
427 lines
41 KiB
HTML
427 lines
41 KiB
HTML
|
||
|
||
<!DOCTYPE html>
|
||
<html class="writer-html5" lang="en" data-content_root="../">
|
||
<head>
|
||
<meta charset="utf-8" /><meta name="viewport" content="width=device-width, initial-scale=1" />
|
||
|
||
<meta name="viewport" content="width=device-width, initial-scale=1.0" />
|
||
<title>Graph Rewriting Module — 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=9bcbadda"></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="Inference Request" href="inference-request.html" />
|
||
<link rel="prev" title="Executor API" href="executor.html" />
|
||
</head>
|
||
|
||
<body class="wy-body-for-nav">
|
||
<div class="wy-grid-for-nav">
|
||
<nav data-toggle="wy-nav-shift" class="wy-nav-side">
|
||
<div class="wy-side-scroll">
|
||
<div class="wy-side-nav-search" >
|
||
|
||
|
||
|
||
<a href="../index.html" class="icon icon-home">
|
||
tensorrt_llm
|
||
</a>
|
||
<div role="search">
|
||
<form id="rtd-search-form" class="wy-form" action="../search.html" method="get">
|
||
<input type="text" name="q" placeholder="Search docs" aria-label="Search docs" />
|
||
<input type="hidden" name="check_keywords" value="yes" />
|
||
<input type="hidden" name="area" value="default" />
|
||
</form>
|
||
</div>
|
||
</div><div class="wy-menu wy-menu-vertical" data-spy="affix" role="navigation" aria-label="Navigation menu">
|
||
<p class="caption" role="heading"><span class="caption-text">Getting Started</span></p>
|
||
<ul>
|
||
<li class="toctree-l1"><a class="reference internal" href="../overview.html">Overview</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="../quick-start-guide.html">Quick Start Guide</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="../key-features.html">Key Features</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="../release-notes.html">Release Notes</a></li>
|
||
</ul>
|
||
<p class="caption" role="heading"><span class="caption-text">Installation</span></p>
|
||
<ul>
|
||
<li class="toctree-l1"><a class="reference internal" href="../installation/linux.html">Installing on Linux</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="../installation/build-from-source-linux.html">Building from Source Code on Linux</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="../installation/windows.html">Installing on Windows</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="../installation/build-from-source-windows.html">Building from Source Code on Windows</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="../installation/grace-hopper.html">Installing on Grace Hopper</a></li>
|
||
</ul>
|
||
<p class="caption" role="heading"><span class="caption-text">LLM API</span></p>
|
||
<ul>
|
||
<li class="toctree-l1"><a class="reference internal" href="../llm-api/index.html">API Introduction</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="../llm-api/reference.html">API Reference</a></li>
|
||
</ul>
|
||
<p class="caption" role="heading"><span class="caption-text">LLM API Examples</span></p>
|
||
<ul>
|
||
<li class="toctree-l1"><a class="reference internal" href="../llm-api-examples/index.html">LLM Examples Introduction</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="../llm-api-examples/customization.html">Common Customizations</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="../llm-api-examples/llm_api_examples.html">Examples</a></li>
|
||
</ul>
|
||
<p class="caption" role="heading"><span class="caption-text">Model Definition API</span></p>
|
||
<ul>
|
||
<li class="toctree-l1"><a class="reference internal" href="../python-api/tensorrt_llm.layers.html">Layers</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="../python-api/tensorrt_llm.functional.html">Functionals</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="../python-api/tensorrt_llm.models.html">Models</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="../python-api/tensorrt_llm.plugin.html">Plugin</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="../python-api/tensorrt_llm.quantization.html">Quantization</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="../python-api/tensorrt_llm.runtime.html">Runtime</a></li>
|
||
</ul>
|
||
<p class="caption" role="heading"><span class="caption-text">C++ API</span></p>
|
||
<ul>
|
||
<li class="toctree-l1"><a class="reference internal" href="../_cpp_gen/executor.html">Executor</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="../_cpp_gen/runtime.html">Runtime</a></li>
|
||
</ul>
|
||
<p class="caption" role="heading"><span class="caption-text">Command-Line Reference</span></p>
|
||
<ul>
|
||
<li class="toctree-l1"><a class="reference internal" href="../commands/trtllm-build.html">trtllm-build</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="../commands/trtllm-serve.html">trtllm-serve</a></li>
