mirror of
https://github.com/NVIDIA/TensorRT-LLM.git
synced 2026-01-14 06:27:45 +08:00
2499 lines
366 KiB
HTML
2499 lines
366 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=e59714d7" />
|
|
<link rel="stylesheet" type="text/css" href="../../../_static/copybutton.css?v=76b2166b" />
|
|
|
|
|
|
<script src="../../../_static/jquery.js?v=5d32c60e"></script>
|
|
<script src="../../../_static/_sphinx_javascript_frameworks_compat.js?v=2cd50e6c"></script>
|
|
<script src="../../../_static/documentation_options.js?v=5929fcd5"></script>
|
|
<script src="../../../_static/doctools.js?v=9bcbadda"></script>
|
|
<script src="../../../_static/sphinx_highlight.js?v=dc90522c"></script>
|
|
<script src="../../../_static/clipboard.min.js?v=a7894cd8"></script>
|
|
<script src="../../../_static/copybutton.js?v=65e89d2a"></script>
|
|
<script src="../../../_static/js/theme.js"></script>
|
|
<link rel="index" title="Index" href="../../../genindex.html" />
|
|
<link rel="search" title="Search" href="../../../search.html" />
|
|
</head>
|
|
|
|
<body class="wy-body-for-nav">
|
|
<div class="wy-grid-for-nav">
|
|
<nav data-toggle="wy-nav-shift" class="wy-nav-side">
|
|
<div class="wy-side-scroll">
|
|
<div class="wy-side-nav-search" >
|
|
|
|
|
|
|
|
<a href="../../../index.html" class="icon icon-home">
|
|
tensorrt_llm
|
|
</a>
|
|
<div role="search">
|
|
<form id="rtd-search-form" class="wy-form" action="../../../search.html" method="get">
|
|
<input type="text" name="q" placeholder="Search docs" aria-label="Search docs" />
|
|
<input type="hidden" name="check_keywords" value="yes" />
|
|
<input type="hidden" name="area" value="default" />
|
|
</form>
|
|
</div>
|
|
</div><div class="wy-menu wy-menu-vertical" data-spy="affix" role="navigation" aria-label="Navigation menu">
|
|
<p class="caption" role="heading"><span class="caption-text">Getting Started</span></p>
|
|
<ul>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../overview.html">Overview</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../quick-start-guide.html">Quick Start Guide</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../key-features.html">Key Features</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../release-notes.html">Release Notes</a></li>
|
|
</ul>
|
|
<p class="caption" role="heading"><span class="caption-text">Installation</span></p>
|
|
<ul>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../installation/linux.html">Installing on Linux</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../installation/build-from-source-linux.html">Building from Source Code on Linux</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../installation/windows.html">Installing on Windows</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../installation/build-from-source-windows.html">Building from Source Code on Windows</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../installation/grace-hopper.html">Installing on Grace Hopper</a></li>
|
|
</ul>
|
|
<p class="caption" role="heading"><span class="caption-text">LLM API</span></p>
|
|
<ul>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../llm-api/index.html">API Introduction</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../llm-api/reference.html">API Reference</a></li>
|
|
</ul>
|
|
<p class="caption" role="heading"><span class="caption-text">LLM API Examples</span></p>
|
|
<ul>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../llm-api-examples/index.html">LLM Examples Introduction</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../llm-api-examples/customization.html">Common Customizations</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../llm-api-examples/llm_api_examples.html">Examples</a></li>
|
|
</ul>
|
|
<p class="caption" role="heading"><span class="caption-text">Model Definition API</span></p>
|
|
<ul>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../python-api/tensorrt_llm.layers.html">Layers</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../python-api/tensorrt_llm.functional.html">Functionals</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../python-api/tensorrt_llm.models.html">Models</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../python-api/tensorrt_llm.plugin.html">Plugin</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../python-api/tensorrt_llm.quantization.html">Quantization</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../python-api/tensorrt_llm.runtime.html">Runtime</a></li>
|
|
</ul>
|
|
<p class="caption" role="heading"><span class="caption-text">C++ API</span></p>
|
|
<ul>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../_cpp_gen/executor.html">Executor</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../_cpp_gen/runtime.html">Runtime</a></li>
|
|
</ul>
|
|
<p class="caption" role="heading"><span class="caption-text">Command-Line Reference</span></p>
|
|
<ul>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../commands/trtllm-build.html">trtllm-build</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../commands/trtllm-serve.html">trtllm-serve</a></li>
|
|
</ul>
|
|
<p class="caption" role="heading"><span class="caption-text">Architecture</span></p>
|
|
<ul>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../architecture/overview.html">TensorRT-LLM Architecture</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../architecture/core-concepts.html">Model Definition</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../architecture/core-concepts.html#compilation">Compilation</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../architecture/core-concepts.html#runtime">Runtime</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../architecture/core-concepts.html#multi-gpu-and-multi-node-support">Multi-GPU and Multi-Node Support</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../architecture/checkpoint.html">TensorRT-LLM Checkpoint</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../architecture/workflow.html">TensorRT-LLM Build Workflow</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../architecture/add-model.html">Adding a Model</a></li>
|
|
</ul>
|
|
<p class="caption" role="heading"><span class="caption-text">Advanced</span></p>
|
|
<ul>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../advanced/gpt-attention.html">Multi-Head, Multi-Query, and Group-Query Attention</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../advanced/gpt-runtime.html">C++ GPT Runtime</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../advanced/executor.html">Executor API</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../advanced/graph-rewriting.html">Graph Rewriting Module</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../advanced/inference-request.html">Inference Request</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../advanced/inference-request.html#responses">Responses</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../advanced/lora.html">Run gpt-2b + LoRA using GptManager / cpp runtime</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../advanced/expert-parallelism.html">Expert Parallelism in TensorRT-LLM</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../advanced/kv-cache-reuse.html">KV cache reuse</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../advanced/speculative-decoding.html">Speculative Sampling</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../advanced/disaggregated-service.html">Disaggregated-Service (experimental)</a></li>
|
|
</ul>
|
|
<p class="caption" role="heading"><span class="caption-text">Performance</span></p>
|
|
<ul>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../performance/perf-overview.html">Overview</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../performance/perf-benchmarking.html">Benchmarking</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../performance/perf-best-practices.html">Best Practices</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../performance/perf-analysis.html">Performance Analysis</a></li>
|
|
</ul>
|
|
<p class="caption" role="heading"><span class="caption-text">Reference</span></p>
|
|
<ul>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../reference/troubleshooting.html">Troubleshooting</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../reference/support-matrix.html">Support Matrix</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../reference/precision.html">Numerical Precision</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../reference/memory.html">Memory Usage of TensorRT-LLM</a></li>
|
|
</ul>
|
|
<p class="caption" role="heading"><span class="caption-text">Blogs</span></p>
|
|
<ul>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../blogs/H100vsA100.html">H100 has 4.6x A100 Performance in TensorRT-LLM, achieving 10,000 tok/s at 100ms to first token</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../blogs/H200launch.html">H200 achieves nearly 12,000 tokens/sec on Llama2-13B with TensorRT-LLM</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../blogs/Falcon180B-H200.html">Falcon-180B on a single H200 GPU with INT4 AWQ, and 6.7x faster Llama-70B over A100</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../blogs/quantization-in-TRT-LLM.html">Speed up inference with SOTA quantization techniques in TRT-LLM</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../blogs/XQA-kernel.html">New XQA-kernel provides 2.4x more Llama-70B throughput within the same latency budget</a></li>
|
|
</ul>
|
|
|
|
</div>
|
|
</div>
|
|
</nav>
|
|
|
|
<section data-toggle="wy-nav-shift" class="wy-nav-content-wrap"><nav class="wy-nav-top" aria-label="Mobile navigation menu" >
|
|
<i data-toggle="wy-nav-top" class="fa fa-bars"></i>
|
|
<a href="../../../index.html">tensorrt_llm</a>
|
|
</nav>
|
|
|
|
<div class="wy-nav-content">
|
|
<div class="rst-content">
|
|
<div role="navigation" aria-label="Page navigation">
|
|
<ul class="wy-breadcrumbs">
|
|
<li><a href="../../../index.html" class="icon icon-home" aria-label="Home"></a></li>
|
|
<li class="breadcrumb-item"><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) 2022-2024 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">import</span> <span class="nn">torch</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="p">(</span><span class="n">fp32_array</span><span class="p">,</span> <span class="n">int32_array</span><span class="p">,</span> <span class="n">is_same_dtype</span><span class="p">,</span> <span class="n">set_obj_attrs</span><span class="p">,</span>
|
|
<span class="n">trt_dtype_to_np</span><span class="p">,</span> <span class="n">trt_dtype_to_str</span><span class="p">)</span>
|
|
|
|
<span class="c1"># isort: off</span>
|
|
<span class="kn">from</span> <span class="nn">..functional</span> <span class="kn">import</span> <span class="p">(</span>
|
|
<span class="n">ACT2FN</span><span class="p">,</span> <span class="n">AllReduceParams</span><span class="p">,</span> <span class="n">AttentionMaskType</span><span class="p">,</span> <span class="n">Conditional</span><span class="p">,</span> <span class="n">LayerNormType</span><span class="p">,</span>
|
|
<span class="n">PositionEmbeddingType</span><span class="p">,</span> <span class="n">RopeEmbeddingUtils</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">allgather</span><span class="p">,</span> <span class="n">arange</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</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">generate_logn_scaling</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">minimum</span><span class="p">,</span> <span class="n">repeat_interleave</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">unsqueeze</span><span class="p">,</span> <span class="n">where</span><span class="p">)</span>
|
|
<span class="c1"># isort: on</span>
|
|
<span class="kn">from</span> <span class="nn">..mapping</span> <span class="kn">import</span> <span class="n">Mapping</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.functional</span> <span class="kn">import</span> <span class="n">dequantize</span><span class="p">,</span> <span class="n">quantize</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">LoraRuntimeParams</span>
|
|
<span class="kn">from</span> <span class="nn">.normalization</span> <span class="kn">import</span> <span class="n">GroupNorm</span><span class="p">,</span> <span class="n">LayerNorm</span><span class="p">,</span> <span class="n">RmsNorm</span>
|
|
|
|
<span class="kn">from</span> <span class="nn">..functional</span> <span class="kn">import</span> <span class="n">maximum</span> <span class="c1"># isort:skip</span>
|
|
|
|
<span class="n">layernorm_map</span> <span class="o">=</span> <span class="p">{</span>
|
|
<span class="n">LayerNormType</span><span class="o">.</span><span class="n">LayerNorm</span><span class="p">:</span> <span class="n">LayerNorm</span><span class="p">,</span>
|
|
<span class="n">LayerNormType</span><span class="o">.</span><span class="n">RmsNorm</span><span class="p">:</span> <span class="n">RmsNorm</span><span class="p">,</span>
|
|
<span class="n">LayerNormType</span><span class="o">.</span><span class="n">GroupNorm</span><span class="p">:</span> <span class="n">GroupNorm</span><span class="p">,</span>
|
|
<span class="p">}</span>
|
|
|
|
|
|
<div class="viewcode-block" id="make_causal_mask">
|
|
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.layers.html#tensorrt_llm.layers.attention.make_causal_mask">[docs]</a>
|
|
<span class="k">def</span> <span class="nf">make_causal_mask</span><span class="p">(</span><span class="n">bsz</span><span class="p">,</span> <span class="n">tgt_len</span><span class="p">,</span> <span class="n">past_key_values_length</span><span class="p">,</span> <span class="n">dtype</span><span class="p">):</span>
|
|
<span class="n">_range</span> <span class="o">=</span> <span class="n">arange</span><span class="p">(</span><span class="n">start</span><span class="o">=</span><span class="n">constant</span><span class="p">(</span><span class="n">int32_array</span><span class="p">(</span><span class="mi">0</span><span class="p">)),</span>
|
|
<span class="n">end</span><span class="o">=</span><span class="n">tgt_len</span><span class="p">,</span>
|
|
<span class="n">dtype</span><span class="o">=</span><span class="n">trt_dtype_to_str</span><span class="p">(</span><span class="n">dtype</span><span class="p">))</span>
|
|
<span class="n">mask</span> <span class="o">=</span> <span class="n">repeat_interleave</span><span class="p">(</span><span class="n">_range</span><span class="p">,</span> <span class="n">tgt_len</span><span class="p">,</span> <span class="mi">0</span><span class="p">)</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">tgt_len</span><span class="p">,</span>
|
|
<span class="n">tgt_len</span><span class="p">]))</span>
|
|
<span class="n">mask</span> <span class="o">=</span> <span class="n">where</span><span class="p">(</span><span class="n">mask</span> <span class="o"><</span> <span class="n">mask</span><span class="o">.</span><span class="n">transpose</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="p">),</span> <span class="mf">1.0</span><span class="p">,</span> <span class="mf">0.0</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">fp32_array</span><span class="p">(</span><span class="mi">0</span><span class="p">))</span>
|
|
<span class="n">zero</span> <span class="o">=</span> <span class="n">expand_dims</span><span class="p">(</span><span class="n">zero</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</span> <span class="o">=</span> <span class="n">expand</span><span class="p">(</span><span class="n">zero</span><span class="p">,</span> <span class="n">concat</span><span class="p">([</span><span class="n">tgt_len</span><span class="p">,</span> <span class="n">past_key_values_length</span><span class="p">]))</span>
|
|
<span class="n">mask</span> <span class="o">=</span> <span class="n">concat</span><span class="p">([</span><span class="n">zero</span><span class="p">,</span> <span class="n">mask</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">mask</span> <span class="o">*=</span> <span class="n">np</span><span class="o">.</span><span class="n">finfo</span><span class="p">(</span><span class="n">trt_dtype_to_np</span><span class="p">(</span><span class="n">dtype</span><span class="p">))</span><span class="o">.</span><span class="n">min</span><span class="o">.</span><span class="n">item</span><span class="p">()</span>
|
|
<span class="n">mask</span> <span class="o">=</span> <span class="n">mask</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="mi">1</span><span class="p">,</span> <span class="mi">1</span><span class="p">,</span> <span class="n">tgt_len</span><span class="p">,</span> <span class="n">tgt_len</span> <span class="o">+</span> <span class="n">past_key_values_length</span><span class="p">]))</span>
|
|
<span class="n">mask</span> <span class="o">=</span> <span class="n">expand</span><span class="p">(</span><span class="n">mask</span><span class="p">,</span>
|
|
<span class="n">concat</span><span class="p">([</span><span class="n">bsz</span><span class="p">,</span> <span class="mi">1</span><span class="p">,</span> <span class="n">tgt_len</span><span class="p">,</span> <span class="n">tgt_len</span> <span class="o">+</span> <span class="n">past_key_values_length</span><span class="p">]))</span>
|
|
<span class="k">return</span> <span class="n">mask</span></div>
|
|
|
|
|
|
|
|
<div class="viewcode-block" id="compute_relative_bias">
|
|
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.layers.html#tensorrt_llm.layers.attention.compute_relative_bias">[docs]</a>
|
|
<span class="k">def</span> <span class="nf">compute_relative_bias</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">num_buckets</span><span class="p">,</span>
|
|
<span class="n">max_distance</span><span class="p">,</span>
|
|
<span class="n">bidirectional</span><span class="p">,</span>
|
|
<span class="n">rel_attn_table</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_group</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">tp_rank</span><span class="o">=</span><span class="kc">None</span><span class="p">):</span>
|
|
|
|
<span class="k">def</span> <span class="nf">make_relative_position_bucket</span><span class="p">(</span><span class="n">relative_position</span><span class="p">,</span> <span class="n">bidirectional</span><span class="p">,</span>
|
|
<span class="n">num_buckets</span><span class="p">,</span> <span class="n">max_distance</span><span class="p">):</span>
|
|
<span class="n">relative_buckets</span> <span class="o">=</span> <span class="mi">0</span>
|
|
<span class="k">if</span> <span class="n">bidirectional</span><span class="p">:</span>
|
|
<span class="n">num_buckets</span> <span class="o">//=</span> <span class="mi">2</span>
|
|
<span class="n">relative_buckets</span> <span class="o">+=</span> <span class="n">where</span><span class="p">(</span><span class="n">relative_position</span> <span class="o">></span> <span class="mi">0</span><span class="p">,</span> <span class="n">num_buckets</span><span class="p">,</span> <span class="mi">0</span><span class="p">)</span>
|
|
<span class="n">relative_position</span> <span class="o">=</span> <span class="n">relative_position</span><span class="o">.</span><span class="n">abs</span><span class="p">()</span>
|
|
<span class="k">else</span><span class="p">:</span>
|
|
<span class="n">relative_position</span> <span class="o">=</span> <span class="mi">0</span> <span class="o">-</span> <span class="n">minimum</span><span class="p">(</span><span class="n">relative_position</span><span class="p">,</span> <span class="mi">0</span><span class="p">)</span>
|
|
|
|
<span class="n">max_exact</span> <span class="o">=</span> <span class="n">num_buckets</span> <span class="o">//</span> <span class="mi">2</span>
|
|
<span class="n">is_small</span> <span class="o">=</span> <span class="n">relative_position</span> <span class="o"><</span> <span class="n">max_exact</span>
|
|
|
|
<span class="n">max_exact_fp</span> <span class="o">=</span> <span class="n">constant</span><span class="p">(</span><span class="n">fp32_array</span><span class="p">(</span><span class="n">max_exact</span><span class="p">))</span>
|
|
<span class="n">tmp</span> <span class="o">=</span> <span class="n">cast</span><span class="p">(</span><span class="n">relative_position</span><span class="p">,</span> <span class="s2">"float32"</span><span class="p">)</span> <span class="o">/</span> <span class="n">max_exact_fp</span>
|
|
<span class="n">tmp</span> <span class="o">=</span> <span class="n">tmp</span><span class="o">.</span><span class="n">log</span><span class="p">()</span>
|
|
<span class="n">const1</span> <span class="o">=</span> <span class="n">math</span><span class="o">.</span><span class="n">log</span><span class="p">(</span><span class="n">max_distance</span> <span class="o">/</span> <span class="n">max_exact</span><span class="p">)</span>
|
|
<span class="n">const2</span> <span class="o">=</span> <span class="n">constant</span><span class="p">(</span><span class="n">fp32_array</span><span class="p">(</span><span class="n">num_buckets</span> <span class="o">-</span> <span class="n">max_exact</span><span class="p">))</span>
|
|
<span class="n">relative_position_if_large</span> <span class="o">=</span> <span class="n">tmp</span> <span class="o">/</span> <span class="n">const1</span> <span class="o">*</span> <span class="n">const2</span>
|
|
<span class="n">relative_position_if_large</span> <span class="o">=</span> <span class="n">cast</span><span class="p">(</span><span class="n">relative_position_if_large</span><span class="p">,</span> <span class="s2">"int32"</span><span class="p">)</span>
|
|
<span class="n">relative_position_if_large</span> <span class="o">=</span> <span class="n">max_exact</span> <span class="o">+</span> <span class="n">relative_position_if_large</span>
|
|
<span class="n">relative_position_if_large</span> <span class="o">=</span> <span class="n">minimum</span><span class="p">(</span><span class="n">relative_position_if_large</span><span class="p">,</span>
|
|
<span class="n">num_buckets</span> <span class="o">-</span> <span class="mi">1</span><span class="p">)</span>
|
|
|
|
<span class="n">relative_buckets</span> <span class="o">+=</span> <span class="n">where</span><span class="p">(</span><span class="n">is_small</span><span class="p">,</span> <span class="n">relative_position</span><span class="p">,</span>
|
|
<span class="n">relative_position_if_large</span><span class="p">)</span>
|
|
<span class="k">return</span> <span class="n">relative_buckets</span>
|
|
|
|
<span class="n">context_position</span> <span class="o">=</span> <span class="n">arange</span><span class="p">(</span><span class="n">start</span><span class="o">=</span><span class="n">constant</span><span class="p">(</span><span class="n">int32_array</span><span class="p">(</span><span class="mi">0</span><span class="p">)),</span>
