mirror of
https://github.com/NVIDIA/TensorRT-LLM.git
synced 2026-01-14 06:27:45 +08:00
2282 lines
277 KiB
HTML
2282 lines
277 KiB
HTML
|
|
|
|
<!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.llmapi.llm_utils — 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=a746c00c" />
|
|
<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" />
|
|
|
|
<!-- 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=888ff710"></script>
|
|
<script src="../../../_static/sphinx_highlight.js?v=dc90522c"></script>
|
|
<script src="../../../_static/clipboard.min.js?v=a7894cd8"></script>
|
|
<script src="../../../_static/copybutton.js?v=65e89d2a"></script>
|
|
<script>DOCUMENTATION_OPTIONS.pagename = '_modules/tensorrt_llm/llmapi/llm_utils';</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="" />
|
|
|
|
|
|
</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="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__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">LLM API Examples</span></p>
|
|
<ul class="nav bd-sidenav">
|
|
<li class="toctree-l1 has-children"><a class="reference internal" href="../../../llm-api-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="../../../llm-api-examples/llm_multilora.html">Generate text with multiple LoRA adapters</a></li>
|
|
<li class="toctree-l2"><a class="reference internal" href="../../../llm-api-examples/llm_quantization.html">Generation with Quantization</a></li>
|
|
<li class="toctree-l2"><a class="reference internal" href="../../../llm-api-examples/llm_inference_async.html">Generate Text Asynchronously</a></li>
|
|
<li class="toctree-l2"><a class="reference internal" href="../../../llm-api-examples/llm_guided_decoding.html">Generate text with guided decoding</a></li>
|
|
<li class="toctree-l2"><a class="reference internal" href="../../../llm-api-examples/llm_inference_customize.html">Generate text with customization</a></li>
|
|
<li class="toctree-l2"><a class="reference internal" href="../../../llm-api-examples/llm_logits_processor.html">Control generated text using logits post processor</a></li>
|
|
<li class="toctree-l2"><a class="reference internal" href="../../../llm-api-examples/llm_inference_async_streaming.html">Generate Text in Streaming</a></li>
|
|
<li class="toctree-l2"><a class="reference internal" href="../../../llm-api-examples/llm_inference_distributed.html">Distributed LLM Generation</a></li>
|
|
<li class="toctree-l2"><a class="reference internal" href="../../../llm-api-examples/llm_medusa_decoding.html">Generate Text Using Medusa Decoding</a></li>
|
|
<li class="toctree-l2"><a class="reference internal" href="../../../llm-api-examples/llm_lookahead_decoding.html">Generate Text Using Lookahead Decoding</a></li>
|
|
<li class="toctree-l2"><a class="reference internal" href="../../../llm-api-examples/llm_inference.html">Generate text</a></li>
|
|
<li class="toctree-l2"><a class="reference internal" href="../../../llm-api-examples/llm_auto_parallel.html">Automatic Parallelism with LLM</a></li>
|
|
</ul>
|
|
</details></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../llm-api-examples/customization.html">Common Customizations</a></li>
|
|
<li class="toctree-l1 has-children"><a class="reference internal" href="../../../llm-api-examples/llm_api_examples.html">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="../../../llm-api-examples/llm_multilora.html">Generate text with multiple LoRA adapters</a></li>
|
|
<li class="toctree-l2"><a class="reference internal" href="../../../llm-api-examples/llm_quantization.html">Generation with Quantization</a></li>
|
|
<li class="toctree-l2"><a class="reference internal" href="../../../llm-api-examples/llm_inference_async.html">Generate Text Asynchronously</a></li>
|
|
<li class="toctree-l2"><a class="reference internal" href="../../../llm-api-examples/llm_guided_decoding.html">Generate text with guided decoding</a></li>
|
|
<li class="toctree-l2"><a class="reference internal" href="../../../llm-api-examples/llm_inference_customize.html">Generate text with customization</a></li>
|
|
<li class="toctree-l2"><a class="reference internal" href="../../../llm-api-examples/llm_logits_processor.html">Control generated text using logits post processor</a></li>
|
|
<li class="toctree-l2"><a class="reference internal" href="../../../llm-api-examples/llm_inference_async_streaming.html">Generate Text in Streaming</a></li>
|
|
<li class="toctree-l2"><a class="reference internal" href="../../../llm-api-examples/llm_inference_distributed.html">Distributed LLM Generation</a></li>
|
|
<li class="toctree-l2"><a class="reference internal" href="../../../llm-api-examples/llm_medusa_decoding.html">Generate Text Using Medusa Decoding</a></li>
|
|
<li class="toctree-l2"><a class="reference internal" href="../../../llm-api-examples/llm_lookahead_decoding.html">Generate Text Using Lookahead Decoding</a></li>
|
|
<li class="toctree-l2"><a class="reference internal" href="../../../llm-api-examples/llm_inference.html">Generate text</a></li>
|
|
<li class="toctree-l2"><a class="reference internal" href="../../../llm-api-examples/llm_auto_parallel.html">Automatic Parallelism with LLM</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/inference-request.html">Inference Request</a></li>
|
|
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../advanced/lora.html">Run gpt-2b + LoRA using GptManager / cpp runtime</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../advanced/expert-parallelism.html">Expert Parallelism in TensorRT-LLM</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../advanced/kv-cache-reuse.html">KV cache reuse</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../advanced/speculative-decoding.html">Speculative Sampling</a></li>
|
|
<li class="toctree-l1"><a class="reference internal" href="../../../advanced/disaggregated-service.html">Disaggregated-Service (experimental)</a></li>
|
|
</ul>
|
|
<p 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>
|
|
</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.llmapi.llm_utils</span></li>
|
|
</ul>
|
|
</nav>
|
|
</div>
|
|
|
|
</div>
|
|
|
|
|
|
</div>
|
|
</div>
|
|
|
|
|
|
|
|
|
|
<div id="searchbox"></div>
|
|
<article class="bd-article">
|
|
|
|
<h1>Source code for tensorrt_llm.llmapi.llm_utils</h1><div class="highlight"><pre>
|
|
<span></span><span class="n">__all__</span> <span class="o">=</span> <span class="p">[</span>
|
|
<span class="s1">'LlmArgs'</span><span class="p">,</span>
|
|
<span class="s1">'LlmBuildStats'</span><span class="p">,</span>
|
|
<span class="s1">'ModelLoader'</span><span class="p">,</span>
|
|
<span class="s1">'_ModelRuntimeContext'</span><span class="p">,</span>
|
|
<span class="s1">'_ModelInfo'</span><span class="p">,</span>
|
|
<span class="s1">'_ParallelConfig'</span><span class="p">,</span>
|
|
<span class="s1">'_ModelFormatKind'</span><span class="p">,</span>
|
|
<span class="s1">'_ModelWrapper'</span><span class="p">,</span>
|
|
<span class="s1">'BatchingType'</span><span class="p">,</span>
|
|
<span class="s1">'ExecutorConfig'</span><span class="p">,</span>
|
|
<span class="s1">'SchedulerConfig'</span><span class="p">,</span>
|
|
<span class="s1">'KvCacheConfig'</span><span class="p">,</span>
|
|
<span class="s1">'LookaheadDecodingConfig'</span><span class="p">,</span>
|
|
<span class="s1">'MedusaDecodingConfig'</span><span class="p">,</span>
|
|
<span class="s1">'ContextChunkingPolicy'</span><span class="p">,</span>
|
|
<span class="s1">'CapacitySchedulerPolicy'</span><span class="p">,</span>
|
|
<span class="s1">'BuildConfig'</span><span class="p">,</span>
|
|
<span class="s1">'BuildCacheConfig'</span><span class="p">,</span>
|
|
<span class="s1">'QuantConfig'</span><span class="p">,</span>
|
|
<span class="s1">'CalibConfig'</span><span class="p">,</span>
|
|
<span class="s1">'CachedModelLoader'</span><span class="p">,</span>
|
|
<span class="s1">'ConfigArbitrateError'</span><span class="p">,</span>
|
|
<span class="s1">'_ConfigArbitrator'</span><span class="p">,</span>
|
|
<span class="p">]</span>
|
|
|
|
<span class="kn">import</span> <span class="nn">copy</span>
|
|
<span class="kn">import</span> <span class="nn">json</span>
|
|
<span class="kn">import</span> <span class="nn">os</span>
|
|
<span class="kn">import</span> <span class="nn">shutil</span>
|
|
<span class="kn">import</span> <span class="nn">tempfile</span>
|
|
<span class="kn">import</span> <span class="nn">time</span>
|
|
<span class="kn">import</span> <span class="nn">weakref</span>
|
|
<span class="kn">from</span> <span class="nn">argparse</span> <span class="kn">import</span> <span class="n">Namespace</span>
|
|
<span class="kn">from</span> <span class="nn">dataclasses</span> <span class="kn">import</span> <span class="n">asdict</span><span class="p">,</span> <span class="n">dataclass</span><span class="p">,</span> <span class="n">field</span><span class="p">,</span> <span class="n">fields</span>
|
|
<span class="kn">from</span> <span class="nn">enum</span> <span class="kn">import</span> <span class="n">Enum</span>
|
|
<span class="kn">from</span> <span class="nn">pathlib</span> <span class="kn">import</span> <span class="n">Path</span>
|
|
<span class="kn">from</span> <span class="nn">typing</span> <span class="kn">import</span> <span class="n">Any</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">List</span><span class="p">,</span> <span class="n">Literal</span><span class="p">,</span> <span class="n">Optional</span><span class="p">,</span> <span class="n">Tuple</span><span class="p">,</span> <span class="n">Union</span>
|
|
|
|
<span class="kn">import</span> <span class="nn">torch</span>
|
|
<span class="kn">from</span> <span class="nn">tqdm</span> <span class="kn">import</span> <span class="n">tqdm</span>
|
|
<span class="kn">from</span> <span class="nn">transformers</span> <span class="kn">import</span> <span class="n">PreTrainedTokenizerBase</span>
|
|
|
|
<span class="kn">from</span> <span class="nn">.._utils</span> <span class="kn">import</span> <span class="n">mpi_barrier</span><span class="p">,</span> <span class="n">mpi_broadcast</span><span class="p">,</span> <span class="n">mpi_rank</span><span class="p">,</span> <span class="n">release_gc</span>
|
|
<span class="kn">from</span> <span class="nn">..auto_parallel</span> <span class="kn">import</span> <span class="n">AutoParallelConfig</span><span class="p">,</span> <span class="n">infer_cluster_config</span>
|
|
<span class="kn">from</span> <span class="nn">..bindings.executor</span> <span class="kn">import</span> <span class="p">(</span><span class="n">BatchingType</span><span class="p">,</span> <span class="n">CapacitySchedulerPolicy</span><span class="p">,</span>
|
|
<span class="n">ContextChunkingPolicy</span><span class="p">,</span> <span class="n">DecodingConfig</span><span class="p">,</span>
|
|
<span class="n">DecodingMode</span><span class="p">,</span> <span class="n">ExecutorConfig</span><span class="p">,</span>
|
|
<span class="n">ExtendedRuntimePerfKnobConfig</span><span class="p">,</span> <span class="n">KvCacheConfig</span><span class="p">,</span>
|
|
<span class="n">LookaheadDecodingConfig</span><span class="p">,</span> <span class="n">PeftCacheConfig</span><span class="p">,</span>
|
|
<span class="n">SchedulerConfig</span><span class="p">)</span>
|
|
<span class="kn">from</span> <span class="nn">..builder</span> <span class="kn">import</span> <span class="n">BuildConfig</span><span class="p">,</span> <span class="n">Engine</span><span class="p">,</span> <span class="n">EngineConfig</span><span class="p">,</span> <span class="n">build</span>
|
|
<span class="kn">from</span> <span class="nn">..logger</span> <span class="kn">import</span> <span class="n">logger</span>
|
|
<span class="kn">from</span> <span class="nn">..mapping</span> <span class="kn">import</span> <span class="n">Mapping</span>
|
|
<span class="kn">from</span> <span class="nn">..models.automodel</span> <span class="kn">import</span> <span class="n">MODEL_MAP</span><span class="p">,</span> <span class="n">AutoConfig</span><span class="p">,</span> <span class="n">AutoModelForCausalLM</span>
|
|
<span class="kn">from</span> <span class="nn">..models.modeling_utils</span> <span class="kn">import</span> <span class="p">(</span><span class="n">PretrainedConfig</span><span class="p">,</span> <span class="n">QuantAlgo</span><span class="p">,</span> <span class="n">QuantConfig</span><span class="p">,</span>
|
|
<span class="n">SpeculativeDecodingMode</span><span class="p">)</span>
|
|
<span class="kn">from</span> <span class="nn">..module</span> <span class="kn">import</span> <span class="n">Module</span>
|
|
<span class="kn">from</span> <span class="nn">.build_cache</span> <span class="kn">import</span> <span class="p">(</span><span class="n">BuildCache</span><span class="p">,</span> <span class="n">BuildCacheConfig</span><span class="p">,</span> <span class="n">CachedStage</span><span class="p">,</span>
|
|
<span class="n">get_build_cache_config_from_env</span><span class="p">)</span>
|
|
<span class="kn">from</span> <span class="nn">.mpi_session</span> <span class="kn">import</span> <span class="n">MPINodeState</span><span class="p">,</span> <span class="n">MpiSession</span>
|
|
<span class="kn">from</span> <span class="nn">.tokenizer</span> <span class="kn">import</span> <span class="n">TokenizerBase</span><span class="p">,</span> <span class="n">TransformersTokenizer</span><span class="p">,</span> <span class="n">tokenizer_factory</span>
|
|
<span class="c1"># TODO[chunweiy]: move the following symbols back to utils scope, and remove the following import</span>
|
|
<span class="kn">from</span> <span class="nn">.utils</span> <span class="kn">import</span> <span class="p">(</span><span class="n">GpuArch</span><span class="p">,</span> <span class="n">append_docstring</span><span class="p">,</span> <span class="n">download_hf_model</span><span class="p">,</span>
|
|
<span class="n">download_hf_pretrained_config</span><span class="p">,</span> <span class="n">enable_llm_debug</span><span class="p">,</span>
|
|
<span class="n">get_directory_size_in_gb</span><span class="p">,</span> <span class="n">print_colored</span><span class="p">,</span>
|
|
<span class="n">print_traceback_on_error</span><span class="p">)</span>
|
|
|
|
|
|
<span class="nd">@dataclass</span>
|
|
<span class="k">class</span> <span class="nc">_ParallelConfig</span><span class="p">:</span>
|
|
<span class="w"> </span><span class="sd">''' The model distribution configs for LLM. '''</span>
|
|
<span class="n">tp_size</span><span class="p">:</span> <span class="nb">int</span> <span class="o">=</span> <span class="mi">1</span>
|
|
<span class="n">pp_size</span><span class="p">:</span> <span class="nb">int</span> <span class="o">=</span> <span class="mi">1</span>
|
|
<span class="n">cp_size</span><span class="p">:</span> <span class="nb">int</span> <span class="o">=</span> <span class="mi">1</span>
|
|
<span class="n">moe_tp_size</span><span class="p">:</span> <span class="nb">int</span> <span class="o">=</span> <span class="mi">1</span>
|
|
<span class="n">moe_ep_size</span><span class="p">:</span> <span class="nb">int</span> <span class="o">=</span> <span class="mi">1</span>
|
|
<span class="n">cp_config</span><span class="p">:</span> <span class="nb">dict</span> <span class="o">=</span> <span class="n">field</span><span class="p">(</span><span class="n">default_factory</span><span class="o">=</span><span class="nb">dict</span><span class="p">)</span>
|
|
<span class="n">auto_parallel</span><span class="p">:</span> <span class="nb">bool</span> <span class="o">=</span> <span class="kc">False</span>
|
|
|
|
<span class="n">_world_size</span><span class="p">:</span> <span class="nb">int</span> <span class="o">=</span> <span class="n">field</span><span class="p">(</span><span class="n">default</span><span class="o">=</span><span class="mi">1</span><span class="p">,</span> <span class="n">init</span><span class="o">=</span><span class="kc">False</span><span class="p">)</span>
|
|
<span class="n">_devices</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">int</span><span class="p">]]</span> <span class="o">=</span> <span class="n">field</span><span class="p">(</span><span class="n">default</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span> <span class="n">init</span><span class="o">=</span><span class="kc">False</span><span class="p">)</span>
|
|
|
|
<span class="nd">@property</span>
|
|
<span class="k">def</span> <span class="nf">devices</span><span class="p">(</span><span class="bp">self</span><span class="p">)</span> <span class="o">-></span> <span class="n">List</span><span class="p">[</span><span class="nb">int</span><span class="p">]:</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">_devices</span> <span class="ow">is</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="k">return</span> <span class="nb">list</span><span class="p">(</span><span class="nb">range</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">world_size</span><span class="p">))</span>
|
|
<span class="k">return</span> <span class="bp">self</span><span class="o">.</span><span class="n">_devices</span>
|
|
|
|
<span class="nd">@devices</span><span class="o">.</span><span class="n">setter</span>
|
|
<span class="k">def</span> <span class="nf">devices</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">devices</span><span class="p">:</span> <span class="n">List</span><span class="p">[</span><span class="nb">int</span><span class="p">]):</span>
|
|
<span class="k">if</span> <span class="nb">len</span><span class="p">(</span><span class="n">devices</span><span class="p">)</span> <span class="o">!=</span> <span class="bp">self</span><span class="o">.</span><span class="n">world_size</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">"devices </span><span class="si">{</span><span class="n">devices</span><span class="si">}</span><span class="s2"> should have the same length as world_size </span><span class="si">{</span><span class="bp">self</span><span class="o">.</span><span class="n">world_size</span><span class="si">}</span><span class="s2">"</span>
|
|
<span class="p">)</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">_devices</span> <span class="o">=</span> <span class="n">devices</span>
|
|
|
|
<span class="nd">@property</span>
|
|
<span class="k">def</span> <span class="nf">world_size</span><span class="p">(</span><span class="bp">self</span><span class="p">)</span> <span class="o">-></span> <span class="nb">bool</span><span class="p">:</span>
|
|
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">auto_parallel</span><span class="p">:</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">tp_size</span> <span class="o">></span> <span class="mi">1</span> <span class="ow">or</span> <span class="bp">self</span><span class="o">.</span><span class="n">pp_size</span> <span class="o">></span> <span class="mi">1</span> <span class="ow">or</span> <span class="bp">self</span><span class="o">.</span><span class="n">cp_size</span> <span class="o">></span> <span class="mi">1</span><span class="p">:</span>
|
|
<span class="k">raise</span> <span class="ne">RuntimeError</span><span class="p">(</span>
|
|
<span class="s2">"manually TP and PP are not supported in auto parallel mode."</span>
|
|
<span class="p">)</span>
|
|
<span class="k">return</span> <span class="bp">self</span><span class="o">.</span><span class="n">_world_size</span>
|
|
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">_world_size</span> <span class="o">></span> <span class="mi">1</span><span class="p">:</span>
|
|
<span class="k">raise</span> <span class="ne">RuntimeError</span><span class="p">(</span>
|
|
<span class="s2">"world_size > 1 is only supported in auto parallel mode."</span><span class="p">)</span>
|
|
<span class="k">return</span> <span class="bp">self</span><span class="o">.</span><span class="n">tp_size</span> <span class="o">*</span> <span class="bp">self</span><span class="o">.</span><span class="n">pp_size</span> <span class="o">*</span> <span class="bp">self</span><span class="o">.</span><span class="n">cp_size</span>
|
|
|
|
<span class="nd">@world_size</span><span class="o">.</span><span class="n">setter</span>
