kanshan/TensorRT-LLMs
1
0
Fork 0
mirror of https://github.com/NVIDIA/TensorRT-LLM.git synced 2026-01-14 06:27:45 +08:00
Code Issues Actions 1 Packages Projects Releases Wiki Activity
40924d2a42
TensorRT-LLMs/gpt_attention.html
石晓伟 40924d2a42
Update GitHub pages (#544)
2023-12-04 18:59:41 +08:00

421 lines
33 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>Multi-head, Multi-query and Group-query Attention &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=19f00094" />
<!--[if lt IE 9]>
<script src="_static/js/html5shiv.min.js"></script>
<![endif]-->
<script src="_static/jquery.js?v=5d32c60e"></script>
<script src="_static/_sphinx_javascript_frameworks_compat.js?v=2cd50e6c"></script>
<script src="_static/documentation_options.js?v=5929fcd5"></script>
<script src="_static/doctools.js?v=888ff710"></script>
<script src="_static/sphinx_highlight.js?v=dc90522c"></script>
<script src="_static/js/theme.js"></script>
<link rel="index" title="Index" href="genindex.html" />
<link rel="search" title="Search" href="search.html" />
<link rel="next" title="Numerical Precision" href="precision.html" />
<link rel="prev" title="The Batch Manager in TensorRT-LLM" href="batch_manager.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">Contents:</span></p>
<ul class="current">
<li class="toctree-l1"><a class="reference internal" href="architecture.html">TensorRT-LLM Architecture</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="batch_manager.html">The Batch Manager in TensorRT-LLM</a></li>
<li class="toctree-l1 current"><a class="current reference internal" href="#">Multi-head, Multi-query and Group-query Attention</a><ul>
<li class="toctree-l2"><a class="reference internal" href="#important-note">Important Note</a></li>
<li class="toctree-l2"><a class="reference internal" href="#padded-and-packed-tensors">Padded and Packed Tensors</a></li>
<li class="toctree-l2"><a class="reference internal" href="#context-and-generation-phases">Context and Generation Phases</a><ul>
<li class="toctree-l3"><a class="reference internal" href="#context-phase">Context Phase</a></li>
<li class="toctree-l3"><a class="reference internal" href="#generation-phase">Generation Phase</a></li>
</ul>
</li>
<li class="toctree-l2"><a class="reference internal" href="#inflight-batching">Inflight batching</a></li>
<li class="toctree-l2"><a class="reference internal" href="#kv-cache-s">KV Cache(s)</a><ul>
<li class="toctree-l3"><a class="reference internal" href="#contiguous-kv-cache">Contiguous KV Cache</a></li>
<li class="toctree-l3"><a class="reference internal" href="#paged-kv-cache">Paged KV Cache</a></li>
</ul>
</li>
<li class="toctree-l2"><a class="reference internal" href="#int8-fp8-kv-caches">INT8/FP8 KV Caches</a></li>
<li class="toctree-l2"><a class="reference internal" href="#sliding-window-attention-cyclic-rolling-buffer-kv-cache">Sliding Window Attention, Cyclic (Rolling Buffer) KV Cache</a></li>
<li class="toctree-l2"><a class="reference internal" href="#beam-search">Beam-Search</a></li>
<li class="toctree-l2"><a class="reference internal" href="#input-qkv-tensor">Input QKV tensor</a></li>
<li class="toctree-l2"><a class="reference internal" href="#additional-features">Additional Features</a><ul>
<li class="toctree-l3"><a class="reference internal" href="#rotary-positional-embedding-rope">Rotary Positional Embedding (RoPE)</a></li>
<li class="toctree-l3"><a class="reference internal" href="#alibi">ALiBi</a></li>
<li class="toctree-l3"><a class="reference internal" href="#scaling-factor-s">Scaling factor(s)</a></li>
<li class="toctree-l3"><a class="reference internal" href="#cross-attention">Cross Attention</a></li>
<li class="toctree-l3"><a class="reference internal" href="#relative-attention-bias-rab">Relative Attention Bias (RAB)</a></li>
</ul>
</li>
</ul>
</li>
<li class="toctree-l1"><a class="reference internal" href="precision.html">Numerical Precision</a></li>
<li class="toctree-l1"><a class="reference internal" href="installation.html">Build TensorRT-LLM</a></li>
<li class="toctree-l1"><a class="reference internal" href="performance.html">Performance of TensorRT-LLM</a></li>
<li class="toctree-l1"><a class="reference internal" href="2023-05-19-how-to-debug.html">How to debug</a></li>