|
|
<span class="n">end</span><span class="o">=</span><span class="n">query_length</span><span class="p">,</span>
|
|
<span class="n">dtype</span><span class="o">=</span><span class="n">trt_dtype_to_str</span><span class="p">(</span><span class="n">trt</span><span class="o">.</span><span class="n">int32</span><span class="p">))</span>
|
|
<span class="n">context_position</span> <span class="o">=</span> <span class="n">unsqueeze</span><span class="p">(</span><span class="n">context_position</span><span class="p">,</span> <span class="o">-</span><span class="mi">1</span><span class="p">)</span>
|
|
<span class="n">memory_position</span> <span class="o">=</span> <span class="n">arange</span><span class="p">(</span><span class="n">start</span><span class="o">=</span><span class="n">constant</span><span class="p">(</span><span class="n">int32_array</span><span class="p">(</span><span class="mi">0</span><span class="p">)),</span>
|
|
<span class="n">end</span><span class="o">=</span><span class="n">key_length</span><span class="p">,</span>
|
|
<span class="n">dtype</span><span class="o">=</span><span class="n">trt_dtype_to_str</span><span class="p">(</span><span class="n">trt</span><span class="o">.</span><span class="n">int32</span><span class="p">))</span>
|
|
<span class="n">memory_position</span> <span class="o">=</span> <span class="n">unsqueeze</span><span class="p">(</span><span class="n">memory_position</span><span class="p">,</span> <span class="mi">0</span><span class="p">)</span>
|
|
<span class="n">relative_position</span> <span class="o">=</span> <span class="n">memory_position</span> <span class="o">-</span> <span class="n">context_position</span>
|
|
<span class="n">relative_position_bucket</span> <span class="o">=</span> <span class="n">make_relative_position_bucket</span><span class="p">(</span>
|
|
<span class="n">relative_position</span><span class="p">,</span> <span class="c1"># shape (query_length, key_length)</span>
|
|
<span class="n">bidirectional</span><span class="p">,</span>
|
|
<span class="n">num_buckets</span><span class="p">,</span>
|
|
<span class="n">max_distance</span><span class="p">,</span>
|
|
<span class="p">)</span>
|
|
<span class="c1"># shape (query_length, key_length, num_heads)</span>
|
|
<span class="n">values</span> <span class="o">=</span> <span class="n">embedding</span><span class="p">(</span><span class="n">relative_position_bucket</span><span class="p">,</span>
|
|
<span class="n">rel_attn_table</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">tp_group</span><span class="o">=</span><span class="n">tp_group</span><span class="p">,</span>
|
|
<span class="n">tp_rank</span><span class="o">=</span><span class="n">tp_rank</span><span class="p">)</span>
|
|
<span class="c1"># shape (1, num_heads, query_length, key_length)</span>
|
|
<span class="n">values</span> <span class="o">=</span> <span class="n">unsqueeze</span><span class="p">(</span><span class="n">values</span><span class="o">.</span><span class="n">permute</span><span class="p">([</span><span class="mi">2</span><span class="p">,</span> <span class="mi">0</span><span class="p">,</span> <span class="mi">1</span><span class="p">]),</span> <span class="mi">0</span><span class="p">)</span>
|
|
<span class="k">return</span> <span class="n">values</span></div>
|
|
|
|
|
|
|
|
<div class="viewcode-block" id="AttentionMaskParams">
|
|
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.layers.html#tensorrt_llm.layers.attention.AttentionMaskParams">[docs]</a>
|
|
<span class="k">class</span> <span class="nc">AttentionMaskParams</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">self_attention_mask</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">self_attention_packed_mask</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">cross_attention_mask</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">cross_attention_packed_mask</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">self_attention_mask</span> <span class="o">=</span> <span class="n">self_attention_mask</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">self_attention_packed_mask</span> <span class="o">=</span> <span class="n">self_attention_packed_mask</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">cross_attention_mask</span> <span class="o">=</span> <span class="n">cross_attention_mask</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">cross_attention_packed_mask</span> <span class="o">=</span> <span class="n">cross_attention_packed_mask</span></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="n">host_runtime_perf_knobs</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_progress</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>
|
|
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">host_runtime_perf_knobs</span> <span class="o">=</span> <span class="n">host_runtime_perf_knobs</span>
|
|
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">host_context_progress</span> <span class="o">=</span> <span class="n">host_context_progress</span>
|
|
|
|
<span class="c1"># const parameters that will be reused by all layers.</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_inv_freq</span> <span class="o">=</span> <span class="kc">None</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">embed_positions_for_gpt_attention</span> <span class="o">=</span> <span class="kc">None</span>
|
|
<span class="c1"># long rope const parameters</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">long_rope_embed_positions</span> <span class="o">=</span> <span class="kc">None</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">long_rope_rotary_inv_freq</span> <span class="o">=</span> <span class="kc">None</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">long_rope_embed_positions_for_gpt_attention</span> <span class="o">=</span> <span class="kc">None</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">short_mscale</span> <span class="o">=</span> <span class="mf">1.0</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">long_mscale</span> <span class="o">=</span> <span class="mf">1.0</span>
|
|
|
|
<div class="viewcode-block" id="AttentionParams.fill_attention_const_params_for_rope">
|
|
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.layers.html#tensorrt_llm.layers.attention.AttentionParams.fill_attention_const_params_for_rope">[docs]</a>
|
|
<span class="k">def</span> <span class="nf">fill_attention_const_params_for_rope</span><span class="p">(</span>
|
|
<span class="bp">self</span><span class="p">,</span>
|
|
<span class="n">embed_positions</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">rotary_inv_freq</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">embed_positions_for_gpt_attention</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">embed_positions</span> <span class="o">=</span> <span class="n">embed_positions</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">rotary_inv_freq</span> <span class="o">=</span> <span class="n">rotary_inv_freq</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">embed_positions_for_gpt_attention</span> <span class="o">=</span> <span class="n">embed_positions_for_gpt_attention</span>
|
|
<span class="k">return</span> <span class="bp">self</span></div>
|
|
|
|
|
|
<div class="viewcode-block" id="AttentionParams.fill_attention_const_params_for_long_rope">
|
|
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.layers.html#tensorrt_llm.layers.attention.AttentionParams.fill_attention_const_params_for_long_rope">[docs]</a>
|
|
<span class="k">def</span> <span class="nf">fill_attention_const_params_for_long_rope</span><span class="p">(</span>
|
|
<span class="bp">self</span><span class="p">,</span> <span class="n">embed_positions</span><span class="p">,</span> <span class="n">long_rope_embed_positions</span><span class="p">,</span> <span class="n">rotary_inv_freq</span><span class="p">,</span>
|
|
<span class="n">long_rope_rotary_inv_freq</span><span class="p">,</span> <span class="n">embed_positions_for_gpt_attention</span><span class="p">,</span>
|
|
<span class="n">long_rope_embed_positions_for_gpt_attention</span><span class="p">,</span> <span class="n">short_mscale</span><span class="p">,</span>
|
|
<span class="n">long_mscale</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">embed_positions</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">long_rope_embed_positions</span> <span class="o">=</span> <span class="n">long_rope_embed_positions</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">rotary_inv_freq</span> <span class="o">=</span> <span class="n">rotary_inv_freq</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">long_rope_rotary_inv_freq</span> <span class="o">=</span> <span class="n">long_rope_rotary_inv_freq</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">embed_positions_for_gpt_attention</span> <span class="o">=</span> <span class="n">embed_positions_for_gpt_attention</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">long_rope_embed_positions_for_gpt_attention</span> <span class="o">=</span> <span class="n">long_rope_embed_positions_for_gpt_attention</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">short_mscale</span> <span class="o">=</span> <span class="n">short_mscale</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">long_mscale</span> <span class="o">=</span> <span class="n">long_mscale</span>
|
|
<span class="k">return</span> <span class="bp">self</span></div>
|
|
|
|
|
|
<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="n">use_kv_cache</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="n">use_kv_cache</span> <span class="ow">and</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="bp">self</span><span class="o">.</span><span class="n">host_runtime_perf_knobs</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_context_progress</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="SpecDecodingParams">
|
|
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.layers.html#tensorrt_llm.layers.attention.SpecDecodingParams">[docs]</a>
|
|
<span class="k">class</span> <span class="nc">SpecDecodingParams</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">spec_decoding_is_generation_length_variable</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">spec_decoding_max_generation_length</span><span class="p">:</span> <span class="nb">int</span> <span class="o">=</span> <span class="mi">1</span><span class="p">,</span>
|
|
<span class="n">spec_decoding_generation_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">spec_decoding_position_offsets</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">spec_decoding_packed_mask</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">spec_decoding_is_generation_length_variable</span> <span class="o">=</span> <span class="n">spec_decoding_is_generation_length_variable</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">spec_decoding_max_generation_length</span> <span class="o">=</span> <span class="n">spec_decoding_max_generation_length</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">spec_decoding_generation_lengths</span> <span class="o">=</span> <span class="n">spec_decoding_generation_lengths</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">spec_decoding_position_offsets</span> <span class="o">=</span> <span class="n">spec_decoding_position_offsets</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">spec_decoding_packed_mask</span> <span class="o">=</span> <span class="n">spec_decoding_packed_mask</span></div>
|
|
|
|
|
|
|
|
<div class="viewcode-block" id="MropeParams">
|
|
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.layers.html#tensorrt_llm.layers.attention.MropeParams">[docs]</a>
|
|
<span class="k">class</span> <span class="nc">MropeParams</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">mrope_rotary_sin_cos</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">mrope_position_deltas</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="p">):</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">mrope_rotary_sin_cos</span> <span class="o">=</span> <span class="n">mrope_rotary_sin_cos</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">mrope_position_deltas</span> <span class="o">=</span> <span class="n">mrope_position_deltas</span></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_attention_window_sizes</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_sink_token_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">kv_cache_block_offsets</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_kv_cache_block_offsets</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_kv_cache_pool_pointers</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_kv_cache_pool_mapping</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">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="n">cross_kv_cache_block_offsets</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_cross_kv_cache_block_offsets</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_cross_kv_cache_pool_pointers</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_cross_kv_cache_pool_mapping</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_attention_window_sizes</span> <span class="o">=</span> <span class="n">host_max_attention_window_sizes</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">host_sink_token_length</span> <span class="o">=</span> <span class="n">host_sink_token_length</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">kv_cache_block_offsets</span> <span class="o">=</span> <span class="n">kv_cache_block_offsets</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">host_kv_cache_block_offsets</span> <span class="o">=</span> <span class="n">host_kv_cache_block_offsets</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">host_kv_cache_pool_pointers</span> <span class="o">=</span> <span class="n">host_kv_cache_pool_pointers</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">host_kv_cache_pool_mapping</span> <span class="o">=</span> <span class="n">host_kv_cache_pool_mapping</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">cross_kv_cache_block_offsets</span> <span class="o">=</span> <span class="n">cross_kv_cache_block_offsets</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">host_cross_kv_cache_block_offsets</span> <span class="o">=</span> <span class="n">host_cross_kv_cache_block_offsets</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">host_cross_kv_cache_pool_pointers</span> <span class="o">=</span> <span class="n">host_cross_kv_cache_pool_pointers</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">host_cross_kv_cache_pool_mapping</span> <span class="o">=</span> <span class="n">host_cross_kv_cache_pool_mapping</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.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></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_attention_window_sizes</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_sink_token_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">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="BlockSparseAttnParams">
|
|
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.layers.html#tensorrt_llm.layers.attention.BlockSparseAttnParams">[docs]</a>
|
|
<span class="k">class</span> <span class="nc">BlockSparseAttnParams</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">block_size</span><span class="p">:</span> <span class="nb">int</span> <span class="o">=</span> <span class="mi">64</span><span class="p">,</span>
|
|
<span class="n">homo_head_pattern</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_local_blocks</span><span class="p">:</span> <span class="nb">int</span> <span class="o">=</span> <span class="mi">16</span><span class="p">,</span>
|
|
<span class="n">vertical_stride</span><span class="p">:</span> <span class="nb">int</span> <span class="o">=</span> <span class="mi">8</span><span class="p">):</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">block_size</span> <span class="o">=</span> <span class="n">block_size</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">homo_head_pattern</span> <span class="o">=</span> <span class="n">homo_head_pattern</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">num_local_blocks</span> <span class="o">=</span> <span class="n">num_local_blocks</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">vertical_stride</span> <span class="o">=</span> <span class="n">vertical_stride</span></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="o">*</span><span class="p">,</span>
|
|
<span class="n">local_layer_idx</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">qk_layernorm</span><span class="o">=</span><span class="kc">False</span><span class="p">,</span>
|
|
<span class="n">layernorm_type</span><span class="o">=</span><span class="n">LayerNormType</span><span class="o">.</span><span class="n">LayerNorm</span><span class="p">,</span>
|
|
<span class="n">layernorm_share</span><span class="o">=</span><span class="kc">True</span><span class="p">,</span>
|
|
<span class="n">inner_layernorm</span><span class="o">=</span><span class="kc">False</span><span class="p">,</span>
|
|
<span class="n">eps</span><span class="o">=</span><span class="mf">1e-05</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">rotary_embedding_percentage</span><span class="o">=</span><span class="mf">1.0</span><span class="p">,</span>
|
|
<span class="n">rope_scaling_short_factors</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">rope_scaling_long_factors</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">rope_scaling_short_mscale</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">rope_scaling_long_mscale</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">original_max_position_embeddings</span><span class="o">=</span><span class="mi">1024</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">dense_bias</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">clip_qkv</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">alibi_bias_max</span><span class="o">=</span><span class="mi">8</span><span class="p">,</span>
|
|
<span class="n">skip_cross_kv</span><span class="o">=</span><span class="kc">False</span><span class="p">,</span>
|
|
<span class="n">max_attn_value</span><span class="o">=</span><span class="mf">0.0</span><span class="p">,</span>
|
|
<span class="n">block_sparse_params</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">use_implicit_relative_attention</span><span class="o">=</span><span class="kc">False</span><span class="p">,</span>
|
|
<span class="n">reorder</span><span class="o">=</span><span class="kc">False</span><span class="p">,</span>
|
|
<span class="n">layer_idx_in_cache_pool</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">enable_qkv</span><span class="o">=</span><span class="kc">True</span><span class="p">,</span>
|
|
<span class="n">cp_group</span><span class="o">=</span><span class="p">[</span><span class="mi">0</span><span class="p">],</span>
|
|
<span class="n">cp_size</span><span class="o">=</span><span class="mi">1</span><span class="p">,</span>
|
|
<span class="n">cp_rank</span><span class="o">=</span><span class="mi">0</span><span class="p">,</span>
|
|
<span class="n">max_seqlen_for_logn_scaling</span><span class="o">=</span><span class="mi">8192</span><span class="p">,</span>
|
|
<span class="n">use_logn_scaling</span><span class="o">=</span><span class="kc">False</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">local_layer_idx</span> <span class="o">=</span> <span class="n">local_layer_idx</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="bp">self</span><span class="o">.</span><span class="n">num_kv_heads</span> <span class="o">=</span> <span class="n">num_kv_heads</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">layer_idx_in_cache_pool</span> <span class="o">=</span> <span class="n">layer_idx_in_cache_pool</span> <span class="k">if</span> <span class="n">layer_idx_in_cache_pool</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span> <span class="k">else</span> <span class="n">local_layer_idx</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="bp">self</span><span class="o">.</span><span class="n">attention_hidden_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_heads</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">original_max_position_embeddings</span> <span class="o">=</span> <span class="n">original_max_position_embeddings</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">bias</span> <span class="o">=</span> <span class="n">bias</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">tp_group</span> <span class="o">=</span> <span class="n">tp_group</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="bp">self</span><span class="o">.</span><span class="n">dense_bias</span> <span class="o">=</span> <span class="n">dense_bias</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="bp">self</span><span class="o">.</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">cp_group</span> <span class="o">=</span> <span class="n">cp_group</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">cp_size</span> <span class="o">=</span> <span class="n">cp_size</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">cp_rank</span> <span class="o">=</span> <span class="n">cp_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="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">alibi_bias_max</span> <span class="o">=</span> <span class="n">alibi_bias_max</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">num_buckets</span> <span class="o">=</span> <span class="n">num_buckets</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_scaling</span> <span class="o">=</span> <span class="n">rotary_embedding_scaling</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="bp">self</span><span class="o">.</span><span class="n">short_mscale</span> <span class="o">=</span> <span class="mf">1.0</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">long_mscale</span> <span class="o">=</span> <span class="mf">1.0</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">rotary_embedding_percentage</span> <span class="o">=</span> <span class="n">rotary_embedding_percentage</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">use_implicit_relative_attention</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">relative_attention</span> <span class="ow">and</span> <span class="n">use_implicit_relative_attention</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">max_seqlen_for_logn_scaling</span> <span class="o">=</span> <span class="n">max_seqlen_for_logn_scaling</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">use_logn_scaling</span> <span class="o">=</span> <span class="n">use_logn_scaling</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="n">rotary_scaling_type</span> <span class="o">=</span> <span class="n">rotary_embedding_scaling</span><span class="o">.</span><span class="n">get</span><span class="p">(</span>
