kanshan/TensorRT-LLMs
1
0
Fork 0
mirror of https://github.com/NVIDIA/TensorRT-LLM.git synced 2026-01-14 06:27:45 +08:00
Code Issues Actions 1 Packages Projects Releases Wiki Activity
9ed1e1fa7d
TensorRT-LLMs/_modules/tensorrt_llm/models/modeling_utils.html
Kaiyu Xie 9ed1e1fa7d Update GitHub pages in root to v0.21.0rc1
2025-06-11 02:46:37 +00:00

2680 lines
350 KiB
HTML
Raw Blame History

<!DOCTYPE html>
<html lang="en" data-content_root="../../../" >
<head>
<meta charset="utf-8" />
<meta name="viewport" content="width=device-width, initial-scale=1.0" />
<title>tensorrt_llm.models.modeling_utils &#8212; TensorRT-LLM</title>
<script data-cfasync="false">
document.documentElement.dataset.mode = localStorage.getItem("mode") || "";
document.documentElement.dataset.theme = localStorage.getItem("theme") || "";
</script>
<!--
this give us a css class that will be invisible only if js is disabled
-->
<noscript>
<style>
.pst-js-only { display: none !important; }
</style>
</noscript>
<!-- Loaded before other Sphinx assets -->
<link href="../../../_static/styles/theme.css?digest=8878045cc6db502f8baf" rel="stylesheet" />
<link href="../../../_static/styles/pydata-sphinx-theme.css?digest=8878045cc6db502f8baf" rel="stylesheet" />
<link rel="stylesheet" type="text/css" href="../../../_static/pygments.css?v=8f2a1f02" />
<link rel="stylesheet" type="text/css" href="../../../_static/styles/nvidia-sphinx-theme.css?v=df3ac72c" />
<link rel="stylesheet" type="text/css" href="../../../_static/copybutton.css?v=76b2166b" />
<link rel="stylesheet" type="text/css" href="../../../_static/autodoc_pydantic.css" />
<!-- So that users can add custom icons -->
<script src="../../../_static/scripts/fontawesome.js?digest=8878045cc6db502f8baf"></script>
<!-- Pre-loaded scripts that we'll load fully later -->
<link rel="preload" as="script" href="../../../_static/scripts/bootstrap.js?digest=8878045cc6db502f8baf" />
<link rel="preload" as="script" href="../../../_static/scripts/pydata-sphinx-theme.js?digest=8878045cc6db502f8baf" />
<script src="../../../_static/documentation_options.js?v=5929fcd5"></script>
<script src="../../../_static/doctools.js?v=9a2dae69"></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>DOCUMENTATION_OPTIONS.pagename = '_modules/tensorrt_llm/models/modeling_utils';</script>
<script>
DOCUMENTATION_OPTIONS.theme_version = '0.16.1';
DOCUMENTATION_OPTIONS.theme_switcher_json_url = './_static/switcher.json';
DOCUMENTATION_OPTIONS.theme_switcher_version_match = '0.21.0rc1';
DOCUMENTATION_OPTIONS.show_version_warning_banner =
false;
</script>
<link rel="icon" href="../../../_static/favicon.png"/>
<link rel="index" title="Index" href="../../../genindex.html" />
<link rel="search" title="Search" href="../../../search.html" />
<meta name="viewport" content="width=device-width, initial-scale=1"/>
<meta name="docsearch:language" content="en"/>
<meta name="docsearch:version" content="0.21.0rc1" />
</head>
<body data-bs-spy="scroll" data-bs-target=".bd-toc-nav" data-offset="180" data-bs-root-margin="0px 0px -60%" data-default-mode="">
<div id="pst-skip-link" class="skip-link d-print-none"><a href="#main-content">Skip to main content</a></div>
<div id="pst-scroll-pixel-helper"></div>
<button type="button" class="btn rounded-pill" id="pst-back-to-top">
<i class="fa-solid fa-arrow-up"></i>Back to top</button>
<dialog id="pst-search-dialog">
<form class="bd-search d-flex align-items-center"
action="../../../search.html"
method="get">
<i class="fa-solid fa-magnifying-glass"></i>
<input type="search"
class="form-control"
name="q"
placeholder="Search the docs ..."
aria-label="Search the docs ..."
autocomplete="off"
autocorrect="off"
autocapitalize="off"
spellcheck="false"/>
<span class="search-button__kbd-shortcut"><kbd class="kbd-shortcut__modifier">Ctrl</kbd>+<kbd>K</kbd></span>
</form>
</dialog>
<div class="pst-async-banner-revealer d-none">
<aside id="bd-header-version-warning" class="d-none d-print-none" aria-label="Version warning"></aside>
</div>
<header class="bd-header navbar navbar-expand-lg bd-navbar d-print-none">
<div class="bd-header__inner bd-page-width">
<button class="pst-navbar-icon sidebar-toggle primary-toggle" aria-label="Site navigation">
<span class="fa-solid fa-bars"></span>
</button>
<div class="col-lg-3 navbar-header-items__start">
<div class="navbar-item">
<a class="navbar-brand logo" href="../../../index.html">
<img src="../../../_static/nvidia-logo-horiz-rgb-blk-for-screen.svg" class="logo__image only-light" alt="TensorRT-LLM - Home"/>
<img src="../../../_static/nvidia-logo-horiz-rgb-wht-for-screen.svg" class="logo__image only-dark pst-js-only" alt="TensorRT-LLM - Home"/>
<p class="title logo__title">TensorRT-LLM</p>
</a></div>
</div>
<div class="col-lg-9 navbar-header-items">
<div class="me-auto navbar-header-items__center">
<div class="navbar-item">
<div class="version-switcher__container dropdown pst-js-only">
<button id="pst-version-switcher-button-2"
type="button"
class="version-switcher__button btn btn-sm dropdown-toggle"
data-bs-toggle="dropdown"
aria-haspopup="listbox"
aria-controls="pst-version-switcher-list-2"
aria-label="Version switcher list"
>
Choose version <!-- this text may get changed later by javascript -->
<span class="caret"></span>
</button>
<div id="pst-version-switcher-list-2"
class="version-switcher__menu dropdown-menu list-group-flush py-0"
role="listbox" aria-labelledby="pst-version-switcher-button-2">
<!-- dropdown will be populated by javascript on page load -->
</div>
</div></div>
</div>
<div class="navbar-header-items__end">
<div class="navbar-item navbar-persistent--container">
<button class="btn search-button-field search-button__button pst-js-only" title="Search" aria-label="Search" data-bs-placement="bottom" data-bs-toggle="tooltip">
<i class="fa-solid fa-magnifying-glass"></i>
<span class="search-button__default-text">Search</span>
<span class="search-button__kbd-shortcut"><kbd class="kbd-shortcut__modifier">Ctrl</kbd>+<kbd class="kbd-shortcut__modifier">K</kbd></span>
</button>
</div>
<div class="navbar-item">
<button class="btn btn-sm nav-link pst-navbar-icon theme-switch-button pst-js-only" aria-label="Color mode" data-bs-title="Color mode" data-bs-placement="bottom" data-bs-toggle="tooltip">
<i class="theme-switch fa-solid fa-sun fa-lg" data-mode="light" title="Light"></i>
<i class="theme-switch fa-solid fa-moon fa-lg" data-mode="dark" title="Dark"></i>
<i class="theme-switch fa-solid fa-circle-half-stroke fa-lg" data-mode="auto" title="System Settings"></i>
</button></div>
</div>
</div>
<div class="navbar-persistent--mobile">
<button class="btn search-button-field search-button__button pst-js-only" title="Search" aria-label="Search" data-bs-placement="bottom" data-bs-toggle="tooltip">
<i class="fa-solid fa-magnifying-glass"></i>
<span class="search-button__default-text">Search</span>
<span class="search-button__kbd-shortcut"><kbd class="kbd-shortcut__modifier">Ctrl</kbd>+<kbd class="kbd-shortcut__modifier">K</kbd></span>
</button>
</div>
</div>
</header>
<div class="bd-container">
<div class="bd-container__inner bd-page-width">
<dialog id="pst-primary-sidebar-modal"></dialog>
<div id="pst-primary-sidebar" class="bd-sidebar-primary bd-sidebar">
<a class="navbar-brand logo" href="../../../index.html">
<img src="../../../_static/nvidia-logo-horiz-rgb-blk-for-screen.svg" class="logo__image only-light" alt="TensorRT-LLM - Home"/>
<img src="../../../_static/nvidia-logo-horiz-rgb-wht-for-screen.svg" class="logo__image only-dark pst-js-only" alt="TensorRT-LLM - Home"/>
<p class="title logo__title">TensorRT-LLM</p>
</a>
<div class="sidebar-header-items sidebar-primary__section">
<div class="sidebar-header-items__center">
<div class="navbar-item">
<div class="version-switcher__container dropdown pst-js-only">
<button id="pst-version-switcher-button-3"
type="button"
class="version-switcher__button btn btn-sm dropdown-toggle"
data-bs-toggle="dropdown"
aria-haspopup="listbox"
aria-controls="pst-version-switcher-list-3"
aria-label="Version switcher list"
>
Choose version <!-- this text may get changed later by javascript -->
<span class="caret"></span>
</button>
<div id="pst-version-switcher-list-3"
class="version-switcher__menu dropdown-menu list-group-flush py-0"
role="listbox" aria-labelledby="pst-version-switcher-button-3">
<!-- dropdown will be populated by javascript on page load -->
</div>
</div></div>
</div>
<div class="sidebar-header-items__end">
<div class="navbar-item">
<button class="btn btn-sm nav-link pst-navbar-icon theme-switch-button pst-js-only" aria-label="Color mode" data-bs-title="Color mode" data-bs-placement="bottom" data-bs-toggle="tooltip">
<i class="theme-switch fa-solid fa-sun fa-lg" data-mode="light" title="Light"></i>
<i class="theme-switch fa-solid fa-moon fa-lg" data-mode="dark" title="Dark"></i>
<i class="theme-switch fa-solid fa-circle-half-stroke fa-lg" data-mode="auto" title="System Settings"></i>
</button></div>
</div>
</div>
<div class="sidebar-primary-items__start sidebar-primary__section">
<div class="sidebar-primary-item">
<nav class="bd-docs-nav bd-links"
aria-label="Table of Contents">
<p class="bd-links__title" role="heading" aria-level="1">Table of Contents</p>
<div class="bd-toc-item navbar-nav"><p aria-level="2" class="caption" role="heading"><span class="caption-text">Getting Started</span></p>
<ul class="nav bd-sidenav">
<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="../../../torch.html">PyTorch Backend</a></li>
<li class="toctree-l1"><a class="reference internal" href="../../../release-notes.html">Release Notes</a></li>
</ul>
<p aria-level="2" class="caption" role="heading"><span class="caption-text">Installation</span></p>
<ul class="nav bd-sidenav">
<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/grace-hopper.html">Installing on Grace Hopper</a></li>
</ul>
<p aria-level="2" class="caption" role="heading"><span class="caption-text">LLM API</span></p>
<ul class="nav bd-sidenav">
<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 aria-level="2" class="caption" role="heading"><span class="caption-text">Examples</span></p>
<ul class="nav bd-sidenav">
<li class="toctree-l1 has-children"><a class="reference internal" href="../../../examples/index.html">LLM Examples Introduction</a><details><summary><span class="toctree-toggle" role="presentation"><i class="fa-solid fa-chevron-down"></i></span></summary><ul>
<li class="toctree-l2"><a class="reference internal" href="../../../examples/llm_medusa_decoding.html">Generate Text Using Medusa Decoding</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../../examples/llm_multilora.html">Generate text with multiple LoRA adapters</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../../examples/llm_eagle_decoding.html">Generate Text Using Eagle Decoding</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../../examples/llm_inference_async.html">Generate Text Asynchronously</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../../examples/llm_inference_distributed.html">Distributed LLM Generation</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../../examples/llm_logits_processor.html">Control generated text using logits processor</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../../examples/llm_eagle2_decoding.html">Generate Text Using Eagle2 Decoding</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../../examples/llm_inference_kv_events.html">Get KV Cache Events</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../../examples/llm_lookahead_decoding.html">Generate Text Using Lookahead Decoding</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../../examples/llm_quantization.html">Generation with Quantization</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../../examples/llm_inference_async_streaming.html">Generate Text in Streaming</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../../examples/llm_guided_decoding.html">Generate text with guided decoding</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../../examples/llm_inference.html">Generate text</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../../examples/llm_inference_customize.html">Generate text with customization</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../../examples/llm_auto_parallel.html">Automatic Parallelism with LLM</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../../examples/llm_mgmn_llm_distributed.html">Llm Mgmn Llm Distributed</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../../examples/llm_mgmn_trtllm_bench.html">Llm Mgmn Trtllm Bench</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../../examples/llm_mgmn_trtllm_serve.html">Llm Mgmn Trtllm Serve</a></li>
</ul>
</details></li>
<li class="toctree-l1"><a class="reference internal" href="../../../examples/customization.html">LLM Common Customizations</a></li>
<li class="toctree-l1 has-children"><a class="reference internal" href="../../../examples/llm_api_examples.html">LLM Examples</a><details><summary><span class="toctree-toggle" role="presentation"><i class="fa-solid fa-chevron-down"></i></span></summary><ul>
<li class="toctree-l2"><a class="reference internal" href="../../../examples/llm_medusa_decoding.html">Generate Text Using Medusa Decoding</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../../examples/llm_multilora.html">Generate text with multiple LoRA adapters</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../../examples/llm_eagle_decoding.html">Generate Text Using Eagle Decoding</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../../examples/llm_inference_async.html">Generate Text Asynchronously</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../../examples/llm_inference_distributed.html">Distributed LLM Generation</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../../examples/llm_logits_processor.html">Control generated text using logits processor</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../../examples/llm_eagle2_decoding.html">Generate Text Using Eagle2 Decoding</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../../examples/llm_inference_kv_events.html">Get KV Cache Events</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../../examples/llm_lookahead_decoding.html">Generate Text Using Lookahead Decoding</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../../examples/llm_quantization.html">Generation with Quantization</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../../examples/llm_inference_async_streaming.html">Generate Text in Streaming</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../../examples/llm_guided_decoding.html">Generate text with guided decoding</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../../examples/llm_inference.html">Generate text</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../../examples/llm_inference_customize.html">Generate text with customization</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../../examples/llm_auto_parallel.html">Automatic Parallelism with LLM</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../../examples/llm_mgmn_llm_distributed.html">Llm Mgmn Llm Distributed</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../../examples/llm_mgmn_trtllm_bench.html">Llm Mgmn Trtllm Bench</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../../examples/llm_mgmn_trtllm_serve.html">Llm Mgmn Trtllm Serve</a></li>
</ul>
</details></li>
<li class="toctree-l1 has-children"><a class="reference internal" href="../../../examples/trtllm_serve_examples.html">Online Serving Examples</a><details><summary><span class="toctree-toggle" role="presentation"><i class="fa-solid fa-chevron-down"></i></span></summary><ul>
<li class="toctree-l2"><a class="reference internal" href="../../../examples/curl_chat_client.html">Curl Chat Client</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../../examples/curl_chat_client_for_multimodal.html">Curl Chat Client For Multimodal</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../../examples/curl_completion_client.html">Curl Completion Client</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../../examples/deepseek_r1_reasoning_parser.html">Deepseek R1 Reasoning Parser</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../../examples/genai_perf_client.html">Genai Perf Client</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../../examples/genai_perf_client_for_multimodal.html">Genai Perf Client For Multimodal</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../../examples/openai_chat_client.html">OpenAI Chat Client</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../../examples/openai_chat_client_for_multimodal.html">OpenAI Chat Client</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../../examples/openai_completion_client.html">OpenAI Completion Client</a></li>
</ul>
</details></li>
</ul>
<p aria-level="2" class="caption" role="heading"><span class="caption-text">Model Definition API</span></p>
<ul class="nav bd-sidenav">
<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 aria-level="2" class="caption" role="heading"><span class="caption-text">C++ API</span></p>
<ul class="nav bd-sidenav">
<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 aria-level="2" class="caption" role="heading"><span class="caption-text">Command-Line Reference</span></p>
<ul class="nav bd-sidenav">
<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 aria-level="2" class="caption" role="heading"><span class="caption-text">Architecture</span></p>
<ul class="nav bd-sidenav">
<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/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 aria-level="2" class="caption" role="heading"><span class="caption-text">Advanced</span></p>
<ul class="nav bd-sidenav">
<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/lora.html">Run gpt-2b + LoRA using Executor / 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-management.html">KV Cache Management: Pools, Blocks, and Events</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 aria-level="2" class="caption" role="heading"><span class="caption-text">Performance</span></p>
<ul class="nav bd-sidenav">
<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 has-children"><a class="reference internal" href="../../../performance/performance-tuning-guide/index.html">Performance Tuning Guide</a><details><summary><span class="toctree-toggle" role="presentation"><i class="fa-solid fa-chevron-down"></i></span></summary><ul>
<li class="toctree-l2"><a class="reference internal" href="../../../performance/performance-tuning-guide/benchmarking-default-performance.html">Benchmarking Default Performance</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../../performance/performance-tuning-guide/useful-build-time-flags.html">Useful Build-Time Flags</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../../performance/performance-tuning-guide/tuning-max-batch-size-and-max-num-tokens.html">Tuning Max Batch Size and Max Num Tokens</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../../performance/performance-tuning-guide/deciding-model-sharding-strategy.html">Deciding Model Sharding Strategy</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../../performance/performance-tuning-guide/fp8-quantization.html">FP8 Quantization</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../../performance/performance-tuning-guide/useful-runtime-flags.html">Useful Runtime Options</a></li>
</ul>
</details></li>
<li class="toctree-l1"><a class="reference internal" href="../../../performance/perf-analysis.html">Performance Analysis</a></li>
</ul>
<p aria-level="2" class="caption" role="heading"><span class="caption-text">Reference</span></p>
<ul class="nav bd-sidenav">
<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 aria-level="2" class="caption" role="heading"><span class="caption-text">Blogs</span></p>
<ul class="nav bd-sidenav">
<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>
<li class="toctree-l1"><a class="reference internal" href="../../../blogs/tech_blog/blog1_Pushing_Latency_Boundaries_Optimizing_DeepSeek-R1_Performance_on_NVIDIA_B200_GPUs.html">Pushing Latency Boundaries: Optimizing DeepSeek-R1 Performance on NVIDIA B200 GPUs</a></li>
<li class="toctree-l1"><a class="reference internal" href="../../../blogs/tech_blog/blog2_DeepSeek_R1_MTP_Implementation_and_Optimization.html">DeepSeek R1 MTP Implementation and Optimization</a></li>
</ul>
</div>
</nav></div>
</div>
<div class="sidebar-primary-items__end sidebar-primary__section">
</div>
</div>
<main id="main-content" class="bd-main" role="main">
<div class="bd-content">
<div class="bd-article-container">
<div class="bd-header-article d-print-none">
<div class="header-article-items header-article__inner">
<div class="header-article-items__start">
<div class="header-article-item">
<nav aria-label="Breadcrumb" class="d-print-none">
<ul class="bd-breadcrumbs">
<li class="breadcrumb-item breadcrumb-home">
<a href="../../../index.html" class="nav-link" aria-label="Home">
<i class="fa-solid fa-home"></i>
</a>
</li>
<li class="breadcrumb-item"><a href="../../index.html" class="nav-link">Module code</a></li>
<li class="breadcrumb-item active" aria-current="page"><span class="ellipsis">tensorrt_llm.models.modeling_utils</span></li>
</ul>
</nav>
</div>
</div>
</div>
</div>
<div id="searchbox"></div>
<article class="bd-article">
<h1>Source code for tensorrt_llm.models.modeling_utils</h1><div class="highlight"><pre>
<span></span><span class="kn">import</span><span class="w"> </span><span class="nn">argparse</span>
<span class="kn">import</span><span class="w"> </span><span class="nn">copy</span>
<span class="kn">import</span><span class="w"> </span><span class="nn">dataclasses</span>
<span class="kn">import</span><span class="w"> </span><span class="nn">fnmatch</span>
<span class="kn">import</span><span class="w"> </span><span class="nn">json</span>
<span class="kn">import</span><span class="w"> </span><span class="nn">os</span>
<span class="kn">import</span><span class="w"> </span><span class="nn">re</span>
<span class="kn">from</span><span class="w"> </span><span class="nn">enum</span><span class="w"> </span><span class="kn">import</span> <span class="n">IntFlag</span><span class="p">,</span> <span class="n">auto</span>
<span class="kn">from</span><span class="w"> </span><span class="nn">functools</span><span class="w"> </span><span class="kn">import</span> <span class="n">cached_property</span>
<span class="kn">from</span><span class="w"> </span><span class="nn">pathlib</span><span class="w"> </span><span class="kn">import</span> <span class="n">Path</span>
<span class="kn">from</span><span class="w"> </span><span class="nn">typing</span><span class="w"> </span><span class="kn">import</span> <span class="p">(</span><span class="n">TYPE_CHECKING</span><span class="p">,</span> <span class="n">Callable</span><span class="p">,</span> <span class="n">Dict</span><span class="p">,</span> <span class="n">Generator</span><span class="p">,</span> <span class="n">List</span><span class="p">,</span> <span class="n">Optional</span><span class="p">,</span>
<span class="n">Union</span><span class="p">)</span>
<span class="kn">import</span><span class="w"> </span><span class="nn">numpy</span><span class="w"> </span><span class="k">as</span><span class="w"> </span><span class="nn">np</span>
<span class="kn">import</span><span class="w"> </span><span class="nn">safetensors</span>
<span class="kn">import</span><span class="w"> </span><span class="nn">torch</span>
<span class="kn">from</span><span class="w"> </span><span class="nn">.._common</span><span class="w"> </span><span class="kn">import</span> <span class="n">default_net</span>
<span class="kn">from</span><span class="w"> </span><span class="nn">.._utils</span><span class="w"> </span><span class="kn">import</span> <span class="p">(</span><span class="n">QuantModeWrapper</span><span class="p">,</span> <span class="n">get_init_params</span><span class="p">,</span> <span class="n">numpy_to_torch</span><span class="p">,</span>
<span class="n">release_gc</span><span class="p">,</span> <span class="n">str_dtype_to_torch</span><span class="p">,</span> <span class="n">str_dtype_to_trt</span><span class="p">,</span>
<span class="n">trt_dtype_to_torch</span><span class="p">)</span>
<span class="kn">from</span><span class="w"> </span><span class="nn">..bindings</span><span class="w"> </span><span class="kn">import</span> <span class="n">KVCacheType</span>
<span class="kn">from</span><span class="w"> </span><span class="nn">..bindings.executor</span><span class="w"> </span><span class="kn">import</span> <span class="n">RuntimeDefaults</span>
<span class="kn">from</span><span class="w"> </span><span class="nn">..functional</span><span class="w"> </span><span class="kn">import</span> <span class="p">(</span><span class="n">PositionEmbeddingType</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">constant</span><span class="p">,</span>
<span class="n">cp_split_plugin</span><span class="p">,</span> <span class="n">gather_last_token_logits</span><span class="p">,</span>
<span class="n">index_select</span><span class="p">,</span> <span class="n">tanh</span><span class="p">,</span> <span class="n">view</span><span class="p">)</span>
<span class="kn">from</span><span class="w"> </span><span class="nn">..layers</span><span class="w"> </span><span class="kn">import</span> <span class="p">(</span><span class="n">MLP</span><span class="p">,</span> <span class="n">AttentionParams</span><span class="p">,</span> <span class="n">Embedding</span><span class="p">,</span> <span class="n">FusedGatedMLP</span><span class="p">,</span>
<span class="n">FusedRgLru</span><span class="p">,</span> <span class="n">GatedMLP</span><span class="p">,</span> <span class="n">KeyValueCacheParams</span><span class="p">,</span> <span class="n">LoraParams</span><span class="p">,</span>
<span class="n">PromptTuningEmbedding</span><span class="p">,</span> <span class="n">RgLru</span><span class="p">)</span>
<span class="kn">from</span><span class="w"> </span><span class="nn">..layers.attention</span><span class="w"> </span><span class="kn">import</span> <span class="n">Attention</span><span class="p">,</span> <span class="n">BertAttention</span>
<span class="kn">from</span><span class="w"> </span><span class="nn">..layers.linear</span><span class="w"> </span><span class="kn">import</span> <span class="n">ColumnLinear</span><span class="p">,</span> <span class="n">Linear</span><span class="p">,</span> <span class="n">RowLinear</span>
<span class="kn">from</span><span class="w"> </span><span class="nn">..layers.lora</span><span class="w"> </span><span class="kn">import</span> <span class="n">Dora</span><span class="p">,</span> <span class="n">Lora</span>
<span class="kn">from</span><span class="w"> </span><span class="nn">..layers.moe</span><span class="w"> </span><span class="kn">import</span> <span class="n">MOE</span><span class="p">,</span> <span class="n">MoeOOTB</span>
<span class="kn">from</span><span class="w"> </span><span class="nn">..logger</span><span class="w"> </span><span class="kn">import</span> <span class="n">logger</span>
<span class="kn">from</span><span class="w"> </span><span class="nn">..mapping</span><span class="w"> </span><span class="kn">import</span> <span class="n">Mapping</span>
<span class="kn">from</span><span class="w"> </span><span class="nn">..module</span><span class="w"> </span><span class="kn">import</span> <span class="n">Module</span><span class="p">,</span> <span class="n">ModuleList</span>
<span class="kn">from</span><span class="w"> </span><span class="nn">..parameter</span><span class="w"> </span><span class="kn">import</span> <span class="n">Parameter</span>
<span class="kn">from</span><span class="w"> </span><span class="nn">..plugin</span><span class="w"> </span><span class="kn">import</span> <span class="n">init_all_reduce_helper</span>
<span class="kn">from</span><span class="w"> </span><span class="nn">..quantization</span><span class="w"> </span><span class="kn">import</span> <span class="n">QuantMode</span>
<span class="kn">from</span><span class="w"> </span><span class="nn">..quantization.functional</span><span class="w"> </span><span class="kn">import</span> <span class="n">preprocess_weights_for_mixed_gemm</span>
<span class="kn">from</span><span class="w"> </span><span class="nn">..quantization.layers</span><span class="w"> </span><span class="kn">import</span> <span class="p">(</span><span class="n">FP8Linear</span><span class="p">,</span> <span class="n">Fp8RowwiseFusedGatedMLP</span><span class="p">,</span>
<span class="n">Fp8RowwiseGatedMLP</span><span class="p">,</span>
<span class="n">WeightOnlyGroupwiseQuantLinear</span><span class="p">,</span>
<span class="n">WeightOnlyGroupwiseQuantRowLinear</span><span class="p">,</span>
<span class="n">WeightOnlyQuantLinear</span><span class="p">,</span>
<span class="n">WeightOnlyQuantRowLinear</span><span class="p">)</span>
<span class="kn">from</span><span class="w"> </span><span class="nn">..quantization.mode</span><span class="w"> </span><span class="kn">import</span> <span class="p">(</span><span class="n">KV_CACHE_QUANT_ALGO_LIST</span><span class="p">,</span> <span class="n">QUANT_ALGO_LIST</span><span class="p">,</span>
<span class="n">W8A8_SQ_PLUGIN_LIST</span><span class="p">,</span> <span class="n">QuantAlgo</span><span class="p">)</span>
<span class="kn">from</span><span class="w"> </span><span class="nn">..quantization.utils</span><span class="w"> </span><span class="kn">import</span> <span class="n">fp4_utils</span>
<span class="kn">from</span><span class="w"> </span><span class="nn">..top_model_mixin</span><span class="w"> </span><span class="kn">import</span> <span class="n">TopModelMixin</span>
<span class="kn">from</span><span class="w"> </span><span class="nn">.convert_utils</span><span class="w"> </span><span class="kn">import</span> <span class="n">weight_only_quantize_dict</span>
<span class="kn">from</span><span class="w"> </span><span class="nn">.generation_mixin</span><span class="w"> </span><span class="kn">import</span> <span class="n">GenerationMixin</span>
<span class="nd">@dataclasses</span><span class="o">.</span><span class="n">dataclass</span><span class="p">(</span><span class="n">kw_only</span><span class="o">=</span><span class="kc">True</span><span class="p">,</span> <span class="n">frozen</span><span class="o">=</span><span class="kc">True</span><span class="p">)</span>
<span class="k">class</span><span class="w"> </span><span class="nc">Gemma2ConfigGroup</span><span class="p">:</span>
<span class="n">query_pre_attn_scalar</span><span class="p">:</span> <span class="nb">int</span>
<span class="n">final_logit_softcapping</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="nb">float</span><span class="p">]</span>
<span class="n">attn_logit_softcapping</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="nb">float</span><span class="p">]</span>
<span class="nd">@classmethod</span>
<span class="k">def</span><span class="w"> </span><span class="nf">keys</span><span class="p">(</span><span class="bp">cls</span><span class="p">):</span>
<span class="k">return</span> <span class="p">{</span><span class="n">f</span><span class="o">.</span><span class="n">name</span> <span class="k">for</span> <span class="n">f</span> <span class="ow">in</span> <span class="n">dataclasses</span><span class="o">.</span><span class="n">fields</span><span class="p">(</span><span class="bp">cls</span><span class="p">)}</span>
<span class="nd">@dataclasses</span><span class="o">.</span><span class="n">dataclass</span><span class="p">(</span><span class="n">kw_only</span><span class="o">=</span><span class="kc">True</span><span class="p">,</span> <span class="n">frozen</span><span class="o">=</span><span class="kc">True</span><span class="p">)</span>
<span class="k">class</span><span class="w"> </span><span class="nc">Gemma3ConfigGroup</span><span class="p">:</span>
<span class="n">query_pre_attn_scalar</span><span class="p">:</span> <span class="nb">float</span>
<span class="n">final_logit_softcapping</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="nb">float</span><span class="p">]</span>
<span class="n">sliding_window_pattern</span><span class="p">:</span> <span class="nb">int</span>
<span class="n">rope_local_base_freq</span><span class="p">:</span> <span class="nb">int</span>
<span class="n">sliding_window</span><span class="p">:</span> <span class="nb">int</span>
<span class="nd">@classmethod</span>
<span class="k">def</span><span class="w"> </span><span class="nf">keys</span><span class="p">(</span><span class="bp">cls</span><span class="p">):</span>
<span class="k">return</span> <span class="p">{</span><span class="n">f</span><span class="o">.</span><span class="n">name</span> <span class="k">for</span> <span class="n">f</span> <span class="ow">in</span> <span class="n">dataclasses</span><span class="o">.</span><span class="n">fields</span><span class="p">(</span><span class="bp">cls</span><span class="p">)}</span>
<span class="k">if</span> <span class="n">TYPE_CHECKING</span><span class="p">:</span>
<span class="kn">from</span><span class="w"> </span><span class="nn">typing</span><span class="w"> </span><span class="kn">import</span> <span class="n">Type</span><span class="p">,</span> <span class="n">TypeVar</span>
<span class="kn">from</span><span class="w"> </span><span class="nn">typing_extensions</span><span class="w"> </span><span class="kn">import</span> <span class="n">Self</span>
<span class="n">ConfigGroups</span> <span class="o">=</span> <span class="n">Union</span><span class="p">[</span><span class="n">Gemma2ConfigGroup</span><span class="p">,</span> <span class="n">Gemma3ConfigGroup</span><span class="p">]</span>
<span class="w"> </span><span class="sd">&quot;&quot;&quot;Groupings of config where, if one of said properties exists, we assume all of the properties exist (even if they are `None`)&quot;&quot;&quot;</span>
<span class="n">CG</span> <span class="o">=</span> <span class="n">TypeVar</span><span class="p">(</span><span class="s2">&quot;CG&quot;</span><span class="p">,</span> <span class="n">bound</span><span class="o">=</span><span class="n">ConfigGroups</span><span class="p">)</span>
<span class="n">RuntimeDefaultsIn</span> <span class="o">=</span> <span class="n">Optional</span><span class="p">[</span><span class="n">Union</span><span class="p">[</span><span class="n">RuntimeDefaults</span><span class="p">,</span> <span class="nb">dict</span><span class="p">]]</span>
<div class="viewcode-block" id="SpeculativeDecodingMode">
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.models.html#tensorrt_llm.llmapi.SpeculativeDecodingMode">[docs]</a>
<span class="k">class</span><span class="w"> </span><span class="nc">SpeculativeDecodingMode</span><span class="p">(</span><span class="n">IntFlag</span><span class="p">):</span>
<span class="c1"># [WARNING] KEEP BELOW DEFINITION IN SYNC WITH cpp/tensorrt_llm/runtime/speculativeDecodingMode.h</span>
<span class="n">NONE</span> <span class="o">=</span> <span class="n">auto</span><span class="p">()</span>
<span class="n">DRAFT_TOKENS_EXTERNAL</span> <span class="o">=</span> <span class="n">auto</span><span class="p">()</span>
<span class="n">MEDUSA</span> <span class="o">=</span> <span class="n">auto</span><span class="p">()</span>
<span class="n">LOOKAHEAD_DECODING</span> <span class="o">=</span> <span class="n">auto</span><span class="p">()</span>
<span class="n">EXPLICIT_DRAFT_TOKENS</span> <span class="o">=</span> <span class="n">auto</span><span class="p">()</span>
<span class="n">EAGLE</span> <span class="o">=</span> <span class="n">auto</span><span class="p">()</span>
<span class="n">NGRAM</span> <span class="o">=</span> <span class="n">auto</span><span class="p">()</span>
<div class="viewcode-block" id="SpeculativeDecodingMode.from_arguments">
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.models.html#tensorrt_llm.llmapi.SpeculativeDecodingMode.from_arguments">[docs]</a>
<span class="nd">@staticmethod</span>
<span class="k">def</span><span class="w"> </span><span class="nf">from_arguments</span><span class="p">(</span><span class="n">args</span><span class="p">:</span> <span class="n">argparse</span><span class="o">.</span><span class="n">Namespace</span><span class="p">):</span>
<span class="k">if</span> <span class="n">args</span><span class="o">.</span><span class="n">speculative_decoding_mode</span> <span class="ow">is</span> <span class="kc">None</span><span class="p">:</span>
<span class="k">return</span> <span class="n">SpeculativeDecodingMode</span><span class="o">.</span><span class="n">NONE</span>
<span class="k">elif</span> <span class="n">args</span><span class="o">.</span><span class="n">speculative_decoding_mode</span> <span class="o">==</span> <span class="s2">&quot;draft_tokens_external&quot;</span><span class="p">:</span>
<span class="k">return</span> <span class="n">SpeculativeDecodingMode</span><span class="o">.</span><span class="n">DRAFT_TOKENS_EXTERNAL</span>
<span class="k">elif</span> <span class="n">args</span><span class="o">.</span><span class="n">speculative_decoding_mode</span> <span class="o">==</span> <span class="s2">&quot;medusa&quot;</span><span class="p">:</span>
<span class="k">return</span> <span class="n">SpeculativeDecodingMode</span><span class="o">.</span><span class="n">MEDUSA</span>
<span class="k">elif</span> <span class="n">args</span><span class="o">.</span><span class="n">speculative_decoding_mode</span> <span class="o">==</span> <span class="s2">&quot;lookahead_decoding&quot;</span><span class="p">:</span>
<span class="k">return</span> <span class="n">SpeculativeDecodingMode</span><span class="o">.</span><span class="n">LOOKAHEAD_DECODING</span>
<span class="k">elif</span> <span class="n">args</span><span class="o">.</span><span class="n">speculative_decoding_mode</span> <span class="o">==</span> <span class="s2">&quot;explicit_draft_tokens&quot;</span><span class="p">:</span>
