mirror of
https://github.com/NVIDIA/TensorRT-LLM.git
synced 2026-01-14 06:27:45 +08:00
1185 lines
171 KiB
HTML
1185 lines
171 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=1.0" />
|
|
<title>tensorrt_llm.layers.attention — 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" />
|
|
</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>
|
|
<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"><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="../../../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"><a href="../../index.html">Module code</a></li>
|
|
<li class="breadcrumb-item active">tensorrt_llm.layers.attention</li>
|
|
<li class="wy-breadcrumbs-aside">
|
|
</li>
|
|
</ul>
|
|
<hr/>
|
|
</div>
|
|
<div role="main" class="document" itemscope="itemscope" itemtype="http://schema.org/Article">
|
|
<div itemprop="articleBody">
|
|
|
|
<h1>Source code for tensorrt_llm.layers.attention</h1><div class="highlight"><pre>
|
|
<span></span><span class="c1"># SPDX-FileCopyrightText: Copyright (c) 2022-2023 NVIDIA CORPORATION & AFFILIATES. All rights reserved.</span>
|
|
<span class="c1"># SPDX-License-Identifier: Apache-2.0</span>
|
|
<span class="c1">#</span>
|
|
<span class="c1"># Licensed under the Apache License, Version 2.0 (the "License");</span>
|
|
<span class="c1"># you may not use this file except in compliance with the License.</span>
|
|
<span class="c1"># You may obtain a copy of the License at</span>
|
|
<span class="c1">#</span>
|
|
<span class="c1"># http://www.apache.org/licenses/LICENSE-2.0</span>
|
|
<span class="c1">#</span>
|
|
<span class="c1"># Unless required by applicable law or agreed to in writing, software</span>
|
|
<span class="c1"># distributed under the License is distributed on an "AS IS" BASIS,</span>
|
|
<span class="c1"># WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.</span>
|
|
<span class="c1"># See the License for the specific language governing permissions and</span>
|
|
<span class="c1"># limitations under the License.</span>
|
|
<span class="kn">import</span> <span class="nn">math</span>
|
|
<span class="kn">from</span> <span class="nn">typing</span> <span class="kn">import</span> <span class="n">List</span><span class="p">,</span> <span class="n">Optional</span>
|
|
|
|
<span class="kn">import</span> <span class="nn">numpy</span> <span class="k">as</span> <span class="nn">np</span>
|
|
<span class="kn">import</span> <span class="nn">tensorrt</span> <span class="k">as</span> <span class="nn">trt</span>
|
|
|
|
<span class="kn">from</span> <span class="nn">.._common</span> <span class="kn">import</span> <span class="n">default_net</span><span class="p">,</span> <span class="n">precision</span>
|
|
<span class="kn">from</span> <span class="nn">.._utils</span> <span class="kn">import</span> <span class="n">numpy_fp32_to_bf16</span><span class="p">,</span> <span class="n">trt_dtype_to_np</span>
|
|
<span class="kn">from</span> <span class="nn">..functional</span> <span class="kn">import</span> <span class="p">(</span><span class="n">AttentionMaskType</span><span class="p">,</span> <span class="n">PositionEmbeddingType</span><span class="p">,</span>
|
|
<span class="n">RotaryScalingType</span><span class="p">,</span> <span class="n">Tensor</span><span class="p">,</span> <span class="n">bert_attention</span><span class="p">,</span> <span class="n">cast</span><span class="p">,</span> <span class="n">clip</span><span class="p">,</span>
|
|
<span class="n">concat</span><span class="p">,</span> <span class="n">constant</span><span class="p">,</span> <span class="n">embedding</span><span class="p">,</span> <span class="n">expand_dims</span><span class="p">,</span> <span class="n">expand_mask</span><span class="p">,</span>
|
|
<span class="n">generate_alibi_biases</span><span class="p">,</span> <span class="n">generate_alibi_slopes</span><span class="p">,</span>
|
|
<span class="n">gpt_attention</span><span class="p">,</span> <span class="n">matmul</span><span class="p">,</span> <span class="n">repeat_interleave</span><span class="p">,</span> <span class="nb">round</span><span class="p">,</span>
|
|
<span class="n">shape</span><span class="p">,</span> <span class="nb">slice</span><span class="p">,</span> <span class="n">softmax</span><span class="p">,</span> <span class="n">split</span><span class="p">,</span> <span class="n">view</span><span class="p">,</span> <span class="n">where</span><span class="p">)</span>
|
|
<span class="kn">from</span> <span class="nn">..module</span> <span class="kn">import</span> <span class="n">Module</span>
|
|
<span class="kn">from</span> <span class="nn">..parameter</span> <span class="kn">import</span> <span class="n">Parameter</span>
|
|
<span class="kn">from</span> <span class="nn">..quantization</span> <span class="kn">import</span> <span class="n">QuantMode</span>
|
|
<span class="kn">from</span> <span class="nn">..quantization.layers</span> <span class="kn">import</span> <span class="n">FP8Linear</span><span class="p">,</span> <span class="n">FP8RowLinear</span>
|
|
<span class="kn">from</span> <span class="nn">.linear</span> <span class="kn">import</span> <span class="n">ColumnLinear</span><span class="p">,</span> <span class="n">RowLinear</span>
|
|
<span class="kn">from</span> <span class="nn">.lora</span> <span class="kn">import</span> <span class="n">Lora</span>
|
|
|
|
|
|
<div class="viewcode-block" id="RopeEmbeddingUtils">
|
|
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.layers.html#tensorrt_llm.layers.attention.RopeEmbeddingUtils">[docs]</a>
|
|
<span class="k">class</span> <span class="nc">RopeEmbeddingUtils</span><span class="p">:</span>
|
|
|
|
<div class="viewcode-block" id="RopeEmbeddingUtils.create_sinusoidal_positions">
|
|
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.layers.html#tensorrt_llm.layers.attention.RopeEmbeddingUtils.create_sinusoidal_positions">[docs]</a>
|
|
<span class="nd">@staticmethod</span>
|
|
<span class="k">def</span> <span class="nf">create_sinusoidal_positions</span><span class="p">(</span><span class="n">num_pos</span><span class="p">:</span> <span class="nb">int</span><span class="p">,</span>
|
|
<span class="n">dim</span><span class="p">:</span> <span class="nb">int</span><span class="p">,</span>
|
|
<span class="n">theta</span><span class="p">:</span> <span class="nb">float</span> <span class="o">=</span> <span class="mf">10000.0</span><span class="p">,</span>
|
|
<span class="n">dtype</span><span class="o">=</span><span class="n">np</span><span class="o">.</span><span class="n">float32</span><span class="p">):</span>
|
|
<span class="n">inv_freq</span> <span class="o">=</span> <span class="mf">1.0</span> <span class="o">/</span> <span class="p">(</span><span class="n">theta</span><span class="o">**</span><span class="p">(</span><span class="n">np</span><span class="o">.</span><span class="n">arange</span><span class="p">(</span><span class="mi">0</span><span class="p">,</span> <span class="n">dim</span><span class="p">,</span> <span class="mi">2</span><span class="p">)</span> <span class="o">/</span> <span class="n">dim</span><span class="p">))</span><span class="o">.</span><span class="n">astype</span><span class="p">(</span><span class="n">dtype</span><span class="p">)</span>
|
|
<span class="n">sinusoid_inp</span> <span class="o">=</span> <span class="n">np</span><span class="o">.</span><span class="n">einsum</span><span class="p">(</span><span class="s2">"i , j -> i j"</span><span class="p">,</span>
|
|
<span class="n">np</span><span class="o">.</span><span class="n">arange</span><span class="p">(</span><span class="n">num_pos</span><span class="p">,</span> <span class="n">dtype</span><span class="o">=</span><span class="n">dtype</span><span class="p">),</span>
|
|
<span class="n">inv_freq</span><span class="p">,</span>
|
|
<span class="n">dtype</span><span class="o">=</span><span class="n">dtype</span><span class="p">)</span>
|
|
<span class="n">concat</span> <span class="o">=</span> <span class="n">np</span><span class="o">.</span><span class="n">concatenate</span><span class="p">((</span><span class="n">np</span><span class="o">.</span><span class="n">sin</span><span class="p">(</span><span class="n">sinusoid_inp</span><span class="p">),</span> <span class="n">np</span><span class="o">.</span><span class="n">cos</span><span class="p">(</span><span class="n">sinusoid_inp</span><span class="p">)),</span>
|
|
<span class="n">axis</span><span class="o">=</span><span class="mi">1</span><span class="p">)</span>
|
|
<span class="k">return</span> <span class="n">np</span><span class="o">.</span><span class="n">expand_dims</span><span class="p">(</span><span class="n">concat</span><span class="p">,</span> <span class="n">axis</span><span class="o">=</span><span class="mi">0</span><span class="p">)</span><span class="o">.</span><span class="n">astype</span><span class="p">(</span><span class="n">np</span><span class="o">.</span><span class="n">float32</span><span class="p">)</span></div>
|
|
|
|
|
|
<div class="viewcode-block" id="RopeEmbeddingUtils.rotate_every_two">
|
|
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.layers.html#tensorrt_llm.layers.attention.RopeEmbeddingUtils.rotate_every_two">[docs]</a>
|
|
<span class="nd">@staticmethod</span>
|
|
<span class="k">def</span> <span class="nf">rotate_every_two</span><span class="p">(</span><span class="n">tensor</span><span class="p">:</span> <span class="n">Tensor</span><span class="p">)</span> <span class="o">-></span> <span class="n">Tensor</span><span class="p">:</span>
|
|
<span class="k">assert</span> <span class="n">tensor</span><span class="o">.</span><span class="n">ndim</span><span class="p">()</span> <span class="o">==</span> <span class="mi">4</span>
|
|
|
|
<span class="n">shape_tensor</span> <span class="o">=</span> <span class="n">concat</span><span class="p">([</span>
|
|
<span class="n">shape</span><span class="p">(</span><span class="n">tensor</span><span class="p">,</span> <span class="n">i</span><span class="p">)</span> <span class="o">/</span> <span class="mi">2</span> <span class="k">if</span> <span class="n">i</span> <span class="o">==</span> <span class="p">(</span><span class="n">tensor</span><span class="o">.</span><span class="n">ndim</span><span class="p">()</span> <span class="o">-</span>
|
|
<span class="mi">1</span><span class="p">)</span> <span class="k">else</span> <span class="n">shape</span><span class="p">(</span><span class="n">tensor</span><span class="p">,</span> <span class="n">i</span><span class="p">)</span>
|
|
<span class="k">for</span> <span class="n">i</span> <span class="ow">in</span> <span class="nb">range</span><span class="p">(</span><span class="n">tensor</span><span class="o">.</span><span class="n">ndim</span><span class="p">())</span>
|
|
<span class="p">])</span>
|
|
<span class="n">x1</span> <span class="o">=</span> <span class="nb">slice</span><span class="p">(</span><span class="n">tensor</span><span class="p">,</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="mi">0</span><span class="p">,</span> <span class="mi">0</span><span class="p">],</span> <span class="n">shape_tensor</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">1</span><span class="p">,</span> <span class="mi">2</span><span class="p">])</span>
|
|
<span class="n">x2</span> <span class="o">=</span> <span class="nb">slice</span><span class="p">(</span><span class="n">tensor</span><span class="p">,</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="mi">0</span><span class="p">,</span> <span class="mi">1</span><span class="p">],</span> <span class="n">shape_tensor</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">1</span><span class="p">,</span> <span class="mi">2</span><span class="p">])</span>
|
|
<span class="n">x1</span> <span class="o">=</span> <span class="n">expand_dims</span><span class="p">(</span><span class="n">x1</span><span class="p">,</span> <span class="mi">4</span><span class="p">)</span>
|
|
<span class="n">x2</span> <span class="o">=</span> <span class="n">expand_dims</span><span class="p">(</span><span class="n">x2</span><span class="p">,</span> <span class="mi">4</span><span class="p">)</span>
|
|
<span class="n">zero</span> <span class="o">=</span> <span class="n">constant</span><span class="p">(</span>
|
|
<span class="n">np</span><span class="o">.</span><span class="n">ascontiguousarray</span><span class="p">(</span><span class="n">np</span><span class="o">.</span><span class="n">zeros</span><span class="p">([</span><span class="mi">1</span><span class="p">],</span>
|
|
<span class="n">dtype</span><span class="o">=</span><span class="n">trt_dtype_to_np</span><span class="p">(</span><span class="n">x2</span><span class="o">.</span><span class="n">dtype</span><span class="p">))))</span>
|
|
<span class="n">x2</span> <span class="o">=</span> <span class="n">zero</span> <span class="o">-</span> <span class="n">x2</span>
|
|
<span class="n">x</span> <span class="o">=</span> <span class="n">concat</span><span class="p">([</span><span class="n">x2</span><span class="p">,</span> <span class="n">x1</span><span class="p">],</span> <span class="mi">4</span><span class="p">)</span>
|
|
<span class="k">return</span> <span class="n">view</span><span class="p">(</span>
|
|
<span class="n">x</span><span class="p">,</span> <span class="n">concat</span><span class="p">([</span><span class="n">shape</span><span class="p">(</span><span class="n">x</span><span class="p">,</span> <span class="mi">0</span><span class="p">),</span>
|
|
<span class="n">shape</span><span class="p">(</span><span class="n">x</span><span class="p">,</span> <span class="mi">1</span><span class="p">),</span>
|
|
<span class="n">shape</span><span class="p">(</span><span class="n">x</span><span class="p">,</span> <span class="mi">2</span><span class="p">),</span>
|
|
<span class="n">shape</span><span class="p">(</span><span class="n">x</span><span class="p">,</span> <span class="mi">3</span><span class="p">)</span> <span class="o">*</span> <span class="mi">2</span><span class="p">]))</span></div>
|
|
|
|
|
|
<div class="viewcode-block" id="RopeEmbeddingUtils.rotate_half">
|
|
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.layers.html#tensorrt_llm.layers.attention.RopeEmbeddingUtils.rotate_half">[docs]</a>
|
|
<span class="nd">@staticmethod</span>
|
|
<span class="k">def</span> <span class="nf">rotate_half</span><span class="p">(</span><span class="n">tensor</span><span class="p">:</span> <span class="n">Tensor</span><span class="p">)</span> <span class="o">-></span> <span class="n">Tensor</span><span class="p">:</span>
|
|
<span class="c1"># [bs, num_attention_kv_heads, seqlen, attention_head_size]</span>
|
|
<span class="k">assert</span> <span class="n">tensor</span><span class="o">.</span><span class="n">ndim</span><span class="p">()</span> <span class="o">==</span> <span class="mi">4</span>
|
|
<span class="n">shape_tensor</span> <span class="o">=</span> <span class="n">concat</span><span class="p">([</span>
|
|
<span class="n">shape</span><span class="p">(</span><span class="n">tensor</span><span class="p">,</span> <span class="n">i</span><span class="p">)</span> <span class="o">/</span> <span class="mi">2</span> <span class="k">if</span> <span class="n">i</span> <span class="o">==</span> <span class="p">(</span><span class="n">tensor</span><span class="o">.</span><span class="n">ndim</span><span class="p">()</span> <span class="o">-</span>
|
|
<span class="mi">1</span><span class="p">)</span> <span class="k">else</span> <span class="n">shape</span><span class="p">(</span><span class="n">tensor</span><span class="p">,</span> <span class="n">i</span><span class="p">)</span>
|
|
<span class="k">for</span> <span class="n">i</span> <span class="ow">in</span> <span class="nb">range</span><span class="p">(</span><span class="n">tensor</span><span class="o">.</span><span class="n">ndim</span><span class="p">())</span>
|
|
<span class="p">])</span>
|
|
<span class="n">last_dim</span> <span class="o">=</span> <span class="n">shape</span><span class="p">(</span><span class="n">tensor</span><span class="p">,</span> <span class="n">tensor</span><span class="o">.</span><span class="n">ndim</span><span class="p">()</span> <span class="o">-</span> <span class="mi">1</span><span class="p">)</span> <span class="o">/</span> <span class="mi">2</span>
|
|
<span class="n">x1</span> <span class="o">=</span> <span class="nb">slice</span><span class="p">(</span><span class="n">tensor</span><span class="p">,</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="mi">0</span><span class="p">,</span> <span class="mi">0</span><span class="p">],</span> <span class="n">shape_tensor</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">1</span><span class="p">,</span> <span class="mi">1</span><span class="p">])</span>
|
|
<span class="n">x2</span> <span class="o">=</span> <span class="nb">slice</span><span class="p">(</span><span class="n">tensor</span><span class="p">,</span> <span class="n">concat</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="mi">0</span><span class="p">,</span> <span class="n">last_dim</span><span class="p">]),</span> <span class="n">shape_tensor</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">1</span><span class="p">,</span> <span class="mi">1</span><span class="p">])</span>
|
|
<span class="n">zero</span> <span class="o">=</span> <span class="n">constant</span><span class="p">(</span>
|
|
<span class="n">np</span><span class="o">.</span><span class="n">ascontiguousarray</span><span class="p">(</span><span class="n">np</span><span class="o">.</span><span class="n">zeros</span><span class="p">([</span><span class="mi">1</span><span class="p">],</span>
|
|
<span class="n">dtype</span><span class="o">=</span><span class="n">trt_dtype_to_np</span><span class="p">(</span><span class="n">x2</span><span class="o">.</span><span class="n">dtype</span><span class="p">))))</span>
|
|
<span class="n">x2</span> <span class="o">=</span> <span class="n">zero</span> <span class="o">-</span> <span class="n">x2</span>
|
|
<span class="n">x</span> <span class="o">=</span> <span class="n">concat</span><span class="p">([</span><span class="n">x2</span><span class="p">,</span> <span class="n">x1</span><span class="p">],</span> <span class="mi">3</span><span class="p">)</span>
|
|
<span class="k">return</span> <span class="n">x</span></div>
|
|