|
||
</ul>
|
||
<p class="caption" role="heading"><span class="caption-text">Architecture</span></p>
|
||
<ul>
|
||
<li class="toctree-l1"><a class="reference internal" href="../architecture/overview.html">TensorRT-LLM Architecture</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="../architecture/core-concepts.html">Model Definition</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="../architecture/core-concepts.html#compilation">Compilation</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="../architecture/core-concepts.html#runtime">Runtime</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="../architecture/core-concepts.html#multi-gpu-and-multi-node-support">Multi-GPU and Multi-Node Support</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="../architecture/checkpoint.html">TensorRT-LLM Checkpoint</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="../architecture/workflow.html">TensorRT-LLM Build Workflow</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="../architecture/add-model.html">Adding a Model</a></li>
|
||
</ul>
|
||
<p class="caption" role="heading"><span class="caption-text">Advanced</span></p>
|
||
<ul class="current">
|
||
<li class="toctree-l1"><a class="reference internal" href="gpt-attention.html">Multi-Head, Multi-Query, and Group-Query Attention</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="gpt-runtime.html">C++ GPT Runtime</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="executor.html">Executor API</a></li>
|
||
<li class="toctree-l1 current"><a class="current reference internal" href="#">Graph Rewriting Module</a><ul>
|
||
<li class="toctree-l2"><a class="reference internal" href="#when-to-use-graph-rewriting">When to Use Graph Rewriting?</a></li>
|
||
<li class="toctree-l2"><a class="reference internal" href="#graph-rewriting-apis">Graph Rewriting APIs</a><ul>
|
||
<li class="toctree-l3"><a class="reference internal" href="#tensor-related-methods">Tensor-Related Methods</a></li>
|
||
<li class="toctree-l3"><a class="reference internal" href="#flayerinfo-for-retrieving-high-level-information-for-a-functional">FLayerInfo for Retrieving High-Level Information for a Functional</a></li>
|
||
<li class="toctree-l3"><a class="reference internal" href="#pattern-and-pattern-manager">Pattern and Pattern Manager</a></li>
|
||
<li class="toctree-l3"><a class="reference internal" href="#record-signature-to-decorate-functionals-requiring-flayerinfo">@record_signature to Decorate Functionals Requiring FLayerInfo</a></li>
|
||
</ul>
|
||
</li>
|
||
<li class="toctree-l2"><a class="reference internal" href="#classical-workflow">Classical Workflow</a></li>
|
||
</ul>
|
||
</li>
|
||
<li class="toctree-l1"><a class="reference internal" href="inference-request.html">Inference Request</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="inference-request.html#responses">Responses</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="lora.html">Run gpt-2b + LoRA using GptManager / cpp runtime</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="expert-parallelism.html">Expert Parallelism in TensorRT-LLM</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="kv-cache-reuse.html">KV cache reuse</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="speculative-decoding.html">Speculative Sampling</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="disaggregated-service.html">Disaggregated-Service (experimental)</a></li>
|
||
</ul>
|
||
<p class="caption" role="heading"><span class="caption-text">Performance</span></p>
|
||
<ul>
|
||
<li class="toctree-l1"><a class="reference internal" href="../performance/perf-overview.html">Overview</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="../performance/perf-benchmarking.html">Benchmarking</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="../performance/perf-best-practices.html">Best Practices</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="../performance/perf-analysis.html">Performance Analysis</a></li>
|
||
</ul>
|
||
<p class="caption" role="heading"><span class="caption-text">Reference</span></p>
|
||
<ul>
|
||
<li class="toctree-l1"><a class="reference internal" href="../reference/troubleshooting.html">Troubleshooting</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="../reference/support-matrix.html">Support Matrix</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="../reference/precision.html">Numerical Precision</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="../reference/memory.html">Memory Usage of TensorRT-LLM</a></li>
|
||
</ul>
|
||
<p class="caption" role="heading"><span class="caption-text">Blogs</span></p>
|
||
<ul>
|
||