<span class="k">return</span> <span class="n">SpeculativeDecodingMode</span><span class="o">.</span><span class="n">EXPLICIT_DRAFT_TOKENS</span>
<span class="k">elif</span> <span class="n">args</span><span class="o">.</span><span class="n">speculative_decoding_mode</span> <span class="o">==</span> <span class="s2">&quot;eagle&quot;</span><span class="p">:</span>
<span class="k">return</span> <span class="n">SpeculativeDecodingMode</span><span class="o">.</span><span class="n">EAGLE</span>
<span class="k">elif</span> <span class="n">args</span><span class="o">.</span><span class="n">speculative_decoding_mode</span> <span class="o">==</span> <span class="s2">&quot;ngram&quot;</span><span class="p">:</span>
<span class="k">return</span> <span class="n">SpeculativeDecodingMode</span><span class="o">.</span><span class="n">NGRAM</span>
<span class="k">else</span><span class="p">:</span>
<span class="k">assert</span> <span class="kc">False</span><span class="p">,</span> <span class="s2">&quot;Unknown speculative_decoding_mode &quot;</span> <span class="o">+</span> <span class="n">args</span><span class="o">.</span><span class="n">speculative_decoding_mode</span></div>
</div>
<div class="viewcode-block" id="QuantConfig">
<a class="viewcode-back" href="../../../llm-api/reference.html#tensorrt_llm.llmapi.QuantConfig">[docs]</a>
<span class="nd">@dataclasses</span><span class="o">.</span><span class="n">dataclass</span>
<span class="k">class</span><span class="w"> </span><span class="nc">QuantConfig</span><span class="p">:</span>
<span class="w"> </span><span class="sd">&quot;&quot;&quot;</span>
<span class="sd"> Serializable quantization configuration class, part of the PretrainedConfig.</span>
<span class="sd"> Args:</span>
<span class="sd"> quant_algo (tensorrt_llm.quantization.mode.QuantAlgo, optional): Quantization algorithm. Defaults to None.</span>
<span class="sd"> kv_cache_quant_algo (tensorrt_llm.quantization.mode.QuantAlgo, optional): KV cache quantization algorithm. Defaults to None.</span>
<span class="sd"> group_size (int): The group size for group-wise quantization. Defaults to 128.</span>
<span class="sd"> smoothquant_val (float): The smoothing parameter alpha used in smooth quant. Defaults to 0.5.</span>
<span class="sd"> clamp_val (List[float], optional): The clamp values used in FP8 rowwise quantization. Defaults to None.</span>
<span class="sd"> use_meta_recipe (bool): Whether to use Meta&#39;s recipe for FP8 rowwise quantization. Defaults to False.</span>
<span class="sd"> has_zero_point (bool): Whether to use zero point for quantization. Defaults to False.</span>
<span class="sd"> pre_quant_scale (bool): Whether to use pre-quant scale for quantization. Defaults to False.</span>
<span class="sd"> exclude_modules (List[str], optional): The module name patterns that are skipped in quantization. Defaults to None.</span>
<span class="sd"> &quot;&quot;&quot;</span>
<span class="n">quant_algo</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="n">QuantAlgo</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</span>
<span class="n">kv_cache_quant_algo</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="n">QuantAlgo</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</span>
<span class="n">group_size</span><span class="p">:</span> <span class="nb">int</span> <span class="o">=</span> <span class="mi">128</span>
<span class="n">smoothquant_val</span><span class="p">:</span> <span class="nb">float</span> <span class="o">=</span> <span class="mf">0.5</span>
<span class="n">clamp_val</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="n">List</span><span class="p">[</span><span class="nb">float</span><span class="p">]]</span> <span class="o">=</span> <span class="kc">None</span>
<span class="n">use_meta_recipe</span><span class="p">:</span> <span class="nb">bool</span> <span class="o">=</span> <span class="kc">False</span>
<span class="n">has_zero_point</span><span class="p">:</span> <span class="nb">bool</span> <span class="o">=</span> <span class="kc">False</span>
<span class="n">pre_quant_scale</span><span class="p">:</span> <span class="nb">bool</span> <span class="o">=</span> <span class="kc">False</span>
<span class="n">exclude_modules</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="n">List</span><span class="p">[</span><span class="nb">str</span><span class="p">]]</span> <span class="o">=</span> <span class="kc">None</span>
<span class="nd">@cached_property</span>
<span class="k">def</span><span class="w"> </span><span class="nf">quant_mode</span><span class="p">(</span><span class="bp">self</span><span class="p">)</span> <span class="o">-&gt;</span> <span class="n">QuantModeWrapper</span><span class="p">:</span>
<span class="n">quant_mode_list</span> <span class="o">=</span> <span class="p">[</span>
<span class="n">QuantMode</span><span class="o">.</span><span class="n">from_quant_algo</span><span class="p">(</span>
<span class="bp">self</span><span class="o">.</span><span class="n">quant_algo</span><span class="p">,</span>
<span class="bp">self</span><span class="o">.</span><span class="n">kv_cache_quant_algo</span><span class="p">,</span>
<span class="p">)</span>
<span class="p">]</span>
<span class="k">return</span> <span class="n">QuantModeWrapper</span><span class="p">(</span><span class="n">quant_mode_list</span><span class="p">)</span>
<span class="nd">@cached_property</span>
<span class="k">def</span><span class="w"> </span><span class="nf">layer_quant_mode</span><span class="p">(</span><span class="bp">self</span><span class="p">)</span> <span class="o">-&gt;</span> <span class="n">QuantMode</span><span class="p">:</span>
<span class="k">return</span> <span class="n">QuantMode</span><span class="o">.</span><span class="n">from_quant_algo</span><span class="p">(</span>
<span class="bp">self</span><span class="o">.</span><span class="n">quant_algo</span><span class="p">,</span>
<span class="bp">self</span><span class="o">.</span><span class="n">kv_cache_quant_algo</span><span class="p">,</span>
<span class="p">)</span>
<span class="nd">@property</span>
<span class="k">def</span><span class="w"> </span><span class="nf">_use_plugin_sq</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
<span class="k">return</span> <span class="bp">self</span><span class="o">.</span><span class="n">quant_algo</span> <span class="ow">in</span> <span class="n">W8A8_SQ_PLUGIN_LIST</span>
<span class="nd">@property</span>
<span class="k">def</span><span class="w"> </span><span class="nf">_requires_calibration</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
<span class="k">return</span> <span class="bp">self</span><span class="o">.</span><span class="n">quant_algo</span> <span class="ow">in</span> <span class="p">(</span><span class="nb">set</span><span class="p">(</span><span class="n">QUANT_ALGO_LIST</span><span class="p">)</span> <span class="o">-</span> <span class="p">{</span>
<span class="n">QuantAlgo</span><span class="o">.</span><span class="n">W8A16</span><span class="p">,</span> <span class="n">QuantAlgo</span><span class="o">.</span><span class="n">W4A16</span><span class="p">,</span>
<span class="n">QuantAlgo</span><span class="o">.</span><span class="n">FP8_PER_CHANNEL_PER_TOKEN</span>
<span class="p">})</span> <span class="ow">or</span> <span class="bp">self</span><span class="o">.</span><span class="n">kv_cache_quant_algo</span> <span class="ow">in</span> <span class="n">KV_CACHE_QUANT_ALGO_LIST</span>
<span class="nd">@property</span>
<span class="k">def</span><span class="w"> </span><span class="nf">_requires_modelopt_quantization</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">quant_algo</span> <span class="ow">in</span> <span class="p">[</span>
<span class="n">QuantAlgo</span><span class="o">.</span><span class="n">NVFP4</span><span class="p">,</span> <span class="n">QuantAlgo</span><span class="o">.</span><span class="n">FP8</span><span class="p">,</span> <span class="n">QuantAlgo</span><span class="o">.</span><span class="n">W4A16_AWQ</span><span class="p">,</span>
<span class="n">QuantAlgo</span><span class="o">.</span><span class="n">W4A8_AWQ</span><span class="p">,</span> <span class="n">QuantAlgo</span><span class="o">.</span><span class="n">W8A8_SQ_PER_CHANNEL</span><span class="p">,</span>
<span class="n">QuantAlgo</span><span class="o">.</span><span class="n">MIXED_PRECISION</span>
<span class="p">]:</span>
<span class="k">return</span> <span class="kc">True</span>
<span class="k">elif</span> <span class="bp">self</span><span class="o">.</span><span class="n">quant_algo</span> <span class="ow">is</span> <span class="kc">None</span> <span class="ow">and</span> <span class="bp">self</span><span class="o">.</span><span class="n">kv_cache_quant_algo</span> <span class="o">==</span> <span class="n">QuantAlgo</span><span class="o">.</span><span class="n">FP8</span><span class="p">:</span>
<span class="k">return</span> <span class="kc">True</span>
<span class="k">else</span><span class="p">:</span>
<span class="k">return</span> <span class="kc">False</span>
<span class="k">def</span><span class="w"> </span><span class="nf">_get_quant_cfg</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">module_name</span><span class="o">=</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">exclude_modules</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
<span class="k">for</span> <span class="n">exclude_module</span> <span class="ow">in</span> <span class="bp">self</span><span class="o">.</span><span class="n">exclude_modules</span><span class="p">:</span>
<span class="k">if</span> <span class="n">exclude_module</span> <span class="o">==</span> <span class="n">module_name</span> <span class="ow">or</span> <span class="p">(</span>
<span class="n">exclude_module</span><span class="o">.</span><span class="n">endswith</span><span class="p">(</span><span class="s1">&#39;*&#39;</span><span class="p">)</span>
<span class="ow">and</span> <span class="n">module_name</span><span class="o">.</span><span class="n">startswith</span><span class="p">(</span><span class="n">exclude_module</span><span class="p">[:</span><span class="o">-</span><span class="mi">1</span><span class="p">])):</span>
<span class="k">return</span> <span class="n">LayerQuantConfig</span><span class="p">(</span><span class="n">quant_algo</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span>
<span class="n">quantized_layers</span><span class="o">=</span><span class="p">{})</span>
<span class="k">return</span> <span class="bp">self</span>
<span class="k">def</span><span class="w"> </span><span class="nf">_get_modelopt_qformat</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
<span class="n">algo_to_modelopt_map</span> <span class="o">=</span> <span class="p">{</span>
<span class="n">QuantAlgo</span><span class="o">.</span><span class="n">W8A16</span><span class="p">:</span> <span class="s2">&quot;int8_wo&quot;</span><span class="p">,</span>
<span class="n">QuantAlgo</span><span class="o">.</span><span class="n">W4A16</span><span class="p">:</span> <span class="s2">&quot;int4_wo&quot;</span><span class="p">,</span>
<span class="n">QuantAlgo</span><span class="o">.</span><span class="n">NVFP4</span><span class="p">:</span> <span class="s2">&quot;nvfp4&quot;</span><span class="p">,</span>
<span class="n">QuantAlgo</span><span class="o">.</span><span class="n">FP8</span><span class="p">:</span> <span class="s2">&quot;fp8&quot;</span><span class="p">,</span>
<span class="n">QuantAlgo</span><span class="o">.</span><span class="n">W4A16_AWQ</span><span class="p">:</span> <span class="s2">&quot;int4_awq&quot;</span><span class="p">,</span>
<span class="n">QuantAlgo</span><span class="o">.</span><span class="n">W4A8_AWQ</span><span class="p">:</span> <span class="s2">&quot;w4a8_awq&quot;</span><span class="p">,</span>
<span class="n">QuantAlgo</span><span class="o">.</span><span class="n">W8A8_SQ_PER_CHANNEL</span><span class="p">:</span> <span class="s2">&quot;int8_sq&quot;</span><span class="p">,</span>
<span class="p">}</span>
<span class="k">assert</span> <span class="bp">self</span><span class="o">.</span><span class="n">quant_algo</span> <span class="o">!=</span> <span class="n">QuantAlgo</span><span class="o">.</span><span class="n">MIXED_PRECISION</span><span class="p">,</span> <span class="sa">f</span><span class="s2">&quot;We don&#39;t support mixed precision in QuantConfig&quot;</span>
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">quant_algo</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="bp">self</span><span class="o">.</span><span class="n">quant_algo</span> <span class="ow">in</span> <span class="n">algo_to_modelopt_map</span><span class="p">,</span> <span class="sa">f</span><span class="s2">&quot;We don&#39;t use Modelopt for quantization algorithm </span><span class="si">{</span><span class="bp">self</span><span class="o">.</span><span class="n">quant_algo</span><span class="si">}</span><span class="s2">, you probably shall not call this&quot;</span>
<span class="k">return</span> <span class="n">algo_to_modelopt_map</span><span class="p">[</span><span class="bp">self</span><span class="o">.</span><span class="n">quant_algo</span><span class="p">]</span>
<span class="k">else</span><span class="p">:</span>
<span class="k">return</span> <span class="s1">&#39;full_prec&#39;</span>
<span class="k">def</span><span class="w"> </span><span class="nf">_get_modelopt_kv_cache_dtype</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
<span class="n">algo_to_modelopt_map</span> <span class="o">=</span> <span class="p">{</span>
<span class="n">QuantAlgo</span><span class="o">.</span><span class="n">FP8</span><span class="p">:</span> <span class="s1">&#39;fp8&#39;</span><span class="p">,</span>
<span class="n">QuantAlgo</span><span class="o">.</span><span class="n">INT8</span><span class="p">:</span> <span class="s1">&#39;int8&#39;</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">kv_cache_quant_algo</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="bp">self</span><span class="o">.</span><span class="n">kv_cache_quant_algo</span> <span class="ow">in</span> <span class="n">algo_to_modelopt_map</span><span class="p">,</span> <span class="sa">f</span><span class="s2">&quot;We don&#39;t use Modelopt for quantization algorithm </span><span class="si">{</span><span class="bp">self</span><span class="o">.</span><span class="n">kv_cache_quant_algo</span><span class="si">}</span><span class="s2">, you probably shall not call this&quot;</span>
<span class="k">return</span> <span class="n">algo_to_modelopt_map</span><span class="p">[</span><span class="bp">self</span><span class="o">.</span><span class="n">kv_cache_quant_algo</span><span class="p">]</span>
<span class="k">else</span><span class="p">:</span>
<span class="k">return</span> <span class="kc">None</span>
<div class="viewcode-block" id="QuantConfig.is_module_excluded_from_quantization">
<a class="viewcode-back" href="../../../llm-api/reference.html#tensorrt_llm.llmapi.QuantConfig.is_module_excluded_from_quantization">[docs]</a>
<span class="k">def</span><span class="w"> </span><span class="nf">is_module_excluded_from_quantization</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">name</span><span class="p">:</span> <span class="nb">str</span><span class="p">)</span> <span class="o">-&gt;</span> <span class="nb">bool</span><span class="p">:</span>
<span class="w"> </span><span class="sd">&quot;&quot;&quot;Check if the module is excluded from quantization.</span>
<span class="sd"> Args:</span>
<span class="sd"> name (str): The name of the module.</span>
<span class="sd"> Returns:</span>
<span class="sd"> bool: True if the module is excluded from quantization, False otherwise.</span>
<span class="sd"> &quot;&quot;&quot;</span>
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">exclude_modules</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
<span class="k">for</span> <span class="n">exclude_module</span> <span class="ow">in</span> <span class="bp">self</span><span class="o">.</span><span class="n">exclude_modules</span><span class="p">:</span>
<span class="k">if</span> <span class="n">fnmatch</span><span class="o">.</span><span class="n">fnmatchcase</span><span class="p">(</span><span class="n">name</span><span class="p">,</span> <span class="n">exclude_module</span><span class="p">):</span>
<span class="k">return</span> <span class="kc">True</span>
<span class="k">return</span> <span class="kc">False</span></div>
<div class="viewcode-block" id="QuantConfig.from_dict">
<a class="viewcode-back" href="../../../llm-api/reference.html#tensorrt_llm.llmapi.QuantConfig.from_dict">[docs]</a>
<span class="nd">@classmethod</span>
<span class="k">def</span><span class="w"> </span><span class="nf">from_dict</span><span class="p">(</span><span class="bp">cls</span><span class="p">,</span> <span class="n">config</span><span class="p">:</span> <span class="nb">dict</span><span class="p">)</span> <span class="o">-&gt;</span> <span class="s1">&#39;QuantConfig&#39;</span><span class="p">:</span>
<span class="w"> </span><span class="sd">&quot;&quot;&quot;Create a QuantConfig instance from a dict.</span>
<span class="sd"> Args:</span>
<span class="sd"> config (dict): The dict used to create QuantConfig.</span>
<span class="sd"> Returns:</span>
<span class="sd"> tensorrt_llm.models.modeling_utils.QuantConfig: The QuantConfig created from dict.</span>
<span class="sd"> &quot;&quot;&quot;</span>
<span class="n">obj</span> <span class="o">=</span> <span class="bp">cls</span><span class="p">(</span><span class="o">**</span><span class="n">config</span><span class="p">)</span>
<span class="k">return</span> <span class="n">obj</span></div>
<div class="viewcode-block" id="QuantConfig.to_dict">
<a class="viewcode-back" href="../../../llm-api/reference.html#tensorrt_llm.llmapi.QuantConfig.to_dict">[docs]</a>
<span class="k">def</span><span class="w"> </span><span class="nf">to_dict</span><span class="p">(</span><span class="bp">self</span><span class="p">)</span> <span class="o">-&gt;</span> <span class="nb">dict</span><span class="p">:</span>
<span class="w"> </span><span class="sd">&quot;&quot;&quot;Dump a QuantConfig instance to a dict.</span>
<span class="sd"> Returns:</span>
<span class="sd"> dict: The dict dumped from QuantConfig.</span>
<span class="sd"> &quot;&quot;&quot;</span>
<span class="k">return</span> <span class="n">dataclasses</span><span class="o">.</span><span class="n">asdict</span><span class="p">(</span><span class="bp">self</span><span class="p">)</span></div>
</div>
<span class="nd">@dataclasses</span><span class="o">.</span><span class="n">dataclass</span>
<span class="k">class</span><span class="w"> </span><span class="nc">LayerQuantConfig</span><span class="p">(</span><span class="n">QuantConfig</span><span class="p">):</span>
<span class="n">quant_algo</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="n">QuantConfig</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</span>
<span class="n">kv_cache_quant_algo</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="n">QuantConfig</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</span>
<span class="n">quantized_layers</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="n">Dict</span><span class="p">[</span><span class="nb">str</span><span class="p">,</span> <span class="n">QuantConfig</span><span class="p">]]</span> <span class="o">=</span> <span class="kc">None</span>
<span class="k">def</span><span class="w"> </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">quant_algo</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="n">QuantConfig</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</span><span class="p">,</span>
<span class="n">kv_cache_quant_algo</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="n">QuantConfig</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</span><span class="p">,</span>
<span class="n">quantized_layers</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="n">Dict</span><span class="p">[</span><span class="nb">str</span><span class="p">,</span> <span class="n">QuantConfig</span><span class="p">]]</span> <span class="o">=</span> <span class="kc">None</span><span class="p">,</span>
<span class="o">**</span><span class="n">kwargs</span><span class="p">):</span>
<span class="bp">self</span><span class="o">.</span><span class="n">quant_algo</span> <span class="o">=</span> <span class="n">quant_algo</span>
<span class="bp">self</span><span class="o">.</span><span class="n">quantized_layers</span> <span class="o">=</span> <span class="n">quantized_layers</span>
<span class="bp">self</span><span class="o">.</span><span class="n">kv_cache_quant_algo</span> <span class="o">=</span> <span class="n">kv_cache_quant_algo</span>
<span class="bp">self</span><span class="o">.</span><span class="n">auto_quant_mode</span> <span class="o">=</span> <span class="p">{}</span>
<span class="k">for</span> <span class="n">name</span><span class="p">,</span> <span class="n">layer_config</span> <span class="ow">in</span> <span class="bp">self</span><span class="o">.</span><span class="n">quantized_layers</span><span class="o">.</span><span class="n">items</span><span class="p">():</span>
<span class="bp">self</span><span class="o">.</span><span class="n">auto_quant_mode</span><span class="o">.</span><span class="n">update</span><span class="p">({</span>
<span class="n">name</span><span class="p">:</span>
<span class="n">QuantMode</span><span class="o">.</span><span class="n">from_quant_algo</span><span class="p">(</span>
<span class="n">layer_config</span><span class="o">.</span><span class="n">quant_algo</span><span class="p">,</span>
<span class="bp">self</span><span class="o">.</span><span class="n">kv_cache_quant_algo</span><span class="p">,</span>
<span class="p">)</span>
<span class="p">})</span>
<span class="k">for</span> <span class="n">key</span> <span class="ow">in</span> <span class="n">kwargs</span><span class="p">:</span>
<span class="n">logger</span><span class="o">.</span><span class="n">warning</span><span class="p">(</span>
<span class="sa">f</span><span class="s2">&quot;Warning: Unrecognized parameter &#39;</span><span class="si">{</span><span class="n">key</span><span class="si">}</span><span class="s2">&#39; with value &#39;</span><span class="si">{</span><span class="n">kwargs</span><span class="p">[</span><span class="n">key</span><span class="p">]</span><span class="si">}</span><span class="s2">&#39;&quot;</span>
<span class="p">)</span>
<span class="nd">@cached_property</span>
<span class="k">def</span><span class="w"> </span><span class="nf">quant_mode</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
<span class="n">quant_mode_list</span> <span class="o">=</span> <span class="nb">list</span><span class="p">(</span><span class="nb">set</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">auto_quant_mode</span><span class="o">.</span><span class="n">values</span><span class="p">()))</span>
<span class="k">return</span> <span class="n">QuantModeWrapper</span><span class="p">(</span><span class="n">quant_mode_list</span><span class="p">)</span>
<span class="c1">#@lru_cache(maxsize=None)</span>
<span class="k">def</span><span class="w"> </span><span class="nf">layer_quant_mode</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">layer_name</span><span class="p">)</span> <span class="o">-&gt;</span> <span class="n">QuantMode</span><span class="p">:</span>
<span class="k">for</span> <span class="n">name</span><span class="p">,</span> <span class="n">quant_mode</span> <span class="ow">in</span> <span class="bp">self</span><span class="o">.</span><span class="n">auto_quant_mode</span><span class="o">.</span><span class="n">items</span><span class="p">():</span>
<span class="k">if</span> <span class="n">fnmatch</span><span class="o">.</span><span class="n">fnmatch</span><span class="p">(</span><span class="n">layer_name</span><span class="p">,</span> <span class="n">name</span><span class="p">):</span>
<span class="k">return</span> <span class="n">quant_mode</span>
<span class="k">return</span> <span class="n">QuantMode</span><span class="p">(</span><span class="mi">0</span><span class="p">)</span>
<span class="nd">@cached_property</span>
<span class="k">def</span><span class="w"> </span><span class="nf">auto_quant_list</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
<span class="n">quant_list</span> <span class="o">=</span> <span class="p">[]</span>
<span class="k">for</span> <span class="n">_</span><span class="p">,</span> <span class="n">layer_config</span> <span class="ow">in</span> <span class="bp">self</span><span class="o">.</span><span class="n">quantized_layers</span><span class="o">.</span><span class="n">items</span><span class="p">():</span>
<span class="n">quant_list</span><span class="o">.</span><span class="n">append</span><span class="p">(</span><span class="n">layer_config</span><span class="o">.</span><span class="n">quant_algo</span><span class="p">)</span>
<span class="k">return</span> <span class="nb">list</span><span class="p">(</span><span class="nb">set</span><span class="p">(</span><span class="n">quant_list</span><span class="p">))</span>
<span class="nd">@classmethod</span>
<span class="k">def</span><span class="w"> </span><span class="nf">from_dict</span><span class="p">(</span><span class="bp">cls</span><span class="p">,</span> <span class="n">config</span><span class="p">:</span> <span class="nb">dict</span><span class="p">):</span>
<span class="n">quantized_layers</span> <span class="o">=</span> <span class="n">config</span><span class="o">.</span><span class="n">pop</span><span class="p">(</span><span class="s1">&#39;quantized_layers&#39;</span><span class="p">,</span> <span class="p">{})</span>
<span class="n">quantized_layers_dict</span> <span class="o">=</span> <span class="p">{</span>
<span class="n">layer_name</span><span class="p">:</span> <span class="n">QuantConfig</span><span class="p">(</span><span class="o">**</span><span class="n">layer_config</span><span class="p">)</span>
<span class="k">for</span> <span class="n">layer_name</span><span class="p">,</span> <span class="n">layer_config</span> <span class="ow">in</span> <span class="n">quantized_layers</span><span class="o">.</span><span class="n">items</span><span class="p">()</span>
<span class="p">}</span>
<span class="n">obj</span> <span class="o">=</span> <span class="bp">cls</span><span class="p">(</span><span class="n">quantized_layers</span><span class="o">=</span><span class="n">quantized_layers_dict</span><span class="p">,</span> <span class="o">**</span><span class="n">config</span><span class="p">)</span>
<span class="k">return</span> <span class="n">obj</span>
<span class="c1">#@lru_cache(maxsize=None)</span>
<span class="k">def</span><span class="w"> </span><span class="nf">_get_quant_cfg</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">module_name</span><span class="p">):</span>
<span class="n">quant_res</span> <span class="o">=</span> <span class="n">QuantConfig</span><span class="p">()</span>
<span class="k">for</span> <span class="n">name</span><span class="p">,</span> <span class="n">quant_cfg</span> <span class="ow">in</span> <span class="bp">self</span><span class="o">.</span><span class="n">quantized_layers</span><span class="o">.</span><span class="n">items</span><span class="p">():</span>
<span class="k">if</span> <span class="n">fnmatch</span><span class="o">.</span><span class="n">fnmatch</span><span class="p">(</span><span class="n">module_name</span><span class="p">,</span> <span class="n">name</span><span class="p">):</span>
<span class="n">quant_res</span> <span class="o">=</span> <span class="n">quant_cfg</span>
<span class="k">break</span>
<span class="k">return</span> <span class="n">quant_res</span>
<span class="k">def</span><span class="w"> </span><span class="nf">_get_modelopt_qformat</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
<span class="n">algo_to_modelopt_map</span> <span class="o">=</span> <span class="p">{</span>
<span class="n">QuantAlgo</span><span class="o">.</span><span class="n">NVFP4</span><span class="p">:</span> <span class="s2">&quot;nvfp4&quot;</span><span class="p">,</span>
<span class="n">QuantAlgo</span><span class="o">.</span><span class="n">FP8</span><span class="p">:</span> <span class="s2">&quot;fp8&quot;</span><span class="p">,</span>
<span class="n">QuantAlgo</span><span class="o">.</span><span class="n">W4A16_AWQ</span><span class="p">:</span> <span class="s2">&quot;int4_awq&quot;</span><span class="p">,</span>
<span class="n">QuantAlgo</span><span class="o">.</span><span class="n">W4A8_AWQ</span><span class="p">:</span> <span class="s2">&quot;w4a8_awq&quot;</span><span class="p">,</span>
<span class="n">QuantAlgo</span><span class="o">.</span><span class="n">W8A8_SQ_PER_CHANNEL</span><span class="p">:</span> <span class="s2">&quot;int8_sq&quot;</span><span class="p">,</span>
<span class="p">}</span>
<span class="k">assert</span> <span class="bp">self</span><span class="o">.</span><span class="n">quant_algo</span> <span class="o">==</span> <span class="n">QuantAlgo</span><span class="o">.</span><span class="n">MIXED_PRECISION</span><span class="p">,</span> <span class="sa">f</span><span class="s2">&quot;We only support mixed precision quantization in LayerQuantConfig&quot;</span>
<span class="n">autoq_format</span> <span class="o">=</span> <span class="s1">&#39;,&#39;</span><span class="o">.</span><span class="n">join</span><span class="p">(</span>
<span class="p">[</span><span class="n">algo_to_modelopt_map</span><span class="p">[</span><span class="n">item</span><span class="p">]</span> <span class="k">for</span> <span class="n">item</span> <span class="ow">in</span> <span class="bp">self</span><span class="o">.</span><span class="n">auto_quant_list</span><span class="p">])</span>
<span class="k">return</span> <span class="n">autoq_format</span>
<span class="k">def</span><span class="w"> </span><span class="nf">to_dict</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
<span class="n">output</span> <span class="o">=</span> <span class="n">copy</span><span class="o">.</span><span class="n">deepcopy</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="vm">__dict__</span><span class="p">)</span>
<span class="n">output</span><span class="o">.</span><span class="n">pop</span><span class="p">(</span><span class="s1">&#39;auto_quant_mode&#39;</span><span class="p">,</span> <span class="kc">None</span><span class="p">)</span>
<span class="n">output</span><span class="o">.</span><span class="n">pop</span><span class="p">(</span><span class="s1">&#39;quant_mode&#39;</span><span class="p">,</span> <span class="kc">None</span><span class="p">)</span>
<span class="k">for</span> <span class="n">name</span><span class="p">,</span> <span class="n">per_layer_config</span> <span class="ow">in</span> <span class="n">output</span><span class="p">[</span><span class="s1">&#39;quantized_layers&#39;</span><span class="p">]</span><span class="o">.</span><span class="n">items</span><span class="p">():</span>
<span class="n">per_layer_config</span> <span class="o">=</span> <span class="n">per_layer_config</span><span class="o">.</span><span class="n">to_dict</span><span class="p">()</span>
<span class="n">output</span><span class="p">[</span><span class="s1">&#39;quantized_layers&#39;</span><span class="p">][</span><span class="n">name</span><span class="p">]</span> <span class="o">=</span> <span class="n">per_layer_config</span>
<span class="k">return</span> <span class="n">output</span>
<div class="viewcode-block" id="PretrainedConfig">
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.models.html#tensorrt_llm.llmapi.PretrainedConfig">[docs]</a>
<span class="k">class</span><span class="w"> </span><span class="nc">PretrainedConfig</span><span class="p">:</span>
<span class="k">def</span><span class="w"> </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">architecture</span><span class="p">:</span> <span class="nb">str</span><span class="p">,</span>
<span class="n">dtype</span><span class="p">:</span> <span class="nb">str</span><span class="p">,</span>
<span class="n">hidden_size</span><span class="p">:</span> <span class="nb">int</span><span class="p">,</span>
<span class="n">num_hidden_layers</span><span class="p">:</span> <span class="nb">int</span><span class="p">,</span>
<span class="n">num_attention_heads</span><span class="p">:</span> <span class="nb">int</span><span class="p">,</span>
<span class="n">vocab_size</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="nb">int</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</span><span class="p">,</span>
<span class="n">hidden_act</span><span class="p">:</span> <span class="nb">str</span> <span class="o">=</span> <span class="s1">&#39;gelu&#39;</span><span class="p">,</span>
<span class="n">logits_dtype</span><span class="p">:</span> <span class="nb">str</span> <span class="o">=</span> <span class="s1">&#39;float32&#39;</span><span class="p">,</span>
<span class="n">norm_epsilon</span><span class="p">:</span> <span class="nb">float</span> <span class="o">=</span> <span class="mf">1e-5</span><span class="p">,</span>
<span class="n">position_embedding_type</span><span class="p">:</span> <span class="n">Union</span><span class="p">[</span>
<span class="n">PositionEmbeddingType</span><span class="p">,</span>
<span class="nb">str</span><span class="p">]</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="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="nb">int</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</span><span class="p">,</span>
<span class="n">rotary_embedding_dim</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="nb">int</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</span><span class="p">,</span>
<span class="n">num_key_value_heads</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="nb">int</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</span><span class="p">,</span>
<span class="n">intermediate_size</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="nb">int</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</span><span class="p">,</span>
<span class="n">mapping</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="n">Union</span><span class="p">[</span><span class="n">Mapping</span><span class="p">,</span> <span class="nb">dict</span><span class="p">]]</span> <span class="o">=</span> <span class="kc">None</span><span class="p">,</span>
<span class="n">quantization</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="n">Union</span><span class="p">[</span><span class="n">QuantConfig</span><span class="p">,</span> <span class="nb">dict</span><span class="p">]]</span> <span class="o">=</span> <span class="kc">None</span><span class="p">,</span>
<span class="n">use_parallel_embedding</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">embedding_sharding_dim</span><span class="p">:</span> <span class="nb">int</span> <span class="o">=</span> <span class="mi">0</span><span class="p">,</span>
<span class="n">head_size</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="nb">int</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</span><span class="p">,</span>
<span class="n">qk_layernorm</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">runtime_defaults</span><span class="p">:</span> <span class="s2">&quot;RuntimeDefaultsIn&quot;</span> <span class="o">=</span> <span class="kc">None</span><span class="p">,</span>
<span class="o">**</span><span class="n">kwargs</span><span class="p">):</span>
<span class="bp">self</span><span class="o">.</span><span class="n">architecture</span> <span class="o">=</span> <span class="n">architecture</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">vocab_size</span> <span class="o">=</span> <span class="n">vocab_size</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">num_hidden_layers</span> <span class="o">=</span> <span class="n">num_hidden_layers</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="bp">self</span><span class="o">.</span><span class="n">hidden_act</span> <span class="o">=</span> <span class="n">hidden_act</span>
<span class="bp">self</span><span class="o">.</span><span class="n">logits_dtype</span> <span class="o">=</span> <span class="n">logits_dtype</span>
<span class="bp">self</span><span class="o">.</span><span class="n">norm_epsilon</span> <span class="o">=</span> <span class="n">norm_epsilon</span>
<span class="bp">self</span><span class="o">.</span><span class="n">runtime_defaults</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">create_runtime_defaults</span><span class="p">(</span><span class="n">runtime_defaults</span><span class="p">)</span>
<span class="k">if</span> <span class="nb">isinstance</span><span class="p">(</span><span class="n">position_embedding_type</span><span class="p">,</span> <span class="nb">str</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">from_string</span><span class="p">(</span>
<span class="n">position_embedding_type</span><span class="p">)</span>
<span class="k">assert</span> <span class="nb">isinstance</span><span class="p">(</span><span class="n">position_embedding_type</span><span class="p">,</span> <span class="n">PositionEmbeddingType</span><span class="p">)</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="k">if</span> <span class="n">num_key_value_heads</span> <span class="ow">is</span> <span class="kc">None</span><span class="p">:</span>
<span class="n">num_key_value_heads</span> <span class="o">=</span> <span class="n">num_attention_heads</span>
<span class="bp">self</span><span class="o">.</span><span class="n">num_key_value_heads</span> <span class="o">=</span> <span class="n">num_key_value_heads</span>
<span class="k">if</span> <span class="n">intermediate_size</span> <span class="ow">is</span> <span class="kc">None</span><span class="p">:</span>