|
|
<span class="k">def</span> <span class="nf">world_size</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">world_size</span><span class="p">:</span> <span class="nb">int</span><span class="p">):</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">auto_parallel</span><span class="p">:</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">_world_size</span> <span class="o">=</span> <span class="n">world_size</span>
|
|
<span class="k">elif</span> <span class="p">(</span><span class="ow">not</span> <span class="bp">self</span><span class="o">.</span><span class="n">auto_parallel</span>
|
|
<span class="p">)</span> <span class="ow">and</span> <span class="n">world_size</span> <span class="o">!=</span> <span class="bp">self</span><span class="o">.</span><span class="n">tp_size</span> <span class="o">*</span> <span class="bp">self</span><span class="o">.</span><span class="n">pp_size</span> <span class="o">*</span> <span class="bp">self</span><span class="o">.</span><span class="n">cp_size</span><span class="p">:</span>
|
|
<span class="k">raise</span> <span class="ne">ValueError</span><span class="p">(</span>
|
|
<span class="sa">f</span><span class="s2">"world_size </span><span class="si">{</span><span class="n">world_size</span><span class="si">}</span><span class="s2"> should be equal to tp_size * pp_size </span><span class="si">{</span><span class="bp">self</span><span class="o">.</span><span class="n">tp_size</span><span class="w"> </span><span class="o">*</span><span class="w"> </span><span class="bp">self</span><span class="o">.</span><span class="n">pp_size</span><span class="w"> </span><span class="o">*</span><span class="w"> </span><span class="bp">self</span><span class="o">.</span><span class="n">cp_size</span><span class="si">}</span><span class="s2"> "</span>
|
|
<span class="p">)</span>
|
|
|
|
<span class="nd">@property</span>
|
|
<span class="k">def</span> <span class="nf">is_multi_gpu</span><span class="p">(</span><span class="bp">self</span><span class="p">)</span> <span class="o">-></span> <span class="nb">bool</span><span class="p">:</span>
|
|
<span class="k">return</span> <span class="bp">self</span><span class="o">.</span><span class="n">world_size</span> <span class="o">></span> <span class="mi">1</span>
|
|
|
|
<span class="k">def</span> <span class="nf">to_mapping</span><span class="p">(</span><span class="bp">self</span><span class="p">)</span> <span class="o">-></span> <span class="n">Mapping</span><span class="p">:</span>
|
|
<span class="k">return</span> <span class="n">Mapping</span><span class="p">(</span><span class="n">world_size</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">world_size</span><span class="p">,</span>
|
|
<span class="n">rank</span><span class="o">=</span><span class="n">mpi_rank</span><span class="p">(),</span>
|
|
<span class="n">tp_size</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">tp_size</span><span class="p">,</span>
|
|
<span class="n">pp_size</span><span class="o">=</span><span class="bp">self</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="bp">self</span><span class="o">.</span><span class="n">cp_size</span><span class="p">,</span>
|
|
<span class="n">cp_config</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">cp_config</span><span class="p">,</span>
|
|
<span class="n">moe_tp_size</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">moe_tp_size</span><span class="p">,</span>
|
|
<span class="n">moe_ep_size</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">moe_ep_size</span><span class="p">,</span>
|
|
<span class="n">auto_parallel</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">auto_parallel</span><span class="p">)</span>
|
|
|
|
|
|
<div class="viewcode-block" id="CalibConfig">
|
|
<a class="viewcode-back" href="../../../llm-api/reference.html#tensorrt_llm.llmapi.CalibConfig">[docs]</a>
|
|
<span class="nd">@dataclass</span><span class="p">(</span><span class="n">slots</span><span class="o">=</span><span class="kc">True</span><span class="p">)</span>
|
|
<span class="k">class</span> <span class="nc">CalibConfig</span><span class="p">:</span>
|
|
<span class="w"> </span><span class="sd">"""</span>
|
|
<span class="sd"> Calibration configuration.</span>
|
|
|
|
<span class="sd"> Args:</span>
|
|
<span class="sd"> device (Literal['cuda', 'cpu'], default='cuda'): The device to run calibration.</span>
|
|
<span class="sd"> calib_dataset (str, default='cnn_dailymail'): The name or local path of calibration dataset.</span>
|
|
<span class="sd"> calib_batches (int, default=512): The number of batches that the calibration runs.</span>
|
|
<span class="sd"> calib_batch_size (int, default=1): The batch size that the calibration runs.</span>
|
|
<span class="sd"> calib_max_seq_length (int, default=512): The maximum sequence length that the calibration runs.</span>
|
|
<span class="sd"> random_seed (int, default=1234): The random seed used for calibration.</span>
|
|
<span class="sd"> tokenizer_max_seq_length (int, default=2048): The maximum sequence length to initialize tokenizer for calibration.</span>
|
|
<span class="sd"> """</span>
|
|
<span class="n">device</span><span class="p">:</span> <span class="n">Literal</span><span class="p">[</span><span class="s1">'cuda'</span><span class="p">,</span> <span class="s1">'cpu'</span><span class="p">]</span> <span class="o">=</span> <span class="s1">'cuda'</span>
|
|
<span class="n">calib_dataset</span><span class="p">:</span> <span class="nb">str</span> <span class="o">=</span> <span class="s1">'cnn_dailymail'</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="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="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="n">random_seed</span><span class="p">:</span> <span class="nb">int</span> <span class="o">=</span> <span class="mi">1234</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>
|
|
|
|
<div class="viewcode-block" id="CalibConfig.from_dict">
|
|
<a class="viewcode-back" href="../../../llm-api/reference.html#tensorrt_llm.llmapi.CalibConfig.from_dict">[docs]</a>
|
|
<span class="nd">@classmethod</span>
|
|
<span class="k">def</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="k">return</span> <span class="bp">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="CalibConfig.to_dict">
|
|
<a class="viewcode-back" href="../../../llm-api/reference.html#tensorrt_llm.llmapi.CalibConfig.to_dict">[docs]</a>
|
|
<span class="k">def</span> <span class="nf">to_dict</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
|
|
<span class="k">return</span> <span class="n">asdict</span><span class="p">(</span><span class="bp">self</span><span class="p">)</span></div>
|
|
</div>
|
|
|
|
|
|
|
|
<span class="k">class</span> <span class="nc">_ModelFormatKind</span><span class="p">(</span><span class="n">Enum</span><span class="p">):</span>
|
|
<span class="n">HF</span> <span class="o">=</span> <span class="mi">0</span>
|
|
<span class="n">TLLM_CKPT</span> <span class="o">=</span> <span class="mi">1</span>
|
|
<span class="n">TLLM_ENGINE</span> <span class="o">=</span> <span class="mi">2</span>
|
|
|
|
|
|
<span class="nd">@dataclass</span>
|
|
<span class="k">class</span> <span class="nc">_ModelInfo</span><span class="p">:</span>
|
|
<span class="n">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="n">architecture</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="nd">@property</span>
|
|
<span class="k">def</span> <span class="nf">model_name</span><span class="p">(</span><span class="bp">self</span><span class="p">)</span> <span class="o">-></span> <span class="nb">str</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="ow">is</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="k">raise</span> <span class="ne">RuntimeError</span><span class="p">(</span><span class="s2">"The architecture is not set yet."</span><span class="p">)</span>
|
|
|
|
<span class="k">return</span> <span class="bp">self</span><span class="o">.</span><span class="n">architecture</span>
|
|
|
|
<span class="nd">@classmethod</span>
|
|
<span class="k">def</span> <span class="nf">from_pretrained_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">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">architecture</span><span class="o">=</span><span class="n">config</span><span class="o">.</span><span class="n">architecture</span><span class="p">)</span>
|
|
|
|
<span class="nd">@classmethod</span>
|
|
<span class="k">def</span> <span class="nf">from_builder_config_json</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="k">if</span> <span class="s1">'version'</span> <span class="ow">in</span> <span class="n">config</span><span class="p">:</span>
|
|
<span class="c1"># The Dict format is { 'builder_config':..., 'plugin_config':...}</span>
|
|
<span class="n">dtype</span> <span class="o">=</span> <span class="n">config</span><span class="p">[</span><span class="s1">'plugin_config'</span><span class="p">][</span><span class="s1">'gpt_attention_plugin'</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">config</span><span class="p">[</span><span class="s1">'pretrained_config'</span><span class="p">][</span><span class="s1">'dtype'</span><span class="p">]</span>
|
|
|
|
<span class="k">return</span> <span class="bp">cls</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">architecture</span><span class="o">=</span><span class="n">config</span><span class="p">[</span><span class="s1">'builder_config'</span><span class="p">][</span><span class="s1">'name'</span><span class="p">])</span>
|
|
|
|
<span class="nd">@classmethod</span>
|
|
<span class="k">def</span> <span class="nf">from_module</span><span class="p">(</span><span class="bp">cls</span><span class="p">,</span> <span class="n">module</span><span class="p">:</span> <span class="n">Module</span><span class="p">):</span>
|
|
<span class="k">raise</span> <span class="ne">NotImplementedError</span><span class="p">()</span>
|
|
|
|
|
|
<div class="viewcode-block" id="MedusaDecodingConfig">
|
|
<a class="viewcode-back" href="../../../llm-api/reference.html#tensorrt_llm.llmapi.MedusaDecodingConfig">[docs]</a>
|
|
<span class="nd">@dataclass</span>
|
|
<span class="k">class</span> <span class="nc">MedusaDecodingConfig</span><span class="p">:</span>
|
|
<span class="n">medusa_choices</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="n">List</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="n">num_medusa_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></div>
|
|
|
|
|
|
|
|
<span class="nd">@dataclass</span>
|
|
<span class="k">class</span> <span class="nc">_ModelWrapper</span><span class="p">:</span>
|
|
<span class="n">model</span><span class="p">:</span> <span class="n">Union</span><span class="p">[</span><span class="nb">str</span><span class="p">,</span> <span class="n">Path</span><span class="p">]</span>
|
|
|
|
<span class="k">def</span> <span class="nf">__post_init__</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
|
|
<span class="k">if</span> <span class="ow">not</span> <span class="bp">self</span><span class="o">.</span><span class="n">model</span><span class="p">:</span>
|
|
<span class="k">raise</span> <span class="ne">ValueError</span><span class="p">(</span><span class="s2">"model should be provided."</span><span class="p">)</span>
|
|
<span class="k">assert</span> <span class="nb">isinstance</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">model</span><span class="p">,</span>
|
|
<span class="p">(</span><span class="nb">str</span><span class="p">,</span> <span class="n">Path</span><span class="p">)),</span> <span class="sa">f</span><span class="s2">"Invalid model: </span><span class="si">{</span><span class="bp">self</span><span class="o">.</span><span class="n">model</span><span class="si">}</span><span class="s2">"</span>
|
|
|
|
<span class="n">model_dir</span> <span class="o">=</span> <span class="n">Path</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">model</span><span class="p">)</span>
|
|
|
|
<span class="k">if</span> <span class="n">model_dir</span><span class="o">.</span><span class="n">exists</span><span class="p">()</span> <span class="ow">and</span> <span class="n">model_dir</span><span class="o">.</span><span class="n">is_dir</span><span class="p">():</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">model</span> <span class="o">=</span> <span class="n">model_dir</span>
|
|
|
|
<span class="nd">@property</span>
|
|
<span class="k">def</span> <span class="nf">is_hub_model</span><span class="p">(</span><span class="bp">self</span><span class="p">)</span> <span class="o">-></span> <span class="nb">bool</span><span class="p">:</span>
|
|
<span class="k">return</span> <span class="ow">not</span> <span class="bp">self</span><span class="o">.</span><span class="n">is_local_model</span>
|
|
|
|
<span class="nd">@property</span>
|
|
<span class="k">def</span> <span class="nf">is_local_model</span><span class="p">(</span><span class="bp">self</span><span class="p">)</span> <span class="o">-></span> <span class="nb">bool</span><span class="p">:</span>
|
|
<span class="k">return</span> <span class="nb">isinstance</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">model</span><span class="p">,</span> <span class="n">Path</span><span class="p">)</span>
|
|
|
|
<span class="nd">@property</span>
|
|
<span class="k">def</span> <span class="nf">model_dir</span><span class="p">(</span><span class="bp">self</span><span class="p">)</span> <span class="o">-></span> <span class="n">Path</span><span class="p">:</span>
|
|
<span class="k">assert</span> <span class="bp">self</span><span class="o">.</span><span class="n">is_local_model</span><span class="p">,</span> <span class="sa">f</span><span class="s2">"model_dir is only available for local model, </span><span class="si">{</span><span class="bp">self</span><span class="o">.</span><span class="n">model</span><span class="si">}</span><span class="s2">."</span>
|
|
<span class="k">return</span> <span class="bp">self</span><span class="o">.</span><span class="n">model</span>
|
|
|
|
<span class="nd">@model_dir</span><span class="o">.</span><span class="n">setter</span>
|
|
<span class="k">def</span> <span class="nf">model_dir</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">model_dir</span><span class="p">:</span> <span class="n">Union</span><span class="p">[</span><span class="nb">str</span><span class="p">,</span> <span class="n">Path</span><span class="p">]):</span>
|
|
<span class="n">model_dir</span> <span class="o">=</span> <span class="n">Path</span><span class="p">(</span><span class="n">model_dir</span><span class="p">)</span>
|
|
<span class="k">assert</span> <span class="n">model_dir</span><span class="o">.</span><span class="n">exists</span><span class="p">()</span> <span class="ow">and</span> <span class="n">model_dir</span><span class="o">.</span><span class="n">is_dir</span><span class="p">(</span>
|
|
<span class="p">),</span> <span class="sa">f</span><span class="s2">"model_dir is not a valid path, </span><span class="si">{</span><span class="n">model_dir</span><span class="si">}</span><span class="s2">"</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">model</span> <span class="o">=</span> <span class="n">model_dir</span>
|
|
|
|
<span class="nd">@property</span>
|
|
<span class="k">def</span> <span class="nf">model_name</span><span class="p">(</span><span class="bp">self</span><span class="p">)</span> <span class="o">-></span> <span class="n">Union</span><span class="p">[</span><span class="nb">str</span><span class="p">,</span> <span class="n">Path</span><span class="p">]:</span>
|
|
<span class="k">return</span> <span class="bp">self</span><span class="o">.</span><span class="n">model</span> <span class="k">if</span> <span class="nb">isinstance</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">model</span><span class="p">,</span> <span class="nb">str</span><span class="p">)</span> <span class="k">else</span> <span class="kc">None</span>
|
|
|
|
|
|
<span class="c1"># The docstring for LlmArgs and LLM; will be appended to the two classes' apidocs.</span>
|
|
<span class="n">LLMARGS_DOCSTRING</span> <span class="o">=</span> <span class="sa">r</span><span class="s2">"""</span>
|
|
<span class="s2"> model (str or Path): The model name or a local model directory.</span>
|
|
<span class="s2"> Note that if the value could be both a model name or a local model directory,</span>
|
|
<span class="s2"> the local model directory will be prioritized.</span>
|
|
|
|
<span class="s2"> tokenizer (str, Path, TokenizerBase, PreTrainedTokenizerBase, optional):</span>
|
|
<span class="s2"> The name or path of a HuggingFace Transformers tokenizer, or the loaded tokenizer.</span>
|
|
<span class="s2"> Defaults to None.</span>
|
|
|
|
<span class="s2"> tokenizer_mode (Literal['auto', 'slow']): The tokenizer mode.</span>
|
|
<span class="s2"> 'auto' will use the fast tokenizer if available, and 'slow' will always use the slow tokenizer.</span>
|
|
<span class="s2"> The fast tokenizer is based on Huggingface's Rust library tokenizers, which achieves a significant speed-up compared to its slow counterpart.</span>
|
|
<span class="s2"> Defaults to 'auto'.</span>
|
|
|
|
<span class="s2"> skip_tokenizer_init (bool):</span>
|
|
<span class="s2"> If true, skip initialization of tokenizer and detokenizer.</span>
|
|
<span class="s2"> LLM.generate and LLM.generate_async will accept prompt token ids as input only.</span>
|
|
<span class="s2"> Defaults to False.</span>
|
|
|
|
<span class="s2"> trust_remote_code (bool): Whether to trust remote code when downloading model and tokenizer from Hugging Face. Defaults to False.</span>
|
|
|
|
<span class="s2"> tensor_parallel_size(int): The number of processes for tensor parallelism. Defaults to 1.</span>
|
|
|
|
<span class="s2"> dtype (str): The data type for the model weights and activations.</span>
|
|
<span class="s2"> Can be "float16", "bfloat16", "float32", or "auto". If "auto", the data type</span>
|
|
<span class="s2"> will be automatically inferred from the source model. If the source data type</span>
|
|
<span class="s2"> is "float32", it will be converted to "float16". Defaults to "auto".</span>
|
|
|
|
<span class="s2"> revision (str, optional): The revision of the model to use. Defaults to None.</span>
|
|
|
|
<span class="s2"> tokenizer_revision (str, optional): The revision of the tokenizer to use. Defaults to None.</span>
|
|
|
|
<span class="s2"> pipeline_parallel_size (int): The pipeline parallel size. Defaults to 1.</span>
|
|
|
|
<span class="s2"> context_parallel_size (int): The context parallel size. Defaults to 1.</span>
|
|
|
|
<span class="s2"> load_format (Literal['auto', 'dummy']): The format of the model weights to load.</span>
|
|
<span class="s2"> * 'auto' will try to load the weights from the provided checkpoint.</span>
|
|
<span class="s2"> * 'dummy' will initialize the weights with random values, which is mainly for profiling.</span>
|
|
<span class="s2"> Defaults to 'auto'.</span>
|
|
|
|
<span class="s2"> enable_tqdm (bool): Whether to display a progress bar during model building. Defaults to False.</span>
|
|
|
|
<span class="s2"> enable_lora (bool): Enable LoRA adapters. Defaults to False.</span>
|
|
|
|
<span class="s2"> max_lora_rank (int, optional): Maximum LoRA rank. If specified, it overrides `build_config.lora_config.max_lora_rank`. Defaults to None.</span>
|
|
|
|
<span class="s2"> max_loras (int): Maximum number of LoRA adapters to be stored in GPU memory. Defaults to 4.</span>
|
|
|
|
<span class="s2"> max_cpu_loras (int): Maximum number of LoRA adapters to be stored in CPU memory. Defaults to 4.</span>
|
|
|
|
<span class="s2"> enable_prompt_adapter (bool): Enable prompt adapters. Defaults to False.</span>
|
|
|
|
<span class="s2"> max_prompt_adapter_token (int): Maximum number of prompt adapter tokens. Defaults to 0.</span>
|
|
|
|
<span class="s2"> quant_config (QuantConfig, optional): The quantization configuration for the model. Defaults to None.</span>
|
|
|
|
<span class="s2"> calib_config (CalibConfig, optional): The calibration configuration for the model. Defaults to None.</span>
|
|
|
|