<li class="toctree-l1"><a class="reference internal" href="../blogs/H100vsA100.html">H100 has 4.6x A100 Performance in TensorRT-LLM, achieving 10,000 tok/s at 100ms to first token</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="../blogs/H200launch.html">H200 achieves nearly 12,000 tokens/sec on Llama2-13B with TensorRT-LLM</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="../blogs/Falcon180B-H200.html">Falcon-180B on a single H200 GPU with INT4 AWQ, and 6.7x faster Llama-70B over A100</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="../blogs/quantization-in-TRT-LLM.html">Speed up inference with SOTA quantization techniques in TRT-LLM</a></li>
|
||
<li class="toctree-l1"><a class="reference internal" href="../blogs/XQA-kernel.html">New XQA-kernel provides 2.4x more Llama-70B throughput within the same latency budget</a></li>
|
||
</ul>
|
||
|
||
</div>
|
||
</div>
|
||
</nav>
|
||
|
||
<section data-toggle="wy-nav-shift" class="wy-nav-content-wrap"><nav class="wy-nav-top" aria-label="Mobile navigation menu" >
|
||
<i data-toggle="wy-nav-top" class="fa fa-bars"></i>
|
||
<a href="../index.html">tensorrt_llm</a>
|
||
</nav>
|
||
|
||
<div class="wy-nav-content">
|
||
<div class="rst-content">
|
||
<div role="navigation" aria-label="Page navigation">
|
||
<ul class="wy-breadcrumbs">
|
||
<li><a href="../index.html" class="icon icon-home" aria-label="Home"></a></li>
|
||
<li class="breadcrumb-item active">Graph Rewriting Module</li>
|
||
<li class="wy-breadcrumbs-aside">
|
||
<a href="../_sources/advanced/graph-rewriting.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="graph-rewriting-module">
|
||
<span id="graph-rewriting"></span><h1>Graph Rewriting Module<a class="headerlink" href="#graph-rewriting-module" title="Link to this heading"></a></h1>
|
||
<p>TensorRT-LLM uses a declarative approach to define neural networks and contains
|
||
techniques to optimize the underlying graph. It provides a wrapper similar to PyTorch’s Module. When a user invokes the <code class="docutils literal notranslate"><span class="pre">forward</span></code> method, the layers are lowered to TensorRT’s <code class="docutils literal notranslate"><span class="pre">ILayer</span></code>s and become part of an <code class="docutils literal notranslate"><span class="pre">INetworkDefinition</span></code>. The Graph Rewriting (GW) module can be used to manipulate the network at the <code class="docutils literal notranslate"><span class="pre">ILayer</span></code>/<code class="docutils literal notranslate"><span class="pre">INetworkDefinition</span></code> level.</p>
|
||
<section id="when-to-use-graph-rewriting">
|
||
<h2>When to Use Graph Rewriting?<a class="headerlink" href="#when-to-use-graph-rewriting" title="Link to this heading"></a></h2>
|
||
<p>For network manipulation, there are two options in TensorRT-LLM:</p>
|
||
<ol class="arabic simple">
|
||
<li><p><strong>Module Rewriting:</strong> This method modifies the members of <code class="docutils literal notranslate"><span class="pre">Module</span></code> instances before triggering the <code class="docutils literal notranslate"><span class="pre">forward</span></code> method (that is, creating the TensorRT graph). It works on the highest level of the network representation and facilitates the modification of sequences of operations (like modifying the GEMM + activation for SmoothQuant),</p></li>
|
||
<li><p><strong>Graph Rewriting:</strong> Graph Rewriting manipulates TensorRT’s <code class="docutils literal notranslate"><span class="pre">INetworkDefinition</span></code> after the <code class="docutils literal notranslate"><span class="pre">forward</span></code> method is triggered. It operates at a finer-grained <code class="docutils literal notranslate"><span class="pre">ILayer</span></code> level and can alter the structure across multiple Module instances. It is typically used for layer fusion.</p></li>
|
||
</ol>
|
||
<p>Graph Rewriting (GW) is ideally used in the following conditions:</p>
|
||
<ol class="arabic simple">
|
||
<li><p>When only <code class="docutils literal notranslate"><span class="pre">ILayer</span></code>/<code class="docutils literal notranslate"><span class="pre">INetworkDefinition</span></code> is available,</p></li>
|
||
<li><p>When Module Rewriting would lead to nested control flow or scattered functionality.</p></li>
|
||
</ol>
|
||
</section>
|
||
<section id="graph-rewriting-apis">
|
||
<h2>Graph Rewriting APIs<a class="headerlink" href="#graph-rewriting-apis" title="Link to this heading"></a></h2>
|
||
<p>Several core APIs are provided for Graph Rewriting:</p>
|
||
<section id="tensor-related-methods">
|
||
<h3>Tensor-Related Methods<a class="headerlink" href="#tensor-related-methods" title="Link to this heading"></a></h3>
|
||
<ul class="simple">
|
||