|
|
|
|
<div class="viewcode-block" id="RopeEmbeddingUtils.apply_rotary_pos_emb">
|
|
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.layers.html#tensorrt_llm.layers.attention.RopeEmbeddingUtils.apply_rotary_pos_emb">[docs]</a>
|
|
<span class="nd">@staticmethod</span>
|
|
<span class="k">def</span> <span class="nf">apply_rotary_pos_emb</span><span class="p">(</span>
|
|
<span class="n">tensor</span><span class="p">:</span> <span class="n">Tensor</span><span class="p">,</span>
|
|
<span class="n">position_embedding</span><span class="p">:</span> <span class="n">List</span><span class="p">[</span><span class="n">Tensor</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">pos_emb_type</span><span class="p">:</span> <span class="n">PositionEmbeddingType</span> <span class="o">=</span> <span class="n">PositionEmbeddingType</span><span class="o">.</span><span class="n">rope_gptj</span>
|
|
<span class="p">)</span> <span class="o">-></span> <span class="n">Tensor</span><span class="p">:</span>
|
|
|
|
<span class="n">rotate_func</span> <span class="o">=</span> <span class="kc">None</span>
|
|
<span class="k">if</span> <span class="n">pos_emb_type</span> <span class="o">==</span> <span class="n">PositionEmbeddingType</span><span class="o">.</span><span class="n">rope_gpt_neox</span><span class="p">:</span>
|
|
<span class="k">assert</span> <span class="nb">len</span><span class="p">(</span><span class="n">position_embedding</span><span class="p">)</span> <span class="o">==</span> <span class="mi">2</span>
|
|
<span class="n">cos</span><span class="p">,</span> <span class="n">sin</span> <span class="o">=</span> <span class="n">position_embedding</span>
|
|
<span class="n">sin</span> <span class="o">=</span> <span class="n">expand_dims</span><span class="p">(</span><span class="n">sin</span><span class="p">,</span> <span class="mi">2</span><span class="p">)</span>
|
|
<span class="n">cos</span> <span class="o">=</span> <span class="n">expand_dims</span><span class="p">(</span><span class="n">cos</span><span class="p">,</span> <span class="mi">2</span><span class="p">)</span>
|
|
<span class="n">sin</span> <span class="o">=</span> <span class="n">concat</span><span class="p">([</span><span class="n">sin</span><span class="p">,</span> <span class="n">sin</span><span class="p">],</span> <span class="mi">3</span><span class="p">)</span>
|
|
<span class="n">cos</span> <span class="o">=</span> <span class="n">concat</span><span class="p">([</span><span class="n">cos</span><span class="p">,</span> <span class="n">cos</span><span class="p">],</span> <span class="mi">3</span><span class="p">)</span>
|
|
<span class="n">rotate_func</span> <span class="o">=</span> <span class="n">RopeEmbeddingUtils</span><span class="o">.</span><span class="n">rotate_half</span>
|
|
<span class="k">elif</span> <span class="n">pos_emb_type</span> <span class="o">==</span> <span class="n">PositionEmbeddingType</span><span class="o">.</span><span class="n">rope_gptj</span><span class="p">:</span>
|
|
<span class="k">assert</span> <span class="nb">len</span><span class="p">(</span><span class="n">position_embedding</span><span class="p">)</span> <span class="o">==</span> <span class="mi">2</span>
|
|
<span class="n">cos</span><span class="p">,</span> <span class="n">sin</span> <span class="o">=</span> <span class="n">position_embedding</span>
|
|
<span class="n">sin</span> <span class="o">=</span> <span class="n">expand_dims</span><span class="p">(</span><span class="n">sin</span><span class="p">,</span> <span class="mi">2</span><span class="p">)</span>
|
|
<span class="n">cos</span> <span class="o">=</span> <span class="n">expand_dims</span><span class="p">(</span><span class="n">cos</span><span class="p">,</span> <span class="mi">2</span><span class="p">)</span>
|
|
<span class="n">sin</span> <span class="o">=</span> <span class="n">repeat_interleave</span><span class="p">(</span><span class="n">sin</span><span class="p">,</span> <span class="mi">2</span><span class="p">,</span> <span class="mi">3</span><span class="p">)</span>
|
|
<span class="n">cos</span> <span class="o">=</span> <span class="n">repeat_interleave</span><span class="p">(</span><span class="n">cos</span><span class="p">,</span> <span class="mi">2</span><span class="p">,</span> <span class="mi">3</span><span class="p">)</span>
|
|
<span class="n">rotate_func</span> <span class="o">=</span> <span class="n">RopeEmbeddingUtils</span><span class="o">.</span><span class="n">rotate_every_two</span>
|
|
<span class="k">elif</span> <span class="n">pos_emb_type</span> <span class="o">==</span> <span class="n">PositionEmbeddingType</span><span class="o">.</span><span class="n">chatglm</span><span class="p">:</span>
|
|
<span class="k">assert</span> <span class="nb">len</span><span class="p">(</span><span class="n">position_embedding</span><span class="p">)</span> <span class="o">==</span> <span class="mi">4</span>
|
|
<span class="n">cos0</span><span class="p">,</span> <span class="n">cos1</span><span class="p">,</span> <span class="n">sin0</span><span class="p">,</span> <span class="n">sin1</span> <span class="o">=</span> <span class="n">position_embedding</span>
|
|
|
|
<span class="n">shape_tensor</span> <span class="o">=</span> <span class="n">concat</span><span class="p">([</span>
|
|
<span class="n">shape</span><span class="p">(</span><span class="n">tensor</span><span class="p">,</span> <span class="n">i</span><span class="p">)</span> <span class="o">/</span> <span class="mi">2</span> <span class="k">if</span> <span class="n">i</span> <span class="o">==</span> <span class="p">(</span><span class="n">tensor</span><span class="o">.</span><span class="n">ndim</span><span class="p">()</span> <span class="o">-</span>
|
|
<span class="mi">1</span><span class="p">)</span> <span class="k">else</span> <span class="n">shape</span><span class="p">(</span><span class="n">tensor</span><span class="p">,</span> <span class="n">i</span><span class="p">)</span>
|
|
<span class="k">for</span> <span class="n">i</span> <span class="ow">in</span> <span class="nb">range</span><span class="p">(</span><span class="n">tensor</span><span class="o">.</span><span class="n">ndim</span><span class="p">())</span>
|
|
<span class="p">])</span>
|
|
<span class="n">last_dim</span> <span class="o">=</span> <span class="n">shape</span><span class="p">(</span><span class="n">tensor</span><span class="p">,</span> <span class="n">tensor</span><span class="o">.</span><span class="n">ndim</span><span class="p">()</span> <span class="o">-</span> <span class="mi">1</span><span class="p">)</span> <span class="o">/</span> <span class="mi">2</span>
|
|
<span class="n">x_part0</span> <span class="o">=</span> <span class="nb">slice</span><span class="p">(</span><span class="n">tensor</span><span class="p">,</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="mi">0</span><span class="p">,</span> <span class="mi">0</span><span class="p">],</span> <span class="n">shape_tensor</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">1</span><span class="p">,</span> <span class="mi">1</span><span class="p">])</span>
|
|
<span class="n">x_part1</span> <span class="o">=</span> <span class="nb">slice</span><span class="p">(</span><span class="n">tensor</span><span class="p">,</span> <span class="n">concat</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="mi">0</span><span class="p">,</span> <span class="n">last_dim</span><span class="p">]),</span> <span class="n">shape_tensor</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">1</span><span class="p">,</span> <span class="mi">1</span><span class="p">])</span>
|
|
|
|
<span class="n">y_part0</span> <span class="o">=</span> <span class="p">(</span><span class="n">x_part0</span> <span class="o">*</span>
|
|
<span class="n">cos0</span><span class="p">)</span> <span class="o">+</span> <span class="p">(</span><span class="n">RopeEmbeddingUtils</span><span class="o">.</span><span class="n">rotate_half</span><span class="p">(</span><span class="n">x_part0</span><span class="p">)</span> <span class="o">*</span> <span class="n">sin0</span><span class="p">)</span>
|
|
<span class="n">y_part1</span> <span class="o">=</span> <span class="p">(</span><span class="n">x_part1</span> <span class="o">*</span>
|
|
<span class="n">cos1</span><span class="p">)</span> <span class="o">+</span> <span class="p">(</span><span class="n">RopeEmbeddingUtils</span><span class="o">.</span><span class="n">rotate_half</span><span class="p">(</span><span class="n">x_part1</span><span class="p">)</span> <span class="o">*</span> <span class="n">sin1</span><span class="p">)</span>
|
|
|
|
<span class="n">result</span> <span class="o">=</span> <span class="n">concat</span><span class="p">([</span><span class="n">y_part0</span><span class="p">,</span> <span class="n">y_part1</span><span class="p">],</span> <span class="n">dim</span><span class="o">=</span><span class="mi">3</span><span class="p">)</span>
|
|
<span class="k">return</span> <span class="n">result</span><span class="o">.</span><span class="n">view</span><span class="p">(</span><span class="n">shape</span><span class="p">(</span><span class="n">tensor</span><span class="p">))</span>
|
|
|
|
<span class="k">else</span><span class="p">:</span>
|
|
<span class="k">raise</span> <span class="ne">ValueError</span><span class="p">(</span><span class="s1">'The PositionEmbeddingType is not RoPE'</span><span class="p">)</span>
|
|
<span class="k">return</span> <span class="p">(</span><span class="n">tensor</span> <span class="o">*</span> <span class="n">cos</span><span class="p">)</span> <span class="o">+</span> <span class="p">(</span><span class="n">rotate_func</span><span class="p">(</span><span class="n">tensor</span><span class="p">)</span> <span class="o">*</span> <span class="n">sin</span><span class="p">)</span></div>
|
|
|
|
|
|
<div class="viewcode-block" id="RopeEmbeddingUtils.apply_rotary_pos_emb_chatglm">
|
|
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.layers.html#tensorrt_llm.layers.attention.RopeEmbeddingUtils.apply_rotary_pos_emb_chatglm">[docs]</a>
|
|
<span class="nd">@staticmethod</span>
|
|
<span class="k">def</span> <span class="nf">apply_rotary_pos_emb_chatglm</span><span class="p">(</span>
|
|
<span class="n">qkv</span><span class="p">,</span>
|
|
<span class="n">position_embedding</span><span class="p">,</span>
|
|
<span class="n">num_attention_heads</span><span class="p">,</span>
|
|
<span class="n">attention_head_size</span><span class="p">,</span>
|
|
<span class="n">max_position_embeddings</span><span class="p">,</span>
|
|
<span class="n">rotary_embedding_scale</span><span class="p">,</span>
|
|
<span class="p">)</span> <span class="o">-></span> <span class="n">Tensor</span><span class="p">:</span>
|
|
|
|
<span class="n">half_head_size</span> <span class="o">=</span> <span class="n">attention_head_size</span> <span class="o">//</span> <span class="mi">2</span>
|
|
<span class="n">qkv_shape</span> <span class="o">=</span> <span class="n">shape</span><span class="p">(</span><span class="n">qkv</span><span class="p">)</span>
|
|
<span class="n">qkv</span> <span class="o">=</span> <span class="n">qkv</span><span class="o">.</span><span class="n">view</span><span class="p">(</span>
|
|
<span class="n">concat</span><span class="p">([</span>
|
|
<span class="n">shape</span><span class="p">(</span><span class="n">qkv</span><span class="p">,</span> <span class="mi">0</span><span class="p">),</span>
|
|
<span class="n">shape</span><span class="p">(</span><span class="n">qkv</span><span class="p">,</span> <span class="mi">1</span><span class="p">),</span>
|
|
<span class="n">num_attention_heads</span><span class="p">,</span>
|
|
<span class="mi">3</span><span class="p">,</span>
|
|
<span class="n">attention_head_size</span><span class="p">,</span>
|
|
<span class="p">]))</span>
|
|
<span class="n">query</span><span class="p">,</span> <span class="n">key</span><span class="p">,</span> <span class="n">value</span> <span class="o">=</span> <span class="n">split</span><span class="p">(</span><span class="n">qkv</span><span class="p">,</span> <span class="mi">1</span><span class="p">,</span> <span class="n">dim</span><span class="o">=</span><span class="mi">3</span><span class="p">)</span>
|
|
<span class="n">q_shape</span> <span class="o">=</span> <span class="n">concat</span><span class="p">([</span>
|
|
<span class="n">shape</span><span class="p">(</span><span class="n">qkv</span><span class="p">,</span> <span class="mi">0</span><span class="p">),</span>
|
|
<span class="n">shape</span><span class="p">(</span><span class="n">qkv</span><span class="p">,</span> <span class="mi">1</span><span class="p">),</span>
|
|
<span class="n">num_attention_heads</span><span class="p">,</span>
|
|
<span class="n">attention_head_size</span><span class="p">,</span>
|
|
<span class="p">])</span>
|
|
<span class="n">query</span> <span class="o">=</span> <span class="n">query</span><span class="o">.</span><span class="n">view</span><span class="p">(</span><span class="n">q_shape</span><span class="p">)</span>
|
|
<span class="n">key</span> <span class="o">=</span> <span class="n">key</span><span class="o">.</span><span class="n">view</span><span class="p">(</span><span class="n">q_shape</span><span class="p">)</span>
|
|
<span class="n">value</span> <span class="o">=</span> <span class="n">value</span><span class="o">.</span><span class="n">view</span><span class="p">(</span><span class="n">q_shape</span><span class="p">)</span>
|
|
|
|
<span class="n">embedding_weight</span> <span class="o">=</span> <span class="n">RopeEmbeddingUtils</span><span class="o">.</span><span class="n">create_sinusoidal_positions</span><span class="p">(</span>
|
|
<span class="n">max_position_embeddings</span><span class="p">,</span> <span class="n">half_head_size</span><span class="p">)</span>
|
|
<span class="n">embedding_weight</span> <span class="o">/=</span> <span class="n">rotary_embedding_scale</span>
|
|
<span class="n">embedding_weight</span> <span class="o">=</span> <span class="n">np</span><span class="o">.</span><span class="n">split</span><span class="p">(</span><span class="n">embedding_weight</span><span class="o">.</span><span class="n">squeeze</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="n">axis</span><span class="o">=</span><span class="mi">1</span><span class="p">)</span>
|
|
<span class="n">embedding_weight</span> <span class="o">=</span> <span class="n">np</span><span class="o">.</span><span class="n">concatenate</span><span class="p">(</span>
|
|
<span class="p">[</span>
|
|
<span class="n">embedding_weight</span><span class="p">[</span><span class="mi">0</span><span class="p">],</span>
|
|
<span class="n">embedding_weight</span><span class="p">[</span><span class="mi">0</span><span class="p">],</span>
|
|
<span class="n">embedding_weight</span><span class="p">[</span><span class="mi">1</span><span class="p">],</span>
|
|
<span class="n">embedding_weight</span><span class="p">[</span><span class="mi">1</span><span class="p">],</span>
|
|
<span class="p">],</span>
|
|
<span class="n">axis</span><span class="o">=</span><span class="mi">1</span><span class="p">,</span>
|
|
<span class="p">)</span>
|
|
|
|
<span class="n">embedding_weight</span> <span class="o">=</span> <span class="n">constant</span><span class="p">(</span><span class="n">embedding_weight</span><span class="p">)</span>
|
|
<span class="n">position_embedding</span> <span class="o">=</span> <span class="n">embedding</span><span class="p">(</span><span class="n">position_embedding</span><span class="p">,</span> <span class="n">embedding_weight</span><span class="p">)</span>
|
|
<span class="n">position_embedding</span><span class="p">,</span> <span class="n">block_embedding</span> <span class="o">=</span> <span class="n">split</span><span class="p">(</span>
|
|
<span class="n">position_embedding</span><span class="p">,</span>
|
|
<span class="mi">1</span><span class="p">,</span>
|
|
<span class="n">dim</span><span class="o">=</span><span class="mi">1</span><span class="p">,</span>
|
|
<span class="p">)</span>
|
|
<span class="n">sin0</span><span class="p">,</span> <span class="n">cos0</span> <span class="o">=</span> <span class="n">split</span><span class="p">(</span><span class="n">position_embedding</span><span class="p">,</span> <span class="n">half_head_size</span><span class="p">,</span> <span class="n">dim</span><span class="o">=</span><span class="mi">3</span><span class="p">)</span>
|
|
<span class="n">sin1</span><span class="p">,</span> <span class="n">cos1</span> <span class="o">=</span> <span class="n">split</span><span class="p">(</span><span class="n">block_embedding</span><span class="p">,</span> <span class="n">half_head_size</span><span class="p">,</span> <span class="n">dim</span><span class="o">=</span><span class="mi">3</span><span class="p">)</span>
|
|
|
|
<span class="n">new_shape</span> <span class="o">=</span> <span class="n">concat</span><span class="p">([</span>
|
|
<span class="n">shape</span><span class="p">(</span><span class="n">qkv</span><span class="p">,</span> <span class="mi">0</span><span class="p">),</span>
|
|
<span class="n">shape</span><span class="p">(</span><span class="n">qkv</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="n">half_head_size</span><span class="p">,</span>
|
|
<span class="p">])</span>
|
|
<span class="n">position_embedding</span> <span class="o">=</span> <span class="p">[</span>
|
|
<span class="n">tensor</span><span class="o">.</span><span class="n">view</span><span class="p">(</span><span class="n">new_shape</span><span class="p">)</span> <span class="k">for</span> <span class="n">tensor</span> <span class="ow">in</span> <span class="p">[</span><span class="n">cos0</span><span class="p">,</span> <span class="n">cos1</span><span class="p">,</span> <span class="n">sin0</span><span class="p">,</span> <span class="n">sin1</span><span class="p">]</span>
|
|
<span class="p">]</span>
|
|
|
|
<span class="n">query</span> <span class="o">=</span> <span class="n">RopeEmbeddingUtils</span><span class="o">.</span><span class="n">apply_rotary_pos_emb</span><span class="p">(</span>
|
|
<span class="n">tensor</span><span class="o">=</span><span class="n">query</span><span class="p">,</span>
|
|
<span class="n">position_embedding</span><span class="o">=</span><span class="n">position_embedding</span><span class="p">,</span>
