mirror of
https://github.com/NVIDIA/TensorRT-LLM.git
synced 2026-01-14 06:27:45 +08:00
404 lines
35 KiB
HTML
404 lines
35 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>Run gpt-2b + LoRA using GptManager / cpp runtime — 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=888ff710"></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="Expert Parallelism in TensorRT-LLM" href="expert-parallelism.html" />
|
|
<link rel="prev" title="Inference Request" href="inference-request.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"><a class="reference internal" href="graph-rewriting.html">Graph Rewriting Module</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="inference-request.html">Inference Request</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="inference-request.html#responses">Responses</a></li>
|
|
<li class="toctree-l1 current"><a class="current reference internal" href="#">Run gpt-2b + LoRA using GptManager / cpp runtime</a><ul>
|
|
<li class="toctree-l2"><a class="reference internal" href="#lora-tensor-format-details">LoRA tensor format details</a><ul>
|
|
<li class="toctree-l3"><a class="reference internal" href="#example-lora-tensors">Example LoRA tensors</a></li>
|
|
<li class="toctree-l3"><a class="reference internal" href="#lora-module-id-mapping">LoRA Module id mapping</a></li>
|
|
<li class="toctree-l3"><a class="reference internal" href="#loracache-configuration">LoraCache configuration</a></li>
|
|
<li class="toctree-l3"><a class="reference internal" href="#lora-with-tensor-parallel">LoRA with tensor parallel</a></li>
|
|
</ul>
|
|
</li>
|
|
</ul>
|
|
</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>
|
|
</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">Run gpt-2b + LoRA using GptManager / cpp runtime</li>
|
|
<li class="wy-breadcrumbs-aside">
|
|
<a href="../_sources/advanced/lora.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="run-gpt-2b-lora-using-gptmanager-cpp-runtime">
|
|
<span id="lora"></span><h1>Run gpt-2b + LoRA using GptManager / cpp runtime<a class="headerlink" href="#run-gpt-2b-lora-using-gptmanager-cpp-runtime" title="Link to this heading"></a></h1>
|
|
<p>First build a model with LoRA and inflight-batching enabled.</p>
|
|
<div class="highlight-bash notranslate"><div class="highlight"><pre><span></span>git-lfs<span class="w"> </span>clone<span class="w"> </span>https://huggingface.co/qychen/luotuo-lora-7b-0.1
|
|
git-lfs<span class="w"> </span>clone<span class="w"> </span>https://huggingface.co/kunishou/Japanese-Alpaca-LoRA-7b-v0
|
|
<span class="nv">BASE_MODEL</span><span class="o">=</span>llama-7b-hf
|
|
|
|
python<span class="w"> </span>examples/llama/convert_checkpoint.py<span class="w"> </span>--model_dir<span class="w"> </span><span class="si">${</span><span class="nv">BASE_MODEL</span><span class="si">}</span><span class="w"> </span><span class="se">\</span>
|
|
<span class="w"> </span>--output_dir<span class="w"> </span>/tmp/llama_7b/trt_ckpt/fp16/1-gpu/<span class="w"> </span><span class="se">\</span>
|
|
<span class="w"> </span>--dtype<span class="w"> </span>float16
|
|
|
|
trtllm-build<span class="w"> </span>--checkpoint_dir<span class="w"> </span>/tmp/llama_7b/trt_ckpt/fp16/1-gpu/<span class="w"> </span><span class="se">\</span>
|
|
<span class="w"> </span>--output_dir<span class="w"> </span>/tmp/llama_7b_with_lora_qkv/trt_engines/fp16/1-gpu/<span class="w"> </span><span class="se">\</span>
|
|
<span class="w"> </span>--remove_input_padding<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>float16<span class="w"> </span><span class="se">\</span>
|
|
<span class="w"> </span>--context_fmha<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>--gemm_plugin<span class="w"> </span>float16<span class="w"> </span><span class="se">\</span>
|
|
<span class="w"> </span>--lora_plugin<span class="w"> </span>float16<span class="w"> </span><span class="se">\</span>
|
|
<span class="w"> </span>--max_batch_size<span class="w"> </span><span class="m">128</span><span class="w"> </span><span class="se">\</span>
|
|
<span class="w"> </span>--max_input_len<span class="w"> </span><span class="m">512</span><span class="w"> </span><span class="se">\</span>
|
|
<span class="w"> </span>--max_seq_len<span class="w"> </span><span class="m">562</span><span class="w"> </span><span class="se">\</span>
|
|
<span class="w"> </span>--lora_dir<span class="w"> </span>Japanese-Alpaca-LoRA-7b-v0<span class="w"> </span><span class="se">\</span>
|
|
<span class="w"> </span>--max_lora_rank<span class="w"> </span><span class="m">8</span><span class="w"> </span><span class="se">\</span>
|
|
<span class="w"> </span>--lora_target_modules<span class="w"> </span><span class="s2">"attn_q"</span><span class="w"> </span><span class="s2">"attn_k"</span><span class="w"> </span><span class="s2">"attn_v"</span>
|
|
</pre></div>
|
|
</div>
|
|
<p>To pass LoRAs into the cpp runtime they must be converted to the format below.