<li><p><code class="docutils literal notranslate"><span class="pre">Tensor.get_parent</span></code>: Get the <code class="docutils literal notranslate"><span class="pre">ILayer</span></code> that produces this tensor,</p></li>
|
||
<li><p><code class="docutils literal notranslate"><span class="pre">Tensor.get_users</span></code>: Get the consumer <code class="docutils literal notranslate"><span class="pre">ILayer</span></code>s of this tensor,</p></li>
|
||
<li><p><code class="docutils literal notranslate"><span class="pre">replace_all_uses_with</span></code>: Replace this tensor with another tensor in all consumer <code class="docutils literal notranslate"><span class="pre">ILayer</span></code>s.</p></li>
|
||
</ul>
|
||
</section>
|
||
<section id="flayerinfo-for-retrieving-high-level-information-for-a-functional">
|
||
<h3>FLayerInfo for Retrieving High-Level Information for a Functional<a class="headerlink" href="#flayerinfo-for-retrieving-high-level-information-for-a-functional" title="Link to this heading"></a></h3>
|
||
<p>For all the layers located in <code class="docutils literal notranslate"><span class="pre">functional.py</span></code>, the original input information is missing once lowered to <code class="docutils literal notranslate"><span class="pre">INetworkDefinition</span></code>, especially for TensorRT plugins, which are opaque in the Python world. <code class="docutils literal notranslate"><span class="pre">FLayerInfo</span></code> holds their original information as a high-level signature containing inputs like <code class="docutils literal notranslate"><span class="pre">Tensor</span></code>, Python attributes, and more. There is a Network-wise singleton called <code class="docutils literal notranslate"><span class="pre">FLayerInfoMemo</span></code> to map each <code class="docutils literal notranslate"><span class="pre">ILayer</span></code> to its corresponding <code class="docutils literal notranslate"><span class="pre">FLayerInfo</span></code>.</p>
|
||
<p>For <code class="docutils literal notranslate"><span class="pre">FLayerInfo</span></code>:</p>
|
||
<ul class="simple">
|
||
<li><p><code class="docutils literal notranslate"><span class="pre">FLayerInfo.replace_input_with</span></code>: Replace some input tensor with another tensor,</p></li>
|
||
<li><p><code class="docutils literal notranslate"><span class="pre">FLayerInfo.replace_output_uses_with</span></code>: Redirect the usage of the original output tensors to a set of new tensors.</p></li>
|
||
</ul>
|
||
<p>For <code class="docutils literal notranslate"><span class="pre">FLayerInfoMemo</span></code>:</p>
|
||
<ul class="simple">
|
||
<li><p><code class="docutils literal notranslate"><span class="pre">FLayerInfoMemo.instance()</span></code>: Get the singleton instance,</p></li>
|
||
<li><p><code class="docutils literal notranslate"><span class="pre">FLayerInfoMemo.get</span></code>: Get the corresponding <code class="docutils literal notranslate"><span class="pre">FLayerInfo</span></code> for an <code class="docutils literal notranslate"><span class="pre">ILayer</span></code>.</p></li>
|
||
</ul>
|
||
<p><code class="docutils literal notranslate"><span class="pre">FLayerInfo</span></code> remains consistent with the actual <code class="docutils literal notranslate"><span class="pre">ILayer</span></code> during GW, making it safe to use.</p>
|
||
</section>
|
||
<section id="pattern-and-pattern-manager">
|
||
<h3>Pattern and Pattern Manager<a class="headerlink" href="#pattern-and-pattern-manager" title="Link to this heading"></a></h3>
|
||
<p>There are two kinds of patterns:</p>
|
||
<ul class="simple">
|
||
<li><p><code class="docutils literal notranslate"><span class="pre">PatternRewriter</span></code>: Used for defining a rewriting pattern, which actually alters the network.</p>
|
||
<ul>
|
||
<li><p><code class="docutils literal notranslate"><span class="pre">match</span></code>: Match the pattern; returns true if a layer is matched,</p></li>
|
||
<li><p><code class="docutils literal notranslate"><span class="pre">rewrite</span></code>: Manipulate a layer,</p></li>
|
||
<li><p><code class="docutils literal notranslate"><span class="pre">match_and_rewrite</span></code>: Combines both <code class="docutils literal notranslate"><span class="pre">match</span></code> and <code class="docutils literal notranslate"><span class="pre">rewrite</span></code>, used for complex states that need to pass from <code class="docutils literal notranslate"><span class="pre">match</span></code> to <code class="docutils literal notranslate"><span class="pre">rewrite</span></code>.</p></li>
|
||
</ul>
|
||
</li>
|
||
<li><p><code class="docutils literal notranslate"><span class="pre">PatternAnalyzer</span></code>: Used for defining an analysis pattern, which collects information from the network.</p>
|
||
<ul>
|
||