<span class="s2"> build_config (BuildConfig, optional)): The build configuration for the model. Defaults to None.</span>
|
|
|
|
<span class="s2"> kv_cache_config (KvCacheConfig, optional): The key-value cache configuration for the model. Defaults to None.</span>
|
|
|
|
<span class="s2"> enable_chunked_prefill (bool): Whether to enable chunked prefill. Defaults to False.</span>
|
|
|
|
<span class="s2"> decoding_config (DecodingConfig, optional): The decoding configuration for the model. Defaults to None.</span>
|
|
|
|
<span class="s2"> guided_decoding_backend (str, optional): The guided decoding backend, currently supports 'xgrammar'. Defaults to None.</span>
|
|
|
|
<span class="s2"> logits_post_processor_map (Dict[str, Callable], optional): A map of logit post-processing functions. Defaults to None.</span>
|
|
|
|
<span class="s2"> iter_stats_max_iterations (int, optional): The maximum number of iterations for iteration statistics. Defaults to None.</span>
|
|
|
|
<span class="s2"> request_stats_max_iterations (int, optional): The maximum number of iterations for request statistics. Defaults to None.</span>
|
|
|
|
<span class="s2"> workspace(str, optional): The directory to store intermediate files. Defaults to None.</span>
|
|
|
|
<span class="s2"> embedding_parallel_mode (str): The parallel mode for embeddings. Defaults to 'SHARDING_ALONG_VOCAB'.</span>
|
|
|
|
<span class="s2"> auto_parallel (bool): Enable auto parallel mode. Defaults to False.</span>
|
|
|
|
<span class="s2"> auto_parallel_world_size (int): The MPI world size for auto parallel. Defaults to 1.</span>
|
|
|
|
<span class="s2"> moe_tensor_parallel_size (int, optional): The tensor parallel size for MoE models's expert weights.</span>
|
|
|
|
<span class="s2"> moe_expert_parallel_size (int, optional): The expert parallel size for MoE models's expert weights.</span>
|
|
|
|
<span class="s2"> fast_build: (bool): Enable features for faster engine building.</span>
|
|
<span class="s2"> This may cause some performance degradation and is currently incompatible with int8/int4 quantization.</span>
|
|
<span class="s2"> Defaults to False.</span>
|
|
|
|
<span class="s2"> enable_build_cache (bool, BuildCacheConfig, optional): Whether to enable build caching for the model. Defaults to None.</span>
|
|
|
|
<span class="s2"> peft_cache_config (PeftCacheConfig, optional): The PEFT cache configuration for the model. Defaults to None.</span>
|
|
|
|
<span class="s2"> scheduler_config (SchedulerConfig, optional): The scheduler configuration for the model. Defaults to None.</span>
|
|
|
|
<span class="s2"> speculative_config (LookaheadDecodingConfig or other speculative configurations, optional): The speculative decoding configuration. Defaults to None.</span>
|
|
|
|
<span class="s2"> batching_type (BatchingType, optional): The batching type for the model. Defaults to None.</span>
|
|
|
|
<span class="s2"> normalize_log_probs (bool): Whether to normalize log probabilities for the model. Defaults to False.</span>
|
|
|
|
|
|
<span class="s2"> max_batch_size (int, optional): The maximum batch size for runtime. Defaults to None.</span>
|
|
|
|
<span class="s2"> max_num_tokens (int, optional): The maximum number of tokens for runtime. Defaults to None.</span>
|
|
|
|
<span class="s2"> extended_runtime_perf_knob_config (ExtendedRuntimePerfKnobConfig, optional): The extended runtime performance knob configuration for the model. Defaults to None.</span>
|
|
|
|
<span class="s2">"""</span>
|
|
|
|
|
|
<span class="nd">@append_docstring</span><span class="p">(</span><span class="n">LLMARGS_DOCSTRING</span><span class="p">)</span>
|
|
<span class="nd">@dataclass</span>
|
|
<span class="k">class</span> <span class="nc">LlmArgs</span><span class="p">:</span>
|
|
<span class="w"> </span><span class="sd">"""The arguments for constructing a LLM instance.</span>
|
|
|
|
<span class="sd"> Parameters:</span>
|
|
<span class="sd"> """</span>
|
|
<span class="c1"># Explicit arguments</span>
|
|
<span class="n">model</span><span class="p">:</span> <span class="n">Union</span><span class="p">[</span><span class="nb">str</span><span class="p">,</span> <span class="n">Path</span><span class="p">]</span>
|
|
|
|
<span class="n">speculative_model</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="nb">str</span><span class="p">,</span> <span class="n">Path</span><span class="p">]]</span> <span class="o">=</span> <span class="kc">None</span>
|
|
|
|
<span class="n">tokenizer</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="nb">str</span><span class="p">,</span> <span class="n">Path</span><span class="p">,</span> <span class="n">TokenizerBase</span><span class="p">,</span>
|
|
<span class="n">PreTrainedTokenizerBase</span><span class="p">]]</span> <span class="o">=</span> <span class="kc">None</span>
|
|
|
|
<span class="n">tokenizer_mode</span><span class="p">:</span> <span class="n">Literal</span><span class="p">[</span><span class="s1">'auto'</span><span class="p">,</span> <span class="s1">'slow'</span><span class="p">]</span> <span class="o">=</span> <span class="s1">'auto'</span>
|
|
|
|
<span class="n">skip_tokenizer_init</span><span class="p">:</span> <span class="nb">bool</span> <span class="o">=</span> <span class="kc">False</span>
|
|
|
|
<span class="n">trust_remote_code</span><span class="p">:</span> <span class="nb">bool</span> <span class="o">=</span> <span class="kc">False</span>
|
|
|
|
<span class="n">tensor_parallel_size</span><span class="p">:</span> <span class="nb">int</span> <span class="o">=</span> <span class="mi">1</span>
|
|
|
|
<span class="n">dtype</span><span class="p">:</span> <span class="nb">str</span> <span class="o">=</span> <span class="s2">"auto"</span>
|
|
|
|
<span class="n">revision</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="n">tokenizer_revision</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="c1"># Below are all remaining arguments</span>
|
|
<span class="n">pipeline_parallel_size</span><span class="p">:</span> <span class="nb">int</span> <span class="o">=</span> <span class="mi">1</span>
|
|
|
|
<span class="n">context_parallel_size</span><span class="p">:</span> <span class="nb">int</span> <span class="o">=</span> <span class="mi">1</span>
|
|
|
|
<span class="n">moe_tensor_parallel_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="n">moe_expert_parallel_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="n">cp_config</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="nb">dict</span><span class="p">]</span> <span class="o">=</span> <span class="n">field</span><span class="p">(</span><span class="n">default_factory</span><span class="o">=</span><span class="nb">dict</span><span class="p">)</span>
|
|
|
|
<span class="n">auto_parallel</span><span class="p">:</span> <span class="nb">bool</span> <span class="o">=</span> <span class="kc">False</span>
|
|
|
|
<span class="n">auto_parallel_world_size</span><span class="p">:</span> <span class="nb">int</span> <span class="o">=</span> <span class="mi">1</span>
|
|
|
|
<span class="n">load_format</span><span class="p">:</span> <span class="n">Literal</span><span class="p">[</span><span class="s1">'auto'</span><span class="p">,</span> <span class="s1">'dummy'</span><span class="p">]</span> <span class="o">=</span> <span class="s1">'auto'</span>
|
|
|
|
<span class="n">enable_tqdm</span><span class="p">:</span> <span class="nb">bool</span> <span class="o">=</span> <span class="kc">False</span>
|
|
|
|
<span class="c1"># LoRA arguments</span>
|
|
<span class="n">enable_lora</span><span class="p">:</span> <span class="nb">bool</span> <span class="o">=</span> <span class="kc">False</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="n">max_loras</span><span class="p">:</span> <span class="nb">int</span> <span class="o">=</span> <span class="mi">4</span>
|
|
|
|
<span class="n">max_cpu_loras</span><span class="p">:</span> <span class="nb">int</span> <span class="o">=</span> <span class="mi">4</span>
|
|
|
|
<span class="c1"># Prompt adapter arguments</span>
|
|
<span class="n">enable_prompt_adapter</span><span class="p">:</span> <span class="nb">bool</span> <span class="o">=</span> <span class="kc">False</span>
|
|
|
|
<span class="n">max_prompt_adapter_token</span><span class="p">:</span> <span class="nb">int</span> <span class="o">=</span> <span class="mi">0</span>
|
|
|
|
<span class="c1"># Quantization and calibration configurations</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="n">calib_config</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="n">CalibConfig</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</span>
|
|
|
|
<span class="c1"># BuildConfig is introduced to give users a familiar interface to configure the model building.</span>
|
|
<span class="n">build_config</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="n">BuildConfig</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</span>
|
|
|
|
<span class="c1"># Several options from ExecutorConfig, expanded here for less hierarchy</span>
|
|
<span class="n">kv_cache_config</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="n">KvCacheConfig</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</span>
|
|
|
|
<span class="n">enable_chunked_prefill</span><span class="p">:</span> <span class="nb">bool</span> <span class="o">=</span> <span class="kc">False</span>
|
|
|
|
<span class="c1"># TODO[enweiz]: this might affect medusa, and could be removed in the future for API consistency</span>
|
|
<span class="n">decoding_config</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="n">DecodingConfig</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</span>
|
|
|
|
<span class="n">guided_decoding_backend</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="n">logits_post_processor_map</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">Callable</span><span class="p">]]</span> <span class="o">=</span> <span class="kc">None</span>
|
|
|
|
<span class="n">iter_stats_max_iterations</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="n">request_stats_max_iterations</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="n">workspace</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="c1"># A handful of options from PretrainedConfig</span>
|
|
<span class="n">embedding_parallel_mode</span><span class="p">:</span> <span class="nb">str</span> <span class="o">=</span> <span class="s1">'SHARDING_ALONG_VOCAB'</span>
|
|
|
|
<span class="n">fast_build</span><span class="p">:</span> <span class="nb">bool</span> <span class="o">=</span> <span class="kc">False</span>
|
|
|
|
<span class="c1"># Once set, the model will reuse the build_cache</span>
|
|
<span class="n">enable_build_cache</span><span class="p">:</span> <span class="n">Union</span><span class="p">[</span><span class="n">BuildCacheConfig</span><span class="p">,</span> <span class="nb">bool</span><span class="p">]</span> <span class="o">=</span> <span class="kc">False</span>
|
|
|
|
<span class="n">peft_cache_config</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="n">PeftCacheConfig</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</span>
|
|
|
|
<span class="n">scheduler_config</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="n">SchedulerConfig</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</span>
|
|
|
|
<span class="c1"># Speculative decoding parameters</span>
|
|
<span class="n">speculative_config</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">LookaheadDecodingConfig</span><span class="p">]]</span> <span class="o">=</span> <span class="kc">None</span>
|
|
|
|
<span class="n">batching_type</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="n">BatchingType</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</span>
|
|
|
|
<span class="n">normalize_log_probs</span><span class="p">:</span> <span class="nb">bool</span> <span class="o">=</span> <span class="kc">False</span>
|
|
|
|
<span class="n">extended_runtime_perf_knob_config</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span>
|
|
<span class="n">ExtendedRuntimePerfKnobConfig</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</span>
|
|
|
|
<span class="c1"># TODO: remove this option in the future</span>
|
|
<span class="n">use_runtime_defaults</span><span class="p">:</span> <span class="nb">bool</span> <span class="o">=</span> <span class="kc">True</span>
|
|
|
|
<span class="n">max_batch_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="n">max_num_tokens</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="c1"># backend to use</span>
|
|
<span class="n">backend</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="k">def</span> <span class="nf">__post_init__</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
|
|
<span class="c1"># TODO[chunweiy]: Enable this option in the future</span>
|
|
<span class="c1"># Currently we want LLMAPI to be consistent with the lower APIs in the model building, thus disable this to avoid</span>
|
|
<span class="c1"># magics.</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">perform_config_arbitration</span> <span class="o">=</span> <span class="kc">False</span>
|
|
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">skip_tokenizer_init</span><span class="p">:</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">tokenizer</span> <span class="o">=</span> <span class="kc">None</span>
|
|
<span class="k">else</span><span class="p">:</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">tokenizer</span> <span class="o">=</span> <span class="n">tokenizer_factory</span><span class="p">(</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">tokenizer</span><span class="p">,</span>
|
|
<span class="n">trust_remote_code</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">trust_remote_code</span><span class="p">,</span>
|
|
<span class="n">use_fast</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">tokenizer_mode</span> <span class="o">!=</span> <span class="s1">'slow'</span><span class="p">)</span>
|
|
|
|
<span class="k">if</span> <span class="n">torch</span><span class="o">.</span><span class="n">cuda</span><span class="o">.</span><span class="n">get_device_properties</span><span class="p">(</span><span class="mi">0</span><span class="p">)</span><span class="o">.</span><span class="n">major</span> <span class="o"><</span> <span class="mi">8</span><span class="p">:</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">dtype</span> <span class="o">==</span> <span class="s1">'auto'</span><span class="p">:</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">dtype</span> <span class="o">=</span> <span class="s1">'float16'</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">dtype</span> <span class="o">==</span> <span class="s1">'bfloat16'</span><span class="p">:</span>
|
|
<span class="k">raise</span> <span class="ne">RuntimeError</span><span class="p">(</span><span class="s2">"Pre SM 80 GPUs do not support bfloat16"</span><span class="p">)</span>
|
|
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">moe_tensor_parallel_size</span> <span class="ow">is</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">moe_tensor_parallel_size</span> <span class="o">=</span> <span class="o">-</span><span class="mi">1</span>
|
|
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">moe_expert_parallel_size</span> <span class="ow">is</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">moe_expert_parallel_size</span> <span class="o">=</span> <span class="o">-</span><span class="mi">1</span>
|
|
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">parallel_config</span> <span class="o">=</span> <span class="n">_ParallelConfig</span><span class="p">(</span>
|
|
<span class="n">tp_size</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">tensor_parallel_size</span><span class="p">,</span>
|
|
<span class="n">pp_size</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">pipeline_parallel_size</span><span class="p">,</span>
|
|
<span class="n">cp_size</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">context_parallel_size</span><span class="p">,</span>
|
|
<span class="n">moe_tp_size</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">moe_tensor_parallel_size</span><span class="p">,</span>
|
|
<span class="n">moe_ep_size</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">moe_expert_parallel_size</span><span class="p">,</span>
|
|
<span class="n">cp_config</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">cp_config</span><span class="p">,</span>
|
|
<span class="n">auto_parallel</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">auto_parallel</span><span class="p">)</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">parallel_config</span><span class="o">.</span><span class="n">auto_parallel</span><span class="p">:</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">parallel_config</span><span class="o">.</span><span class="n">world_size</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">auto_parallel_world_size</span>
|
|
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">auto_parallel_config</span> <span class="o">=</span> <span class="n">AutoParallelConfig</span><span class="p">(</span>
|
|
<span class="n">sharded_io_allowlist</span><span class="o">=</span><span class="p">[</span>
|
|
<span class="s2">"past_key_value_</span><span class="se">\\</span><span class="s2">d+"</span><span class="p">,</span>
|
|
<span class="s2">"present_key_value_</span><span class="se">\\</span><span class="s2">d*"</span><span class="p">,</span>
|
|
<span class="p">],</span>
|
|
<span class="n">same_buffer_io</span><span class="o">=</span><span class="p">{</span>
|
|
<span class="s2">"past_key_value_(</span><span class="se">\\</span><span class="s2">d+)"</span><span class="p">:</span> <span class="s2">"present_key_value_</span><span class="se">\\</span><span class="s2">1"</span><span class="p">,</span>
|
|
<span class="p">},</span>
|
|
<span class="o">**</span><span class="n">infer_cluster_config</span><span class="p">(),</span>
|
|
<span class="p">)</span>
|
|
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">kv_cache_config</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">kv_cache_config</span> <span class="ow">or</span> <span class="n">KvCacheConfig</span><span class="p">()</span>
|
|
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">scheduler_config</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">scheduler_config</span> <span class="ow">or</span> <span class="n">SchedulerConfig</span><span class="p">()</span>
|
|
|
|
<span class="c1"># This is used to hold th options for convert_checkpoint</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">_convert_checkpoint_options</span> <span class="o">=</span> <span class="p">{}</span>
|
|
|
|
<span class="nd">@classmethod</span>
|
|
<span class="k">def</span> <span class="nf">from_kwargs</span><span class="p">(</span><span class="bp">cls</span><span class="p">,</span> <span class="o">**</span><span class="n">kwargs</span><span class="p">)</span> <span class="o">-></span> <span class="s2">"LlmArgs"</span><span class="p">:</span>
|
|
<span class="n">LlmArgs</span><span class="o">.</span><span class="n">_check_executor_config_options_consistency</span><span class="p">()</span>
|
|
<span class="n">ret</span> <span class="o">=</span> <span class="bp">cls</span><span class="p">(</span><span class="o">**</span><span class="n">kwargs</span><span class="p">)</span>
|
|
<span class="n">ret</span><span class="o">.</span><span class="n">setup</span><span class="p">()</span>
|
|
<span class="k">return</span> <span class="n">ret</span>
|
|
|
|
<span class="k">def</span> <span class="nf">to_dict</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
|
|
<span class="k">return</span> <span class="nb">dict</span><span class="p">(</span>
|
|
<span class="p">(</span><span class="n">field</span><span class="o">.</span><span class="n">name</span><span class="p">,</span> <span class="nb">getattr</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">field</span><span class="o">.</span><span class="n">name</span><span class="p">))</span> <span class="k">for</span> <span class="n">field</span> <span class="ow">in</span> <span class="n">fields</span><span class="p">(</span><span class="bp">self</span><span class="p">))</span>