|
|
The script below will convert a Hugging Face LoRA model to the correct NumPy tensor.</p>
|
|
<div class="highlight-bash notranslate"><div class="highlight"><pre><span></span>python3<span class="w"> </span>tensorrt_llm/examples/hf_lora_convert.py<span class="w"> </span>-i<span class="w"> </span>Japanese-Alpaca-LoRA-7b-v0<span class="w"> </span>-o<span class="w"> </span>Japanese-Alpaca-LoRA-7b-v0-weights<span class="w"> </span>--storage-type<span class="w"> </span>float16
|
|
python3<span class="w"> </span>tensorrt_llm/examples/hf_lora_convert.py<span class="w"> </span>-i<span class="w"> </span>luotuo-lora-7b-0.1<span class="w"> </span>-o<span class="w"> </span>luotuo-lora-7b-0.1-weights<span class="w"> </span>--storage-type<span class="w"> </span>float16
|
|
</pre></div>
|
|
</div>
|
|
<p>Refer to the <a class="reference external" href="https://github.com/triton-inference-server/tensorrtllm_backend/blob/main/docs/lora.md">tensorrtllm_backend documentation</a> for a Multi-LoRA example using Triton.</p>
|
|
<section id="lora-tensor-format-details">
|
|
<h2>LoRA tensor format details<a class="headerlink" href="#lora-tensor-format-details" title="Link to this heading"></a></h2>
|
|
<p>To run inference with <code class="docutils literal notranslate"><span class="pre">LoraWeights</span></code> using <code class="docutils literal notranslate"><span class="pre">GptManager</span></code>, <code class="docutils literal notranslate"><span class="pre">InferenceRequests</span></code> must have <code class="docutils literal notranslate"><span class="pre">LoraWeights</span></code> (<code class="docutils literal notranslate"><span class="pre">lora_weights</span></code>) and <code class="docutils literal notranslate"><span class="pre">LoraConfig</span></code> (<code class="docutils literal notranslate"><span class="pre">lora_config</span></code>) parameters.</p>
|
|
<p><code class="docutils literal notranslate"><span class="pre">LoraTaskId</span></code> the unique task ID for the given LoRA.</p>
|
|
<p>To perform inference with a specific LoRA for the first time, <code class="docutils literal notranslate"><span class="pre">lora_task_id</span></code>, <code class="docutils literal notranslate"><span class="pre">lora_weights</span></code>, and <code class="docutils literal notranslate"><span class="pre">lora_config</span></code> must all be given. The LoRA will be cached, so that subsequent requests for the same task only require <code class="docutils literal notranslate"><span class="pre">lora_task_id</span></code>.
|
|
If the cache is full, the oldest LoRA will be evicted to make space for new ones. An error is returned if <code class="docutils literal notranslate"><span class="pre">lora_task_id</span></code> is not cached.</p>
|
|
<p><code class="docutils literal notranslate"><span class="pre">LoraWeights</span></code> contains the weights for all the LoRAs. Currently, this should include weights for all TP and PP ranks.
|
|
The weights tensor has the shape <code class="docutils literal notranslate"><span class="pre">[num_lora_modules_layers,</span> <span class="pre">D</span> <span class="pre">x</span> <span class="pre">Hi</span> <span class="pre">+</span> <span class="pre">Ho</span> <span class="pre">x</span> <span class="pre">D</span> <span class="pre">]</span></code>. The last dimension holds the in / out adapter weights for the associated module (for example, <code class="docutils literal notranslate"><span class="pre">attn_qkv</span></code>) and model layer.</p>
|
|
<p>Each of the in / out tensors are first flattened and then concatenated together in the format above.