<span class="n">intermediate_size</span> <span class="o">=</span> <span class="n">hidden_size</span> <span class="o">*</span> <span class="mi">4</span>
<span class="bp">self</span><span class="o">.</span><span class="n">intermediate_size</span> <span class="o">=</span> <span class="n">intermediate_size</span>
<span class="bp">self</span><span class="o">.</span><span class="n">max_position_embeddings</span> <span class="o">=</span> <span class="n">max_position_embeddings</span>
<span class="k">if</span> <span class="n">mapping</span> <span class="ow">is</span> <span class="kc">None</span><span class="p">:</span>
<span class="n">mapping</span> <span class="o">=</span> <span class="n">Mapping</span><span class="p">()</span>
<span class="k">elif</span> <span class="nb">isinstance</span><span class="p">(</span><span class="n">mapping</span><span class="p">,</span> <span class="nb">dict</span><span class="p">):</span>
<span class="n">mapping</span> <span class="o">=</span> <span class="n">Mapping</span><span class="o">.</span><span class="n">from_dict</span><span class="p">(</span><span class="n">mapping</span><span class="p">)</span>
<span class="k">assert</span> <span class="nb">isinstance</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="bp">self</span><span class="o">.</span><span class="n">mapping</span> <span class="o">=</span> <span class="n">mapping</span>
<span class="k">if</span> <span class="n">quantization</span> <span class="ow">is</span> <span class="kc">None</span><span class="p">:</span>
<span class="n">quantization</span> <span class="o">=</span> <span class="n">QuantConfig</span><span class="p">()</span>
<span class="k">elif</span> <span class="nb">isinstance</span><span class="p">(</span><span class="n">quantization</span><span class="p">,</span> <span class="nb">dict</span><span class="p">):</span>
<span class="n">quantization</span> <span class="o">=</span> <span class="n">QuantConfig</span><span class="o">.</span><span class="n">from_dict</span><span class="p">(</span><span class="n">quantization</span><span class="p">)</span>
<span class="k">assert</span> <span class="nb">isinstance</span><span class="p">(</span><span class="n">quantization</span><span class="p">,</span> <span class="n">QuantConfig</span><span class="p">)</span>
<span class="bp">self</span><span class="o">.</span><span class="n">quantization</span> <span class="o">=</span> <span class="n">quantization</span>
<span class="bp">self</span><span class="o">.</span><span class="n">use_parallel_embedding</span> <span class="o">=</span> <span class="n">use_parallel_embedding</span>
<span class="bp">self</span><span class="o">.</span><span class="n">embedding_sharding_dim</span> <span class="o">=</span> <span class="n">embedding_sharding_dim</span>
<span class="k">if</span> <span class="n">head_size</span> <span class="ow">is</span> <span class="kc">None</span><span class="p">:</span>
<span class="n">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="bp">self</span><span class="o">.</span><span class="n">head_size</span> <span class="o">=</span> <span class="n">head_size</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="k">if</span> <span class="n">rotary_embedding_dim</span> <span class="ow">is</span> <span class="kc">None</span><span class="p">:</span>
<span class="n">rotary_embedding_percentage</span> <span class="o">=</span> <span class="n">kwargs</span><span class="o">.</span><span class="n">get</span><span class="p">(</span><span class="s1">&#39;rotary_pct&#39;</span><span class="p">,</span> <span class="mf">1.0</span><span class="p">)</span>
<span class="n">rotary_embedding_dim</span> <span class="o">=</span> <span class="n">kwargs</span><span class="o">.</span><span class="n">get</span><span class="p">(</span>
<span class="s1">&#39;rotary_dim&#39;</span><span class="p">,</span> <span class="nb">int</span><span class="p">(</span><span class="n">head_size</span> <span class="o">*</span> <span class="n">rotary_embedding_percentage</span><span class="p">))</span>
<span class="bp">self</span><span class="o">.</span><span class="n">rotary_embedding_dim</span> <span class="o">=</span> <span class="n">rotary_embedding_dim</span>
<span class="k">for</span> <span class="n">key</span><span class="p">,</span> <span class="n">value</span> <span class="ow">in</span> <span class="n">kwargs</span><span class="o">.</span><span class="n">items</span><span class="p">():</span>
<span class="k">try</span><span class="p">:</span>
<span class="nb">setattr</span><span class="p">(</span><span class="bp">self</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">logger</span><span class="o">.</span><span class="n">warning</span><span class="p">(</span>
<span class="sa">f</span><span class="s2">&quot;Implicitly setting </span><span class="si">{</span><span class="bp">self</span><span class="o">.</span><span class="vm">__class__</span><span class="o">.</span><span class="vm">__name__</span><span class="si">}</span><span class="s2">.</span><span class="si">{</span><span class="n">key</span><span class="si">}</span><span class="s2"> = </span><span class="si">{</span><span class="n">value</span><span class="si">}</span><span class="s2">&quot;</span>
<span class="p">)</span>
<span class="k">except</span> <span class="ne">AttributeError</span> <span class="k">as</span> <span class="n">err</span><span class="p">:</span>
<span class="k">raise</span> <span class="n">err</span>
<div class="viewcode-block" id="PretrainedConfig.create_runtime_defaults">
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.models.html#tensorrt_llm.llmapi.PretrainedConfig.create_runtime_defaults">[docs]</a>
<span class="nd">@staticmethod</span>
<span class="k">def</span><span class="w"> </span><span class="nf">create_runtime_defaults</span><span class="p">(</span>
<span class="n">defaults</span><span class="p">:</span> <span class="s2">&quot;RuntimeDefaultsIn&quot;</span> <span class="o">=</span> <span class="kc">None</span><span class="p">)</span> <span class="o">-&gt;</span> <span class="n">Optional</span><span class="p">[</span><span class="n">RuntimeDefaults</span><span class="p">]:</span>
<span class="k">if</span> <span class="nb">isinstance</span><span class="p">(</span><span class="n">defaults</span><span class="p">,</span> <span class="nb">dict</span><span class="p">):</span>
<span class="k">return</span> <span class="n">RuntimeDefaults</span><span class="p">(</span><span class="o">**</span><span class="n">defaults</span><span class="p">)</span>
<span class="k">return</span> <span class="n">defaults</span></div>
<span class="nd">@property</span>
<span class="k">def</span><span class="w"> </span><span class="nf">kv_dtype</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
<span class="c1"># TODO: need to align the kv dtype</span>
<span class="c1"># now assume the kv cache is for all layers</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_int8_kv_cache</span><span class="p">():</span>
<span class="k">return</span> <span class="s1">&#39;int8&#39;</span>
<span class="k">elif</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">return</span> <span class="s1">&#39;fp8&#39;</span>
<span class="k">elif</span> <span class="bp">self</span><span class="o">.</span><span class="n">quant_mode</span><span class="o">.</span><span class="n">has_fp4_kv_cache</span><span class="p">():</span>
<span class="k">return</span> <span class="s1">&#39;fp4&#39;</span>
<span class="k">else</span><span class="p">:</span>
<span class="k">return</span> <span class="bp">self</span><span class="o">.</span><span class="n">dtype</span>
<div class="viewcode-block" id="PretrainedConfig.set_if_not_exist">
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.models.html#tensorrt_llm.llmapi.PretrainedConfig.set_if_not_exist">[docs]</a>
<span class="k">def</span><span class="w"> </span><span class="nf">set_if_not_exist</span><span class="p">(</span><span class="bp">self</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="k">if</span> <span class="ow">not</span> <span class="nb">hasattr</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">key</span><span class="p">):</span>
<span class="nb">setattr</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">key</span><span class="p">,</span> <span class="n">value</span><span class="p">)</span></div>
<div class="viewcode-block" id="PretrainedConfig.from_dict">
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.models.html#tensorrt_llm.llmapi.PretrainedConfig.from_dict">[docs]</a>
<span class="nd">@classmethod</span>
<span class="k">def</span><span class="w"> </span><span class="nf">from_dict</span><span class="p">(</span><span class="bp">cls</span><span class="p">,</span> <span class="n">config</span><span class="p">:</span> <span class="nb">dict</span><span class="p">):</span>
<span class="c1"># Maybe we need AutoConfig for this</span>
<span class="kn">from</span><span class="w"> </span><span class="nn">.</span><span class="w"> </span><span class="kn">import</span> <span class="n">MODEL_MAP</span>
<span class="n">model_cls</span> <span class="o">=</span> <span class="n">MODEL_MAP</span><span class="p">[</span><span class="n">config</span><span class="p">[</span><span class="s1">&#39;architecture&#39;</span><span class="p">]]</span>
<span class="n">config_cls</span> <span class="o">=</span> <span class="nb">getattr</span><span class="p">(</span><span class="n">model_cls</span><span class="p">,</span> <span class="s1">&#39;config_class&#39;</span><span class="p">,</span> <span class="bp">cls</span><span class="p">)</span>
<span class="k">return</span> <span class="n">config_cls</span><span class="p">(</span><span class="o">**</span><span class="n">config</span><span class="p">)</span></div>
<div class="viewcode-block" id="PretrainedConfig.to_dict">
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.models.html#tensorrt_llm.llmapi.PretrainedConfig.to_dict">[docs]</a>
<span class="k">def</span><span class="w"> </span><span class="nf">to_dict</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
<span class="n">output</span> <span class="o">=</span> <span class="n">copy</span><span class="o">.</span><span class="n">deepcopy</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="vm">__dict__</span><span class="p">)</span>
<span class="n">output</span><span class="p">[</span><span class="s1">&#39;position_embedding_type&#39;</span><span class="p">]</span> <span class="o">=</span> <span class="nb">str</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">output</span><span class="p">[</span><span class="s1">&#39;mapping&#39;</span><span class="p">]</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">mapping</span><span class="o">.</span><span class="n">to_dict</span><span class="p">()</span>
<span class="n">output</span><span class="p">[</span><span class="s1">&#39;mapping&#39;</span><span class="p">]</span><span class="o">.</span><span class="n">pop</span><span class="p">(</span><span class="s1">&#39;rank&#39;</span><span class="p">)</span>
<span class="n">output</span><span class="p">[</span><span class="s1">&#39;quantization&#39;</span><span class="p">]</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">quantization</span><span class="o">.</span><span class="n">to_dict</span><span class="p">()</span>
<span class="k">return</span> <span class="n">output</span></div>
<div class="viewcode-block" id="PretrainedConfig.from_json_file">
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.models.html#tensorrt_llm.llmapi.PretrainedConfig.from_json_file">[docs]</a>
<span class="nd">@classmethod</span>
<span class="k">def</span><span class="w"> </span><span class="nf">from_json_file</span><span class="p">(</span><span class="bp">cls</span><span class="p">,</span> <span class="n">config_file</span><span class="p">:</span> <span class="nb">str</span><span class="p">):</span>
<span class="k">with</span> <span class="nb">open</span><span class="p">(</span><span class="n">config_file</span><span class="p">)</span> <span class="k">as</span> <span class="n">f</span><span class="p">:</span>
<span class="n">config</span> <span class="o">=</span> <span class="n">json</span><span class="o">.</span><span class="n">load</span><span class="p">(</span><span class="n">f</span><span class="p">)</span>
<span class="n">obj</span> <span class="o">=</span> <span class="bp">cls</span><span class="o">.</span><span class="n">from_dict</span><span class="p">(</span><span class="n">config</span><span class="p">)</span>
<span class="k">if</span> <span class="n">obj</span><span class="o">.</span><span class="n">quantization</span><span class="o">.</span><span class="n">quant_algo</span> <span class="o">==</span> <span class="n">QuantAlgo</span><span class="o">.</span><span class="n">MIXED_PRECISION</span><span class="p">:</span>
<span class="k">try</span><span class="p">:</span>
<span class="n">layer_config_path</span> <span class="o">=</span> <span class="nb">str</span><span class="p">(</span><span class="n">config_file</span><span class="p">)</span><span class="o">.</span><span class="n">replace</span><span class="p">(</span>
<span class="s1">&#39;config.json&#39;</span><span class="p">,</span> <span class="s1">&#39;quant_cfg.json&#39;</span><span class="p">)</span>
<span class="n">obj</span><span class="o">.</span><span class="n">to_layer_quant_config</span><span class="p">(</span><span class="n">layer_config_path</span><span class="p">)</span>
<span class="k">except</span> <span class="ne">Exception</span> <span class="k">as</span> <span class="n">e</span><span class="p">:</span>
<span class="k">raise</span> <span class="ne">RuntimeError</span><span class="p">(</span>
<span class="sa">f</span><span class="s2">&quot;Encounter error &#39;</span><span class="si">{</span><span class="n">e</span><span class="si">}</span><span class="s2">&#39; for read quantization config &#39;</span><span class="si">{</span><span class="n">layer_config_path</span><span class="si">}</span><span class="s2">&#39;&quot;</span>
<span class="p">)</span>
<span class="k">return</span> <span class="n">obj</span></div>
<div class="viewcode-block" id="PretrainedConfig.from_checkpoint">
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.models.html#tensorrt_llm.llmapi.PretrainedConfig.from_checkpoint">[docs]</a>
<span class="nd">@classmethod</span>
<span class="k">def</span><span class="w"> </span><span class="nf">from_checkpoint</span><span class="p">(</span><span class="bp">cls</span><span class="p">,</span> <span class="n">ckpt_dir</span><span class="p">:</span> <span class="nb">str</span><span class="p">):</span>
<span class="k">return</span> <span class="bp">cls</span><span class="o">.</span><span class="n">from_json_file</span><span class="p">(</span><span class="n">os</span><span class="o">.</span><span class="n">path</span><span class="o">.</span><span class="n">join</span><span class="p">(</span><span class="n">ckpt_dir</span><span class="p">,</span> <span class="s1">&#39;config.json&#39;</span><span class="p">))</span></div>
<div class="viewcode-block" id="PretrainedConfig.to_json_file">
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.models.html#tensorrt_llm.llmapi.PretrainedConfig.to_json_file">[docs]</a>
<span class="k">def</span><span class="w"> </span><span class="nf">to_json_file</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">config_file</span><span class="p">:</span> <span class="nb">str</span><span class="p">):</span>
<span class="k">with</span> <span class="nb">open</span><span class="p">(</span><span class="n">config_file</span><span class="p">,</span> <span class="s1">&#39;w&#39;</span><span class="p">)</span> <span class="k">as</span> <span class="n">f</span><span class="p">:</span>
<span class="n">json</span><span class="o">.</span><span class="n">dump</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">to_dict</span><span class="p">(),</span> <span class="n">f</span><span class="p">,</span> <span class="n">indent</span><span class="o">=</span><span class="mi">4</span><span class="p">)</span></div>
<div class="viewcode-block" id="PretrainedConfig.to_layer_quant_config">
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.models.html#tensorrt_llm.llmapi.PretrainedConfig.to_layer_quant_config">[docs]</a>
<span class="k">def</span><span class="w"> </span><span class="nf">to_layer_quant_config</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">config_file</span><span class="p">:</span> <span class="nb">str</span><span class="p">):</span>
<span class="k">with</span> <span class="nb">open</span><span class="p">(</span><span class="n">config_file</span><span class="p">)</span> <span class="k">as</span> <span class="n">f</span><span class="p">:</span>
<span class="n">config</span> <span class="o">=</span> <span class="n">json</span><span class="o">.</span><span class="n">load</span><span class="p">(</span><span class="n">f</span><span class="p">)</span>
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">architecture</span> <span class="o">==</span> <span class="s2">&quot;MixtralForCausalLM&quot;</span><span class="p">:</span>
<span class="k">for</span> <span class="n">layer_name</span> <span class="ow">in</span> <span class="nb">list</span><span class="p">(</span><span class="n">config</span><span class="p">[</span><span class="s2">&quot;quantized_layers&quot;</span><span class="p">]</span><span class="o">.</span><span class="n">keys</span><span class="p">()):</span>
<span class="n">quant_cfg</span> <span class="o">=</span> <span class="n">config</span><span class="p">[</span><span class="s2">&quot;quantized_layers&quot;</span><span class="p">][</span><span class="n">layer_name</span><span class="p">]</span>
<span class="k">if</span> <span class="s2">&quot;mlp.fc&quot;</span> <span class="ow">in</span> <span class="n">layer_name</span> <span class="ow">or</span> <span class="s2">&quot;mlp.proj&quot;</span> <span class="ow">in</span> <span class="n">layer_name</span><span class="p">:</span>
<span class="n">moe_name</span><span class="p">,</span> <span class="n">_</span> <span class="o">=</span> <span class="n">layer_name</span><span class="o">.</span><span class="n">rsplit</span><span class="p">(</span><span class="s1">&#39;.&#39;</span><span class="p">,</span> <span class="mi">1</span><span class="p">)</span>
<span class="k">if</span> <span class="n">moe_name</span> <span class="ow">not</span> <span class="ow">in</span> <span class="n">config</span><span class="p">[</span><span class="s2">&quot;quantized_layers&quot;</span><span class="p">]:</span>
<span class="n">config</span><span class="p">[</span><span class="s2">&quot;quantized_layers&quot;</span><span class="p">][</span><span class="n">moe_name</span><span class="p">]</span> <span class="o">=</span> <span class="n">quant_cfg</span>
<span class="k">else</span><span class="p">:</span>
<span class="k">assert</span> <span class="n">quant_cfg</span> <span class="o">==</span> <span class="n">config</span><span class="p">[</span><span class="s2">&quot;quantized_layers&quot;</span><span class="p">][</span>
<span class="n">moe_name</span><span class="p">],</span> <span class="s2">&quot;MoE module needs to have the same quantization format for non-rounter sub-modules&quot;</span>
<span class="bp">self</span><span class="o">.</span><span class="n">quantization</span> <span class="o">=</span> <span class="n">LayerQuantConfig</span><span class="o">.</span><span class="n">from_dict</span><span class="p">(</span><span class="n">config</span><span class="p">)</span></div>
<span class="nd">@property</span>
<span class="k">def</span><span class="w"> </span><span class="nf">quant_mode</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
<span class="k">return</span> <span class="bp">self</span><span class="o">.</span><span class="n">quantization</span><span class="o">.</span><span class="n">quant_mode</span>
<span class="nd">@property</span>
<span class="k">def</span><span class="w"> </span><span class="nf">quant_algo</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
<span class="k">return</span> <span class="bp">self</span><span class="o">.</span><span class="n">quantization</span><span class="o">.</span><span class="n">quant_algo</span>
<span class="k">def</span><span class="w"> </span><span class="nf">_get_quant_cfg</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">module_name</span><span class="p">:</span> <span class="nb">str</span><span class="p">):</span>
<span class="k">return</span> <span class="bp">self</span><span class="o">.</span><span class="n">quantization</span><span class="o">.</span><span class="n">_get_quant_cfg</span><span class="p">(</span><span class="n">module_name</span><span class="p">)</span>
<div class="viewcode-block" id="PretrainedConfig.set_rank">
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.models.html#tensorrt_llm.llmapi.PretrainedConfig.set_rank">[docs]</a>
<span class="k">def</span><span class="w"> </span><span class="nf">set_rank</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">rank</span><span class="p">:</span> <span class="nb">int</span><span class="p">):</span>
<span class="bp">self</span><span class="o">.</span><span class="n">mapping</span><span class="o">.</span><span class="n">rank</span> <span class="o">=</span> <span class="n">rank</span></div>
<div class="viewcode-block" id="PretrainedConfig.get_config_group">
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.models.html#tensorrt_llm.llmapi.PretrainedConfig.get_config_group">[docs]</a>
<span class="k">def</span><span class="w"> </span><span class="nf">get_config_group</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">group_cls</span><span class="p">:</span> <span class="s2">&quot;Type[CG]&quot;</span><span class="p">)</span> <span class="o">-&gt;</span> <span class="s2">&quot;CG&quot;</span><span class="p">:</span>
<span class="n">cfg</span> <span class="o">=</span> <span class="p">{</span><span class="n">k</span><span class="p">:</span> <span class="n">v</span> <span class="k">for</span> <span class="n">k</span><span class="p">,</span> <span class="n">v</span> <span class="ow">in</span> <span class="bp">self</span><span class="o">.</span><span class="n">to_dict</span><span class="p">()</span><span class="o">.</span><span class="n">items</span><span class="p">()</span> <span class="k">if</span> <span class="n">k</span> <span class="ow">in</span> <span class="n">group_cls</span><span class="o">.</span><span class="n">keys</span><span class="p">()}</span>
<span class="k">return</span> <span class="n">group_cls</span><span class="p">(</span><span class="o">**</span><span class="n">cfg</span><span class="p">)</span></div>
<div class="viewcode-block" id="PretrainedConfig.has_config_group">
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.models.html#tensorrt_llm.llmapi.PretrainedConfig.has_config_group">[docs]</a>
<span class="k">def</span><span class="w"> </span><span class="nf">has_config_group</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">group_cls</span><span class="p">:</span> <span class="s2">&quot;Type[CG]&quot;</span><span class="p">)</span> <span class="o">-&gt;</span> <span class="s2">&quot;bool&quot;</span><span class="p">:</span>
<span class="k">return</span> <span class="nb">all</span><span class="p">(</span><span class="nb">hasattr</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">key</span><span class="p">)</span> <span class="k">for</span> <span class="n">key</span> <span class="ow">in</span> <span class="n">group_cls</span><span class="o">.</span><span class="n">keys</span><span class="p">())</span></div>
<div class="viewcode-block" id="PretrainedConfig.for_each_rank">
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.models.html#tensorrt_llm.llmapi.PretrainedConfig.for_each_rank">[docs]</a>
<span class="k">def</span><span class="w"> </span><span class="nf">for_each_rank</span><span class="p">(</span><span class="bp">self</span><span class="p">)</span> <span class="o">-&gt;</span> <span class="s2">&quot;Generator[Self, None, None]&quot;</span><span class="p">:</span>
<span class="k">for</span> <span class="n">rank</span> <span class="ow">in</span> <span class="nb">range</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">mapping</span><span class="o">.</span><span class="n">world_size</span><span class="p">):</span>
<span class="n">config_copy</span> <span class="o">=</span> <span class="n">copy</span><span class="o">.</span><span class="n">deepcopy</span><span class="p">(</span><span class="bp">self</span><span class="p">)</span>
<span class="n">config_copy</span><span class="o">.</span><span class="n">set_rank</span><span class="p">(</span><span class="n">rank</span><span class="p">)</span>
<span class="k">yield</span> <span class="n">config_copy</span></div>
</div>
<span class="k">class</span><span class="w"> </span><span class="nc">DecoderLayerList</span><span class="p">(</span><span class="n">ModuleList</span><span class="p">):</span>
<span class="k">def</span><span class="w"> </span><span class="fm">__init__</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="bp">cls</span><span class="p">,</span> <span class="n">config</span><span class="p">):</span>
<span class="bp">self</span><span class="o">.</span><span class="n">num_hidden_layers</span> <span class="o">=</span> <span class="n">config</span><span class="o">.</span><span class="n">num_hidden_layers</span>
<span class="bp">self</span><span class="o">.</span><span class="n">layer_list</span> <span class="o">=</span> <span class="n">config</span><span class="o">.</span><span class="n">mapping</span><span class="o">.</span><span class="n">pp_layers</span><span class="p">(</span><span class="n">config</span><span class="o">.</span><span class="n">num_hidden_layers</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">config</span><span class="o">.</span><span class="n">quant_mode</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">cls</span><span class="p">(</span><span class="n">config</span><span class="p">,</span> <span class="n">idx</span><span class="p">)</span> <span class="k">for</span> <span class="n">idx</span> <span class="ow">in</span> <span class="bp">self</span><span class="o">.</span><span class="n">layer_list</span><span class="p">])</span>
<span class="k">def</span><span class="w"> </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">use_cache</span><span class="o">=</span><span class="kc">False</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">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">mrope_params</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span>
<span class="n">position_ids</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span>
<span class="n">lora_params</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span>
<span class="n">spec_decoding_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">kv_cache_params</span><span class="o">.</span><span class="n">fill_none_tensor_list</span><span class="p">(</span><span class="nb">len</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">layer_list</span><span class="p">))</span>
<span class="k">if</span> <span class="n">use_cache</span><span class="p">:</span>
<span class="n">presents</span> <span class="o">=</span> <span class="p">[]</span>
<span class="k">for</span> <span class="n">layer_idx</span><span class="p">,</span> <span class="p">(</span><span class="n">layer</span><span class="p">,</span> <span class="n">past</span><span class="p">)</span> <span class="ow">in</span> <span class="nb">enumerate</span><span class="p">(</span>
<span class="nb">zip</span><span class="p">(</span><span class="bp">self</span><span class="p">,</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="n">lora_layer_params</span> <span class="o">=</span> <span class="kc">None</span>
<span class="k">if</span> <span class="n">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">lora_params</span><span class="o">.</span><span class="n">lora_ranks</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
<span class="n">lora_layer_params</span> <span class="o">=</span> <span class="n">lora_params</span><span class="o">.</span><span class="n">get_layer_params</span><span class="p">(</span><span class="n">layer_idx</span><span class="p">)</span>
<span class="n">kwargs</span> <span class="o">=</span> <span class="p">{}</span>
<span class="k">if</span> <span class="n">position_ids</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
<span class="n">kwargs</span><span class="p">[</span><span class="s1">&#39;position_ids&#39;</span><span class="p">]</span> <span class="o">=</span> <span class="n">position_ids</span>
<span class="k">if</span> <span class="n">vision_token_mask</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
<span class="n">kwargs</span><span class="p">[</span><span class="s1">&#39;vision_token_mask&#39;</span><span class="p">]</span> <span class="o">=</span> <span class="n">vision_token_mask</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">kwargs</span><span class="p">[</span><span class="s1">&#39;lora_layer_params&#39;</span><span class="p">]</span> <span class="o">=</span> <span class="n">lora_layer_params</span>
<span class="k">if</span> <span class="n">spec_decoding_params</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
<span class="n">kwargs</span><span class="p">[</span><span class="s1">&#39;spec_decoding_params&#39;</span><span class="p">]</span> <span class="o">=</span> <span class="n">spec_decoding_params</span>
<span class="k">if</span> <span class="n">mrope_params</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
<span class="n">kwargs</span><span class="p">[</span><span class="s1">&#39;mrope_params&#39;</span><span class="p">]</span> <span class="o">=</span> <span class="n">mrope_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">reduce_fusion</span><span class="p">:</span>
<span class="k">if</span> <span class="n">layer_idx</span> <span class="o">+</span> <span class="bp">self</span><span class="o">.</span><span class="n">layer_list</span><span class="p">[</span><span class="mi">0</span><span class="p">]</span> <span class="o">&lt;</span> <span class="bp">self</span><span class="o">.</span><span class="n">layer_list</span><span class="p">[</span><span class="o">-</span><span class="mi">1</span><span class="p">]:</span>
<span class="n">qkv_activation_scaling_factor</span> <span class="o">=</span> <span class="kc">None</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">user_buffer</span><span class="p">:</span>
<span class="n">qkv_linear</span> <span class="o">=</span> <span class="bp">self</span><span class="p">[</span><span class="n">layer_idx</span> <span class="o">+</span> <span class="mi">1</span><span class="p">]</span><span class="o">.</span><span class="n">attention</span><span class="o">.</span><span class="n">qkv</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_qdq</span><span class="p">():</span>
<span class="n">qkv_activation_scaling_factor</span> <span class="o">=</span> <span class="n">constant</span><span class="p">(</span>
<span class="n">qkv_linear</span><span class="o">.</span><span class="n">activation_scaling_factor</span><span class="o">.</span><span class="n">raw_value</span><span class="o">.</span>
<span class="n">copy</span><span class="p">())</span>
<span class="k">elif</span> <span class="bp">self</span><span class="o">.</span><span class="n">quant_mode</span><span class="o">.</span><span class="n">has_nvfp4</span><span class="p">():</span>
<span class="n">qkv_activation_scaling_factor</span> <span class="o">=</span> <span class="n">constant</span><span class="p">(</span>
<span class="n">qkv_linear</span><span class="o">.</span><span class="n">activation_global_scaling_factor</span><span class="o">.</span>
<span class="n">raw_value</span><span class="o">.</span><span class="n">copy</span><span class="p">())</span>
<span class="n">kwargs</span><span class="p">[</span><span class="s1">&#39;next_layer_input_layernorm_args&#39;</span><span class="p">]</span> <span class="o">=</span> <span class="p">(</span>
<span class="bp">self</span><span class="p">[</span><span class="n">layer_idx</span> <span class="o">+</span> <span class="mi">1</span><span class="p">]</span><span class="o">.</span><span class="n">input_layernorm</span><span class="o">.</span><span class="n">weight</span><span class="o">.</span><span class="n">value</span><span class="p">,</span>
<span class="bp">self</span><span class="p">[</span><span class="n">layer_idx</span> <span class="o">+</span> <span class="mi">1</span><span class="p">]</span><span class="o">.</span><span class="n">input_layernorm</span><span class="o">.</span><span class="n">eps</span><span class="p">,</span>
<span class="n">qkv_activation_scaling_factor</span><span class="p">)</span>
<span class="k">else</span><span class="p">:</span>
<span class="n">kwargs</span><span class="p">[</span><span class="s1">&#39;next_layer_input_layernorm_args&#39;</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</span>
<span class="k">elif</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">norm_quant_fusion</span><span class="p">:</span>
<span class="k">if</span> <span class="n">layer_idx</span> <span class="o">&lt;</span> <span class="bp">self</span><span class="o">.</span><span class="n">layer_list</span><span class="p">[</span><span class="o">-</span><span class="mi">1</span><span class="p">]</span> <span class="o">-</span> <span class="bp">self</span><span class="o">.</span><span class="n">layer_list</span><span class="p">[</span><span class="mi">0</span><span class="p">]:</span>
<span class="k">try</span><span class="p">:</span>
<span class="n">activation_scaling_factor</span> <span class="o">=</span> <span class="n">constant</span><span class="p">(</span>
<span class="bp">self</span><span class="p">[</span><span class="n">layer_idx</span> <span class="o">+</span> <span class="mi">1</span><span class="p">]</span><span class="o">.</span><span class="n">attention</span><span class="o">.</span><span class="n">qkv</span><span class="o">.</span>
<span class="n">activation_global_scaling_factor</span><span class="o">.</span><span class="n">raw_value</span><span class="o">.</span><span class="n">copy</span><span class="p">())</span>
<span class="k">except</span><span class="p">:</span>
<span class="n">activation_scaling_factor</span> <span class="o">=</span> <span class="kc">None</span>
<span class="n">kwargs</span><span class="p">[</span><span class="s1">&#39;next_layer_input_layernorm_args&#39;</span><span class="p">]</span> <span class="o">=</span> <span class="p">(</span>
<span class="bp">self</span><span class="p">[</span><span class="n">layer_idx</span> <span class="o">+</span> <span class="mi">1</span><span class="p">]</span><span class="o">.</span><span class="n">input_layernorm</span><span class="o">.</span><span class="n">weight</span><span class="o">.</span><span class="n">value</span><span class="p">,</span>
<span class="bp">self</span><span class="p">[</span><span class="n">layer_idx</span> <span class="o">+</span> <span class="mi">1</span><span class="p">]</span><span class="o">.</span><span class="n">input_layernorm</span><span class="o">.</span><span class="n">eps</span><span class="p">,</span>
<span class="n">activation_scaling_factor</span><span class="p">)</span>
<span class="k">else</span><span class="p">:</span>
<span class="n">kwargs</span><span class="p">[</span><span class="s1">&#39;next_layer_input_layernorm_args&#39;</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</span>
<span class="n">hidden_states</span> <span class="o">=</span> <span class="n">layer</span><span class="p">(</span>
<span class="n">hidden_states</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">attention_mask</span><span class="o">=</span><span class="n">attention_mask</span><span class="p">,</span>
<span class="n">kv_cache_params</span><span class="o">=</span><span class="n">KeyValueCacheParams</span><span class="p">(</span>
<span class="n">past_key_value</span><span class="o">=</span><span class="p">[</span><span class="n">past</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">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">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">attention_params</span><span class="o">=</span><span class="n">attention_params</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">use_cache</span><span class="p">:</span>
<span class="n">presents</span><span class="o">.</span><span class="n">append</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">hidden_states</span> <span class="o">=</span> <span class="n">hidden_states</span><span class="p">[</span><span class="mi">0</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="n">hidden_states</span><span class="p">,</span> <span class="n">presents</span>
<span class="k">return</span> <span class="n">hidden_states</span>
<span class="k">class</span><span class="w"> </span><span class="nc">PostInitCaller</span><span class="p">(</span><span class="nb">type</span><span class="p">):</span>
<span class="k">def</span><span class="w"> </span><span class="fm">__call__</span><span class="p">(</span><span class="bp">cls</span><span class="p">,</span> <span class="o">*</span><span class="n">args</span><span class="p">,</span> <span class="o">**</span><span class="n">kwargs</span><span class="p">):</span>
<span class="n">obj</span> <span class="o">=</span> <span class="nb">type</span><span class="o">.</span><span class="fm">__call__</span><span class="p">(</span><span class="bp">cls</span><span class="p">,</span> <span class="o">*</span><span class="n">args</span><span class="p">,</span> <span class="o">**</span><span class="n">kwargs</span><span class="p">)</span>
<span class="n">obj</span><span class="o">.</span><span class="n">__post_init__</span><span class="p">()</span>
<span class="k">return</span> <span class="n">obj</span>
<div class="viewcode-block" id="PretrainedModel">