<li class="toctree-l1"><a class="reference internal" href="2023-05-17-how-to-add-a-new-model.html">How to add a new model</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="memory.html">Memory Usage of TensorRT-LLM</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">Python 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/runtime.html">Runtime</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">Blogs</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="blogs/H100vsA100.html">H100 has 4.6x A100 Performance in TensorRT-LLM, achieving 10,000 tok/s at 100ms to first token</a></li>
<li class="toctree-l1"><a class="reference internal" href="blogs/H200launch.html">H200 achieves nearly 12,000 tokens/sec on Llama2-13B with TensorRT-LLM</a></li>
</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">Multi-head, Multi-query and Group-query Attention</li>
<li class="wy-breadcrumbs-aside">
<a href="_sources/gpt_attention.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="multi-head-multi-query-and-group-query-attention">
<h1>Multi-head, Multi-query and Group-query Attention<a class="headerlink" href="#multi-head-multi-query-and-group-query-attention" title="Link to this heading"></a></h1>
<p>This document details the implementation of multihead attention (MHA),
multiquery attention (MQA) and group-query attention (GQA) for auto-regressive
GPT-like models in TensorRT-LLM. As a quick reminder, the multihead attention
is the sequence of a batched matmul, a softmax and another batched matmul
described in the
<a class="reference external" href="https://arxiv.org/abs/1706.03762">Attention Is All You Need</a> article.
Multi-query Attention (MQA) [<a class="reference external" href="https://arxiv.org/abs/1911.02150">https://arxiv.org/abs/1911.02150</a>]
Group-query Attention (GQA) [<a class="reference external" href="https://arxiv.org/abs/2307.09288">https://arxiv.org/abs/2307.09288</a>]
are variants of MHA that use fewer, so-called, K/V head than the number of
query heads. TensorRT-LLM, MHA, MQA and GQA are implemented by the operator
<a class="reference external" href="https://github.com/NVIDIA/TensorRT-LLM/tree/rel/tensorrt_llm/functional.py"><code class="docutils literal notranslate"><span class="pre">tensorrt_llm.functional.gpt_attention</span></code></a>.</p>
<section id="important-note">
<h2>Important Note<a class="headerlink" href="#important-note" title="Link to this heading"></a></h2>
<p>As discussed below, the current implementation supports two input modes: Padded
and packed (non-padded). As the packed mode is always more memory-efficient and
faster than the padded mode, <em><strong>support for padded mode may be removed in the
future</strong></em>.</p>
</section>
<section id="padded-and-packed-tensors">
<h2>Padded and Packed Tensors<a class="headerlink" href="#padded-and-packed-tensors" title="Link to this heading"></a></h2>
<p>In TensorRT-LLM, the GPT attention operator supports two different types
of QKV inputs: Padded and packed (i.e. non padded) inputs. The mode is
determined by the global configuration parameter <code class="docutils literal notranslate"><span class="pre">remove_input_padding</span></code> defined
in <a class="reference external" href="https://github.com/NVIDIA/TensorRT-LLM/tree/rel/tensorrt_llm/plugin/plugin.py"><code class="docutils literal notranslate"><span class="pre">tensorrt_llm.plugin</span></code></a>.</p>
<p>When padding is enabled (i.e. <code class="docutils literal notranslate"><span class="pre">remove_input_padding</span></code> is <code class="docutils literal notranslate"><span class="pre">False</span></code>), the sequences
that are shorter than the <code class="docutils literal notranslate"><span class="pre">max_sequence_length</span></code> are padded to that maximum
length. It may result in excessive memory consumption as well as unneeded
computations on padding tokens (in the various matrix multiplications that
surround the MHA block).</p>
<p>To overcome that problem, TensorRT-LLM supports a mode without padding where
the different tokens are packed together and the user provides the operator
with a 1D tensor containing the lengths of the different sequences. It is
recommended that users to always use packed mode (and support for the padded
mode may be removed in the future).</p>
</section>
<section id="context-and-generation-phases">
<h2>Context and Generation Phases<a class="headerlink" href="#context-and-generation-phases" title="Link to this heading"></a></h2>
<p>The GPT attention operator encapsulates different implementations for both