|
|
|
|
<span class="nd">@staticmethod</span>
|
|
<span class="k">def</span> <span class="nf">_check_executor_config_options_consistency</span><span class="p">():</span>
|
|
<span class="c1"># max_beam_width is not included since vague behavior due to lacking the support for dynamic beam width during</span>
|
|
<span class="c1"># generation</span>
|
|
<span class="n">black_list</span> <span class="o">=</span> <span class="nb">set</span><span class="p">([</span><span class="s2">"max_beam_width"</span><span class="p">])</span>
|
|
<span class="n">executor_config_attrs</span> <span class="o">=</span> <span class="nb">set</span><span class="p">(</span><span class="n">attr</span> <span class="k">for</span> <span class="n">attr</span> <span class="ow">in</span> <span class="nb">dir</span><span class="p">(</span><span class="n">ExecutorConfig</span><span class="p">)</span>
|
|
<span class="k">if</span> <span class="ow">not</span> <span class="n">attr</span><span class="o">.</span><span class="n">startswith</span><span class="p">(</span><span class="s1">'_'</span><span class="p">)</span>
|
|
<span class="ow">and</span> <span class="nb">callable</span><span class="p">(</span><span class="nb">getattr</span><span class="p">(</span><span class="n">ExecutorConfig</span><span class="p">,</span> <span class="n">attr</span><span class="p">)))</span>
|
|
<span class="n">executor_config_attrs</span> <span class="o">-=</span> <span class="n">black_list</span>
|
|
<span class="n">llm_args_attr</span> <span class="o">=</span> <span class="nb">set</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">fields</span><span class="p">(</span><span class="n">LlmArgs</span><span class="p">)])</span>
|
|
<span class="c1"># NOTE: When cpp ExecutorConfig add new options, please add the new options into `_LlmArgs` with docs as well</span>
|
|
<span class="c1"># ASK chunweiy for help if you are not sure about the new options.</span>
|
|
<span class="k">assert</span> <span class="n">executor_config_attrs</span><span class="o">.</span><span class="n">issubset</span><span class="p">(</span>
|
|
<span class="n">llm_args_attr</span>
|
|
<span class="p">),</span> <span class="sa">f</span><span class="s2">"New options found in underlying ExecutorConfig: </span><span class="si">{</span><span class="n">llm_args_attr</span><span class="w"> </span><span class="o">-</span><span class="w"> </span><span class="n">executor_config_attrs</span><span class="si">}</span><span class="s2">"</span>
|
|
|
|
<span class="k">def</span> <span class="nf">setup</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
|
|
<span class="w"> </span><span class="sd">''' This method will setup the configs right before building the model.</span>
|
|
<span class="sd"> It will check the consistency of the configs and arbitrate the conflicts.</span>
|
|
<span class="sd"> '''</span>
|
|
|
|
<span class="k">assert</span> <span class="nb">isinstance</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">model</span><span class="p">,</span>
|
|
<span class="p">(</span><span class="nb">str</span><span class="p">,</span> <span class="n">Path</span><span class="p">)),</span> <span class="sa">f</span><span class="s2">"Invalid model: </span><span class="si">{</span><span class="bp">self</span><span class="o">.</span><span class="n">model</span><span class="si">}</span><span class="s2">"</span>
|
|
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">_setup_embedding_parallel_mode</span><span class="p">()</span>
|
|
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">enable_build_cache</span><span class="p">:</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">enable_build_cache</span> <span class="o">=</span> <span class="n">BuildCacheConfig</span><span class="p">()</span> <span class="k">if</span> <span class="nb">isinstance</span><span class="p">(</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">enable_build_cache</span><span class="p">,</span> <span class="nb">bool</span><span class="p">)</span> <span class="k">else</span> <span class="bp">self</span><span class="o">.</span><span class="n">enable_build_cache</span>
|
|
<span class="k">if</span> <span class="ow">not</span> <span class="nb">isinstance</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">enable_build_cache</span><span class="p">,</span> <span class="n">BuildCacheConfig</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">"Invalid build_cache_config: </span><span class="si">{</span><span class="bp">self</span><span class="o">.</span><span class="n">enable_build_cache</span><span class="si">}</span><span class="s2">"</span><span class="p">)</span>
|
|
<span class="n">model_obj</span> <span class="o">=</span> <span class="n">_ModelWrapper</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">model</span><span class="p">)</span>
|
|
<span class="n">speculative_model_obj</span> <span class="o">=</span> <span class="n">_ModelWrapper</span><span class="p">(</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">speculative_model</span>
|
|
<span class="p">)</span> <span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">speculative_model</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span> <span class="k">else</span> <span class="kc">None</span>
|
|
<span class="k">if</span> <span class="n">model_obj</span><span class="o">.</span><span class="n">is_local_model</span> <span class="ow">and</span> <span class="nb">getattr</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="s1">'backend'</span><span class="p">,</span>
|
|
<span class="kc">None</span><span class="p">)</span> <span class="o">!=</span> <span class="s1">'pytorch'</span><span class="p">:</span>
|
|
<span class="c1"># Load parallel_config from the engine.</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">model_format</span> <span class="o">=</span> <span class="n">ModelLoader</span><span class="o">.</span><span class="n">get_model_format</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">model</span><span class="p">)</span>
|
|
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">model_format</span> <span class="ow">is</span> <span class="n">_ModelFormatKind</span><span class="o">.</span><span class="n">TLLM_ENGINE</span><span class="p">:</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">build_config</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</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="s2">"The build_config is ignored for model format of TLLM_ENGINE."</span>
|
|
<span class="p">)</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">_load_config_from_engine</span><span class="p">(</span><span class="n">model_obj</span><span class="o">.</span><span class="n">model_dir</span><span class="p">)</span>
|
|
<span class="n">runtime_defaults</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">_pretrained_config</span><span class="o">.</span><span class="n">runtime_defaults</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">use_runtime_defaults</span> <span class="ow">and</span> <span class="n">runtime_defaults</span><span class="p">:</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">kv_cache_config</span><span class="o">.</span><span class="n">fill_empty_fields_from_runtime_defaults</span><span class="p">(</span>
|
|
<span class="n">runtime_defaults</span><span class="p">)</span>
|
|
|
|
<span class="c1"># Load parallel_config from the checkpoint.</span>
|
|
<span class="k">elif</span> <span class="bp">self</span><span class="o">.</span><span class="n">model_format</span> <span class="ow">is</span> <span class="n">_ModelFormatKind</span><span class="o">.</span><span class="n">TLLM_CKPT</span><span class="p">:</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">_load_config_from_ckpt</span><span class="p">(</span><span class="n">model_obj</span><span class="o">.</span><span class="n">model_dir</span><span class="p">)</span>
|
|
<span class="k">else</span><span class="p">:</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">model_format</span> <span class="o">=</span> <span class="n">_ModelFormatKind</span><span class="o">.</span><span class="n">HF</span>
|
|
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">speculative_model</span> <span class="ow">and</span> <span class="n">speculative_model_obj</span><span class="o">.</span><span class="n">is_local_model</span><span class="p">:</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">speculative_model_format</span> <span class="o">=</span> <span class="n">_ModelFormatKind</span><span class="o">.</span><span class="n">HF</span>
|
|
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">quant_config</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">quant_config</span> <span class="ow">or</span> <span class="n">QuantConfig</span><span class="p">()</span>
|
|
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">calib_config</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">calib_config</span> <span class="ow">or</span> <span class="n">CalibConfig</span><span class="p">()</span>
|
|
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">build_config</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">build_config</span> <span class="ow">or</span> <span class="n">BuildConfig</span><span class="p">()</span>
|
|
|
|
<span class="c1"># TODO(xiweny): remove the checker when manage weights support all data types</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">fast_build</span> <span class="ow">and</span> <span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">quant_config</span><span class="o">.</span><span class="n">quant_algo</span> <span class="ow">is</span> <span class="n">QuantAlgo</span><span class="o">.</span><span class="n">FP8</span>
|
|
<span class="ow">or</span> <span class="bp">self</span><span class="o">.</span><span class="n">quant_config</span><span class="o">.</span><span class="n">quant_algo</span> <span class="ow">is</span> <span class="kc">None</span><span class="p">):</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">_update_plugin_config</span><span class="p">(</span><span class="s2">"manage_weights"</span><span class="p">,</span> <span class="kc">True</span><span class="p">)</span>
|
|
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">parallel_config</span><span class="o">.</span><span class="n">_world_size</span> <span class="o">==</span> <span class="mi">1</span><span class="p">:</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">build_config</span><span class="o">.</span><span class="n">plugin_config</span><span class="o">.</span><span class="n">nccl_plugin</span> <span class="o">=</span> <span class="kc">None</span>
|
|
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">enable_lora</span><span class="p">:</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">build_config</span><span class="o">.</span><span class="n">plugin_config</span><span class="o">.</span><span class="n">lora_plugin</span> <span class="o">=</span> <span class="s1">'auto'</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">max_lora_rank</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">build_config</span><span class="o">.</span><span class="n">lora_config</span><span class="o">.</span><span class="n">max_lora_rank</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">max_lora_rank</span>
|
|
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">enable_prompt_adapter</span><span class="p">:</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">build_config</span><span class="o">.</span><span class="n">max_prompt_embedding_table_size</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">max_prompt_adapter_token</span> <span class="o">*</span> <span class="bp">self</span><span class="o">.</span><span class="n">build_config</span><span class="o">.</span><span class="n">max_batch_size</span>
|
|
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">perform_config_arbitration</span><span class="p">:</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">_perform_config_arbitration</span><span class="p">()</span>
|
|
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">speculative_config</span><span class="p">:</span>
|
|
<span class="k">if</span> <span class="nb">isinstance</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">speculative_config</span><span class="p">,</span> <span class="n">LookaheadDecodingConfig</span><span class="p">):</span>
|
|
<span class="n">lookahead_config</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">speculative_config</span>
|
|
<span class="c1"># Update the build config</span>
|
|
<span class="n">_</span><span class="p">,</span> <span class="n">_</span><span class="p">,</span> <span class="n">max_draft_tokens</span><span class="p">,</span> <span class="n">_</span> <span class="o">=</span> <span class="n">lookahead_config</span><span class="o">.</span><span class="n">calculate_speculative_resource</span><span class="p">(</span>
|
|
<span class="p">)</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">build_config</span><span class="o">.</span><span class="n">speculative_decoding_mode</span> <span class="o">=</span> <span class="n">SpeculativeDecodingMode</span><span class="o">.</span><span class="n">LOOKAHEAD_DECODING</span>
|
|
<span class="k">if</span> <span class="n">max_draft_tokens</span> <span class="o">></span> <span class="bp">self</span><span class="o">.</span><span class="n">build_config</span><span class="o">.</span><span class="n">max_draft_len</span><span class="p">:</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">build_config</span><span class="o">.</span><span class="n">max_draft_len</span> <span class="o">=</span> <span class="n">max_draft_tokens</span>
|
|
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">decoding_config</span> <span class="o">=</span> <span class="n">DecodingConfig</span><span class="p">(</span>
|
|
<span class="n">decoding_mode</span><span class="o">=</span><span class="n">DecodingMode</span><span class="o">.</span><span class="n">Lookahead</span><span class="p">(),</span>
|
|
<span class="n">lookahead_decoding_config</span><span class="o">=</span><span class="n">lookahead_config</span><span class="p">)</span>
|
|
<span class="k">elif</span> <span class="nb">isinstance</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">speculative_config</span><span class="p">,</span> <span class="n">MedusaDecodingConfig</span><span class="p">):</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">decoding_config</span> <span class="o">=</span> <span class="n">DecodingConfig</span><span class="p">(</span>
|
|
<span class="n">decoding_mode</span><span class="o">=</span><span class="n">DecodingMode</span><span class="o">.</span><span class="n">Medusa</span><span class="p">(),</span>
|
|
<span class="n">medusa_choices</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">speculative_config</span><span class="o">.</span><span class="n">medusa_choices</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="s2">"Speculative config type not recognized"</span><span class="p">)</span>
|
|
|
|
<span class="k">def</span> <span class="nf">_perform_config_arbitration</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
|
|
<span class="w"> </span><span class="sd">'''</span>
|
|
<span class="sd"> Arbitrate the configurations for the model building. The configs between different functional or performance</span>
|
|
<span class="sd"> features might be conflicted, and this method will arbitrate the conflicts and raise errors if necessary.</span>
|
|
<span class="sd"> '''</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">_config_arbitrator</span> <span class="o">=</span> <span class="n">_ConfigArbitrator</span><span class="p">()</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">build_config_mutable</span><span class="p">:</span>
|
|
<span class="k">if</span> <span class="ow">not</span> <span class="bp">self</span><span class="o">.</span><span class="n">build_config</span><span class="o">.</span><span class="n">max_num_tokens</span><span class="p">:</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">build_config</span><span class="o">.</span><span class="n">max_num_tokens</span> <span class="o">=</span> <span class="mi">2048</span>
|
|
|
|
<span class="k">if</span> <span class="ow">not</span> <span class="n">GpuArch</span><span class="o">.</span><span class="n">is_post_ampere</span><span class="p">():</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">_config_arbitrator</span><span class="o">.</span><span class="n">setup</span><span class="p">(</span><span class="s2">"pre-ampere not supported"</span><span class="p">,</span>
|
|
<span class="n">config_name</span><span class="o">=</span><span class="s2">"plugin_config"</span><span class="p">,</span>
|
|
<span class="n">use_paged_context_fmha</span><span class="o">=</span><span class="kc">False</span><span class="p">)</span>
|
|
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">_setup_enable_chunked_context</span><span class="p">()</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">_setup_enable_streaming_llm</span><span class="p">()</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">_setup_quant_config</span><span class="p">()</span>
|
|
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">build_config</span><span class="o">.</span><span class="n">max_beam_width</span> <span class="o">></span> <span class="mi">1</span><span class="p">:</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">_config_arbitrator</span><span class="o">.</span><span class="n">claim_func</span><span class="p">(</span>
|
|
<span class="s2">"beam_search (beam_width > 1)"</span><span class="p">,</span>
|
|
<span class="n">config_name</span><span class="o">=</span><span class="s2">"kv_cache_config"</span><span class="p">,</span>
|
|
<span class="n">enable_block_reuse</span><span class="o">=</span><span class="kc">False</span><span class="p">)</span>
|
|
|
|
<span class="k">else</span><span class="p">:</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">_setup_build_config_into_config_arbitrator</span><span class="p">()</span>
|
|
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">_setup_kv_cache_config</span><span class="p">()</span>
|
|
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">_config_arbitrator</span><span class="p">(</span><span class="n">plugin_config</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">build_config</span><span class="o">.</span><span class="n">plugin_config</span><span class="p">,</span>
|
|
<span class="n">kv_cache_config</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">kv_cache_config</span><span class="p">,</span>
|
|
<span class="n">build_config</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">build_config</span><span class="p">)</span>
|
|
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">_config_arbitrator</span> <span class="o">=</span> <span class="kc">None</span>
|
|
|
|
<span class="nd">@property</span>
|
|
<span class="k">def</span> <span class="nf">build_config_mutable</span><span class="p">(</span><span class="bp">self</span><span class="p">)</span> <span class="o">-></span> <span class="nb">bool</span><span class="p">:</span>
|
|
<span class="k">return</span> <span class="bp">self</span><span class="o">.</span><span class="n">model_format</span> <span class="ow">is</span> <span class="ow">not</span> <span class="n">_ModelFormatKind</span><span class="o">.</span><span class="n">TLLM_ENGINE</span>
|
|
|
|
<span class="k">def</span> <span class="nf">_update_plugin_config</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">str</span><span class="p">,</span> <span class="n">value</span><span class="p">:</span> <span class="n">Any</span><span class="p">):</span>
|
|
<span class="nb">setattr</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">build_config</span><span class="o">.</span><span class="n">plugin_config</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">def</span> <span class="nf">_load_config_from_engine</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">engine_dir</span><span class="p">:</span> <span class="n">Path</span><span class="p">):</span>