<li><p><code class="docutils literal notranslate"><span class="pre">match</span></code>: Match the pattern,</p></li>
|
||
<li><p><code class="docutils literal notranslate"><span class="pre">analyze</span></code>: Perform analysis on a list of layers.</p></li>
|
||
</ul>
|
||
</li>
|
||
</ul>
|
||
<p>There are two managers for managing multiple <code class="docutils literal notranslate"><span class="pre">PatternRewriter</span></code> or <code class="docutils literal notranslate"><span class="pre">PatternAnalyzer</span></code>:</p>
|
||
<ul class="simple">
|
||
<li><p><code class="docutils literal notranslate"><span class="pre">RewritePatternManager</span></code>:</p>
|
||
<ul>
|
||
<li><p><code class="docutils literal notranslate"><span class="pre">add</span></code>: Add a pattern with its label and benefit; the benefit specifies its privilege,</p></li>
|
||
<li><p><code class="docutils literal notranslate"><span class="pre">get</span></code>: Get a pattern by label,</p></li>
|
||
<li><p><code class="docutils literal notranslate"><span class="pre">rewrite</span></code>: Apply the rewriting patterns contained to a network.</p></li>
|
||
</ul>
|
||
</li>
|
||
<li><p><code class="docutils literal notranslate"><span class="pre">AnalysisPatternManager</span></code>:</p>
|
||
<ul>
|
||
<li><p><code class="docutils literal notranslate"><span class="pre">add</span></code>: Add a pattern with its label and benefit; the benefit specifies its privilege,</p></li>
|
||
<li><p><code class="docutils literal notranslate"><span class="pre">get</span></code>: Get a pattern by label,</p></li>
|
||
<li><p><code class="docutils literal notranslate"><span class="pre">analyze</span></code>: Apply the analysis patterns contained to a network.</p></li>
|
||
</ul>
|
||
</li>
|
||
</ul>
|
||
</section>
|
||
<section id="record-signature-to-decorate-functionals-requiring-flayerinfo">
|
||
<h3>@record_signature to Decorate Functionals Requiring FLayerInfo<a class="headerlink" href="#record-signature-to-decorate-functionals-requiring-flayerinfo" title="Link to this heading"></a></h3>
|
||
<p>The <code class="docutils literal notranslate"><span class="pre">@record_signature</span></code> decorator is used to record the <code class="docutils literal notranslate"><span class="pre">FLayerInfo</span></code> for a functional. While FLayerInfo is vital for GW when analyzing or rewriting certain functionals, it is used in an “add as needed” manner. If you are adding GW patterns, ensure that the functional requires the <code class="docutils literal notranslate"><span class="pre">@record_signature</span></code> decorator.</p>
|
||
</section>
|
||
</section>
|
||
<section id="classical-workflow">
|
||
<h2>Classical Workflow<a class="headerlink" href="#classical-workflow" title="Link to this heading"></a></h2>
|
||
<p>There are specific routines for defining a GW pattern. Let’s start with a simple example: replacing a sum layer with a subtract layer, which can also be found in the <code class="docutils literal notranslate"><span class="pre">test_graph_rewriting.py</span></code> file.</p>
|
||
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="k">class</span> <span class="nc">NaivePatternRewriter_ReplaceAddWithSub</span><span class="p">(</span><span class="n">PatternRewriter</span><span class="p">):</span>
|
||
|
||
<span class="k">def</span> <span class="fm">__init__</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
|
||
<span class="nb">super</span><span class="p">()</span><span class="o">.</span><span class="fm">__init__</span><span class="p">(</span><span class="s1">'replace_add_with_sub'</span><span class="p">,</span>
|
||
<span class="n">root_layer</span><span class="o">=</span><span class="p">{</span><span class="n">trt</span><span class="o">.</span><span class="n">LayerType</span><span class="o">.</span><span class="n">ELEMENTWISE</span><span class="p">},</span>
|
||
<span class="n">separate_match_rewrite</span><span class="o">=</span><span class="kc">True</span><span class="p">)</span>
|
||
|
||
<span class="k">def</span> <span class="nf">match</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">layer</span><span class="p">:</span> <span class="n">Layer</span><span class="p">):</span>
|
||
<span class="c1"># The rewriter will stop at the first matched layer, and then the Rewriter will enter the rewrite() to do the rewriting.</span>
|
||
<span class="k">return</span> <span class="n">layer</span><span class="o">.</span><span class="n">as_layer</span><span class="p">()</span><span class="o">.</span><span class="n">op</span> <span class="o">==</span> <span class="n">trt</span><span class="o">.</span><span class="n">ElementWiseOperation</span><span class="o">.</span><span class="n">SUM</span>
|
||
|
||