context and generation phases in auto-regressive models like GPT.</p>
<section id="context-phase">
<h3>Context Phase<a class="headerlink" href="#context-phase" title="Link to this heading"></a></h3>
<p>If the <code class="docutils literal notranslate"><span class="pre">context_fmha_type</span></code> is set to <code class="docutils literal notranslate"><span class="pre">disabled</span></code> (see
<a class="reference external" href="https://github.com/NVIDIA/TensorRT-LLM/tree/rel/tensorrt_llm/plugin/plugin.py"><code class="docutils literal notranslate"><span class="pre">tensorrt_llm.plugin</span></code></a>),
the implementation maps to a sequence of GPU kernels that will store the
intermediate <code class="docutils literal notranslate"><span class="pre">Q*K^T</span></code> tensor in memory before calling the softmax operator. It
is the slowest method and the memory footprint is significant (quadratically
depends on the sequence length).</p>
<p>Otherwise, if <code class="docutils literal notranslate"><span class="pre">context_fmha_type</span></code> is set to a <code class="docutils literal notranslate"><span class="pre">enabled</span></code> or
<code class="docutils literal notranslate"><span class="pre">enabled_with_fp32_acc</span></code> (accumulation in the first batched matmul is forced to
FP32), that function will trigger a kernel that performs the MHA/MQA block
using a single kernel. For short sequences, that kernel uses a vanilla
implementation of MHA/MQA. For larger sequences, this kernel uses the Flash
Attention algorithm as described in
<a class="reference external" href="https://arxiv.org/abs/2205.14135">https://arxiv.org/abs/2205.14135</a>
and
<a class="reference external" href="https://arxiv.org/abs/2307.08691">https://arxiv.org/abs/2307.08691</a>.</p>
<p>Currently, the implementation triggers extra kernels that apply pre-processing
to the elements (like RoPE) and populate the KV cache (see below). In a future
release, the number of such kernels is planned on being reduced in order to
improve the overall performance.</p>
</section>
<section id="generation-phase">
<h3>Generation Phase<a class="headerlink" href="#generation-phase" title="Link to this heading"></a></h3>
<p>The generation phase is implemented using a single kernel, called the masked
multihead attention in TensorRT-LLM. That kernel is able to apply
pre-processing on the Q, K and V elements on-the-fly: Add the QKV bias, apply
RoPE, do dequantization/quantization. TensorRT-LLM will continue to add (or
enable) additional features in future releases. For example, enable the support
for ALiBi or IA3.</p>
<p><em>The masked MHA kernel has a special version that distributes the work across
multiple CUDA thread-blocks on the GPU for cases where the GPU occupancy is
low. That mode called multi-block can be enabled using the <code class="docutils literal notranslate"><span class="pre">multi_block_mode</span></code>
flag. Users are recommended to test that mode in scenarios where both the batch
size and the number of heads in the model are relatively small. The exact
definition of small in that context will depend on the model of the GPU and is
hard to predict but to provide with a rule of thumb, it is worth testing that
mode when <code class="docutils literal notranslate"><span class="pre">batch_size</span> <span class="pre">*</span> <span class="pre">num_heads</span></code> is less than the number of multi-processors
on the GPU (that suggestion may evolve in the future as more research is
conducted and the software improves)</em>.</p>
<p><em>Note that even if the multi-block mode is enabled, the attention operator will
not immediately trigger the multi-block version of the GPU kernel. There is a
minimum number of tokens (input + generated) that are required for the
multi-block version to become more efficient than the “vanilla” implementation
that uses a single CUDA thread-block per head. It is controlled by an internal
heuristic.</em></p>
</section>
</section>
<section id="inflight-batching">
<h2>Inflight batching<a class="headerlink" href="#inflight-batching" title="Link to this heading"></a></h2>
<p>TensorRT-LLM supports a feature called in-flight batching. With that feature,
sequences in context phase can be processed together with sequences in
generation phase. The purpose of that technique is to better interleave
requests to reduce latency as well as make better use the of the GPUs.