|
|
<span class="s2">"type"</span><span class="p">,</span> <span class="n">rotary_embedding_scaling</span><span class="o">.</span><span class="n">get</span><span class="p">(</span><span class="s2">"rope_type"</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">from_string</span><span class="p">(</span>
|
|
<span class="n">rotary_scaling_type</span><span class="p">)</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="o">.</span><span class="n">get</span><span class="p">(</span>
|
|
<span class="s2">"factor"</span><span class="p">,</span> <span class="mf">1.0</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">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="k">elif</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_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">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">alibi_bias_max</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">alibi_bias_max</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">'alibi_slopes'</span><span class="p">,</span>
|
|
<span class="n">Parameter</span><span class="p">(</span><span class="n">alibi_slopes</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="n">is_buffer</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">use_logn_scaling</span><span class="p">:</span>
|
|
<span class="n">logn_scaling</span> <span class="o">=</span> <span class="n">generate_logn_scaling</span><span class="p">(</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">max_seqlen_for_logn_scaling</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">register_parameter</span><span class="p">(</span>
|
|
<span class="s1">'logn_scaling'</span><span class="p">,</span>
|
|
<span class="n">Parameter</span><span class="p">(</span><span class="n">logn_scaling</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="n">is_buffer</span><span class="o">=</span><span class="kc">True</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="bp">self</span><span class="o">.</span><span class="n">max_attn_value</span> <span class="o">=</span> <span class="n">max_attn_value</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">register_parameter</span><span class="p">(</span><span class="s1">'kv_cache_scaling_factor'</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">'attention_output_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">block_sparse_params</span> <span class="o">=</span> <span class="n">block_sparse_params</span> <span class="k">if</span> <span class="n">block_sparse_params</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span> <span class="k">else</span> <span class="n">BlockSparseAttnParams</span><span class="p">(</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="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/ByT5/Gemma</span>
|
|
<span class="k">if</span> <span class="n">enable_qkv</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">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="n">is_qkv</span><span class="o">=</span><span class="kc">True</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="bp">self</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="c1"># see optimize_model's add_lora for LoRA initialization</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">qkv_lora</span> <span class="o">=</span> <span class="kc">None</span>
|
|
|
|
<span class="c1"># per-layer relative attention table</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">use_implicit_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="c1"># qk layernorm</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">qk_layernorm</span> <span class="o">=</span> <span class="n">qk_layernorm</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">layernorm_type</span> <span class="o">=</span> <span class="n">layernorm_type</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">layernorm_share</span> <span class="o">=</span> <span class="n">layernorm_share</span>
|
|
<span class="n">ln_type</span> <span class="o">=</span> <span class="n">layernorm_map</span><span class="p">[</span><span class="n">layernorm_type</span><span class="p">]</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">qk_layernorm</span><span class="p">:</span>
|
|
<span class="c1"># layernorm_share indicates whether all the QK head in one layer shares the same norm parameters or not</span>
|
|
<span class="k">if</span> <span class="n">layernorm_share</span><span class="p">:</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">q_layernorm</span> <span class="o">=</span> <span class="n">ln_type</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">eps</span><span class="o">=</span><span class="n">eps</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">k_layernorm</span> <span class="o">=</span> <span class="n">ln_type</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">eps</span><span class="o">=</span><span class="n">eps</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="k">else</span><span class="p">:</span>
|
|
<span class="k">assert</span> <span class="n">ln_type</span> <span class="o">==</span> <span class="n">LayerNorm</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">q_layernorm</span> <span class="o">=</span> <span class="n">ln_type</span><span class="p">(</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">eps</span><span class="o">=</span><span class="n">eps</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">bias</span><span class="o">=</span><span class="kc">False</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">tp_dim</span><span class="o">=</span><span class="mi">0</span><span class="p">)</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">k_layernorm</span> <span class="o">=</span> <span class="n">ln_type</span><span class="p">(</span>
|
|
<span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">num_attention_kv_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">eps</span><span class="o">=</span><span class="n">eps</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">bias</span><span class="o">=</span><span class="kc">False</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">tp_dim</span><span class="o">=</span><span class="mi">0</span><span class="p">)</span>
|
|
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">inner_layernorm</span> <span class="o">=</span> <span class="n">ln_type</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">dtype</span><span class="o">=</span><span class="n">dtype</span><span class="p">,</span>
|
|
<span class="n">eps</span><span class="o">=</span><span class="n">eps</span><span class="p">)</span> <span class="k">if</span> <span class="n">inner_layernorm</span> <span class="k">else</span> <span class="kc">None</span>
|
|
<span class="k">if</span> <span class="n">clip_qkv</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">clip_qkv</span> <span class="o">=</span> <span class="n">fp32_array</span><span class="p">([</span><span class="n">clip_qkv</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">clip_qkv</span> <span class="o">=</span> <span class="kc">None</span>
|
|
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">skip_cross_kv</span> <span class="o">=</span> <span class="n">skip_cross_kv</span>
|
|
|
|
<div class="viewcode-block" id="Attention.create_attention_const_params">
|
|
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.layers.html#tensorrt_llm.layers.attention.Attention.create_attention_const_params">[docs]</a>
|
|
<span class="nd">@staticmethod</span>
|
|
<span class="k">def</span> <span class="nf">create_attention_const_params</span><span class="p">(</span><span class="n">model_cls</span><span class="p">,</span> <span class="n">config</span><span class="p">):</span>
|
|
<span class="c1"># get rotary parameters.</span>
|
|
<span class="n">hidden_size</span> <span class="o">=</span> <span class="n">config</span><span class="o">.</span><span class="n">hidden_size</span>
|
|
<span class="n">num_attention_heads</span> <span class="o">=</span> <span class="n">config</span><span class="o">.</span><span class="n">num_attention_heads</span>
|
|
<span class="n">attention_head_size</span> <span class="o">=</span> <span class="n">config</span><span class="o">.</span><span class="n">head_size</span>
|
|
<span class="n">max_position_embeddings</span> <span class="o">=</span> <span class="n">config</span><span class="o">.</span><span class="n">max_position_embeddings</span>
|
|
<span class="n">position_embedding_type</span> <span class="o">=</span> <span class="n">config</span><span class="o">.</span><span class="n">position_embedding_type</span>
|
|
<span class="n">rotary_embedding_base</span> <span class="o">=</span> <span class="nb">getattr</span><span class="p">(</span><span class="n">config</span><span class="p">,</span> <span class="s1">'rotary_base'</span><span class="p">,</span> <span class="mf">10000.0</span><span class="p">)</span>
|
|
<span class="n">rotary_embedding_scaling</span> <span class="o">=</span> <span class="nb">getattr</span><span class="p">(</span><span class="n">config</span><span class="p">,</span> <span class="s1">'rotary_scaling'</span><span class="p">,</span> <span class="kc">None</span><span class="p">)</span>
|
|
<span class="n">rotary_embedding_percentage</span> <span class="o">=</span> <span class="nb">getattr</span><span class="p">(</span><span class="n">config</span><span class="p">,</span> <span class="s1">'rotary_pct'</span><span class="p">,</span> <span class="mf">1.0</span><span class="p">)</span>
|
|
<span class="c1"># only rope need the const parameters.</span>
|
|
<span class="k">if</span> <span class="ow">not</span> <span class="n">position_embedding_type</span><span class="o">.</span><span class="n">is_rope</span><span class="p">():</span>
|
|
<span class="k">return</span>
|
|
<span class="c1"># attention head size</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="c1"># rotary embedding dim.</span>
|
|
<span class="n">rotary_embedding_dim</span> <span class="o">=</span> <span class="nb">getattr</span><span class="p">(</span>
|
|
<span class="n">config</span><span class="p">,</span> <span class="s1">'rotary_dim'</span><span class="p">,</span>
|
|
<span class="nb">int</span><span class="p">(</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="c1"># rotary scaling.</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="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="n">rotary_scaling_type</span> <span class="o">=</span> <span class="n">rotary_embedding_scaling</span><span class="o">.</span><span class="n">get</span><span class="p">(</span>
|
|
<span class="s2">"type"</span><span class="p">,</span> <span class="n">rotary_embedding_scaling</span><span class="o">.</span><span class="n">get</span><span class="p">(</span><span class="s2">"rope_type"</span><span class="p">))</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">from_string</span><span class="p">(</span>
|
|
<span class="n">rotary_scaling_type</span><span class="p">)</span>
|
|
<span class="n">rotary_embedding_scale</span> <span class="o">=</span> <span class="n">rotary_embedding_scaling</span><span class="o">.</span><span class="n">get</span><span class="p">(</span><span class="s2">"factor"</span><span class="p">,</span> <span class="mf">1.0</span><span class="p">)</span>
|
|
|
|
<span class="k">if</span> <span class="n">position_embedding_type</span> <span class="o">==</span> <span class="n">PositionEmbeddingType</span><span class="o">.</span><span class="n">long_rope</span><span class="p">:</span>
|
|
<span class="n">rope_scaling_short_factors</span><span class="p">,</span> <span class="n">rope_scaling_long_factors</span> <span class="o">=</span> <span class="kc">None</span><span class="p">,</span> <span class="kc">None</span>
|
|
<span class="n">rope_scaling_short_mscale</span><span class="p">,</span> <span class="n">rope_scaling_long_mscale</span> <span class="o">=</span> <span class="kc">None</span><span class="p">,</span> <span class="kc">None</span>
|
|
<span class="n">original_max_position_embeddings</span> <span class="o">=</span> <span class="n">max_position_embeddings</span>
|
|
|
|
<span class="k">if</span> <span class="nb">hasattr</span><span class="p">(</span><span class="n">config</span><span class="p">,</span> <span class="s2">"longrope_scaling_short_factors"</span><span class="p">):</span>
|
|
<span class="n">rope_scaling_short_factors</span> <span class="o">=</span> <span class="n">np</span><span class="o">.</span><span class="n">asarray</span><span class="p">(</span>
|
|
<span class="n">config</span><span class="o">.</span><span class="n">longrope_scaling_short_factors</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>
|
|
<span class="n">rope_scaling_long_factors</span> <span class="o">=</span> <span class="n">np</span><span class="o">.</span><span class="n">asarray</span><span class="p">(</span>
|
|
<span class="n">config</span><span class="o">.</span><span class="n">longrope_scaling_long_factors</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>
|
|
|
|
<span class="n">original_max_position_embeddings</span> <span class="o">=</span> <span class="n">config</span><span class="o">.</span><span class="n">original_max_position_embeddings</span>
|
|
|
|
<span class="k">if</span> <span class="n">config</span><span class="o">.</span><span class="n">architecture</span> <span class="o">==</span> <span class="s2">"Phi3SmallForCausalLM"</span> <span class="ow">or</span> <span class="n">config</span><span class="o">.</span><span class="n">architecture</span> <span class="o">==</span> <span class="s2">"PhiMoEForCausalLM"</span><span class="p">:</span>
|
|
<span class="n">rope_scaling_short_mscale</span> <span class="o">=</span> <span class="n">config</span><span class="o">.</span><span class="n">longrope_short_mscale</span>
|
|
<span class="n">rope_scaling_long_mscale</span> <span class="o">=</span> <span class="n">config</span><span class="o">.</span><span class="n">longrope_long_mscale</span>
|
|
|
|
<span class="n">embed_positions</span><span class="p">,</span> <span class="n">long_rope_embed_positions</span><span class="p">,</span> \
|
|
<span class="p">(</span><span class="n">rotary_inv_freq</span><span class="p">,</span> <span class="n">embed_positions_for_gpt_attention</span><span class="p">),</span> \
|
|
<span class="p">(</span><span class="n">long_rope_rotary_inv_freq</span><span class="p">,</span> <span class="n">long_rope_embed_positions_for_gpt_attention</span><span class="p">),</span> <span class="n">mscale</span> \
|
|
<span class="o">=</span> <span class="n">RopeEmbeddingUtils</span><span class="o">.</span><span class="n">create_sinusoidal_positions_long_rope</span><span class="p">(</span>
|
|
<span class="n">max_position_embeddings</span><span class="p">,</span>
|
|
<span class="n">original_max_position_embeddings</span><span class="p">,</span> <span class="n">rotary_embedding_dim</span><span class="p">,</span>
|
|
<span class="n">rotary_embedding_base</span><span class="p">,</span> <span class="n">rope_scaling_short_factors</span><span class="p">,</span>
|
|
<span class="n">rope_scaling_long_factors</span><span class="p">,</span> <span class="n">rope_scaling_short_mscale</span><span class="p">,</span> <span class="n">rope_scaling_long_mscale</span><span class="p">)</span>
|
|
|
|
<span class="k">if</span> <span class="n">rope_scaling_short_mscale</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">rope_scaling_long_mscale</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span>
|
|
<span class="n">short_mscale</span> <span class="o">=</span> <span class="n">rope_scaling_short_mscale</span>
|
|
<span class="n">long_mscale</span> <span class="o">=</span> <span class="n">rope_scaling_long_mscale</span>
|
|
<span class="k">else</span><span class="p">:</span>
|
|
<span class="n">short_mscale</span> <span class="o">=</span> <span class="n">long_mscale</span> <span class="o">=</span> <span class="n">mscale</span>
|
|
|
|
<span class="n">model_cls</span><span class="o">.</span><span class="n">register_parameter</span><span class="p">(</span>
|
|
<span class="s1">'embed_positions'</span><span class="p">,</span>
|
|
<span class="n">Parameter</span><span class="p">(</span><span class="n">embed_positions</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="n">is_buffer</span><span class="o">=</span><span class="kc">True</span><span class="p">))</span>
|
|
<span class="n">model_cls</span><span class="o">.</span><span class="n">register_parameter</span><span class="p">(</span>
|
|
<span class="s1">'long_rope_embed_positions'</span><span class="p">,</span>
|
|
<span class="n">Parameter</span><span class="p">(</span><span class="n">long_rope_embed_positions</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="n">is_buffer</span><span class="o">=</span><span class="kc">True</span><span class="p">))</span>
|
|
<span class="n">model_cls</span><span class="o">.</span><span class="n">register_parameter</span><span class="p">(</span>
|
|
<span class="s1">'rotary_inv_freq'</span><span class="p">,</span>
|
|
<span class="n">Parameter</span><span class="p">(</span><span class="n">rotary_inv_freq</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="n">is_buffer</span><span class="o">=</span><span class="kc">True</span><span class="p">))</span>
|
|
<span class="n">model_cls</span><span class="o">.</span><span class="n">register_parameter</span><span class="p">(</span>
|
|
<span class="s1">'long_rope_rotary_inv_freq'</span><span class="p">,</span>
|
|
<span class="n">Parameter</span><span class="p">(</span><span class="n">long_rope_rotary_inv_freq</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="n">is_buffer</span><span class="o">=</span><span class="kc">True</span><span class="p">))</span>
|
|
<span class="n">model_cls</span><span class="o">.</span><span class="n">register_parameter</span><span class="p">(</span>
|
|
<span class="s1">'embed_positions_for_gpt_attention'</span><span class="p">,</span>
|
|
<span class="n">Parameter</span><span class="p">(</span><span class="n">embed_positions_for_gpt_attention</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="n">is_buffer</span><span class="o">=</span><span class="kc">True</span><span class="p">))</span>
|
|
<span class="n">model_cls</span><span class="o">.</span><span class="n">register_parameter</span><span class="p">(</span>
|
|
<span class="s1">'long_rope_embed_positions_for_gpt_attention'</span><span class="p">,</span>
|
|
<span class="n">Parameter</span><span class="p">(</span><span class="n">long_rope_embed_positions_for_gpt_attention</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="n">is_buffer</span><span class="o">=</span><span class="kc">True</span><span class="p">))</span>
|
|
<span class="n">model_cls</span><span class="o">.</span><span class="n">short_mscale</span> <span class="o">=</span> <span class="n">short_mscale</span>
|
|
<span class="n">model_cls</span><span class="o">.</span><span class="n">long_mscale</span> <span class="o">=</span> <span class="n">long_mscale</span>
|
|
<span class="k">else</span><span class="p">:</span>
|
|
<span class="c1"># Rotary const weights.</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="n">max_position_embeddings</span><span class="p">,</span>
|
|
<span class="n">rotary_embedding_dim</span><span class="p">,</span>
|
|
<span class="p">)</span>
|
|
<span class="n">rotary_inv_freq</span><span class="p">,</span> <span class="n">embed_positions_for_gpt_attention</span> <span class="o">=</span> <span class="n">RopeEmbeddingUtils</span><span class="o">.</span><span class="n">create_sinusoidal_positions_for_attention_plugin</span><span class="p">(</span>
|
|
<span class="n">max_position_embeddings</span><span class="p">,</span> <span class="n">rotary_embedding_dim</span><span class="p">,</span>
|
|
<span class="n">rotary_embedding_base</span><span class="p">,</span> <span class="n">rotary_embedding_scale</span><span class="p">,</span>
|
|
<span class="n">rotary_embedding_scale_type</span><span class="p">,</span> <span class="n">rotary_embedding_scaling</span><span class="p">)</span>
|
|
<span class="n">model_cls</span><span class="o">.</span><span class="n">register_parameter</span><span class="p">(</span>
|
|
<span class="s1">'embed_positions'</span><span class="p">,</span>
|
|
<span class="n">Parameter</span><span class="p">(</span><span class="n">embed_positions</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="n">is_buffer</span><span class="o">=</span><span class="kc">True</span><span class="p">))</span>
|
|
<span class="n">model_cls</span><span class="o">.</span><span class="n">register_parameter</span><span class="p">(</span>
|
|
<span class="s1">'rotary_inv_freq'</span><span class="p">,</span>
|
|
<span class="n">Parameter</span><span class="p">(</span><span class="n">rotary_inv_freq</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="n">is_buffer</span><span class="o">=</span><span class="kc">True</span><span class="p">))</span>
|
|
<span class="n">model_cls</span><span class="o">.</span><span class="n">register_parameter</span><span class="p">(</span>
|
|
<span class="s1">'embed_positions_for_gpt_attention'</span><span class="p">,</span>
|
|
<span class="n">Parameter</span><span class="p">(</span><span class="n">embed_positions_for_gpt_attention</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="n">is_buffer</span><span class="o">=</span><span class="kc">True</span><span class="p">))</span></div>
|
|
|
|
|
|
<div class="viewcode-block" id="Attention.fill_attention_params">
|
|
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.layers.html#tensorrt_llm.layers.attention.Attention.fill_attention_params">[docs]</a>
|
|
<span class="nd">@staticmethod</span>
|
|
<span class="k">def</span> <span class="nf">fill_attention_params</span><span class="p">(</span><span class="n">model_cls</span><span class="p">,</span> <span class="n">attention_params</span><span class="p">):</span>
|
|
<span class="k">if</span> <span class="n">model_cls</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="k">if</span> <span class="n">attention_params</span> <span class="ow">is</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="n">attention_params</span> <span class="o">=</span> <span class="n">AttentionParams</span><span class="p">()</span>
|
|
<span class="k">if</span> <span class="n">model_cls</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">long_rope</span><span class="p">:</span>
|
|
<span class="k">return</span> <span class="n">attention_params</span><span class="o">.</span><span class="n">fill_attention_const_params_for_long_rope</span><span class="p">(</span>