<span class="k">def</span> <span class="nf">rewrite</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">layer</span><span class="p">:</span> <span class="n">Layer</span><span class="p">)</span> <span class="o">-></span> <span class="kc">None</span><span class="p">:</span>
|
||
<span class="c1"># The layer here should be an Elementwise_SUM layer.</span>
|
||
<span class="k">with</span> <span class="n">net_guard</span><span class="p">(</span><span class="n">layer</span><span class="o">.</span><span class="n">network</span><span class="p">):</span>
|
||
<span class="c1"># There are several stages to replace some subgraph with another subgraph:</span>
|
||
|
||
<span class="c1"># Stage 1: Get the input tensors and output tensors of the subgraph to replace.</span>
|
||
<span class="c1"># - For Elementwise_SUM, there are two inputs and one output.</span>
|
||
<span class="n">a</span><span class="p">,</span> <span class="n">b</span> <span class="o">=</span> <span class="n">layer</span><span class="o">.</span><span class="n">get_inputs</span><span class="p">(</span><span class="mi">0</span><span class="p">,</span> <span class="mi">1</span><span class="p">)</span>
|
||
<span class="n">o</span> <span class="o">=</span> <span class="n">layer</span><span class="o">.</span><span class="n">get_outputs</span><span class="p">(</span><span class="mi">0</span><span class="p">)[</span><span class="mi">0</span><span class="p">]</span>
|
||
|
||
<span class="c1"># Stage 2: Create a new subgraph that takes the old one's inputs.</span>
|
||
<span class="c1"># - Here we insert an Elementwise_SUB layer, and 'c' is the output.</span>
|
||
<span class="n">c</span> <span class="o">=</span> <span class="n">a</span> <span class="o">-</span> <span class="n">b</span>
|
||
|
||
<span class="c1"># Stage 3: Redirect all the layers depending on the outputs of the old subgraph to the new subgraph's.</span>
|
||
<span class="c1"># - After this, the SUM becomes dangling and will be pruned by TensorRT when building the engine.</span>
|
||
<span class="c1"># - Note that there is no API in TensorRT python to remove a layer explicitly; `replace_all_uses_with` is the only way to "remove" a layer.</span>
|
||
<span class="n">o</span><span class="o">.</span><span class="n">replace_all_uses_with</span><span class="p">(</span><span class="n">c</span><span class="p">)</span>
|
||
|
||
<span class="c1"># Stage 4: Mark all the layers in the old subgraph as removed.</span>
|
||
<span class="c1"># - This helps the PatternRewriter to skip the removed layers.</span>
|
||
<span class="n">layer</span><span class="o">.</span><span class="n">mark_as_removed</span><span class="p">()</span>
|
||
</pre></div>
|
||
</div>
|
||
<p>In this example, we deal with <code class="docutils literal notranslate"><span class="pre">ILayer</span></code> rather than Plugins, so <code class="docutils literal notranslate"><span class="pre">FLayerInfo</span></code> is unnecessary. As illustrated in the <code class="docutils literal notranslate"><span class="pre">rewrite</span></code> method, there are four stages that are shared across nearly all rewrite patterns.</p>
|
||
<p>Note that in GW, we <strong>NEVER</strong> rewrite a layer directly. Instead, we do it in two steps: first, create another layer with the same input and deprive all the users of the original outputs, redirecting them to the outputs of the new layers. In this way, the old layer will be dangling and pruned automatically by TensorRT during the engine building phase. This is a limitation of TensorRT since remove-layer-like APIs are not available in Python.</p>
|
||
<p>In Stage 2, we rely on operators and layers commonly used during the network building phase. Ideally, you can replace them with any network structure during GW.</p>
|
||
<p>For the usage of <code class="docutils literal notranslate"><span class="pre">FLayerInfo</span></code>, let’s rewrite the <code class="docutils literal notranslate"><span class="pre">gpt_attention</span></code> to enable the <code class="docutils literal notranslate"><span class="pre">remove-padding</span></code> feature. <code class="docutils literal notranslate"><span class="pre">gpt_attention</span></code> is actually</p>
|
||
<p>a TensorRT plugin, so we need <code class="docutils literal notranslate"><span class="pre">FLayerInfo</span></code> to hold the original Tensor-wise inputs to help create new <code class="docutils literal notranslate"><span class="pre">gpt_attention</span></code> layers.</p>
|
||
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="k">class</span> <span class="nc">GPTAttentionPluginRemovePaddingRewritePass</span><span class="p">(</span><span class="n">PatternRewriter</span><span class="p">):</span>
|
||
|
||
<span class="k">def</span> <span class="fm">__init__</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
|
||