For efficiency reasons (1), the support for inflight batching <em><strong>requires the
input tensors to be packed (no padding)</strong></em>.</p>
<p><em><strong>In the current implementation, the sequences that are going through the
context phase must be before the sequences in the generation phase in the input
tensor. For example, for sequences <code class="docutils literal notranslate"><span class="pre">S0</span></code>, <code class="docutils literal notranslate"><span class="pre">S1</span></code> and <code class="docutils literal notranslate"><span class="pre">S2</span></code>, if <code class="docutils literal notranslate"><span class="pre">S0</span></code> and <code class="docutils literal notranslate"><span class="pre">S2</span></code> are in
context phase (and <code class="docutils literal notranslate"><span class="pre">S1</span></code> in generation), tokens from <code class="docutils literal notranslate"><span class="pre">S0</span></code> and <code class="docutils literal notranslate"><span class="pre">S2</span></code> must appear
before the tokens of <code class="docutils literal notranslate"><span class="pre">S1</span></code> in the input tensor</strong></em>. The constraint may or may not
be relaxed in a future version.</p>
<p><em>(1) Padding sequences in the generation phase, that contain a single token, to
the length of the maximum input sequence is inefficient use of resources</em>.</p>
</section>
<section id="kv-cache-s">
<h2>KV Cache(s)<a class="headerlink" href="#kv-cache-s" title="Link to this heading"></a></h2>
<p>In the generation phase, a common optimization is to provide the MHA kernel
with a cache containing the values of the past K and V elements that have
already been computed. That cache is known as the KV cache. TensorRT-LLM uses
that technique to accelerate its generation phase. In TensorRT-LLM, there is
one KV cache per Transformer layer, which means that there are as many KV
caches as layers in a model. The current version of TensorRT-LLM supports two
different types of KV caches: <strong>contiguous</strong> and <strong>paged</strong> KV caches.</p>
<section id="contiguous-kv-cache">
<h3>Contiguous KV Cache<a class="headerlink" href="#contiguous-kv-cache" title="Link to this heading"></a></h3>
<p>The contiguous KV cache is a monolithic tensor. Its shape is:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="p">[</span><span class="n">max_batch_size</span> <span class="o">*</span> <span class="n">max_beam_width</span><span class="p">,</span> <span class="mi">2</span><span class="p">,</span> <span class="n">num_heads</span><span class="p">,</span> <span class="n">max_seqlen</span><span class="p">,</span> <span class="n">hidden_dim_per_head</span><span class="p">]</span><span class="o">.</span>
</pre></div>
</div>
<p>That implementation uses a lot more memory than needed when the sequences are
shorter than the maximum sequence length (even if they end up close to the
limit after the generation of many output tokens, it may take a lot of steps to
reach that point).</p>
</section>
<section id="paged-kv-cache">
<h3>Paged KV Cache<a class="headerlink" href="#paged-kv-cache" title="Link to this heading"></a></h3>
<p>The paged KV cache decomposes the KV cache into blocks that are distributed to
the different requests by a cache manager during processing. That cache manager
keeps track of the sequences, allocate new blocks from a pool and recycle those
blocks when required. See the simplified implementation of
<a class="reference external" href="https://github.com/NVIDIA/TensorRT-LLM/tree/rel/tensorrt_llm/runtime/kv_cache_manager.py"><code class="docutils literal notranslate"><span class="pre">tensorrt_llm.runtime.KVCacheManager</span></code></a>.