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.models.html#tensorrt_llm.llmapi.PretrainedModel">[docs]</a>
<span class="k">class</span><span class="w"> </span><span class="nc">PretrainedModel</span><span class="p">(</span><span class="n">Module</span><span class="p">,</span>
<span class="n">GenerationMixin</span><span class="p">,</span>
<span class="n">TopModelMixin</span><span class="p">,</span>
<span class="n">metaclass</span><span class="o">=</span><span class="n">PostInitCaller</span><span class="p">):</span>
<span class="k">def</span><span class="w"> </span><span class="fm">__init__</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">config</span><span class="p">:</span> <span class="n">PretrainedConfig</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">init_all_reduce_helper</span><span class="p">()</span>
<span class="bp">self</span><span class="o">.</span><span class="n">config</span> <span class="o">=</span> <span class="n">config</span>
<span class="k">def</span><span class="w"> </span><span class="nf">__post_init__</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
<span class="kn">from</span><span class="w"> </span><span class="nn">..quantization.quantize</span><span class="w"> </span><span class="kn">import</span> <span class="n">quantize</span>
<span class="n">quantize</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="bp">self</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">quantization</span><span class="p">)</span>
<span class="c1"># Currently, use_parallel_embedding must be enabled before weight loading;</span>
<span class="c1"># otherwise, the model will be inconsistent with the weights loaded from checkpoint.</span>
<span class="n">optimize_model</span><span class="p">(</span>
<span class="bp">self</span><span class="p">,</span> <span class="n">use_parallel_embedding</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">use_parallel_embedding</span><span class="p">)</span>
<div class="viewcode-block" id="PretrainedModel.release">
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.models.html#tensorrt_llm.llmapi.PretrainedModel.release">[docs]</a>
<span class="k">def</span><span class="w"> </span><span class="nf">release</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
<span class="n">release_gc</span><span class="p">()</span></div>
<span class="k">def</span><span class="w"> </span><span class="fm">__del__</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
<span class="bp">self</span><span class="o">.</span><span class="n">release</span><span class="p">()</span>
<div class="viewcode-block" id="PretrainedModel.check_config">
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.models.html#tensorrt_llm.llmapi.PretrainedModel.check_config">[docs]</a>
<span class="k">def</span><span class="w"> </span><span class="nf">check_config</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">config</span><span class="p">):</span>
<span class="k">raise</span> <span class="ne">NotImplementedError</span><span class="p">(</span>
<span class="sa">f</span><span class="s2">&quot;</span><span class="si">{</span><span class="bp">self</span><span class="o">.</span><span class="vm">__class__</span><span class="si">}</span><span class="s2"> is an abstract class. Only classes inheriting this class can be called.&quot;</span>
<span class="p">)</span></div>
<div class="viewcode-block" id="PretrainedModel.from_config">
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.models.html#tensorrt_llm.llmapi.PretrainedModel.from_config">[docs]</a>
<span class="nd">@classmethod</span>
<span class="k">def</span><span class="w"> </span><span class="nf">from_config</span><span class="p">(</span><span class="bp">cls</span><span class="p">,</span> <span class="n">config</span><span class="p">:</span> <span class="n">PretrainedConfig</span><span class="p">):</span>
<span class="k">return</span> <span class="bp">cls</span><span class="p">(</span><span class="n">config</span><span class="p">)</span></div>
<div class="viewcode-block" id="PretrainedModel.from_checkpoint">
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.models.html#tensorrt_llm.llmapi.PretrainedModel.from_checkpoint">[docs]</a>
<span class="nd">@classmethod</span>
<span class="k">def</span><span class="w"> </span><span class="nf">from_checkpoint</span><span class="p">(</span>
<span class="bp">cls</span><span class="p">,</span>
<span class="n">ckpt_dir</span><span class="p">:</span> <span class="nb">str</span><span class="p">,</span>
<span class="n">rank</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="nb">int</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</span><span class="p">,</span>
<span class="n">config</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="n">PretrainedConfig</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</span><span class="p">,</span>
<span class="o">*</span><span class="p">,</span>
<span class="n">preprocess_weights_hook</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="n">Callable</span><span class="p">[[</span><span class="n">Dict</span><span class="p">[</span><span class="nb">str</span><span class="p">,</span> <span class="n">Tensor</span><span class="p">]],</span>
<span class="n">Dict</span><span class="p">[</span><span class="nb">str</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="k">if</span> <span class="n">config</span> <span class="ow">is</span> <span class="kc">None</span><span class="p">:</span>
<span class="n">config</span> <span class="o">=</span> <span class="n">PretrainedConfig</span><span class="o">.</span><span class="n">from_json_file</span><span class="p">(</span>
<span class="n">os</span><span class="o">.</span><span class="n">path</span><span class="o">.</span><span class="n">join</span><span class="p">(</span><span class="n">ckpt_dir</span><span class="p">,</span> <span class="s1">&#39;config.json&#39;</span><span class="p">))</span>
<span class="k">if</span> <span class="n">rank</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
<span class="n">config</span><span class="o">.</span><span class="n">set_rank</span><span class="p">(</span><span class="n">rank</span><span class="p">)</span>
<span class="n">rank</span> <span class="o">=</span> <span class="n">config</span><span class="o">.</span><span class="n">mapping</span><span class="o">.</span><span class="n">rank</span>
<span class="k">if</span> <span class="n">config</span><span class="o">.</span><span class="n">mapping</span><span class="o">.</span><span class="n">auto_parallel</span><span class="p">:</span>
<span class="n">rank</span> <span class="o">=</span> <span class="mi">0</span>
<span class="k">elif</span> <span class="n">config</span><span class="o">.</span><span class="n">mapping</span><span class="o">.</span><span class="n">cp_size</span> <span class="o">&gt;</span> <span class="mi">1</span><span class="p">:</span>
<span class="c1"># tp_cp_pp rank -&gt; tp_pp rank: because different cp ranks share the same ckpt</span>
<span class="n">tp_size</span> <span class="o">=</span> <span class="n">config</span><span class="o">.</span><span class="n">mapping</span><span class="o">.</span><span class="n">tp_size</span>
<span class="n">cp_size</span> <span class="o">=</span> <span class="n">config</span><span class="o">.</span><span class="n">mapping</span><span class="o">.</span><span class="n">cp_size</span>
<span class="n">rank</span> <span class="o">=</span> <span class="n">rank</span> <span class="o">%</span> <span class="n">tp_size</span> <span class="o">+</span> <span class="n">rank</span> <span class="o">//</span> <span class="p">(</span><span class="n">tp_size</span> <span class="o">*</span> <span class="n">cp_size</span><span class="p">)</span> <span class="o">*</span> <span class="n">tp_size</span>
<span class="n">weights_path</span> <span class="o">=</span> <span class="n">os</span><span class="o">.</span><span class="n">path</span><span class="o">.</span><span class="n">join</span><span class="p">(</span><span class="n">ckpt_dir</span><span class="p">,</span> <span class="sa">f</span><span class="s1">&#39;rank</span><span class="si">{</span><span class="n">rank</span><span class="si">}</span><span class="s1">.safetensors&#39;</span><span class="p">)</span>
<span class="k">assert</span> <span class="n">os</span><span class="o">.</span><span class="n">path</span><span class="o">.</span><span class="n">isfile</span><span class="p">(</span><span class="n">weights_path</span><span class="p">)</span>
<span class="n">weights</span> <span class="o">=</span> <span class="n">safetensors</span><span class="o">.</span><span class="n">torch</span><span class="o">.</span><span class="n">load_file</span><span class="p">(</span><span class="n">weights_path</span><span class="p">)</span>
<span class="n">is_checkpoint_pruned</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">&#39;is_pruned&#39;</span><span class="p">,</span> <span class="kc">False</span><span class="p">)</span>
<span class="k">if</span> <span class="n">preprocess_weights_hook</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
<span class="n">weights</span> <span class="o">=</span> <span class="n">preprocess_weights_hook</span><span class="p">(</span><span class="n">weights</span><span class="p">)</span>
<span class="n">weights</span> <span class="o">=</span> <span class="n">preprocess_weights</span><span class="p">(</span><span class="n">weights</span><span class="p">,</span>
<span class="n">config</span><span class="p">,</span>
<span class="n">from_pruned</span><span class="o">=</span><span class="n">is_checkpoint_pruned</span><span class="p">)</span>
<span class="n">model</span> <span class="o">=</span> <span class="bp">cls</span><span class="p">(</span><span class="n">config</span><span class="p">)</span>
<span class="n">model</span><span class="o">.</span><span class="n">load</span><span class="p">(</span><span class="n">weights</span><span class="p">,</span> <span class="n">from_pruned</span><span class="o">=</span><span class="n">is_checkpoint_pruned</span><span class="p">)</span>
<span class="k">return</span> <span class="n">model</span></div>
<div class="viewcode-block" id="PretrainedModel.load">
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.models.html#tensorrt_llm.llmapi.PretrainedModel.load">[docs]</a>
<span class="k">def</span><span class="w"> </span><span class="nf">load</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">weights</span><span class="p">,</span> <span class="n">from_pruned</span><span class="o">=</span><span class="kc">False</span><span class="p">):</span>
<span class="n">required_names</span> <span class="o">=</span> <span class="nb">set</span><span class="p">()</span>
<span class="k">for</span> <span class="n">name</span><span class="p">,</span> <span class="n">param</span> <span class="ow">in</span> <span class="bp">self</span><span class="o">.</span><span class="n">named_parameters</span><span class="p">():</span>
<span class="k">if</span> <span class="n">param</span><span class="o">.</span><span class="n">is_inited</span><span class="p">():</span>
<span class="k">continue</span>
<span class="k">if</span> <span class="n">name</span> <span class="ow">not</span> <span class="ow">in</span> <span class="n">weights</span><span class="p">:</span>
<span class="c1"># Exemption for embedding sharing</span>
<span class="k">if</span> <span class="n">name</span><span class="o">.</span><span class="n">endswith</span><span class="p">(</span><span class="s1">&#39;lm_head.weight&#39;</span><span class="p">)</span> <span class="ow">and</span> <span class="nb">any</span><span class="p">(</span>
<span class="n">k</span><span class="o">.</span><span class="n">endswith</span><span class="p">(</span><span class="s1">&#39;vocab_embedding.weight&#39;</span><span class="p">)</span>
<span class="k">for</span> <span class="n">k</span> <span class="ow">in</span> <span class="n">weights</span><span class="o">.</span><span class="n">keys</span><span class="p">()):</span>
<span class="k">continue</span>
<span class="k">if</span> <span class="n">name</span><span class="o">.</span><span class="n">endswith</span><span class="p">(</span><span class="s1">&#39;lm_head.per_channel_scale&#39;</span><span class="p">)</span> <span class="ow">and</span> <span class="nb">any</span><span class="p">(</span>
<span class="n">k</span><span class="o">.</span><span class="n">endswith</span><span class="p">(</span><span class="s1">&#39;vocab_embedding.per_channel_scale&#39;</span><span class="p">)</span>
<span class="k">for</span> <span class="n">k</span> <span class="ow">in</span> <span class="n">weights</span><span class="o">.</span><span class="n">keys</span><span class="p">()):</span>
<span class="k">continue</span>
<span class="n">required_names</span><span class="o">.</span><span class="n">add</span><span class="p">(</span><span class="n">name</span><span class="p">)</span>
<span class="n">provided_names</span> <span class="o">=</span> <span class="nb">set</span><span class="p">(</span><span class="n">weights</span><span class="o">.</span><span class="n">keys</span><span class="p">())</span>
<span class="k">if</span> <span class="ow">not</span> <span class="n">required_names</span><span class="o">.</span><span class="n">issubset</span><span class="p">(</span><span class="n">provided_names</span><span class="p">):</span>
<span class="k">raise</span> <span class="ne">RuntimeError</span><span class="p">(</span>
<span class="sa">f</span><span class="s2">&quot;Required but not provided tensors:</span><span class="si">{</span><span class="n">required_names</span><span class="o">.</span><span class="n">difference</span><span class="p">(</span><span class="n">provided_names</span><span class="p">)</span><span class="si">}</span><span class="s2">&quot;</span>
<span class="p">)</span>
<span class="k">if</span> <span class="ow">not</span> <span class="n">provided_names</span><span class="o">.</span><span class="n">issubset</span><span class="p">(</span><span class="n">required_names</span><span class="p">):</span>
<span class="n">logger</span><span class="o">.</span><span class="n">warning</span><span class="p">(</span>
<span class="sa">f</span><span class="s2">&quot;Provided but not required tensors: </span><span class="si">{</span><span class="n">provided_names</span><span class="o">.</span><span class="n">difference</span><span class="p">(</span><span class="n">required_names</span><span class="p">)</span><span class="si">}</span><span class="s2">&quot;</span>
<span class="p">)</span>
<span class="k">for</span> <span class="n">name</span><span class="p">,</span> <span class="n">param</span> <span class="ow">in</span> <span class="bp">self</span><span class="o">.</span><span class="n">named_parameters</span><span class="p">():</span>
<span class="k">if</span> <span class="n">name</span> <span class="ow">in</span> <span class="n">provided_names</span><span class="p">:</span>
<span class="k">if</span> <span class="ow">not</span> <span class="n">from_pruned</span><span class="p">:</span>
<span class="k">try</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">weights</span><span class="p">[</span><span class="n">name</span><span class="p">]</span>
<span class="k">except</span> <span class="ne">Exception</span> <span class="k">as</span> <span class="n">e</span><span class="p">:</span>
<span class="k">raise</span> <span class="ne">RuntimeError</span><span class="p">(</span>
<span class="sa">f</span><span class="s2">&quot;Encounter error &#39;</span><span class="si">{</span><span class="n">e</span><span class="si">}</span><span class="s2">&#39; for parameter &#39;</span><span class="si">{</span><span class="n">name</span><span class="si">}</span><span class="s2">&#39;&quot;</span><span class="p">)</span>
<span class="k">else</span><span class="p">:</span>
<span class="n">param</span><span class="o">.</span><span class="n">set_value_or_dummy</span><span class="p">(</span><span class="n">weights</span><span class="p">[</span><span class="n">name</span><span class="p">])</span></div>
<div class="viewcode-block" id="PretrainedModel.save_checkpoint">
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.models.html#tensorrt_llm.llmapi.PretrainedModel.save_checkpoint">[docs]</a>
<span class="k">def</span><span class="w"> </span><span class="nf">save_checkpoint</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">output_dir</span><span class="p">,</span> <span class="n">save_config</span><span class="o">=</span><span class="kc">True</span><span class="p">):</span>
<span class="c1"># multiple ranks could share same config.json, so adding a save_config parameter to let user avoiding writing config.json in all ranks</span>
<span class="n">rank</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">mapping</span><span class="o">.</span><span class="n">rank</span>
<span class="n">weights</span> <span class="o">=</span> <span class="p">{</span>
<span class="n">name</span><span class="p">:</span> <span class="n">numpy_to_torch</span><span class="p">(</span><span class="n">param</span><span class="o">.</span><span class="n">raw_value</span><span class="p">)</span>
<span class="k">for</span> <span class="n">name</span><span class="p">,</span> <span class="n">param</span> <span class="ow">in</span> <span class="bp">self</span><span class="o">.</span><span class="n">named_parameters</span><span class="p">()</span>
<span class="p">}</span>
<span class="c1"># If there are some tensors share memory, this will lead to error when we call &quot;save_file&quot;. So, for repeated tensors, we</span>
<span class="c1"># clone the tensors to prevent this issue.</span>
<span class="n">data_ptrs</span> <span class="o">=</span> <span class="nb">set</span><span class="p">()</span>
<span class="k">for</span> <span class="n">name</span><span class="p">,</span> <span class="n">param</span> <span class="ow">in</span> <span class="n">weights</span><span class="o">.</span><span class="n">items</span><span class="p">():</span>
<span class="k">if</span> <span class="n">param</span><span class="o">.</span><span class="n">data_ptr</span><span class="p">()</span> <span class="ow">in</span> <span class="n">data_ptrs</span><span class="p">:</span>
<span class="n">weights</span><span class="p">[</span><span class="n">name</span><span class="p">]</span> <span class="o">=</span> <span class="n">param</span><span class="o">.</span><span class="n">clone</span><span class="p">()</span>
<span class="n">data_ptrs</span><span class="o">.</span><span class="n">add</span><span class="p">(</span><span class="n">weights</span><span class="p">[</span><span class="n">name</span><span class="p">]</span><span class="o">.</span><span class="n">data_ptr</span><span class="p">())</span>
<span class="n">safetensors</span><span class="o">.</span><span class="n">torch</span><span class="o">.</span><span class="n">save_file</span><span class="p">(</span>
<span class="n">weights</span><span class="p">,</span> <span class="n">os</span><span class="o">.</span><span class="n">path</span><span class="o">.</span><span class="n">join</span><span class="p">(</span><span class="n">output_dir</span><span class="p">,</span> <span class="sa">f</span><span class="s1">&#39;rank</span><span class="si">{</span><span class="n">rank</span><span class="si">}</span><span class="s1">.safetensors&#39;</span><span class="p">))</span>
<span class="k">if</span> <span class="n">save_config</span><span class="p">:</span>
<span class="bp">self</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">to_json_file</span><span class="p">(</span><span class="n">os</span><span class="o">.</span><span class="n">path</span><span class="o">.</span><span class="n">join</span><span class="p">(</span><span class="n">output_dir</span><span class="p">,</span> <span class="s1">&#39;config.json&#39;</span><span class="p">))</span></div>
<div class="viewcode-block" id="PretrainedModel.prepare_inputs">
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.models.html#tensorrt_llm.llmapi.PretrainedModel.prepare_inputs">[docs]</a>
<span class="k">def</span><span class="w"> </span><span class="nf">prepare_inputs</span><span class="p">(</span>
<span class="bp">self</span><span class="p">,</span>
<span class="n">max_batch_size</span><span class="p">,</span>
<span class="n">max_input_len</span><span class="p">,</span>
<span class="n">max_seq_len</span><span class="p">,</span>
<span class="n">max_num_tokens</span><span class="p">,</span>
<span class="n">use_cache</span><span class="p">,</span>
<span class="n">max_beam_width</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">opt_num_tokens</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">prompt_embedding_table_size</span><span class="p">:</span> <span class="nb">int</span> <span class="o">=</span> <span class="mi">0</span><span class="p">,</span>
<span class="n">position_encoding_2d</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">max_draft_len</span><span class="p">:</span> <span class="nb">int</span> <span class="o">=</span> <span class="mi">0</span><span class="p">,</span>
<span class="n">speculative_decoding_draft_tokens_external</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_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">gather_context_logits</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">lora_target_modules</span><span class="p">:</span> <span class="n">List</span><span class="p">[</span><span class="nb">str</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</span><span class="p">,</span>
<span class="n">opt_batch_size</span><span class="p">:</span> <span class="nb">int</span> <span class="o">=</span> <span class="mi">0</span><span class="p">,</span>
<span class="n">num_hidden_layers</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">mrope_rotary_cos_sin_size</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="p">):</span>
<span class="w"> </span><span class="sd">&#39;&#39;&#39;@brief: Prepare inputs Tensors for the model, the given sizes are used to determine the</span>
<span class="sd"> ranges of the dimensions of when using TRT dynamic shapes.</span>
<span class="sd"> @return: a list contains values which can be fed into the self.forward()</span>
<span class="sd"> &#39;&#39;&#39;</span>
<span class="c1"># Prepare inputs</span>
<span class="n">remove_input_padding</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">remove_input_padding</span>
<span class="n">use_gpt_attention_plugin</span> <span class="o">=</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">gpt_attention_plugin</span>
<span class="n">use_gemm_plugin</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">gemm_plugin</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="n">tokens_per_block</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">tokens_per_block</span>
<span class="n">use_lora_plugin</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">lora_plugin</span>
<span class="n">multiple_profiles</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">multiple_profiles</span>
<span class="n">streamingllm</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">streamingllm</span>
<span class="n">pp_reduce_scatter</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">pp_reduce_scatter</span>
<span class="n">kv_cache_type</span> <span class="o">=</span> <span class="kc">None</span>
<span class="k">if</span> <span class="ow">not</span> <span class="n">use_cache</span><span class="p">:</span>
<span class="n">kv_cache_type</span> <span class="o">=</span> <span class="n">KVCacheType</span><span class="o">.</span><span class="n">DISABLED</span>
<span class="k">else</span><span class="p">:</span>
<span class="k">if</span> <span class="n">paged_kv_cache</span><span class="p">:</span>
<span class="n">kv_cache_type</span> <span class="o">=</span> <span class="n">KVCacheType</span><span class="o">.</span><span class="n">PAGED</span>
<span class="k">else</span><span class="p">:</span>
<span class="n">kv_cache_type</span> <span class="o">=</span> <span class="n">KVCacheType</span><span class="o">.</span><span class="n">CONTINUOUS</span>
<span class="n">model_inputs</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">prepare_basic_inputs</span><span class="p">(</span>
<span class="n">max_batch_size</span><span class="o">=</span><span class="n">max_batch_size</span><span class="p">,</span>
<span class="n">max_beam_width</span><span class="o">=</span><span class="n">max_beam_width</span><span class="p">,</span>
<span class="n">max_input_len</span><span class="o">=</span><span class="n">max_input_len</span><span class="p">,</span>
<span class="n">max_seq_len</span><span class="o">=</span><span class="n">max_seq_len</span><span class="p">,</span>
<span class="n">hidden_size</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">hidden_size</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">config</span><span class="o">.</span><span class="n">num_key_value_heads</span><span class="p">,</span>
<span class="n">head_size</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">head_size</span><span class="p">,</span>
<span class="n">num_layers</span><span class="o">=</span><span class="n">num_hidden_layers</span>
<span class="k">if</span> <span class="n">num_hidden_layers</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">config</span><span class="o">.</span><span class="n">num_hidden_layers</span><span class="p">,</span>
<span class="n">kv_dtype</span><span class="o">=</span><span class="n">str_dtype_to_trt</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">kv_dtype</span><span class="p">),</span>
<span class="n">remove_input_padding</span><span class="o">=</span><span class="n">remove_input_padding</span><span class="p">,</span>
<span class="n">use_gpt_attention_plugin</span><span class="o">=</span><span class="n">use_gpt_attention_plugin</span><span class="p">,</span>
<span class="n">use_gemm_plugin</span><span class="o">=</span><span class="n">use_gemm_plugin</span><span class="p">,</span>
<span class="n">kv_cache_type</span><span class="o">=</span><span class="n">kv_cache_type</span><span class="p">,</span>
<span class="n">tokens_per_block</span><span class="o">=</span><span class="n">tokens_per_block</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">config</span><span class="o">.</span><span class="n">num_attention_heads</span><span class="p">,</span>
<span class="n">max_num_tokens</span><span class="o">=</span><span class="n">max_num_tokens</span><span class="p">,</span>
<span class="n">opt_num_tokens</span><span class="o">=</span><span class="n">opt_num_tokens</span><span class="p">,</span>
<span class="n">dtype</span><span class="o">=</span><span class="n">str_dtype_to_trt</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">dtype</span><span class="p">),</span>
<span class="n">prompt_embedding_table_size</span><span class="o">=</span><span class="n">prompt_embedding_table_size</span><span class="p">,</span>
<span class="n">position_encoding_2d</span><span class="o">=</span><span class="n">position_encoding_2d</span><span class="p">,</span>
<span class="n">mapping</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">mapping</span><span class="p">,</span>
<span class="n">gather_context_logits</span><span class="o">=</span><span class="n">gather_context_logits</span><span class="p">,</span>
<span class="n">use_lora_plugin</span><span class="o">=</span><span class="n">use_lora_plugin</span><span class="p">,</span>
<span class="n">max_draft_len</span><span class="o">=</span><span class="n">max_draft_len</span><span class="p">,</span>
<span class="n">speculative_decoding_draft_tokens_external</span><span class="o">=</span>
<span class="n">speculative_decoding_draft_tokens_external</span><span class="p">,</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="p">,</span>
<span class="n">lora_target_modules</span><span class="o">=</span><span class="n">lora_target_modules</span><span class="p">,</span>
<span class="n">multiple_profiles</span><span class="o">=</span><span class="n">multiple_profiles</span><span class="p">,</span>
<span class="n">streamingllm</span><span class="o">=</span><span class="n">streamingllm</span><span class="p">,</span>
<span class="n">opt_batch_size</span><span class="o">=</span><span class="n">opt_batch_size</span><span class="p">,</span>
<span class="n">pp_reduce_scatter</span><span class="o">=</span><span class="n">pp_reduce_scatter</span><span class="p">,</span>
<span class="n">mrope_rotary_cos_sin_size</span><span class="o">=</span><span class="n">mrope_rotary_cos_sin_size</span><span class="p">)</span>
<span class="n">result</span> <span class="o">=</span> <span class="p">{</span>
<span class="s1">&#39;input_ids&#39;</span><span class="p">:</span>
<span class="n">model_inputs</span><span class="p">[</span><span class="s1">&#39;input_ids&#39;</span><span class="p">],</span>
<span class="s1">&#39;position_ids&#39;</span><span class="p">:</span>
<span class="n">model_inputs</span><span class="p">[</span><span class="s1">&#39;position_ids&#39;</span><span class="p">],</span>
<span class="s1">&#39;use_cache&#39;</span><span class="p">:</span>
<span class="n">kv_cache_type</span> <span class="o">!=</span> <span class="n">KVCacheType</span><span class="o">.</span><span class="n">DISABLED</span><span class="p">,</span>
<span class="s1">&#39;last_token_ids&#39;</span><span class="p">:</span>
<span class="n">model_inputs</span><span class="p">[</span><span class="s1">&#39;last_token_ids&#39;</span><span class="p">],</span>
<span class="s1">&#39;attention_mask&#39;</span><span class="p">:</span>
<span class="n">model_inputs</span><span class="p">[</span><span class="s1">&#39;attention_mask&#39;</span><span class="p">],</span>
<span class="s1">&#39;kv_cache_params&#39;</span><span class="p">:</span>
<span class="n">KeyValueCacheParams</span><span class="p">(</span>
<span class="n">past_key_value</span><span class="o">=</span><span class="n">model_inputs</span><span class="p">[</span><span class="s1">&#39;past_key_value&#39;</span><span class="p">],</span>
<span class="n">host_past_key_value_lengths</span><span class="o">=</span><span class="n">model_inputs</span><span class="p">[</span>
<span class="s1">&#39;host_past_key_value_lengths&#39;</span><span class="p">],</span>
<span class="n">host_max_attention_window_sizes</span><span class="o">=</span><span class="n">model_inputs</span><span class="p">[</span>
<span class="s1">&#39;host_max_attention_window_sizes&#39;</span><span class="p">],</span>
<span class="n">host_sink_token_length</span><span class="o">=</span><span class="n">model_inputs</span><span class="p">[</span><span class="s1">&#39;host_sink_token_length&#39;</span><span class="p">],</span>
<span class="n">kv_cache_block_offsets</span><span class="o">=</span><span class="n">model_inputs</span><span class="p">[</span><span class="s1">&#39;kv_cache_block_offsets&#39;</span><span class="p">],</span>
<span class="n">host_kv_cache_block_offsets</span><span class="o">=</span><span class="n">model_inputs</span><span class="p">[</span>
<span class="s1">&#39;host_kv_cache_block_offsets&#39;</span><span class="p">],</span>
<span class="n">host_kv_cache_pool_pointers</span><span class="o">=</span><span class="n">model_inputs</span><span class="p">[</span>
<span class="s1">&#39;host_kv_cache_pool_pointers&#39;</span><span class="p">],</span>
<span class="n">host_kv_cache_pool_mapping</span><span class="o">=</span><span class="n">model_inputs</span><span class="p">[</span>
<span class="s1">&#39;host_kv_cache_pool_mapping&#39;</span><span class="p">],</span>
<span class="n">cache_indirection</span><span class="o">=</span><span class="n">model_inputs</span><span class="p">[</span><span class="s1">&#39;cache_indirection&#39;</span><span class="p">],</span>
<span class="p">),</span>
<span class="s1">&#39;attention_params&#39;</span><span class="p">:</span>
<span class="n">AttentionParams</span><span class="p">(</span>
<span class="n">sequence_length</span><span class="o">=</span><span class="n">model_inputs</span><span class="p">[</span><span class="s1">&#39;sequence_length&#39;</span><span class="p">],</span>
<span class="n">context_lengths</span><span class="o">=</span><span class="n">model_inputs</span><span class="p">[</span><span class="s1">&#39;context_lengths&#39;</span><span class="p">],</span>
<span class="n">host_context_lengths</span><span class="o">=</span><span class="n">model_inputs</span><span class="p">[</span><span class="s1">&#39;host_context_lengths&#39;</span><span class="p">],</span>
<span class="n">max_context_length</span><span class="o">=</span><span class="n">max_input_len</span><span class="p">,</span>
<span class="n">host_request_types</span><span class="o">=</span><span class="n">model_inputs</span><span class="p">[</span><span class="s1">&#39;host_request_types&#39;</span><span class="p">],</span>
<span class="n">host_runtime_perf_knobs</span><span class="o">=</span><span class="n">model_inputs</span><span class="p">[</span><span class="s1">&#39;host_runtime_perf_knobs&#39;</span><span class="p">],</span>
<span class="n">host_context_progress</span><span class="o">=</span><span class="n">model_inputs</span><span class="p">[</span><span class="s1">&#39;host_context_progress&#39;</span><span class="p">],</span>
<span class="p">)</span>
<span class="p">}</span>
<span class="k">if</span> <span class="n">prompt_embedding_table_size</span> <span class="o">&gt;</span> <span class="mi">0</span><span class="p">:</span>
<span class="n">result</span><span class="p">[</span><span class="s1">&#39;prompt_embedding_table&#39;</span><span class="p">]</span> <span class="o">=</span> <span class="n">model_inputs</span><span class="p">[</span>
<span class="s1">&#39;prompt_embedding_table&#39;</span><span class="p">]</span>
<span class="n">result</span><span class="p">[</span><span class="s1">&#39;prompt_tasks&#39;</span><span class="p">]</span> <span class="o">=</span> <span class="n">model_inputs</span><span class="p">[</span><span class="s1">&#39;tasks&#39;</span><span class="p">]</span>
<span class="n">result</span><span class="p">[</span><span class="s1">&#39;prompt_vocab_size&#39;</span><span class="p">]</span> <span class="o">=</span> <span class="n">model_inputs</span><span class="p">[</span><span class="s1">&#39;prompt_vocab_size&#39;</span><span class="p">]</span>
<span class="k">if</span> <span class="n">model_inputs</span><span class="p">[</span><span class="s1">&#39;hidden_states_input&#39;</span><span class="p">]</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
<span class="n">result</span><span class="p">[</span><span class="s1">&#39;hidden_states&#39;</span><span class="p">]</span> <span class="o">=</span> <span class="n">model_inputs</span><span class="p">[</span><span class="s1">&#39;hidden_states_input&#39;</span><span class="p">]</span>
<span class="k">if</span> <span class="n">use_lora_plugin</span><span class="p">:</span>
<span class="n">result</span><span class="p">[</span><span class="s1">&#39;lora_params&#39;</span><span class="p">]</span> <span class="o">=</span> <span class="n">LoraParams</span><span class="p">(</span>
<span class="n">model_inputs</span><span class="p">[</span><span class="s1">&#39;lora_ranks&#39;</span><span class="p">],</span>
<span class="n">model_inputs</span><span class="p">[</span><span class="s1">&#39;lora_weights_pointers&#39;</span><span class="p">],</span>
<span class="n">host_context_lengths</span><span class="o">=</span><span class="n">model_inputs</span><span class="p">[</span><span class="s1">&#39;host_context_lengths&#39;</span><span class="p">],</span>
<span class="n">host_request_types</span><span class="o">=</span><span class="n">model_inputs</span><span class="p">[</span><span class="s1">&#39;host_request_types&#39;</span><span class="p">])</span>
<span class="k">if</span> <span class="n">model_inputs</span><span class="p">[</span><span class="s1">&#39;spec_decoding_params&#39;</span><span class="p">]</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
<span class="n">result</span><span class="p">[</span><span class="s1">&#39;spec_decoding_params&#39;</span><span class="p">]</span> <span class="o">=</span> <span class="n">model_inputs</span><span class="p">[</span>
<span class="s1">&#39;spec_decoding_params&#39;</span><span class="p">]</span>
<span class="k">if</span> <span class="n">model_inputs</span><span class="p">[</span><span class="s1">&#39;mrope_params&#39;</span><span class="p">]</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
<span class="n">result</span><span class="p">[</span><span class="s1">&#39;mrope_params&#39;</span><span class="p">]</span> <span class="o">=</span> <span class="n">model_inputs</span><span class="p">[</span><span class="s1">&#39;mrope_params&#39;</span><span class="p">]</span>
<span class="k">return</span> <span class="n">result</span></div>
<div class="viewcode-block" id="PretrainedModel.quantize">
<a class="viewcode-back" href="../../../python-api/tensorrt_llm.models.html#tensorrt_llm.llmapi.PretrainedModel.quantize">[docs]</a>
<span class="nd">@classmethod</span>
<span class="k">def</span><span class="w"> </span><span class="nf">quantize</span><span class="p">(</span>
<span class="bp">cls</span><span class="p">,</span>
<span class="n">hf_model_dir</span><span class="p">:</span> <span class="nb">str</span><span class="p">,</span>
<span class="n">output_dir</span><span class="p">:</span> <span class="nb">str</span><span class="p">,</span>
<span class="n">dtype</span><span class="p">:</span> <span class="nb">str</span> <span class="o">=</span> <span class="s1">&#39;auto&#39;</span><span class="p">,</span>
<span class="n">mapping</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="n">Mapping</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</span><span class="p">,</span>
<span class="n">quant_config</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="n">QuantConfig</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</span><span class="p">,</span>