|
|
The first dimension (of size <code class="docutils literal notranslate"><span class="pre">num_lora_module_layers</span></code>) has an entry for each module-layer (that is, there is an entry for <code class="docutils literal notranslate"><span class="pre">attn_q</span> <span class="pre">layer1</span></code> and another for <code class="docutils literal notranslate"><span class="pre">attn_k</span> <span class="pre">layer1</span></code>).</p>
|
|
<p><code class="docutils literal notranslate"><span class="pre">D=adapter_size</span> <span class="pre">(i.e.</span> <span class="pre">R</span> <span class="pre">value),</span> <span class="pre">Hi=hidden_size_in,</span> <span class="pre">Ho=hidden_size_out.</span></code></p>
|
|
<p><code class="docutils literal notranslate"><span class="pre">LoraConfig</span></code> is a configuration tensor which identifies the moduleId, layerId, and adapter size of each element of <code class="docutils literal notranslate"><span class="pre">LoraWeights</span></code>. It has the shape <code class="docutils literal notranslate"><span class="pre">[num_lora_modules_layers,</span> <span class="pre">3]</span></code>. The last dimension holds <code class="docutils literal notranslate"><span class="pre">[module_id,</span> <span class="pre">layer_idx,</span> <span class="pre">adapter_size</span> <span class="pre">D</span> <span class="pre">(i.e.</span> <span class="pre">R</span> <span class="pre">value)]</span></code>.</p>
|
|
<p>This feature supports LoRAs as described in https://arxiv.org/pdf/2106.09685.pdf</p>
|
|
<section id="example-lora-tensors">
|
|
<h3>Example LoRA tensors<a class="headerlink" href="#example-lora-tensors" title="Link to this heading"></a></h3>
|
|
<p>Here is an example of <code class="docutils literal notranslate"><span class="pre">LoraWeights</span></code> and <code class="docutils literal notranslate"><span class="pre">LoraConfig</span></code> tensors for a model with tp=1, pp=1, 4 layers, and a hidden size of 4.
|
|
The following tensors are for a LoRA which has a <code class="docutils literal notranslate"><span class="pre">q</span></code> and <code class="docutils literal notranslate"><span class="pre">k</span></code> adapter.</p>
|
|
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="c1"># loraConfig</span>
|
|
<span class="p">[</span>
|
|
<span class="p">[</span><span class="mi">1</span><span class="p">,</span> <span class="mi">0</span><span class="p">,</span> <span class="mi">2</span><span class="p">]</span>
|
|
<span class="p">[</span><span class="mi">2</span><span class="p">,</span> <span class="mi">0</span><span class="p">,</span> <span class="mi">4</span><span class="p">]</span>
|
|
<span class="p">[</span><span class="mi">1</span><span class="p">,</span> <span class="mi">1</span><span class="p">,</span> <span class="mi">2</span><span class="p">]</span>
|
|
<span class="p">[</span><span class="mi">2</span><span class="p">,</span> <span class="mi">1</span><span class="p">,</span> <span class="mi">4</span><span class="p">]</span>
|
|
<span class="p">[</span><span class="mi">1</span><span class="p">,</span> <span class="mi">2</span><span class="p">,</span> <span class="mi">2</span><span class="p">]</span> <span class="c1"># Note that the final 2 layers only adapt `q`</span>
|
|
<span class="p">[</span><span class="mi">1</span><span class="p">,</span> <span class="mi">3</span><span class="p">,</span> <span class="mi">8</span><span class="p">]</span>
|
|
<span class="p">]</span>
|
|
<span class="c1"># Note: The loraConfig tensor configures the loraWeights tensor.</span>
|
|
<span class="c1"># The contents of each row of loraWeights is specified be the corresponding row in loraConfig</span>
|
|
|
|
<span class="c1"># loraWeights</span>
|
|
<span class="c1"># Note: that 'in weights' and 'out weights' are 'A' and 'B' in the LoRA paper.</span>
|
|
<span class="p">[</span>
|
|