<span class="nb">super</span><span class="p">()</span><span class="o">.</span><span class="fm">__init__</span><span class="p">(</span><span class="s1">'gpt_attention_plugin_remove_padding'</span><span class="p">,</span>
|
||
<span class="n">root_layer</span><span class="o">=</span><span class="p">{</span><span class="n">trt</span><span class="o">.</span><span class="n">LayerType</span><span class="o">.</span><span class="n">PLUGIN_V2</span><span class="p">})</span>
|
||
|
||
<span class="k">def</span> <span class="nf">match_and_rewrite</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">layer</span><span class="p">:</span> <span class="n">Layer</span><span class="p">)</span> <span class="o">-></span> <span class="nb">bool</span><span class="p">:</span>
|
||
<span class="k">if</span> <span class="n">layer</span><span class="o">.</span><span class="n">as_layer</span><span class="p">()</span><span class="o">.</span><span class="n">type</span> <span class="o">!=</span> <span class="n">trt</span><span class="o">.</span><span class="n">LayerType</span><span class="o">.</span><span class="n">PLUGIN_V2</span> <span class="ow">or</span> \
|
||
<span class="n">layer</span><span class="o">.</span><span class="n">as_layer</span><span class="p">()</span><span class="o">.</span><span class="n">plugin</span><span class="o">.</span><span class="n">plugin_namespace</span> <span class="o">!=</span> <span class="s1">'tensorrt_llm'</span> <span class="ow">or</span> \
|
||
<span class="n">layer</span><span class="o">.</span><span class="n">as_layer</span><span class="p">()</span><span class="o">.</span><span class="n">plugin</span><span class="o">.</span><span class="n">plugin_type</span> <span class="o">!=</span> <span class="s1">'GPTAttention'</span><span class="p">:</span>
|
||
<span class="k">return</span> <span class="kc">False</span>
|
||
|
||
<span class="c1"># Retrieve the FLayerInfo</span>
|
||
<span class="n">flayer</span> <span class="o">=</span> <span class="n">FLayerInfoMemo</span><span class="o">.</span><span class="n">instance</span><span class="p">()</span><span class="o">.</span><span class="n">get</span><span class="p">(</span><span class="n">layer</span><span class="o">.</span><span class="n">name</span><span class="p">)</span>
|
||
<span class="k">assert</span> <span class="n">flayer</span>
|
||
<span class="c1"># Although the layer is a plugin, which is a black box, we get some high-level input information from the FLayerInfo.</span>
|
||
<span class="n">tensor_input</span><span class="p">:</span> <span class="n">Tensor</span> <span class="o">=</span> <span class="n">flayer</span><span class="o">.</span><span class="n">get_input</span><span class="p">(</span><span class="s1">'qkv'</span><span class="p">)</span>
|
||
<span class="k">if</span> <span class="n">tensor_input</span><span class="o">.</span><span class="n">shape</span><span class="p">[</span><span class="mi">0</span><span class="p">]</span> <span class="o">==</span> <span class="mi">1</span><span class="p">:</span> <span class="c1"># Already in remove-padding mode</span>
|
||
<span class="k">return</span> <span class="kc">False</span>
|
||
|
||
<span class="c1"># Some information could be passed in from external</span>
|
||
<span class="k">assert</span> <span class="bp">self</span><span class="o">.</span><span class="n">args</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">,</span> <span class="s2">"args should be passed in from RewritePatternManager.rewrite()"</span>
|
||
<span class="n">batch_size</span><span class="p">,</span> <span class="n">in_len</span><span class="p">,</span> <span class="n">hidden_size</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">args</span><span class="p">[</span><span class="s1">'batch_size'</span><span class="p">],</span> <span class="bp">self</span><span class="o">.</span><span class="n">args</span><span class="p">[</span><span class="s1">'in_len'</span><span class="p">],</span> <span class="bp">self</span><span class="o">.</span><span class="n">args</span><span class="p">[</span><span class="s1">'hidden_size'</span><span class="p">]</span>
|
||
|
||
<span class="k">with</span> <span class="n">net_guard</span><span class="p">(</span><span class="n">layer</span><span class="o">.</span><span class="n">network</span><span class="p">):</span>
|
||
<span class="n">new_inputs</span> <span class="o">=</span> <span class="n">flayer</span><span class="o">.</span><span class="n">clone_inputs</span><span class="p">()</span>
|
||
|
||
<span class="c1"># Step 1: Create new inputs and replace the original arglist.</span>
|
||
<span class="nb">input</span> <span class="o">=</span> <span class="n">Tensor</span><span class="p">(</span>
|
||
<span class="n">name</span><span class="o">=</span><span class="s1">'qkv'</span><span class="p">,</span>
|