|
|
<span class="n">engine_config</span> <span class="o">=</span> <span class="n">EngineConfig</span><span class="o">.</span><span class="n">from_json_file</span><span class="p">(</span><span class="n">engine_dir</span> <span class="o">/</span> <span class="s2">"config.json"</span><span class="p">)</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">_pretrained_config</span> <span class="o">=</span> <span class="n">engine_config</span><span class="o">.</span><span class="n">pretrained_config</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">build_config</span> <span class="o">=</span> <span class="n">engine_config</span><span class="o">.</span><span class="n">build_config</span>
|
|
|
|
<span class="c1"># load and check parallel_config</span>
|
|
<span class="n">mapping</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">_pretrained_config</span><span class="o">.</span><span class="n">mapping</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">parallel_config</span><span class="o">.</span><span class="n">tp_size</span> <span class="ow">not</span> <span class="ow">in</span> <span class="p">(</span><span class="mi">1</span><span class="p">,</span> <span class="n">mapping</span><span class="o">.</span><span class="n">tp_size</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">"tp_size </span><span class="si">{</span><span class="bp">self</span><span class="o">.</span><span class="n">parallel_config</span><span class="o">.</span><span class="n">tp_size</span><span class="si">}</span><span class="s2"> is not consistent with the engine's tp_size </span><span class="si">{</span><span class="n">mapping</span><span class="o">.</span><span class="n">tp_size</span><span class="si">}</span><span class="s2">"</span>
|
|
<span class="p">)</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">parallel_config</span><span class="o">.</span><span class="n">pp_size</span> <span class="ow">not</span> <span class="ow">in</span> <span class="p">(</span><span class="mi">1</span><span class="p">,</span> <span class="n">mapping</span><span class="o">.</span><span class="n">pp_size</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">"pp_size </span><span class="si">{</span><span class="bp">self</span><span class="o">.</span><span class="n">parallel_config</span><span class="o">.</span><span class="n">pp_size</span><span class="si">}</span><span class="s2"> is not consistent with the engine's pp_size </span><span class="si">{</span><span class="n">mapping</span><span class="o">.</span><span class="n">pp_size</span><span class="si">}</span><span class="s2">"</span>
|
|
<span class="p">)</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">parallel_config</span><span class="o">.</span><span class="n">cp_size</span> <span class="ow">not</span> <span class="ow">in</span> <span class="p">(</span><span class="mi">1</span><span class="p">,</span> <span class="n">mapping</span><span class="o">.</span><span class="n">cp_size</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">"cp_size </span><span class="si">{</span><span class="bp">self</span><span class="o">.</span><span class="n">parallel_config</span><span class="o">.</span><span class="n">cp_size</span><span class="si">}</span><span class="s2"> is not consistent with the engine's cp_size </span><span class="si">{</span><span class="n">mapping</span><span class="o">.</span><span class="n">cp_size</span><span class="si">}</span><span class="s2">"</span>
|
|
<span class="p">)</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">parallel_config</span> <span class="o">=</span> <span class="n">_ParallelConfig</span><span class="p">(</span><span class="n">tp_size</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">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">mapping</span><span class="o">.</span><span class="n">cp_size</span><span class="p">,</span>
|
|
<span class="n">moe_tp_size</span><span class="o">=</span><span class="n">mapping</span><span class="o">.</span><span class="n">moe_tp_size</span><span class="p">,</span>
|
|
<span class="n">moe_ep_size</span><span class="o">=</span><span class="n">mapping</span><span class="o">.</span><span class="n">moe_ep_size</span><span class="p">)</span>
|
|
|
|
<span class="k">def</span> <span class="nf">_load_config_from_ckpt</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">ckpt_dir</span><span class="p">:</span> <span class="n">Path</span><span class="p">):</span>
|
|
<span class="n">pretrained_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">ckpt_dir</span> <span class="o">/</span>
|
|
<span class="s2">"config.json"</span><span class="p">)</span>
|
|
<span class="n">tp_size</span> <span class="o">=</span> <span class="n">pretrained_config</span><span class="o">.</span><span class="n">mapping</span><span class="o">.</span><span class="n">tp_size</span>
|
|
<span class="n">pp_size</span> <span class="o">=</span> <span class="n">pretrained_config</span><span class="o">.</span><span class="n">mapping</span><span class="o">.</span><span class="n">pp_size</span>
|
|
<span class="n">cp_size</span> <span class="o">=</span> <span class="n">pretrained_config</span><span class="o">.</span><span class="n">mapping</span><span class="o">.</span><span class="n">cp_size</span>
|
|
<span class="n">moe_tp_size</span> <span class="o">=</span> <span class="n">pretrained_config</span><span class="o">.</span><span class="n">mapping</span><span class="o">.</span><span class="n">moe_tp_size</span>
|
|
<span class="n">moe_ep_size</span> <span class="o">=</span> <span class="n">pretrained_config</span><span class="o">.</span><span class="n">mapping</span><span class="o">.</span><span class="n">moe_ep_size</span>
|
|
<span class="n">world_size</span> <span class="o">=</span> <span class="n">pretrained_config</span><span class="o">.</span><span class="n">mapping</span><span class="o">.</span><span class="n">world_size</span>
|
|
|
|
<span class="c1"># load parallel_config</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">parallel_config</span><span class="o">.</span><span class="n">tp_size</span> <span class="o">!=</span> <span class="mi">1</span> <span class="ow">and</span> <span class="bp">self</span><span class="o">.</span><span class="n">parallel_config</span><span class="o">.</span><span class="n">tp_size</span> <span class="o">!=</span> <span class="n">tp_size</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">"tp_size </span><span class="si">{</span><span class="bp">self</span><span class="o">.</span><span class="n">parallel_config</span><span class="o">.</span><span class="n">tp_size</span><span class="si">}</span><span class="s2"> is not consistent with the checkpoint's tp_size </span><span class="si">{</span><span class="n">tp_size</span><span class="si">}</span><span class="s2">"</span>
|
|
<span class="p">)</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">parallel_config</span><span class="o">.</span><span class="n">pp_size</span> <span class="o">!=</span> <span class="mi">1</span> <span class="ow">and</span> <span class="bp">self</span><span class="o">.</span><span class="n">parallel_config</span><span class="o">.</span><span class="n">pp_size</span> <span class="o">!=</span> <span class="n">pp_size</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">"pp_size </span><span class="si">{</span><span class="bp">self</span><span class="o">.</span><span class="n">parallel_config</span><span class="o">.</span><span class="n">pp_size</span><span class="si">}</span><span class="s2"> is not consistent with the checkpoint's pp_size </span><span class="si">{</span><span class="n">pp_size</span><span class="si">}</span><span class="s2">"</span>
|
|
<span class="p">)</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">parallel_config</span><span class="o">.</span><span class="n">cp_size</span> <span class="o">!=</span> <span class="mi">1</span> <span class="ow">and</span> <span class="bp">self</span><span class="o">.</span><span class="n">parallel_config</span><span class="o">.</span><span class="n">cp_size</span> <span class="o">!=</span> <span class="n">cp_size</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">"cp_size </span><span class="si">{</span><span class="bp">self</span><span class="o">.</span><span class="n">parallel_config</span><span class="o">.</span><span class="n">cp_size</span><span class="si">}</span><span class="s2"> is not consistent with the checkpoint's cp_size </span><span class="si">{</span><span class="n">cp_size</span><span class="si">}</span><span class="s2">"</span>
|
|
<span class="p">)</span>
|
|
<span class="k">if</span> <span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">parallel_config</span><span class="o">.</span><span class="n">auto_parallel</span>
|
|
<span class="ow">and</span> <span class="bp">self</span><span class="o">.</span><span class="n">parallel_config</span><span class="o">.</span><span class="n">world_size</span> <span class="o">!=</span> <span class="mi">1</span> <span class="ow">and</span> <span class="n">world_size</span> <span class="o">!=</span> <span class="mi">1</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">"auto parallel with world_size </span><span class="si">{</span><span class="bp">self</span><span class="o">.</span><span class="n">parallel_config</span><span class="o">.</span><span class="n">world_size</span><span class="si">}</span><span class="s2"> does not support checkpoint with "</span>
|
|
<span class="s2">"world_size </span><span class="si">{world_size}</span><span class="s2"> > 1"</span><span class="p">)</span>
|
|
<span class="k">if</span> <span class="ow">not</span> <span class="bp">self</span><span class="o">.</span><span class="n">parallel_config</span><span class="o">.</span><span class="n">auto_parallel</span><span class="p">:</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">parallel_config</span> <span class="o">=</span> <span class="n">_ParallelConfig</span><span class="p">(</span><span class="n">tp_size</span><span class="o">=</span><span class="n">tp_size</span><span class="p">,</span>
|
|
<span class="n">pp_size</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">cp_size</span><span class="p">,</span>
|
|
<span class="n">moe_tp_size</span><span class="o">=</span><span class="n">moe_tp_size</span><span class="p">,</span>
|
|
<span class="n">moe_ep_size</span><span class="o">=</span><span class="n">moe_ep_size</span><span class="p">)</span>
|
|
|
|
<span class="k">def</span> <span class="nf">_setup_embedding_parallel_mode</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">embedding_parallel_mode</span> <span class="o">==</span> <span class="s1">'NONE'</span><span class="p">:</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">_convert_checkpoint_options</span><span class="p">[</span><span class="s1">'use_parallel_embedding'</span><span class="p">]</span> <span class="o">=</span> <span class="kc">False</span>
|
|
<span class="k">elif</span> <span class="bp">self</span><span class="o">.</span><span class="n">embedding_parallel_mode</span> <span class="o">==</span> <span class="s1">'SHARDING_ALONG_VOCAB'</span><span class="p">:</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">_convert_checkpoint_options</span><span class="p">[</span><span class="s1">'use_parallel_embedding'</span><span class="p">]</span> <span class="o">=</span> <span class="kc">True</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">_convert_checkpoint_options</span><span class="p">[</span><span class="s1">'embedding_sharding_dim'</span><span class="p">]</span> <span class="o">=</span> <span class="mi">0</span>
|
|
<span class="k">elif</span> <span class="bp">self</span><span class="o">.</span><span class="n">embedding_parallel_mode</span> <span class="o">==</span> <span class="s1">'SHARDING_ALONG_HIDDEN'</span><span class="p">:</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">_convert_checkpoint_options</span><span class="p">[</span><span class="s1">'use_parallel_embedding'</span><span class="p">]</span> <span class="o">=</span> <span class="kc">True</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">_convert_checkpoint_options</span><span class="p">[</span><span class="s1">'embedding_sharding_dim'</span><span class="p">]</span> <span class="o">=</span> <span class="mi">1</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="s2">"Invalid embedding_parallel_mode: </span><span class="si">{</span><span class="bp">self</span><span class="o">.</span><span class="n">llm_args</span><span class="o">.</span><span class="n">embedding_parallel_mode</span><span class="si">}</span><span class="s2">"</span>
|
|
<span class="p">)</span>
|
|
|
|
<span class="k">def</span> <span class="nf">_setup_build_config_into_config_arbitrator</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
|
|
<span class="c1"># Setup the ConfigArbitrator with the plugin_config, the runtime configs such as KvCacheConfig should not be</span>
|
|
<span class="c1"># conflict with it.</span>
|
|
<span class="n">build_config</span> <span class="o">=</span> <span class="n">asdict</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">build_config</span><span class="p">)</span>
|
|
<span class="k">del</span> <span class="n">build_config</span><span class="p">[</span><span class="s1">'plugin_config'</span><span class="p">]</span>
|
|
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">_config_arbitrator</span><span class="o">.</span><span class="n">setup</span><span class="p">(</span><span class="s2">"BuildConfig is readonly"</span><span class="p">,</span>
|
|
<span class="n">config_name</span><span class="o">=</span><span class="s2">"build_config"</span><span class="p">,</span>
|
|
<span class="o">**</span><span class="n">build_config</span><span class="p">)</span>
|
|
|
|
<span class="n">plugin_config</span> <span class="o">=</span> <span class="n">asdict</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">build_config</span><span class="o">.</span><span class="n">plugin_config</span><span class="p">)</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">_config_arbitrator</span><span class="o">.</span><span class="n">setup</span><span class="p">(</span><span class="s2">"PluginConfig is readonly"</span><span class="p">,</span>
|
|
<span class="n">config_name</span><span class="o">=</span><span class="s2">"plugin_config"</span><span class="p">,</span>
|
|
<span class="o">**</span><span class="n">plugin_config</span><span class="p">)</span>
|
|
|
|
<span class="k">def</span> <span class="nf">_setup_enable_chunked_context</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
|
|
|
|
<span class="k">def</span> <span class="nf">fallback</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">"Disabling chunked context due to configuration conflict."</span><span class="p">)</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">enable_chunked_prefill</span> <span class="o">=</span> <span class="kc">False</span>
|
|
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">enable_chunked_prefill</span><span class="p">:</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">build_config_mutable</span><span class="p">:</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">_config_arbitrator</span><span class="o">.</span><span class="n">claim_perf</span><span class="p">(</span><span class="s2">"chunked_context"</span><span class="p">,</span>
|
|
<span class="n">config_name</span><span class="o">=</span><span class="s2">"plugin_config"</span><span class="p">,</span>
|
|
<span class="n">use_paged_context_fmha</span><span class="o">=</span><span class="kc">True</span><span class="p">,</span>
|
|
<span class="n">fallback</span><span class="o">=</span><span class="n">fallback</span><span class="p">)</span>
|
|
|
|
<span class="k">def</span> <span class="nf">_setup_enable_streaming_llm</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">build_config</span><span class="o">.</span><span class="n">plugin_config</span><span class="o">.</span><span class="n">streamingllm</span><span class="p">:</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">_validate_kv_cache_config</span><span class="p">()</span>
|
|
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">_config_arbitrator</span><span class="o">.</span><span class="n">claim_func</span><span class="p">(</span><span class="s2">"streamingllm"</span><span class="p">,</span>
|
|
<span class="n">config_name</span><span class="o">=</span><span class="s2">"plugin_config"</span><span class="p">,</span>
|
|
<span class="n">streamingllm</span><span class="o">=</span><span class="kc">True</span><span class="p">,</span>
|
|
<span class="n">use_paged_context_fmha</span><span class="o">=</span><span class="kc">False</span><span class="p">)</span>
|
|
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">_config_arbitrator</span><span class="o">.</span><span class="n">claim_func</span><span class="p">(</span><span class="s2">"streamingllm"</span><span class="p">,</span>
|
|
<span class="n">config_name</span><span class="o">=</span><span class="s2">"kv_cache_config"</span><span class="p">,</span>
|
|
<span class="n">enable_block_reuse</span><span class="o">=</span><span class="kc">False</span><span class="p">)</span>
|
|
|
|
<span class="k">def</span> <span class="nf">_validate_kv_cache_config</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">kv_cache_config</span> <span class="ow">is</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="k">raise</span> <span class="ne">ValueError</span><span class="p">(</span><span class="s2">"KvCacheConfig is required for streaming LLM."</span><span class="p">)</span>
|
|
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">kv_cache_config</span><span class="o">.</span><span class="n">max_attention_window</span> <span class="ow">is</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="k">raise</span> <span class="ne">ValueError</span><span class="p">(</span>
|
|
<span class="s2">"KvCacheConfig.max_attention_window should be set for streaming LLM."</span>
|
|
<span class="p">)</span>
|
|
<span class="k">if</span> <span class="nb">any</span><span class="p">(</span><span class="n">i</span> <span class="o"><=</span> <span class="mi">0</span> <span class="k">for</span> <span class="n">i</span> <span class="ow">in</span> <span class="bp">self</span><span class="o">.</span><span class="n">kv_cache_config</span><span class="o">.</span><span class="n">max_attention_window</span><span class="p">):</span>
|
|
<span class="k">raise</span> <span class="ne">ValueError</span><span class="p">(</span>
|
|
<span class="s2">"Elements in KvCacheConfig.max_attention_window should be greater than 0."</span>
|
|
<span class="p">)</span>
|
|
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">kv_cache_config</span><span class="o">.</span><span class="n">sink_token_length</span> <span class="ow">is</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="k">raise</span> <span class="ne">ValueError</span><span class="p">(</span>
|
|
<span class="s2">"KvCacheConfig.sink_token_length should be set for streaming LLM."</span>
|
|
<span class="p">)</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">kv_cache_config</span><span class="o">.</span><span class="n">sink_token_length</span> <span class="o"><=</span> <span class="mi">0</span><span class="p">:</span>
|
|
<span class="k">raise</span> <span class="ne">ValueError</span><span class="p">(</span>
|
|
<span class="s2">"KvCacheConfig.sink_token_length should be greater than 0."</span><span class="p">)</span>
|
|
|
|
<span class="k">def</span> <span class="nf">_setup_kv_cache_config</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
|
|
<span class="k">assert</span> <span class="bp">self</span><span class="o">.</span><span class="n">kv_cache_config</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span>
|
|
|
|
<span class="k">if</span> <span class="ow">not</span> <span class="n">GpuArch</span><span class="o">.</span><span class="n">is_post_ampere</span><span class="p">():</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">_config_arbitrator</span><span class="o">.</span><span class="n">setup</span><span class="p">(</span><span class="s2">"pre-ampere not supported"</span><span class="p">,</span>