|
|
<span class="n">pos_emb_type</span><span class="o">=</span><span class="n">PositionEmbeddingType</span><span class="o">.</span><span class="n">chatglm</span><span class="p">)</span>
|
|
<span class="n">key</span> <span class="o">=</span> <span class="n">RopeEmbeddingUtils</span><span class="o">.</span><span class="n">apply_rotary_pos_emb</span><span class="p">(</span>
|
|
<span class="n">tensor</span><span class="o">=</span><span class="n">key</span><span class="p">,</span>
|
|
<span class="n">position_embedding</span><span class="o">=</span><span class="n">position_embedding</span><span class="p">,</span>
|
|
<span class="n">pos_emb_type</span><span class="o">=</span><span class="n">PositionEmbeddingType</span><span class="o">.</span><span class="n">chatglm</span><span class="p">)</span>
|
|
|
|
<span class="n">qkv</span> <span class="o">=</span> <span class="n">concat</span><span class="p">([</span><span class="n">query</span><span class="p">,</span> <span class="n">key</span><span class="p">,</span> <span class="n">value</span><span class="p">],</span> <span class="n">dim</span><span class="o">=</span><span class="mi">2</span><span class="p">)</span>
|
|
<span class="n">qkv</span> <span class="o">=</span> <span class="n">qkv</span><span class="o">.</span><span class="n">view</span><span class="p">(</span><span class="n">qkv_shape</span><span class="p">)</span>
|
|
|
|
<span class="k">return</span> <span class="n">qkv</span></div>
|
|
</div>
|
|
|
|
|
|
|
|
<div class="viewcode-block" id="AttentionParams">
|
|
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.layers.html#tensorrt_llm.layers.attention.AttentionParams">[docs]</a>
|
|
<span class="k">class</span> <span class="nc">AttentionParams</span><span class="p">(</span><span class="nb">object</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="n">sequence_length</span><span class="p">:</span> <span class="n">Tensor</span> <span class="o">=</span> <span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">context_lengths</span><span class="p">:</span> <span class="n">Tensor</span> <span class="o">=</span> <span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">host_context_lengths</span><span class="p">:</span> <span class="n">Tensor</span> <span class="o">=</span> <span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">max_context_length</span><span class="p">:</span> <span class="nb">int</span> <span class="o">=</span> <span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">host_request_types</span><span class="p">:</span> <span class="n">Tensor</span> <span class="o">=</span> <span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">encoder_input_lengths</span><span class="p">:</span> <span class="n">Tensor</span> <span class="o">=</span> <span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">encoder_max_input_length</span><span class="p">:</span> <span class="n">Tensor</span> <span class="o">=</span> <span class="kc">None</span><span class="p">):</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">sequence_length</span> <span class="o">=</span> <span class="n">sequence_length</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">context_lengths</span> <span class="o">=</span> <span class="n">context_lengths</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">host_context_lengths</span> <span class="o">=</span> <span class="n">host_context_lengths</span>
|
|
<span class="c1"># max allowed context length. Required to</span>
|
|
<span class="c1"># compute scratch memory size.</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">max_context_length</span> <span class="o">=</span> <span class="n">max_context_length</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">host_request_types</span> <span class="o">=</span> <span class="n">host_request_types</span>
|
|
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">encoder_input_lengths</span> <span class="o">=</span> <span class="n">encoder_input_lengths</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">encoder_max_input_length</span> <span class="o">=</span> <span class="n">encoder_max_input_length</span>
|
|
|
|
<div class="viewcode-block" id="AttentionParams.is_valid_cross_attn">
|
|
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.layers.html#tensorrt_llm.layers.attention.AttentionParams.is_valid_cross_attn">[docs]</a>
|
|
<span class="k">def</span> <span class="nf">is_valid_cross_attn</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">do_cross_attention</span><span class="p">):</span>
|
|
<span class="k">if</span> <span class="n">do_cross_attention</span><span class="p">:</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">encoder_input_lengths</span> <span class="ow">is</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="k">return</span> <span class="kc">False</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">encoder_max_input_length</span> <span class="ow">is</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="k">return</span> <span class="kc">False</span>
|
|
<span class="k">return</span> <span class="kc">True</span></div>
|
|
|
|
|
|
<div class="viewcode-block" id="AttentionParams.is_valid">
|
|
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.layers.html#tensorrt_llm.layers.attention.AttentionParams.is_valid">[docs]</a>
|
|
<span class="k">def</span> <span class="nf">is_valid</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">gpt_attention_plugin</span><span class="p">,</span> <span class="n">remove_input_padding</span><span class="p">):</span>
|
|
<span class="k">if</span> <span class="n">gpt_attention_plugin</span><span class="p">:</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">sequence_length</span> <span class="ow">is</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="k">return</span> <span class="kc">False</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">context_lengths</span> <span class="ow">is</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="k">return</span> <span class="kc">False</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">host_request_types</span> <span class="ow">is</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="k">return</span> <span class="kc">False</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">max_context_length</span> <span class="ow">is</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="k">return</span> <span class="kc">False</span>
|
|
|
|
<span class="k">if</span> <span class="n">remove_input_padding</span><span class="p">:</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">host_context_lengths</span> <span class="ow">is</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="k">return</span> <span class="kc">False</span>
|
|
<span class="k">if</span> <span class="ow">not</span> <span class="n">gpt_attention_plugin</span><span class="p">:</span>
|
|
<span class="k">return</span> <span class="kc">False</span>
|
|
|
|
<span class="k">return</span> <span class="kc">True</span></div>
|
|
</div>
|
|
|
|
|
|
|
|
<div class="viewcode-block" id="KeyValueCacheParams">
|
|
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.layers.html#tensorrt_llm.layers.attention.KeyValueCacheParams">[docs]</a>
|
|
<span class="k">class</span> <span class="nc">KeyValueCacheParams</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="n">past_key_value</span><span class="p">:</span> <span class="n">List</span><span class="p">[</span><span class="n">Tensor</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">host_past_key_value_lengths</span><span class="p">:</span> <span class="n">Tensor</span> <span class="o">=</span> <span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">host_max_kv_cache_lengths</span><span class="p">:</span> <span class="n">List</span><span class="p">[</span><span class="n">Tensor</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">kv_cache_block_pointers</span><span class="p">:</span> <span class="n">List</span><span class="p">[</span><span class="n">Tensor</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">cache_indirection</span><span class="p">:</span> <span class="n">Tensor</span> <span class="o">=</span> <span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">past_key_value_length</span><span class="p">:</span> <span class="n">Tensor</span> <span class="o">=</span> <span class="kc">None</span><span class="p">):</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">past_key_value</span> <span class="o">=</span> <span class="n">past_key_value</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">host_past_key_value_lengths</span> <span class="o">=</span> <span class="n">host_past_key_value_lengths</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">host_max_kv_cache_lengths</span> <span class="o">=</span> <span class="n">host_max_kv_cache_lengths</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">kv_cache_block_pointers</span> <span class="o">=</span> <span class="n">kv_cache_block_pointers</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">cache_indirection</span> <span class="o">=</span> <span class="n">cache_indirection</span>
|
|
<span class="c1"># self.past_key_value_length = past_key_value_length</span>
|
|
|
|
<div class="viewcode-block" id="KeyValueCacheParams.get_first_past_key_value">
|
|
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.layers.html#tensorrt_llm.layers.attention.KeyValueCacheParams.get_first_past_key_value">[docs]</a>
|
|
<span class="k">def</span> <span class="nf">get_first_past_key_value</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">past_key_value</span> <span class="ow">is</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="k">return</span> <span class="kc">None</span>
|
|
<span class="k">return</span> <span class="bp">self</span><span class="o">.</span><span class="n">past_key_value</span><span class="p">[</span><span class="mi">0</span><span class="p">]</span></div>
|
|
|
|
|
|
<div class="viewcode-block" id="KeyValueCacheParams.get_first_kv_cache_block_pointers">
|
|
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.layers.html#tensorrt_llm.layers.attention.KeyValueCacheParams.get_first_kv_cache_block_pointers">[docs]</a>
|
|
<span class="k">def</span> <span class="nf">get_first_kv_cache_block_pointers</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">kv_cache_block_pointers</span> <span class="ow">is</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="k">return</span> <span class="kc">None</span>
|
|
<span class="k">return</span> <span class="bp">self</span><span class="o">.</span><span class="n">kv_cache_block_pointers</span><span class="p">[</span><span class="mi">0</span><span class="p">]</span></div>
|
|
|
|
|
|
<div class="viewcode-block" id="KeyValueCacheParams.fill_none_tensor_list">
|
|
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.layers.html#tensorrt_llm.layers.attention.KeyValueCacheParams.fill_none_tensor_list">[docs]</a>
|
|
<span class="k">def</span> <span class="nf">fill_none_tensor_list</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">list_size</span><span class="p">):</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">past_key_value</span> <span class="ow">is</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">past_key_value</span> <span class="o">=</span> <span class="nb">tuple</span><span class="p">([</span><span class="kc">None</span><span class="p">]</span> <span class="o">*</span> <span class="n">list_size</span><span class="p">)</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">host_max_kv_cache_lengths</span> <span class="ow">is</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">host_max_kv_cache_lengths</span> <span class="o">=</span> <span class="nb">tuple</span><span class="p">([</span><span class="kc">None</span><span class="p">]</span> <span class="o">*</span> <span class="n">list_size</span><span class="p">)</span></div>
|
|
|
|
|
|
<div class="viewcode-block" id="KeyValueCacheParams.is_valid">
|
|
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.layers.html#tensorrt_llm.layers.attention.KeyValueCacheParams.is_valid">[docs]</a>
|
|
<span class="k">def</span> <span class="nf">is_valid</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">gpt_attention_plugin</span><span class="p">):</span>
|
|
<span class="k">if</span> <span class="n">gpt_attention_plugin</span><span class="p">:</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">host_past_key_value_lengths</span> <span class="ow">is</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="k">return</span> <span class="kc">False</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">host_max_kv_cache_lengths</span> <span class="ow">is</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="k">return</span> <span class="kc">False</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">cache_indirection</span> <span class="ow">is</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="k">return</span> <span class="kc">False</span>
|
|
|
|
<span class="k">return</span> <span class="kc">True</span></div>
|
|
</div>
|
|
|
|
|
|
|
|
<div class="viewcode-block" id="Attention">
|
|
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.layers.html#tensorrt_llm.layers.attention.Attention">[docs]</a>
|
|
<span class="k">class</span> <span class="nc">Attention</span><span class="p">(</span><span class="n">Module</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="n">hidden_size</span><span class="p">,</span>
|
|
<span class="n">num_attention_heads</span><span class="p">,</span>
|
|
<span class="n">num_kv_heads</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">max_position_embeddings</span><span class="o">=</span><span class="mi">1024</span><span class="p">,</span>
|
|
<span class="n">num_layers</span><span class="o">=</span><span class="mi">1</span><span class="p">,</span>
|
|
<span class="n">apply_query_key_layer_scaling</span><span class="o">=</span><span class="kc">False</span><span class="p">,</span>
|
|
<span class="n">attention_head_size</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">attention_mask_type</span><span class="o">=</span><span class="n">AttentionMaskType</span><span class="o">.</span><span class="n">padding</span><span class="p">,</span>
|
|
<span class="n">bias</span><span class="o">=</span><span class="kc">True</span><span class="p">,</span>
|
|
<span class="n">dtype</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">position_embedding_type</span><span class="o">=</span><span class="n">PositionEmbeddingType</span><span class="o">.</span><span class="n">learned_absolute</span><span class="p">,</span>
|
|
<span class="n">rotary_embedding_base</span><span class="o">=</span><span class="mf">10000.0</span><span class="p">,</span>
|
|
<span class="n">rotary_embedding_scaling</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">use_int8_kv_cache</span><span class="o">=</span><span class="kc">False</span><span class="p">,</span>
|
|
<span class="n">rotary_embedding_percentage</span><span class="o">=</span><span class="mf">1.0</span><span class="p">,</span>
|
|
<span class="n">tp_group</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">tp_size</span><span class="o">=</span><span class="mi">1</span><span class="p">,</span>
|
|
<span class="n">tp_rank</span><span class="o">=</span><span class="mi">0</span><span class="p">,</span>
|
|
<span class="n">quant_mode</span><span class="p">:</span> <span class="n">QuantMode</span> <span class="o">=</span> <span class="n">QuantMode</span><span class="p">(</span><span class="mi">0</span><span class="p">),</span>
|
|
<span class="n">q_scaling</span><span class="o">=</span><span class="mf">1.0</span><span class="p">,</span>
|
|
<span class="n">cross_attention</span><span class="o">=</span><span class="kc">False</span><span class="p">,</span>
|
|
<span class="n">relative_attention</span><span class="o">=</span><span class="kc">False</span><span class="p">,</span>
|
|
<span class="n">max_distance</span><span class="o">=</span><span class="mi">0</span><span class="p">,</span>
|
|
<span class="n">num_buckets</span><span class="o">=</span><span class="mi">0</span><span class="p">,</span>
|
|
<span class="n">instance_id</span><span class="p">:</span> <span class="nb">int</span> <span class="o">=</span> <span class="mi">0</span><span class="p">,</span>
|
|
<span class="n">dense_bias</span><span class="o">=</span><span class="kc">None</span><span class="p">,</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="bp">self</span><span class="o">.</span><span class="n">cross_attention</span> <span class="o">=</span> <span class="n">cross_attention</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">attention_mask_type</span> <span class="o">=</span> <span class="n">attention_mask_type</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">attention_head_size</span> <span class="o">=</span> <span class="n">hidden_size</span> <span class="o">//</span> <span class="n">num_attention_heads</span> <span class="k">if</span> <span class="n">attention_head_size</span> <span class="ow">is</span> <span class="kc">None</span> <span class="k">else</span> <span class="n">attention_head_size</span>
|
|
<span class="k">assert</span> <span class="n">num_attention_heads</span> <span class="o">%</span> <span class="n">tp_size</span> <span class="o">==</span> <span class="mi">0</span><span class="p">,</span> \
|
|
<span class="s2">"num_attention_heads must be divisible by tp_size"</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">num_attention_heads</span> <span class="o">=</span> <span class="n">num_attention_heads</span> <span class="o">//</span> <span class="n">tp_size</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">num_attention_kv_heads</span> <span class="o">=</span> <span class="p">(</span>
|
|
<span class="n">num_kv_heads</span> <span class="o">+</span> <span class="n">tp_size</span> <span class="o">-</span> <span class="mi">1</span>
|
|