<span class="o">*</span><span class="p">,</span>
<span class="n">device</span><span class="p">:</span> <span class="nb">str</span> <span class="o">=</span> <span class="s1">&#39;cuda&#39;</span><span class="p">,</span>
<span class="n">calib_dataset</span><span class="p">:</span> <span class="nb">str</span> <span class="o">=</span> <span class="s1">&#39;cnn_dailymail&#39;</span><span class="p">,</span>
<span class="n">calib_batches</span><span class="p">:</span> <span class="nb">int</span> <span class="o">=</span> <span class="mi">512</span><span class="p">,</span>
<span class="n">calib_batch_size</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">calib_max_seq_length</span><span class="p">:</span> <span class="nb">int</span> <span class="o">=</span> <span class="mi">512</span><span class="p">,</span>
<span class="n">random_seed</span><span class="p">:</span> <span class="nb">int</span> <span class="o">=</span> <span class="mi">1234</span><span class="p">,</span>
<span class="n">tokenizer_max_seq_length</span><span class="p">:</span> <span class="nb">int</span> <span class="o">=</span> <span class="mi">2048</span><span class="p">,</span>
<span class="o">**</span><span class="n">kwargs</span><span class="p">,</span>
<span class="p">):</span>
<span class="n">config_cls</span> <span class="o">=</span> <span class="nb">getattr</span><span class="p">(</span><span class="bp">cls</span><span class="p">,</span> <span class="s1">&#39;config_class&#39;</span><span class="p">,</span> <span class="kc">None</span><span class="p">)</span>
<span class="k">if</span> <span class="n">config_cls</span> <span class="ow">is</span> <span class="kc">None</span><span class="p">:</span>
<span class="k">raise</span> <span class="ne">NotImplementedError</span><span class="p">(</span>
<span class="sa">f</span><span class="s2">&quot;</span><span class="si">{</span><span class="bp">cls</span><span class="o">.</span><span class="vm">__name__</span><span class="si">}</span><span class="s2"> has not implemented corresponding config class, which is needed for correct config parsing.&quot;</span>
<span class="p">)</span>
<span class="n">config</span><span class="p">:</span> <span class="n">PretrainedConfig</span> <span class="o">=</span> <span class="n">config_cls</span><span class="o">.</span><span class="n">from_hugging_face</span><span class="p">(</span>
<span class="n">hf_model_dir</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">mapping</span><span class="o">=</span><span class="n">mapping</span><span class="p">,</span>
<span class="n">quant_config</span><span class="o">=</span><span class="n">quant_config</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">config</span><span class="o">.</span><span class="n">mapping</span><span class="o">.</span><span class="n">moe_ep_size</span> <span class="o">&gt;</span> <span class="mi">1</span><span class="p">:</span>
<span class="k">raise</span> <span class="ne">NotImplementedError</span><span class="p">(</span>
<span class="s2">&quot;Quantization for expert parallelism is not supported&quot;</span><span class="p">)</span>
<span class="k">if</span> <span class="ow">not</span> <span class="n">config</span><span class="o">.</span><span class="n">quantization</span><span class="o">.</span><span class="n">_requires_modelopt_quantization</span><span class="p">:</span>
<span class="k">raise</span> <span class="ne">ValueError</span><span class="p">(</span>
<span class="sa">f</span><span class="s2">&quot;The quant_config (</span><span class="si">{</span><span class="n">quant_config</span><span class="si">}</span><span class="s2">) should not call modelopt quantization&quot;</span>
<span class="p">)</span>
<span class="kn">from</span><span class="w"> </span><span class="nn">..quantization</span><span class="w"> </span><span class="kn">import</span> <span class="n">quantize_and_export</span>
<span class="n">quantize_and_export</span><span class="p">(</span>
<span class="n">model_dir</span><span class="o">=</span><span class="nb">str</span><span class="p">(</span><span class="n">hf_model_dir</span><span class="p">),</span>
<span class="n">device</span><span class="o">=</span><span class="n">device</span><span class="p">,</span>
<span class="n">calib_dataset</span><span class="o">=</span><span class="n">calib_dataset</span><span class="p">,</span>
<span class="n">dtype</span><span class="o">=</span><span class="n">config</span><span class="o">.</span><span class="n">dtype</span><span class="p">,</span>
<span class="n">qformat</span><span class="o">=</span><span class="n">config</span><span class="o">.</span><span class="n">quantization</span><span class="o">.</span><span class="n">_get_modelopt_qformat</span><span class="p">(),</span>
<span class="n">kv_cache_dtype</span><span class="o">=</span><span class="n">config</span><span class="o">.</span><span class="n">quantization</span><span class="o">.</span><span class="n">_get_modelopt_kv_cache_dtype</span><span class="p">(),</span>
<span class="n">calib_size</span><span class="o">=</span><span class="n">calib_batches</span><span class="p">,</span>
<span class="n">batch_size</span><span class="o">=</span><span class="n">calib_batch_size</span><span class="p">,</span>
<span class="n">calib_max_seq_length</span><span class="o">=</span><span class="n">calib_max_seq_length</span><span class="p">,</span>
<span class="n">awq_block_size</span><span class="o">=</span><span class="n">config</span><span class="o">.</span><span class="n">quantization</span><span class="o">.</span><span class="n">group_size</span><span class="p">,</span>
<span class="n">output_dir</span><span class="o">=</span><span class="n">output_dir</span><span class="p">,</span>
<span class="n">tp_size</span><span class="o">=</span><span class="n">config</span><span class="o">.</span><span class="n">mapping</span><span class="o">.</span><span class="n">tp_size</span><span class="p">,</span>
<span class="n">pp_size</span><span class="o">=</span><span class="n">config</span><span class="o">.</span><span class="n">mapping</span><span class="o">.</span><span class="n">pp_size</span><span class="p">,</span>
<span class="n">cp_size</span><span class="o">=</span><span class="n">config</span><span class="o">.</span><span class="n">mapping</span><span class="o">.</span><span class="n">cp_size</span><span class="p">,</span>
<span class="n">seed</span><span class="o">=</span><span class="n">random_seed</span><span class="p">,</span>
<span class="n">tokenizer_max_seq_length</span><span class="o">=</span><span class="n">tokenizer_max_seq_length</span><span class="p">,</span>
<span class="p">)</span></div>
</div>
<span class="k">class</span><span class="w"> </span><span class="nc">DecoderModelForCausalLM</span><span class="p">(</span><span class="n">PretrainedModel</span><span class="p">):</span>
<span class="k">def</span><span class="w"> </span><span class="fm">__init__</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">config</span><span class="p">:</span> <span class="n">PretrainedConfig</span><span class="p">,</span> <span class="n">transformer</span><span class="p">,</span> <span class="n">lm_head</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">config</span><span class="p">)</span>
<span class="bp">self</span><span class="o">.</span><span class="n">transformer</span> <span class="o">=</span> <span class="n">transformer</span>
<span class="bp">self</span><span class="o">.</span><span class="n">lm_head</span> <span class="o">=</span> <span class="n">lm_head</span>
<span class="bp">self</span><span class="o">.</span><span class="n">mup_width_multiplier</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">&#39;mup_width_multiplier&#39;</span><span class="p">,</span>
<span class="kc">None</span><span class="p">)</span>
<span class="c1"># Create constant attention parameters to be reused by all layers.</span>
<span class="n">Attention</span><span class="o">.</span><span class="n">create_attention_const_params</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">config</span><span class="p">)</span>
<span class="bp">self</span><span class="o">.</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="k">def</span><span class="w"> </span><span class="nf">forward</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span>
<span class="n">input_ids</span><span class="p">:</span> <span class="n">Tensor</span><span class="p">,</span>
<span class="n">position_ids</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">last_token_ids</span><span class="o">=</span><span class="kc">None</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">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">mrope_params</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span>
<span class="n">hidden_states</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span>
<span class="n">prompt_embedding_table</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">prompt_tasks</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">prompt_vocab_size</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">lora_params</span><span class="o">=</span><span class="kc">None</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="c1"># fill attention params.</span>
<span class="n">attention_params</span> <span class="o">=</span> <span class="n">Attention</span><span class="o">.</span><span class="n">fill_attention_params</span><span class="p">(</span>
<span class="bp">self</span><span class="p">,</span> <span class="n">attention_params</span><span class="p">)</span>
<span class="c1"># split the sequence for context parallelism</span>
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">mapping</span><span class="o">.</span><span class="n">cp_size</span> <span class="o">&gt;</span> <span class="mi">1</span><span class="p">:</span>
<span class="k">if</span> <span class="nb">len</span><span class="p">(</span><span class="n">input_ids</span><span class="o">.</span><span class="n">shape</span><span class="p">)</span> <span class="o">==</span> <span class="mi">1</span><span class="p">:</span>
<span class="c1"># input shape is [-1]</span>
<span class="n">input_ids</span><span class="p">,</span> <span class="n">cp_join_index</span> <span class="o">=</span> <span class="n">cp_split_plugin</span><span class="p">(</span>
<span class="n">input_ids</span><span class="p">,</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">attention_params</span><span class="o">.</span><span class="n">host_context_lengths</span><span class="p">,</span>
<span class="bp">self</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">mapping</span><span class="o">.</span><span class="n">cp_size</span><span class="p">,</span>
<span class="bp">self</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">mapping</span><span class="o">.</span><span class="n">cp_rank</span><span class="p">,</span>
<span class="p">)</span>
<span class="k">else</span><span class="p">:</span>
<span class="k">assert</span> <span class="kc">False</span><span class="p">,</span> <span class="s2">&quot;Context parallelism with non-remove-padding is not supported yet.&quot;</span>
<span class="n">is_gemma_2_cg</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">has_config_group</span><span class="p">(</span><span class="n">Gemma2ConfigGroup</span><span class="p">)</span>
<span class="n">is_gemma_3_cg</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">has_config_group</span><span class="p">(</span><span class="n">Gemma3ConfigGroup</span><span class="p">)</span>
<span class="n">kwargs</span> <span class="o">=</span> <span class="p">{</span>
<span class="s1">&#39;input_ids&#39;</span><span class="p">:</span> <span class="n">input_ids</span><span class="p">,</span>
<span class="s1">&#39;position_ids&#39;</span><span class="p">:</span> <span class="n">position_ids</span><span class="p">,</span>
<span class="s1">&#39;use_cache&#39;</span><span class="p">:</span> <span class="n">use_cache</span><span class="p">,</span>
<span class="s1">&#39;attention_mask&#39;</span><span class="p">:</span> <span class="n">attention_mask</span><span class="p">,</span>
<span class="s1">&#39;kv_cache_params&#39;</span><span class="p">:</span> <span class="n">kv_cache_params</span><span class="p">,</span>
<span class="s1">&#39;attention_params&#39;</span><span class="p">:</span> <span class="n">attention_params</span><span class="p">,</span>
<span class="p">}</span>
<span class="k">if</span> <span class="n">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">kwargs</span><span class="p">[</span><span class="s1">&#39;lora_params&#39;</span><span class="p">]</span> <span class="o">=</span> <span class="n">lora_params</span>
<span class="k">if</span> <span class="n">hidden_states</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
<span class="n">kwargs</span><span class="p">[</span><span class="s1">&#39;hidden_states&#39;</span><span class="p">]</span> <span class="o">=</span> <span class="n">hidden_states</span>
<span class="k">if</span> <span class="n">prompt_embedding_table</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
<span class="n">kwargs</span><span class="p">[</span><span class="s1">&#39;prompt_embedding_table&#39;</span><span class="p">]</span> <span class="o">=</span> <span class="n">prompt_embedding_table</span>
<span class="k">if</span> <span class="n">prompt_tasks</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
<span class="n">kwargs</span><span class="p">[</span><span class="s1">&#39;prompt_tasks&#39;</span><span class="p">]</span> <span class="o">=</span> <span class="n">prompt_tasks</span>
<span class="k">if</span> <span class="n">prompt_vocab_size</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
<span class="n">kwargs</span><span class="p">[</span><span class="s1">&#39;prompt_vocab_size&#39;</span><span class="p">]</span> <span class="o">=</span> <span class="n">prompt_vocab_size</span>
<span class="k">if</span> <span class="n">spec_decoding_params</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
<span class="n">kwargs</span><span class="p">[</span><span class="s1">&#39;spec_decoding_params&#39;</span><span class="p">]</span> <span class="o">=</span> <span class="n">spec_decoding_params</span>
<span class="k">if</span> <span class="n">mrope_params</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
<span class="n">kwargs</span><span class="p">[</span><span class="s1">&#39;mrope_params&#39;</span><span class="p">]</span> <span class="o">=</span> <span class="n">mrope_params</span>
<span class="n">hidden_states</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">transformer</span><span class="o">.</span><span class="n">forward</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">use_cache</span><span class="p">:</span>
<span class="n">hidden_states</span><span class="p">,</span> <span class="n">presents</span> <span class="o">=</span> <span class="n">hidden_states</span>
<span class="c1"># All gather and rebuild sequence after transformer layer for context parallelism</span>
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">mapping</span><span class="o">.</span><span class="n">cp_size</span> <span class="o">&gt;</span> <span class="mi">1</span><span class="p">:</span>
<span class="k">if</span> <span class="nb">len</span><span class="p">(</span><span class="n">hidden_states</span><span class="o">.</span><span class="n">shape</span><span class="p">)</span> <span class="o">==</span> <span class="mi">2</span><span class="p">:</span>
<span class="n">hidden_states</span> <span class="o">=</span> <span class="n">allgather</span><span class="p">(</span><span class="n">hidden_states</span><span class="p">,</span>
<span class="bp">self</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">mapping</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">0</span><span class="p">)</span>
<span class="n">hidden_states</span> <span class="o">=</span> <span class="n">view</span><span class="p">(</span><span class="n">hidden_states</span><span class="p">,</span>
<span class="p">[</span><span class="o">-</span><span class="mi">1</span><span class="p">,</span> <span class="n">hidden_states</span><span class="o">.</span><span class="n">shape</span><span class="p">[</span><span class="o">-</span><span class="mi">1</span><span class="p">]])</span>
<span class="n">hidden_states</span> <span class="o">=</span> <span class="n">index_select</span><span class="p">(</span><span class="n">hidden_states</span><span class="p">,</span> <span class="mi">0</span><span class="p">,</span> <span class="n">cp_join_index</span><span class="p">)</span>
<span class="k">else</span><span class="p">:</span>
<span class="k">assert</span> <span class="kc">False</span><span class="p">,</span> <span class="s2">&quot;Context parallelism with non-remove-padding is not supported yet.&quot;</span>
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">mapping</span><span class="o">.</span><span class="n">is_last_pp_rank</span><span class="p">():</span>
<span class="n">all_hidden_states</span> <span class="o">=</span> <span class="n">hidden_states</span>
<span class="n">hidden_states</span> <span class="o">=</span> <span class="n">gather_last_token_logits</span><span class="p">(</span>
<span class="n">hidden_states</span><span class="p">,</span> <span class="n">last_token_ids</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="c1"># [batch_size, hidden_size] -&gt; [batch_size, vocab_size]</span>
<span class="n">lm_logits</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">lm_head</span><span class="p">(</span><span class="n">hidden_states</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="o">.</span><span class="n">config</span><span class="p">,</span> <span class="s1">&#39;output_multiplier_scale&#39;</span><span class="p">):</span>
<span class="n">lm_logits</span> <span class="o">*=</span> <span class="nb">getattr</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">config</span><span class="p">,</span> <span class="s1">&#39;output_multiplier_scale&#39;</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">mup_width_multiplier</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
<span class="n">lm_logits</span> <span class="o">=</span> <span class="n">lm_logits</span> <span class="o">/</span> <span class="bp">self</span><span class="o">.</span><span class="n">mup_width_multiplier</span>
<span class="k">if</span> <span class="n">is_gemma_2_cg</span> <span class="ow">or</span> <span class="n">is_gemma_3_cg</span><span class="p">:</span>
<span class="n">softcap</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">get_config_group</span><span class="p">(</span>
<span class="n">Gemma2ConfigGroup</span> <span class="k">if</span> <span class="ow">not</span> <span class="n">is_gemma_3_cg</span> <span class="k">else</span>
<span class="n">Gemma3ConfigGroup</span><span class="p">)</span><span class="o">.</span><span class="n">final_logit_softcapping</span>
<span class="k">if</span> <span class="n">softcap</span><span class="p">:</span>
<span class="n">lm_logits</span> <span class="o">=</span> <span class="n">lm_logits</span> <span class="o">*</span> <span class="nb">float</span><span class="p">(</span><span class="mi">1</span> <span class="o">/</span> <span class="n">softcap</span><span class="p">)</span>
<span class="n">lm_logits</span> <span class="o">=</span> <span class="n">tanh</span><span class="p">(</span><span class="n">lm_logits</span><span class="p">)</span> <span class="o">*</span> <span class="nb">float</span><span class="p">(</span><span class="n">softcap</span><span class="p">)</span>
<span class="n">lm_logits</span><span class="o">.</span><span class="n">mark_output</span><span class="p">(</span><span class="s1">&#39;logits&#39;</span><span class="p">,</span> <span class="bp">self</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">logits_dtype</span><span class="p">)</span>
<span class="k">else</span><span class="p">:</span>
<span class="n">hidden_states</span><span class="o">.</span><span class="n">mark_output</span><span class="p">(</span><span class="s1">&#39;hidden_states_output&#39;</span><span class="p">,</span> <span class="bp">self</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">dtype</span><span class="p">)</span>
<span class="k">if</span> <span class="n">use_cache</span> <span class="ow">and</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">paged_kv_cache</span><span class="p">:</span>
<span class="k">for</span> <span class="n">i</span><span class="p">,</span> <span class="n">present</span> <span class="ow">in</span> <span class="nb">zip</span><span class="p">(</span>
<span class="bp">self</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">mapping</span><span class="o">.</span><span class="n">pp_layers</span><span class="p">(</span>
<span class="bp">self</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">num_hidden_layers</span><span class="p">),</span> <span class="n">presents</span><span class="p">):</span>
<span class="n">present</span><span class="o">.</span><span class="n">mark_output</span><span class="p">(</span><span class="sa">f</span><span class="s1">&#39;present_key_value_</span><span class="si">{</span><span class="n">i</span><span class="si">}</span><span class="s1">&#39;</span><span class="p">,</span>
<span class="bp">self</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">kv_dtype</span><span class="p">)</span>
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">mapping</span><span class="o">.</span><span class="n">is_last_pp_rank</span><span class="p">():</span>
<span class="k">return</span> <span class="p">(</span><span class="n">lm_logits</span><span class="p">,</span> <span class="n">presents</span><span class="p">,</span> <span class="n">hidden_states</span><span class="p">)</span>
<span class="k">return</span> <span class="p">(</span><span class="n">hidden_states</span><span class="p">,</span> <span class="n">presents</span><span class="p">)</span>
<span class="k">else</span><span class="p">:</span>
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">mapping</span><span class="o">.</span><span class="n">is_last_pp_rank</span><span class="p">():</span>
<span class="k">return</span> <span class="n">lm_logits</span><span class="p">,</span> <span class="n">hidden_states</span><span class="p">,</span> <span class="n">all_hidden_states</span>
<span class="k">return</span> <span class="n">hidden_states</span>
<span class="k">def</span><span class="w"> </span><span class="nf">fuse_gate_mlp</span><span class="p">(</span>
<span class="n">model</span><span class="p">:</span> <span class="n">PretrainedModel</span><span class="p">,</span>
<span class="n">gemm_swiglu_plugin_dtype</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="nb">str</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</span><span class="p">,</span>
<span class="n">low_latency_gemm_swiglu_plugin_dtype</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="nb">str</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</span><span class="p">,</span>
<span class="p">)</span> <span class="o">-&gt;</span> <span class="n">PretrainedModel</span><span class="p">:</span>
<span class="kn">from</span><span class="w"> </span><span class="nn">..quantization.quantize</span><span class="w"> </span><span class="kn">import</span> <span class="n">fp8_quantize</span>
<span class="k">for</span> <span class="n">name</span><span class="p">,</span> <span class="n">mlp</span><span class="p">,</span> <span class="n">layer</span> <span class="ow">in</span> <span class="n">model</span><span class="o">.</span><span class="n">named_modules_with_parent</span><span class="p">():</span>
<span class="k">if</span> <span class="nb">isinstance</span><span class="p">(</span><span class="n">mlp</span><span class="p">,</span> <span class="n">GatedMLP</span><span class="p">):</span>
<span class="n">init_params</span> <span class="o">=</span> <span class="n">get_init_params</span><span class="p">(</span><span class="n">mlp</span><span class="p">)</span>
<span class="n">hidden_act</span> <span class="o">=</span> <span class="n">init_params</span><span class="p">[</span><span class="s2">&quot;hidden_act&quot;</span><span class="p">]</span>
<span class="k">if</span> <span class="n">hidden_act</span> <span class="ow">not</span> <span class="ow">in</span> <span class="p">[</span><span class="s2">&quot;silu&quot;</span><span class="p">,</span> <span class="s2">&quot;gelu&quot;</span><span class="p">]:</span>
<span class="n">logger</span><span class="o">.</span><span class="n">warning</span><span class="p">(</span>
<span class="sa">f</span><span class="s2">&quot;fuse_gate_mlp cannot be done for </span><span class="si">{</span><span class="n">name</span><span class="si">}</span><span class="s2"> due to unsupported activation </span><span class="si">{</span><span class="n">hidden_act</span><span class="si">}</span><span class="s2">. Skipping.&quot;</span>
<span class="p">)</span>
<span class="k">continue</span>
<span class="n">init_params</span><span class="p">[</span><span class="s2">&quot;inner_layernorm&quot;</span><span class="p">]</span> <span class="o">=</span> <span class="n">mlp</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="n">fused_layer</span> <span class="o">=</span> <span class="n">FusedGatedMLP</span><span class="p">(</span><span class="o">**</span><span class="n">init_params</span><span class="p">)</span>
<span class="n">fc_name</span> <span class="o">=</span> <span class="n">name</span> <span class="o">+</span> <span class="s1">&#39;.fc&#39;</span>
<span class="n">layer_quant_cfg</span> <span class="o">=</span> <span class="n">model</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">_get_quant_cfg</span><span class="p">(</span><span class="n">fc_name</span><span class="p">)</span>
<span class="n">layer_quant_algo</span> <span class="o">=</span> <span class="n">layer_quant_cfg</span><span class="o">.</span><span class="n">quant_algo</span>
<span class="k">if</span> <span class="n">layer_quant_algo</span> <span class="o">!=</span> <span class="n">QuantAlgo</span><span class="o">.</span><span class="n">FP8</span> <span class="ow">and</span> <span class="n">layer_quant_algo</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
<span class="k">continue</span>
<span class="k">if</span> <span class="nb">isinstance</span><span class="p">(</span><span class="n">model</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">quantization</span><span class="o">.</span><span class="n">exclude_modules</span><span class="p">,</span> <span class="nb">list</span><span class="p">)</span> \
<span class="ow">and</span> <span class="n">fc_name</span> <span class="ow">in</span> <span class="n">model</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">quantization</span><span class="o">.</span><span class="n">exclude_modules</span><span class="p">:</span>
<span class="n">layer_quant_algo</span> <span class="o">=</span> <span class="kc">None</span>
<span class="k">if</span> <span class="n">layer_quant_algo</span> <span class="o">==</span> <span class="n">QuantAlgo</span><span class="o">.</span><span class="n">FP8</span><span class="p">:</span>
<span class="n">fused_layer</span> <span class="o">=</span> <span class="n">fp8_quantize</span><span class="p">(</span><span class="n">fused_layer</span><span class="p">,</span> <span class="n">layer_quant_cfg</span><span class="p">)</span>
<span class="k">if</span> <span class="nb">isinstance</span><span class="p">(</span><span class="n">mlp</span><span class="o">.</span><span class="n">dtype</span><span class="p">,</span> <span class="nb">str</span><span class="p">):</span>
<span class="n">dtype</span> <span class="o">=</span> <span class="n">str_dtype_to_torch</span><span class="p">(</span><span class="n">mlp</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="n">dtype</span> <span class="o">=</span> <span class="n">trt_dtype_to_torch</span><span class="p">(</span><span class="n">mlp</span><span class="o">.</span><span class="n">dtype</span><span class="p">)</span>
<span class="n">gate_weight</span> <span class="o">=</span> <span class="n">numpy_to_torch</span><span class="p">(</span><span class="n">mlp</span><span class="o">.</span><span class="n">gate</span><span class="o">.</span><span class="n">weight</span><span class="o">.</span><span class="n">raw_value</span><span class="p">)</span>
<span class="n">fc_weight</span> <span class="o">=</span> <span class="n">numpy_to_torch</span><span class="p">(</span><span class="n">mlp</span><span class="o">.</span><span class="n">fc</span><span class="o">.</span><span class="n">weight</span><span class="o">.</span><span class="n">raw_value</span><span class="p">)</span>
<span class="k">assert</span> <span class="n">gate_weight</span><span class="o">.</span><span class="n">dtype</span> <span class="o">==</span> <span class="n">fc_weight</span><span class="o">.</span><span class="n">dtype</span>
<span class="n">need_qdq</span> <span class="o">=</span> <span class="n">gate_weight</span><span class="o">.</span><span class="n">dtype</span> <span class="o">==</span> <span class="n">torch</span><span class="o">.</span><span class="n">float8_e4m3fn</span>
<span class="n">gate_weight</span> <span class="o">=</span> <span class="n">gate_weight</span><span class="o">.</span><span class="n">to</span><span class="p">(</span><span class="n">dtype</span><span class="p">)</span>
<span class="n">fc_weight</span> <span class="o">=</span> <span class="n">fc_weight</span><span class="o">.</span><span class="n">to</span><span class="p">(</span><span class="n">dtype</span><span class="p">)</span>
<span class="c1"># dequantize if needed</span>
<span class="k">if</span> <span class="n">need_qdq</span><span class="p">:</span>
<span class="n">gate_weight</span> <span class="o">=</span> <span class="n">gate_weight</span><span class="o">.</span><span class="n">to</span><span class="p">(</span><span class="n">dtype</span><span class="p">)</span> <span class="o">*</span> <span class="n">numpy_to_torch</span><span class="p">(</span>
<span class="n">mlp</span><span class="o">.</span><span class="n">gate</span><span class="o">.</span><span class="n">weights_scaling_factor</span><span class="o">.</span><span class="n">raw_value</span><span class="p">)</span>
<span class="n">fc_weight</span> <span class="o">=</span> <span class="n">fc_weight</span><span class="o">.</span><span class="n">to</span><span class="p">(</span><span class="n">dtype</span><span class="p">)</span> <span class="o">*</span> <span class="n">numpy_to_torch</span><span class="p">(</span>
<span class="n">mlp</span><span class="o">.</span><span class="n">fc</span><span class="o">.</span><span class="n">weights_scaling_factor</span><span class="o">.</span><span class="n">raw_value</span><span class="p">)</span>
<span class="c1"># concat</span>
<span class="n">fused_weight</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">cat</span><span class="p">([</span><span class="n">gate_weight</span><span class="p">,</span> <span class="n">fc_weight</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">fused_weight_scaling_factor</span> <span class="o">=</span> <span class="n">numpy_to_torch</span><span class="p">(</span>
<span class="nb">max</span><span class="p">(</span>
<span class="n">mlp</span><span class="o">.</span><span class="n">gate</span><span class="o">.</span><span class="n">weights_scaling_factor</span><span class="o">.</span><span class="n">raw_value</span><span class="p">,</span>
<span class="n">mlp</span><span class="o">.</span><span class="n">fc</span><span class="o">.</span><span class="n">weights_scaling_factor</span><span class="o">.</span><span class="n">raw_value</span><span class="p">,</span>
<span class="p">))</span>
<span class="c1"># quantize if needed</span>
<span class="k">if</span> <span class="n">need_qdq</span><span class="p">:</span>
<span class="n">fused_weight</span> <span class="o">=</span> <span class="p">(</span><span class="n">fused_weight</span> <span class="o">/</span>
<span class="n">fused_weight_scaling_factor</span><span class="p">)</span><span class="o">.</span><span class="n">to</span><span class="p">(</span>
<span class="n">torch</span><span class="o">.</span><span class="n">float8_e4m3fn</span><span class="p">)</span>
<span class="k">if</span> <span class="n">gemm_swiglu_plugin_dtype</span> <span class="o">==</span> <span class="s1">&#39;fp8&#39;</span> <span class="ow">or</span> <span class="n">low_latency_gemm_swiglu_plugin_dtype</span> <span class="o">==</span> <span class="s1">&#39;fp8&#39;</span><span class="p">:</span>
<span class="c1"># gemm_swiglu_plugin needs (k, n) weights</span>
<span class="c1"># but weights should still be k-major for fp8</span>
<span class="n">fused_layer</span><span class="o">.</span><span class="n">fused_fc</span><span class="o">.</span><span class="n">weight</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">fused_layer</span><span class="o">.</span><span class="n">fused_fc</span><span class="o">.</span><span class="n">in_features</span><span class="p">,</span>
<span class="n">fused_layer</span><span class="o">.</span><span class="n">fused_fc</span><span class="o">.</span><span class="n">out_features</span><span class="p">),</span>
<span class="n">dtype</span><span class="o">=</span><span class="s1">&#39;fp8&#39;</span><span class="p">)</span>
<span class="n">fused_layer</span><span class="o">.</span><span class="n">fused_fc</span><span class="o">.</span><span class="n">weight</span><span class="o">.</span><span class="n">value</span> <span class="o">=</span> <span class="n">fused_weight</span><span class="o">.</span><span class="n">view</span><span class="p">(</span>
<span class="n">fused_layer</span><span class="o">.</span><span class="n">fused_fc</span><span class="o">.</span><span class="n">in_features</span><span class="p">,</span>
<span class="n">fused_layer</span><span class="o">.</span><span class="n">fused_fc</span><span class="o">.</span><span class="n">out_features</span><span class="p">)</span>
<span class="k">else</span><span class="p">:</span>
<span class="n">fused_layer</span><span class="o">.</span><span class="n">fused_fc</span><span class="o">.</span><span class="n">weight</span><span class="o">.</span><span class="n">value</span> <span class="o">=</span> <span class="n">fused_weight</span>
<span class="n">fused_layer</span><span class="o">.</span><span class="n">fused_fc</span><span class="o">.</span><span class="n">weights_scaling_factor</span><span class="o">.</span><span class="n">value</span> <span class="o">=</span> <span class="n">fused_weight_scaling_factor</span>
<span class="n">fused_layer</span><span class="o">.</span><span class="n">fused_fc</span><span class="o">.</span><span class="n">activation_scaling_factor</span><span class="o">.</span><span class="n">value</span> <span class="o">=</span> <span class="nb">max</span><span class="p">(</span>
<span class="n">mlp</span><span class="o">.</span><span class="n">gate</span><span class="o">.</span><span class="n">activation_scaling_factor</span><span class="o">.</span><span class="n">raw_value</span><span class="p">,</span>
<span class="n">mlp</span><span class="o">.</span><span class="n">fc</span><span class="o">.</span><span class="n">activation_scaling_factor</span><span class="o">.</span><span class="n">raw_value</span><span class="p">,</span>
<span class="p">)</span>
<span class="k">elif</span> <span class="n">layer_quant_algo</span> <span class="ow">is</span> <span class="kc">None</span><span class="p">:</span>
<span class="n">fused_layer</span><span class="o">.</span><span class="n">fused_fc</span><span class="o">.</span><span class="n">weight</span><span class="o">.</span><span class="n">value</span> <span class="o">=</span> <span class="n">np</span><span class="o">.</span><span class="n">concatenate</span><span class="p">(</span>
<span class="p">[</span>
<span class="n">mlp</span><span class="o">.</span><span class="n">gate</span><span class="o">.</span><span class="n">weight</span><span class="o">.</span><span class="n">raw_value</span><span class="p">,</span>
<span class="n">mlp</span><span class="o">.</span><span class="n">fc</span><span class="o">.</span><span class="n">weight</span><span class="o">.</span><span class="n">raw_value</span><span class="p">,</span>
<span class="p">],</span>
<span class="n">axis</span><span class="o">=</span><span class="mi">0</span><span class="p">,</span>
<span class="p">)</span>
<span class="k">if</span> <span class="n">mlp</span><span class="o">.</span><span class="n">bias</span><span class="p">:</span>
<span class="n">fused_layer</span><span class="o">.</span><span class="n">fused_fc</span><span class="o">.</span><span class="n">bias</span><span class="o">.</span><span class="n">value</span> <span class="o">=</span> <span class="n">np</span><span class="o">.</span><span class="n">concatenate</span><span class="p">(</span>
<span class="p">[</span><span class="n">mlp</span><span class="o">.</span><span class="n">gate</span><span class="o">.</span><span class="n">bias</span><span class="o">.</span><span class="n">raw_value</span><span class="p">,</span> <span class="n">mlp</span><span class="o">.</span><span class="n">fc</span><span class="o">.</span><span class="n">bias</span><span class="o">.</span><span class="n">raw_value</span><span class="p">],</span>
<span class="n">axis</span><span class="o">=</span><span class="mi">0</span><span class="p">)</span>
<span class="k">else</span><span class="p">:</span>
<span class="k">raise</span> <span class="ne">ValueError</span><span class="p">(</span><span class="sa">f</span><span class="s1">&#39;Unsupported quant algo: </span><span class="si">{</span><span class="n">layer_quant_algo</span><span class="si">}</span><span class="s1">&#39;</span><span class="p">)</span>
<span class="n">fused_layer</span><span class="o">.</span><span class="n">proj</span> <span class="o">=</span> <span class="n">mlp</span><span class="o">.</span><span class="n">proj</span>