|
|
<span class="n">config_name</span><span class="o">=</span><span class="s2">"kv_cache_config"</span><span class="p">,</span>
|
|
<span class="n">enable_block_reuse</span><span class="o">=</span><span class="kc">False</span><span class="p">)</span>
|
|
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">kv_cache_config</span><span class="o">.</span><span class="n">enable_block_reuse</span><span class="p">:</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">_config_arbitrator</span><span class="o">.</span><span class="n">claim_func</span><span class="p">(</span><span class="s2">"enable_block_reuse"</span><span class="p">,</span>
|
|
<span class="n">config_name</span><span class="o">=</span><span class="s2">"kv_cache_config"</span><span class="p">,</span>
|
|
<span class="n">enable_block_reuse</span><span class="o">=</span><span class="kc">True</span><span class="p">)</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">_config_arbitrator</span><span class="o">.</span><span class="n">claim_func</span><span class="p">(</span><span class="s2">"enable_block_reuse"</span><span class="p">,</span>
|
|
<span class="n">config_name</span><span class="o">=</span><span class="s2">"plugin_config"</span><span class="p">,</span>
|
|
<span class="n">use_paged_context_fmha</span><span class="o">=</span><span class="kc">True</span><span class="p">)</span>
|
|
|
|
<span class="k">def</span> <span class="nf">_setup_quant_config</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_config</span><span class="o">.</span><span class="n">quant_algo</span> <span class="ow">is</span> <span class="n">QuantAlgo</span><span class="o">.</span><span class="n">FP8</span><span class="p">:</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">_config_arbitrator</span><span class="o">.</span><span class="n">claim_func</span><span class="p">(</span><span class="s2">"fp8_quant"</span><span class="p">,</span>
|
|
<span class="n">config_name</span><span class="o">=</span><span class="s2">"plugin_config"</span><span class="p">,</span>
|
|
<span class="n">use_paged_context_fmha</span><span class="o">=</span><span class="kc">False</span><span class="p">)</span>
|
|
|
|
<span class="k">def</span> <span class="nf">__setstate__</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">state</span><span class="p">):</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="vm">__dict__</span><span class="o">.</span><span class="n">update</span><span class="p">(</span><span class="n">state</span><span class="p">)</span>
|
|
|
|
<span class="k">def</span> <span class="nf">__getstate__</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
|
|
<span class="n">state</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="vm">__dict__</span><span class="o">.</span><span class="n">copy</span><span class="p">()</span>
|
|
<span class="k">if</span> <span class="s1">'_config_arbitrator'</span> <span class="ow">in</span> <span class="n">state</span><span class="p">:</span>
|
|
<span class="k">del</span> <span class="n">state</span><span class="p">[</span><span class="s1">'_config_arbitrator'</span><span class="p">]</span>
|
|
<span class="k">return</span> <span class="n">state</span>
|
|
|
|
|
|
<span class="k">class</span> <span class="nc">ConfigArbitrateError</span><span class="p">(</span><span class="ne">Exception</span><span class="p">):</span>
|
|
<span class="w"> </span><span class="sd">''' Exception raised when there is a conflict in configurations. '''</span>
|
|
|
|
<span class="k">def</span> <span class="fm">__init__</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">message</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">message</span><span class="p">)</span>
|
|
|
|
|
|
<span class="k">class</span> <span class="nc">_ConfigArbitrator</span><span class="p">:</span>
|
|
<span class="w"> </span><span class="sd">''' The ConfigArbitrator will arbitrate the options from different sources and raise errors if there are conflicts. '''</span>
|
|
|
|
<span class="k">def</span> <span class="fm">__init__</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
|
|
<span class="c1"># Dict of configs, the format is {config_name: {option: value}}</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">virtual_configs</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">Dict</span><span class="p">[</span><span class="nb">str</span><span class="p">,</span> <span class="n">Any</span><span class="p">]]</span> <span class="o">=</span> <span class="p">{}</span>
|
|
<span class="c1"># The claims for functionalities, the format is {config_name: [(func_name, {option: value})]}</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">func_claims</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">List</span><span class="p">[</span><span class="n">Tuple</span><span class="p">[</span><span class="nb">str</span><span class="p">,</span> <span class="nb">dict</span><span class="p">]]]</span> <span class="o">=</span> <span class="p">{}</span>
|
|
<span class="c1"># The claims for performances, the format is {perf_name: [(config_name, {option: value}, fallback)]},</span>
|
|
<span class="c1"># the fallback is a callback function to be called when the performance is abandoned.</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">perf_claims</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">List</span><span class="p">[</span><span class="n">Tuple</span><span class="p">[</span><span class="nb">str</span><span class="p">,</span> <span class="nb">dict</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="kc">None</span><span class="p">]]]]]</span> <span class="o">=</span> <span class="p">{}</span>
|
|
<span class="c1"># Track where the option settings came from, this will be used for messages when encountered conflicts.</span>
|
|
<span class="c1"># The format is {config_name: {option: error_information}}</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">option_sources</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">Dict</span><span class="p">[</span><span class="nb">str</span><span class="p">,</span> <span class="nb">str</span><span class="p">]]</span> <span class="o">=</span> <span class="p">{}</span>
|
|
|
|
<span class="k">def</span> <span class="fm">__call__</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="o">**</span><span class="n">configs</span><span class="p">)</span> <span class="o">-></span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="w"> </span><span class="sd">'''</span>
|
|
<span class="sd"> Args:</span>
|
|
<span class="sd"> configs: name to config instance for each config need to be arbitrated.</span>
|
|
<span class="sd"> '''</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">_arbitrate</span><span class="p">()</span>
|
|
|
|
<span class="c1"># Apply the successfully arbitrated virtual configs to the real configs</span>
|
|
<span class="k">for</span> <span class="n">name</span><span class="p">,</span> <span class="n">config</span> <span class="ow">in</span> <span class="n">configs</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">in</span> <span class="bp">self</span><span class="o">.</span><span class="n">virtual_configs</span><span class="p">:</span>
|
|
<span class="n">virtual_config</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">virtual_configs</span><span class="p">[</span><span class="n">name</span><span class="p">]</span>
|
|
<span class="k">for</span> <span class="n">option</span><span class="p">,</span> <span class="n">value</span> <span class="ow">in</span> <span class="n">virtual_config</span><span class="o">.</span><span class="n">items</span><span class="p">():</span>
|
|
<span class="nb">setattr</span><span class="p">(</span><span class="n">config</span><span class="p">,</span> <span class="n">option</span><span class="p">,</span> <span class="n">value</span><span class="p">)</span>
|
|
|
|
<span class="k">def</span> <span class="nf">setup</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">info</span><span class="p">:</span> <span class="nb">str</span><span class="p">,</span> <span class="n">config_name</span><span class="p">:</span> <span class="nb">str</span><span class="p">,</span> <span class="o">**</span><span class="n">kwargs</span><span class="p">):</span>
|
|
<span class="w"> </span><span class="sd">''' Setup with some pre-defined configs comes from environment such as GPU arch. '''</span>
|
|
<span class="n">config</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">virtual_configs</span><span class="o">.</span><span class="n">setdefault</span><span class="p">(</span><span class="n">config_name</span><span class="p">,</span> <span class="p">{})</span>
|
|
<span class="n">option_sources</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">option_sources</span><span class="o">.</span><span class="n">setdefault</span><span class="p">(</span><span class="n">config_name</span><span class="p">,</span> <span class="p">{})</span>
|
|
<span class="k">for</span> <span class="n">option</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">assert</span> <span class="n">config</span><span class="o">.</span><span class="n">get</span><span class="p">(</span><span class="n">option</span><span class="p">,</span> <span class="n">value</span><span class="p">)</span> <span class="o">==</span> <span class="n">value</span>
|
|
<span class="n">config</span><span class="p">[</span><span class="n">option</span><span class="p">]</span> <span class="o">=</span> <span class="n">value</span>
|
|
<span class="n">option_sources</span><span class="p">[</span><span class="n">option</span><span class="p">]</span> <span class="o">=</span> <span class="n">info</span>
|
|
|
|
<span class="k">def</span> <span class="nf">claim_func</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">func</span><span class="p">:</span> <span class="nb">str</span><span class="p">,</span> <span class="n">config_name</span><span class="p">:</span> <span class="nb">str</span><span class="p">,</span> <span class="o">**</span><span class="n">options</span><span class="p">):</span>
|
|
<span class="w"> </span><span class="sd">''' Claim a functionality demanding with configs and options.</span>
|
|
<span class="sd"> The functionality should be fulfilled, or errors will be raised. '''</span>
|
|
|
|
<span class="n">claims</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">func_claims</span><span class="o">.</span><span class="n">setdefault</span><span class="p">(</span><span class="n">config_name</span><span class="p">,</span> <span class="p">[])</span>
|
|
<span class="n">claims</span><span class="o">.</span><span class="n">append</span><span class="p">((</span><span class="n">func</span><span class="p">,</span> <span class="n">options</span><span class="p">))</span>
|
|
|
|
<span class="k">def</span> <span class="nf">claim_perf</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span>
|
|
<span class="n">perf</span><span class="p">:</span> <span class="nb">str</span><span class="p">,</span>
|
|
<span class="n">config_name</span><span class="p">:</span> <span class="nb">str</span><span class="p">,</span>
|
|
<span class="n">fallback</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="kc">None</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">options</span><span class="p">):</span>
|
|
<span class="w"> </span><span class="sd">''' Claim a performance demanding for configs and options.</span>
|
|
<span class="sd"> The performance could be abandoned if the demanding is not available.'''</span>
|
|
<span class="n">claims</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">perf_claims</span><span class="o">.</span><span class="n">setdefault</span><span class="p">(</span><span class="n">perf</span><span class="p">,</span> <span class="p">[])</span>
|
|
<span class="n">claims</span><span class="o">.</span><span class="n">append</span><span class="p">((</span><span class="n">config_name</span><span class="p">,</span> <span class="n">options</span><span class="p">,</span> <span class="n">fallback</span><span class="p">))</span>
|
|
|
|
<span class="k">def</span> <span class="nf">_arbitrate</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
|
|
<span class="w"> </span><span class="sd">''' Arbitrate the configs for all the functionalities and performances. '''</span>
|
|
|
|
<span class="c1"># Resolve functionality claims</span>
|
|
<span class="k">for</span> <span class="n">config_name</span><span class="p">,</span> <span class="n">funcs</span> <span class="ow">in</span> <span class="bp">self</span><span class="o">.</span><span class="n">func_claims</span><span class="o">.</span><span class="n">items</span><span class="p">():</span>
|
|
<span class="n">virtual_config</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">virtual_configs</span><span class="o">.</span><span class="n">setdefault</span><span class="p">(</span><span class="n">config_name</span><span class="p">,</span> <span class="p">{})</span>
|
|
<span class="n">option_sources</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">option_sources</span><span class="o">.</span><span class="n">setdefault</span><span class="p">(</span><span class="n">config_name</span><span class="p">,</span> <span class="p">{})</span>
|
|
<span class="k">for</span> <span class="n">func</span><span class="p">,</span> <span class="n">options</span> <span class="ow">in</span> <span class="n">funcs</span><span class="p">:</span>
|
|
<span class="k">for</span> <span class="n">option</span><span class="p">,</span> <span class="n">value</span> <span class="ow">in</span> <span class="n">options</span><span class="o">.</span><span class="n">items</span><span class="p">():</span>
|
|
<span class="k">if</span> <span class="n">option</span> <span class="ow">in</span> <span class="n">virtual_config</span><span class="p">:</span>
|
|
<span class="k">if</span> <span class="n">virtual_config</span><span class="p">[</span><span class="n">option</span><span class="p">]</span> <span class="o">!=</span> <span class="n">value</span><span class="p">:</span>
|
|
<span class="n">existing_func</span> <span class="o">=</span> <span class="n">option_sources</span><span class="p">[</span><span class="n">option</span><span class="p">]</span>
|
|
<span class="k">raise</span> <span class="n">ConfigArbitrateError</span><span class="p">(</span>
|
|
<span class="sa">f</span><span class="s2">"Cannot set '</span><span class="si">{</span><span class="n">option</span><span class="si">}</span><span class="s2">' to be '</span><span class="si">{</span><span class="n">value</span><span class="si">}</span><span class="s2">' when enabling '</span><span class="si">{</span><span class="n">func</span><span class="si">}</span><span class="s2">', "</span>
|
|
<span class="sa">f</span><span class="s2">"since '</span><span class="si">{</span><span class="n">existing_func</span><span class="si">}</span><span class="s2">' has set it to be '</span><span class="si">{</span><span class="n">virtual_config</span><span class="p">[</span><span class="n">option</span><span class="p">]</span><span class="si">}</span><span class="s2">'."</span>
|
|
<span class="p">)</span>
|
|
<span class="k">else</span><span class="p">:</span>
|
|
<span class="n">virtual_config</span><span class="p">[</span><span class="n">option</span><span class="p">]</span> <span class="o">=</span> <span class="n">value</span>
|
|
<span class="c1"># Track where the setting came from</span>
|
|
<span class="n">option_sources</span><span class="p">[</span><span class="n">option</span><span class="p">]</span> <span class="o">=</span> <span class="n">func</span>
|
|
|
|
<span class="c1"># copy for restore</span>
|
|
<span class="c1"># Resolve performance claims</span>
|
|
<span class="k">for</span> <span class="n">perf</span><span class="p">,</span> <span class="n">options</span> <span class="ow">in</span> <span class="bp">self</span><span class="o">.</span><span class="n">perf_claims</span><span class="o">.</span><span class="n">items</span><span class="p">():</span>
|
|
<span class="n">option_sources</span> <span class="o">=</span> <span class="n">copy</span><span class="o">.</span><span class="n">copy</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">option_sources</span><span class="p">)</span>
|
|
<span class="n">virtual_configs</span> <span class="o">=</span> <span class="n">copy</span><span class="o">.</span><span class="n">copy</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">virtual_configs</span><span class="p">)</span>
|
|
<span class="n">restore</span> <span class="o">=</span> <span class="kc">False</span>
|
|
<span class="k">for</span> <span class="n">config_name</span><span class="p">,</span> <span class="n">options</span><span class="p">,</span> <span class="n">fallback</span> <span class="ow">in</span> <span class="n">options</span><span class="p">:</span>
|
|
<span class="n">virtual_config</span> <span class="o">=</span> <span class="n">virtual_configs</span><span class="o">.</span><span class="n">setdefault</span><span class="p">(</span><span class="n">config_name</span><span class="p">,</span> <span class="p">{})</span>
|
|
<span class="n">option_source</span> <span class="o">=</span> <span class="n">option_sources</span><span class="o">.</span><span class="n">setdefault</span><span class="p">(</span><span class="n">config_name</span><span class="p">,</span> <span class="p">{})</span>
|
|
<span class="k">for</span> <span class="n">option</span><span class="p">,</span> <span class="n">value</span> <span class="ow">in</span> <span class="n">options</span><span class="o">.</span><span class="n">items</span><span class="p">():</span>
|
|
<span class="k">if</span> <span class="n">option</span> <span class="ow">in</span> <span class="n">virtual_config</span> <span class="ow">and</span> <span class="n">virtual_config</span><span class="p">[</span>
|
|
<span class="n">option</span><span class="p">]</span> <span class="o">!=</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">"Ignoring performance claim '</span><span class="si">{</span><span class="n">perf</span><span class="si">}</span><span class="s2">' for option '</span><span class="si">{</span><span class="n">option</span><span class="si">}</span><span class="s2">' due to conflict."</span>
|
|
<span class="p">)</span>
|
|
<span class="n">restore</span> <span class="o">=</span> <span class="kc">True</span>
|
|
<span class="k">else</span><span class="p">:</span>
|
|
<span class="n">virtual_config</span><span class="p">[</span><span class="n">option</span><span class="p">]</span> <span class="o">=</span> <span class="n">value</span>
|
|
<span class="n">option_source</span><span class="p">[</span><span class="n">option</span><span class="p">]</span> <span class="o">=</span> <span class="n">perf</span>
|
|
<span class="k">if</span> <span class="n">restore</span><span class="p">:</span> <span class="k">break</span>
|
|
<span class="k">if</span> <span class="n">restore</span><span class="p">:</span>
|
|
<span class="k">if</span> <span class="n">fallback</span><span class="p">:</span> <span class="n">fallback</span><span class="p">()</span>
|
|
<span class="k">break</span>
|
|
|
|
<span class="k">if</span> <span class="ow">not</span> <span class="n">restore</span><span class="p">:</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">option_sources</span> <span class="o">=</span> <span class="n">option_sources</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">virtual_configs</span> <span class="o">=</span> <span class="n">virtual_configs</span>
|
|
|
|
|
|
<span class="nd">@dataclass</span>
|
|
<span class="k">class</span> <span class="nc">_ModelRuntimeContext</span><span class="p">:</span>
|
|
<span class="w"> </span><span class="sd">''' _ModelRuntimeContext holds the minimum runtime resources for running a model.</span>
|
|
<span class="sd"> It could be a runtime cache in MPI nodes.</span>
|
|
<span class="sd"> '''</span>
|
|
<span class="n">engine</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="n">Engine</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</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="n">model_info</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="n">_ModelInfo</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</span>