<span class="p">)</span> <span class="o">//</span> <span class="n">tp_size</span> <span class="k">if</span> <span class="n">num_kv_heads</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span> <span class="k">else</span> <span class="bp">self</span><span class="o">.</span><span class="n">num_attention_heads</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">hidden_size</span> <span class="o">=</span> <span class="n">hidden_size</span> <span class="o">//</span> <span class="n">tp_size</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">max_position_embeddings</span> <span class="o">=</span> <span class="n">max_position_embeddings</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">tp_size</span> <span class="o">=</span> <span class="n">tp_size</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">tp_rank</span> <span class="o">=</span> <span class="n">tp_rank</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">dtype</span> <span class="o">=</span> <span class="n">dtype</span>
|
|
<span class="k">if</span> <span class="n">dense_bias</span> <span class="ow">is</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="n">dense_bias</span> <span class="o">=</span> <span class="n">bias</span>
|
|
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">num_layers</span> <span class="o">=</span> <span class="n">num_layers</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">apply_query_key_layer_scaling</span> <span class="o">=</span> <span class="n">apply_query_key_layer_scaling</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">norm_factor</span> <span class="o">=</span> <span class="n">math</span><span class="o">.</span><span class="n">sqrt</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">attention_head_size</span><span class="p">)</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">q_scaling</span> <span class="o">=</span> <span class="n">q_scaling</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">apply_query_key_layer_scaling</span><span class="p">:</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">norm_factor</span> <span class="o">*=</span> <span class="bp">self</span><span class="o">.</span><span class="n">num_layers</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">q_scaling</span> <span class="o">*=</span> <span class="bp">self</span><span class="o">.</span><span class="n">num_layers</span>
|
|
<span class="c1"># Whether to scale ALiBi bias. Mathematically, it's equivalent to</span>
|
|
<span class="c1"># normalizing QK after adding bias.</span>
|
|
<span class="c1"># - False, inv_sqrt_Dh * Q*K^T + alibi_bias</span>
|
|
<span class="c1"># - True, inv_sqrt_Dh * Q*K^T + inv_sqrt_Dh * alibi_bias</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">scale_alibi_bias</span> <span class="o">=</span> <span class="n">position_embedding_type</span> <span class="o">==</span> <span class="n">PositionEmbeddingType</span><span class="o">.</span><span class="n">alibi_with_scale</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">position_embedding_type</span> <span class="o">=</span> <span class="n">position_embedding_type</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">relative_attention</span> <span class="o">=</span> <span class="n">relative_attention</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">max_distance</span> <span class="o">=</span> <span class="n">max_distance</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">rotary_embedding_base</span> <span class="o">=</span> <span class="n">rotary_embedding_base</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">rotary_embedding_scale_type</span> <span class="o">=</span> <span class="n">RotaryScalingType</span><span class="o">.</span><span class="n">none</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">rotary_embedding_scale</span> <span class="o">=</span> <span class="mf">1.0</span>
|
|
<span class="k">if</span> <span class="n">rotary_embedding_scaling</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="k">assert</span> <span class="n">rotary_embedding_scaling</span><span class="p">[</span><span class="s2">"type"</span><span class="p">]</span> <span class="ow">in</span> <span class="p">[</span><span class="s2">"linear"</span><span class="p">,</span> <span class="s2">"dynamic"</span><span class="p">]</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">rotary_embedding_scale_type</span> <span class="o">=</span> <span class="n">RotaryScalingType</span><span class="o">.</span><span class="n">linear</span> <span class="k">if</span> <span class="n">rotary_embedding_scaling</span><span class="p">[</span>
|
|
<span class="s2">"type"</span><span class="p">]</span> <span class="o">==</span> <span class="s2">"linear"</span> <span class="k">else</span> <span class="n">RotaryScalingType</span><span class="o">.</span><span class="n">dynamic</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">rotary_embedding_scale</span> <span class="o">=</span> <span class="n">rotary_embedding_scaling</span><span class="p">[</span><span class="s2">"factor"</span><span class="p">]</span>
|
|
<span class="k">assert</span> <span class="bp">self</span><span class="o">.</span><span class="n">rotary_embedding_scale</span> <span class="o">></span> <span class="mf">1.0</span>
|
|
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">embed_positions</span> <span class="o">=</span> <span class="kc">None</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">rotary_enabled</span> <span class="o">=</span> <span class="kc">False</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">rotary_embedding_dim</span> <span class="o">=</span> <span class="mi">0</span>
|
|
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">position_embedding_type</span><span class="o">.</span><span class="n">is_rope</span><span class="p">():</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">rotary_embedding_dim</span> <span class="o">=</span> <span class="nb">int</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">attention_head_size</span> <span class="o">*</span>
|
|
<span class="n">rotary_embedding_percentage</span><span class="p">)</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">rotary_enabled</span> <span class="o">=</span> <span class="kc">True</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">embed_positions</span> <span class="o">=</span> <span class="n">RopeEmbeddingUtils</span><span class="o">.</span><span class="n">create_sinusoidal_positions</span><span class="p">(</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">max_position_embeddings</span><span class="p">,</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">rotary_embedding_dim</span><span class="p">,</span>
|
|
<span class="p">)</span>
|
|
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">quant_mode</span> <span class="o">=</span> <span class="n">quant_mode</span>
|
|
<span class="k">if</span> <span class="n">use_int8_kv_cache</span><span class="p">:</span>
|
|
<span class="c1"># TODO: remove use_int8_kv_cache as can be replaced by quant_mode.has_kv_cache_quant()</span>
|
|
<span class="c1"># Merge int8 setting into quant_mode</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">quant_mode</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">quant_mode</span><span class="o">.</span><span class="n">set_int8_kv_cache</span><span class="p">()</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">use_int8_kv_cache</span> <span class="o">=</span> <span class="n">use_int8_kv_cache</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">quant_mode</span><span class="o">.</span><span class="n">has_kv_cache_quant</span><span class="p">():</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">kv_orig_quant_scale</span> <span class="o">=</span> <span class="n">Parameter</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="p">),</span> <span class="n">dtype</span><span class="o">=</span><span class="s1">'float32'</span><span class="p">)</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">kv_quant_orig_scale</span> <span class="o">=</span> <span class="n">Parameter</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="p">),</span> <span class="n">dtype</span><span class="o">=</span><span class="s1">'float32'</span><span class="p">)</span>
|
|
<span class="k">else</span><span class="p">:</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">register_parameter</span><span class="p">(</span><span class="s1">'kv_orig_quant_scale'</span><span class="p">,</span> <span class="kc">None</span><span class="p">)</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">register_parameter</span><span class="p">(</span><span class="s1">'kv_quant_orig_scale'</span><span class="p">,</span> <span class="kc">None</span><span class="p">)</span>
|
|
|
|
<span class="c1"># The output feature size is therefore (h/tp + 2*kvh/tp) * d, where h is num_heads,</span>
|
|
<span class="c1"># d is head_size, kvh is the num_kv_heads and tp is tensor_parallel_size.</span>
|
|
<span class="c1"># In ColumnLinear op, the output dim is calculated by (h + 2*kvh) * d / tp,</span>
|
|
<span class="c1"># which matches the desired output size (h/tp + 2*kvh/tp) * d after splitting</span>
|
|
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">use_fp8_qdq</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">quant_mode</span><span class="o">.</span><span class="n">has_fp8_qdq</span><span class="p">()</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">use_fp8_qdq</span><span class="p">:</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">qkv</span> <span class="o">=</span> <span class="n">FP8Linear</span><span class="p">(</span>
|
|
<span class="n">hidden_size</span><span class="p">,</span>
|
|
<span class="n">tp_size</span> <span class="o">*</span> <span class="bp">self</span><span class="o">.</span><span class="n">num_attention_heads</span> <span class="o">*</span> <span class="bp">self</span><span class="o">.</span><span class="n">attention_head_size</span> <span class="o">+</span>
|
|
<span class="p">(</span><span class="mi">2</span> <span class="o">*</span> <span class="n">tp_size</span> <span class="o">*</span> <span class="bp">self</span><span class="o">.</span><span class="n">num_attention_kv_heads</span> <span class="o">*</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">attention_head_size</span><span class="p">),</span>
|
|
<span class="n">bias</span><span class="o">=</span><span class="n">bias</span><span class="p">,</span>
|
|
<span class="n">dtype</span><span class="o">=</span><span class="n">dtype</span><span class="p">,</span>
|
|
<span class="n">tp_group</span><span class="o">=</span><span class="n">tp_group</span><span class="p">,</span>
|
|
<span class="n">tp_size</span><span class="o">=</span><span class="n">tp_size</span><span class="p">,</span>
|
|
<span class="n">gather_output</span><span class="o">=</span><span class="kc">False</span><span class="p">)</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">dense</span> <span class="o">=</span> <span class="n">FP8RowLinear</span><span class="p">(</span><span class="n">hidden_size</span><span class="p">,</span>
|
|
<span class="n">hidden_size</span><span class="p">,</span>
|
|
<span class="n">bias</span><span class="o">=</span><span class="n">dense_bias</span><span class="p">,</span>
|
|
<span class="n">dtype</span><span class="o">=</span><span class="n">dtype</span><span class="p">,</span>
|
|
<span class="n">tp_group</span><span class="o">=</span><span class="n">tp_group</span><span class="p">,</span>
|
|
<span class="n">tp_size</span><span class="o">=</span><span class="n">tp_size</span><span class="p">,</span>
|
|
<span class="n">instance_id</span><span class="o">=</span><span class="n">instance_id</span><span class="p">)</span>
|
|
<span class="k">else</span><span class="p">:</span>
|
|
<span class="c1"># out dim is not necessarily hidden_size + kv specific size (in MQA/GQA), but num_heads * heads_size</span>
|
|
<span class="c1"># example: d_model != num_heads * head_size in Flan-T5</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">qkv</span> <span class="o">=</span> <span class="n">ColumnLinear</span><span class="p">(</span>
|
|
<span class="n">hidden_size</span><span class="p">,</span>
|
|
<span class="n">tp_size</span> <span class="o">*</span> <span class="bp">self</span><span class="o">.</span><span class="n">num_attention_heads</span> <span class="o">*</span> <span class="bp">self</span><span class="o">.</span><span class="n">attention_head_size</span> <span class="o">+</span>
|
|
<span class="p">(</span><span class="mi">2</span> <span class="o">*</span> <span class="n">tp_size</span> <span class="o">*</span> <span class="bp">self</span><span class="o">.</span><span class="n">num_attention_kv_heads</span> <span class="o">*</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">attention_head_size</span><span class="p">),</span>
|
|
<span class="n">bias</span><span class="o">=</span><span class="n">bias</span><span class="p">,</span>
|
|
<span class="n">dtype</span><span class="o">=</span><span class="n">dtype</span><span class="p">,</span>
|
|
<span class="n">tp_group</span><span class="o">=</span><span class="n">tp_group</span><span class="p">,</span>
|
|
<span class="n">tp_size</span><span class="o">=</span><span class="n">tp_size</span><span class="p">,</span>
|
|
<span class="n">gather_output</span><span class="o">=</span><span class="kc">False</span><span class="p">)</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">dense</span> <span class="o">=</span> <span class="n">RowLinear</span><span class="p">(</span><span class="n">tp_size</span> <span class="o">*</span> <span class="bp">self</span><span class="o">.</span><span class="n">num_attention_heads</span> <span class="o">*</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">attention_head_size</span><span class="p">,</span>
|
|
<span class="n">hidden_size</span><span class="p">,</span>
|
|
<span class="n">bias</span><span class="o">=</span><span class="n">dense_bias</span><span class="p">,</span>
|
|
<span class="n">dtype</span><span class="o">=</span><span class="n">dtype</span><span class="p">,</span>
|
|
<span class="n">tp_group</span><span class="o">=</span><span class="n">tp_group</span><span class="p">,</span>
|
|
<span class="n">tp_size</span><span class="o">=</span><span class="n">tp_size</span><span class="p">,</span>
|
|
<span class="n">instance_id</span><span class="o">=</span><span class="n">instance_id</span><span class="p">)</span>
|
|
|
|
<span class="c1"># per-layer relative attention table</span>
|
|
<span class="k">if</span> <span class="n">relative_attention</span><span class="p">:</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">rel_attn_table</span> <span class="o">=</span> <span class="n">Parameter</span><span class="p">(</span><span class="n">shape</span><span class="o">=</span><span class="p">(</span><span class="n">num_attention_heads</span> <span class="o">//</span>
|
|
<span class="n">tp_size</span><span class="p">,</span> <span class="n">num_buckets</span><span class="p">),</span>
|
|
<span class="n">dtype</span><span class="o">=</span><span class="n">dtype</span><span class="p">)</span>
|
|
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">qkv_lora</span> <span class="o">=</span> <span class="n">Lora</span><span class="p">(</span>
|
|
<span class="n">in_hidden_size</span><span class="o">=</span><span class="n">hidden_size</span><span class="p">,</span>
|
|
<span class="n">out_hidden_size</span><span class="o">=</span><span class="n">hidden_size</span> <span class="o">+</span>
|
|
<span class="p">(</span><span class="mi">2</span> <span class="o">*</span> <span class="n">tp_size</span> <span class="o">*</span> <span class="bp">self</span><span class="o">.</span><span class="n">num_attention_kv_heads</span> <span class="o">*</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">attention_head_size</span><span class="p">),</span>
|
|
<span class="n">max_low_rank</span><span class="o">=</span><span class="n">hidden_size</span><span class="p">,</span>
|
|
<span class="p">)</span>
|
|
|
|
<div class="viewcode-block" id="Attention.forward">
|
|
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.layers.html#tensorrt_llm.layers.attention.Attention.forward">[docs]</a>
|
|
<span class="k">def</span> <span class="nf">forward</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span>
|
|
<span class="n">hidden_states</span><span class="p">:</span> <span class="n">Tensor</span><span class="p">,</span>
|
|
<span class="n">attention_mask</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">use_cache</span><span class="o">=</span><span class="kc">False</span><span class="p">,</span>
|
|
<span class="n">kv_cache_params</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">attention_params</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">encoder_output</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="n">Tensor</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">workspace</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">position_embedding</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">norm_before_bmm1</span><span class="o">=</span><span class="kc">False</span><span class="p">,</span>
|
|
<span class="n">lora_params</span><span class="o">=</span><span class="kc">None</span><span class="p">):</span>
|
|
|
|
<span class="k">assert</span> <span class="nb">isinstance</span><span class="p">(</span><span class="n">hidden_states</span><span class="p">,</span> <span class="n">Tensor</span><span class="p">)</span>
|
|
|
|
<span class="n">alibi_slopes</span> <span class="o">=</span> <span class="kc">None</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">position_embedding_type</span><span class="o">.</span><span class="n">is_alibi</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">float32</span>
|
|
<span class="k">if</span> <span class="n">default_net</span><span class="p">()</span><span class="o">.</span><span class="n">plugin_config</span><span class="o">.</span><span class="n">gpt_attention_plugin</span><span class="p">:</span>
|
|
<span class="n">dtype</span> <span class="o">=</span> <span class="n">hidden_states</span><span class="o">.</span><span class="n">dtype</span>
|
|
<span class="n">alibi_scale</span> <span class="o">=</span> <span class="mf">1.</span> <span class="o">/</span> <span class="bp">self</span><span class="o">.</span><span class="n">norm_factor</span> <span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">scale_alibi_bias</span> <span class="k">else</span> <span class="mf">1.</span>
|
|
<span class="n">alibi_slopes</span> <span class="o">=</span> <span class="n">generate_alibi_slopes</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">num_attention_heads</span> <span class="o">*</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">tp_size</span><span class="p">,</span>