|
|
<span class="n">model_cls</span><span class="o">.</span><span class="n">embed_positions</span><span class="o">.</span><span class="n">value</span><span class="p">,</span>
|
|
<span class="n">model_cls</span><span class="o">.</span><span class="n">long_rope_embed_positions</span><span class="o">.</span><span class="n">value</span><span class="p">,</span>
|
|
<span class="n">model_cls</span><span class="o">.</span><span class="n">rotary_inv_freq</span><span class="o">.</span><span class="n">value</span><span class="p">,</span>
|
|
<span class="n">model_cls</span><span class="o">.</span><span class="n">long_rope_rotary_inv_freq</span><span class="o">.</span><span class="n">value</span><span class="p">,</span>
|
|
<span class="n">model_cls</span><span class="o">.</span><span class="n">embed_positions_for_gpt_attention</span><span class="o">.</span><span class="n">value</span><span class="p">,</span>
|
|
<span class="n">model_cls</span><span class="o">.</span><span class="n">long_rope_embed_positions_for_gpt_attention</span><span class="o">.</span><span class="n">value</span><span class="p">,</span>
|
|
<span class="n">model_cls</span><span class="o">.</span><span class="n">short_mscale</span><span class="p">,</span> <span class="n">model_cls</span><span class="o">.</span><span class="n">long_mscale</span><span class="p">)</span>
|
|
<span class="k">else</span><span class="p">:</span>
|
|
<span class="k">return</span> <span class="n">attention_params</span><span class="o">.</span><span class="n">fill_attention_const_params_for_rope</span><span class="p">(</span>
|
|
<span class="n">model_cls</span><span class="o">.</span><span class="n">embed_positions</span><span class="o">.</span><span class="n">value</span><span class="p">,</span>
|
|
<span class="n">model_cls</span><span class="o">.</span><span class="n">rotary_inv_freq</span><span class="o">.</span><span class="n">value</span><span class="p">,</span>
|
|
<span class="n">model_cls</span><span class="o">.</span><span class="n">embed_positions_for_gpt_attention</span><span class="o">.</span><span class="n">value</span><span class="p">)</span>
|
|
<span class="c1"># Fill nothing.</span>
|
|
<span class="k">return</span> <span class="n">attention_params</span></div>
|
|
|
|
|
|
<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">attention_packed_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">spec_decoding_params</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">mrope_params</span><span class="o">=</span><span class="kc">None</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">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_layer_params</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">cross_kv_cache_gen</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">cross_kv_reuse</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">all_reduce_params</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="n">AllReduceParams</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">skip_attn</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span>
|
|
<span class="p">):</span>
|
|
<span class="n">attention_input</span> <span class="o">=</span> <span class="n">hidden_states</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">spec_decoding_params</span> <span class="o">=</span> <span class="n">SpecDecodingParams</span><span class="p">(</span>
|
|
<span class="p">)</span> <span class="k">if</span> <span class="n">spec_decoding_params</span> <span class="ow">is</span> <span class="kc">None</span> <span class="k">else</span> <span class="n">spec_decoding_params</span>
|
|
|
|
<span class="n">mrope_params</span> <span class="o">=</span> <span class="n">MropeParams</span><span class="p">()</span> <span class="k">if</span> <span class="n">mrope_params</span> <span class="ow">is</span> <span class="kc">None</span> <span class="k">else</span> <span class="n">mrope_params</span>
|
|
<span class="n">logn_scaling</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">use_logn_scaling</span><span class="p">:</span>
|
|
<span class="n">logn_scaling</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">logn_scaling</span><span class="o">.</span><span class="n">value</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">alibi_slopes</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">alibi_slopes</span><span class="o">.</span><span class="n">value</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">alibi_slopes</span> <span class="o">=</span> <span class="n">cast</span><span class="p">(</span><span class="n">alibi_slopes</span><span class="p">,</span> <span class="n">hidden_states</span><span class="o">.</span><span class="n">dtype</span><span class="p">)</span>
|
|
|
|
<span class="n">qkv_lora_params</span> <span class="o">=</span> <span class="kc">None</span>
|
|
<span class="k">if</span> <span class="n">lora_layer_params</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="k">if</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">qkv_lora_params</span> <span class="o">=</span> <span class="n">lora_layer_params</span><span class="o">.</span><span class="n">get_runtime_params</span><span class="p">(</span>
|
|
<span class="mi">0</span><span class="p">,</span> <span class="s2">"attn_qkv"</span><span class="p">)</span>
|
|
<span class="k">else</span><span class="p">:</span>
|
|
<span class="n">qkv_lora_params</span> <span class="o">=</span> <span class="n">lora_layer_params</span><span class="o">.</span><span class="n">get_runtime_params</span><span class="p">(</span>
|
|
<span class="mi">0</span><span class="p">,</span> <span class="s2">"cross_attn_qkv"</span><span class="p">)</span>
|
|
|
|
<span class="n">unfuse_qkv_gemm</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">qkv</span> <span class="ow">is</span> <span class="kc">None</span>
|
|
<span class="k">if</span> <span class="n">unfuse_qkv_gemm</span><span class="p">:</span>
|
|
<span class="n">qkv_gemm</span> <span class="o">=</span> <span class="p">[</span><span class="bp">self</span><span class="o">.</span><span class="n">q</span><span class="p">,</span> <span class="bp">self</span><span class="o">.</span><span class="n">k</span><span class="p">,</span> <span class="bp">self</span><span class="o">.</span><span class="n">v</span><span class="p">]</span>
|
|
<span class="n">qkv</span> <span class="o">=</span> <span class="p">[</span><span class="n">gemm</span><span class="p">(</span><span class="n">hidden_states</span><span class="p">)</span> <span class="k">for</span> <span class="n">gemm</span> <span class="ow">in</span> <span class="n">qkv_gemm</span><span class="p">]</span>
|
|
<span class="k">if</span> <span class="n">default_net</span><span class="p">(</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="ow">and</span> <span class="n">qkv_lora_params</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="n">lora</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">qkv</span><span class="o">.</span><span class="n">lora</span><span class="p">(</span><span class="n">hidden_states</span><span class="p">,</span> <span class="n">qkv_lora_params</span><span class="p">)</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">qkv_lora</span> <span class="o">=</span> <span class="n">split</span><span class="p">(</span><span class="n">lora</span><span class="p">,</span>
|
|
<span class="p">[</span><span class="bp">self</span><span class="o">.</span><span class="n">attention_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">1</span><span class="p">)</span>
|
|
<span class="n">qkv</span> <span class="o">=</span> <span class="p">[</span><span class="n">tensor</span> <span class="o">+</span> <span class="n">lora</span> <span class="k">for</span> <span class="n">tensor</span><span class="p">,</span> <span class="n">lora</span> <span class="ow">in</span> <span class="nb">zip</span><span class="p">(</span><span class="n">qkv</span><span class="p">,</span> <span class="n">qkv_lora</span><span class="p">)]</span>
|
|
<span class="k">else</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="n">qkv_lora_params</span><span class="p">)</span>
|
|
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">clip_qkv</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="n">qkv</span> <span class="o">=</span> <span class="n">clip</span><span class="p">(</span><span class="n">qkv</span><span class="p">,</span> <span class="o">-</span><span class="bp">self</span><span class="o">.</span><span class="n">clip_qkv</span><span class="p">,</span> <span class="bp">self</span><span class="o">.</span><span class="n">clip_qkv</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">remove_input_padding</span><span class="p">:</span>
|
|
<span class="k">if</span> <span class="n">unfuse_qkv_gemm</span><span class="p">:</span>
|
|
<span class="k">for</span> <span class="n">tensor</span> <span class="ow">in</span> <span class="n">qkv</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">2</span>
|
|
<span class="k">else</span><span class="p">:</span>
|
|
<span class="k">assert</span> <span class="n">qkv</span><span class="o">.</span><span class="n">ndim</span><span class="p">()</span> <span class="o">==</span> <span class="mi">2</span>
|
|
|
|
<span class="k">if</span> <span class="n">default_net</span><span class="p">(</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="ow">and</span> <span class="n">qkv_lora_params</span> <span class="ow">is</span> <span class="kc">None</span> <span class="ow">and</span> <span class="n">lora_layer_params</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="k">if</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">q_lora_params</span> <span class="o">=</span> <span class="n">lora_layer_params</span><span class="o">.</span><span class="n">get_runtime_params</span><span class="p">(</span>
|
|
<span class="mi">0</span><span class="p">,</span> <span class="s2">"attn_q"</span><span class="p">)</span>
|
|
<span class="n">k_lora_params</span> <span class="o">=</span> <span class="n">lora_layer_params</span><span class="o">.</span><span class="n">get_runtime_params</span><span class="p">(</span>
|
|
<span class="mi">0</span><span class="p">,</span> <span class="s2">"attn_k"</span><span class="p">)</span>
|
|
<span class="n">v_lora_params</span> <span class="o">=</span> <span class="n">lora_layer_params</span><span class="o">.</span><span class="n">get_runtime_params</span><span class="p">(</span>
|
|
<span class="mi">0</span><span class="p">,</span> <span class="s2">"attn_v"</span><span class="p">)</span>
|
|
<span class="k">else</span><span class="p">:</span>
|
|
<span class="n">q_lora_params</span> <span class="o">=</span> <span class="n">lora_layer_params</span><span class="o">.</span><span class="n">get_runtime_params</span><span class="p">(</span>
|
|
<span class="mi">0</span><span class="p">,</span> <span class="s2">"cross_attn_q"</span><span class="p">)</span>
|
|
<span class="n">k_lora_params</span> <span class="o">=</span> <span class="n">lora_layer_params</span><span class="o">.</span><span class="n">get_runtime_params</span><span class="p">(</span>
|
|
<span class="mi">0</span><span class="p">,</span> <span class="s2">"cross_attn_k"</span><span class="p">)</span>
|
|
<span class="n">v_lora_params</span> <span class="o">=</span> <span class="n">lora_layer_params</span><span class="o">.</span><span class="n">get_runtime_params</span><span class="p">(</span>
|
|
<span class="mi">0</span><span class="p">,</span> <span class="s2">"cross_attn_v"</span><span class="p">)</span>
|
|
|
|
<span class="k">assert</span> <span class="p">(</span><span class="n">q_lora_params</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span> <span class="ow">and</span> <span class="n">k_lora_params</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span> <span class="ow">and</span> <span class="n">v_lora_params</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">)</span> <span class="ow">or</span> \
|
|
<span class="p">(</span><span class="n">q_lora_params</span> <span class="ow">is</span> <span class="kc">None</span> <span class="ow">and</span> <span class="n">k_lora_params</span> <span class="ow">is</span> <span class="kc">None</span> <span class="ow">and</span> <span class="n">v_lora_params</span> <span class="ow">is</span> <span class="kc">None</span><span class="p">),</span> <span class="s2">"q_lora_params, k_lora_params and v_lora_params should be all enabled or all disabled at the same time."</span>
|
|
|
|
<span class="k">if</span> <span class="n">q_lora_params</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span> <span class="ow">and</span> <span class="n">k_lora_params</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span> <span class="ow">and</span> <span class="n">v_lora_params</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="n">qkv_lora_runtime_params</span> <span class="o">=</span> <span class="n">LoraRuntimeParams</span><span class="p">(</span>
|
|
<span class="n">lora_ranks</span><span class="o">=</span><span class="p">[</span>
|
|
<span class="n">q_lora_params</span><span class="o">.</span><span class="n">lora_ranks</span><span class="p">[</span><span class="mi">0</span><span class="p">],</span>
|
|
<span class="n">k_lora_params</span><span class="o">.</span><span class="n">lora_ranks</span><span class="p">[</span><span class="mi">0</span><span class="p">],</span>
|
|
<span class="n">v_lora_params</span><span class="o">.</span><span class="n">lora_ranks</span><span class="p">[</span><span class="mi">0</span><span class="p">],</span>
|
|
<span class="p">],</span>
|
|
<span class="n">lora_weights_pointers</span><span class="o">=</span><span class="p">[</span>
|
|
<span class="n">q_lora_params</span><span class="o">.</span><span class="n">lora_weights_pointers</span><span class="p">[</span><span class="mi">0</span><span class="p">],</span>
|
|
<span class="n">k_lora_params</span><span class="o">.</span><span class="n">lora_weights_pointers</span><span class="p">[</span><span class="mi">0</span><span class="p">],</span>
|
|
<span class="n">v_lora_params</span><span class="o">.</span><span class="n">lora_weights_pointers</span><span class="p">[</span><span class="mi">0</span><span class="p">],</span>
|
|
<span class="p">],</span>
|
|
<span class="n">host_request_types</span><span class="o">=</span><span class="n">q_lora_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">q_lora_params</span><span class="o">.</span><span class="n">host_context_lengths</span><span class="p">,</span>
|
|
<span class="n">max_encoder_context_length</span><span class="o">=</span><span class="n">q_lora_params</span><span class="o">.</span>
|
|
<span class="n">max_encoder_context_length</span><span class="p">,</span>
|
|
<span class="n">host_encoder_input_lengths</span><span class="o">=</span><span class="n">q_lora_params</span><span class="o">.</span>
|
|
<span class="n">host_encoder_input_lengths</span><span class="p">,</span>
|
|
<span class="p">)</span>
|
|
|
|
<span class="n">q_lora</span><span class="p">,</span> <span class="n">k_lora</span><span class="p">,</span> <span class="n">v_lora</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">qkv_lora_runtime_params</span><span class="p">)</span>
|
|
<span class="n">qkv_lora</span> <span class="o">=</span> <span class="n">concat</span><span class="p">([</span><span class="n">q_lora</span><span class="p">,</span> <span class="n">k_lora</span><span class="p">,</span> <span class="n">v_lora</span><span class="p">],</span>
|
|
<span class="n">dim</span><span class="o">=</span><span class="n">q_lora</span><span class="o">.</span><span class="n">rank</span><span class="p">()</span> <span class="o">-</span> <span class="mi">1</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">qkv_lora</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">qk_layernorm</span><span class="p">:</span>
|
|
<span class="n">base_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="mi">0</span><span class="p">)</span> <span class="k">if</span> <span class="n">qkv</span><span class="o">.</span><span class="n">ndim</span><span class="p">()</span> <span class="o">==</span> <span class="mi">2</span> <span class="k">else</span> <span class="n">concat</span><span class="p">(</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">qkv_sections</span> <span class="o">=</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">num_attention_kv_heads</span><span class="p">,</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">total_heads</span> <span class="o">=</span> <span class="nb">sum</span><span class="p">(</span><span class="n">qkv_sections</span><span class="p">)</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="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">base_shape</span><span class="p">,</span> <span class="n">total_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">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="n">qkv_sections</span><span class="p">,</span> <span class="n">dim</span><span class="o">=</span><span class="n">qkv</span><span class="o">.</span><span class="n">ndim</span><span class="p">()</span> <span class="o">-</span> <span class="mi">2</span><span class="p">)</span>
|
|
<span class="k">else</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">base_shape</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="mi">3</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">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="n">qkv</span><span class="o">.</span><span class="n">ndim</span><span class="p">()</span> <span class="o">-</span> <span class="mi">2</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">base_shape</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">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">normalized_shape</span> <span class="o">=</span> <span class="kc">None</span>
|
|
<span class="k">if</span> <span class="ow">not</span> <span class="bp">self</span><span class="o">.</span><span class="n">layernorm_share</span><span class="p">:</span>
|
|
<span class="n">normalized_shape</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">attention_head_size</span>
|
|
<span class="n">query</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">q_layernorm</span><span class="p">(</span><span class="n">query</span><span class="p">,</span> <span class="n">normalized_shape</span><span class="o">=</span><span class="n">normalized_shape</span><span class="p">)</span>
|
|
<span class="n">key</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">k_layernorm</span><span class="p">(</span><span class="n">key</span><span class="p">,</span> <span class="n">normalized_shape</span><span class="o">=</span><span class="n">normalized_shape</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="n">query</span><span class="o">.</span><span class="n">ndim</span><span class="p">()</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">concat</span><span class="p">([</span><span class="n">base_shape</span><span class="p">,</span> <span class="n">total_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="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="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="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="n">use_cache</span><span class="p">)</span>
|
|
|
|
<span class="k">if</span> <span class="n">use_cache</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="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 steps.</span>
|
|
<span class="c1"># 1st and >1st steps are distinguished by a boolean tensor `cross_kv_cache_gen` passed at runtime</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_kv</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">cross_attention</span> <span class="ow">and</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="k">def</span> <span class="nf">compute_cross_kv</span><span class="p">(</span><span class="n">encoder_output</span><span class="p">):</span>
|
|
<span class="k">if</span> <span class="nb">hasattr</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="s1">'kv'</span><span class="p">):</span>
|
|
<span class="c1"># We optimize the graph by adding kv in the cross attention layer, preventing computing the</span>
|
|
<span class="c1"># query of encoder_output.</span>
|
|
<span class="k">assert</span> <span class="n">qkv_lora_params</span> <span class="o">==</span> <span class="kc">None</span><span class="p">,</span> <span class="s2">"Not support LoRA when we only compute key/value in cross atteniton"</span>
|
|
<span class="c1"># see optimization_model's optimize_cross_qkv</span>
|
|
<span class="n">cross_kv</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">kv</span><span class="p">(</span><span class="n">encoder_output</span><span class="p">,</span> <span class="n">qkv_lora_params</span><span class="p">)</span>
|
|
<span class="n">base_shape</span> <span class="o">=</span> <span class="n">shape</span><span class="p">(</span>
|
|
<span class="n">cross_kv</span><span class="p">,</span> <span class="mi">0</span><span class="p">)</span> <span class="k">if</span> <span class="n">cross_kv</span><span class="o">.</span><span class="n">ndim</span><span class="p">()</span> <span class="o">==</span> <span class="mi">2</span> <span class="k">else</span> <span class="n">concat</span><span class="p">(</span>
|
|
<span class="p">[</span><span class="n">shape</span><span class="p">(</span><span class="n">cross_kv</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">cross_kv</span><span class="p">,</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">qk_layernorm</span><span class="p">:</span>
|
|
<span class="n">cross_kv</span> <span class="o">=</span> <span class="n">cross_kv</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">base_shape</span><span class="p">,</span> <span class="mi">2</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="bp">self</span><span class="o">.</span><span class="n">attention_head_size</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">cross_kv</span><span class="p">,</span> <span class="p">[</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">num_attention_kv_heads</span><span class="p">,</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">dim</span><span class="o">=</span><span class="n">cross_kv</span><span class="o">.</span><span class="n">ndim</span><span class="p">()</span> <span class="o">-</span> <span class="mi">2</span><span class="p">)</span>
|
|
|
|
<span class="n">key</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">k_layernorm</span><span class="p">(</span><span class="n">key</span><span class="p">)</span>
|
|
<span class="n">cross_kv</span> <span class="o">=</span> <span class="n">concat</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="n">key</span><span class="o">.</span><span class="n">ndim</span><span class="p">()</span> <span class="o">-</span> <span class="mi">2</span><span class="p">)</span>