A more efficient C++ implementation is included in the
<a class="reference external" href="https://github.com/NVIDIA/TensorRT-LLM/tree/rel/cpp/include/tensorrt_llm/batch_manager">Batch Manager</a>.</p>
</section>
</section>
<section id="int8-fp8-kv-caches">
<h2>INT8/FP8 KV Caches<a class="headerlink" href="#int8-fp8-kv-caches" title="Link to this heading"></a></h2>
<p>In its current implementation, even if the rest of the network runs in INT8 or
FP8, the GPT attention operator works with FP32, FP16, and BFloat16 inputs and
outputs. However, TensorRT-LLM supports INT8 and FP8
(<code class="docutils literal notranslate"><span class="pre">kv_cache_quant_mode=QuantMode.INT8_KV_CACHE</span></code> and
<code class="docutils literal notranslate"><span class="pre">kv_cache_quant_mode=QuantMode.FP8_KV_CACHE</span></code>) KV caches.</p>
<p>The GPT attention operator populates the KV cache. When INT8 or FP8 KV caches
are enabled, the input values have to be quantized to 8 bits using a scaling
factor. For quantization, the scaling factor is stored in the
<code class="docutils literal notranslate"><span class="pre">kv_orig_quant_scale</span></code> tensor. Its shape is <code class="docutils literal notranslate"><span class="pre">[1]</span></code> and only per-tensor
quantization is supported in the current version.</p>
<p>During generation, the values read from the cache are dequantized on-the-fly in
the MHA/MQA kernel. The scaling factor to dequantize those values is stored in
the <code class="docutils literal notranslate"><span class="pre">kv_quant_orig_scale</span></code> tensor. That tensor contains a single value (per
tensor scaling).</p>
</section>
<section id="sliding-window-attention-cyclic-rolling-buffer-kv-cache">
<h2>Sliding Window Attention, Cyclic (Rolling Buffer) KV Cache<a class="headerlink" href="#sliding-window-attention-cyclic-rolling-buffer-kv-cache" title="Link to this heading"></a></h2>
<p>TensorRT-LLM has a feature called <code class="docutils literal notranslate"><span class="pre">Cyclic</span> <span class="pre">KV</span> <span class="pre">Cache</span></code>, which treats the kv cache
as a circular buffer. This means that it only stores the kv cache for the last N
tokens, where N is determined by the <code class="docutils literal notranslate"><span class="pre">max_kv_cache_length</span></code> parameter in
<code class="docutils literal notranslate"><span class="pre">GenerationSession.setup</span></code>. You can see examples of this in the <code class="docutils literal notranslate"><span class="pre">run.py</span></code> or
<code class="docutils literal notranslate"><span class="pre">summarize.py</span></code> files. When the cache is full, new tokens’ kv cache will
overwrite the “least recently used” caches.</p>
<p>In the context phase, if the input length surpasses the <code class="docutils literal notranslate"><span class="pre">max_kv_cache_length</span></code>,
<code class="docutils literal notranslate"><span class="pre">Sliding</span> <span class="pre">Window</span> <span class="pre">Attention</span></code> will be activated. This serves the same function as
the <code class="docutils literal notranslate"><span class="pre">sliding</span> <span class="pre">window_size</span></code>.</p>
<p>This feature helps to reduce the memory footprint of the kv cache when
dealing with very long sequences.</p>
<p><em>Note that when using beam search, cyclic kv cache may not perform as well as
full kv cache when the current step exceeds <code class="docutils literal notranslate"><span class="pre">max_kv_cache_length</span></code>.
This issue will be addressed in future releases.</em></p>
<p><em>The experimental feature, which allows different <code class="docutils literal notranslate"><span class="pre">max_kv_cache_length</span></code> values
for each layer, is also supported. To utilize this feature, simply provide an
<code class="docutils literal notranslate"><span class="pre">int32</span> <span class="pre">torch.Tensor</span></code> with a shape of <code class="docutils literal notranslate"><span class="pre">[num_layers]</span></code> to the <code class="docutils literal notranslate"><span class="pre">GenerationSession.setup</span></code>.
This tensor will serve as the buffer for <code class="docutils literal notranslate"><span class="pre">max_kv_cache_length</span></code>,
setting unique values for each layer. However, it’s important to note that the
memory allocation for the kv cache still relies on the buffer’s maximum value.</em></p>
</section>
<section id="beam-search">
<h2>Beam-Search<a class="headerlink" href="#beam-search" title="Link to this heading"></a></h2>
<p>The GPT attention operator supports beam-search. In the context phase, a single
beam is computed per input sequence. In the generation phase, the MHA/MQA/GQA
kernel uses an additional tensor to reconstruct the correct path for each beam.