|
|
|
|
<span class="c1"># This is only used when build-cache is enabled</span>
|
|
<span class="n">engine_path</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="nd">@property</span>
|
|
<span class="k">def</span> <span class="nf">model_arch</span><span class="p">(</span><span class="bp">self</span><span class="p">)</span> <span class="o">-></span> <span class="nb">str</span><span class="p">:</span>
|
|
<span class="c1"># "LlaMACausalForLM" or "OPTForCausalLM" and so on</span>
|
|
<span class="k">return</span> <span class="bp">self</span><span class="o">.</span><span class="n">engine</span><span class="o">.</span><span class="n">config</span><span class="o">.</span><span class="n">pretrained_config</span><span class="p">[</span><span class="s1">'architecture'</span><span class="p">]</span>
|
|
|
|
|
|
<span class="k">class</span> <span class="nc">ModelLoader</span><span class="p">:</span>
|
|
<span class="w"> </span><span class="sd">''' The ModelLoader is used to build an end-to-end model for a single-gpu.</span>
|
|
<span class="sd"> It accepts model name or a local model dir, and will download the model if necessary.</span>
|
|
<span class="sd"> '''</span>
|
|
|
|
<span class="k">def</span> <span class="fm">__init__</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span>
|
|
<span class="n">llm_args</span><span class="p">:</span> <span class="n">LlmArgs</span><span class="p">,</span>
|
|
<span class="n">workspace</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="nb">str</span> <span class="o">|</span> <span class="n">tempfile</span><span class="o">.</span><span class="n">TemporaryDirectory</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">llm_build_stats</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="s2">"LlmBuildStats"</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</span><span class="p">):</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">llm_args</span> <span class="o">=</span> <span class="n">llm_args</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">_workspace</span> <span class="o">=</span> <span class="n">workspace</span> <span class="ow">or</span> <span class="n">tempfile</span><span class="o">.</span><span class="n">TemporaryDirectory</span><span class="p">()</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">llm_build_stats</span> <span class="o">=</span> <span class="n">llm_build_stats</span> <span class="ow">or</span> <span class="n">LlmBuildStats</span><span class="p">()</span>
|
|
|
|
<span class="k">assert</span> <span class="bp">self</span><span class="o">.</span><span class="n">llm_args</span><span class="o">.</span><span class="n">build_config</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">build_config</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">llm_args</span><span class="o">.</span><span class="n">build_config</span>
|
|
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">model_obj</span> <span class="o">=</span> <span class="n">_ModelWrapper</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">llm_args</span><span class="o">.</span><span class="n">model</span><span class="p">)</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">speculative_model_obj</span> <span class="o">=</span> <span class="n">_ModelWrapper</span><span class="p">(</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">llm_args</span><span class="o">.</span><span class="n">speculative_model</span>
|
|
<span class="p">)</span> <span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">llm_args</span><span class="o">.</span><span class="n">speculative_model</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span> <span class="k">else</span> <span class="kc">None</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">convert_checkpoint_options</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">llm_args</span><span class="o">.</span><span class="n">_convert_checkpoint_options</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">rank</span> <span class="o">=</span> <span class="n">mpi_rank</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">llm_args</span><span class="o">.</span><span class="n">parallel_config</span><span class="o">.</span><span class="n">to_mapping</span><span class="p">()</span>
|
|
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">_build_pipeline</span> <span class="o">=</span> <span class="p">[]</span>
|
|
|
|
<span class="c1"># For model from hub, the _model_dir is None, and will updated once downloaded</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">_model_dir</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span>
|
|
<span class="n">Path</span><span class="p">]</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">model_obj</span><span class="o">.</span><span class="n">model_dir</span> <span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">model_obj</span><span class="o">.</span><span class="n">is_local_model</span> <span class="k">else</span> <span class="kc">None</span>
|
|
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">_speculative_model_dir</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span>
|
|
<span class="n">Path</span><span class="p">]</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">speculative_model_obj</span><span class="o">.</span><span class="n">model_dir</span> <span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">speculative_model_obj</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span> <span class="ow">and</span> <span class="bp">self</span><span class="o">.</span><span class="n">model_obj</span><span class="o">.</span><span class="n">is_local_model</span> <span class="k">else</span> <span class="kc">None</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">_model_info</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="n">_ModelInfo</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">_model_format</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">llm_args</span><span class="o">.</span><span class="n">model_format</span>
|
|
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">auto_parallel_config</span> <span class="o">=</span> <span class="n">AutoParallelConfig</span><span class="p">(</span>
|
|
<span class="n">world_size</span><span class="o">=</span><span class="n">llm_args</span><span class="o">.</span><span class="n">parallel_config</span><span class="o">.</span><span class="n">world_size</span> <span class="k">if</span> <span class="n">llm_args</span><span class="o">.</span>
|
|
<span class="n">parallel_config</span><span class="o">.</span><span class="n">auto_parallel</span> <span class="k">else</span> <span class="mi">1</span><span class="p">)</span>
|
|
<span class="n">default_config</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">llm_args</span><span class="o">.</span><span class="n">auto_parallel_config</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">auto_parallel_config</span><span class="o">.</span><span class="n">set_defaults</span><span class="p">(</span>
|
|
<span class="n">cluster_key</span><span class="o">=</span><span class="n">default_config</span><span class="o">.</span><span class="n">cluster_key</span><span class="p">,</span>
|
|
<span class="n">cluster_info</span><span class="o">=</span><span class="n">default_config</span><span class="o">.</span><span class="n">cluster_info</span><span class="p">,</span>
|
|
<span class="n">same_buffer_io</span><span class="o">=</span><span class="n">default_config</span><span class="o">.</span><span class="n">same_buffer_io</span><span class="p">,</span>
|
|
<span class="n">sharded_io_allowlist</span><span class="o">=</span><span class="n">default_config</span><span class="o">.</span><span class="n">sharded_io_allowlist</span><span class="p">,</span>
|
|
<span class="p">)</span>
|
|
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">_gather_build_steps</span><span class="p">()</span>
|
|
|
|
<span class="k">def</span> <span class="nf">_gather_build_steps</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
|
|
<span class="c1"># Prepare the model processing pipeline</span>
|
|
<span class="k">if</span> <span class="nb">isinstance</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">llm_args</span><span class="o">.</span><span class="n">model</span><span class="p">,</span> <span class="n">Module</span><span class="p">):</span>
|
|
<span class="c1"># Build engine from user provided model</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">_build_pipeline</span><span class="o">.</span><span class="n">append</span><span class="p">(</span>
|
|
<span class="p">(</span><span class="s2">"Build TensorRT-LLM engine"</span><span class="p">,</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">_build_engine_from_inmemory_model</span><span class="p">))</span>
|
|
<span class="k">return</span>
|
|
|
|
<span class="k">if</span> <span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">model_obj</span><span class="o">.</span><span class="n">is_hub_model</span>
|
|
<span class="ow">and</span> <span class="bp">self</span><span class="o">.</span><span class="n">_model_format</span> <span class="ow">is</span> <span class="ow">not</span> <span class="n">_ModelFormatKind</span><span class="o">.</span><span class="n">TLLM_ENGINE</span><span class="p">)</span> <span class="ow">or</span> <span class="p">(</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">speculative_model_obj</span>
|
|
<span class="ow">and</span> <span class="bp">self</span><span class="o">.</span><span class="n">speculative_model_obj</span><span class="o">.</span><span class="n">is_hub_model</span><span class="p">):</span>
|
|
<span class="c1"># Download HF model if necessary</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">model_obj</span><span class="o">.</span><span class="n">model_name</span> <span class="ow">is</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="k">raise</span> <span class="ne">ValueError</span><span class="p">(</span>
|
|
<span class="s2">"Either model_dir or model should be provided to ModelConfig."</span>
|
|
<span class="p">)</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">_build_pipeline</span><span class="o">.</span><span class="n">append</span><span class="p">(</span>
|
|
<span class="p">(</span><span class="s2">"Downloading HF model"</span><span class="p">,</span> <span class="bp">self</span><span class="o">.</span><span class="n">_download_hf_model</span><span class="p">))</span>
|
|
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">_model_format</span> <span class="ow">is</span> <span class="n">_ModelFormatKind</span><span class="o">.</span><span class="n">HF</span><span class="p">:</span>
|
|
<span class="c1"># HF -> TRT checkpoints -> engine</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">_build_pipeline</span><span class="o">.</span><span class="n">append</span><span class="p">(</span>
|
|
<span class="p">(</span><span class="s2">"Loading HF model to memory"</span><span class="p">,</span> <span class="bp">self</span><span class="o">.</span><span class="n">_load_model_from_hf</span><span class="p">))</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">_build_pipeline</span><span class="o">.</span><span class="n">append</span><span class="p">(</span>
|
|
<span class="p">(</span><span class="s2">"Building TRT-LLM engine"</span><span class="p">,</span> <span class="bp">self</span><span class="o">.</span><span class="n">_build_engine</span><span class="p">))</span>
|
|
<span class="k">elif</span> <span class="bp">self</span><span class="o">.</span><span class="n">_model_format</span> <span class="ow">is</span> <span class="n">_ModelFormatKind</span><span class="o">.</span><span class="n">TLLM_CKPT</span><span class="p">:</span>
|
|
<span class="c1"># TRT checkpoints -> engine</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">_build_pipeline</span><span class="o">.</span><span class="n">append</span><span class="p">((</span><span class="s2">"Loading TRT checkpoints to memory"</span><span class="p">,</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">_load_model_from_ckpt</span><span class="p">))</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">_build_pipeline</span><span class="o">.</span><span class="n">append</span><span class="p">(</span>
|
|
<span class="p">(</span><span class="s2">"Build TRT-LLM engine"</span><span class="p">,</span> <span class="bp">self</span><span class="o">.</span><span class="n">_build_engine</span><span class="p">))</span>
|
|
<span class="k">elif</span> <span class="bp">self</span><span class="o">.</span><span class="n">_model_format</span> <span class="ow">is</span> <span class="n">_ModelFormatKind</span><span class="o">.</span><span class="n">TLLM_ENGINE</span><span class="p">:</span>
|
|
<span class="c1"># Nothing need to do</span>
|
|
<span class="k">pass</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="s2">"Unknown model format </span><span class="si">{</span><span class="bp">self</span><span class="o">.</span><span class="n">_model_format</span><span class="si">}</span><span class="s2">"</span><span class="p">)</span>
|
|
|
|
<span class="k">class</span> <span class="nc">BuildPipeline</span><span class="p">:</span>
|
|
|
|
<span class="k">def</span> <span class="fm">__init__</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">enable_tqdm</span><span class="p">:</span> <span class="nb">bool</span><span class="p">,</span> <span class="n">labels</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="n">step_handlers</span><span class="p">:</span> <span class="n">List</span><span class="p">[</span><span class="n">Callable</span><span class="p">],</span>
|
|
<span class="n">llm_build_stats</span><span class="p">:</span> <span class="s2">"LlmBuildStats"</span><span class="p">):</span>
|
|
<span class="k">assert</span> <span class="nb">len</span><span class="p">(</span><span class="n">labels</span><span class="p">)</span> <span class="o">==</span> <span class="nb">len</span><span class="p">(</span><span class="n">step_handlers</span><span class="p">)</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">labels</span> <span class="o">=</span> <span class="n">labels</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">step_handlers</span> <span class="o">=</span> <span class="n">step_handlers</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">llm_build_stats</span> <span class="o">=</span> <span class="n">llm_build_stats</span>
|
|
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">to_log</span> <span class="o">=</span> <span class="n">mpi_rank</span><span class="p">()</span> <span class="o">==</span> <span class="mi">0</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">counter</span> <span class="o">=</span> <span class="mi">0</span>
|
|
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">progress_bar</span> <span class="o">=</span> <span class="n">tqdm</span><span class="p">(</span>
|
|
<span class="n">total</span><span class="o">=</span><span class="nb">len</span><span class="p">(</span><span class="n">labels</span><span class="p">))</span> <span class="k">if</span> <span class="n">enable_tqdm</span> <span class="ow">and</span> <span class="bp">self</span><span class="o">.</span><span class="n">to_log</span> <span class="k">else</span> <span class="kc">None</span>
|
|
|
|
<span class="k">def</span> <span class="fm">__call__</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
|
|
<span class="n">start_time</span> <span class="o">=</span> <span class="n">time</span><span class="o">.</span><span class="n">time</span><span class="p">()</span>
|
|
|
|
<span class="k">for</span> <span class="n">i</span> <span class="ow">in</span> <span class="nb">range</span><span class="p">(</span><span class="nb">len</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">labels</span><span class="p">)):</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">step_forward</span><span class="p">()</span>
|
|
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">to_log</span><span class="p">:</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">progress_bar</span><span class="p">:</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">progress_bar</span><span class="o">.</span><span class="n">close</span><span class="p">()</span>
|
|
<span class="k">else</span><span class="p">:</span>
|
|
<span class="n">overall_latency</span> <span class="o">=</span> <span class="n">time</span><span class="o">.</span><span class="n">time</span><span class="p">()</span> <span class="o">-</span> <span class="n">start_time</span>
|
|
<span class="n">print_colored</span><span class="p">(</span><span class="s2">"Loading model done.</span><span class="se">\n</span><span class="s2">"</span><span class="p">,</span> <span class="s1">'bold_green'</span><span class="p">)</span>
|
|
<span class="n">print_colored</span><span class="p">(</span>
|
|
<span class="s1">'Total latency: </span><span class="si">{:.3f}</span><span class="s1">s</span><span class="se">\n</span><span class="s1">'</span><span class="o">.</span><span class="n">format</span><span class="p">(</span><span class="n">overall_latency</span><span class="p">),</span>
|
|
<span class="s1">'grey'</span><span class="p">)</span>
|
|
|
|
<span class="k">def</span> <span class="nf">step_forward</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
|
|
<span class="n">n_steps</span> <span class="o">=</span> <span class="nb">len</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">labels</span><span class="p">)</span>
|
|
|
|
<span class="n">label</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">labels</span><span class="p">[</span><span class="bp">self</span><span class="o">.</span><span class="n">counter</span><span class="p">]</span>
|
|
|
|
<span class="c1"># display step information</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">to_log</span><span class="p">:</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">progress_bar</span><span class="p">:</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">progress_bar</span><span class="o">.</span><span class="n">set_description</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">labels</span><span class="p">[</span><span class="bp">self</span><span class="o">.</span><span class="n">counter</span><span class="p">])</span>
|
|
<span class="k">else</span><span class="p">:</span>
|
|
<span class="n">print_colored</span><span class="p">(</span><span class="s2">"Loading Model: "</span><span class="p">)</span>
|
|
<span class="n">print_colored</span><span class="p">(</span><span class="sa">f</span><span class="s2">"[</span><span class="si">{</span><span class="bp">self</span><span class="o">.</span><span class="n">counter</span><span class="o">+</span><span class="mi">1</span><span class="si">}</span><span class="s2">/</span><span class="si">{</span><span class="n">n_steps</span><span class="si">}</span><span class="s2">]</span><span class="se">\t</span><span class="s2">"</span><span class="p">,</span>
|
|
<span class="s1">'bold_green'</span><span class="p">)</span>
|
|
<span class="n">print_colored</span><span class="p">(</span><span class="sa">f</span><span class="s2">"</span><span class="si">{</span><span class="n">label</span><span class="si">}</span><span class="se">\n</span><span class="s2">"</span><span class="p">)</span>
|
|
|
|
<span class="c1"># execute the step</span>
|
|
<span class="n">start_time</span> <span class="o">=</span> <span class="n">time</span><span class="o">.</span><span class="n">time</span><span class="p">()</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">step_handlers</span><span class="p">[</span><span class="bp">self</span><span class="o">.</span><span class="n">counter</span><span class="p">]()</span>
|
|
<span class="c1"># release resource after each step</span>
|
|
<span class="n">release_gc</span><span class="p">()</span>
|
|
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">progress_bar</span><span class="p">:</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">progress_bar</span><span class="o">.</span><span class="n">update</span><span class="p">(</span><span class="mi">1</span><span class="p">)</span>