|
|
<span class="k">else</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="n">qkv_lora_params</span><span class="p">)</span>
|
|
<span class="n">base_shape</span> <span class="o">=</span> <span class="n">shape</span><span class="p">(</span>
|
|
<span class="n">cross_qkv</span><span class="p">,</span> <span class="mi">0</span><span class="p">)</span> <span class="k">if</span> <span class="n">cross_qkv</span><span class="o">.</span><span class="n">ndim</span><span class="p">()</span> <span class="o">==</span> <span class="mi">2</span> <span class="k">else</span> <span class="n">concat</span><span class="p">(</span>
|
|
<span class="p">[</span><span class="n">shape</span><span class="p">(</span><span class="n">cross_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">cross_qkv</span><span class="p">,</span> <span class="mi">1</span><span class="p">)])</span>
|
|
|
|
<span class="n">cross_qkv</span> <span class="o">=</span> <span class="n">cross_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">base_shape</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="mi">2</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="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="bp">self</span><span class="o">.</span><span class="n">qk_layernorm</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">cross_qkv</span><span class="p">,</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">num_attention_kv_heads</span><span class="p">,</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">dim</span><span class="o">=</span><span class="n">cross_qkv</span><span class="o">.</span><span class="n">ndim</span><span class="p">()</span> <span class="o">-</span> <span class="mi">2</span><span class="p">)</span>
|
|
|
|
<span class="n">key</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">k_layernorm</span><span class="p">(</span><span class="n">key</span><span class="p">)</span>
|
|
<span class="n">cross_kv</span> <span class="o">=</span> <span class="n">concat</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="n">key</span><span class="o">.</span><span class="n">ndim</span><span class="p">()</span> <span class="o">-</span> <span class="mi">2</span><span class="p">)</span>
|
|
<span class="k">else</span><span class="p">:</span>
|
|
<span class="n">_</span><span class="p">,</span> <span class="n">cross_kv</span> <span class="o">=</span> <span class="n">split</span><span class="p">(</span><span class="n">cross_qkv</span><span class="p">,</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">num_attention_kv_heads</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="n">cross_qkv</span><span class="o">.</span><span class="n">ndim</span><span class="p">()</span> <span class="o">-</span> <span class="mi">2</span><span class="p">)</span>
|
|
<span class="n">cross_kv</span> <span class="o">=</span> <span class="n">cross_kv</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">base_shape</span><span class="p">,</span> <span class="mi">2</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="k">if</span> <span class="n">default_net</span><span class="p">(</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="ow">and</span> <span class="n">qkv_lora_params</span> <span class="ow">is</span> <span class="kc">None</span> <span class="ow">and</span> <span class="n">lora_layer_params</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="n">_</span><span class="p">,</span> <span class="n">cross_k_lora</span><span class="p">,</span> <span class="n">cross_v_lora</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">encoder_output</span><span class="p">,</span>
|
|
<span class="n">qkv_lora_runtime_params</span><span class="p">,</span>
|
|
<span class="n">is_cross_attention</span><span class="o">=</span><span class="kc">True</span><span class="p">)</span>
|
|
<span class="n">cross_kv_lora</span> <span class="o">=</span> <span class="n">concat</span><span class="p">([</span><span class="n">cross_k_lora</span><span class="p">,</span> <span class="n">cross_v_lora</span><span class="p">],</span>
|
|
<span class="n">dim</span><span class="o">=</span><span class="n">cross_k_lora</span><span class="o">.</span><span class="n">rank</span><span class="p">()</span> <span class="o">-</span> <span class="mi">1</span><span class="p">)</span>
|
|
<span class="n">cross_kv</span> <span class="o">=</span> <span class="n">cross_kv</span> <span class="o">+</span> <span class="n">cross_kv_lora</span>
|
|
|
|
<span class="k">return</span> <span class="n">cross_kv</span>
|
|
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">skip_cross_kv</span><span class="p">:</span>
|
|
<span class="n">conditional</span> <span class="o">=</span> <span class="n">Conditional</span><span class="p">(</span><span class="n">cross_kv_cache_gen</span><span class="p">)</span>
|
|
<span class="n">cond_in1</span> <span class="o">=</span> <span class="n">conditional</span><span class="o">.</span><span class="n">add_input</span><span class="p">(</span><span class="n">encoder_output</span><span class="p">)</span>
|
|
<span class="n">cond_in2</span> <span class="o">=</span> <span class="n">conditional</span><span class="o">.</span><span class="n">add_input</span><span class="p">(</span><span class="n">cross_kv_reuse</span><span class="p">)</span>
|
|
|
|
<span class="c1">## True branch: context phase, compute cross qkv</span>
|
|
<span class="n">cross_kv_true</span> <span class="o">=</span> <span class="n">compute_cross_kv</span><span class="p">(</span><span class="n">cond_in1</span><span class="p">)</span>
|
|
|
|
<span class="c1">## False branch: generation phase, no compute but need to obey shape constraints</span>
|
|
<span class="c1"># because TRT's IfConditional requires the output shape of two subgraphs to be identical</span>
|
|
<span class="c1"># our 1st attempt was to stack encoder_output [B, S, H] or [N, H] --> cross qkv [B, S, 3*H] or [N, 3*H],</span>
|
|
<span class="c1"># but it still introduces unnecessary concat. A better solution is to create a dummy torch tensor `cross_kv_resue`</span>
|
|
<span class="c1"># with the correct shape and reuse it in every generation step</span>
|
|
<span class="n">cross_kv_false</span> <span class="o">=</span> <span class="n">cond_in2</span>
|
|
<span class="n">cross_kv</span> <span class="o">=</span> <span class="n">conditional</span><span class="o">.</span><span class="n">add_output</span><span class="p">(</span><span class="n">cross_kv_true</span><span class="p">,</span> <span class="n">cross_kv_false</span><span class="p">)</span>
|
|
<span class="k">else</span><span class="p">:</span>
|
|
<span class="n">cross_kv</span> <span class="o">=</span> <span class="n">compute_cross_kv</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">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="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="n">AttentionMaskType</span><span class="o">.</span><span class="n">blocksparse</span>
|
|
<span class="p">],</span> <span class="s1">'Plugin only support masked MHA.'</span>
|
|
|
|
<span class="c1"># KV cache scales.</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">kv_cache_scaling_factor</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="n">kv_orig_quant_scale</span> <span class="o">=</span> <span class="n">constant</span><span class="p">(</span><span class="n">fp32_array</span><span class="p">(</span>
|
|
<span class="p">[</span><span class="mf">1.0</span><span class="p">]))</span> <span class="o">/</span> <span class="bp">self</span><span class="o">.</span><span class="n">kv_cache_scaling_factor</span><span class="o">.</span><span class="n">value</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_cache_scaling_factor</span><span class="o">.</span><span class="n">value</span>
|
|
<span class="k">else</span><span class="p">:</span>
|
|
<span class="n">kv_orig_quant_scale</span> <span class="o">=</span> <span class="kc">None</span>
|
|
<span class="n">kv_quant_orig_scale</span> <span class="o">=</span> <span class="kc">None</span>
|
|
|
|
<span class="c1"># Attention output scales</span>
|
|
<span class="k">assert</span> <span class="p">(</span>
|
|
<span class="ow">not</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">use_fp8_context_fmha</span>
|
|
<span class="p">)</span> <span class="ow">or</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="p">),</span> <span class="s2">"FP8 Context FMHA must be used together with the fp8 quantization workflow."</span>
|
|
|
|
<span class="n">attention_output_orig_quant_scale</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">attention_output_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">attention_output_orig_quant_scale</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span> <span class="k">else</span> <span class="kc">None</span>
|
|
|
|
<span class="c1"># The rotary inv freq can be pre-computed.</span>
|
|
<span class="n">rotary_inv_freq</span> <span class="o">=</span> <span class="nb">getattr</span><span class="p">(</span><span class="n">attention_params</span><span class="p">,</span> <span class="s2">"rotary_inv_freq"</span><span class="p">,</span> <span class="kc">None</span><span class="p">)</span>
|
|
<span class="c1"># Rotary cos/sin cache.</span>
|
|
<span class="n">rotary_cos_sin</span> <span class="o">=</span> <span class="nb">getattr</span><span class="p">(</span><span class="n">attention_params</span><span class="p">,</span>
|
|
<span class="s2">"embed_positions_for_gpt_attention"</span><span class="p">,</span> <span class="kc">None</span><span class="p">)</span>
|
|
|
|
<span class="n">long_rope_rotary_inv_freq</span> <span class="o">=</span> <span class="nb">getattr</span><span class="p">(</span><span class="n">attention_params</span><span class="p">,</span>
|
|
<span class="s2">"long_rope_rotary_inv_freq"</span><span class="p">,</span>
|
|
<span class="kc">None</span><span class="p">)</span>
|
|
<span class="n">long_rope_rotary_cos_sin</span> <span class="o">=</span> <span class="nb">getattr</span><span class="p">(</span>
|
|
<span class="n">attention_params</span><span class="p">,</span> <span class="s2">"long_rope_embed_positions_for_gpt_attention"</span><span class="p">,</span>
|
|
<span class="kc">None</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">learned_absolute</span><span class="p">:</span>
|
|
<span class="n">rotary_inv_freq</span> <span class="o">=</span> <span class="kc">None</span>
|
|
<span class="n">rotary_cos_sin</span> <span class="o">=</span> <span class="kc">None</span>
|
|
|
|
<span class="c1"># check if the cache is provided.</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="k">assert</span> <span class="p">(</span><span class="n">rotary_inv_freq</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">)</span> <span class="ow">and</span> <span class="p">(</span>
|
|
<span class="n">rotary_cos_sin</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span>
|
|
<span class="p">),</span> <span class="s2">"rotary_inv_freq and embed_positions_for_gpt_attention must be provided."</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">long_rope</span><span class="p">:</span>
|
|
<span class="k">assert</span> <span class="n">long_rope_rotary_inv_freq</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span>
|
|
<span class="k">assert</span> <span class="n">long_rope_rotary_cos_sin</span> <span class="ow">is</span> <span class="ow">not</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">qkv</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">attention_mask</span><span class="o">=</span><span class="n">attention_mask</span><span class="p">,</span>
|
|
<span class="n">attention_packed_mask</span><span class="o">=</span><span class="n">attention_packed_mask</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_attention_window_sizes</span><span class="o">=</span><span class="n">kv_cache_params</span><span class="o">.</span>
|
|
<span class="n">host_max_attention_window_sizes</span><span class="p">,</span>
|
|
<span class="n">host_sink_token_length</span><span class="o">=</span><span class="n">kv_cache_params</span><span class="o">.</span><span class="n">host_sink_token_length</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">layer_idx</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">local_layer_idx</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">layer_idx_in_cache_pool</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">layer_idx_in_cache_pool</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_short_m_scale</span><span class="o">=</span><span class="n">attention_params</span><span class="o">.</span><span class="n">short_mscale</span><span class="p">,</span>
|
|
<span class="n">rotary_embedding_long_m_scale</span><span class="o">=</span><span class="n">attention_params</span><span class="o">.</span><span class="n">long_mscale</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">rotary_embedding_original_max_positions</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span>
|
|
<span class="n">original_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">rotary_inv_freq</span><span class="o">=</span><span class="n">rotary_inv_freq</span><span class="p">,</span>
|
|
<span class="n">rotary_cos_sin</span><span class="o">=</span><span class="n">rotary_cos_sin</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">attention_output_orig_quant_scale</span><span class="o">=</span>
|
|
<span class="n">attention_output_orig_quant_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">block_sparse_block_size</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">block_sparse_params</span><span class="o">.</span><span class="n">block_size</span><span class="p">,</span>
|
|
<span class="n">block_sparse_homo_head_pattern</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">block_sparse_params</span><span class="o">.</span>
|
|
<span class="n">homo_head_pattern</span><span class="p">,</span>
|
|
<span class="n">block_sparse_num_local_blocks</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">block_sparse_params</span><span class="o">.</span>
|
|
<span class="n">num_local_blocks</span><span class="p">,</span>
|
|
<span class="n">block_sparse_vertical_stride</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">block_sparse_params</span><span class="o">.</span>
|
|
<span class="n">vertical_stride</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_offsets</span><span class="o">=</span><span class="n">kv_cache_params</span><span class="o">.</span><span class="n">kv_cache_block_offsets</span>
|
|
<span class="k">if</span> <span class="ow">not</span> <span class="bp">self</span><span class="o">.</span><span class="n">cross_attention</span> <span class="k">else</span>
|
|
<span class="n">kv_cache_params</span><span class="o">.</span><span class="n">cross_kv_cache_block_offsets</span><span class="p">,</span>
|
|
<span class="n">host_kv_cache_block_offsets</span><span class="o">=</span><span class="n">kv_cache_params</span><span class="o">.</span>
|
|
<span class="n">host_kv_cache_block_offsets</span> <span class="k">if</span> <span class="ow">not</span> <span class="bp">self</span><span class="o">.</span><span class="n">cross_attention</span> <span class="k">else</span>
|
|
<span class="n">kv_cache_params</span><span class="o">.</span><span class="n">host_cross_kv_cache_block_offsets</span><span class="p">,</span>
|
|
<span class="n">host_kv_cache_pool_pointers</span><span class="o">=</span><span class="n">kv_cache_params</span><span class="o">.</span>
|
|
<span class="n">host_kv_cache_pool_pointers</span> <span class="k">if</span> <span class="ow">not</span> <span class="bp">self</span><span class="o">.</span><span class="n">cross_attention</span> <span class="k">else</span>
|
|
<span class="n">kv_cache_params</span><span class="o">.</span><span class="n">host_cross_kv_cache_pool_pointers</span><span class="p">,</span>
|
|
<span class="n">host_kv_cache_pool_mapping</span><span class="o">=</span><span class="n">kv_cache_params</span><span class="o">.</span>
|
|
<span class="n">host_kv_cache_pool_mapping</span> <span class="k">if</span> <span class="ow">not</span> <span class="bp">self</span><span class="o">.</span><span class="n">cross_attention</span> <span class="k">else</span>
|
|
<span class="n">kv_cache_params</span><span class="o">.</span><span class="n">host_cross_kv_cache_pool_mapping</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_kv</span><span class="o">=</span><span class="n">cross_kv</span><span class="p">,</span>
|
|
<span class="n">cross_kv_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">logn_scaling</span><span class="o">=</span><span class="n">logn_scaling</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="n">use_cache</span><span class="o">=</span><span class="n">use_cache</span><span class="p">,</span>
|
|
<span class="n">spec_decoding_is_generation_length_variable</span><span class="o">=</span><span class="n">spec_decoding_params</span>
|
|
<span class="o">.</span><span class="n">spec_decoding_is_generation_length_variable</span><span class="p">,</span>
|
|
<span class="n">spec_decoding_max_generation_length</span><span class="o">=</span><span class="n">spec_decoding_params</span><span class="o">.</span>
|
|
<span class="n">spec_decoding_max_generation_length</span><span class="p">,</span>
|
|
<span class="n">spec_decoding_generation_lengths</span><span class="o">=</span><span class="n">spec_decoding_params</span><span class="o">.</span>
|
|
<span class="n">spec_decoding_generation_lengths</span><span class="p">,</span>
|
|
<span class="n">spec_decoding_position_offsets</span><span class="o">=</span><span class="n">spec_decoding_params</span><span class="o">.</span>
|
|
<span class="n">spec_decoding_position_offsets</span><span class="p">,</span>
|
|
<span class="n">spec_decoding_packed_mask</span><span class="o">=</span><span class="n">spec_decoding_params</span><span class="o">.</span>
|
|
<span class="n">spec_decoding_packed_mask</span><span class="p">,</span>
|
|
<span class="n">long_rope_rotary_inv_freq</span><span class="o">=</span><span class="n">long_rope_rotary_inv_freq</span><span class="p">,</span>
|
|
<span class="n">long_rope_rotary_cos_sin</span><span class="o">=</span><span class="n">long_rope_rotary_cos_sin</span><span class="p">,</span>
|
|
<span class="n">mrope_rotary_sin_cos</span><span class="o">=</span><span class="n">mrope_params</span><span class="o">.</span><span class="n">mrope_rotary_sin_cos</span><span class="p">,</span>
|
|
<span class="n">mrope_position_deltas</span><span class="o">=</span><span class="n">mrope_params</span><span class="o">.</span><span class="n">mrope_position_deltas</span><span class="p">,</span>
|
|
<span class="n">attn_logit_softcapping_scale</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">max_attn_value</span><span class="p">,</span>
|
|
<span class="n">host_runtime_perf_knobs</span><span class="o">=</span><span class="n">attention_params</span><span class="o">.</span>
|
|
<span class="n">host_runtime_perf_knobs</span><span class="p">,</span>
|
|
<span class="n">host_context_progress</span><span class="o">=</span><span class="n">attention_params</span><span class="o">.</span><span class="n">host_context_progress</span><span class="p">,</span>
|
|
<span class="n">skip_attn</span><span class="o">=</span><span class="n">skip_attn</span><span class="p">,</span>
|
|
<span class="n">cp_size</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">cp_size</span><span class="p">,</span>
|
|
<span class="n">cp_rank</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">cp_rank</span><span class="p">,</span>
|
|
<span class="n">cp_group</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">cp_group</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="k">assert</span> <span class="n">logn_scaling</span> <span class="ow">is</span> <span class="kc">None</span><span class="p">,</span> <span class="s2">"plan TensorRT mode does not support logn scaling now"</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="k">if</span> <span class="n">unfuse_qkv_gemm</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">qkv</span><span class="p">[</span><span class="mi">0</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">qkv</span><span class="p">[</span><span class="mi">2</span><span class="p">]</span>
|
|
<span class="k">else</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="p">[</span><span class="bp">self</span><span class="o">.</span><span class="n">attention_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">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">cross_kv</span><span class="p">,</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">position_embedding_type</span><span class="o">.</span><span class="n">is_rope</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">position_embedding_type</span><span class="o">.</span><span class="n">is_rope</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">position_embedding_type</span><span class="o">.</span><span class="n">is_rope</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">long_rope</span><span class="p">:</span>
|
|
<span class="n">sequence_length</span> <span class="o">=</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="n">floor_seq_length</span> <span class="o">=</span> <span class="n">maximum</span><span class="p">(</span>
|
|
<span class="n">sequence_length</span><span class="p">,</span> <span class="bp">self</span><span class="o">.</span><span class="n">original_max_position_embeddings</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="mi">0</span><span class="p">])</span>
|
|
<span class="n">shapes</span> <span class="o">=</span> <span class="n">concat</span><span class="p">(</span>
|
|
<span class="p">[</span><span class="mi">1</span><span class="p">,</span> <span class="n">floor_seq_length</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">short</span> <span class="o">=</span> <span class="nb">slice</span><span class="p">(</span><span class="n">attention_params</span><span class="o">.</span><span class="n">embed_positions</span><span class="p">,</span> <span class="n">starts</span><span class="p">,</span>
|
|
<span class="n">shapes</span><span class="p">)</span>
|
|
<span class="n">long</span> <span class="o">=</span> <span class="nb">slice</span><span class="p">(</span><span class="n">attention_params</span><span class="o">.</span><span class="n">long_rope_embed_positions</span><span class="p">,</span>
|
|
<span class="n">starts</span><span class="p">,</span> <span class="n">shapes</span><span class="p">)</span>
|
|
|
|