|
|
<span class="n">dtype</span><span class="o">=</span><span class="n">dtype</span><span class="p">,</span>
|
|
<span class="n">tp_size</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">tp_size</span><span class="p">,</span>
|
|
<span class="n">tp_rank</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">tp_rank</span><span class="p">,</span>
|
|
<span class="n">alibi_scale</span><span class="o">=</span><span class="n">alibi_scale</span><span class="p">)</span>
|
|
|
|
<span class="n">qkv</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">qkv</span><span class="p">(</span><span class="n">hidden_states</span><span class="p">)</span>
|
|
|
|
<span class="k">if</span> <span class="n">default_net</span><span class="p">()</span><span class="o">.</span><span class="n">plugin_config</span><span class="o">.</span><span class="n">lora_plugin</span><span class="p">:</span>
|
|
<span class="n">qkv</span> <span class="o">=</span> <span class="n">qkv</span> <span class="o">+</span> <span class="bp">self</span><span class="o">.</span><span class="n">qkv_lora</span><span class="p">(</span>
|
|
<span class="n">hidden_states</span><span class="p">,</span>
|
|
<span class="n">host_request_types</span><span class="o">=</span><span class="n">attention_params</span><span class="o">.</span><span class="n">host_request_types</span><span class="p">,</span>
|
|
<span class="n">host_context_lengths</span><span class="o">=</span><span class="n">attention_params</span><span class="o">.</span><span class="n">host_context_lengths</span><span class="p">,</span>
|
|
<span class="n">max_context_length</span><span class="o">=</span><span class="n">attention_params</span><span class="o">.</span><span class="n">max_context_length</span><span class="p">,</span>
|
|
<span class="n">lora_ranks</span><span class="o">=</span><span class="n">lora_params</span><span class="o">.</span><span class="n">lora_ranks</span><span class="p">,</span>
|
|
<span class="n">lora_weights_pointers</span><span class="o">=</span><span class="n">lora_params</span><span class="o">.</span><span class="n">lora_weights_pointers_list</span><span class="p">[</span><span class="mi">0</span><span class="p">])</span>
|
|
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">position_embedding_type</span> <span class="o">==</span> <span class="n">PositionEmbeddingType</span><span class="o">.</span><span class="n">chatglm</span><span class="p">:</span>
|
|
<span class="n">qkv</span> <span class="o">=</span> <span class="n">RopeEmbeddingUtils</span><span class="o">.</span><span class="n">apply_rotary_pos_emb_chatglm</span><span class="p">(</span>
|
|
<span class="n">qkv</span><span class="p">,</span>
|
|
<span class="n">position_embedding</span><span class="p">,</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">num_attention_heads</span><span class="p">,</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">attention_head_size</span><span class="p">,</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">max_position_embeddings</span><span class="p">,</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">rotary_embedding_scale</span><span class="p">,</span>
|
|
<span class="p">)</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">rotary_embedding_scale_type</span> <span class="o">=</span> <span class="n">RotaryScalingType</span><span class="o">.</span><span class="n">none</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">rotary_embedding_scale</span> <span class="o">=</span> <span class="mf">1.0</span>
|
|
|
|
<span class="n">paged_kv_cache</span> <span class="o">=</span> <span class="n">default_net</span><span class="p">()</span><span class="o">.</span><span class="n">plugin_config</span><span class="o">.</span><span class="n">paged_kv_cache</span>
|
|
|
|
<span class="k">assert</span> <span class="n">attention_params</span> <span class="ow">is</span> <span class="kc">None</span> <span class="ow">or</span> <span class="n">attention_params</span><span class="o">.</span><span class="n">is_valid</span><span class="p">(</span>
|
|
<span class="n">default_net</span><span class="p">()</span><span class="o">.</span><span class="n">plugin_config</span><span class="o">.</span><span class="n">gpt_attention_plugin</span><span class="p">,</span>
|
|
<span class="n">default_net</span><span class="p">()</span><span class="o">.</span><span class="n">plugin_config</span><span class="o">.</span><span class="n">remove_input_padding</span><span class="p">)</span>
|
|
<span class="k">assert</span> <span class="n">kv_cache_params</span> <span class="ow">is</span> <span class="kc">None</span> <span class="ow">or</span> <span class="n">kv_cache_params</span><span class="o">.</span><span class="n">is_valid</span><span class="p">(</span>
|
|
<span class="n">default_net</span><span class="p">()</span><span class="o">.</span><span class="n">plugin_config</span><span class="o">.</span><span class="n">gpt_attention_plugin</span><span class="p">)</span>
|
|
|
|
<span class="n">past_key_value</span> <span class="o">=</span> <span class="kc">None</span> <span class="k">if</span> <span class="n">kv_cache_params</span> <span class="ow">is</span> <span class="kc">None</span> <span class="k">else</span> <span class="n">kv_cache_params</span><span class="o">.</span><span class="n">get_first_past_key_value</span><span class="p">(</span>
|
|
<span class="p">)</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">cross_attention</span> <span class="ow">and</span> <span class="p">(</span><span class="n">past_key_value</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">):</span>
|
|
<span class="n">past_key_value</span> <span class="o">=</span> <span class="n">kv_cache_params</span><span class="o">.</span><span class="n">past_key_value</span><span class="p">[</span><span class="mi">1</span><span class="p">]</span>
|
|
|
|
<span class="c1"># if cross attention, cross QKV only needs to be calculated once in the</span>
|
|
<span class="c1"># 1st decoding step --> write to cross KV cache --> remains constant</span>
|
|
<span class="c1"># during the entire decoding. 1st and >1 steps are distinguished by</span>
|
|
<span class="c1"># whether past_key_value exists or not</span>
|
|
<span class="c1"># also, cross KV cache max length is set from encoder output seqlen,</span>
|
|
<span class="c1"># this maps to the max context length concept in decoder-only models</span>
|
|
<span class="n">cross_qkv</span> <span class="o">=</span> <span class="kc">None</span>
|
|
<span class="c1"># get length data in every run</span>
|
|
<span class="k">if</span> <span class="n">encoder_output</span><span class="p">:</span>
|
|
<span class="k">assert</span> <span class="nb">isinstance</span><span class="p">(</span><span class="n">encoder_output</span><span class="p">,</span> <span class="n">Tensor</span><span class="p">)</span>
|
|
<span class="c1"># but only do projection once at 1st decoding step</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">cross_attention</span> <span class="ow">and</span> <span class="n">encoder_output</span><span class="p">:</span>
|
|
<span class="n">cross_qkv</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">qkv</span><span class="p">(</span><span class="n">encoder_output</span><span class="p">)</span>
|
|
|
|
<span class="k">if</span> <span class="n">default_net</span><span class="p">()</span><span class="o">.</span><span class="n">plugin_config</span><span class="o">.</span><span class="n">gpt_attention_plugin</span><span class="p">:</span>
|
|
<span class="k">assert</span> <span class="bp">self</span><span class="o">.</span><span class="n">attention_mask_type</span> <span class="ow">in</span> <span class="p">[</span>
|
|
<span class="n">AttentionMaskType</span><span class="o">.</span><span class="n">causal</span><span class="p">,</span> <span class="n">AttentionMaskType</span><span class="o">.</span><span class="n">bidirectional</span><span class="p">,</span>
|
|
<span class="n">AttentionMaskType</span><span class="o">.</span><span class="n">bidirectionalglm</span>
|
|
<span class="p">],</span> <span class="s1">'Plugin only support masked MHA.'</span>
|
|
<span class="n">kv_orig_quant_scale</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">kv_orig_quant_scale</span><span class="o">.</span><span class="n">value</span> <span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">quant_mode</span><span class="o">.</span><span class="n">has_kv_cache_quant</span><span class="p">(</span>
|
|
<span class="p">)</span> <span class="k">else</span> <span class="kc">None</span>
|
|
<span class="n">kv_quant_orig_scale</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">kv_quant_orig_scale</span><span class="o">.</span><span class="n">value</span> <span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">quant_mode</span><span class="o">.</span><span class="n">has_kv_cache_quant</span><span class="p">(</span>
|
|
<span class="p">)</span> <span class="k">else</span> <span class="kc">None</span>
|
|
<span class="n">context</span><span class="p">,</span> <span class="n">past_key_value</span> <span class="o">=</span> <span class="n">gpt_attention</span><span class="p">(</span>
|
|
<span class="n">tensor</span><span class="o">=</span><span class="n">qkv</span><span class="p">,</span>
|
|
<span class="n">past_key_value</span><span class="o">=</span><span class="n">past_key_value</span><span class="p">,</span>
|
|
<span class="n">sequence_length</span><span class="o">=</span><span class="n">attention_params</span><span class="o">.</span><span class="n">sequence_length</span><span class="p">,</span>
|
|
<span class="n">host_past_key_value_lengths</span><span class="o">=</span><span class="n">kv_cache_params</span><span class="o">.</span>
|
|
<span class="n">host_past_key_value_lengths</span><span class="p">,</span>
|
|
<span class="n">host_max_kv_cache_lengths</span><span class="o">=</span><span class="n">kv_cache_params</span><span class="o">.</span>
|
|
<span class="n">host_max_kv_cache_lengths</span><span class="p">,</span>
|
|
<span class="n">context_lengths</span><span class="o">=</span><span class="n">attention_params</span><span class="o">.</span><span class="n">context_lengths</span><span class="p">,</span>
|
|
<span class="n">cache_indirection</span><span class="o">=</span><span class="n">kv_cache_params</span><span class="o">.</span><span class="n">cache_indirection</span><span class="p">,</span>
|
|
<span class="n">host_request_types</span><span class="o">=</span><span class="n">attention_params</span><span class="o">.</span><span class="n">host_request_types</span><span class="p">,</span>
|
|
<span class="n">num_heads</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">num_attention_heads</span><span class="p">,</span>
|
|
<span class="n">num_kv_heads</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">num_attention_kv_heads</span><span class="p">,</span>
|
|
<span class="n">hidden_size_per_head</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">attention_head_size</span><span class="p">,</span>
|
|
<span class="n">q_scaling</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">q_scaling</span><span class="p">,</span>
|
|
<span class="n">rotary_embedding_dim</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">rotary_embedding_dim</span><span class="p">,</span>
|
|
<span class="n">rotary_embedding_base</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">rotary_embedding_base</span><span class="p">,</span>
|
|
<span class="n">rotary_embedding_scale_type</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">rotary_embedding_scale_type</span><span class="p">,</span>
|
|
<span class="n">rotary_embedding_scale</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">rotary_embedding_scale</span><span class="p">,</span>
|
|
<span class="n">rotary_embedding_max_positions</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">max_position_embeddings</span><span class="p">,</span>
|
|
<span class="n">position_embedding_type</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">position_embedding_type</span><span class="p">,</span>
|
|
<span class="n">kv_orig_quant_scale</span><span class="o">=</span><span class="n">kv_orig_quant_scale</span><span class="p">,</span>
|
|
<span class="n">kv_quant_orig_scale</span><span class="o">=</span><span class="n">kv_quant_orig_scale</span><span class="p">,</span>
|
|
<span class="n">kv_cache_quant_mode</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">quant_mode</span><span class="p">,</span>
|
|
<span class="n">max_context_length</span><span class="o">=</span><span class="n">attention_params</span><span class="o">.</span><span class="n">max_context_length</span><span class="p">,</span>
|
|
<span class="n">mask_type</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">attention_mask_type</span><span class="p">,</span>
|
|
<span class="n">alibi_slopes</span><span class="o">=</span><span class="n">alibi_slopes</span><span class="p">,</span>
|
|
<span class="n">tp_size</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">tp_size</span><span class="p">,</span>
|
|
<span class="n">tp_rank</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">tp_rank</span><span class="p">,</span>
|
|
<span class="n">kv_cache_block_pointers</span><span class="o">=</span><span class="n">kv_cache_params</span><span class="o">.</span>
|
|
<span class="n">get_first_kv_cache_block_pointers</span><span class="p">(),</span>
|
|
<span class="n">do_cross_attention</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">cross_attention</span><span class="p">,</span>
|
|
<span class="n">cross_qkv</span><span class="o">=</span><span class="n">cross_qkv</span><span class="p">,</span>
|
|
<span class="n">cross_qkv_length</span><span class="o">=</span><span class="n">attention_params</span><span class="o">.</span><span class="n">encoder_max_input_length</span><span class="p">,</span>
|
|
<span class="n">encoder_input_lengths</span><span class="o">=</span><span class="n">attention_params</span><span class="o">.</span><span class="n">encoder_input_lengths</span><span class="p">,</span>
|
|
<span class="n">relative_attention_bias</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">rel_attn_table</span><span class="o">.</span><span class="n">value</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">relative_attention</span> <span class="k">else</span> <span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">max_distance</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">max_distance</span><span class="p">,</span>
|
|
<span class="n">host_context_lengths</span><span class="o">=</span><span class="n">attention_params</span><span class="o">.</span><span class="n">host_context_lengths</span><span class="p">,</span>
|
|
<span class="p">)</span>
|
|
|
|
<span class="k">else</span><span class="p">:</span>
|
|
<span class="c1"># plain TensorRT mode</span>
|
|
<span class="k">assert</span> <span class="n">paged_kv_cache</span> <span class="o">==</span> <span class="kc">False</span>
|
|
<span class="n">past_key_value</span> <span class="o">=</span> <span class="kc">None</span> <span class="k">if</span> <span class="n">kv_cache_params</span> <span class="ow">is</span> <span class="kc">None</span> <span class="k">else</span> <span class="n">kv_cache_params</span><span class="o">.</span><span class="n">get_first_past_key_value</span><span class="p">(</span>
|
|
<span class="p">)</span>
|
|
|
|
<span class="k">def</span> <span class="nf">transpose_for_scores</span><span class="p">(</span><span class="n">x</span><span class="p">,</span>
|
|
<span class="n">rotary</span><span class="p">:</span> <span class="nb">bool</span> <span class="o">=</span> <span class="kc">False</span><span class="p">,</span>
|
|
<span class="n">is_kv</span><span class="p">:</span> <span class="nb">bool</span> <span class="o">=</span> <span class="kc">False</span><span class="p">):</span>
|
|
<span class="n">_num_attention_heads</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">num_attention_kv_heads</span> <span class="k">if</span> <span class="n">is_kv</span> <span class="k">else</span> <span class="bp">self</span><span class="o">.</span><span class="n">num_attention_heads</span>
|
|
<span class="n">new_x_shape</span> <span class="o">=</span> <span class="n">concat</span><span class="p">([</span>
|
|
<span class="n">shape</span><span class="p">(</span><span class="n">x</span><span class="p">,</span> <span class="mi">0</span><span class="p">),</span>
|
|
<span class="n">shape</span><span class="p">(</span><span class="n">x</span><span class="p">,</span> <span class="mi">1</span><span class="p">),</span> <span class="n">_num_attention_heads</span><span class="p">,</span> <span class="bp">self</span><span class="o">.</span><span class="n">attention_head_size</span>
|
|
<span class="p">])</span>
|
|
<span class="k">if</span> <span class="n">rotary</span><span class="p">:</span>
|
|
<span class="k">return</span> <span class="n">x</span><span class="o">.</span><span class="n">view</span><span class="p">(</span><span class="n">new_x_shape</span><span class="p">)</span>
|
|
<span class="k">else</span><span class="p">:</span>
|
|
<span class="k">return</span> <span class="n">x</span><span class="o">.</span><span class="n">view</span><span class="p">(</span><span class="n">new_x_shape</span><span class="p">)</span><span class="o">.</span><span class="n">permute</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="mi">1</span><span class="p">,</span> <span class="mi">3</span><span class="p">])</span>
|
|
|
|
<span class="c1"># qkv after projection is of shape</span>
|
|
<span class="c1"># [bs, seqlen, (num_attention_heads + 2 * num_attention_kv_heads), attention_head_size].</span>
|
|
<span class="c1"># The projected and split qkv after transpose_for_scores():</span>
|
|
<span class="c1"># Q[bs, num_attention_heads, seqlen, attention_head_size]</span>
|
|
<span class="c1"># K[bs, num_attention_kv_heads, seqlen, attention_head_size]</span>
|
|
<span class="c1"># V[bs, num_attention_kv_heads, seqlen, attention_head_size]</span>
|
|
<span class="n">kv_size</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">attention_head_size</span> <span class="o">*</span> <span class="bp">self</span><span class="o">.</span><span class="n">num_attention_kv_heads</span>