That tensor is called the <code class="docutils literal notranslate"><span class="pre">cache_indirection</span></code>. Its shape is <code class="docutils literal notranslate"><span class="pre">[batch_size,</span> <span class="pre">beam_width,</span> <span class="pre">max_seqlen]</span></code>.</p>
<p>For a sequence <code class="docutils literal notranslate"><span class="pre">si</span></code>, a beam <code class="docutils literal notranslate"><span class="pre">bi</span></code> and a token <code class="docutils literal notranslate"><span class="pre">ti</span></code>, the element
<code class="docutils literal notranslate"><span class="pre">cache_indirection[si][bi][ti]</span></code> is an integer between <code class="docutils literal notranslate"><span class="pre">0</span></code> and <code class="docutils literal notranslate"><span class="pre">beam_width-1</span></code>
that indicates which path in the beam to read the K and V elements from in the
KV cache. This tensor is populated in the sampling stage.</p>
</section>
<section id="input-qkv-tensor">
<h2>Input QKV tensor<a class="headerlink" href="#input-qkv-tensor" title="Link to this heading"></a></h2>
<p>The input QKV tensor packs the Q, K and V tensors (concatenated along the last
dimension) after the projection of the hidden states. It is a 3D tensor. RoPE
and quantization to INT8 or FP8 (when needed) are performed by the GPT
attention operator.</p>
<p>In padded mode, its shape is <code class="docutils literal notranslate"><span class="pre">[batch_beam_size,</span> <span class="pre">max_seqlen,</span> <span class="pre">3</span> <span class="pre">*</span> <span class="pre">hidden_dim]</span></code>
where <code class="docutils literal notranslate"><span class="pre">batch_beam_size</span></code> is the batch size (number of sequences) for the context
phase and the batch size multiplied by the beam width for the generation phase.
Having different beam widths per sequence in padded mode is not supported.</p>
<p>In packed mode, its shape is <code class="docutils literal notranslate"><span class="pre">[1,</span> <span class="pre">num_tokens,</span> <span class="pre">3</span> <span class="pre">*</span> <span class="pre">hidden_dim]</span></code> where
<code class="docutils literal notranslate"><span class="pre">num_tokens</span></code> is the total number of tokens in the batch. For the sequences in
context phase, the number of tokens of a sequence corresponds to its input
length (even if the beam width is greater than <code class="docutils literal notranslate"><span class="pre">1</span></code> for beam search). For the
sequences in generation phase, there are <code class="docutils literal notranslate"><span class="pre">beam_width</span></code> tokens per sequence. The
beam width can be different for each sequence.</p>
<p>In other words, the pseudo-code to compute the number of tokens is:</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="n">num_tokens</span> <span class="o">=</span> <span class="mi">0</span>
<span class="c1"># Add the length of each sequence in context phase.</span>
<span class="k">for</span> <span class="n">seq</span> <span class="ow">in</span> <span class="n">context_phase</span><span class="p">:</span>
<span class="n">num_tokens</span> <span class="o">+=</span> <span class="n">seq</span><span class="o">.</span><span class="n">length</span>
<span class="c1"># Add the width of the beam for each sequence in generation phase.</span>
<span class="k">for</span> <span class="n">seq</span> <span class="ow">in</span> <span class="n">generation_phase</span><span class="p">:</span>
<span class="n">num_tokens</span> <span class="o">+=</span> <span class="n">seq</span><span class="o">.</span><span class="n">beam_width</span>
</pre></div>
</div>
<p>In a future release of TensorRT-LLM, the rank of that packed input tensor
may be reduced from 3 to 2. The current rank is to maintain the homogeneity
between padded and packed modes. It is no longer justified if support for
padded mode is removed.</p>
</section>
<section id="additional-features">
<h2>Additional Features<a class="headerlink" href="#additional-features" title="Link to this heading"></a></h2>
<section id="rotary-positional-embedding-rope">
<h3>Rotary Positional Embedding (RoPE)<a class="headerlink" href="#rotary-positional-embedding-rope" title="Link to this heading"></a></h3>
<p>The GPT attention operation can perform the computation of the Rotary
Positional Embedding (RoPE). When that operation is enabled,