<span class="n">embed_positions</span> <span class="o">=</span> <span class="n">concat</span><span class="p">([</span><span class="n">short</span><span class="p">,</span> <span class="n">long</span><span class="p">],</span> <span class="n">dim</span><span class="o">=</span><span class="mi">0</span><span class="p">)</span>
|
|
<span class="n">select</span> <span class="o">=</span> <span class="n">where</span><span class="p">(</span>
|
|
<span class="n">sequence_length</span> <span class="o"><=</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">original_max_position_embeddings</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">embed_positions</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="n">select</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">sizes</span><span class="o">=</span><span class="n">shape</span><span class="p">(</span><span class="n">short</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="bp">self</span><span class="o">.</span><span class="n">dtype</span><span class="p">)</span>
|
|
<span class="k">elif</span> <span class="n">is_same_dtype</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">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">cast</span><span class="p">(</span><span class="n">attention_params</span><span class="o">.</span><span class="n">embed_positions</span><span class="p">,</span>
|
|
<span class="n">trt</span><span class="o">.</span><span class="n">bfloat16</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">cast</span><span class="p">(</span><span class="n">attention_params</span><span class="o">.</span><span class="n">embed_positions</span><span class="p">,</span>
|
|
<span class="n">query</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="k">if</span> <span class="ow">not</span> <span class="n">use_cache</span><span class="p">:</span>
|
|
<span class="c1"># Only context phase is involved when kv cache is disabled.</span>
|
|
<span class="n">start</span> <span class="o">=</span> <span class="mi">0</span>
|
|
<span class="k">else</span><span class="p">:</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="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="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="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">kv_cache_scaling_factor</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">dequantize</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_cache_scaling_factor</span><span class="o">.</span><span class="n">value</span><span class="p">,</span>
|
|
<span class="n">output_type</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">dtype</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="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="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="c1"># TRT quantizes the tensor value by doing `cast(clip(fp_value / scale))` while</span>
|
|
<span class="c1"># the plugin quantizes it by doing `cast(clip(fp_value * scale))`.</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">kv_cache_scaling_factor</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">quantize</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_cache_scaling_factor</span><span class="o">.</span><span class="n">value</span><span class="p">,</span>
|
|
<span class="n">dtype</span><span class="o">=</span><span class="s1">'fp8'</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_fp8_kv_cache</span><span class="p">()</span> <span class="k">else</span> <span class="s1">'int8'</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="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="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">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="c1"># bsz, tatget_length, past_key_value_length</span>
|
|
<span class="n">buffer</span> <span class="o">=</span> <span class="n">make_causal_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">0</span><span class="p">),</span> <span class="n">query_length</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="n">trt</span><span class="o">.</span><span class="n">float32</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="mi">0</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">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">else</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="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">long_rope</span><span class="p">:</span>
|
|
<span class="n">buf_shape</span> <span class="o">=</span> <span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">original_max_position_embeddings</span><span class="p">,</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">original_max_position_embeddings</span><span class="p">)</span>
|
|
<span class="k">else</span><span class="p">:</span>
|
|
<span class="n">buf_shape</span> <span class="o">=</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">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="n">buf_shape</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="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">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="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">cross_attention</span><span class="p">:</span>
|
|
<span class="n">batch_size</span> <span class="o">=</span> <span class="n">shape</span><span class="p">(</span><span class="n">attention_mask</span><span class="p">,</span> <span class="mi">0</span><span class="p">)</span>
|
|
<span class="n">query_len</span> <span class="o">=</span> <span class="n">shape</span><span class="p">(</span><span class="n">attention_mask</span><span class="p">,</span> <span class="mi">1</span><span class="p">)</span>
|
|
<span class="n">encoder_input_len</span> <span class="o">=</span> <span class="n">shape</span><span class="p">(</span><span class="n">attention_mask</span><span class="p">,</span> <span class="mi">2</span><span class="p">)</span>
|
|
<span class="n">attention_mask</span> <span class="o">=</span> <span class="n">attention_mask</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">batch_size</span><span class="p">,</span> <span class="mi">1</span><span class="p">,</span> <span class="n">query_len</span><span class="p">,</span> <span class="n">encoder_input_len</span><span class="p">]))</span>
|
|
<span class="n">attention_mask</span> <span class="o">=</span> <span class="n">where</span><span class="p">(</span><span class="n">attention_mask</span> <span class="o">==</span> <span class="mi">0</span><span class="p">,</span> <span class="nb">float</span><span class="p">(</span><span class="s1">'-inf'</span><span class="p">),</span>
|
|
<span class="mf">0.0</span><span class="p">)</span>
|
|
<span class="k">else</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="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">relative_attention</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="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">use_implicit_relative_attention</span><span class="p">:</span>
|
|
<span class="n">relative_bias</span> <span class="o">=</span> <span class="n">compute_relative_bias</span><span class="p">(</span>
|
|
<span class="n">query_length</span> <span class="o">+</span> <span class="n">key_length</span> <span class="o">-</span> <span class="mi">1</span><span class="p">,</span>
|
|
<span class="n">key_length</span><span class="p">,</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">num_buckets</span><span class="p">,</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">max_distance</span><span class="p">,</span>
|
|
<span class="kc">False</span><span class="p">,</span> <span class="c1"># bidirectional</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="o">.</span><span class="n">transpose</span><span class="p">(</span><span class="mi">1</span><span class="p">,</span> <span class="mi">0</span><span class="p">),</span>
|
|
<span class="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_group</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">tp_group</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="k">else</span><span class="p">:</span>
|
|
<span class="n">relative_bias</span> <span class="o">=</span> <span class="n">unsqueeze</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">value</span><span class="p">,</span> <span class="mi">0</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">query_length</span> <span class="o">+</span> <span class="n">key_length</span> <span class="o">-</span> <span class="mi">2</span><span class="p">,</span> <span class="mi">0</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="n">shape</span><span class="p">(</span><span class="n">relative_bias</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">relative_bias</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">key_length</span>
|
|
<span class="p">])</span>
|
|
<span class="n">relative_bias</span> <span class="o">=</span> <span class="nb">slice</span><span class="p">(</span><span class="n">relative_bias</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="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="p">(</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">norm_factor</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="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="p">(</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">norm_factor</span><span class="p">)</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">max_attn_value</span> <span class="o">></span> <span class="mi">0</span><span class="p">:</span>
|
|
<span class="n">attention_scores</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">max_attn_value</span> <span class="o">*</span> <span class="n">ACT2FN</span><span class="p">[</span><span class="s1">'tanh'</span><span class="p">](</span>
|
|
<span class="n">attention_scores</span> <span class="o">/</span> <span class="bp">self</span><span class="o">.</span><span class="n">max_attn_value</span><span class="p">)</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="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">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="p">:</span>
|
|
<span class="n">bias</span> <span class="o">=</span> <span class="n">cast</span><span class="p">(</span><span class="n">bias</span><span class="p">,</span> <span class="n">attention_scores</span><span class="o">.</span><span class="n">dtype</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="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">relative_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">relative_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="c1"># A dummy reshape WAR for mha fusion</span>
|
|
<span class="n">attention_probs</span> <span class="o">=</span> <span class="n">attention_probs</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">attention_probs</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">attention_probs</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">attention_probs</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">value</span><span class="p">,</span> <span class="mi">2</span><span class="p">)</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="n">use_fp32_acc</span><span class="o">=</span><span class="kc">False</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">attention_hidden_size</span>
|
|
<span class="p">]))</span>
|
|
|
|
<span class="n">dense_lora_params</span> <span class="o">=</span> <span class="kc">None</span>
|
|
<span class="k">if</span> <span class="n">lora_layer_params</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="n">dense_lora_params</span> <span class="o">=</span> <span class="n">lora_layer_params</span><span class="o">.</span><span class="n">get_runtime_params</span><span class="p">(</span>
|
|
<span class="mi">0</span><span class="p">,</span> <span class="s2">"attn_dense"</span><span class="p">)</span>
|
|
|
|
<span class="k">if</span> <span class="n">skip_attn</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="c1"># This case is used when we can skip this attention layer directly.</span>
|
|
<span class="c1"># The output would be undefined and not used if skip_attn is not None</span>
|
|
<span class="c1"># and set skip_attn as True during runtime</span>
|
|
|
|
<span class="n">dense_conditional</span> <span class="o">=</span> <span class="n">Conditional</span><span class="p">(</span><span class="n">skip_attn</span><span class="p">)</span>
|
|
<span class="n">skip_case</span> <span class="o">=</span> <span class="n">dense_conditional</span><span class="o">.</span><span class="n">add_input</span><span class="p">(</span><span class="n">attention_input</span><span class="p">)</span>
|
|
<span class="n">context</span> <span class="o">=</span> <span class="n">dense_conditional</span><span class="o">.</span><span class="n">add_input</span><span class="p">(</span><span class="n">context</span><span class="p">)</span>
|
|
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">inner_layernorm</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</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">inner_layernorm</span><span class="p">(</span><span class="n">context</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">lora_runtime_params</span><span class="o">=</span><span class="n">dense_lora_params</span><span class="p">,</span>
|
|
<span class="n">all_reduce_params</span><span class="o">=</span><span class="n">all_reduce_params</span><span class="p">)</span>
|
|
|
|
<span class="k">if</span> <span class="n">skip_attn</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="n">context</span> <span class="o">=</span> <span class="n">dense_conditional</span><span class="o">.</span><span class="n">add_output</span><span class="p">(</span><span class="n">skip_case</span><span class="p">,</span> <span class="n">context</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 class="viewcode-block" id="Attention.set_rel_attn_table">
|
|
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.layers.html#tensorrt_llm.layers.attention.Attention.set_rel_attn_table">[docs]</a>
|
|
<span class="k">def</span> <span class="nf">set_rel_attn_table</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">max_seq_len</span><span class="p">,</span> <span class="n">precomputed_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="bp">self</span><span class="o">.</span><span class="n">num_attention_heads</span><span class="p">,</span>
|
|
<span class="n">max_seq_len</span> <span class="o">+</span> <span class="mi">1</span><span class="p">,</span>
|
|
<span class="n">max_seq_len</span> <span class="o">+</span> <span class="mi">1</span><span class="p">),</span>
|
|
<span class="n">dtype</span><span class="o">=</span><span class="bp">self</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">rel_attn_table</span><span class="o">.</span><span class="n">value</span> <span class="o">=</span> <span class="n">precomputed_relative_attention</span></div>
|
|
|
|
|
|
<div class="viewcode-block" id="Attention.postprocess">
|
|
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.layers.html#tensorrt_llm.layers.attention.Attention.postprocess">[docs]</a>
|
|
<span class="k">def</span> <span class="nf">postprocess</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">tllm_key</span><span class="p">,</span> <span class="n">weights</span><span class="p">,</span> <span class="o">**</span><span class="n">kwargs</span><span class="p">):</span>
|
|
<span class="k">if</span> <span class="n">tllm_key</span><span class="o">.</span><span class="n">endswith</span><span class="p">(</span><span class="s2">"kv_cache_scaling_factor"</span><span class="p">):</span>
|
|
<span class="k">if</span> <span class="n">weights</span> <span class="ow">is</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="k">return</span> <span class="p">{</span><span class="n">tllm_key</span><span class="p">:</span> <span class="n">torch</span><span class="o">.</span><span class="n">ones</span><span class="p">(</span><span class="mi">1</span><span class="p">,</span> <span class="p">)}</span>
|
|
<span class="k">elif</span> <span class="nb">isinstance</span><span class="p">(</span><span class="n">weights</span><span class="p">,</span> <span class="n">torch</span><span class="o">.</span><span class="n">Tensor</span><span class="p">):</span>
|
|
<span class="k">return</span> <span class="p">{</span><span class="n">tllm_key</span><span class="p">:</span> <span class="n">weights</span><span class="p">}</span>
|
|
<span class="k">elif</span> <span class="kc">None</span> <span class="ow">in</span> <span class="n">weights</span><span class="p">:</span>
|
|
<span class="k">return</span> <span class="p">{</span><span class="n">tllm_key</span><span class="p">:</span> <span class="n">torch</span><span class="o">.</span><span class="n">ones</span><span class="p">(</span><span class="mi">1</span><span class="p">,</span> <span class="p">)}</span>
|
|
<span class="k">else</span><span class="p">:</span>
|
|
<span class="k">return</span> <span class="p">{</span><span class="n">tllm_key</span><span class="p">:</span> <span class="nb">max</span><span class="p">(</span><span class="n">weights</span><span class="p">)}</span>
|
|
<span class="k">else</span><span class="p">:</span>
|
|
<span class="k">return</span> <span class="p">{</span><span class="n">tllm_key</span><span class="p">:</span> <span class="n">weights</span><span class="p">}</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">cp_group</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">cp_size</span><span class="o">=</span><span class="mi">1</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">quant_mode</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="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="bp">self</span><span class="o">.</span><span class="n">attention_hidden_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_heads</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_group</span> <span class="o">=</span> <span class="n">tp_group</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">cp_group</span> <span class="o">=</span> <span class="n">cp_group</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">cp_size</span> <span class="o">=</span> <span class="n">cp_size</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="c1"># add quant mode to control quantization</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="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">num_buckets</span> <span class="o">=</span> <span class="n">num_buckets</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">attention_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">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="n">is_qkv</span><span class="o">=</span><span class="kc">True</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"># see optimize_model's add_lora for LoRA initialization</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">qkv_lora</span> <span class="o">=</span> <span class="kc">None</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">max_input_length</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">lora_layer_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">qkv_lora_params</span> <span class="o">=</span> <span class="kc">None</span>
|
|
<span class="k">if</span> <span class="n">lora_layer_params</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="n">qkv_lora_params</span> <span class="o">=</span> <span class="n">lora_layer_params</span><span class="o">.</span><span class="n">get_runtime_params</span><span class="p">(</span>
|
|
<span class="mi">0</span><span class="p">,</span> <span class="s2">"attn_qkv"</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="n">qkv_lora_params</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">remove_input_padding</span><span class="p">:</span>
|
|
<span class="k">assert</span> <span class="n">qkv</span><span class="o">.</span><span class="n">ndim</span><span class="p">()</span> <span class="o">==</span> <span class="mi">2</span>
|
|
|
|
<span class="k">if</span> <span class="n">default_net</span><span class="p">(</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="ow">and</span> <span class="n">qkv_lora_params</span> <span class="ow">is</span> <span class="kc">None</span> <span class="ow">and</span> <span class="n">lora_layer_params</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="n">q_lora_params</span> <span class="o">=</span> <span class="n">lora_layer_params</span><span class="o">.</span><span class="n">get_runtime_params</span><span class="p">(</span><span class="mi">0</span><span class="p">,</span> <span class="s2">"attn_q"</span><span class="p">)</span>
|
|
<span class="n">k_lora_params</span> <span class="o">=</span> <span class="n">lora_layer_params</span><span class="o">.</span><span class="n">get_runtime_params</span><span class="p">(</span><span class="mi">0</span><span class="p">,</span> <span class="s2">"attn_k"</span><span class="p">)</span>
|
|
<span class="n">v_lora_params</span> <span class="o">=</span> <span class="n">lora_layer_params</span><span class="o">.</span><span class="n">get_runtime_params</span><span class="p">(</span><span class="mi">0</span><span class="p">,</span> <span class="s2">"attn_v"</span><span class="p">)</span>
|
|
|
|
<span class="k">assert</span> <span class="p">(</span><span class="n">q_lora_params</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span> <span class="ow">and</span> <span class="n">k_lora_params</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span> <span class="ow">and</span> <span class="n">v_lora_params</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">)</span> <span class="ow">or</span> \
|
|
<span class="p">(</span><span class="n">q_lora_params</span> <span class="ow">is</span> <span class="kc">None</span> <span class="ow">and</span> <span class="n">k_lora_params</span> <span class="ow">is</span> <span class="kc">None</span> <span class="ow">and</span> <span class="n">v_lora_params</span> <span class="ow">is</span> <span class="kc">None</span><span class="p">),</span> <span class="s2">"q_lora_params, k_lora_params and v_lora_params should be all enabled or all disabled at the same time."</span>
|
|
|
|
<span class="k">if</span> <span class="n">q_lora_params</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span> <span class="ow">and</span> <span class="n">k_lora_params</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span> <span class="ow">and</span> <span class="n">v_lora_params</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="n">qkv_lora_params</span> <span class="o">=</span> <span class="n">LoraRuntimeParams</span><span class="p">(</span>
|
|
<span class="n">lora_ranks</span><span class="o">=</span><span class="p">[</span>
|
|
<span class="n">q_lora_params</span><span class="o">.</span><span class="n">lora_ranks</span><span class="p">[</span><span class="mi">0</span><span class="p">],</span>
|
|