<span class="p">[</span> <span class="o"><</span><span class="mi">2</span> <span class="n">x</span> <span class="mi">4</span> <span class="ow">in</span> <span class="n">weights</span><span class="o">></span><span class="p">,</span> <span class="o"><</span><span class="mi">4</span> <span class="n">x</span> <span class="mi">2</span> <span class="n">out</span> <span class="n">weights</span><span class="o">></span> <span class="o"><</span><span class="n">padding</span><span class="o">></span> <span class="p">]</span> <span class="c1"># `q` adapter for layer 0</span>
|
|
<span class="p">[</span> <span class="o"><</span><span class="mi">4</span> <span class="n">x</span> <span class="mi">4</span> <span class="ow">in</span> <span class="n">weights</span><span class="o">></span><span class="p">,</span> <span class="o"><</span><span class="mi">4</span> <span class="n">x</span> <span class="mi">4</span> <span class="n">out</span> <span class="n">weights</span><span class="o">></span> <span class="o"><</span><span class="n">padding</span><span class="o">></span> <span class="p">]</span> <span class="c1"># `k` adapter for layer 0</span>
|
|
<span class="p">[</span> <span class="o"><</span><span class="mi">2</span> <span class="n">x</span> <span class="mi">4</span> <span class="ow">in</span> <span class="n">weights</span><span class="o">></span><span class="p">,</span> <span class="o"><</span><span class="mi">4</span> <span class="n">x</span> <span class="mi">2</span> <span class="n">out</span> <span class="n">weights</span><span class="o">></span> <span class="o"><</span><span class="n">padding</span><span class="o">></span> <span class="p">]</span> <span class="c1"># `q` adapter for layer 1</span>
|
|
<span class="p">[</span> <span class="o"><</span><span class="mi">4</span> <span class="n">x</span> <span class="mi">4</span> <span class="ow">in</span> <span class="n">weights</span><span class="o">></span><span class="p">,</span> <span class="o"><</span><span class="mi">4</span> <span class="n">x</span> <span class="mi">4</span> <span class="n">out</span> <span class="n">weights</span><span class="o">></span> <span class="o"><</span><span class="n">padding</span><span class="o">></span> <span class="p">]</span> <span class="c1"># `k` adapter for layer 1</span>
|
|
<span class="p">[</span> <span class="o"><</span><span class="mi">2</span> <span class="n">x</span> <span class="mi">4</span> <span class="ow">in</span> <span class="n">weights</span><span class="o">></span><span class="p">,</span> <span class="o"><</span><span class="mi">4</span> <span class="n">x</span> <span class="mi">2</span> <span class="n">out</span> <span class="n">weights</span><span class="o">></span> <span class="o"><</span><span class="n">padding</span><span class="o">></span> <span class="p">]</span> <span class="c1"># `q` adapter for layer 2</span>
|
|
<span class="p">[</span> <span class="o"><</span><span class="mi">8</span> <span class="n">x</span> <span class="mi">4</span> <span class="ow">in</span> <span class="n">weights</span><span class="o">></span><span class="p">,</span> <span class="o"><</span><span class="mi">4</span> <span class="n">x</span> <span class="mi">8</span> <span class="n">out</span> <span class="n">weights</span><span class="o">></span> <span class="p">]</span> <span class="c1"># `q` adapter for layer 3. Note the final layer has a adapter size of 8</span>
|
|
<span class="p">]</span>
|
|
|
|
</pre></div>
|
|
</div>
|
|
</section>
|
|
<section id="lora-module-id-mapping">
|
|
<h3>LoRA Module id mapping<a class="headerlink" href="#lora-module-id-mapping" title="Link to this heading"></a></h3>
|
|
<table class="docutils align-default">
|
|
<thead>
|
|
<tr class="row-odd"><th class="head"><p>module name (as specified in <code class="docutils literal notranslate"><span class="pre">convert_checkpoint.py</span></code> scripts)</p></th>
|
|
<th class="head"><p>module id</p></th>
|
|
<th class="head"><p>description</p></th>
|
|
</tr>
|
|
</thead>
|
|