<span class="n">fused_layer</span><span class="o">.</span><span class="n">inner_layernorm</span> <span class="o">=</span> <span class="n">mlp</span><span class="o">.</span><span class="n">inner_layernorm</span>
<span class="n">_</span><span class="p">,</span> <span class="n">mlp_name</span> <span class="o">=</span> <span class="n">name</span><span class="o">.</span><span class="n">rsplit</span><span class="p">(</span><span class="s1">&#39;.&#39;</span><span class="p">,</span> <span class="mi">1</span><span class="p">)</span>
<span class="nb">setattr</span><span class="p">(</span><span class="n">layer</span><span class="p">,</span> <span class="n">mlp_name</span><span class="p">,</span> <span class="n">fused_layer</span><span class="p">)</span>
<span class="k">elif</span> <span class="nb">isinstance</span><span class="p">(</span><span class="n">mlp</span><span class="p">,</span> <span class="n">Fp8RowwiseGatedMLP</span><span class="p">):</span>
<span class="n">init_params</span> <span class="o">=</span> <span class="n">get_init_params</span><span class="p">(</span><span class="n">mlp</span><span class="p">)</span>
<span class="n">hidden_act</span> <span class="o">=</span> <span class="n">init_params</span><span class="p">[</span><span class="s2">&quot;hidden_act&quot;</span><span class="p">]</span>
<span class="k">if</span> <span class="n">hidden_act</span> <span class="ow">not</span> <span class="ow">in</span> <span class="p">[</span><span class="s2">&quot;silu&quot;</span><span class="p">,</span> <span class="s2">&quot;gelu&quot;</span><span class="p">]:</span>
<span class="n">logger</span><span class="o">.</span><span class="n">warning</span><span class="p">(</span>
<span class="sa">f</span><span class="s2">&quot;fuse_gate_mlp cannot be done for </span><span class="si">{</span><span class="n">name</span><span class="si">}</span><span class="s2"> due to unsupported activation </span><span class="si">{</span><span class="n">hidden_act</span><span class="si">}</span><span class="s2">. Skipping.&quot;</span>
<span class="p">)</span>
<span class="k">continue</span>
<span class="k">if</span> <span class="n">mlp</span><span class="o">.</span><span class="n">clamp_val</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
<span class="n">init_params</span><span class="p">[</span><span class="s2">&quot;clamp_val&quot;</span><span class="p">]</span> <span class="o">=</span> <span class="n">mlp</span><span class="o">.</span><span class="n">clamp_val</span><span class="o">.</span><span class="n">raw_value</span><span class="o">.</span><span class="n">tolist</span><span class="p">()</span>
<span class="n">fused_layer</span> <span class="o">=</span> <span class="n">Fp8RowwiseFusedGatedMLP</span><span class="p">(</span><span class="o">**</span><span class="n">init_params</span><span class="p">)</span>
<span class="n">fused_layer</span><span class="o">.</span><span class="n">fused_fc</span><span class="o">.</span><span class="n">weight</span><span class="o">.</span><span class="n">value</span> <span class="o">=</span> <span class="n">np</span><span class="o">.</span><span class="n">concatenate</span><span class="p">(</span>
<span class="p">[</span>
<span class="n">mlp</span><span class="o">.</span><span class="n">gate</span><span class="o">.</span><span class="n">weight</span><span class="o">.</span><span class="n">raw_value</span><span class="p">,</span>
<span class="n">mlp</span><span class="o">.</span><span class="n">fc</span><span class="o">.</span><span class="n">weight</span><span class="o">.</span><span class="n">raw_value</span><span class="p">,</span>
<span class="p">],</span>
<span class="n">axis</span><span class="o">=</span><span class="mi">0</span><span class="p">,</span>
<span class="p">)</span>
<span class="n">fused_layer</span><span class="o">.</span><span class="n">fused_fc</span><span class="o">.</span><span class="n">per_channel_scale</span><span class="o">.</span><span class="n">value</span> <span class="o">=</span> <span class="n">np</span><span class="o">.</span><span class="n">concatenate</span><span class="p">(</span>
<span class="p">[</span>
<span class="n">mlp</span><span class="o">.</span><span class="n">gate</span><span class="o">.</span><span class="n">per_channel_scale</span><span class="o">.</span><span class="n">raw_value</span><span class="p">,</span>
<span class="n">mlp</span><span class="o">.</span><span class="n">fc</span><span class="o">.</span><span class="n">per_channel_scale</span><span class="o">.</span><span class="n">raw_value</span><span class="p">,</span>
<span class="p">],</span>
<span class="n">axis</span><span class="o">=</span><span class="mi">0</span><span class="p">,</span>
<span class="p">)</span>
<span class="k">if</span> <span class="n">mlp</span><span class="o">.</span><span class="n">bias</span><span class="p">:</span>
<span class="n">fused_layer</span><span class="o">.</span><span class="n">fused_fc</span><span class="o">.</span><span class="n">bias</span><span class="o">.</span><span class="n">value</span> <span class="o">=</span> <span class="n">np</span><span class="o">.</span><span class="n">concatenate</span><span class="p">(</span>
<span class="p">[</span><span class="n">mlp</span><span class="o">.</span><span class="n">gate</span><span class="o">.</span><span class="n">bias</span><span class="o">.</span><span class="n">raw_value</span><span class="p">,</span> <span class="n">mlp</span><span class="o">.</span><span class="n">fc</span><span class="o">.</span><span class="n">bias</span><span class="o">.</span><span class="n">raw_value</span><span class="p">],</span> <span class="n">axis</span><span class="o">=</span><span class="mi">0</span><span class="p">)</span>
<span class="n">fused_layer</span><span class="o">.</span><span class="n">proj</span> <span class="o">=</span> <span class="n">mlp</span><span class="o">.</span><span class="n">proj</span>
<span class="n">_</span><span class="p">,</span> <span class="n">mlp_name</span> <span class="o">=</span> <span class="n">name</span><span class="o">.</span><span class="n">rsplit</span><span class="p">(</span><span class="s1">&#39;.&#39;</span><span class="p">,</span> <span class="mi">1</span><span class="p">)</span>
<span class="nb">setattr</span><span class="p">(</span><span class="n">layer</span><span class="p">,</span> <span class="n">mlp_name</span><span class="p">,</span> <span class="n">fused_layer</span><span class="p">)</span>
<span class="k">return</span> <span class="n">model</span>
<span class="k">def</span><span class="w"> </span><span class="nf">unfuse_qkv_gemm</span><span class="p">(</span><span class="n">model</span><span class="p">:</span> <span class="n">PretrainedModel</span><span class="p">)</span> <span class="o">-&gt;</span> <span class="n">PretrainedModel</span><span class="p">:</span>
<span class="w"> </span><span class="sd">&#39;&#39;&#39;Split all the models&#39; Attention layer&#39;s QKV GEMM into 3 GEMMs layer.q layer.k, layer.v and return the changed model</span>
<span class="sd"> &#39;&#39;&#39;</span>
<span class="kn">from</span><span class="w"> </span><span class="nn">..quantization.quantize</span><span class="w"> </span><span class="kn">import</span> <span class="n">quantize</span>
<span class="k">for</span> <span class="n">name</span><span class="p">,</span> <span class="n">layer</span> <span class="ow">in</span> <span class="n">model</span><span class="o">.</span><span class="n">named_modules</span><span class="p">():</span>
<span class="k">if</span> <span class="nb">isinstance</span><span class="p">(</span><span class="n">layer</span><span class="p">,</span> <span class="n">Attention</span><span class="p">)</span> <span class="ow">and</span> <span class="ow">not</span> <span class="n">layer</span><span class="o">.</span><span class="n">cross_attention</span><span class="p">:</span>
<span class="k">assert</span> <span class="n">layer</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="s2">&quot;please disable manual tp when enable auto parallel&quot;</span>
<span class="k">if</span> <span class="n">layer</span><span class="o">.</span><span class="n">qkv</span> <span class="ow">is</span> <span class="kc">None</span><span class="p">:</span>
<span class="k">continue</span>
<span class="n">qkv_params</span> <span class="o">=</span> <span class="n">get_init_params</span><span class="p">(</span><span class="n">layer</span><span class="o">.</span><span class="n">qkv</span><span class="p">,</span> <span class="n">ColumnLinear</span><span class="p">)</span>
<span class="n">qkv_params</span><span class="p">[</span><span class="s2">&quot;bias&quot;</span><span class="p">]</span> <span class="o">=</span> <span class="n">qkv_params</span><span class="p">[</span><span class="s2">&quot;bias&quot;</span><span class="p">]</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span>
<span class="n">qkv_params</span><span class="p">[</span><span class="s2">&quot;strict_dtype&quot;</span><span class="p">]</span> <span class="o">=</span> <span class="n">qkv_params</span><span class="o">.</span><span class="n">get</span><span class="p">(</span>
<span class="s2">&quot;strict_dtype&quot;</span><span class="p">)</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span>
<span class="n">q</span> <span class="o">=</span> <span class="n">ColumnLinear</span><span class="p">(</span>
<span class="o">**</span><span class="p">{</span>
<span class="o">**</span><span class="n">qkv_params</span><span class="p">,</span>
<span class="s2">&quot;out_features&quot;</span><span class="p">:</span>
<span class="n">layer</span><span class="o">.</span><span class="n">tp_size</span> <span class="o">*</span> <span class="n">layer</span><span class="o">.</span><span class="n">num_attention_heads</span> <span class="o">*</span>
<span class="n">layer</span><span class="o">.</span><span class="n">attention_head_size</span><span class="p">,</span>
<span class="p">})</span>
<span class="n">k</span> <span class="o">=</span> <span class="n">ColumnLinear</span><span class="p">(</span>
<span class="o">**</span><span class="p">{</span>
<span class="o">**</span><span class="n">qkv_params</span><span class="p">,</span>
<span class="s2">&quot;out_features&quot;</span><span class="p">:</span>
<span class="n">layer</span><span class="o">.</span><span class="n">tp_size</span> <span class="o">*</span> <span class="n">layer</span><span class="o">.</span><span class="n">num_attention_kv_heads</span> <span class="o">*</span>
<span class="n">layer</span><span class="o">.</span><span class="n">attention_head_size</span><span class="p">,</span>
<span class="p">})</span>
<span class="n">v</span> <span class="o">=</span> <span class="n">ColumnLinear</span><span class="p">(</span>
<span class="o">**</span><span class="p">{</span>
<span class="o">**</span><span class="n">qkv_params</span><span class="p">,</span>
<span class="s2">&quot;out_features&quot;</span><span class="p">:</span>
<span class="n">layer</span><span class="o">.</span><span class="n">tp_size</span> <span class="o">*</span> <span class="n">layer</span><span class="o">.</span><span class="n">num_attention_kv_heads</span> <span class="o">*</span>
<span class="n">layer</span><span class="o">.</span><span class="n">attention_head_size</span><span class="p">,</span>
<span class="p">})</span>
<span class="n">layer_quant_cfg</span> <span class="o">=</span> <span class="n">model</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">_get_quant_cfg</span><span class="p">(</span><span class="n">name</span> <span class="o">+</span> <span class="s1">&#39;.qkv&#39;</span><span class="p">)</span>
<span class="n">q</span> <span class="o">=</span> <span class="n">quantize</span><span class="p">(</span><span class="n">q</span><span class="p">,</span> <span class="n">layer_quant_cfg</span><span class="p">)</span>
<span class="n">k</span> <span class="o">=</span> <span class="n">quantize</span><span class="p">(</span><span class="n">k</span><span class="p">,</span> <span class="n">layer_quant_cfg</span><span class="p">)</span>
<span class="n">v</span> <span class="o">=</span> <span class="n">quantize</span><span class="p">(</span><span class="n">v</span><span class="p">,</span> <span class="n">layer_quant_cfg</span><span class="p">)</span>
<span class="n">out_features</span> <span class="o">=</span> <span class="n">q</span><span class="o">.</span><span class="n">out_features</span> <span class="o">+</span> <span class="n">k</span><span class="o">.</span><span class="n">out_features</span> <span class="o">+</span> <span class="n">v</span><span class="o">.</span><span class="n">out_features</span>
<span class="k">if</span> <span class="nb">isinstance</span><span class="p">(</span><span class="n">layer</span><span class="o">.</span><span class="n">qkv</span><span class="p">,</span> <span class="p">(</span>
<span class="n">WeightOnlyQuantLinear</span><span class="p">,</span>
<span class="n">WeightOnlyQuantRowLinear</span><span class="p">,</span>
<span class="n">WeightOnlyGroupwiseQuantLinear</span><span class="p">,</span>
<span class="n">WeightOnlyGroupwiseQuantRowLinear</span><span class="p">,</span>
<span class="p">)):</span>
<span class="n">out_dim</span> <span class="o">=</span> <span class="mi">1</span>
<span class="k">else</span><span class="p">:</span>
<span class="n">out_dim</span> <span class="o">=</span> <span class="mi">0</span>
<span class="k">if</span> <span class="n">layer</span><span class="o">.</span><span class="n">qkv</span><span class="o">.</span><span class="n">weight</span><span class="o">.</span><span class="n">is_inited</span><span class="p">():</span>
<span class="n">qkv_weight</span> <span class="o">=</span> <span class="n">layer</span><span class="o">.</span><span class="n">qkv</span><span class="o">.</span><span class="n">weight</span><span class="o">.</span><span class="n">raw_value</span>
<span class="n">weights</span> <span class="o">=</span> <span class="n">np</span><span class="o">.</span><span class="n">split</span><span class="p">(</span><span class="n">qkv_weight</span><span class="p">,</span> <span class="p">[</span>
<span class="n">qkv_weight</span><span class="o">.</span><span class="n">shape</span><span class="p">[</span><span class="n">out_dim</span><span class="p">]</span> <span class="o">*</span> <span class="n">q</span><span class="o">.</span><span class="n">out_features</span> <span class="o">//</span> <span class="n">out_features</span><span class="p">,</span>
<span class="n">qkv_weight</span><span class="o">.</span><span class="n">shape</span><span class="p">[</span><span class="n">out_dim</span><span class="p">]</span> <span class="o">*</span>
<span class="p">(</span><span class="n">q</span><span class="o">.</span><span class="n">out_features</span> <span class="o">+</span> <span class="n">k</span><span class="o">.</span><span class="n">out_features</span><span class="p">)</span> <span class="o">//</span> <span class="n">out_features</span><span class="p">,</span>
<span class="p">],</span>
<span class="n">axis</span><span class="o">=</span><span class="n">out_dim</span><span class="p">)</span>
<span class="k">for</span> <span class="n">gemm</span><span class="p">,</span> <span class="n">weight</span> <span class="ow">in</span> <span class="nb">zip</span><span class="p">([</span><span class="n">q</span><span class="p">,</span> <span class="n">k</span><span class="p">,</span> <span class="n">v</span><span class="p">],</span> <span class="n">weights</span><span class="p">):</span>
<span class="n">gemm</span><span class="o">.</span><span class="n">weight</span><span class="o">.</span><span class="n">value</span> <span class="o">=</span> <span class="n">weight</span>
<span class="k">if</span> <span class="n">layer</span><span class="o">.</span><span class="n">qkv</span><span class="o">.</span><span class="n">bias</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span> <span class="ow">and</span> <span class="n">layer</span><span class="o">.</span><span class="n">qkv</span><span class="o">.</span><span class="n">bias</span><span class="o">.</span><span class="n">is_inited</span><span class="p">():</span>
<span class="n">qkv_bias</span> <span class="o">=</span> <span class="n">layer</span><span class="o">.</span><span class="n">qkv</span><span class="o">.</span><span class="n">bias</span><span class="o">.</span><span class="n">raw_value</span>
<span class="n">biases</span> <span class="o">=</span> <span class="n">np</span><span class="o">.</span><span class="n">split</span><span class="p">(</span><span class="n">qkv_bias</span><span class="p">,</span> <span class="p">[</span>
<span class="n">qkv_bias</span><span class="o">.</span><span class="n">shape</span><span class="p">[</span><span class="n">out_dim</span><span class="p">]</span> <span class="o">*</span> <span class="n">q</span><span class="o">.</span><span class="n">out_features</span> <span class="o">//</span> <span class="n">out_features</span><span class="p">,</span>
<span class="n">qkv_bias</span><span class="o">.</span><span class="n">shape</span><span class="p">[</span><span class="n">out_dim</span><span class="p">]</span> <span class="o">*</span>
<span class="p">(</span><span class="n">q</span><span class="o">.</span><span class="n">out_features</span> <span class="o">+</span> <span class="n">k</span><span class="o">.</span><span class="n">out_features</span><span class="p">)</span> <span class="o">//</span> <span class="n">out_features</span><span class="p">,</span>
<span class="p">],</span>
<span class="n">axis</span><span class="o">=</span><span class="n">out_dim</span><span class="p">)</span>
<span class="k">for</span> <span class="n">gemm</span><span class="p">,</span> <span class="n">bias</span> <span class="ow">in</span> <span class="nb">zip</span><span class="p">([</span><span class="n">q</span><span class="p">,</span> <span class="n">k</span><span class="p">,</span> <span class="n">v</span><span class="p">],</span> <span class="n">biases</span><span class="p">):</span>
<span class="n">gemm</span><span class="o">.</span><span class="n">bias</span><span class="o">.</span><span class="n">value</span> <span class="o">=</span> <span class="n">bias</span>
<span class="k">for</span> <span class="n">name</span><span class="p">,</span> <span class="n">parameter</span> <span class="ow">in</span> <span class="n">layer</span><span class="o">.</span><span class="n">qkv</span><span class="o">.</span><span class="n">_parameters</span><span class="o">.</span><span class="n">items</span><span class="p">():</span>
<span class="k">if</span> <span class="n">name</span> <span class="ow">not</span> <span class="ow">in</span> <span class="p">[</span><span class="s2">&quot;weight&quot;</span><span class="p">,</span> <span class="s2">&quot;bias&quot;</span><span class="p">]:</span>
<span class="k">for</span> <span class="n">gemm</span> <span class="ow">in</span> <span class="p">[</span><span class="n">q</span><span class="p">,</span> <span class="n">k</span><span class="p">,</span> <span class="n">v</span><span class="p">]:</span>
<span class="nb">setattr</span><span class="p">(</span><span class="n">gemm</span><span class="p">,</span> <span class="n">name</span><span class="p">,</span> <span class="n">parameter</span><span class="p">)</span>
<span class="n">layer</span><span class="o">.</span><span class="n">q</span> <span class="o">=</span> <span class="n">q</span>
<span class="n">layer</span><span class="o">.</span><span class="n">k</span> <span class="o">=</span> <span class="n">k</span>
<span class="n">layer</span><span class="o">.</span><span class="n">v</span> <span class="o">=</span> <span class="n">v</span>
<span class="n">layer</span><span class="o">.</span><span class="n">qkv</span> <span class="o">=</span> <span class="kc">None</span>
<span class="k">return</span> <span class="n">model</span>
<span class="k">def</span><span class="w"> </span><span class="nf">fuse_rg_lru</span><span class="p">(</span><span class="n">model</span><span class="p">:</span> <span class="n">PretrainedModel</span><span class="p">)</span> <span class="o">-&gt;</span> <span class="n">PretrainedModel</span><span class="p">:</span>
<span class="k">for</span> <span class="n">name</span><span class="p">,</span> <span class="n">rg_lru</span><span class="p">,</span> <span class="n">parent</span> <span class="ow">in</span> <span class="n">model</span><span class="o">.</span><span class="n">named_modules_with_parent</span><span class="p">():</span>
<span class="k">if</span> <span class="nb">isinstance</span><span class="p">(</span><span class="n">rg_lru</span><span class="p">,</span> <span class="n">RgLru</span><span class="p">):</span>
<span class="n">fused_layer</span> <span class="o">=</span> <span class="n">FusedRgLru</span><span class="p">(</span><span class="o">**</span><span class="n">get_init_params</span><span class="p">(</span><span class="n">rg_lru</span><span class="p">))</span>
<span class="n">fused_layer</span><span class="o">.</span><span class="n">gate</span><span class="o">.</span><span class="n">weight</span><span class="o">.</span><span class="n">value</span> <span class="o">=</span> <span class="n">np</span><span class="o">.</span><span class="n">concatenate</span><span class="p">(</span>
<span class="p">[</span>
<span class="n">rg_lru</span><span class="o">.</span><span class="n">input_gate</span><span class="o">.</span><span class="n">weight</span><span class="o">.</span><span class="n">raw_value</span><span class="p">,</span>
<span class="n">rg_lru</span><span class="o">.</span><span class="n">recurrent_gate</span><span class="o">.</span><span class="n">weight</span><span class="o">.</span><span class="n">raw_value</span><span class="p">,</span>
<span class="p">],</span>
<span class="n">axis</span><span class="o">=-</span><span class="mi">1</span><span class="p">,</span>
<span class="p">)</span>
<span class="n">fused_layer</span><span class="o">.</span><span class="n">gate</span><span class="o">.</span><span class="n">bias</span><span class="o">.</span><span class="n">value</span> <span class="o">=</span> <span class="n">np</span><span class="o">.</span><span class="n">concatenate</span><span class="p">(</span>
<span class="p">[</span>
<span class="n">rg_lru</span><span class="o">.</span><span class="n">input_gate</span><span class="o">.</span><span class="n">bias</span><span class="o">.</span><span class="n">raw_value</span><span class="p">,</span>
<span class="n">rg_lru</span><span class="o">.</span><span class="n">recurrent_gate</span><span class="o">.</span><span class="n">bias</span><span class="o">.</span><span class="n">raw_value</span><span class="p">,</span>
<span class="p">],</span>
<span class="n">axis</span><span class="o">=-</span><span class="mi">1</span><span class="p">,</span>
<span class="p">)</span>
<span class="n">fused_layer</span><span class="o">.</span><span class="n">recurrent_param</span><span class="o">.</span><span class="n">value</span> <span class="o">=</span> <span class="n">rg_lru</span><span class="o">.</span><span class="n">recurrent_param</span><span class="o">.</span><span class="n">raw_value</span>
<span class="n">rg_lru_name</span> <span class="o">=</span> <span class="n">name</span><span class="o">.</span><span class="n">rsplit</span><span class="p">(</span><span class="s1">&#39;.&#39;</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="nb">setattr</span><span class="p">(</span><span class="n">parent</span><span class="p">,</span> <span class="n">rg_lru_name</span><span class="p">,</span> <span class="n">fused_layer</span><span class="p">)</span>
<span class="k">return</span> <span class="n">model</span>
<span class="k">def</span><span class="w"> </span><span class="nf">set_prompt_tuning</span><span class="p">(</span><span class="n">model</span><span class="p">:</span> <span class="n">PretrainedModel</span><span class="p">)</span> <span class="o">-&gt;</span> <span class="n">PretrainedModel</span><span class="p">:</span>
<span class="w"> </span><span class="sd">&#39;&#39;&#39;Replace the given models embedding layer with a PromptTuningEmbedding layer in-place, return the changed model</span>
<span class="sd"> Pre-conditions: vocab_embedding exists</span>
<span class="sd"> Post-conditions: isinstance(vocab_embedding, PromptTuningEmbedding)</span>
<span class="sd"> &#39;&#39;&#39;</span>
<span class="k">for</span> <span class="n">name</span><span class="p">,</span> <span class="n">embedding</span><span class="p">,</span> <span class="n">parent</span> <span class="ow">in</span> <span class="n">model</span><span class="o">.</span><span class="n">named_modules_with_parent</span><span class="p">():</span>
<span class="n">layer_name</span> <span class="o">=</span> <span class="n">name</span><span class="o">.</span><span class="n">rsplit</span><span class="p">(</span><span class="s1">&#39;.&#39;</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="k">if</span> <span class="n">layer_name</span> <span class="o">==</span> <span class="s2">&quot;vocab_embedding&quot;</span> <span class="ow">and</span> <span class="nb">isinstance</span><span class="p">(</span><span class="n">embedding</span><span class="p">,</span> <span class="n">Embedding</span><span class="p">):</span>
<span class="n">ptuning_embedding</span> <span class="o">=</span> <span class="n">PromptTuningEmbedding</span><span class="p">(</span>
<span class="o">**</span><span class="n">get_init_params</span><span class="p">(</span><span class="n">embedding</span><span class="p">))</span>
<span class="n">ptuning_embedding</span><span class="o">.</span><span class="n">weight</span><span class="o">.</span><span class="n">value</span> <span class="o">=</span> <span class="n">embedding</span><span class="o">.</span><span class="n">weight</span><span class="o">.</span><span class="n">raw_value</span>
<span class="n">parent</span><span class="o">.</span><span class="n">vocab_embedding</span> <span class="o">=</span> <span class="n">ptuning_embedding</span>
<span class="k">return</span> <span class="n">model</span>
<span class="k">def</span><span class="w"> </span><span class="nf">add_lora</span><span class="p">(</span><span class="n">model</span><span class="p">:</span> <span class="n">PretrainedModel</span><span class="p">,</span>
<span class="n">max_lora_rank</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="nb">int</span><span class="p">],</span>
<span class="n">with_dora</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="o">-&gt;</span> <span class="n">PretrainedModel</span><span class="p">:</span>
<span class="w"> </span><span class="sd">&#39;&#39;&#39; Add lora layers to the Attention/BertAttention/Linear/RowLinear/FusedGatedMLP layers to the given model, return the changed model</span>
<span class="sd"> &#39;&#39;&#39;</span>
<span class="k">for</span> <span class="n">name</span><span class="p">,</span> <span class="n">layer</span> <span class="ow">in</span> <span class="n">model</span><span class="o">.</span><span class="n">named_modules</span><span class="p">():</span>
<span class="n">max_rank</span> <span class="o">=</span> <span class="n">max_lora_rank</span>
<span class="k">if</span> <span class="nb">isinstance</span><span class="p">(</span><span class="n">layer</span><span class="p">,</span> <span class="p">(</span><span class="n">Attention</span><span class="p">,</span> <span class="n">BertAttention</span><span class="p">)):</span>
<span class="k">if</span> <span class="n">max_rank</span> <span class="ow">is</span> <span class="kc">None</span><span class="p">:</span>
<span class="n">max_rank</span> <span class="o">=</span> <span class="nb">min</span><span class="p">(</span>
<span class="n">layer</span><span class="o">.</span><span class="n">hidden_size</span><span class="p">,</span>
<span class="n">layer</span><span class="o">.</span><span class="n">num_attention_heads</span> <span class="o">*</span> <span class="n">layer</span><span class="o">.</span><span class="n">attention_head_size</span><span class="p">,</span>
<span class="n">layer</span><span class="o">.</span><span class="n">num_attention_kv_heads</span> <span class="o">*</span> <span class="n">layer</span><span class="o">.</span><span class="n">attention_head_size</span><span class="p">)</span>
<span class="n">layer</span><span class="o">.</span><span class="n">qkv_lora</span> <span class="o">=</span> <span class="n">Lora</span><span class="p">(</span>
<span class="n">in_hidden_size</span><span class="o">=</span><span class="n">layer</span><span class="o">.</span><span class="n">hidden_size</span><span class="p">,</span>
<span class="n">out_hidden_sizes</span><span class="o">=</span><span class="p">[</span>
<span class="n">layer</span><span class="o">.</span><span class="n">num_attention_heads</span> <span class="o">*</span> <span class="n">layer</span><span class="o">.</span><span class="n">attention_head_size</span><span class="p">,</span>
<span class="n">layer</span><span class="o">.</span><span class="n">num_attention_kv_heads</span> <span class="o">*</span> <span class="n">layer</span><span class="o">.</span><span class="n">attention_head_size</span><span class="p">,</span>
<span class="n">layer</span><span class="o">.</span><span class="n">num_attention_kv_heads</span> <span class="o">*</span> <span class="n">layer</span><span class="o">.</span><span class="n">attention_head_size</span>
<span class="p">],</span>
<span class="n">max_low_rank</span><span class="o">=</span><span class="n">max_rank</span><span class="p">,</span>
<span class="p">)</span>
<span class="k">if</span> <span class="n">with_dora</span><span class="p">:</span>
<span class="n">layer</span><span class="o">.</span><span class="n">qkv_dora</span> <span class="o">=</span> <span class="n">Dora</span><span class="p">(</span><span class="n">out_hidden_sizes</span><span class="o">=</span><span class="p">[</span>
<span class="n">layer</span><span class="o">.</span><span class="n">num_attention_heads</span> <span class="o">*</span> <span class="n">layer</span><span class="o">.</span><span class="n">attention_head_size</span><span class="p">,</span>
<span class="n">layer</span><span class="o">.</span><span class="n">num_attention_kv_heads</span> <span class="o">*</span> <span class="n">layer</span><span class="o">.</span><span class="n">attention_head_size</span><span class="p">,</span>
<span class="n">layer</span><span class="o">.</span><span class="n">num_attention_kv_heads</span> <span class="o">*</span> <span class="n">layer</span><span class="o">.</span><span class="n">attention_head_size</span>
<span class="p">],</span> <span class="p">)</span>
<span class="k">if</span> <span class="nb">isinstance</span><span class="p">(</span><span class="n">layer</span><span class="p">,</span> <span class="p">(</span><span class="n">Linear</span><span class="p">,</span> <span class="n">RowLinear</span><span class="p">)):</span>
<span class="k">if</span> <span class="n">max_rank</span> <span class="ow">is</span> <span class="kc">None</span><span class="p">:</span>
<span class="n">max_rank</span> <span class="o">=</span> <span class="nb">min</span><span class="p">(</span><span class="n">layer</span><span class="o">.</span><span class="n">in_features</span><span class="p">,</span> <span class="n">layer</span><span class="o">.</span><span class="n">out_features</span><span class="p">)</span>
<span class="n">layer</span><span class="o">.</span><span class="n">lora</span> <span class="o">=</span> <span class="n">Lora</span><span class="p">(</span>
<span class="n">in_hidden_size</span><span class="o">=</span><span class="n">layer</span><span class="o">.</span><span class="n">in_features</span><span class="p">,</span>
<span class="n">out_hidden_sizes</span><span class="o">=</span><span class="p">[</span><span class="n">layer</span><span class="o">.</span><span class="n">out_features</span><span class="p">],</span>
<span class="n">max_low_rank</span><span class="o">=</span><span class="n">max_rank</span><span class="p">,</span>
<span class="p">)</span>
<span class="k">if</span> <span class="n">with_dora</span><span class="p">:</span>
<span class="n">layer</span><span class="o">.</span><span class="n">dora</span> <span class="o">=</span> <span class="n">Dora</span><span class="p">(</span><span class="n">out_hidden_sizes</span><span class="o">=</span><span class="p">[</span><span class="n">layer</span><span class="o">.</span><span class="n">out_features</span><span class="p">])</span>
<span class="k">if</span> <span class="nb">isinstance</span><span class="p">(</span><span class="n">layer</span><span class="p">,</span> <span class="p">(</span><span class="n">MLP</span><span class="p">,</span> <span class="n">FusedGatedMLP</span><span class="p">)):</span>
<span class="k">if</span> <span class="n">max_rank</span> <span class="ow">is</span> <span class="kc">None</span><span class="p">:</span>
<span class="n">max_rank</span> <span class="o">=</span> <span class="nb">min</span><span class="p">(</span><span class="n">layer</span><span class="o">.</span><span class="n">hidden_size</span><span class="p">,</span>
<span class="n">layer</span><span class="o">.</span><span class="n">ffn_hidden_size</span> <span class="o">//</span> <span class="n">layer</span><span class="o">.</span><span class="n">tp_size</span><span class="p">)</span>
<span class="n">layer</span><span class="o">.</span><span class="n">lora</span> <span class="o">=</span> <span class="n">Lora</span><span class="p">(</span>
<span class="n">in_hidden_size</span><span class="o">=</span><span class="n">layer</span><span class="o">.</span><span class="n">hidden_size</span><span class="p">,</span>
<span class="n">out_hidden_sizes</span><span class="o">=</span><span class="p">[</span>
<span class="n">layer</span><span class="o">.</span><span class="n">ffn_hidden_size</span> <span class="o">//</span> <span class="n">layer</span><span class="o">.</span><span class="n">tp_size</span><span class="p">,</span>
<span class="n">layer</span><span class="o">.</span><span class="n">ffn_hidden_size</span> <span class="o">//</span> <span class="n">layer</span><span class="o">.</span><span class="n">tp_size</span>
<span class="p">],</span>
<span class="n">max_low_rank</span><span class="o">=</span><span class="n">max_rank</span><span class="p">,</span>
<span class="p">)</span>
<span class="k">if</span> <span class="nb">isinstance</span><span class="p">(</span><span class="n">layer</span><span class="p">,</span> <span class="n">FusedGatedMLP</span><span class="p">):</span>
<span class="n">layer</span><span class="o">.</span><span class="n">fused_gate_up_lora</span> <span class="o">=</span> <span class="n">Lora</span><span class="p">(</span>
<span class="n">in_hidden_size</span><span class="o">=</span><span class="n">layer</span><span class="o">.</span><span class="n">hidden_size</span><span class="p">,</span>
<span class="n">out_hidden_sizes</span><span class="o">=</span><span class="p">[</span>
<span class="n">layer</span><span class="o">.</span><span class="n">ffn_hidden_size</span> <span class="o">*</span> <span class="mi">2</span> <span class="o">//</span> <span class="n">layer</span><span class="o">.</span><span class="n">tp_size</span>
<span class="p">],</span>
<span class="n">max_low_rank</span><span class="o">=</span><span class="n">max_rank</span><span class="p">,</span>
<span class="p">)</span>
<span class="k">if</span> <span class="n">with_dora</span><span class="p">:</span>
<span class="n">layer</span><span class="o">.</span><span class="n">dora</span> <span class="o">=</span> <span class="n">Dora</span><span class="p">(</span><span class="n">out_hidden_sizes</span><span class="o">=</span><span class="p">[</span>
<span class="n">layer</span><span class="o">.</span><span class="n">ffn_hidden_size</span> <span class="o">//</span> <span class="n">layer</span><span class="o">.</span><span class="n">tp_size</span><span class="p">,</span>
<span class="n">layer</span><span class="o">.</span><span class="n">ffn_hidden_size</span> <span class="o">//</span> <span class="n">layer</span><span class="o">.</span><span class="n">tp_size</span>
<span class="p">],</span> <span class="p">)</span>
<span class="k">if</span> <span class="nb">isinstance</span><span class="p">(</span><span class="n">layer</span><span class="p">,</span> <span class="n">FusedGatedMLP</span><span class="p">):</span>
<span class="n">layer</span><span class="o">.</span><span class="n">fused_gate_up_dora</span> <span class="o">=</span> <span class="n">Dora</span><span class="p">(</span><span class="n">out_hidden_sizes</span><span class="o">=</span><span class="p">[</span>
<span class="n">layer</span><span class="o">.</span><span class="n">ffn_hidden_size</span> <span class="o">*</span> <span class="mi">2</span> <span class="o">//</span> <span class="n">layer</span><span class="o">.</span><span class="n">tp_size</span>
<span class="p">],</span> <span class="p">)</span>
<span class="k">if</span> <span class="nb">isinstance</span><span class="p">(</span><span class="n">layer</span><span class="p">,</span> <span class="n">MOE</span><span class="p">):</span>
<span class="k">if</span> <span class="n">max_rank</span> <span class="ow">is</span> <span class="kc">None</span><span class="p">:</span>
<span class="n">max_rank</span> <span class="o">=</span> <span class="nb">min</span><span class="p">(</span><span class="n">layer</span><span class="o">.</span><span class="n">hidden_size</span><span class="p">,</span>
<span class="n">layer</span><span class="o">.</span><span class="n">ffn_hidden_size</span> <span class="o">//</span> <span class="n">layer</span><span class="o">.</span><span class="n">tp_size</span><span class="p">)</span>
<span class="n">layer</span><span class="o">.</span><span class="n">max_low_rank</span> <span class="o">=</span> <span class="n">max_rank</span>
<span class="k">return</span> <span class="n">model</span>