<span class="n">k_lora_params</span><span class="o">.</span><span class="n">lora_ranks</span><span class="p">[</span><span class="mi">0</span><span class="p">],</span>
|
|
<span class="n">v_lora_params</span><span class="o">.</span><span class="n">lora_ranks</span><span class="p">[</span><span class="mi">0</span><span class="p">],</span>
|
|
<span class="p">],</span>
|
|
<span class="n">lora_weights_pointers</span><span class="o">=</span><span class="p">[</span>
|
|
<span class="n">q_lora_params</span><span class="o">.</span><span class="n">lora_weights_pointers</span><span class="p">[</span><span class="mi">0</span><span class="p">],</span>
|
|
<span class="n">k_lora_params</span><span class="o">.</span><span class="n">lora_weights_pointers</span><span class="p">[</span><span class="mi">0</span><span class="p">],</span>
|
|
<span class="n">v_lora_params</span><span class="o">.</span><span class="n">lora_weights_pointers</span><span class="p">[</span><span class="mi">0</span><span class="p">],</span>
|
|
<span class="p">],</span>
|
|
<span class="n">host_request_types</span><span class="o">=</span><span class="n">q_lora_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">q_lora_params</span><span class="o">.</span><span class="n">host_context_lengths</span><span class="p">)</span>
|
|
|
|
<span class="n">q_lora</span><span class="p">,</span> <span class="n">k_lora</span><span class="p">,</span> <span class="n">v_lora</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">qkv_lora_params</span><span class="p">)</span>
|
|
<span class="n">qkv_lora</span> <span class="o">=</span> <span class="n">concat</span><span class="p">([</span><span class="n">q_lora</span><span class="p">,</span> <span class="n">k_lora</span><span class="p">,</span> <span class="n">v_lora</span><span class="p">],</span>
|
|
<span class="n">dim</span><span class="o">=</span><span class="n">q_lora</span><span class="o">.</span><span class="n">rank</span><span class="p">()</span> <span class="o">-</span> <span class="mi">1</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">qkv_lora</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="k">assert</span> <span class="bp">self</span><span class="o">.</span><span class="n">cp_size</span> <span class="o">==</span> <span class="mi">1</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">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">attention_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="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">cp_size</span> <span class="o">></span> <span class="mi">1</span> <span class="ow">and</span> <span class="bp">self</span><span class="o">.</span><span class="n">cp_group</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="n">key</span> <span class="o">=</span> <span class="n">allgather</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">cp_group</span><span class="p">,</span> <span class="n">gather_dim</span><span class="o">=</span><span class="mi">1</span><span class="p">)</span>
|
|
<span class="n">value</span> <span class="o">=</span> <span class="n">allgather</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">cp_group</span><span class="p">,</span> <span class="n">gather_dim</span><span class="o">=</span><span class="mi">1</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">use_fp32_acc</span><span class="o">=</span><span class="kc">False</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="p">(</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">norm_factor</span><span class="p">)</span>
|
|
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">relative_attention</span><span class="p">:</span>
|
|
<span class="n">query_len</span> <span class="o">=</span> <span class="n">shape</span><span class="p">(</span><span class="n">attention_scores</span><span class="p">,</span> <span class="mi">2</span><span class="p">)</span>
|
|
<span class="n">key_len</span> <span class="o">=</span> <span class="n">shape</span><span class="p">(</span><span class="n">attention_scores</span><span class="p">,</span> <span class="mi">3</span><span class="p">)</span>
|
|
<span class="n">bias</span> <span class="o">=</span> <span class="n">compute_relative_bias</span><span class="p">(</span>
|
|
<span class="n">query_len</span><span class="p">,</span>
|
|
<span class="n">key_len</span><span class="p">,</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">num_buckets</span><span class="p">,</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">max_distance</span><span class="p">,</span>
|
|
<span class="kc">True</span><span class="p">,</span> <span class="c1"># bidirectional</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="o">.</span><span class="n">transpose</span><span class="p">(</span><span class="mi">1</span><span class="p">,</span> <span class="mi">0</span><span class="p">),</span>
|
|
<span class="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_group</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">tp_group</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">attention_scores</span> <span class="o">=</span> <span class="n">attention_scores</span> <span class="o">+</span> <span class="n">bias</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">attention_mask</span> <span class="o">=</span> <span class="n">cast</span><span class="p">(</span><span class="n">attention_mask</span><span class="p">,</span> <span class="n">attention_scores</span><span class="o">.</span><span class="n">dtype</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="n">use_fp32_acc</span><span class="o">=</span><span class="kc">False</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">attention_hidden_size</span>
|
|
<span class="p">]))</span>
|
|
|
|
<span class="n">dense_lora_params</span> <span class="o">=</span> <span class="kc">None</span>
|
|
<span class="k">if</span> <span class="n">lora_layer_params</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="n">dense_lora_params</span> <span class="o">=</span> <span class="n">lora_layer_params</span><span class="o">.</span><span class="n">get_runtime_params</span><span class="p">(</span>
|
|
<span class="mi">0</span><span class="p">,</span> <span class="s2">"attn_dense"</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">lora_runtime_params</span><span class="o">=</span><span class="n">dense_lora_params</span><span class="p">)</span>
|
|
|
|
<span class="k">return</span> <span class="n">context</span></div>
|
|
</div>
|
|
|
|
|
|
|
|
<div class="viewcode-block" id="CogVLMAttention">
|
|
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.layers.html#tensorrt_llm.layers.attention.CogVLMAttention">[docs]</a>
|
|
<span class="k">class</span> <span class="nc">CogVLMAttention</span><span class="p">(</span><span class="n">Attention</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="o">*</span><span class="p">,</span>
|
|
<span class="n">local_layer_idx</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">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">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">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">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="n">local_layer_idx</span><span class="o">=</span><span class="n">local_layer_idx</span><span class="p">,</span>
|
|
<span class="n">hidden_size</span><span class="o">=</span><span class="n">hidden_size</span><span class="p">,</span>
|
|
<span class="n">num_attention_heads</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="n">num_kv_heads</span><span class="p">,</span>
|
|
<span class="n">max_position_embeddings</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">dtype</span><span class="p">,</span>
|
|
<span class="n">attention_mask_type</span><span class="o">=</span><span class="n">attention_mask_type</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">position_embedding_type</span><span class="o">=</span><span class="n">position_embedding_type</span><span class="p">,</span>
|
|
<span class="n">rotary_embedding_base</span><span class="o">=</span><span class="n">rotary_embedding_base</span><span class="p">,</span>
|
|
<span class="n">rotary_embedding_scaling</span><span class="o">=</span><span class="n">rotary_embedding_scaling</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">tp_rank</span><span class="o">=</span><span class="n">tp_rank</span><span class="p">,</span>
|
|
<span class="n">quant_mode</span><span class="o">=</span><span class="n">quant_mode</span><span class="p">)</span>
|
|
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">vis_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="n">is_qkv</span><span class="o">=</span><span class="kc">True</span><span class="p">)</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">vis_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="bp">self</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>
|
|
|
|
<div class="viewcode-block" id="CogVLMAttention.forward">
|
|
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.layers.html#tensorrt_llm.layers.attention.CogVLMAttention.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">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">vision_token_mask</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="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="k">assert</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">vision_qkv</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">vis_qkv</span><span class="p">(</span><span class="n">hidden_states</span><span class="p">)</span>
|
|
<span class="n">language_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="n">qkv</span> <span class="o">=</span> <span class="n">where</span><span class="p">(</span><span class="n">vision_token_mask</span><span class="p">,</span> <span class="n">vision_qkv</span><span class="p">,</span> <span class="n">language_qkv</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_cogvlm</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="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">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="n">use_cache</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="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">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="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="c1"># KV cache scales.</span>
|
|
<span class="n">kv_orig_quant_scale</span> <span class="o">=</span> <span class="n">constant</span><span class="p">(</span>
|
|
<span class="n">fp32_array</span><span class="p">([</span><span class="mf">1.0</span><span class="p">])</span>
|
|
<span class="p">)</span> <span class="o">/</span> <span class="bp">self</span><span class="o">.</span><span class="n">kv_cache_scaling_factor</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_cache_scaling_factor</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="c1"># Attention output scales</span>
|
|
<span class="k">assert</span> <span class="p">(</span>
|
|
<span class="ow">not</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">use_fp8_context_fmha</span>
|
|
<span class="p">)</span> <span class="ow">or</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="p">),</span> <span class="s2">"FP8 Context FMHA must be used together with the fp8 quantization workflow."</span>
|
|
|
|
<span class="n">attention_output_orig_quant_scale</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">attention_output_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">attention_output_orig_quant_scale</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</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">qkv</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_attention_window_sizes</span><span class="o">=</span><span class="n">kv_cache_params</span><span class="o">.</span>
|
|
<span class="n">host_max_attention_window_sizes</span><span class="p">,</span>
|
|
<span class="n">host_sink_token_length</span><span class="o">=</span><span class="n">kv_cache_params</span><span class="o">.</span><span class="n">host_sink_token_length</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">layer_idx</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">local_layer_idx</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">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">attention_output_orig_quant_scale</span><span class="o">=</span>
|
|
<span class="n">attention_output_orig_quant_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="kc">None</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_offsets</span><span class="o">=</span><span class="n">kv_cache_params</span><span class="o">.</span><span class="n">kv_cache_block_offsets</span><span class="p">,</span>
|
|
<span class="n">host_kv_cache_block_offsets</span><span class="o">=</span><span class="n">kv_cache_params</span><span class="o">.</span>
|
|
<span class="n">host_kv_cache_block_offsets</span><span class="p">,</span>
|
|
<span class="n">host_kv_cache_pool_pointers</span><span class="o">=</span><span class="n">kv_cache_params</span><span class="o">.</span>
|
|
<span class="n">host_kv_cache_pool_pointers</span><span class="p">,</span>
|
|
<span class="n">host_kv_cache_pool_mapping</span><span class="o">=</span><span class="n">kv_cache_params</span><span class="o">.</span>
|
|
<span class="n">host_kv_cache_pool_mapping</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_kv</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">cross_kv_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="n">use_cache</span><span class="o">=</span><span class="n">use_cache</span><span class="p">,</span>
|
|
<span class="n">spec_decoding_position_offsets</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">spec_decoding_packed_mask</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">mrope_rotary_sin_cos</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">mrope_position_deltas</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">host_runtime_perf_knobs</span><span class="o">=</span><span class="n">attention_params</span><span class="o">.</span>
|
|
<span class="n">host_runtime_perf_knobs</span><span class="p">,</span>
|
|
<span class="n">host_context_progress</span><span class="o">=</span><span class="n">attention_params</span><span class="o">.</span><span class="n">host_context_progress</span><span class="p">,</span>
|
|
<span class="p">)</span>
|
|
|
|
<span class="n">vision_dense</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">vis_dense</span><span class="p">(</span><span class="n">context</span><span class="p">)</span>
|
|
<span class="n">language_dense</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">context</span> <span class="o">=</span> <span class="n">where</span><span class="p">(</span><span class="n">vision_token_mask</span><span class="p">,</span> <span class="n">vision_dense</span><span class="p">,</span> <span class="n">language_dense</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="DeepseekV2Attention">
|
|
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.layers.html#tensorrt_llm.layers.attention.DeepseekV2Attention">[docs]</a>
|
|
<span class="k">class</span> <span class="nc">DeepseekV2Attention</span><span class="p">(</span><span class="n">Attention</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="o">*</span><span class="p">,</span>
|
|
<span class="n">local_layer_idx</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">q_lora_rank</span><span class="p">,</span>
|
|
<span class="n">kv_lora_rank</span><span class="p">,</span>
|
|
<span class="n">qk_nope_head_dim</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">qk_rope_head_dim</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">v_head_dim</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">eps</span><span class="o">=</span><span class="mf">1e-06</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">causal</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">max_position_embeddings</span><span class="o">=</span><span class="mi">1024</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">rotary_embedding_beta_fast</span><span class="o">=</span><span class="mi">32</span><span class="p">,</span>
|
|
<span class="n">rotary_embedding_beta_slow</span><span class="o">=</span><span class="mi">1</span><span class="p">,</span>
|
|
<span class="n">rotary_embedding_mscale</span><span class="o">=</span><span class="mi">1</span><span class="p">,</span>
|
|
<span class="n">rotary_embedding_mscale_all_dim</span><span class="o">=</span><span class="mi">0</span><span class="p">,</span>
|
|
<span class="n">rotary_embedding_origin_max_position</span><span class="o">=</span><span class="mi">4096</span><span class="p">,</span>
|
|
<span class="n">rotary_scaling</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">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="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="n">local_layer_idx</span><span class="o">=</span><span class="n">local_layer_idx</span><span class="p">,</span>
|
|
<span class="n">hidden_size</span><span class="o">=</span><span class="n">hidden_size</span><span class="p">,</span>
|
|
<span class="n">num_attention_heads</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="mi">1</span><span class="p">,</span>
|
|
<span class="n">max_position_embeddings</span><span class="o">=</span><span class="n">max_position_embeddings</span><span class="p">,</span>
|
|
<span class="n">attention_head_size</span><span class="o">=</span><span class="n">kv_lora_rank</span> <span class="o">+</span> <span class="n">qk_rope_head_dim</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">attention_mask_type</span><span class="o">=</span><span class="n">attention_mask_type</span><span class="p">,</span>
|
|
<span class="n">position_embedding_type</span><span class="o">=</span><span class="n">position_embedding_type</span><span class="p">,</span>
|
|
<span class="n">rotary_embedding_base</span><span class="o">=</span><span class="n">rotary_embedding_base</span><span class="p">,</span>
|
|
<span class="n">rotary_embedding_scaling</span><span class="o">=</span><span class="n">rotary_embedding_scaling</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">tp_rank</span><span class="o">=</span><span class="n">tp_rank</span><span class="p">,</span>
|
|
<span class="n">quant_mode</span><span class="o">=</span><span class="n">quant_mode</span><span class="p">,</span>
|
|
<span class="n">bias</span><span class="o">=</span><span class="kc">False</span><span class="p">,</span>
|
|
<span class="n">dense_bias</span><span class="o">=</span><span class="kc">False</span><span class="p">,</span>
|
|
<span class="n">enable_qkv</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">tp_size</span> <span class="o">=</span> <span class="n">tp_size</span>
|
|
|
|
<span class="k">if</span> <span class="n">q_lora_rank</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">q_lora_rank</span> <span class="o">=</span> <span class="n">hidden_size</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">is_deepseek_v2_lite</span> <span class="o">=</span> <span class="kc">True</span>
|
|
<span class="k">else</span><span class="p">:</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">q_lora_rank</span> <span class="o">=</span> <span class="n">q_lora_rank</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">is_deepseek_v2_lite</span> <span class="o">=</span> <span class="kc">False</span>
|
|
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">kv_lora_rank</span> <span class="o">=</span> <span class="n">kv_lora_rank</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">qk_nope_head_dim</span> <span class="o">=</span> <span class="n">qk_nope_head_dim</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">qk_rope_head_dim</span> <span class="o">=</span> <span class="n">qk_rope_head_dim</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">v_head_dim</span> <span class="o">=</span> <span class="n">v_head_dim</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="bp">self</span><span class="o">.</span><span class="n">rotary_scaling</span> <span class="o">=</span> <span class="n">rotary_scaling</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">shard_dim</span> <span class="o">=</span> <span class="mi">1</span>
|
|
|
|
<span class="k">def</span> <span class="nf">yarn_get_mscale</span><span class="p">(</span><span class="n">scale</span><span class="o">=</span><span class="mi">1</span><span class="p">,</span> <span class="n">mscale</span><span class="o">=</span><span class="mi">1</span><span class="p">):</span>
|
|
<span class="k">if</span> <span class="n">scale</span> <span class="o"><=</span> <span class="mi">1</span><span class="p">:</span>
|
|
<span class="k">return</span> <span class="mf">1.0</span>
|
|
<span class="k">return</span> <span class="mf">0.1</span> <span class="o">*</span> <span class="n">mscale</span> <span class="o">*</span> <span class="n">math</span><span class="o">.</span><span class="n">log</span><span class="p">(</span><span class="n">scale</span><span class="p">)</span> <span class="o">+</span> <span class="mf">1.0</span>
|
|
|
|
<span class="k">assert</span> <span class="bp">self</span><span class="o">.</span><span class="n">rotary_scaling</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">rotary_scaling</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="n">mscale_all_dim</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">rotary_scaling</span><span class="o">.</span><span class="n">get</span><span class="p">(</span><span class="s2">"mscale_all_dim"</span><span class="p">,</span> <span class="mi">0</span><span class="p">)</span>
|
|
<span class="n">scaling_factor</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">rotary_scaling</span><span class="p">[</span><span class="s2">"factor"</span><span class="p">]</span>
|
|
<span class="k">if</span> <span class="n">mscale_all_dim</span><span class="p">:</span>
|
|
<span class="n">mscale</span> <span class="o">=</span> <span class="n">yarn_get_mscale</span><span class="p">(</span><span class="n">scaling_factor</span><span class="p">,</span> <span class="n">mscale_all_dim</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="mf">1.0</span> <span class="o">/</span> <span class="p">(</span><span class="n">mscale</span> <span class="o">*</span> <span class="n">mscale</span><span class="p">)</span>
|
|
|
|
<span class="n">embed_positions_for_gpt_attention</span> <span class="o">=</span> <span class="n">RopeEmbeddingUtils</span><span class="o">.</span><span class="n">create_sinusoidal_positions_for_deepseek_attention_plugin</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">qk_rope_head_dim</span><span class="p">,</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">rotary_embedding_base</span><span class="p">,</span> <span class="bp">self</span><span class="o">.</span><span class="n">rotary_scaling</span><span class="p">[</span><span class="s2">"factor"</span><span class="p">],</span>