<tbody>
|
|
<tr class="row-even"><td><p>attn_qkv</p></td>
|
|
<td><p>0</p></td>
|
|
<td><p>compbined qkv adapter</p></td>
|
|
</tr>
|
|
<tr class="row-odd"><td><p>attn_q</p></td>
|
|
<td><p>1</p></td>
|
|
<td><p>q adapter</p></td>
|
|
</tr>
|
|
<tr class="row-even"><td><p>attn_k</p></td>
|
|
<td><p>2</p></td>
|
|
<td><p>k adapter</p></td>
|
|
</tr>
|
|
<tr class="row-odd"><td><p>attn_v</p></td>
|
|
<td><p>3</p></td>
|
|
<td><p>v adapter</p></td>
|
|
</tr>
|
|
<tr class="row-even"><td><p>attn_dense</p></td>
|
|
<td><p>4</p></td>
|
|
<td><p>adapter for the dense layer in attention</p></td>
|
|
</tr>
|
|
<tr class="row-odd"><td><p>mlp_h_to_4h</p></td>
|
|
<td><p>5</p></td>
|
|
<td><p>for llama2 adapter for gated mlp layer after attention / RMSNorm: up projection</p></td>
|
|
</tr>
|
|
<tr class="row-even"><td><p>mlp_4h_to_h</p></td>
|
|
<td><p>6</p></td>
|
|
<td><p>for llama2 adapter for gated mlp layer after attention / RMSNorm: down projection</p></td>
|
|
</tr>
|
|
<tr class="row-odd"><td><p>mlp_gate</p></td>
|
|
<td><p>7</p></td>
|
|
<td><p>for llama2 adapter for gated mlp later after attention / RMSNorm: gate</p></td>
|
|
</tr>
|
|
<tr class="row-even"><td><p>cross_attn_qkv</p></td>
|
|
<td><p>8</p></td>
|
|
<td><p>compbined qkv adapter for cross attention</p></td>
|
|
</tr>
|
|
<tr class="row-odd"><td><p>cross_attn_q</p></td>
|
|
<td><p>9</p></td>
|
|
<td><p>q adapter for cross attention</p></td>
|
|
</tr>
|
|
<tr class="row-even"><td><p>cross_attn_k</p></td>
|
|
<td><p>10</p></td>
|
|
<td><p>k adapter for cross attention</p></td>
|
|
</tr>
|
|
<tr class="row-odd"><td><p>cross_attn_v</p></td>
|
|
<td><p>11</p></td>
|
|
<td><p>v adapter for cross attention</p></td>
|
|
</tr>
|
|
<tr class="row-even"><td><p>cross_attn_dense</p></td>
|
|
<td><p>12</p></td>
|
|
<td><p>adapter for the dense layer in cross attention</p></td>
|
|
</tr>
|
|
<tr class="row-odd"><td><p>moe_h_to_4h</p></td>
|
|
<td><p>13</p></td>
|
|
<td><p>for mixtral adapter for expert mlp layer: up projection</p></td>
|
|
</tr>
|
|
<tr class="row-even"><td><p>moe_4h_to_h</p></td>
|
|
<td><p>14</p></td>
|
|
<td><p>for mixtral adapter for expert mlp layer: down projection</p></td>
|
|
</tr>
|
|
<tr class="row-odd"><td><p>moe_gate</p></td>
|
|
<td><p>15</p></td>
|
|
<td><p>for mixtral adapter for expert mlp layer: gate</p></td>
|
|
</tr>
|
|
<tr class="row-even"><td><p>moe_router</p></td>
|
|
<td><p>16</p></td>
|
|
<td><p>for mixtral adapter for expert router layer</p></td>
|
|
</tr>
|
|
<tr class="row-odd"><td><p>mlp_router</p></td>
|
|
<td><p>17</p></td>
|
|
<td><p>for qwen2-moe adapter for shared expert gate layer</p></td>
|
|
</tr>
|
|
</tbody>
|
|
</table>
|
|
</section>
|
|
<section id="loracache-configuration">
|
|
<h3>LoraCache configuration<a class="headerlink" href="#loracache-configuration" title="Link to this heading"></a></h3>
|
|
<p>The core idea is that we will have a fixed size, 2-level LoRA cache in TRT-LLM. The higher level cache resides on the host and the lower level is on GPU (distinct from the existing KV cache). Sizes of both are user configurable.</p>
|
|
<p>The CPU cache is configured to be a max size. The GPU cache is configured to a percentage of free GPU memory after engine load. As requests come in LoRAs are stored in the host cache.</p>
|
|
<p>As requests are scheduled for execution LoRAs are loaded into the GPU cache.</p>
|
|
</section>
|
|
<section id="lora-with-tensor-parallel">
|
|
<h3>LoRA with tensor parallel<a class="headerlink" href="#lora-with-tensor-parallel" title="Link to this heading"></a></h3>
|
|
<p>The partition of tensor parallel for LoRA is special. There are two cases: <code class="docutils literal notranslate"><span class="pre">RowLinear</span></code> and <code class="docutils literal notranslate"><span class="pre">ColumnLinear</span></code>. Assume we have a linear layer and the input feature size is <code class="docutils literal notranslate"><span class="pre">K</span></code> and the output feature size is <code class="docutils literal notranslate"><span class="pre">N</span></code>. Then, the shape of the weight is <code class="docutils literal notranslate"><span class="pre">[K,</span> <span class="pre">N]</span></code>.</p>