|
|
<span class="n">query</span><span class="p">,</span> <span class="n">key</span><span class="p">,</span> <span class="n">value</span> <span class="o">=</span> <span class="n">split</span><span class="p">(</span><span class="n">qkv</span><span class="p">,</span> <span class="p">[</span><span class="bp">self</span><span class="o">.</span><span class="n">hidden_size</span><span class="p">,</span> <span class="n">kv_size</span><span class="p">,</span> <span class="n">kv_size</span><span class="p">],</span>
|
|
<span class="n">dim</span><span class="o">=</span><span class="mi">2</span><span class="p">)</span>
|
|
|
|
<span class="c1"># in cross attention mode, replace kv by encoder_output</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">cross_attention</span> <span class="ow">and</span> <span class="n">encoder_output</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="n">encoder_qkv</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">qkv</span><span class="p">(</span><span class="n">encoder_output</span><span class="p">)</span>
|
|
<span class="n">_</span><span class="p">,</span> <span class="n">key</span><span class="p">,</span> <span class="n">value</span> <span class="o">=</span> <span class="n">split</span><span class="p">(</span><span class="n">encoder_qkv</span><span class="p">,</span>
|
|
<span class="p">[</span><span class="bp">self</span><span class="o">.</span><span class="n">hidden_size</span><span class="p">,</span> <span class="n">kv_size</span><span class="p">,</span> <span class="n">kv_size</span><span class="p">],</span>
|
|
<span class="n">dim</span><span class="o">=</span><span class="mi">2</span><span class="p">)</span>
|
|
|
|
<span class="n">query</span> <span class="o">=</span> <span class="n">transpose_for_scores</span><span class="p">(</span><span class="n">query</span><span class="p">,</span> <span class="n">rotary</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">rotary_enabled</span><span class="p">)</span>
|
|
<span class="n">key</span> <span class="o">=</span> <span class="n">transpose_for_scores</span><span class="p">(</span><span class="n">key</span><span class="p">,</span>
|
|
<span class="n">is_kv</span><span class="o">=</span><span class="kc">True</span><span class="p">,</span>
|
|
<span class="n">rotary</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">rotary_enabled</span><span class="p">)</span>
|
|
<span class="n">value</span> <span class="o">=</span> <span class="n">transpose_for_scores</span><span class="p">(</span><span class="n">value</span><span class="p">,</span> <span class="n">is_kv</span><span class="o">=</span><span class="kc">True</span><span class="p">)</span>
|
|
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">rotary_enabled</span><span class="p">:</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">dtype</span> <span class="o">==</span> <span class="n">trt</span><span class="o">.</span><span class="n">bfloat16</span><span class="p">:</span>
|
|
<span class="n">embed_positions</span> <span class="o">=</span> <span class="n">numpy_fp32_to_bf16</span><span class="p">(</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">embed_positions</span><span class="o">.</span><span class="n">astype</span><span class="p">(</span><span class="n">np</span><span class="o">.</span><span class="n">float32</span><span class="p">))</span>
|
|
<span class="n">embed_positions</span> <span class="o">=</span> <span class="n">constant</span><span class="p">(</span><span class="n">embed_positions</span><span class="p">)</span>
|
|
<span class="k">else</span><span class="p">:</span>
|
|
<span class="n">embed_positions</span> <span class="o">=</span> <span class="n">constant</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">embed_positions</span><span class="p">)</span>
|
|
|
|
<span class="k">if</span> <span class="n">default_net</span><span class="p">()</span><span class="o">.</span><span class="n">strongly_typed</span> <span class="ow">and</span> <span class="p">(</span><span class="n">embed_positions</span><span class="o">.</span><span class="n">dtype</span> <span class="o">!=</span>
|
|
<span class="n">value</span><span class="o">.</span><span class="n">dtype</span><span class="p">):</span>
|
|
<span class="n">embed_positions</span> <span class="o">=</span> <span class="n">cast</span><span class="p">(</span><span class="n">embed_positions</span><span class="p">,</span> <span class="n">value</span><span class="o">.</span><span class="n">dtype</span><span class="p">)</span>
|
|
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">rotary_embedding_dim</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="c1"># When shape(hidden_states, 1) > 1(Context phase), the embedding start from 0,</span>
|
|
<span class="c1"># otherwise (Generation phase) move start to position</span>
|
|
<span class="n">start</span> <span class="o">=</span> <span class="n">where</span><span class="p">(</span>
|
|
<span class="n">shape</span><span class="p">(</span><span class="n">hidden_states</span><span class="p">,</span> <span class="mi">1</span><span class="p">)</span> <span class="o">></span> <span class="mi">1</span><span class="p">,</span> <span class="mi">0</span><span class="p">,</span>
|
|
<span class="n">shape</span><span class="p">(</span><span class="n">past_key_value</span><span class="p">,</span> <span class="mi">3</span><span class="p">))</span>
|
|
<span class="n">size</span> <span class="o">=</span> <span class="n">where</span><span class="p">(</span>
|
|
<span class="n">shape</span><span class="p">(</span><span class="n">hidden_states</span><span class="p">,</span> <span class="mi">1</span><span class="p">)</span> <span class="o">></span> <span class="mi">1</span><span class="p">,</span> <span class="n">shape</span><span class="p">(</span><span class="n">hidden_states</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="n">sincos</span> <span class="o">=</span> <span class="nb">slice</span><span class="p">(</span><span class="n">embed_positions</span><span class="p">,</span> <span class="n">concat</span><span class="p">([</span><span class="mi">0</span><span class="p">,</span> <span class="n">start</span><span class="p">,</span> <span class="mi">0</span><span class="p">]),</span>
|
|
<span class="n">concat</span><span class="p">([</span><span class="mi">1</span><span class="p">,</span> <span class="n">size</span><span class="p">,</span> <span class="bp">self</span><span class="o">.</span><span class="n">rotary_embedding_dim</span><span class="p">]))</span>
|
|
<span class="n">sin</span><span class="p">,</span> <span class="n">cos</span> <span class="o">=</span> <span class="n">split</span><span class="p">(</span><span class="n">sincos</span><span class="p">,</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">rotary_embedding_dim</span> <span class="o">//</span> <span class="mi">2</span><span class="p">,</span>
|
|
<span class="n">dim</span><span class="o">=-</span><span class="mi">1</span><span class="p">)</span>
|
|
|
|
<span class="n">key_rot_size</span> <span class="o">=</span> <span class="n">concat</span><span class="p">([</span>
|
|
<span class="n">shape</span><span class="p">(</span><span class="n">key</span><span class="p">,</span> <span class="mi">0</span><span class="p">),</span>
|
|
<span class="n">shape</span><span class="p">(</span><span class="n">key</span><span class="p">,</span> <span class="mi">1</span><span class="p">),</span>
|
|
<span class="n">shape</span><span class="p">(</span><span class="n">key</span><span class="p">,</span> <span class="mi">2</span><span class="p">),</span> <span class="bp">self</span><span class="o">.</span><span class="n">rotary_embedding_dim</span>
|
|
<span class="p">])</span>
|
|
<span class="n">query_rot_size</span> <span class="o">=</span> <span class="n">concat</span><span class="p">([</span>
|
|
<span class="n">shape</span><span class="p">(</span><span class="n">query</span><span class="p">,</span> <span class="mi">0</span><span class="p">),</span>
|
|
<span class="n">shape</span><span class="p">(</span><span class="n">query</span><span class="p">,</span> <span class="mi">1</span><span class="p">),</span>
|
|
<span class="n">shape</span><span class="p">(</span><span class="n">query</span><span class="p">,</span> <span class="mi">2</span><span class="p">),</span> <span class="bp">self</span><span class="o">.</span><span class="n">rotary_embedding_dim</span>
|
|
<span class="p">])</span>
|
|
<span class="n">remaining</span> <span class="o">=</span> <span class="n">shape</span><span class="p">(</span><span class="n">key</span><span class="p">,</span> <span class="mi">3</span><span class="p">)</span> <span class="o">-</span> <span class="bp">self</span><span class="o">.</span><span class="n">rotary_embedding_dim</span>
|
|
<span class="n">key_pass_size</span> <span class="o">=</span> <span class="n">concat</span><span class="p">([</span>
|
|
<span class="n">shape</span><span class="p">(</span><span class="n">key</span><span class="p">,</span> <span class="mi">0</span><span class="p">),</span>
|
|
<span class="n">shape</span><span class="p">(</span><span class="n">key</span><span class="p">,</span> <span class="mi">1</span><span class="p">),</span>
|
|
<span class="n">shape</span><span class="p">(</span><span class="n">key</span><span class="p">,</span> <span class="mi">2</span><span class="p">),</span> <span class="n">remaining</span>
|
|
<span class="p">])</span>
|
|
<span class="n">query_pass_size</span> <span class="o">=</span> <span class="n">concat</span><span class="p">([</span>
|
|
<span class="n">shape</span><span class="p">(</span><span class="n">query</span><span class="p">,</span> <span class="mi">0</span><span class="p">),</span>
|
|
<span class="n">shape</span><span class="p">(</span><span class="n">query</span><span class="p">,</span> <span class="mi">1</span><span class="p">),</span>
|
|
<span class="n">shape</span><span class="p">(</span><span class="n">query</span><span class="p">,</span> <span class="mi">2</span><span class="p">),</span> <span class="n">remaining</span>
|
|
<span class="p">])</span>
|
|
<span class="n">k_rot</span> <span class="o">=</span> <span class="nb">slice</span><span class="p">(</span><span class="n">key</span><span class="p">,</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="mi">0</span><span class="p">,</span> <span class="mi">0</span><span class="p">],</span> <span class="n">key_rot_size</span><span class="p">)</span>
|
|
<span class="n">k_pass</span> <span class="o">=</span> <span class="nb">slice</span><span class="p">(</span><span class="n">key</span><span class="p">,</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="mi">0</span><span class="p">,</span> <span class="bp">self</span><span class="o">.</span><span class="n">rotary_embedding_dim</span><span class="p">],</span>
|
|
<span class="n">key_pass_size</span><span class="p">)</span>
|
|
|
|
<span class="n">q_rot</span> <span class="o">=</span> <span class="nb">slice</span><span class="p">(</span><span class="n">query</span><span class="p">,</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="mi">0</span><span class="p">,</span> <span class="mi">0</span><span class="p">],</span> <span class="n">query_rot_size</span><span class="p">)</span>
|
|
<span class="n">q_pass</span> <span class="o">=</span> <span class="nb">slice</span><span class="p">(</span><span class="n">query</span><span class="p">,</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="mi">0</span><span class="p">,</span> <span class="bp">self</span><span class="o">.</span><span class="n">rotary_embedding_dim</span><span class="p">],</span>
|
|
<span class="n">query_pass_size</span><span class="p">)</span>
|
|
|
|
<span class="n">k_rot</span> <span class="o">=</span> <span class="n">RopeEmbeddingUtils</span><span class="o">.</span><span class="n">apply_rotary_pos_emb</span><span class="p">(</span>
|
|
<span class="n">k_rot</span><span class="p">,</span> <span class="p">[</span><span class="n">cos</span><span class="p">,</span> <span class="n">sin</span><span class="p">],</span> <span class="bp">self</span><span class="o">.</span><span class="n">position_embedding_type</span><span class="p">)</span>
|
|
<span class="n">q_rot</span> <span class="o">=</span> <span class="n">RopeEmbeddingUtils</span><span class="o">.</span><span class="n">apply_rotary_pos_emb</span><span class="p">(</span>
|
|
<span class="n">q_rot</span><span class="p">,</span> <span class="p">[</span><span class="n">cos</span><span class="p">,</span> <span class="n">sin</span><span class="p">],</span> <span class="bp">self</span><span class="o">.</span><span class="n">position_embedding_type</span><span class="p">)</span>
|
|
|
|
<span class="n">key</span> <span class="o">=</span> <span class="n">concat</span><span class="p">([</span><span class="n">k_rot</span><span class="p">,</span> <span class="n">k_pass</span><span class="p">],</span> <span class="n">dim</span><span class="o">=</span><span class="mi">3</span><span class="p">)</span>
|
|
<span class="n">query</span> <span class="o">=</span> <span class="n">concat</span><span class="p">([</span><span class="n">q_rot</span><span class="p">,</span> <span class="n">q_pass</span><span class="p">],</span> <span class="n">dim</span><span class="o">=</span><span class="mi">3</span><span class="p">)</span>
|
|
<span class="k">else</span><span class="p">:</span>
|
|
<span class="n">key</span> <span class="o">=</span> <span class="n">RopeEmbeddingUtils</span><span class="o">.</span><span class="n">apply_rotary_pos_emb</span><span class="p">(</span>
|
|
<span class="n">key</span><span class="p">,</span> <span class="p">[</span><span class="n">cos</span><span class="p">,</span> <span class="n">sin</span><span class="p">],</span> <span class="bp">self</span><span class="o">.</span><span class="n">position_embedding_type</span><span class="p">)</span>
|
|
<span class="n">query</span> <span class="o">=</span> <span class="n">RopeEmbeddingUtils</span><span class="o">.</span><span class="n">apply_rotary_pos_emb</span><span class="p">(</span>
|
|
<span class="n">query</span><span class="p">,</span> <span class="p">[</span><span class="n">cos</span><span class="p">,</span> <span class="n">sin</span><span class="p">],</span> <span class="bp">self</span><span class="o">.</span><span class="n">position_embedding_type</span><span class="p">)</span>
|
|
|
|
<span class="n">key</span> <span class="o">=</span> <span class="n">key</span><span class="o">.</span><span class="n">permute</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="mi">1</span><span class="p">,</span> <span class="mi">3</span><span class="p">])</span>
|
|
<span class="n">query</span> <span class="o">=</span> <span class="n">query</span><span class="o">.</span><span class="n">permute</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="mi">1</span><span class="p">,</span> <span class="mi">3</span><span class="p">])</span>
|
|
|
|
<span class="n">past_key_value</span> <span class="o">=</span> <span class="kc">None</span> <span class="k">if</span> <span class="n">kv_cache_params</span> <span class="ow">is</span> <span class="kc">None</span> <span class="k">else</span> <span class="n">kv_cache_params</span><span class="o">.</span><span class="n">get_first_past_key_value</span><span class="p">(</span>
|
|
<span class="p">)</span>
|
|
<span class="k">if</span> <span class="n">past_key_value</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
|
|
|
|
<span class="k">def</span> <span class="nf">dequantize_tensor</span><span class="p">(</span><span class="n">x</span><span class="p">,</span> <span class="n">scale</span><span class="p">):</span>
|
|
<span class="c1"># Cast from int8 to dtype</span>
|
|
<span class="n">casted_x</span> <span class="o">=</span> <span class="n">cast</span><span class="p">(</span><span class="n">x</span><span class="p">,</span> <span class="bp">self</span><span class="o">.</span><span class="n">dtype</span><span class="p">)</span>
|
|
<span class="k">return</span> <span class="n">casted_x</span> <span class="o">*</span> <span class="n">scale</span>
|
|
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">use_int8_kv_cache</span><span class="p">:</span>
|
|
<span class="n">past_key_value</span> <span class="o">=</span> <span class="n">dequantize_tensor</span><span class="p">(</span>
|
|
<span class="n">past_key_value</span><span class="p">,</span> <span class="bp">self</span><span class="o">.</span><span class="n">kv_quant_orig_scale</span><span class="o">.</span><span class="n">value</span><span class="p">)</span>
|
|
|
|
<span class="c1"># past_key_value [bs, 2, num_heads, max_seq_len, head_dim]</span>
|
|
<span class="n">past_key</span><span class="p">,</span> <span class="n">past_value</span> <span class="o">=</span> <span class="n">split</span><span class="p">(</span><span class="n">past_key_value</span><span class="p">,</span> <span class="mi">1</span><span class="p">,</span> <span class="n">dim</span><span class="o">=</span><span class="mi">1</span><span class="p">)</span>
|
|
|
|
<span class="n">key_shape</span> <span class="o">=</span> <span class="n">concat</span><span class="p">([</span>
|
|
<span class="n">shape</span><span class="p">(</span><span class="n">past_key</span><span class="p">,</span> <span class="mi">0</span><span class="p">),</span>
|
|
<span class="n">shape</span><span class="p">(</span><span class="n">past_key</span><span class="p">,</span> <span class="mi">2</span><span class="p">),</span>
|
|
<span class="n">shape</span><span class="p">(</span><span class="n">past_key</span><span class="p">,</span> <span class="mi">3</span><span class="p">),</span>
|
|
<span class="n">shape</span><span class="p">(</span><span class="n">past_key</span><span class="p">,</span> <span class="mi">4</span><span class="p">)</span>
|
|
<span class="p">])</span>
|
|
<span class="n">past_key</span> <span class="o">=</span> <span class="n">past_key</span><span class="o">.</span><span class="n">view</span><span class="p">(</span><span class="n">key_shape</span><span class="p">,</span> <span class="n">zero_is_placeholder</span><span class="o">=</span><span class="kc">False</span><span class="p">)</span>
|
|