<span class="k">def</span><span class="w"> </span><span class="nf">to_ootb_moe</span><span class="p">(</span><span class="n">model</span><span class="p">:</span> <span class="n">PretrainedModel</span><span class="p">)</span> <span class="o">-&gt;</span> <span class="n">PretrainedModel</span><span class="p">:</span>
<span class="w"> </span><span class="sd">&#39;&#39;&#39; Use OOTB MoE instead of MoE plugin, return the changed model</span>
<span class="sd"> &#39;&#39;&#39;</span>
<span class="k">for</span> <span class="n">name</span><span class="p">,</span> <span class="n">layer</span><span class="p">,</span> <span class="n">parent</span> <span class="ow">in</span> <span class="n">model</span><span class="o">.</span><span class="n">named_modules_with_parent</span><span class="p">():</span>
<span class="k">if</span> <span class="nb">isinstance</span><span class="p">(</span><span class="n">layer</span><span class="p">,</span> <span class="n">MOE</span><span class="p">):</span>
<span class="n">layer_name</span> <span class="o">=</span> <span class="n">name</span><span class="o">.</span><span class="n">rsplit</span><span class="p">(</span><span class="s1">&#39;.&#39;</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">ootb_layer</span> <span class="o">=</span> <span class="n">layer</span><span class="o">.</span><span class="n">to</span><span class="p">(</span><span class="n">MoeOOTB</span><span class="p">,</span> <span class="n">model</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">quantization</span><span class="p">)</span>
<span class="nb">setattr</span><span class="p">(</span><span class="n">parent</span><span class="p">,</span> <span class="n">layer_name</span><span class="p">,</span> <span class="n">ootb_layer</span><span class="p">)</span>
<span class="k">return</span> <span class="n">model</span>
<span class="k">def</span><span class="w"> </span><span class="nf">parallelize_embedding</span><span class="p">(</span><span class="n">model</span><span class="p">:</span> <span class="n">PretrainedModel</span><span class="p">)</span> <span class="o">-&gt;</span> <span class="n">PretrainedModel</span><span class="p">:</span>
<span class="k">for</span> <span class="n">name</span><span class="p">,</span> <span class="n">embedding</span><span class="p">,</span> <span class="n">parent</span> <span class="ow">in</span> <span class="n">model</span><span class="o">.</span><span class="n">named_modules_with_parent</span><span class="p">():</span>
<span class="n">layer_name</span> <span class="o">=</span> <span class="n">name</span><span class="o">.</span><span class="n">rsplit</span><span class="p">(</span><span class="s1">&#39;.&#39;</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="k">if</span> <span class="nb">isinstance</span><span class="p">(</span><span class="n">embedding</span><span class="p">,</span> <span class="n">Embedding</span><span class="p">)</span> <span class="ow">and</span> <span class="n">embedding</span><span class="o">.</span><span class="n">tp_group</span> <span class="ow">is</span> <span class="kc">None</span><span class="p">:</span>
<span class="n">init_params</span> <span class="o">=</span> <span class="n">get_init_params</span><span class="p">(</span><span class="n">embedding</span><span class="p">)</span>
<span class="n">init_params</span><span class="p">[</span><span class="s2">&quot;tp_group&quot;</span><span class="p">]</span> <span class="o">=</span> <span class="n">model</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">mapping</span><span class="o">.</span><span class="n">tp_group</span>
<span class="n">init_params</span><span class="p">[</span><span class="s2">&quot;tp_size&quot;</span><span class="p">]</span> <span class="o">=</span> <span class="n">model</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">mapping</span><span class="o">.</span><span class="n">tp_size</span>
<span class="n">init_params</span><span class="p">[</span><span class="s2">&quot;tp_rank&quot;</span><span class="p">]</span> <span class="o">=</span> <span class="n">model</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">mapping</span><span class="o">.</span><span class="n">tp_rank</span>
<span class="n">init_params</span><span class="p">[</span><span class="s2">&quot;sharding_dim&quot;</span><span class="p">]</span> <span class="o">=</span> <span class="n">model</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">embedding_sharding_dim</span>
<span class="n">new_embedding</span> <span class="o">=</span> <span class="n">embedding</span><span class="o">.</span><span class="vm">__class__</span><span class="p">(</span><span class="o">**</span><span class="n">init_params</span><span class="p">)</span>
<span class="nb">setattr</span><span class="p">(</span><span class="n">parent</span><span class="p">,</span> <span class="n">layer_name</span><span class="p">,</span> <span class="n">new_embedding</span><span class="p">)</span>
<span class="k">return</span> <span class="n">model</span>
<span class="k">def</span><span class="w"> </span><span class="nf">share_embedding</span><span class="p">(</span><span class="n">model</span><span class="p">:</span> <span class="n">PretrainedModel</span><span class="p">)</span> <span class="o">-&gt;</span> <span class="n">PretrainedModel</span><span class="p">:</span>
<span class="n">lm_head</span> <span class="o">=</span> <span class="kc">None</span>
<span class="n">vocab_embedding</span> <span class="o">=</span> <span class="kc">None</span>
<span class="k">for</span> <span class="n">name</span><span class="p">,</span> <span class="n">layer</span> <span class="ow">in</span> <span class="n">model</span><span class="o">.</span><span class="n">named_modules</span><span class="p">():</span>
<span class="n">layer_name</span> <span class="o">=</span> <span class="n">name</span><span class="o">.</span><span class="n">rsplit</span><span class="p">(</span><span class="s1">&#39;.&#39;</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="k">if</span> <span class="n">layer_name</span> <span class="o">==</span> <span class="s2">&quot;lm_head&quot;</span><span class="p">:</span>
<span class="n">lm_head</span> <span class="o">=</span> <span class="n">layer</span>
<span class="k">if</span> <span class="n">layer_name</span> <span class="o">==</span> <span class="s2">&quot;vocab_embedding&quot;</span><span class="p">:</span>
<span class="n">vocab_embedding</span> <span class="o">=</span> <span class="n">layer</span>
<span class="k">if</span> <span class="n">lm_head</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span> <span class="ow">and</span> <span class="n">vocab_embedding</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
<span class="k">break</span>
<span class="c1"># Cannot find either lm_head or vocab_embedding, e.g., pipeline parallel</span>
<span class="k">if</span> <span class="n">lm_head</span> <span class="ow">is</span> <span class="kc">None</span> <span class="ow">or</span> <span class="n">vocab_embedding</span> <span class="ow">is</span> <span class="kc">None</span><span class="p">:</span>
<span class="k">return</span> <span class="n">model</span>
<span class="c1"># lm_head and vocab_embedding have different shapes, e.g., tensor parallel without embedding parallel</span>
<span class="k">if</span> <span class="n">lm_head</span><span class="o">.</span><span class="n">weight</span><span class="o">.</span><span class="n">shape</span> <span class="o">!=</span> <span class="n">vocab_embedding</span><span class="o">.</span><span class="n">weight</span><span class="o">.</span><span class="n">shape</span><span class="p">:</span>
<span class="k">return</span> <span class="n">model</span>
<span class="c1"># lm_head can have a different type if quantized</span>
<span class="k">if</span> <span class="n">lm_head</span><span class="o">.</span><span class="n">weight</span><span class="o">.</span><span class="n">dtype</span> <span class="o">!=</span> <span class="n">vocab_embedding</span><span class="o">.</span><span class="n">weight</span><span class="o">.</span><span class="n">dtype</span><span class="p">:</span>
<span class="k">return</span> <span class="n">model</span>
<span class="c1"># Don&#39;t assume weight can be shared if vocab_embedding is not initialized, e.g., dummy weights</span>
<span class="k">if</span> <span class="ow">not</span> <span class="n">vocab_embedding</span><span class="o">.</span><span class="n">weight</span><span class="o">.</span><span class="n">is_inited</span><span class="p">():</span>
<span class="k">return</span> <span class="n">model</span>
<span class="k">if</span> <span class="n">lm_head</span><span class="o">.</span><span class="n">weight</span><span class="o">.</span><span class="n">is_inited</span><span class="p">():</span>
<span class="n">lm_head_weight</span> <span class="o">=</span> <span class="n">numpy_to_torch</span><span class="p">(</span><span class="n">lm_head</span><span class="o">.</span><span class="n">weight</span><span class="o">.</span><span class="n">raw_value</span><span class="p">)</span>
<span class="n">vocab_embed_weight</span> <span class="o">=</span> <span class="n">numpy_to_torch</span><span class="p">(</span><span class="n">vocab_embedding</span><span class="o">.</span><span class="n">weight</span><span class="o">.</span><span class="n">raw_value</span><span class="p">)</span>
<span class="c1"># The lm_head and vocab_embedding have different weights</span>
<span class="k">if</span> <span class="p">(</span><span class="n">lm_head_weight</span> <span class="o">-</span> <span class="n">vocab_embed_weight</span><span class="p">)</span><span class="o">.</span><span class="n">abs</span><span class="p">()</span><span class="o">.</span><span class="n">max</span><span class="p">()</span><span class="o">.</span><span class="n">item</span><span class="p">()</span> <span class="o">&gt;</span> <span class="mf">1e-6</span><span class="p">:</span>
<span class="k">return</span> <span class="n">model</span>
<span class="n">lm_head</span><span class="o">.</span><span class="n">weight</span> <span class="o">=</span> <span class="n">vocab_embedding</span><span class="o">.</span><span class="n">weight</span>
<span class="k">if</span> <span class="nb">getattr</span><span class="p">(</span><span class="n">lm_head</span><span class="p">,</span> <span class="s1">&#39;per_channel_scale&#39;</span><span class="p">,</span> <span class="kc">None</span><span class="p">)</span> <span class="ow">and</span> <span class="nb">getattr</span><span class="p">(</span>
<span class="n">vocab_embedding</span><span class="p">,</span> <span class="s1">&#39;per_channel_scale&#39;</span><span class="p">,</span> <span class="kc">None</span><span class="p">):</span>
<span class="n">lm_head</span><span class="o">.</span><span class="n">per_channel_scale</span> <span class="o">=</span> <span class="n">vocab_embedding</span><span class="o">.</span><span class="n">per_token_scale</span>
<span class="k">return</span> <span class="n">model</span>
<span class="k">def</span><span class="w"> </span><span class="nf">set_fp8_context_fhma</span><span class="p">(</span><span class="n">model</span><span class="p">:</span> <span class="n">PretrainedModel</span><span class="p">)</span> <span class="o">-&gt;</span> <span class="n">PretrainedModel</span><span class="p">:</span>
<span class="k">for</span> <span class="n">name</span><span class="p">,</span> <span class="n">layer</span> <span class="ow">in</span> <span class="n">model</span><span class="o">.</span><span class="n">named_modules</span><span class="p">():</span>
<span class="k">if</span> <span class="nb">isinstance</span><span class="p">(</span><span class="n">layer</span><span class="p">,</span> <span class="n">Attention</span><span class="p">)</span> <span class="ow">and</span> <span class="nb">hasattr</span><span class="p">(</span>
<span class="n">layer</span><span class="o">.</span><span class="n">dense</span><span class="p">,</span> <span class="s1">&#39;activation_scaling_factor&#39;</span><span class="p">):</span>
<span class="n">scale</span> <span class="o">=</span> <span class="p">[</span><span class="mf">1.0</span><span class="p">]</span> <span class="o">/</span> <span class="n">layer</span><span class="o">.</span><span class="n">dense</span><span class="o">.</span><span class="n">activation_scaling_factor</span><span class="o">.</span><span class="n">raw_value</span>
<span class="n">layer</span><span class="o">.</span><span class="n">attention_output_orig_quant_scale</span> <span class="o">=</span> <span class="n">Parameter</span><span class="p">(</span>
<span class="n">value</span><span class="o">=</span><span class="n">scale</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">dtype</span><span class="o">=</span><span class="s1">&#39;float32&#39;</span><span class="p">)</span>
<span class="k">elif</span> <span class="nb">isinstance</span><span class="p">(</span><span class="n">layer</span><span class="p">,</span> <span class="n">Attention</span><span class="p">)</span> <span class="ow">and</span> <span class="nb">hasattr</span><span class="p">(</span>
<span class="n">layer</span><span class="o">.</span><span class="n">dense</span><span class="p">,</span> <span class="s1">&#39;activation_global_scaling_factor&#39;</span><span class="p">):</span>
<span class="n">scale</span> <span class="o">=</span> <span class="p">[</span><span class="mf">1.0</span>
<span class="p">]</span> <span class="o">/</span> <span class="n">layer</span><span class="o">.</span><span class="n">dense</span><span class="o">.</span><span class="n">activation_global_scaling_factor</span><span class="o">.</span><span class="n">raw_value</span>
<span class="n">layer</span><span class="o">.</span><span class="n">attention_output_orig_quant_scale</span> <span class="o">=</span> <span class="n">Parameter</span><span class="p">(</span>
<span class="n">value</span><span class="o">=</span><span class="n">scale</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">dtype</span><span class="o">=</span><span class="s1">&#39;float32&#39;</span><span class="p">)</span>
<span class="k">return</span> <span class="n">model</span>
<span class="k">def</span><span class="w"> </span><span class="nf">set_fuse_fp4_quant</span><span class="p">(</span><span class="n">model</span><span class="p">:</span> <span class="n">PretrainedModel</span><span class="p">)</span> <span class="o">-&gt;</span> <span class="n">PretrainedModel</span><span class="p">:</span>
<span class="k">for</span> <span class="n">name</span><span class="p">,</span> <span class="n">layer</span> <span class="ow">in</span> <span class="n">model</span><span class="o">.</span><span class="n">named_modules</span><span class="p">():</span>
<span class="k">if</span> <span class="nb">isinstance</span><span class="p">(</span><span class="n">layer</span><span class="p">,</span> <span class="n">Attention</span><span class="p">)</span> <span class="ow">and</span> <span class="nb">hasattr</span><span class="p">(</span>
<span class="n">layer</span><span class="o">.</span><span class="n">dense</span><span class="p">,</span> <span class="s1">&#39;activation_global_scaling_factor&#39;</span><span class="p">):</span>
<span class="n">scale</span> <span class="o">=</span> <span class="p">[</span><span class="mf">1.0</span>
<span class="p">]</span> <span class="o">/</span> <span class="n">layer</span><span class="o">.</span><span class="n">dense</span><span class="o">.</span><span class="n">activation_global_scaling_factor</span><span class="o">.</span><span class="n">raw_value</span>
<span class="n">layer</span><span class="o">.</span><span class="n">attention_output_sf_scale</span> <span class="o">=</span> <span class="n">Parameter</span><span class="p">(</span><span class="n">value</span><span class="o">=</span><span class="n">scale</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">dtype</span><span class="o">=</span><span class="s1">&#39;float32&#39;</span><span class="p">)</span>
<span class="k">return</span> <span class="n">model</span>
<span class="k">def</span><span class="w"> </span><span class="nf">optimize_model</span><span class="p">(</span>
<span class="n">model</span><span class="p">:</span> <span class="n">PretrainedModel</span><span class="p">,</span>
<span class="n">use_parallel_embedding</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">share_embedding_table</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">use_ootb_moe</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">use_fused_mlp</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">gemm_swiglu_plugin_dtype</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="nb">str</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</span><span class="p">,</span>
<span class="n">low_latency_gemm_swiglu_plugin_dtype</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="nb">str</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</span><span class="p">,</span>
<span class="n">use_fused_rg_lru</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">use_unfused_qkv_gemm</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">use_prompt_tuning</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">use_lora</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">max_lora_rank</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="nb">int</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</span><span class="p">,</span>
<span class="n">use_fp8_context_fmha</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">fuse_fp4_quant</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">use_optimize_cross_qkv</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">use_dora</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="p">)</span> <span class="o">-&gt;</span> <span class="n">PretrainedModel</span><span class="p">:</span>
<span class="w"> </span><span class="sd">&quot;&quot;&quot;</span>
<span class="sd"> Run optimization passes on model.</span>
<span class="sd"> There are dependencies between some passes,</span>
<span class="sd"> so we always run passes in the order of arguments to guarantee the execution order.</span>
<span class="sd"> &quot;&quot;&quot;</span>
<span class="c1"># before weight loading</span>
<span class="k">if</span> <span class="n">use_parallel_embedding</span><span class="p">:</span>
<span class="n">model</span> <span class="o">=</span> <span class="n">parallelize_embedding</span><span class="p">(</span><span class="n">model</span><span class="p">)</span>
<span class="k">if</span> <span class="n">share_embedding_table</span><span class="p">:</span>
<span class="c1"># if share_embedding_table is enabled, only one copy of the embedding table is store in converted ckpt</span>
<span class="c1"># this pass is required to make lm_head.weight and vocab_embedding.weight point to the same tensor</span>
<span class="c1"># however even if share_embedding_table is not enabled, trt would still only keep one copy of the table if the weights are identical</span>
<span class="n">model</span> <span class="o">=</span> <span class="n">share_embedding</span><span class="p">(</span><span class="n">model</span><span class="p">)</span>
<span class="c1"># After weight loading</span>
<span class="k">if</span> <span class="n">use_ootb_moe</span><span class="p">:</span>
<span class="n">model</span> <span class="o">=</span> <span class="n">to_ootb_moe</span><span class="p">(</span><span class="n">model</span><span class="p">)</span>
<span class="k">if</span> <span class="n">use_fused_mlp</span><span class="p">:</span>
<span class="n">model</span> <span class="o">=</span> <span class="n">fuse_gate_mlp</span><span class="p">(</span><span class="n">model</span><span class="p">,</span> <span class="n">gemm_swiglu_plugin_dtype</span><span class="p">,</span>
<span class="n">low_latency_gemm_swiglu_plugin_dtype</span><span class="p">)</span>
<span class="k">if</span> <span class="n">use_fused_rg_lru</span><span class="p">:</span>
<span class="n">model</span> <span class="o">=</span> <span class="n">fuse_rg_lru</span><span class="p">(</span><span class="n">model</span><span class="p">)</span>
<span class="k">if</span> <span class="n">use_unfused_qkv_gemm</span><span class="p">:</span>
<span class="n">model</span> <span class="o">=</span> <span class="n">unfuse_qkv_gemm</span><span class="p">(</span><span class="n">model</span><span class="p">)</span>
<span class="k">if</span> <span class="n">use_prompt_tuning</span><span class="p">:</span>
<span class="n">model</span> <span class="o">=</span> <span class="n">set_prompt_tuning</span><span class="p">(</span><span class="n">model</span><span class="p">)</span>
<span class="k">if</span> <span class="n">use_lora</span><span class="p">:</span>
<span class="n">model</span> <span class="o">=</span> <span class="n">add_lora</span><span class="p">(</span><span class="n">model</span><span class="p">,</span> <span class="n">max_lora_rank</span><span class="p">,</span> <span class="n">with_dora</span><span class="o">=</span><span class="n">use_dora</span><span class="p">)</span>
<span class="k">if</span> <span class="n">use_fp8_context_fmha</span><span class="p">:</span>
<span class="n">model</span> <span class="o">=</span> <span class="n">set_fp8_context_fhma</span><span class="p">(</span><span class="n">model</span><span class="p">)</span>
<span class="k">if</span> <span class="n">fuse_fp4_quant</span><span class="p">:</span>
<span class="n">model</span> <span class="o">=</span> <span class="n">set_fuse_fp4_quant</span><span class="p">(</span><span class="n">model</span><span class="p">)</span>
<span class="k">if</span> <span class="ow">not</span> <span class="n">use_lora</span> <span class="ow">and</span> <span class="n">use_optimize_cross_qkv</span> <span class="ow">is</span> <span class="kc">True</span><span class="p">:</span>
<span class="c1"># This optimization is not supported when we use lora</span>
<span class="n">model</span> <span class="o">=</span> <span class="n">optimize_cross_qkv</span><span class="p">(</span><span class="n">model</span><span class="p">)</span>
<span class="k">return</span> <span class="n">model</span>
<span class="k">def</span><span class="w"> </span><span class="nf">optimize_cross_qkv</span><span class="p">(</span><span class="n">model</span><span class="p">):</span>
<span class="w"> </span><span class="sd">&quot;&quot;&quot;</span>
<span class="sd"> For cross attention layer, we can skip computing the query of encoder_output.</span>
<span class="sd"> So, add a new attribute &#39;kv&#39; in the cross_attention layer. This might lead to</span>
<span class="sd"> additional memory cost on model size, but save the memory usage on runtime.</span>
<span class="sd"> Currently, this function only detect the ColumnLinear and FP8Linear. It does not supports</span>
<span class="sd"> other quantization now.</span>
<span class="sd"> &quot;&quot;&quot;</span>
<span class="k">for</span> <span class="n">name</span><span class="p">,</span> <span class="n">attn</span><span class="p">,</span> <span class="n">layer</span> <span class="ow">in</span> <span class="n">model</span><span class="o">.</span><span class="n">named_modules_with_parent</span><span class="p">():</span>
<span class="k">if</span> <span class="nb">isinstance</span><span class="p">(</span><span class="n">attn</span><span class="p">,</span> <span class="n">Attention</span><span class="p">)</span> <span class="ow">and</span> <span class="n">attn</span><span class="o">.</span><span class="n">cross_attention</span> <span class="ow">and</span> \
<span class="p">(</span><span class="nb">type</span><span class="p">(</span><span class="n">attn</span><span class="o">.</span><span class="n">qkv</span><span class="p">)</span> <span class="o">==</span> <span class="n">ColumnLinear</span> <span class="ow">or</span> <span class="nb">type</span><span class="p">(</span><span class="n">attn</span><span class="o">.</span><span class="n">qkv</span><span class="p">)</span> <span class="o">==</span> <span class="n">FP8Linear</span><span class="p">):</span>
<span class="n">old_qkv</span> <span class="o">=</span> <span class="n">attn</span><span class="o">.</span><span class="n">qkv</span>
<span class="n">linear_class</span> <span class="o">=</span> <span class="nb">type</span><span class="p">(</span><span class="n">old_qkv</span><span class="p">)</span>
<span class="n">new_kv</span> <span class="o">=</span> <span class="n">linear_class</span><span class="p">(</span>
<span class="n">in_features</span><span class="o">=</span><span class="n">attn</span><span class="o">.</span><span class="n">hidden_size</span><span class="p">,</span>
<span class="n">out_features</span><span class="o">=</span><span class="mi">2</span> <span class="o">*</span> <span class="n">attn</span><span class="o">.</span><span class="n">tp_size</span> <span class="o">*</span> <span class="n">attn</span><span class="o">.</span><span class="n">num_attention_kv_heads</span> <span class="o">*</span>
<span class="n">attn</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">old_qkv</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">old_qkv</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">old_qkv</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">old_qkv</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="n">old_qkv</span><span class="o">.</span><span class="n">gather_output</span><span class="p">,</span>
<span class="n">prefer_managed_weight</span><span class="o">=</span><span class="n">old_qkv</span><span class="o">.</span><span class="n">prefer_managed_weight</span><span class="p">,</span>
<span class="n">is_qkv</span><span class="o">=</span><span class="n">old_qkv</span><span class="o">.</span><span class="n">is_qkv</span><span class="p">,</span>
<span class="p">)</span>
<span class="n">old_qkv_weight_value</span> <span class="o">=</span> <span class="n">old_qkv</span><span class="o">.</span><span class="n">weight</span><span class="o">.</span><span class="n">raw_value</span>
<span class="k">if</span> <span class="p">(</span><span class="n">old_qkv_weight_value</span><span class="o">.</span><span class="n">shape</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="p">(</span><span class="n">attn</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="n">attn</span><span class="o">.</span><span class="n">num_attention_kv_heads</span><span class="p">)</span> <span class="o">*</span>
<span class="n">attn</span><span class="o">.</span><span class="n">attention_head_size</span><span class="p">,</span> <span class="n">attn</span><span class="o">.</span><span class="n">hidden_size</span>
<span class="p">]))</span><span class="o">.</span><span class="n">all</span><span class="p">():</span>
<span class="n">q_weight</span><span class="p">,</span> <span class="n">kv_weight</span> <span class="o">=</span> <span class="n">np</span><span class="o">.</span><span class="n">array_split</span><span class="p">(</span>
<span class="n">old_qkv_weight_value</span><span class="o">.</span><span class="n">reshape</span><span class="p">(</span>
<span class="n">attn</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="n">attn</span><span class="o">.</span><span class="n">num_attention_kv_heads</span><span class="p">,</span>
<span class="n">attn</span><span class="o">.</span><span class="n">attention_head_size</span><span class="p">,</span> <span class="n">attn</span><span class="o">.</span><span class="n">hidden_size</span><span class="p">),</span>
<span class="p">[</span><span class="n">attn</span><span class="o">.</span><span class="n">num_attention_heads</span><span class="p">],</span>
<span class="n">axis</span><span class="o">=</span><span class="mi">0</span><span class="p">)</span>
<span class="n">new_kv</span><span class="o">.</span><span class="n">weight</span><span class="o">.</span><span class="n">value</span> <span class="o">=</span> <span class="n">kv_weight</span><span class="o">.</span><span class="n">reshape</span><span class="p">([</span>
<span class="mi">2</span> <span class="o">*</span> <span class="n">attn</span><span class="o">.</span><span class="n">num_attention_kv_heads</span> <span class="o">*</span> <span class="n">attn</span><span class="o">.</span><span class="n">attention_head_size</span><span class="p">,</span>
<span class="n">attn</span><span class="o">.</span><span class="n">hidden_size</span>
<span class="p">])</span>
<span class="k">elif</span> <span class="p">(</span><span class="n">old_qkv_weight_value</span><span class="o">.</span><span class="n">shape</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">attn</span><span class="o">.</span><span class="n">hidden_size</span><span class="p">,</span>
<span class="p">(</span><span class="n">attn</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="n">attn</span><span class="o">.</span><span class="n">num_attention_kv_heads</span><span class="p">)</span> <span class="o">*</span>
<span class="n">attn</span><span class="o">.</span><span class="n">attention_head_size</span>
<span class="p">]))</span><span class="o">.</span><span class="n">all</span><span class="p">():</span>
<span class="n">q_weight</span><span class="p">,</span> <span class="n">kv_weight</span> <span class="o">=</span> <span class="n">np</span><span class="o">.</span><span class="n">array_split</span><span class="p">(</span>
<span class="n">old_qkv_weight_value</span><span class="o">.</span><span class="n">reshape</span><span class="p">(</span>
<span class="n">attn</span><span class="o">.</span><span class="n">hidden_size</span><span class="p">,</span> <span class="n">attn</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="n">attn</span><span class="o">.</span><span class="n">num_attention_kv_heads</span><span class="p">,</span>
<span class="n">attn</span><span class="o">.</span><span class="n">attention_head_size</span><span class="p">),</span> <span class="p">[</span><span class="n">attn</span><span class="o">.</span><span class="n">num_attention_heads</span><span class="p">],</span>
<span class="n">axis</span><span class="o">=</span><span class="mi">1</span><span class="p">)</span>
<span class="n">new_kv</span><span class="o">.</span><span class="n">weight</span><span class="o">.</span><span class="n">value</span> <span class="o">=</span> <span class="n">kv_weight</span><span class="o">.</span><span class="n">reshape</span><span class="p">([</span>
<span class="n">attn</span><span class="o">.</span><span class="n">hidden_size</span><span class="p">,</span>
<span class="mi">2</span> <span class="o">*</span> <span class="n">attn</span><span class="o">.</span><span class="n">num_attention_kv_heads</span> <span class="o">*</span> <span class="n">attn</span><span class="o">.</span><span class="n">attention_head_size</span>
<span class="p">])</span>
<span class="k">else</span><span class="p">:</span>
<span class="k">assert</span> <span class="kc">False</span>
<span class="k">if</span> <span class="nb">isinstance</span><span class="p">(</span><span class="n">attn</span><span class="o">.</span><span class="n">qkv</span><span class="p">,</span> <span class="n">FP8Linear</span><span class="p">):</span>
<span class="n">new_kv</span><span class="o">.</span><span class="n">activation_scaling_factor</span><span class="o">.</span><span class="n">value</span> <span class="o">=</span> <span class="n">old_qkv</span><span class="o">.</span><span class="n">activation_scaling_factor</span><span class="o">.</span><span class="n">raw_value</span>
<span class="n">new_kv</span><span class="o">.</span><span class="n">weights_scaling_factor</span><span class="o">.</span><span class="n">value</span> <span class="o">=</span> <span class="n">old_qkv</span><span class="o">.</span><span class="n">weights_scaling_factor</span><span class="o">.</span><span class="n">raw_value</span>
<span class="k">if</span> <span class="n">old_qkv</span><span class="o">.</span><span class="n">bias</span><span class="p">:</span>
<span class="n">q_bias</span><span class="p">,</span> <span class="n">kv_bias</span> <span class="o">=</span> <span class="n">np</span><span class="o">.</span><span class="n">array_split</span><span class="p">(</span><span class="n">old_qkv</span><span class="o">.</span><span class="n">bias</span><span class="o">.</span><span class="n">raw_value</span><span class="o">.</span><span class="n">reshape</span><span class="p">(</span>
<span class="n">attn</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="n">attn</span><span class="o">.</span><span class="n">num_attention_kv_heads</span><span class="p">,</span>
<span class="n">attn</span><span class="o">.</span><span class="n">attention_head_size</span><span class="p">),</span> <span class="p">[</span><span class="n">attn</span><span class="o">.</span><span class="n">num_attention_heads</span><span class="p">],</span>
<span class="n">axis</span><span class="o">=</span><span class="mi">0</span><span class="p">)</span>
<span class="n">new_kv</span><span class="o">.</span><span class="n">bias</span><span class="o">.</span><span class="n">value</span> <span class="o">=</span> <span class="n">kv_bias</span><span class="o">.</span><span class="n">reshape</span><span class="p">([</span>
<span class="mi">2</span> <span class="o">*</span> <span class="n">attn</span><span class="o">.</span><span class="n">num_attention_kv_heads</span> <span class="o">*</span> <span class="n">attn</span><span class="o">.</span><span class="n">attention_head_size</span>
<span class="p">])</span>
<span class="nb">setattr</span><span class="p">(</span><span class="n">attn</span><span class="p">,</span> <span class="s2">&quot;kv&quot;</span><span class="p">,</span> <span class="n">new_kv</span><span class="p">)</span>
<span class="k">return</span> <span class="n">model</span>
<span class="k">def</span><span class="w"> </span><span class="nf">preprocess_perlayer_weights</span><span class="p">(</span><span class="n">weights</span><span class="p">,</span>
<span class="n">model_config</span><span class="p">,</span>
<span class="n">quant_algo</span><span class="p">,</span>
<span class="n">from_pruned</span><span class="o">=</span><span class="kc">False</span><span class="p">):</span>
<span class="n">exclude_modules</span> <span class="o">=</span> <span class="n">model_config</span><span class="o">.</span><span class="n">quantization</span><span class="o">.</span><span class="n">exclude_modules</span>
<span class="c1"># INT4_AWQ</span>
<span class="k">if</span> <span class="n">quant_algo</span> <span class="o">==</span> <span class="n">QuantAlgo</span><span class="o">.</span><span class="n">W4A8_AWQ</span> <span class="ow">or</span> <span class="n">quant_algo</span> <span class="o">==</span> <span class="n">QuantAlgo</span><span class="o">.</span><span class="n">W4A16_AWQ</span><span class="p">:</span>
<span class="n">preprocessor</span> <span class="o">=</span> <span class="n">preprocess_weights_for_mixed_gemm</span>
<span class="k">if</span> <span class="n">quant_algo</span> <span class="o">==</span> <span class="n">QuantAlgo</span><span class="o">.</span><span class="n">W4A8_AWQ</span><span class="p">:</span>
<span class="n">activation_type</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">float8_e4m3fn</span>
<span class="k">elif</span> <span class="n">quant_algo</span> <span class="o">==</span> <span class="n">QuantAlgo</span><span class="o">.</span><span class="n">W4A16_AWQ</span><span class="p">:</span>
<span class="n">activation_type</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">float16</span>
<span class="k">for</span> <span class="n">name</span><span class="p">,</span> <span class="n">param</span> <span class="ow">in</span> <span class="n">weights</span><span class="o">.</span><span class="n">items</span><span class="p">():</span>
<span class="k">if</span> <span class="n">from_pruned</span> <span class="ow">and</span> <span class="n">param</span><span class="o">.</span><span class="n">numel</span><span class="p">()</span> <span class="o">==</span> <span class="mi">0</span><span class="p">:</span>
<span class="k">continue</span>
<span class="k">if</span> <span class="n">name</span><span class="o">.</span><span class="n">endswith</span><span class="p">(</span><span class="s1">&#39;weight&#39;</span><span class="p">)</span> <span class="ow">and</span> <span class="n">param</span><span class="o">.</span><span class="n">dtype</span> <span class="o">==</span> <span class="n">torch</span><span class="o">.</span><span class="n">int8</span><span class="p">:</span>
<span class="n">dtype</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">float16</span>
<span class="k">if</span> <span class="n">model_config</span><span class="o">.</span><span class="n">dtype</span> <span class="o">==</span> <span class="s2">&quot;bfloat16&quot;</span><span class="p">:</span>
<span class="n">dtype</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">bfloat16</span>
<span class="n">weights</span><span class="p">[</span><span class="n">name</span><span class="p">]</span> <span class="o">=</span> <span class="n">preprocessor</span><span class="p">(</span><span class="n">param</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="n">torch</span><span class="o">.</span><span class="n">quint4x2</span><span class="p">,</span>
<span class="n">activation_type</span><span class="p">)</span><span class="o">.</span><span class="n">view</span><span class="p">(</span><span class="n">dtype</span><span class="p">)</span>
<span class="k">if</span> <span class="n">name</span><span class="o">.</span><span class="n">endswith</span><span class="p">(</span><span class="s1">&#39;weights_scaling_factor&#39;</span><span class="p">):</span>
<span class="n">weights</span><span class="p">[</span><span class="n">name</span><span class="p">]</span> <span class="o">=</span> <span class="n">param</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="o">.</span><span class="n">contiguous</span><span class="p">()</span><span class="o">.</span><span class="n">to</span><span class="p">(</span>