||
<span class="n">dtype</span><span class="o">=</span><span class="n">trt</span><span class="o">.</span><span class="n">float16</span><span class="p">,</span>
|
||
<span class="n">shape</span><span class="o">=</span><span class="p">(</span><span class="mi">1</span><span class="p">,</span> <span class="n">batch_size</span> <span class="o">*</span> <span class="n">in_len</span><span class="p">,</span> <span class="n">hidden_size</span><span class="p">),</span>
|
||
<span class="p">)</span>
|
||
<span class="n">new_inputs</span><span class="p">[</span><span class="s1">'qkv'</span><span class="p">]</span> <span class="o">=</span> <span class="nb">input</span>
|
||
|
||
<span class="c1"># Step 2: Create a new plugin instance.</span>
|
||
<span class="n">new_outs</span> <span class="o">=</span> <span class="n">gpt_attention</span><span class="p">(</span><span class="o">**</span><span class="n">new_inputs</span><span class="p">)</span>
|
||
|
||
<span class="c1"># Step 3: Deprive all the users of the old plugin instance.</span>
|
||
<span class="n">flayer</span><span class="o">.</span><span class="n">replace_outputs_uses_with</span><span class="p">(</span><span class="n">layer</span><span class="o">.</span><span class="n">network</span><span class="p">,</span> <span class="n">new_outs</span><span class="p">)</span>
|
||
|
||
<span class="c1"># Step 4: Remove the old plugin instance.</span>
|
||
<span class="n">layer</span><span class="o">.</span><span class="n">mark_as_removed</span><span class="p">()</span>
|
||
|
||
<span class="k">return</span> <span class="kc">True</span>
|
||
</pre></div>
|
||
</div>
|
||
<p>This is quite similar to the first example, with the focus on the <code class="docutils literal notranslate"><span class="pre">FLayerInfo</span></code> part. Through the code below, we can get the original inputs of this layer, enabling us to alter the inputs related to remove-padding and create a new layer to replace it.</p>
|
||
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="n">flayer</span> <span class="o">=</span> <span class="n">FLayerInfoMemo</span><span class="o">.</span><span class="n">instance</span><span class="p">()</span><span class="o">.</span><span class="n">get</span><span class="p">(</span><span class="n">layer</span><span class="o">.</span><span class="n">name</span><span class="p">)</span>
|
||
<span class="k">assert</span> <span class="n">flayer</span>
|
||
</pre></div>
|
||
</div>
|
||
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="n">new_inputs</span> <span class="o">=</span> <span class="n">flayer</span><span class="o">.</span><span class="n">clone_inputs</span><span class="p">()</span>
|
||
|
||
<span class="c1"># Step 1: Create new inputs and replace the original arglist.</span>
|
||
<span class="nb">input</span> <span class="o">=</span> <span class="n">Tensor</span><span class="p">(</span>
|
||
<span class="n">name</span><span class="o">=</span><span class="s1">'tensor'</span><span class="p">,</span>
|
||
<span class="n">dtype</span><span class="o">=</span><span class="n">trt</span><span class="o">.</span><span class="n">float16</span><span class="p">,</span>
|
||
<span class="n">shape</span><span class="o">=</span><span class="p">(</span><span class="mi">1</span><span class="p">,</span> <span class="n">batch_size</span> <span class="o">*</span> <span class="n">in_len</span><span class="p">,</span> <span class="n">hidden_size</span><span class="p">),</span>
|
||
<span class="p">)</span>
|
||
<span class="n">new_inputs</span><span class="p">[</span><span class="s1">'tensor'</span><span class="p">]</span> <span class="o">=</span> <span class="nb">input</span>
|
||
|
||
<span class="c1"># Step 2: Create a new plugin instance.</span>
|
||
<span class="n">new_outs</span> <span class="o">=</span> <span class="n">gpt_attention</span><span class="p">(</span><span class="o">**</span><span class="n">new_inputs</span><span class="p">)</span>
|
||
</pre></div>
|
||
</div>
|
||
<p>For real examples, please refer to the <code class="docutils literal notranslate"><span class="pre">FuseAttentionWithBiasPass</span></code> in the <code class="docutils literal notranslate"><span class="pre">graph_rewriting.py</span></code>.</p>
|
||
</section>
|
||
</section>
|
||
|
||
|
||
</div>
|
||
</div>
|
||
<footer><div class="rst-footer-buttons" role="navigation" aria-label="Footer">
|
||
<a href="executor.html" class="btn btn-neutral float-left" title="Executor API" accesskey="p" rel="prev"><span class="fa fa-arrow-circle-left" aria-hidden="true"></span> Previous</a>
|
||
<a href="inference-request.html" class="btn btn-neutral float-right" title="Inference Request" 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 0x7f1ac799a000>
|
||
|
||
<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> |