<span class="n">past_value</span> <span class="o">=</span> <span class="n">past_value</span><span class="o">.</span><span class="n">view</span><span class="p">(</span><span class="n">key_shape</span><span class="p">,</span>
|
|
<span class="n">zero_is_placeholder</span><span class="o">=</span><span class="kc">False</span><span class="p">)</span>
|
|
|
|
<span class="n">key</span> <span class="o">=</span> <span class="n">concat</span><span class="p">([</span><span class="n">past_key</span><span class="p">,</span> <span class="n">key</span><span class="p">],</span> <span class="n">dim</span><span class="o">=</span><span class="mi">2</span><span class="p">)</span><span class="o">.</span><span class="n">cast</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">dtype</span><span class="p">)</span>
|
|
<span class="n">value</span> <span class="o">=</span> <span class="n">concat</span><span class="p">([</span><span class="n">past_value</span><span class="p">,</span> <span class="n">value</span><span class="p">],</span> <span class="n">dim</span><span class="o">=</span><span class="mi">2</span><span class="p">)</span><span class="o">.</span><span class="n">cast</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">dtype</span><span class="p">)</span>
|
|
|
|
<span class="k">if</span> <span class="n">use_cache</span><span class="p">:</span>
|
|
<span class="n">key_inflated_shape</span> <span class="o">=</span> <span class="n">concat</span><span class="p">([</span>
|
|
<span class="n">shape</span><span class="p">(</span><span class="n">key</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">shape</span><span class="p">(</span><span class="n">key</span><span class="p">,</span> <span class="mi">1</span><span class="p">),</span>
|
|
<span class="n">shape</span><span class="p">(</span><span class="n">key</span><span class="p">,</span> <span class="mi">2</span><span class="p">),</span>
|
|
<span class="n">shape</span><span class="p">(</span><span class="n">key</span><span class="p">,</span> <span class="mi">3</span><span class="p">)</span>
|
|
<span class="p">])</span>
|
|
<span class="n">inflated_key</span> <span class="o">=</span> <span class="n">key</span><span class="o">.</span><span class="n">view</span><span class="p">(</span><span class="n">key_inflated_shape</span><span class="p">,</span>
|
|
<span class="n">zero_is_placeholder</span><span class="o">=</span><span class="kc">False</span><span class="p">)</span>
|
|
<span class="n">inflated_value</span> <span class="o">=</span> <span class="n">value</span><span class="o">.</span><span class="n">view</span><span class="p">(</span><span class="n">key_inflated_shape</span><span class="p">,</span>
|
|
<span class="n">zero_is_placeholder</span><span class="o">=</span><span class="kc">False</span><span class="p">)</span>
|
|
<span class="n">past_key_value</span> <span class="o">=</span> <span class="n">concat</span><span class="p">([</span><span class="n">inflated_key</span><span class="p">,</span> <span class="n">inflated_value</span><span class="p">],</span> <span class="n">dim</span><span class="o">=</span><span class="mi">1</span><span class="p">)</span>
|
|
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">use_int8_kv_cache</span><span class="p">:</span>
|
|
|
|
<span class="k">def</span> <span class="nf">quantize_tensor</span><span class="p">(</span><span class="n">x</span><span class="p">,</span> <span class="n">scale</span><span class="p">):</span>
|
|
<span class="n">scaled</span> <span class="o">=</span> <span class="n">x</span> <span class="o">*</span> <span class="n">scale</span>
|
|
<span class="n">rounded</span> <span class="o">=</span> <span class="nb">round</span><span class="p">(</span><span class="n">scaled</span><span class="p">)</span>
|
|
<span class="n">clipped</span> <span class="o">=</span> <span class="n">clip</span><span class="p">(</span><span class="n">rounded</span><span class="p">,</span> <span class="o">-</span><span class="mi">128</span><span class="p">,</span> <span class="mi">127</span><span class="p">)</span>
|
|
<span class="n">quantized</span> <span class="o">=</span> <span class="n">cast</span><span class="p">(</span><span class="n">clipped</span><span class="p">,</span> <span class="s1">'int8'</span><span class="p">)</span>
|
|
<span class="k">return</span> <span class="n">quantized</span>
|
|
|
|
<span class="n">past_key_value</span> <span class="o">=</span> <span class="n">quantize_tensor</span><span class="p">(</span>
|
|
<span class="n">past_key_value</span><span class="p">,</span> <span class="bp">self</span><span class="o">.</span><span class="n">kv_orig_quant_scale</span><span class="o">.</span><span class="n">value</span><span class="p">)</span>
|
|
|
|
<span class="c1"># MQA broadcast</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">num_attention_heads</span> <span class="o">//</span> <span class="bp">self</span><span class="o">.</span><span class="n">num_attention_kv_heads</span> <span class="o">></span> <span class="mi">1</span><span class="p">:</span>
|
|
<span class="n">key</span> <span class="o">=</span> <span class="n">repeat_interleave</span><span class="p">(</span>
|
|
<span class="n">key</span><span class="p">,</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">num_attention_heads</span> <span class="o">//</span> <span class="bp">self</span><span class="o">.</span><span class="n">num_attention_kv_heads</span><span class="p">,</span> <span class="mi">1</span><span class="p">)</span>
|
|
<span class="n">value</span> <span class="o">=</span> <span class="n">repeat_interleave</span><span class="p">(</span>
|
|
<span class="n">value</span><span class="p">,</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">num_attention_heads</span> <span class="o">//</span> <span class="bp">self</span><span class="o">.</span><span class="n">num_attention_kv_heads</span><span class="p">,</span> <span class="mi">1</span><span class="p">)</span>
|
|
|
|
<span class="n">key_length</span> <span class="o">=</span> <span class="n">shape</span><span class="p">(</span><span class="n">key</span><span class="p">,</span> <span class="mi">2</span><span class="p">)</span>
|
|
|
|
<span class="c1"># The following code creates a 2D tensor with 0s in the lower triangular (including the diagonal) and</span>
|
|
<span class="c1"># +INF in the upper triangular parts. This bias tensor will be added to the output of the Q*K^T matrix</span>
|
|
<span class="c1"># multiplication (BMM1). The +INF elements will be transformed to 0s by the Softmax operator that</span>
|
|
<span class="c1"># follows. The elements that corresponds to 0s in the bias are unaffected by the bias tensor.</span>
|
|
<span class="c1">#</span>
|
|
<span class="c1"># Note that when we added to another bias tensor B (for example, with AliBi), the values in the lower-</span>
|
|
<span class="c1"># triangular part of the B tensor are not affected and the upper-triangular ones are set to +INF.</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">attention_mask_type</span> <span class="o">==</span> <span class="n">AttentionMaskType</span><span class="o">.</span><span class="n">causal</span><span class="p">:</span>
|
|
<span class="n">query_length</span> <span class="o">=</span> <span class="n">shape</span><span class="p">(</span><span class="n">query</span><span class="p">,</span> <span class="mi">2</span><span class="p">)</span>
|
|
<span class="n">starts</span> <span class="o">=</span> <span class="n">concat</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="n">key_length</span> <span class="o">-</span> <span class="n">query_length</span><span class="p">,</span> <span class="mi">0</span><span class="p">])</span>
|
|
<span class="n">sizes</span> <span class="o">=</span> <span class="n">concat</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="n">query_length</span><span class="p">,</span> <span class="n">key_length</span><span class="p">])</span>
|
|
<span class="n">select_buf</span> <span class="o">=</span> <span class="n">np</span><span class="o">.</span><span class="n">expand_dims</span><span class="p">(</span>
|
|
<span class="n">np</span><span class="o">.</span><span class="n">tril</span><span class="p">(</span>
|
|
<span class="n">np</span><span class="o">.</span><span class="n">ones</span><span class="p">((</span><span class="bp">self</span><span class="o">.</span><span class="n">max_position_embeddings</span><span class="p">,</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">max_position_embeddings</span><span class="p">)))</span><span class="o">.</span><span class="n">astype</span><span class="p">(</span><span class="nb">bool</span><span class="p">),</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">select_buf</span> <span class="o">=</span> <span class="n">np</span><span class="o">.</span><span class="n">logical_not</span><span class="p">(</span><span class="n">select_buf</span><span class="p">)</span>
|
|
<span class="n">mask_buf</span> <span class="o">=</span> <span class="n">np</span><span class="o">.</span><span class="n">zeros_like</span><span class="p">(</span><span class="n">select_buf</span><span class="p">,</span> <span class="n">np</span><span class="o">.</span><span class="n">float32</span><span class="p">)</span>
|
|
<span class="n">mask_buf</span><span class="p">[</span><span class="n">select_buf</span><span class="p">]</span> <span class="o">=</span> <span class="nb">float</span><span class="p">(</span><span class="s1">'-inf'</span><span class="p">)</span>
|
|
<span class="n">buffer</span> <span class="o">=</span> <span class="n">constant</span><span class="p">(</span><span class="n">mask_buf</span><span class="p">)</span>
|
|
<span class="n">generated_mask</span> <span class="o">=</span> <span class="nb">slice</span><span class="p">(</span><span class="n">buffer</span><span class="p">,</span> <span class="n">starts</span><span class="p">,</span> <span class="n">sizes</span><span class="p">)</span>
|
|
|
|
<span class="k">elif</span> <span class="bp">self</span><span class="o">.</span><span class="n">attention_mask_type</span> <span class="o">==</span> <span class="n">AttentionMaskType</span><span class="o">.</span><span class="n">bidirectional</span><span class="p">:</span>
|
|
<span class="n">query_length</span> <span class="o">=</span> <span class="n">shape</span><span class="p">(</span><span class="n">query</span><span class="p">,</span> <span class="mi">2</span><span class="p">)</span>
|
|
<span class="n">zero_buf</span> <span class="o">=</span> <span class="n">np</span><span class="o">.</span><span class="n">expand_dims</span><span class="p">(</span>
|
|
<span class="n">np</span><span class="o">.</span><span class="n">zeros</span><span class="p">((</span><span class="bp">self</span><span class="o">.</span><span class="n">max_position_embeddings</span><span class="p">,</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">max_position_embeddings</span><span class="p">),</span>
|
|
<span class="n">dtype</span><span class="o">=</span><span class="n">np</span><span class="o">.</span><span class="n">float32</span><span class="p">),</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">zero_buf</span><span class="p">[:,</span> <span class="p">:,</span> <span class="p">:</span><span class="o">-</span><span class="mi">1</span><span class="p">,</span> <span class="o">-</span><span class="mi">1</span><span class="p">]</span> <span class="o">=</span> <span class="mi">1</span>
|
|
<span class="n">zero_buf</span> <span class="o">*=</span> <span class="o">-</span><span class="mi">10000</span>
|
|
|
|
<span class="n">mask</span> <span class="o">=</span> <span class="n">constant</span><span class="p">(</span><span class="n">zero_buf</span><span class="p">)</span>
|
|
|
|
<span class="c1"># context phase, query_length</span>
|
|
<span class="n">mask_size</span> <span class="o">=</span> <span class="n">where</span><span class="p">(</span><span class="n">query_length</span> <span class="o">></span> <span class="mi">1</span><span class="p">,</span> <span class="n">query_length</span><span class="p">,</span> <span class="mi">1</span><span class="p">)</span>
|
|
<span class="n">mask_start</span> <span class="o">=</span> <span class="n">where</span><span class="p">(</span><span class="n">query_length</span> <span class="o">></span> <span class="mi">1</span><span class="p">,</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">max_position_embeddings</span> <span class="o">-</span> <span class="n">mask_size</span><span class="p">,</span> <span class="mi">1</span><span class="p">)</span>
|
|
<span class="n">start</span> <span class="o">=</span> <span class="n">concat</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="n">mask_start</span><span class="p">,</span> <span class="n">mask_start</span><span class="p">])</span>
|
|
<span class="n">size</span> <span class="o">=</span> <span class="n">concat</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="n">mask_size</span><span class="p">,</span> <span class="n">mask_size</span><span class="p">])</span>
|
|
<span class="n">generated_mask</span> <span class="o">=</span> <span class="nb">slice</span><span class="p">(</span><span class="n">mask</span><span class="p">,</span> <span class="n">start</span><span class="p">,</span> <span class="n">size</span><span class="p">)</span>
|
|
|
|
<span class="k">if</span> <span class="n">attention_mask</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="n">attention_mask</span> <span class="o">=</span> <span class="n">expand_mask</span><span class="p">(</span><span class="n">attention_mask</span><span class="p">,</span> <span class="n">shape</span><span class="p">(</span><span class="n">query</span><span class="p">,</span> <span class="mi">2</span><span class="p">))</span>
|
|
<span class="n">bias</span> <span class="o">=</span> <span class="n">attention_mask</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">position_embedding_type</span><span class="o">.</span><span class="n">is_alibi</span><span class="p">():</span>
|
|
<span class="n">alibi_biases</span> <span class="o">=</span> <span class="n">generate_alibi_biases</span><span class="p">(</span><span class="n">alibi_slopes</span><span class="p">,</span> <span class="n">key_length</span><span class="p">)</span>
|
|
<span class="n">bias</span> <span class="o">=</span> <span class="n">alibi_biases</span> <span class="k">if</span> <span class="n">bias</span> <span class="ow">is</span> <span class="kc">None</span> <span class="k">else</span> <span class="n">bias</span> <span class="o">+</span> <span class="n">alibi_biases</span>
|
|
|
|
<span class="n">key</span> <span class="o">=</span> <span class="n">key</span><span class="o">.</span><span class="n">permute</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="mi">3</span><span class="p">,</span> <span class="mi">2</span><span class="p">])</span>
|
|
<span class="k">with</span> <span class="n">precision</span><span class="p">(</span><span class="s1">'float32'</span><span class="p">):</span>
|
|
<span class="k">if</span> <span class="n">norm_before_bmm1</span><span class="p">:</span>
|
|
<span class="c1"># Apply norm on query earlier to prevent matmul fp16 overflow.</span>
|
|
<span class="n">query</span> <span class="o">/=</span> <span class="bp">self</span><span class="o">.</span><span class="n">norm_factor</span>
|
|
<span class="n">attention_scores</span> <span class="o">=</span> <span class="n">matmul</span><span class="p">(</span><span class="n">cast</span><span class="p">(</span><span class="n">query</span><span class="p">,</span> <span class="s1">'float32'</span><span class="p">),</span>
|
|
<span class="n">cast</span><span class="p">(</span><span class="n">key</span><span class="p">,</span> <span class="s1">'float32'</span><span class="p">))</span>
|
|
<span class="k">if</span> <span class="ow">not</span> <span class="n">norm_before_bmm1</span><span class="p">:</span>
|
|
<span class="n">attention_scores</span> <span class="o">=</span> <span class="n">attention_scores</span> <span class="o">/</span> <span class="bp">self</span><span class="o">.</span><span class="n">norm_factor</span>
|
|
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">attention_mask_type</span> <span class="ow">in</span> <span class="p">[</span>
|
|
<span class="n">AttentionMaskType</span><span class="o">.</span><span class="n">causal</span><span class="p">,</span>
|
|
<span class="n">AttentionMaskType</span><span class="o">.</span><span class="n">bidirectional</span>
|
|
<span class="p">]:</span>
|
|
<span class="n">bias</span> <span class="o">=</span> <span class="n">generated_mask</span> <span class="k">if</span> <span class="n">bias</span> <span class="ow">is</span> <span class="kc">None</span> <span class="k">else</span> <span class="n">bias</span> <span class="o">+</span> <span class="n">generated_mask</span>
|
|
|
|
<span class="k">if</span> <span class="n">bias</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span> <span class="ow">and</span> <span class="ow">not</span> <span class="bp">self</span><span class="o">.</span><span class="n">cross_attention</span><span class="p">:</span>
|
|
<span class="n">attention_scores</span> <span class="o">=</span> <span class="n">attention_scores</span> <span class="o">+</span> <span class="n">bias</span>
|
|
|
|
<span class="n">attention_probs</span> <span class="o">=</span> <span class="n">softmax</span><span class="p">(</span><span class="n">attention_scores</span><span class="p">,</span> <span class="n">dim</span><span class="o">=-</span><span class="mi">1</span><span class="p">)</span>
|
|
|
|
<span class="k">if</span> <span class="n">default_net</span><span class="p">()</span><span class="o">.</span><span class="n">strongly_typed</span> <span class="ow">and</span> <span class="p">(</span><span class="n">attention_probs</span><span class="o">.</span><span class="n">dtype</span> <span class="o">!=</span>
|
|
<span class="n">value</span><span class="o">.</span><span class="n">dtype</span><span class="p">):</span>
|
|
<span class="n">attention_probs</span> <span class="o">=</span> <span class="n">cast</span><span class="p">(</span><span class="n">attention_probs</span><span class="p">,</span> <span class="n">value</span><span class="o">.</span><span class="n">dtype</span><span class="p">)</span>
|
|
|
|