<code class="docutils literal notranslate"><span class="pre">rotary_embedding_dim</span></code> is set to a value greater than 0, it is fused with other
operations. The GPT operator supports GPT-NeoX and GPT-J forms of RoPE by
setting <code class="docutils literal notranslate"><span class="pre">position_embedding_type</span></code> to <code class="docutils literal notranslate"><span class="pre">PositionEmbeddingType.rope_gpt_neox</span></code>
or <code class="docutils literal notranslate"><span class="pre">PositionEmbeddingType.rope_gptj</span></code>.</p>
</section>
<section id="alibi">
<h3>ALiBi<a class="headerlink" href="#alibi" title="Link to this heading"></a></h3>
<p>The GPT attention operator can apply ALiBi to the result of the <code class="docutils literal notranslate"><span class="pre">Q*K^T</span></code>
product. The bias is computed on-the-fly from the ALiBi slopes in the optimized
kernel.</p>
</section>
<section id="scaling-factor-s">
<h3>Scaling factor(s)<a class="headerlink" href="#scaling-factor-s" title="Link to this heading"></a></h3>
<p>In MHA, the output of the <code class="docutils literal notranslate"><span class="pre">Q*K^T</span></code> product is scaled by a constant value that
is computed as:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="n">scaling</span> <span class="o">=</span> <span class="mf">1.</span><span class="n">f</span> <span class="o">/</span> <span class="p">(</span><span class="n">q_scaling</span> <span class="o">*</span> <span class="n">sqrt</span><span class="p">(</span><span class="n">head_size</span><span class="p">))</span><span class="o">.</span>
</pre></div>
</div>
</section>
<section id="cross-attention">
<h3>Cross Attention<a class="headerlink" href="#cross-attention" title="Link to this heading"></a></h3>
<p>On top of the MHA as self attention needed by GPT-style decoder-only models, <code class="docutils literal notranslate"><span class="pre">gpt_attention</span></code> also supports cross attention.</p>
<p>This enables using <code class="docutils literal notranslate"><span class="pre">gpt_attention</span></code> in a broader aspect as a generic decoder component. For example, the Encoder-Decoder model uses <code class="docutils literal notranslate"><span class="pre">gpt_attention</span></code> to issue both the self attention and cross attention modules in its Decoder.</p>
</section>
<section id="relative-attention-bias-rab">
<h3>Relative Attention Bias (RAB)<a class="headerlink" href="#relative-attention-bias-rab" title="Link to this heading"></a></h3>
<p>Relative attention bias (RAB) is a kind of relative position modeling, adding an attention bias (<code class="docutils literal notranslate"><span class="pre">Q*K^T+bias</span></code>) according to relative positions. RAB is a lightweight method to include the information of relative positions, and is used in the popular Encoder-Decoder model <a class="reference external" href="https://huggingface.co/docs/transformers/model_doc/t5">T5</a> and also other models in the T5 family.</p>
<p>RAB is supported in two modes: i) regular mode which user passes in relative attention bias computed ahead of MHA. ii) implicit mode which computes the relative attention bias on the fly in MHA. The implicit mode suits the case when the relative attention bias is too large to fit in memory and can be turned on by passing in <code class="docutils literal notranslate"><span class="pre">max_distance</span></code>.</p>
</section>
</section>
</section>
</div>
</div>
<footer><div class="rst-footer-buttons" role="navigation" aria-label="Footer">
<a href="batch_manager.html" class="btn btn-neutral float-left" title="The Batch Manager in TensorRT-LLM" accesskey="p" rel="prev"><span class="fa fa-arrow-circle-left" aria-hidden="true"></span> Previous</a>
<a href="precision.html" class="btn btn-neutral float-right" title="Numerical Precision" accesskey="n" rel="next">Next <span class="fa fa-arrow-circle-right" aria-hidden="true"></span></a>
</div>
<hr/>
<div role="contentinfo">
<p>&#169; Copyright 2023, NVidia.</p>
</div>
Built with <a href="https://www.sphinx-doc.org/">Sphinx</a> using a
<a href="https://github.com/readthedocs/sphinx_rtd_theme">theme</a>
provided by <a href="https://readthedocs.org">Read the Docs</a>.
</footer>
</div>
</div>
</section>
</div>
<script>
jQuery(function () {
SphinxRtdTheme.Navigation.enable(true);
});
</script>
</body>
</html>
Reference in New Issue View Git Blame Copy Permalink