|
|
<span class="n">rotary_embedding_origin_max_position</span><span class="p">,</span> <span class="n">rotary_embedding_beta_fast</span><span class="p">,</span>
|
|
<span class="n">rotary_embedding_beta_slow</span><span class="p">,</span> <span class="n">rotary_embedding_mscale</span><span class="p">,</span>
|
|
<span class="n">rotary_embedding_mscale_all_dim</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">'embed_positions_for_gpt_attention'</span><span class="p">,</span>
|
|
<span class="n">Parameter</span><span class="p">(</span><span class="n">embed_positions_for_gpt_attention</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">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="bp">self</span><span class="o">.</span><span class="n">is_deepseek_v2_lite</span><span class="p">:</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">fused_a</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">kv_lora_rank</span> <span class="o">+</span> <span class="n">qk_rope_head_dim</span><span class="p">,</span>
|
|
<span class="n">bias</span><span class="o">=</span><span class="bp">self</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="p">)</span>
|
|
<span class="k">else</span><span class="p">:</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">fused_a</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">q_lora_rank</span> <span class="o">+</span> <span class="n">kv_lora_rank</span> <span class="o">+</span> <span class="n">qk_rope_head_dim</span><span class="p">,</span>
|
|
<span class="n">bias</span><span class="o">=</span><span class="bp">self</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="p">)</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">q_a_layernorm</span> <span class="o">=</span> <span class="n">RmsNorm</span><span class="p">(</span><span class="n">q_lora_rank</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">eps</span><span class="o">=</span><span class="n">eps</span><span class="p">)</span>
|
|
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">kv_a_layernorm</span> <span class="o">=</span> <span class="n">RmsNorm</span><span class="p">(</span><span class="n">kv_lora_rank</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">eps</span><span class="o">=</span><span class="n">eps</span><span class="p">)</span>
|
|
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">fused_q_proj</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="bp">self</span><span class="o">.</span><span class="n">num_attention_heads</span> <span class="o">*</span>
|
|
<span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">kv_lora_rank</span> <span class="o">+</span> <span class="bp">self</span><span class="o">.</span><span class="n">qk_rope_head_dim</span><span class="p">),</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">q_lora_rank</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">kv_b_proj</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="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">qk_nope_head_dim</span> <span class="o">*</span> <span class="mi">2</span><span class="p">,</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">kv_lora_rank</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">q_b_proj</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="bp">self</span><span class="o">.</span><span class="n">num_attention_heads</span> <span class="o">*</span>
|
|
<span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">qk_nope_head_dim</span> <span class="o">+</span> <span class="bp">self</span><span class="o">.</span><span class="n">qk_rope_head_dim</span><span class="p">),</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">q_lora_rank</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">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">v_head_dim</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="bp">self</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">set_obj_attrs</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">fused_q_proj</span><span class="p">,</span> <span class="p">{</span>
|
|
<span class="s2">"weight_loader"</span><span class="p">:</span> <span class="bp">self</span><span class="o">.</span><span class="n">weight_loader</span><span class="p">,</span>
|
|
<span class="p">})</span>
|
|
<span class="n">set_obj_attrs</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">q_b_proj</span><span class="p">,</span> <span class="p">{</span>
|
|
<span class="s2">"weight_loader"</span><span class="p">:</span> <span class="bp">self</span><span class="o">.</span><span class="n">weight_loader</span><span class="p">,</span>
|
|
<span class="p">})</span>
|
|
<span class="n">set_obj_attrs</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">kv_b_proj</span><span class="p">,</span> <span class="p">{</span>
|
|
<span class="s2">"weight_loader"</span><span class="p">:</span> <span class="bp">self</span><span class="o">.</span><span class="n">weight_loader</span><span class="p">,</span>
|
|
<span class="p">})</span>
|
|
|
|
<div class="viewcode-block" id="DeepseekV2Attention.weight_loader">
|
|
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.layers.html#tensorrt_llm.layers.attention.DeepseekV2Attention.weight_loader">[docs]</a>
|
|
<span class="k">def</span> <span class="nf">weight_loader</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">mapping</span><span class="p">:</span> <span class="n">Mapping</span><span class="p">,</span> <span class="n">param</span><span class="p">:</span> <span class="n">Parameter</span><span class="p">,</span>
|
|
<span class="n">loaded_weight</span><span class="p">:</span> <span class="n">torch</span><span class="o">.</span><span class="n">Tensor</span><span class="p">):</span>
|
|
<span class="c1"># use_parallel_embedding</span>
|
|
<span class="n">tp_rank</span> <span class="o">=</span> <span class="n">mapping</span><span class="o">.</span><span class="n">tp_rank</span>
|
|
<span class="k">if</span> <span class="bp">self</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="n">sharding_dim</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">sharding_dim</span>
|
|
<span class="n">shard_size</span> <span class="o">=</span> <span class="n">param</span><span class="o">.</span><span class="n">_shape</span><span class="p">[</span><span class="n">sharding_dim</span><span class="p">]</span>
|
|
<span class="n">start_idx</span> <span class="o">=</span> <span class="n">tp_rank</span> <span class="o">*</span> <span class="n">shard_size</span>
|
|
<span class="n">loaded_weight</span> <span class="o">=</span> <span class="n">loaded_weight</span><span class="o">.</span><span class="n">narrow</span><span class="p">(</span><span class="n">sharding_dim</span><span class="p">,</span> <span class="n">start_idx</span><span class="p">,</span>
|
|
<span class="n">shard_size</span><span class="p">)</span>
|
|
<span class="n">param</span><span class="o">.</span><span class="n">value</span> <span class="o">=</span> <span class="n">loaded_weight</span></div>
|
|
|
|
|
|
<div class="viewcode-block" id="DeepseekV2Attention.forward">
|
|
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.layers.html#tensorrt_llm.layers.attention.DeepseekV2Attention.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">use_cache</span><span class="o">=</span><span class="kc">False</span><span class="p">,</span>
|
|
<span class="n">spec_decoding_params</span><span class="o">=</span><span class="kc">None</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="k">assert</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="n">spec_decoding_params</span> <span class="o">=</span> <span class="n">SpecDecodingParams</span><span class="p">(</span>
|
|
<span class="p">)</span> <span class="k">if</span> <span class="n">spec_decoding_params</span> <span class="ow">is</span> <span class="kc">None</span> <span class="k">else</span> <span class="n">spec_decoding_params</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">remove_input_padding</span><span class="p">:</span>
|
|
<span class="k">assert</span> <span class="n">hidden_states</span><span class="o">.</span><span class="n">ndim</span><span class="p">()</span> <span class="o">==</span> <span class="mi">2</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="n">use_cache</span><span class="p">)</span>
|
|
|
|
<span class="k">if</span> <span class="n">use_cache</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">is_deepseek_v2_lite</span><span class="p">:</span>
|
|
<span class="n">compressed_kv</span><span class="p">,</span> <span class="n">k_pe</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">fused_a</span><span class="p">(</span><span class="n">hidden_states</span><span class="p">)</span><span class="o">.</span><span class="n">split</span><span class="p">(</span>
|
|
<span class="p">[</span><span class="bp">self</span><span class="o">.</span><span class="n">kv_lora_rank</span><span class="p">,</span> <span class="bp">self</span><span class="o">.</span><span class="n">qk_rope_head_dim</span><span class="p">],</span> <span class="o">-</span><span class="mi">1</span><span class="p">)</span>
|
|
<span class="n">compressed_kv</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">kv_a_layernorm</span><span class="p">(</span><span class="n">compressed_kv</span><span class="p">)</span>
|
|
<span class="n">input_qkv</span> <span class="o">=</span> <span class="n">concat</span><span class="p">([</span><span class="n">hidden_states</span><span class="p">,</span> <span class="n">compressed_kv</span><span class="p">,</span> <span class="n">k_pe</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">else</span><span class="p">:</span>
|
|
<span class="n">compressed_q</span><span class="p">,</span> <span class="n">compressed_kv</span><span class="p">,</span> <span class="n">k_pe</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">fused_a</span><span class="p">(</span>
|
|
<span class="n">hidden_states</span><span class="p">)</span><span class="o">.</span><span class="n">split</span><span class="p">([</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">q_lora_rank</span><span class="p">,</span> <span class="bp">self</span><span class="o">.</span><span class="n">kv_lora_rank</span><span class="p">,</span> <span class="bp">self</span><span class="o">.</span><span class="n">qk_rope_head_dim</span>
|
|
<span class="p">],</span> <span class="o">-</span><span class="mi">1</span><span class="p">)</span>
|
|
<span class="n">compressed_q</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">q_a_layernorm</span><span class="p">(</span><span class="n">compressed_q</span><span class="p">)</span>
|
|
<span class="n">compressed_kv</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">kv_a_layernorm</span><span class="p">(</span><span class="n">compressed_kv</span><span class="p">)</span>
|
|
<span class="n">input_qkv</span> <span class="o">=</span> <span class="n">concat</span><span class="p">([</span><span class="n">compressed_q</span><span class="p">,</span> <span class="n">compressed_kv</span><span class="p">,</span> <span class="n">k_pe</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">plugin_config</span><span class="o">.</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">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="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="p">],</span> <span class="s1">'Plugin only support masked MHA.'</span>
|
|
|
|
<span class="c1"># KV cache scales.</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">kv_cache_scaling_factor</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="n">kv_orig_quant_scale</span> <span class="o">=</span> <span class="n">constant</span><span class="p">(</span><span class="n">fp32_array</span><span class="p">(</span>
|
|
<span class="p">[</span><span class="mf">1.0</span><span class="p">]))</span> <span class="o">/</span> <span class="bp">self</span><span class="o">.</span><span class="n">kv_cache_scaling_factor</span><span class="o">.</span><span class="n">value</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_cache_scaling_factor</span><span class="o">.</span><span class="n">value</span>
|
|
<span class="k">else</span><span class="p">:</span>
|
|
<span class="n">kv_orig_quant_scale</span> <span class="o">=</span> <span class="kc">None</span>
|
|
<span class="n">kv_quant_orig_scale</span> <span class="o">=</span> <span class="kc">None</span>
|
|
|
|
<span class="c1"># Attention output scales</span>
|
|
<span class="k">assert</span> <span class="p">(</span>
|
|
<span class="ow">not</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">use_fp8_context_fmha</span>
|
|
<span class="p">)</span> <span class="ow">or</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="p">),</span> <span class="s2">"FP8 Context FMHA must be used together with the fp8 quantization workflow."</span>
|
|
|
|
<span class="n">attention_output_orig_quant_scale</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">attention_output_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">attention_output_orig_quant_scale</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span> <span class="k">else</span> <span class="kc">None</span>
|
|
|
|
<span class="n">rotary_cos_sin</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">embed_positions_for_gpt_attention</span><span class="o">.</span><span class="n">value</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">qkv</span><span class="o">=</span><span class="n">input_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_attention_window_sizes</span><span class="o">=</span><span class="n">kv_cache_params</span><span class="o">.</span>
|
|
<span class="n">host_max_attention_window_sizes</span><span class="p">,</span>
|
|
<span class="n">host_sink_token_length</span><span class="o">=</span><span class="n">kv_cache_params</span><span class="o">.</span><span class="n">host_sink_token_length</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">layer_idx</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">local_layer_idx</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="mi">1</span><span class="p">,</span>
|
|
<span class="n">layer_idx_in_cache_pool</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">layer_idx_in_cache_pool</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">kv_lora_rank</span> <span class="o">+</span> <span class="bp">self</span><span class="o">.</span><span class="n">qk_rope_head_dim</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">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">rotary_inv_freq</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">rotary_cos_sin</span><span class="o">=</span><span class="n">rotary_cos_sin</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">attention_output_orig_quant_scale</span><span class="o">=</span>
|
|
<span class="n">attention_output_orig_quant_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">block_sparse_block_size</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">block_sparse_params</span><span class="o">.</span><span class="n">block_size</span><span class="p">,</span>
|
|
<span class="n">block_sparse_homo_head_pattern</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">block_sparse_params</span><span class="o">.</span>
|
|
<span class="n">homo_head_pattern</span><span class="p">,</span>
|
|
<span class="n">block_sparse_num_local_blocks</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">block_sparse_params</span><span class="o">.</span>
|
|
<span class="n">num_local_blocks</span><span class="p">,</span>
|
|
<span class="n">block_sparse_vertical_stride</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">block_sparse_params</span><span class="o">.</span>
|
|
<span class="n">vertical_stride</span><span class="p">,</span>
|
|
<span class="n">alibi_slopes</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="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_offsets</span><span class="o">=</span><span class="n">kv_cache_params</span><span class="o">.</span><span class="n">kv_cache_block_offsets</span>
|
|
<span class="k">if</span> <span class="ow">not</span> <span class="bp">self</span><span class="o">.</span><span class="n">cross_attention</span> <span class="k">else</span>
|
|
<span class="n">kv_cache_params</span><span class="o">.</span><span class="n">cross_kv_cache_block_offsets</span><span class="p">,</span>
|
|
<span class="n">host_kv_cache_block_offsets</span><span class="o">=</span><span class="n">kv_cache_params</span><span class="o">.</span>
|
|
<span class="n">host_kv_cache_block_offsets</span> <span class="k">if</span> <span class="ow">not</span> <span class="bp">self</span><span class="o">.</span><span class="n">cross_attention</span> <span class="k">else</span>
|
|
<span class="n">kv_cache_params</span><span class="o">.</span><span class="n">host_cross_kv_cache_block_offsets</span><span class="p">,</span>
|
|
<span class="n">host_kv_cache_pool_pointers</span><span class="o">=</span><span class="n">kv_cache_params</span><span class="o">.</span>
|
|
<span class="n">host_kv_cache_pool_pointers</span> <span class="k">if</span> <span class="ow">not</span> <span class="bp">self</span><span class="o">.</span><span class="n">cross_attention</span> <span class="k">else</span>
|
|
<span class="n">kv_cache_params</span><span class="o">.</span><span class="n">host_cross_kv_cache_pool_pointers</span><span class="p">,</span>
|
|
<span class="n">host_kv_cache_pool_mapping</span><span class="o">=</span><span class="n">kv_cache_params</span><span class="o">.</span>
|
|
<span class="n">host_kv_cache_pool_mapping</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_kv</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">cross_kv_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="n">use_cache</span><span class="o">=</span><span class="n">use_cache</span><span class="p">,</span>
|
|
<span class="n">spec_decoding_is_generation_length_variable</span><span class="o">=</span><span class="n">spec_decoding_params</span>
|
|
<span class="o">.</span><span class="n">spec_decoding_is_generation_length_variable</span><span class="p">,</span>
|
|
<span class="n">spec_decoding_max_generation_length</span><span class="o">=</span><span class="n">spec_decoding_params</span><span class="o">.</span>
|
|
<span class="n">spec_decoding_max_generation_length</span><span class="p">,</span>
|
|
<span class="n">spec_decoding_generation_lengths</span><span class="o">=</span><span class="n">spec_decoding_params</span><span class="o">.</span>
|
|
<span class="n">spec_decoding_generation_lengths</span><span class="p">,</span>
|
|
<span class="n">spec_decoding_position_offsets</span><span class="o">=</span><span class="n">spec_decoding_params</span><span class="o">.</span>
|
|
<span class="n">spec_decoding_position_offsets</span><span class="p">,</span>
|
|
<span class="n">spec_decoding_packed_mask</span><span class="o">=</span><span class="n">spec_decoding_params</span><span class="o">.</span>
|
|
<span class="n">spec_decoding_packed_mask</span><span class="p">,</span>
|
|
<span class="n">attn_logit_softcapping_scale</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">max_attn_value</span><span class="p">,</span>
|
|
<span class="n">host_runtime_perf_knobs</span><span class="o">=</span><span class="n">attention_params</span><span class="o">.</span>
|
|
<span class="n">host_runtime_perf_knobs</span><span class="p">,</span>
|
|
<span class="n">host_context_progress</span><span class="o">=</span><span class="n">attention_params</span><span class="o">.</span><span class="n">host_context_progress</span><span class="p">,</span>
|
|
<span class="n">is_mla_enabled_flag</span><span class="o">=</span><span class="kc">True</span><span class="p">,</span>
|
|
<span class="n">q_lora_rank</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">q_lora_rank</span><span class="p">,</span>
|
|
<span class="n">kv_lora_rank</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">kv_lora_rank</span><span class="p">,</span>
|
|
<span class="n">qk_nope_head_dim</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">qk_nope_head_dim</span><span class="p">,</span>
|
|
<span class="n">qk_rope_head_dim</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">qk_rope_head_dim</span><span class="p">,</span>
|
|
<span class="n">v_head_dim</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">v_head_dim</span><span class="p">,</span>
|
|
<span class="n">fused_q_proj</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">fused_q_proj</span><span class="o">.</span><span class="n">value</span><span class="p">,</span>
|
|
<span class="n">q_b_proj</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">q_b_proj</span><span class="o">.</span><span class="n">value</span><span class="p">,</span>
|
|
<span class="n">kv_b_proj</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">kv_b_proj</span><span class="o">.</span><span class="n">value</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="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>
|
|
|
|
</pre></div>
|
|
|
|
</div>
|
|
</div>
|
|
<footer>
|
|
|
|
<hr/>
|
|
|
|
<div role="contentinfo">
|
|
<jinja2.runtime.BlockReference object at 0x7f1ac6834e00>
|
|
|
|
<div class="footer">
|
|
<p>
|
|
Copyright © 2024 NVIDIA Corporation
|
|
</p>
|
|
<p>
|
|
<a class="Link" href="https://www.nvidia.com/en-us/about-nvidia/privacy-policy/" target="_blank" rel="noopener"
|
|
data-cms-ai="0">Privacy Policy</a> |
|
|
<a class="Link" href="https://www.nvidia.com/en-us/about-nvidia/privacy-center/" target="_blank" rel="noopener"
|
|
data-cms-ai="0">Manage My Privacy</a> |
|
|
<a class="Link" href="https://www.nvidia.com/en-us/preferences/start/" target="_blank" rel="noopener"
|
|
data-cms-ai="0">Do Not Sell or Share My Data</a> |
|
|
<a class="Link" href="https://www.nvidia.com/en-us/about-nvidia/terms-of-service/" target="_blank"
|
|
rel="noopener" data-cms-ai="0">Terms of Service</a> |
|
|
<a class="Link" href="https://www.nvidia.com/en-us/about-nvidia/accessibility/" target="_blank" rel="noopener"
|
|
data-cms-ai="0">Accessibility</a> |
|
|
<a class="Link" href="https://www.nvidia.com/en-us/about-nvidia/company-policies/" target="_blank"
|
|
rel="noopener" data-cms-ai="0">Corporate Policies</a> |
|
|
<a class="Link" href="https://www.nvidia.com/en-us/product-security/" target="_blank" rel="noopener"
|
|
data-cms-ai="0">Product Security</a> |
|
|
<a class="Link" href="https://www.nvidia.com/en-us/contact/" target="_blank" rel="noopener"
|
|
data-cms-ai="0">Contact</a>
|
|
</p>
|
|
</div>
|
|
|
|
|
|
</div>
|
|
|
|
|
|
|
|
</footer>
|
|
</div>
|
|
</div>
|
|
</section>
|
|
</div>
|
|
<script>
|
|
jQuery(function () {
|
|
SphinxRtdTheme.Navigation.enable(true);
|
|
});
|
|
</script>
|
|
|
|
</body>
|
|
</html> |