<span class="n">str_dtype_to_torch</span><span class="p">(</span><span class="n">model_config</span><span class="o">.</span><span class="n">dtype</span><span class="p">))</span>
<span class="k">if</span> <span class="n">name</span><span class="o">.</span><span class="n">endswith</span><span class="p">(</span><span class="s1">&#39;prequant_scaling_factor&#39;</span><span class="p">):</span>
<span class="k">if</span> <span class="nb">len</span><span class="p">(</span><span class="n">weights</span><span class="p">[</span><span class="n">name</span><span class="p">]</span><span class="o">.</span><span class="n">shape</span><span class="p">)</span> <span class="o">==</span> <span class="mi">2</span><span class="p">:</span>
<span class="c1"># MoE experts share the same scaling factor.</span>
<span class="n">param</span> <span class="o">=</span> <span class="n">param</span><span class="p">[</span><span class="mi">0</span><span class="p">,</span> <span class="p">:]</span>
<span class="n">weights</span><span class="p">[</span><span class="n">name</span><span class="p">]</span> <span class="o">=</span> <span class="n">param</span><span class="o">.</span><span class="n">reshape</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="k">if</span> <span class="n">model_config</span><span class="o">.</span><span class="n">mapping</span><span class="o">.</span><span class="n">tp_rank</span> <span class="o">&gt;</span> <span class="mi">0</span><span class="p">:</span>
<span class="k">if</span> <span class="n">name</span><span class="o">.</span><span class="n">endswith</span><span class="p">(</span><span class="s1">&#39;attention.dense.bias&#39;</span><span class="p">)</span> <span class="ow">or</span> <span class="n">name</span><span class="o">.</span><span class="n">endswith</span><span class="p">(</span>
<span class="s1">&#39;mlp.proj.bias&#39;</span><span class="p">):</span>
<span class="n">weights</span><span class="p">[</span><span class="n">name</span><span class="p">]</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">zeros_like</span><span class="p">(</span><span class="n">param</span><span class="p">)</span>
<span class="k">if</span> <span class="n">quant_algo</span> <span class="o">==</span> <span class="n">QuantAlgo</span><span class="o">.</span><span class="n">W4A8_AWQ</span><span class="p">:</span>
<span class="k">for</span> <span class="n">name</span> <span class="ow">in</span> <span class="nb">list</span><span class="p">(</span><span class="n">weights</span><span class="p">):</span>
<span class="k">if</span> <span class="n">name</span><span class="o">.</span><span class="n">endswith</span><span class="p">(</span><span class="s1">&#39;weights_scaling_factor&#39;</span><span class="p">):</span>
<span class="n">activation_scaling_factor</span> <span class="o">=</span> <span class="n">weights</span><span class="o">.</span><span class="n">pop</span><span class="p">(</span>
<span class="n">name</span><span class="o">.</span><span class="n">replace</span><span class="p">(</span><span class="s1">&#39;weights_scaling_factor&#39;</span><span class="p">,</span>
<span class="s1">&#39;activation_scaling_factor&#39;</span><span class="p">))</span>
<span class="n">weights_scaling_factor_2</span> <span class="o">=</span> <span class="n">weights</span><span class="o">.</span><span class="n">pop</span><span class="p">(</span>
<span class="n">name</span><span class="o">.</span><span class="n">replace</span><span class="p">(</span><span class="s1">&#39;weights_scaling_factor&#39;</span><span class="p">,</span>
<span class="s1">&#39;weights_scaling_factor_2&#39;</span><span class="p">))</span>
<span class="n">weights</span><span class="p">[</span><span class="n">name</span><span class="p">]</span> <span class="o">/=</span> <span class="n">weights_scaling_factor_2</span>
<span class="n">weights</span><span class="p">[</span><span class="n">name</span><span class="p">]</span> <span class="o">=</span> <span class="n">weights</span><span class="p">[</span><span class="n">name</span><span class="p">]</span><span class="o">.</span><span class="n">to</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">float16</span><span class="p">)</span><span class="o">.</span><span class="n">view</span><span class="p">(</span>
<span class="n">str_dtype_to_torch</span><span class="p">(</span><span class="n">model_config</span><span class="o">.</span><span class="n">dtype</span><span class="p">))</span>
<span class="n">weights</span><span class="p">[</span><span class="n">name</span><span class="o">.</span><span class="n">replace</span><span class="p">(</span>
<span class="s1">&#39;weights_scaling_factor&#39;</span><span class="p">,</span>
<span class="s1">&#39;prequant_scaling_factor&#39;</span><span class="p">)]</span> <span class="o">/=</span> <span class="n">activation_scaling_factor</span>
<span class="n">weights</span><span class="p">[</span><span class="n">name</span><span class="o">.</span><span class="n">replace</span><span class="p">(</span>
<span class="s1">&#39;weights_scaling_factor&#39;</span><span class="p">,</span> <span class="s1">&#39;alpha&#39;</span>
<span class="p">)]</span> <span class="o">=</span> <span class="n">activation_scaling_factor</span> <span class="o">*</span> <span class="n">weights_scaling_factor_2</span>
<span class="n">weights</span><span class="p">[</span><span class="n">name</span><span class="o">.</span><span class="n">replace</span><span class="p">(</span><span class="s1">&#39;weights_scaling_factor&#39;</span><span class="p">,</span>
<span class="s1">&#39;activation_scaling_factor&#39;</span>
<span class="p">)]</span> <span class="o">=</span> <span class="n">activation_scaling_factor</span>
<span class="c1"># FP8</span>
<span class="k">elif</span> <span class="n">quant_algo</span> <span class="o">==</span> <span class="n">QuantAlgo</span><span class="o">.</span><span class="n">FP8</span><span class="p">:</span>
<span class="k">for</span> <span class="n">name</span><span class="p">,</span> <span class="n">param</span> <span class="ow">in</span> <span class="n">weights</span><span class="o">.</span><span class="n">items</span><span class="p">():</span>
<span class="k">if</span> <span class="n">name</span><span class="o">.</span><span class="n">endswith</span><span class="p">(</span><span class="s1">&#39;weight&#39;</span><span class="p">)</span> <span class="ow">and</span> <span class="n">param</span><span class="o">.</span><span class="n">dtype</span> <span class="o">==</span> <span class="n">torch</span><span class="o">.</span><span class="n">int8</span><span class="p">:</span>
<span class="n">weights</span><span class="p">[</span><span class="n">name</span><span class="p">]</span> <span class="o">=</span> <span class="n">param</span><span class="o">.</span><span class="n">view</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">float8_e4m3fn</span><span class="p">)</span>
<span class="c1"># lm_head is not always quantized to FP8</span>
<span class="k">if</span> <span class="s2">&quot;lm_head.weight&quot;</span> <span class="ow">in</span> <span class="n">weights</span> <span class="ow">and</span> <span class="n">weights</span><span class="p">[</span>
<span class="s1">&#39;lm_head.weight&#39;</span><span class="p">]</span><span class="o">.</span><span class="n">dtype</span> <span class="ow">is</span> <span class="ow">not</span> <span class="n">torch</span><span class="o">.</span><span class="n">float8_e4m3fn</span><span class="p">:</span>
<span class="n">weights</span><span class="o">.</span><span class="n">pop</span><span class="p">(</span><span class="s1">&#39;lm_head.weights_scaling_factor&#39;</span><span class="p">,</span> <span class="kc">None</span><span class="p">)</span>
<span class="n">weights</span><span class="o">.</span><span class="n">pop</span><span class="p">(</span><span class="s1">&#39;lm_head.activation_scaling_factor&#39;</span><span class="p">,</span> <span class="kc">None</span><span class="p">)</span>
<span class="k">elif</span> <span class="n">quant_algo</span> <span class="o">==</span> <span class="n">QuantAlgo</span><span class="o">.</span><span class="n">FP8_PER_CHANNEL_PER_TOKEN</span><span class="p">:</span>
<span class="k">for</span> <span class="n">name</span><span class="p">,</span> <span class="n">param</span> <span class="ow">in</span> <span class="n">weights</span><span class="o">.</span><span class="n">items</span><span class="p">():</span>
<span class="k">if</span> <span class="n">name</span><span class="o">.</span><span class="n">endswith</span><span class="p">(</span><span class="s1">&#39;weight&#39;</span><span class="p">)</span> <span class="ow">and</span> <span class="n">param</span><span class="o">.</span><span class="n">dtype</span> <span class="o">==</span> <span class="n">torch</span><span class="o">.</span><span class="n">int8</span><span class="p">:</span>
<span class="n">weights</span><span class="p">[</span><span class="n">name</span><span class="p">]</span> <span class="o">=</span> <span class="n">param</span><span class="o">.</span><span class="n">view</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">float8_e4m3fn</span><span class="p">)</span>
<span class="c1"># lm_head is not quantized to FP8</span>
<span class="k">if</span> <span class="s2">&quot;lm_head.weight&quot;</span> <span class="ow">in</span> <span class="n">weights</span><span class="p">:</span>
<span class="k">assert</span> <span class="n">weights</span><span class="p">[</span><span class="s1">&#39;lm_head.weight&#39;</span><span class="p">]</span><span class="o">.</span><span class="n">dtype</span> <span class="o">==</span> <span class="n">str_dtype_to_torch</span><span class="p">(</span>
<span class="n">model_config</span><span class="o">.</span><span class="n">dtype</span><span class="p">)</span>
<span class="n">weights</span><span class="o">.</span><span class="n">pop</span><span class="p">(</span><span class="s1">&#39;lm_head.weights_scaling_factor&#39;</span><span class="p">,</span> <span class="kc">None</span><span class="p">)</span>
<span class="n">weights</span><span class="o">.</span><span class="n">pop</span><span class="p">(</span><span class="s1">&#39;lm_head.activation_scaling_factor&#39;</span><span class="p">,</span> <span class="kc">None</span><span class="p">)</span>
<span class="c1"># FP4</span>
<span class="k">elif</span> <span class="n">quant_algo</span> <span class="o">==</span> <span class="n">QuantAlgo</span><span class="o">.</span><span class="n">NVFP4</span><span class="p">:</span>
<span class="c1"># Interleave block scale for NVFP4 plugin.</span>
<span class="k">for</span> <span class="n">name</span> <span class="ow">in</span> <span class="nb">list</span><span class="p">(</span><span class="n">weights</span><span class="p">):</span>
<span class="k">if</span> <span class="n">name</span><span class="o">.</span><span class="n">endswith</span><span class="p">(</span><span class="s1">&#39;weights_scaling_factor&#39;</span><span class="p">):</span>
<span class="n">out_features</span><span class="p">,</span> <span class="n">in_features</span> <span class="o">=</span> <span class="n">weights</span><span class="p">[</span><span class="n">name</span><span class="p">]</span><span class="o">.</span><span class="n">shape</span>
<span class="n">nrows</span> <span class="o">=</span> <span class="n">fp4_utils</span><span class="o">.</span><span class="n">pad_up</span><span class="p">(</span><span class="n">out_features</span><span class="p">,</span> <span class="mi">128</span><span class="p">)</span>
<span class="n">ncols</span> <span class="o">=</span> <span class="n">fp4_utils</span><span class="o">.</span><span class="n">pad_up</span><span class="p">(</span><span class="n">in_features</span><span class="p">,</span> <span class="mi">4</span><span class="p">)</span>
<span class="n">new_name</span> <span class="o">=</span> <span class="n">name</span><span class="o">.</span><span class="n">replace</span><span class="p">(</span><span class="s1">&#39;weights_scaling_factor&#39;</span><span class="p">,</span>
<span class="s1">&#39;weights_block_scaling_factor&#39;</span><span class="p">)</span>
<span class="n">weights</span><span class="p">[</span><span class="n">new_name</span><span class="p">]</span> <span class="o">=</span> <span class="n">weights</span><span class="p">[</span><span class="n">name</span><span class="p">]</span>
<span class="n">weights</span><span class="p">[</span>
<span class="n">new_name</span> <span class="o">+</span>
<span class="s2">&quot;_interleaved&quot;</span><span class="p">]</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">ops</span><span class="o">.</span><span class="n">tensorrt_llm</span><span class="o">.</span><span class="n">nvfp4_block_scale_interleave</span><span class="p">(</span>
<span class="n">weights</span><span class="p">[</span><span class="n">name</span><span class="p">]</span><span class="o">.</span><span class="n">view</span><span class="p">(</span><span class="n">fp4_utils</span><span class="o">.</span><span class="n">float4_sf_dtype</span><span class="p">)</span><span class="o">.</span><span class="n">cpu</span><span class="p">(</span>
<span class="p">)</span><span class="o">.</span><span class="n">contiguous</span><span class="p">())</span><span class="o">.</span><span class="n">reshape</span><span class="p">(</span><span class="n">nrows</span><span class="p">,</span> <span class="n">ncols</span><span class="p">)</span><span class="o">.</span><span class="n">view</span><span class="p">(</span>
<span class="n">fp4_utils</span><span class="o">.</span><span class="n">float4_sf_dtype</span><span class="p">)</span>
<span class="n">weights</span><span class="o">.</span><span class="n">pop</span><span class="p">(</span><span class="n">name</span><span class="p">)</span>
<span class="k">if</span> <span class="n">name</span><span class="o">.</span><span class="n">endswith</span><span class="p">(</span><span class="s1">&#39;weights_scaling_factor_2&#39;</span><span class="p">):</span>
<span class="n">new_name</span> <span class="o">=</span> <span class="n">name</span><span class="o">.</span><span class="n">replace</span><span class="p">(</span><span class="s1">&#39;weights_scaling_factor_2&#39;</span><span class="p">,</span>
<span class="s1">&#39;weights_global_scaling_factor&#39;</span><span class="p">)</span>
<span class="n">weights</span><span class="p">[</span><span class="n">new_name</span><span class="p">]</span> <span class="o">=</span> <span class="n">weights</span><span class="p">[</span><span class="n">name</span><span class="p">]</span>
<span class="n">weights</span><span class="o">.</span><span class="n">pop</span><span class="p">(</span><span class="n">name</span><span class="p">)</span>
<span class="k">if</span> <span class="n">name</span><span class="o">.</span><span class="n">endswith</span><span class="p">(</span><span class="s1">&#39;activation_scaling_factor&#39;</span><span class="p">):</span>
<span class="n">new_name</span> <span class="o">=</span> <span class="n">name</span><span class="o">.</span><span class="n">replace</span><span class="p">(</span><span class="s1">&#39;activation_scaling_factor&#39;</span><span class="p">,</span>
<span class="s1">&#39;activation_global_scaling_factor&#39;</span><span class="p">)</span>
<span class="n">weights</span><span class="p">[</span><span class="n">new_name</span><span class="p">]</span> <span class="o">=</span> <span class="n">weights</span><span class="p">[</span><span class="n">name</span><span class="p">]</span>
<span class="n">weights</span><span class="o">.</span><span class="n">pop</span><span class="p">(</span><span class="n">name</span><span class="p">)</span>
<span class="k">for</span> <span class="n">name</span> <span class="ow">in</span> <span class="nb">list</span><span class="p">(</span><span class="n">weights</span><span class="p">):</span>
<span class="k">if</span> <span class="n">name</span><span class="o">.</span><span class="n">endswith</span><span class="p">(</span><span class="s1">&#39;weights_global_scaling_factor&#39;</span><span class="p">):</span>
<span class="n">weight_global_sf</span> <span class="o">=</span> <span class="n">weights</span><span class="p">[</span><span class="n">name</span><span class="p">]</span>
<span class="n">act_global_sf</span> <span class="o">=</span> <span class="n">weights</span><span class="p">[</span><span class="n">name</span><span class="o">.</span><span class="n">replace</span><span class="p">(</span>
<span class="s1">&#39;weights_global_scaling_factor&#39;</span><span class="p">,</span>
<span class="s1">&#39;activation_global_scaling_factor&#39;</span><span class="p">)]</span>
<span class="n">weights</span><span class="p">[</span><span class="n">name</span><span class="o">.</span><span class="n">replace</span><span class="p">(</span>
<span class="s1">&#39;weights_global_scaling_factor&#39;</span><span class="p">,</span>
<span class="s1">&#39;alpha&#39;</span><span class="p">)]</span> <span class="o">=</span> <span class="n">act_global_sf</span> <span class="o">*</span> <span class="n">weight_global_sf</span>
<span class="k">elif</span> <span class="n">quant_algo</span> <span class="ow">in</span> <span class="p">[</span><span class="n">QuantAlgo</span><span class="o">.</span><span class="n">W4A16</span><span class="p">,</span> <span class="n">QuantAlgo</span><span class="o">.</span><span class="n">W8A16</span><span class="p">]:</span>
<span class="n">weights</span> <span class="o">=</span> <span class="n">weight_only_quantize_dict</span><span class="p">(</span><span class="n">weights</span><span class="o">=</span><span class="n">weights</span><span class="p">,</span>
<span class="n">quant_algo</span><span class="o">=</span><span class="n">quant_algo</span><span class="p">,</span>
<span class="n">exclude_modules</span><span class="o">=</span><span class="n">exclude_modules</span><span class="p">,</span>
<span class="n">plugin</span><span class="o">=</span><span class="kc">True</span><span class="p">)</span>
<span class="k">def</span><span class="w"> </span><span class="nf">preprocess_weights</span><span class="p">(</span><span class="n">weights</span><span class="p">:</span> <span class="n">Dict</span><span class="p">[</span><span class="nb">str</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="n">model_config</span><span class="p">:</span> <span class="n">PretrainedConfig</span><span class="p">,</span>
<span class="n">from_pruned</span><span class="o">=</span><span class="kc">False</span><span class="p">)</span> <span class="o">-&gt;</span> <span class="kc">None</span><span class="p">:</span>
<span class="w"> </span><span class="sd">&quot;&quot;&quot;This function in-place modifies weights and model_config, making them compatible with each other.</span>
<span class="sd"> Note: Typically, it should be called before model creation and weight loading. For example,</span>
<span class="sd"> preprocess_weights(weights, model_config)</span>
<span class="sd"> model = XXXForCausalLM(model_config)</span>
<span class="sd"> model.load(weights)</span>
<span class="sd"> &quot;&quot;&quot;</span>
<span class="n">quant_config</span> <span class="o">=</span> <span class="n">model_config</span><span class="o">.</span><span class="n">quantization</span>
<span class="n">quant_algo</span> <span class="o">=</span> <span class="n">quant_config</span><span class="o">.</span><span class="n">quant_algo</span>
<span class="n">pattern_info</span> <span class="o">=</span> <span class="p">[</span><span class="s1">&#39;fc&#39;</span><span class="p">,</span> <span class="s1">&#39;gate&#39;</span><span class="p">,</span> <span class="s1">&#39;proj&#39;</span><span class="p">,</span> <span class="s1">&#39;qkv&#39;</span><span class="p">,</span> <span class="s1">&#39;dense&#39;</span><span class="p">]</span>
<span class="k">def</span><span class="w"> </span><span class="nf">process_kv_scaling_factor</span><span class="p">(</span><span class="n">weights</span><span class="p">:</span> <span class="n">Dict</span><span class="p">[</span><span class="nb">str</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="n">new_entries</span> <span class="o">=</span> <span class="p">{}</span>
<span class="n">names_to_delete</span> <span class="o">=</span> <span class="nb">set</span><span class="p">()</span>
<span class="c1"># If k, v cache scaling factors are stored separately, combine them into kv cache scaling factor.</span>
<span class="k">for</span> <span class="n">name</span><span class="p">,</span> <span class="n">param</span> <span class="ow">in</span> <span class="n">weights</span><span class="o">.</span><span class="n">items</span><span class="p">():</span>
<span class="k">if</span> <span class="n">name</span><span class="o">.</span><span class="n">endswith</span><span class="p">(</span><span class="s1">&#39;.k_cache_scaling_factor&#39;</span><span class="p">):</span>
<span class="n">v_name</span> <span class="o">=</span> <span class="n">name</span><span class="o">.</span><span class="n">replace</span><span class="p">(</span><span class="s1">&#39;k_cache_scaling_factor&#39;</span><span class="p">,</span>
<span class="s1">&#39;v_cache_scaling_factor&#39;</span><span class="p">)</span>
<span class="k">assert</span> <span class="n">v_name</span> <span class="ow">in</span> <span class="n">weights</span><span class="p">,</span> <span class="sa">f</span><span class="s2">&quot;</span><span class="si">{</span><span class="n">v_name</span><span class="si">}</span><span class="s2"> not found&quot;</span>
<span class="n">kv_name</span> <span class="o">=</span> <span class="n">name</span><span class="o">.</span><span class="n">replace</span><span class="p">(</span><span class="s1">&#39;k_cache_scaling_factor&#39;</span><span class="p">,</span>
<span class="s1">&#39;kv_cache_scaling_factor&#39;</span><span class="p">)</span>
<span class="n">new_entries</span><span class="p">[</span><span class="n">kv_name</span><span class="p">]</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">max</span><span class="p">(</span><span class="n">weights</span><span class="p">[</span><span class="n">name</span><span class="p">],</span> <span class="n">weights</span><span class="p">[</span><span class="n">v_name</span><span class="p">])</span>
<span class="n">names_to_delete</span><span class="o">.</span><span class="n">update</span><span class="p">([</span><span class="n">name</span><span class="p">,</span> <span class="n">v_name</span><span class="p">])</span>
<span class="n">weights</span><span class="o">.</span><span class="n">update</span><span class="p">(</span><span class="n">new_entries</span><span class="p">)</span>
<span class="k">for</span> <span class="n">k</span> <span class="ow">in</span> <span class="n">names_to_delete</span><span class="p">:</span>
<span class="k">del</span> <span class="n">weights</span><span class="p">[</span><span class="n">k</span><span class="p">]</span>
<span class="n">new_entries</span> <span class="o">=</span> <span class="p">[]</span>
<span class="c1"># The unified converter generate_tllm_weights() already generates these rcp weights, but legacy</span>
<span class="c1"># converters do not. Handle it here.</span>
<span class="k">for</span> <span class="n">name</span><span class="p">,</span> <span class="n">param</span> <span class="ow">in</span> <span class="n">weights</span><span class="o">.</span><span class="n">items</span><span class="p">():</span>
<span class="k">if</span> <span class="n">name</span><span class="o">.</span><span class="n">endswith</span><span class="p">(</span><span class="s1">&#39;.kv_cache_scaling_factor&#39;</span><span class="p">):</span>
<span class="n">rcp_name</span> <span class="o">=</span> <span class="n">name</span><span class="o">.</span><span class="n">replace</span><span class="p">(</span><span class="s1">&#39;kv_cache_scaling_factor&#39;</span><span class="p">,</span>
<span class="s1">&#39;kv_cache_rcp_scaling_factor&#39;</span><span class="p">)</span>
<span class="k">if</span> <span class="n">rcp_name</span> <span class="ow">not</span> <span class="ow">in</span> <span class="n">weights</span><span class="p">:</span>
<span class="n">new_entries</span><span class="o">.</span><span class="n">append</span><span class="p">((</span><span class="n">rcp_name</span><span class="p">,</span> <span class="n">torch</span><span class="o">.</span><span class="n">reciprocal</span><span class="p">(</span><span class="n">param</span><span class="p">)))</span>
<span class="n">weights</span><span class="o">.</span><span class="n">update</span><span class="p">(</span><span class="n">new_entries</span><span class="p">)</span>
<span class="n">process_kv_scaling_factor</span><span class="p">(</span><span class="n">weights</span><span class="p">)</span>
<span class="n">per_layer_weights</span> <span class="o">=</span> <span class="p">{}</span>
<span class="k">for</span> <span class="n">name</span><span class="p">,</span> <span class="n">param</span> <span class="ow">in</span> <span class="n">weights</span><span class="o">.</span><span class="n">items</span><span class="p">():</span>
<span class="n">in_mode</span> <span class="o">=</span> <span class="kc">False</span>
<span class="k">for</span> <span class="n">info</span> <span class="ow">in</span> <span class="n">pattern_info</span><span class="p">:</span>
<span class="n">pattern</span> <span class="o">=</span> <span class="sa">rf</span><span class="s1">&#39;(.*?</span><span class="si">{</span><span class="n">info</span><span class="si">}</span><span class="s1">.*?)&#39;</span>
<span class="n">pattern_match</span> <span class="o">=</span> <span class="n">re</span><span class="o">.</span><span class="n">match</span><span class="p">(</span><span class="n">pattern</span><span class="p">,</span> <span class="n">name</span><span class="p">)</span>
<span class="k">if</span> <span class="n">pattern_match</span><span class="p">:</span>
<span class="n">base_name</span> <span class="o">=</span> <span class="n">pattern_match</span><span class="o">.</span><span class="n">group</span><span class="p">(</span><span class="mi">1</span><span class="p">)</span>
<span class="k">if</span> <span class="n">base_name</span> <span class="ow">not</span> <span class="ow">in</span> <span class="n">per_layer_weights</span><span class="o">.</span><span class="n">keys</span><span class="p">():</span>
<span class="n">per_layer_weights</span><span class="p">[</span><span class="n">base_name</span><span class="p">]</span> <span class="o">=</span> <span class="p">{}</span>
<span class="n">per_layer_weights</span><span class="p">[</span><span class="n">base_name</span><span class="p">][</span><span class="n">name</span><span class="p">]</span> <span class="o">=</span> <span class="n">param</span>
<span class="n">in_mode</span> <span class="o">=</span> <span class="kc">True</span>
<span class="k">break</span>
<span class="k">if</span> <span class="ow">not</span> <span class="n">in_mode</span><span class="p">:</span>
<span class="c1"># [lm_head.weight, ln_f.weight, vocab_embedding.weight]</span>
<span class="n">base_name</span> <span class="o">=</span> <span class="n">name</span><span class="o">.</span><span class="n">rsplit</span><span class="p">(</span><span class="s1">&#39;.&#39;</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">if</span> <span class="n">base_name</span> <span class="ow">not</span> <span class="ow">in</span> <span class="n">per_layer_weights</span><span class="o">.</span><span class="n">keys</span><span class="p">():</span>
<span class="n">per_layer_weights</span><span class="p">[</span><span class="n">base_name</span><span class="p">]</span> <span class="o">=</span> <span class="p">{}</span>
<span class="n">per_layer_weights</span><span class="p">[</span><span class="n">base_name</span><span class="p">][</span><span class="n">name</span><span class="p">]</span> <span class="o">=</span> <span class="n">param</span>
<span class="n">new_weights</span> <span class="o">=</span> <span class="p">{}</span>
<span class="k">for</span> <span class="n">base_name</span><span class="p">,</span> <span class="n">layer_weights</span> <span class="ow">in</span> <span class="n">per_layer_weights</span><span class="o">.</span><span class="n">items</span><span class="p">():</span>
<span class="k">if</span> <span class="n">quant_algo</span> <span class="o">!=</span> <span class="n">QuantAlgo</span><span class="o">.</span><span class="n">MIXED_PRECISION</span><span class="p">:</span>
<span class="n">layer_quant_algo</span> <span class="o">=</span> <span class="n">quant_algo</span>
<span class="k">else</span><span class="p">:</span>
<span class="n">quant_cfg</span> <span class="o">=</span> <span class="n">quant_config</span><span class="o">.</span><span class="n">_get_quant_cfg</span><span class="p">(</span><span class="n">base_name</span><span class="p">)</span>
<span class="k">if</span> <span class="ow">not</span> <span class="n">quant_cfg</span><span class="o">.</span><span class="n">quant_algo</span><span class="p">:</span>
<span class="n">new_weights</span><span class="o">.</span><span class="n">update</span><span class="p">(</span><span class="n">layer_weights</span><span class="p">)</span>
<span class="k">continue</span>
<span class="n">layer_quant_algo</span> <span class="o">=</span> <span class="n">quant_cfg</span><span class="o">.</span><span class="n">quant_algo</span>
<span class="n">preprocess_perlayer_weights</span><span class="p">(</span><span class="n">layer_weights</span><span class="p">,</span> <span class="n">model_config</span><span class="p">,</span>
<span class="n">layer_quant_algo</span><span class="p">,</span> <span class="n">from_pruned</span><span class="p">)</span>
<span class="n">new_weights</span><span class="o">.</span><span class="n">update</span><span class="p">(</span><span class="n">layer_weights</span><span class="p">)</span>
<span class="n">weights</span> <span class="o">=</span> <span class="n">new_weights</span>
<span class="k">for</span> <span class="n">name</span><span class="p">,</span> <span class="n">param</span> <span class="ow">in</span> <span class="n">weights</span><span class="o">.</span><span class="n">items</span><span class="p">():</span>
<span class="k">if</span> <span class="n">model_config</span><span class="o">.</span><span class="n">architecture</span> <span class="o">==</span> <span class="s1">&#39;GPTJForCausalLM&#39;</span><span class="p">:</span>
<span class="k">if</span> <span class="n">model_config</span><span class="o">.</span><span class="n">mapping</span><span class="o">.</span><span class="n">tp_rank</span> <span class="o">&gt;</span> <span class="mi">0</span><span class="p">:</span>
<span class="k">if</span> <span class="s1">&#39;attention.dense.bias&#39;</span> <span class="ow">in</span> <span class="n">name</span> <span class="ow">or</span> <span class="s1">&#39;mlp.proj.bias&#39;</span> <span class="ow">in</span> <span class="n">name</span><span class="p">:</span>
<span class="n">weights</span><span class="p">[</span><span class="n">name</span><span class="p">]</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">zeros_like</span><span class="p">(</span><span class="n">param</span><span class="p">)</span>
<span class="k">return</span> <span class="n">weights</span>
<span class="k">def</span><span class="w"> </span><span class="nf">get_kv_cache_type_from_legacy</span><span class="p">(</span><span class="n">use_cache</span><span class="p">:</span> <span class="nb">bool</span><span class="p">,</span>
<span class="n">paged_kv_cache</span><span class="p">:</span> <span class="nb">bool</span><span class="p">)</span> <span class="o">-&gt;</span> <span class="n">KVCacheType</span><span class="p">:</span>
<span class="k">if</span> <span class="n">use_cache</span><span class="p">:</span>
<span class="k">if</span> <span class="n">paged_kv_cache</span><span class="p">:</span>
<span class="k">return</span> <span class="n">KVCacheType</span><span class="o">.</span><span class="n">PAGED</span>
<span class="k">else</span><span class="p">:</span>
<span class="k">return</span> <span class="n">KVCacheType</span><span class="o">.</span><span class="n">CONTINUOUS</span>
<span class="k">else</span><span class="p">:</span>
<span class="k">return</span> <span class="n">KVCacheType</span><span class="o">.</span><span class="n">DISABLED</span>
<span class="k">def</span><span class="w"> </span><span class="nf">save_config</span><span class="p">(</span><span class="n">config</span><span class="p">:</span> <span class="n">PretrainedConfig</span><span class="p">,</span> <span class="o">*</span><span class="p">,</span> <span class="n">output_dir</span><span class="p">:</span> <span class="nb">str</span><span class="p">,</span>
<span class="n">log</span><span class="p">:</span> <span class="nb">bool</span><span class="p">)</span> <span class="o">-&gt;</span> <span class="kc">None</span><span class="p">:</span>
<span class="n">config_path</span> <span class="o">=</span> <span class="n">Path</span><span class="p">(</span><span class="n">output_dir</span><span class="p">)</span> <span class="o">/</span> <span class="s2">&quot;config.json&quot;</span>
<span class="k">if</span> <span class="n">log</span><span class="p">:</span>
<span class="n">logger</span><span class="o">.</span><span class="n">debug</span><span class="p">(</span><span class="sa">f</span><span class="s2">&quot;Saving TensorRT-LLM configuration to </span><span class="si">{</span><span class="n">config_path</span><span class="si">}</span><span class="s2">&quot;</span><span class="p">)</span>
<span class="n">config_path</span><span class="o">.</span><span class="n">parent</span><span class="o">.</span><span class="n">mkdir</span><span class="p">(</span><span class="n">exist_ok</span><span class="o">=</span><span class="kc">True</span><span class="p">,</span> <span class="n">parents</span><span class="o">=</span><span class="kc">True</span><span class="p">)</span>
<span class="n">config_path</span><span class="o">.</span><span class="n">write_text</span><span class="p">(</span><span class="n">json</span><span class="o">.</span><span class="n">dumps</span><span class="p">(</span><span class="n">config</span><span class="o">.</span><span class="n">to_dict</span><span class="p">(),</span> <span class="n">indent</span><span class="o">=</span><span class="mi">4</span><span class="p">))</span>
<span class="k">def</span><span class="w"> </span><span class="nf">save_checkpoint</span><span class="p">(</span><span class="o">*</span><span class="p">,</span> <span class="n">output_dir</span><span class="p">:</span> <span class="nb">str</span><span class="p">,</span> <span class="n">weights</span><span class="p">:</span> <span class="nb">dict</span><span class="p">,</span> <span class="n">rank</span><span class="p">:</span> <span class="nb">int</span><span class="p">)</span> <span class="o">-&gt;</span> <span class="kc">None</span><span class="p">:</span>
<span class="w"> </span><span class="sd">&quot;&quot;&quot; Checkpoint saver for weight loader.&quot;&quot;&quot;</span>
<span class="n">safetensors</span><span class="o">.</span><span class="n">torch</span><span class="o">.</span><span class="n">save_file</span><span class="p">(</span>
<span class="n">weights</span><span class="p">,</span> <span class="n">os</span><span class="o">.</span><span class="n">path</span><span class="o">.</span><span class="n">join</span><span class="p">(</span><span class="n">output_dir</span><span class="p">,</span> <span class="sa">f</span><span class="s1">&#39;rank</span><span class="si">{</span><span class="n">rank</span><span class="si">}</span><span class="s1">.safetensors&#39;</span><span class="p">))</span>
</pre></div>
</article>
<footer class="prev-next-footer d-print-none">
<div class="prev-next-area">
</div>
</footer>
</div>
<div class="bd-sidebar-secondary"></div>
</div>
<footer class="bd-footer-content">
</footer>
</main>
</div>
</div>
<!-- Scripts loaded after <body> so the DOM is not blocked -->
<script defer src="../../../_static/scripts/bootstrap.js?digest=8878045cc6db502f8baf"></script>
<script defer src="../../../_static/scripts/pydata-sphinx-theme.js?digest=8878045cc6db502f8baf"></script>
<footer class="bd-footer">
<div class="bd-footer__inner bd-page-width">
<div class="footer-items__start">
<div class="footer-item">
<a class="footer-brand logo" href="https://www.nvidia.com">
<img src="../../../_static/nvidia-logo-horiz-rgb-1c-blk-for-screen.svg" class="logo__image only-light" alt="NVIDIA"/>
<img src="../../../_static/nvidia-logo-horiz-rgb-1c-wht-for-screen.svg" class="logo__image only-dark" alt="NVIDIA"/>
</a></div>
<div class="footer-item">
<div class="footer-links">
<a class="external" href="https://www.nvidia.com/en-us/about-nvidia/privacy-policy/">Privacy Policy</a>
|
<a class="external" href="https://www.nvidia.com/en-us/about-nvidia/privacy-center/">Manage My Privacy</a>
|
<a class="external" href="https://www.nvidia.com/en-us/preferences/start/">Do Not Sell or Share My Data</a>
|
<a class="external" href="https://www.nvidia.com/en-us/about-nvidia/terms-of-service/">Terms of Service</a>
|
<a class="external" href="https://www.nvidia.com/en-us/about-nvidia/accessibility/">Accessibility</a>
|
<a class="external" href="https://www.nvidia.com/en-us/about-nvidia/company-policies/">Corporate Policies</a>
|
<a class="external" href="https://www.nvidia.com/en-us/product-security/">Product Security</a>
|
<a class="external" href="https://www.nvidia.com/en-us/contact/">Contact</a>
</div>
</div>
<div class="footer-item">
<p class="copyright">
Copyright © 2025, NVidia.
<br/>
</p>
</div>
<div class="footer-item">
<div class="extra_footer">
<p>Last updated on June 09, 2025.</p>
<p>This page is generated by TensorRT-LLM commit <a href="https://github.com/NVIDIA/TensorRT-LLM/tree/9c012d5">9c012d5</a>.</p>
</div></div>
</div>
</div>
</footer>
</body>
</html>
Reference in New Issue View Git Blame Copy Permalink