<span class="n">context</span> <span class="o">=</span> <span class="n">matmul</span><span class="p">(</span><span class="n">attention_probs</span><span class="p">,</span> <span class="n">value</span><span class="p">)</span><span class="o">.</span><span class="n">permute</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="mi">1</span><span class="p">,</span> <span class="mi">3</span><span class="p">])</span>
|
|
<span class="n">context</span> <span class="o">=</span> <span class="n">context</span><span class="o">.</span><span class="n">view</span><span class="p">(</span>
|
|
<span class="n">concat</span><span class="p">([</span><span class="n">shape</span><span class="p">(</span><span class="n">context</span><span class="p">,</span> <span class="mi">0</span><span class="p">),</span>
|
|
<span class="n">shape</span><span class="p">(</span><span class="n">context</span><span class="p">,</span> <span class="mi">1</span><span class="p">),</span> <span class="bp">self</span><span class="o">.</span><span class="n">hidden_size</span><span class="p">]))</span>
|
|
|
|
<span class="n">context</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">dense</span><span class="p">(</span><span class="n">context</span><span class="p">,</span> <span class="n">workspace</span><span class="p">)</span>
|
|
|
|
<span class="k">if</span> <span class="n">use_cache</span><span class="p">:</span>
|
|
<span class="k">return</span> <span class="p">(</span><span class="n">context</span><span class="p">,</span> <span class="n">past_key_value</span><span class="p">)</span>
|
|
<span class="k">else</span><span class="p">:</span>
|
|
<span class="k">return</span> <span class="n">context</span></div>
|
|
</div>
|
|
|
|
|
|
|
|
<div class="viewcode-block" id="BertAttention">
|
|
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.layers.html#tensorrt_llm.layers.attention.BertAttention">[docs]</a>
|
|
<span class="k">class</span> <span class="nc">BertAttention</span><span class="p">(</span><span class="n">Module</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="n">hidden_size</span><span class="p">,</span>
|
|
<span class="n">num_attention_heads</span><span class="p">,</span>
|
|
<span class="n">max_position_embeddings</span><span class="o">=</span><span class="mi">1024</span><span class="p">,</span>
|
|
<span class="n">num_layers</span><span class="o">=</span><span class="mi">1</span><span class="p">,</span>
|
|
<span class="n">attention_head_size</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">num_kv_heads</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">q_scaling</span><span class="o">=</span><span class="mf">1.0</span><span class="p">,</span>
|
|
<span class="n">apply_query_key_layer_scaling</span><span class="o">=</span><span class="kc">False</span><span class="p">,</span>
|
|
<span class="n">bias</span><span class="o">=</span><span class="kc">True</span><span class="p">,</span>
|
|
<span class="n">dtype</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">tp_group</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">tp_size</span><span class="o">=</span><span class="mi">1</span><span class="p">,</span>
|
|
<span class="n">tp_rank</span><span class="o">=</span><span class="mi">0</span><span class="p">,</span>
|
|
<span class="n">relative_attention</span><span class="o">=</span><span class="kc">False</span><span class="p">,</span>
|
|
<span class="n">max_distance</span><span class="o">=</span><span class="mi">0</span><span class="p">,</span>
|
|
<span class="n">num_buckets</span><span class="o">=</span><span class="mi">0</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="bp">self</span><span class="o">.</span><span class="n">attention_head_size</span> <span class="o">=</span> <span class="n">hidden_size</span> <span class="o">//</span> <span class="n">num_attention_heads</span> <span class="k">if</span> <span class="n">attention_head_size</span> <span class="ow">is</span> <span class="kc">None</span> <span class="k">else</span> <span class="n">attention_head_size</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">num_attention_heads</span> <span class="o">=</span> <span class="n">num_attention_heads</span> <span class="o">//</span> <span class="n">tp_size</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">num_attention_kv_heads</span> <span class="o">=</span> <span class="p">(</span>
|
|
<span class="n">num_kv_heads</span> <span class="o">+</span> <span class="n">tp_size</span> <span class="o">-</span> <span class="mi">1</span>
|
|
<span class="p">)</span> <span class="o">//</span> <span class="n">tp_size</span> <span class="k">if</span> <span class="n">num_kv_heads</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span> <span class="k">else</span> <span class="bp">self</span><span class="o">.</span><span class="n">num_attention_heads</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">hidden_size</span> <span class="o">=</span> <span class="n">hidden_size</span> <span class="o">//</span> <span class="n">tp_size</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">max_position_embeddings</span> <span class="o">=</span> <span class="n">max_position_embeddings</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">norm_factor</span> <span class="o">=</span> <span class="n">math</span><span class="o">.</span><span class="n">sqrt</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">attention_head_size</span><span class="p">)</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">tp_size</span> <span class="o">=</span> <span class="n">tp_size</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">tp_rank</span> <span class="o">=</span> <span class="n">tp_rank</span>
|
|
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">num_layers</span> <span class="o">=</span> <span class="n">num_layers</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">apply_query_key_layer_scaling</span> <span class="o">=</span> <span class="n">apply_query_key_layer_scaling</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">norm_factor</span> <span class="o">=</span> <span class="n">math</span><span class="o">.</span><span class="n">sqrt</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">attention_head_size</span><span class="p">)</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">q_scaling</span> <span class="o">=</span> <span class="n">q_scaling</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">apply_query_key_layer_scaling</span><span class="p">:</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">norm_factor</span> <span class="o">*=</span> <span class="bp">self</span><span class="o">.</span><span class="n">num_layers</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">q_scaling</span> <span class="o">*=</span> <span class="bp">self</span><span class="o">.</span><span class="n">num_layers</span>
|
|
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">dtype</span> <span class="o">=</span> <span class="n">dtype</span>
|
|
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">relative_attention</span> <span class="o">=</span> <span class="n">relative_attention</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">max_distance</span> <span class="o">=</span> <span class="n">max_distance</span>
|
|
|
|
<span class="c1"># out dim is not necessarily hidden_size + kv specific size (in MQA/GQA), but num_heads * heads_size</span>
|
|
<span class="c1"># example: d_model != num_heads * head_size in Flan-T5</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">qkv</span> <span class="o">=</span> <span class="n">ColumnLinear</span><span class="p">(</span>
|
|
<span class="n">hidden_size</span><span class="p">,</span>
|
|
<span class="n">tp_size</span> <span class="o">*</span> <span class="bp">self</span><span class="o">.</span><span class="n">num_attention_heads</span> <span class="o">*</span> <span class="bp">self</span><span class="o">.</span><span class="n">attention_head_size</span> <span class="o">+</span>
|
|
<span class="p">(</span><span class="mi">2</span> <span class="o">*</span> <span class="n">tp_size</span> <span class="o">*</span> <span class="bp">self</span><span class="o">.</span><span class="n">num_attention_kv_heads</span> <span class="o">*</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">attention_head_size</span><span class="p">),</span>
|
|
<span class="n">bias</span><span class="o">=</span><span class="n">bias</span><span class="p">,</span>
|
|
<span class="n">dtype</span><span class="o">=</span><span class="n">dtype</span><span class="p">,</span>
|
|
<span class="n">tp_group</span><span class="o">=</span><span class="n">tp_group</span><span class="p">,</span>
|
|
<span class="n">tp_size</span><span class="o">=</span><span class="n">tp_size</span><span class="p">,</span>
|
|
<span class="n">gather_output</span><span class="o">=</span><span class="kc">False</span><span class="p">)</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">dense</span> <span class="o">=</span> <span class="n">RowLinear</span><span class="p">(</span><span class="n">tp_size</span> <span class="o">*</span> <span class="bp">self</span><span class="o">.</span><span class="n">num_attention_heads</span> <span class="o">*</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">attention_head_size</span><span class="p">,</span>
|
|
<span class="n">hidden_size</span><span class="p">,</span>
|
|
<span class="n">bias</span><span class="o">=</span><span class="n">bias</span><span class="p">,</span>
|
|
<span class="n">dtype</span><span class="o">=</span><span class="n">dtype</span><span class="p">,</span>
|
|
<span class="n">tp_group</span><span class="o">=</span><span class="n">tp_group</span><span class="p">,</span>
|
|
<span class="n">tp_size</span><span class="o">=</span><span class="n">tp_size</span><span class="p">)</span>
|
|
|
|
<span class="c1"># per-layer relative attention table</span>
|
|
<span class="k">if</span> <span class="n">relative_attention</span><span class="p">:</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">rel_attn_table</span> <span class="o">=</span> <span class="n">Parameter</span><span class="p">(</span><span class="n">shape</span><span class="o">=</span><span class="p">(</span><span class="n">num_attention_heads</span> <span class="o">//</span>
|
|
<span class="n">tp_size</span><span class="p">,</span> <span class="n">num_buckets</span><span class="p">),</span>
|
|
<span class="n">dtype</span><span class="o">=</span><span class="n">dtype</span><span class="p">)</span>
|
|
|
|
<div class="viewcode-block" id="BertAttention.forward">
|
|
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.layers.html#tensorrt_llm.layers.attention.BertAttention.forward">[docs]</a>
|
|
<span class="k">def</span> <span class="nf">forward</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span>
|
|
<span class="n">hidden_states</span><span class="p">:</span> <span class="n">Tensor</span><span class="p">,</span>
|
|
<span class="n">attention_mask</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">input_lengths</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">workspace</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">max_input_length</span><span class="o">=</span><span class="kc">None</span><span class="p">):</span>
|
|
<span class="k">assert</span> <span class="nb">isinstance</span><span class="p">(</span><span class="n">hidden_states</span><span class="p">,</span> <span class="n">Tensor</span><span class="p">)</span>
|
|
|
|
<span class="n">qkv</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">qkv</span><span class="p">(</span><span class="n">hidden_states</span><span class="p">)</span>
|
|
|
|
<span class="k">if</span> <span class="n">default_net</span><span class="p">()</span><span class="o">.</span><span class="n">plugin_config</span><span class="o">.</span><span class="n">bert_attention_plugin</span><span class="p">:</span>
|
|
<span class="c1"># TRT plugin mode</span>
|
|
<span class="k">assert</span> <span class="n">input_lengths</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span>
|
|
<span class="n">context</span> <span class="o">=</span> <span class="n">bert_attention</span><span class="p">(</span>
|
|
<span class="n">qkv</span><span class="p">,</span>
|
|
<span class="n">input_lengths</span><span class="p">,</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">num_attention_heads</span><span class="p">,</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">attention_head_size</span><span class="p">,</span>
|
|
<span class="n">q_scaling</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">q_scaling</span><span class="p">,</span>
|
|
<span class="n">relative_attention</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">relative_attention</span><span class="p">,</span>
|
|
<span class="n">max_distance</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">max_distance</span><span class="p">,</span>
|
|
<span class="n">relative_attention_bias</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">rel_attn_table</span><span class="o">.</span><span class="n">value</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">relative_attention</span> <span class="k">else</span> <span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">max_input_length</span><span class="o">=</span><span class="n">max_input_length</span><span class="p">)</span>
|
|
<span class="k">else</span><span class="p">:</span>
|
|
<span class="c1"># plain TRT mode</span>
|
|
<span class="k">def</span> <span class="nf">transpose_for_scores</span><span class="p">(</span><span class="n">x</span><span class="p">):</span>
|
|
<span class="n">new_x_shape</span> <span class="o">=</span> <span class="n">concat</span><span class="p">([</span>
|
|
<span class="n">shape</span><span class="p">(</span><span class="n">x</span><span class="p">,</span> <span class="mi">0</span><span class="p">),</span>
|
|
<span class="n">shape</span><span class="p">(</span><span class="n">x</span><span class="p">,</span> <span class="mi">1</span><span class="p">),</span> <span class="bp">self</span><span class="o">.</span><span class="n">num_attention_heads</span><span class="p">,</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">attention_head_size</span>
|
|
<span class="p">])</span>
|
|
<span class="k">return</span> <span class="n">x</span><span class="o">.</span><span class="n">view</span><span class="p">(</span><span class="n">new_x_shape</span><span class="p">)</span><span class="o">.</span><span class="n">permute</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="mi">1</span><span class="p">,</span> <span class="mi">3</span><span class="p">])</span>
|
|
|
|
<span class="n">query</span><span class="p">,</span> <span class="n">key</span><span class="p">,</span> <span class="n">value</span> <span class="o">=</span> <span class="n">split</span><span class="p">(</span><span class="n">qkv</span><span class="p">,</span> <span class="bp">self</span><span class="o">.</span><span class="n">hidden_size</span><span class="p">,</span> <span class="n">dim</span><span class="o">=</span><span class="mi">2</span><span class="p">)</span>
|
|
<span class="n">query</span> <span class="o">=</span> <span class="n">transpose_for_scores</span><span class="p">(</span><span class="n">query</span><span class="p">)</span>
|
|
<span class="n">key</span> <span class="o">=</span> <span class="n">transpose_for_scores</span><span class="p">(</span><span class="n">key</span><span class="p">)</span>
|
|
<span class="n">value</span> <span class="o">=</span> <span class="n">transpose_for_scores</span><span class="p">(</span><span class="n">value</span><span class="p">)</span>
|
|
|
|
<span class="n">key</span> <span class="o">=</span> <span class="n">key</span><span class="o">.</span><span class="n">permute</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="mi">3</span><span class="p">,</span> <span class="mi">2</span><span class="p">])</span>
|
|
<span class="n">attention_scores</span> <span class="o">=</span> <span class="n">matmul</span><span class="p">(</span><span class="n">query</span><span class="p">,</span> <span class="n">key</span><span class="p">)</span>
|
|
<span class="n">attention_scores</span> <span class="o">=</span> <span class="n">attention_scores</span> <span class="o">/</span> <span class="bp">self</span><span class="o">.</span><span class="n">norm_factor</span>
|
|
|
|
<span class="k">if</span> <span class="n">attention_mask</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="n">attention_scores</span> <span class="o">=</span> <span class="n">attention_scores</span> <span class="o">+</span> <span class="n">attention_mask</span>
|
|
|
|
<span class="n">attention_probs</span> <span class="o">=</span> <span class="n">softmax</span><span class="p">(</span><span class="n">attention_scores</span><span class="p">,</span> <span class="n">dim</span><span class="o">=-</span><span class="mi">1</span><span class="p">)</span>
|
|
|
|
<span class="n">context</span> <span class="o">=</span> <span class="n">matmul</span><span class="p">(</span><span class="n">attention_probs</span><span class="p">,</span> <span class="n">value</span><span class="p">)</span><span class="o">.</span><span class="n">permute</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="mi">1</span><span class="p">,</span> <span class="mi">3</span><span class="p">])</span>
|
|
<span class="n">context</span> <span class="o">=</span> <span class="n">context</span><span class="o">.</span><span class="n">view</span><span class="p">(</span>
|
|
<span class="n">concat</span><span class="p">([</span><span class="n">shape</span><span class="p">(</span><span class="n">context</span><span class="p">,</span> <span class="mi">0</span><span class="p">),</span>
|
|
<span class="n">shape</span><span class="p">(</span><span class="n">context</span><span class="p">,</span> <span class="mi">1</span><span class="p">),</span> <span class="bp">self</span><span class="o">.</span><span class="n">hidden_size</span><span class="p">]))</span>
|
|
|
|
<span class="n">context</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">dense</span><span class="p">(</span><span class="n">context</span><span class="p">,</span> <span class="n">workspace</span><span class="p">)</span>
|
|
|
|
<span class="k">return</span> <span class="n">context</span></div>
|
|
</div>
|
|
|
|
</pre></div>
|
|
|
|
</div>
|
|
</div>
|
|
<footer>
|
|
|
|
<hr/>
|
|
|
|
<div role="contentinfo">
|
|
<p>© 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> |