|
|
<p>First, consider this linear layer is a <code class="docutils literal notranslate"><span class="pre">ColumnLinear</span></code> layer. When we partition the weight, we split the weight by column with <code class="docutils literal notranslate"><span class="pre">tp_size</span></code>. Then, there are <code class="docutils literal notranslate"><span class="pre">tp_size</span></code> split weights and the shapes of these weights are <code class="docutils literal notranslate"><span class="pre">[K,</span> <span class="pre">N</span> <span class="pre">//</span> <span class="pre">tp_size]</span></code>. When we apply LoRA adapter on such <code class="docutils literal notranslate"><span class="pre">ColumnLinear</span></code> layer, the shapes of original two weights are <code class="docutils literal notranslate"><span class="pre">[K,</span> <span class="pre">lora_rank]</span></code> and <code class="docutils literal notranslate"><span class="pre">[lora_rank,</span> <span class="pre">N]</span></code>. So, we only partition the second weight and get <code class="docutils literal notranslate"><span class="pre">tp_size</span></code> split weights with shapes <code class="docutils literal notranslate"><span class="pre">[lora_rank,</span> <span class="pre">N</span> <span class="pre">//</span> <span class="pre">tp_size]</span></code>. For the first weight, each GPU maintains the same entire weight (with shape <code class="docutils literal notranslate"><span class="pre">[K,</span> <span class="pre">lora_rank]</span></code>).</p>
|
|
<p>Next, consider this linear layer is a <code class="docutils literal notranslate"><span class="pre">RowLinear</span></code> layer. When we partition the weight, we split the weight by row with <code class="docutils literal notranslate"><span class="pre">tp_size</span></code>. Then, there are <code class="docutils literal notranslate"><span class="pre">tp_size</span></code> split weights and the shapes of these weights are <code class="docutils literal notranslate"><span class="pre">[K</span> <span class="pre">//</span> <span class="pre">tp_size,</span> <span class="pre">N]</span></code>. When we apply LoRA adapter on such <code class="docutils literal notranslate"><span class="pre">RowLinear</span></code> layer, the shapes of original two weights are <code class="docutils literal notranslate"><span class="pre">[K,</span> <span class="pre">lora_rank]</span></code> and <code class="docutils literal notranslate"><span class="pre">[lora_rank,</span> <span class="pre">N]</span></code>. So, we only partition the first weight and get <code class="docutils literal notranslate"><span class="pre">tp_size</span></code> split weights with shapes <code class="docutils literal notranslate"><span class="pre">[K</span> <span class="pre">//</span> <span class="pre">tp_size,</span> <span class="pre">lora_rank]</span></code>. For the second weight, each GPU maintains the same entire weight (with shape <code class="docutils literal notranslate"><span class="pre">[lora_rank,</span> <span class="pre">N]</span></code>).</p>
|
|
</section>
|
|
</section>
|
|
</section>
|
|
|
|
|
|
</div>
|
|
</div>
|
|
<footer><div class="rst-footer-buttons" role="navigation" aria-label="Footer">
|
|
<a href="inference-request.html" class="btn btn-neutral float-left" title="Inference Request" accesskey="p" rel="prev"><span class="fa fa-arrow-circle-left" aria-hidden="true"></span> Previous</a>
|
|
<a href="expert-parallelism.html" class="btn btn-neutral float-right" title="Expert Parallelism in TensorRT-LLM" 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 0x7fed9c653a10>
|
|
|
|
<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> |