|
|
|
|
<span class="n">latency</span> <span class="o">=</span> <span class="n">time</span><span class="o">.</span><span class="n">time</span><span class="p">()</span> <span class="o">-</span> <span class="n">start_time</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">to_log</span> <span class="ow">and</span> <span class="ow">not</span> <span class="bp">self</span><span class="o">.</span><span class="n">progress_bar</span><span class="p">:</span>
|
|
<span class="n">print_colored</span><span class="p">(</span><span class="s2">"Time: </span><span class="si">{:.3f}</span><span class="s2">s</span><span class="se">\n</span><span class="s2">"</span><span class="o">.</span><span class="n">format</span><span class="p">(</span><span class="n">latency</span><span class="p">),</span> <span class="s1">'grey'</span><span class="p">)</span>
|
|
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">llm_build_stats</span><span class="o">.</span><span class="n">build_steps_info</span><span class="o">.</span><span class="n">append</span><span class="p">((</span><span class="n">label</span><span class="p">,</span> <span class="n">latency</span><span class="p">))</span>
|
|
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">counter</span> <span class="o">+=</span> <span class="mi">1</span>
|
|
|
|
<span class="k">def</span> <span class="fm">__call__</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">engine_dir</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="n">Path</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">Path</span><span class="p">:</span>
|
|
<span class="w"> </span><span class="sd">'''</span>
|
|
<span class="sd"> The engine_dir is the path to save the built engine.</span>
|
|
<span class="sd"> '''</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">llm_args</span><span class="o">.</span><span class="n">model_format</span> <span class="ow">is</span> <span class="n">_ModelFormatKind</span><span class="o">.</span><span class="n">TLLM_ENGINE</span><span class="p">:</span>
|
|
<span class="k">return</span> <span class="bp">self</span><span class="o">.</span><span class="n">model_obj</span><span class="o">.</span><span class="n">model_dir</span>
|
|
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">llm_args</span><span class="o">.</span><span class="n">parallel_config</span><span class="o">.</span><span class="n">is_multi_gpu</span><span class="p">:</span>
|
|
<span class="n">torch</span><span class="o">.</span><span class="n">cuda</span><span class="o">.</span><span class="n">set_device</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">rank</span><span class="p">)</span>
|
|
|
|
<span class="n">pipeline</span> <span class="o">=</span> <span class="n">ModelLoader</span><span class="o">.</span><span class="n">BuildPipeline</span><span class="p">(</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">llm_args</span><span class="o">.</span><span class="n">enable_tqdm</span><span class="p">,</span>
|
|
<span class="p">[</span><span class="n">label</span> <span class="k">for</span> <span class="n">label</span><span class="p">,</span> <span class="n">_</span> <span class="ow">in</span> <span class="bp">self</span><span class="o">.</span><span class="n">_build_pipeline</span><span class="p">],</span>
|
|
<span class="p">[</span><span class="n">handler</span> <span class="k">for</span> <span class="n">_</span><span class="p">,</span> <span class="n">handler</span> <span class="ow">in</span> <span class="bp">self</span><span class="o">.</span><span class="n">_build_pipeline</span><span class="p">],</span>
|
|
<span class="n">llm_build_stats</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">llm_build_stats</span><span class="p">,</span>
|
|
<span class="p">)</span>
|
|
<span class="n">pipeline</span><span class="p">()</span>
|
|
|
|
<span class="k">assert</span> <span class="n">engine_dir</span>
|
|
|
|
<span class="n">runtime_context</span> <span class="o">=</span> <span class="n">_ModelRuntimeContext</span><span class="p">(</span>
|
|
<span class="n">engine</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">_engine</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">mapping</span><span class="p">,</span>
|
|
<span class="n">model_info</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">_model_info</span><span class="p">,</span>
|
|
<span class="p">)</span>
|
|
<span class="n">ModelLoader</span><span class="o">.</span><span class="n">save</span><span class="p">(</span><span class="n">runtime_context</span><span class="p">,</span> <span class="bp">self</span><span class="o">.</span><span class="n">model_obj</span><span class="o">.</span><span class="n">model_dir</span><span class="p">,</span> <span class="n">engine_dir</span><span class="p">)</span>
|
|
<span class="k">return</span> <span class="n">engine_dir</span>
|
|
|
|
<span class="k">def</span> <span class="fm">__enter__</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="k">def</span> <span class="fm">__exit__</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">exc_type</span><span class="p">,</span> <span class="n">exc_value</span><span class="p">,</span> <span class="n">traceback</span><span class="p">):</span>
|
|
<span class="k">for</span> <span class="n">attr_name</span> <span class="ow">in</span> <span class="nb">dir</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
|
|
<span class="k">if</span> <span class="ow">not</span> <span class="nb">callable</span><span class="p">(</span><span class="nb">getattr</span><span class="p">(</span>
|
|
<span class="bp">self</span><span class="p">,</span> <span class="n">attr_name</span><span class="p">))</span> <span class="ow">and</span> <span class="ow">not</span> <span class="n">attr_name</span><span class="o">.</span><span class="n">startswith</span><span class="p">(</span><span class="s2">"__"</span><span class="p">):</span>
|
|
<span class="k">if</span> <span class="n">attr_name</span> <span class="ow">not</span> <span class="ow">in</span> <span class="p">(</span><span class="s1">'model_format'</span><span class="p">,</span> <span class="s1">'workspace'</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">attr_name</span><span class="p">,</span> <span class="kc">None</span><span class="p">)</span>
|
|
|
|
<span class="n">release_gc</span><span class="p">()</span>
|
|
|
|
<span class="nd">@property</span>
|
|
<span class="k">def</span> <span class="nf">workspace</span><span class="p">(</span><span class="bp">self</span><span class="p">)</span> <span class="o">-></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">_workspace</span>
|
|
|
|
<span class="nd">@property</span>
|
|
<span class="k">def</span> <span class="nf">model_format</span><span class="p">(</span><span class="bp">self</span><span class="p">)</span> <span class="o">-></span> <span class="n">_ModelFormatKind</span><span class="p">:</span>
|
|
<span class="k">return</span> <span class="bp">self</span><span class="o">.</span><span class="n">_model_format</span>
|
|
|
|
<span class="nd">@staticmethod</span>
|
|
<span class="k">def</span> <span class="nf">save</span><span class="p">(</span>
|
|
<span class="n">model</span><span class="p">:</span> <span class="n">_ModelRuntimeContext</span><span class="p">,</span>
|
|
<span class="n">model_dir</span><span class="p">:</span> <span class="nb">str</span><span class="p">,</span>
|
|
<span class="n">engine_dir</span><span class="p">:</span> <span class="nb">str</span><span class="p">,</span>
|
|
<span class="p">):</span>
|
|
<span class="w"> </span><span class="sd">''' Save the built engine on a single GPU to the given path. '''</span>
|
|
<span class="n">mapping</span> <span class="o">=</span> <span class="n">model</span><span class="o">.</span><span class="n">mapping</span>
|
|
<span class="n">rank</span> <span class="o">=</span> <span class="n">mapping</span><span class="o">.</span><span class="n">rank</span>
|
|
|
|
<span class="k">def</span> <span class="nf">copy_hf_tokenizer_data_to_engine_dir</span><span class="p">():</span>
|
|
<span class="c1"># Copy the HF tokenizer stuff to the engine dir so that we can use the engine dir as a standalone model dir</span>
|
|
<span class="c1"># supports end-to-end task.</span>
|
|
<span class="c1"># This is only for HF model for now, not available for users' customized tokenizers.</span>
|
|
<span class="kn">import</span> <span class="nn">shutil</span>
|
|
|
|
<span class="k">for</span> <span class="n">name</span> <span class="ow">in</span> <span class="n">os</span><span class="o">.</span><span class="n">listdir</span><span class="p">(</span><span class="n">model_dir</span><span class="p">):</span>
|
|
<span class="n">src</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">model_dir</span><span class="p">,</span> <span class="n">name</span><span class="p">)</span>
|
|
<span class="n">dst</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">engine_dir</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">startswith</span><span class="p">(</span><span class="s1">'tokenizer'</span><span class="p">):</span>
|
|
<span class="n">src</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">realpath</span><span class="p">(</span><span class="n">src</span><span class="p">)</span> <span class="k">if</span> <span class="n">os</span><span class="o">.</span><span class="n">path</span><span class="o">.</span><span class="n">islink</span><span class="p">(</span><span class="n">src</span><span class="p">)</span> <span class="k">else</span> <span class="n">src</span>
|
|
<span class="k">if</span> <span class="n">os</span><span class="o">.</span><span class="n">path</span><span class="o">.</span><span class="n">isdir</span><span class="p">(</span><span class="n">src</span><span class="p">):</span>
|
|
<span class="n">shutil</span><span class="o">.</span><span class="n">copytree</span><span class="p">(</span><span class="n">src</span><span class="p">,</span> <span class="n">dst</span><span class="p">,</span> <span class="n">dirs_exist_ok</span><span class="o">=</span><span class="kc">True</span><span class="p">)</span>
|
|
<span class="k">else</span><span class="p">:</span>
|
|
<span class="n">shutil</span><span class="o">.</span><span class="n">copy2</span><span class="p">(</span><span class="n">src</span><span class="p">,</span> <span class="n">dst</span><span class="p">)</span>
|
|
|
|
<span class="n">model</span><span class="o">.</span><span class="n">engine</span><span class="o">.</span><span class="n">save</span><span class="p">(</span><span class="n">engine_dir</span><span class="p">)</span>
|
|
<span class="k">if</span> <span class="n">rank</span> <span class="o">==</span> <span class="mi">0</span><span class="p">:</span>
|
|
<span class="n">copy_hf_tokenizer_data_to_engine_dir</span><span class="p">()</span>
|
|
|
|
<span class="nd">@staticmethod</span>
|
|
<span class="k">def</span> <span class="nf">get_model_format</span><span class="p">(</span><span class="n">model_dir</span><span class="p">:</span> <span class="nb">str</span><span class="p">)</span> <span class="o">-></span> <span class="n">_ModelFormatKind</span><span class="p">:</span>
|
|
<span class="w"> </span><span class="sd">''' Get the format of the model. '''</span>
|
|
<span class="k">if</span> <span class="ow">not</span> <span class="p">(</span><span class="n">Path</span><span class="p">(</span><span class="n">model_dir</span><span class="p">)</span> <span class="o">/</span> <span class="s1">'config.json'</span><span class="p">)</span><span class="o">.</span><span class="n">exists</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">"Failed to infer model format because no config.json exists in </span><span class="si">{</span><span class="n">model_dir</span><span class="si">}</span><span class="s2">"</span>
|
|
<span class="p">)</span>
|
|
|
|
<span class="k">with</span> <span class="nb">open</span><span class="p">(</span><span class="n">Path</span><span class="p">(</span><span class="n">model_dir</span><span class="p">)</span> <span class="o">/</span> <span class="s1">'config.json'</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">try</span><span class="p">:</span>
|
|
<span class="k">if</span> <span class="s1">'pretrained_config'</span> <span class="ow">in</span> <span class="n">config</span> <span class="ow">and</span> <span class="s1">'build_config'</span> <span class="ow">in</span> <span class="n">config</span><span class="p">:</span>
|
|
<span class="n">model_format</span> <span class="o">=</span> <span class="n">_ModelFormatKind</span><span class="o">.</span><span class="n">TLLM_ENGINE</span>
|
|
<span class="n">EngineConfig</span><span class="o">.</span><span class="n">from_json_file</span><span class="p">(</span><span class="n">Path</span><span class="p">(</span><span class="n">model_dir</span><span class="p">)</span> <span class="o">/</span> <span class="s1">'config.json'</span><span class="p">)</span>
|
|
<span class="k">elif</span> <span class="s1">'architecture'</span> <span class="ow">in</span> <span class="n">config</span> <span class="ow">and</span> <span class="s1">'dtype'</span> <span class="ow">in</span> <span class="n">config</span><span class="p">:</span>
|
|
<span class="n">model_format</span> <span class="o">=</span> <span class="n">_ModelFormatKind</span><span class="o">.</span><span class="n">TLLM_CKPT</span>
|
|
<span class="n">PretrainedConfig</span><span class="o">.</span><span class="n">from_checkpoint</span><span class="p">(</span><span class="n">model_dir</span><span class="p">)</span>
|
|
<span class="k">else</span><span class="p">:</span>
|
|
<span class="n">model_format</span> <span class="o">=</span> <span class="n">_ModelFormatKind</span><span class="o">.</span><span class="n">HF</span>
|
|
<span class="n">AutoConfig</span><span class="o">.</span><span class="n">from_hugging_face</span><span class="p">(</span><span class="n">model_dir</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">ValueError</span><span class="p">(</span>
|
|
<span class="sa">f</span><span class="s2">"Inferred model format </span><span class="si">{</span><span class="n">model_format</span><span class="si">}</span><span class="s2">, but failed to load config.json: </span><span class="si">{</span><span class="n">e</span><span class="si">}</span><span class="s2">"</span>
|
|
<span class="p">)</span>
|
|
<span class="k">else</span><span class="p">:</span>
|
|
<span class="k">return</span> <span class="n">model_format</span>
|
|
|
|
<span class="k">def</span> <span class="nf">_download_hf_model</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
|
|
<span class="w"> </span><span class="sd">''' Download HF model from third-party model hub like www.modelscope.cn or huggingface. '''</span>
|
|
<span class="n">model_dir</span> <span class="o">=</span> <span class="kc">None</span>
|
|
<span class="c1"># Only the rank0 are allowed to download model</span>
|
|
<span class="k">if</span> <span class="n">mpi_rank</span><span class="p">()</span> <span class="o">==</span> <span class="mi">0</span><span class="p">:</span>
|
|
<span class="k">assert</span> <span class="bp">self</span><span class="o">.</span><span class="n">_workspace</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span>
|
|
<span class="k">assert</span> <span class="nb">isinstance</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">model_obj</span><span class="o">.</span><span class="n">model_name</span><span class="p">,</span> <span class="nb">str</span><span class="p">)</span>
|
|
<span class="c1"># this will download only once when multiple MPI processes are running</span>
|
|
|
|
<span class="n">model_dir</span> <span class="o">=</span> <span class="n">download_hf_model</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">model_obj</span><span class="o">.</span><span class="n">model_name</span><span class="p">,</span>
|
|
<span class="n">revision</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">llm_args</span><span class="o">.</span><span class="n">revision</span><span class="p">)</span>
|
|
<span class="n">print_colored</span><span class="p">(</span><span class="sa">f</span><span class="s2">"Downloaded model to </span><span class="si">{</span><span class="n">model_dir</span><span class="si">}</span><span class="se">\n</span><span class="s2">"</span><span class="p">,</span> <span class="s1">'grey'</span><span class="p">)</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">speculative_model_obj</span><span class="p">:</span>
|
|
<span class="n">speculative_model_dir</span> <span class="o">=</span> <span class="n">download_hf_model</span><span class="p">(</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">speculative_model_obj</span><span class="o">.</span><span class="n">model_name</span><span class="p">)</span>
|
|
<span class="n">print_colored</span><span class="p">(</span><span class="sa">f</span><span class="s2">"Downloaded model to </span><span class="si">{</span><span class="n">speculative_model_dir</span><span class="si">}</span><span class="se">\n</span><span class="s2">"</span><span class="p">,</span>
|
|
<span class="s1">'grey'</span><span class="p">)</span>
|
|
<span class="c1"># Make all the processes got the same model_dir</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">_model_dir</span> <span class="o">=</span> <span class="n">mpi_broadcast</span><span class="p">(</span><span class="n">model_dir</span><span class="p">,</span> <span class="n">root</span><span class="o">=</span><span class="mi">0</span><span class="p">)</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">model_obj</span><span class="o">.</span><span class="n">model_dir</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">_model_dir</span> <span class="c1"># mark as a local model</span>
|
|
<span class="k">assert</span> <span class="bp">self</span><span class="o">.</span><span class="n">model_obj</span><span class="o">.</span><span class="n">is_local_model</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">speculative_model_obj</span><span class="p">:</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">_speculative_model_dir</span> <span class="o">=</span> <span class="n">mpi_broadcast</span><span class="p">(</span><span class="n">speculative_model_dir</span><span class="p">,</span>
|
|
<span class="n">root</span><span class="o">=</span><span class="mi">0</span><span class="p">)</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">speculative_model_obj</span><span class="o">.</span><span class="n">model_dir</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">_speculative_model_dir</span>
|
|
|
|
<span class="k">assert</span> <span class="bp">self</span><span class="o">.</span><span class="n">speculative_model_obj</span><span class="o">.</span><span class="n">is_local_model</span>
|
|
|
|
<span class="k">def</span> <span class="nf">_load_model_from_hf</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
|
|
<span class="w"> </span><span class="sd">''' Load a TRT-LLM model from a HF model. '''</span>
|
|
<span class="k">assert</span> <span class="bp">self</span><span class="o">.</span><span class="n">_model_dir</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span>
|
|
<span class="n">model_cls</span> <span class="o">=</span> <span class="n">AutoModelForCausalLM</span><span class="o">.</span><span class="n">get_trtllm_model_class</span><span class="p">(</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">_model_dir</span><span class="p">,</span> <span class="bp">self</span><span class="o">.</span><span class="n">llm_args</span><span class="o">.</span><span class="n">trust_remote_code</span><span class="p">,</span>
|
|
<span class="nb">hasattr</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">llm_args</span><span class="p">,</span> <span class="s2">"speculative_model"</span><span class="p">)</span>
|
|
<span class="ow">and</span> <span class="bp">self</span><span class="o">.</span><span class="n">llm_args</span><span class="o">.</span><span class="n">speculative_model</span><span class="p">)</span>
|
|
|
|
<span class="c1"># Update quant_config if it's ModelOpt quantized ckpt</span>
|
|
<span class="n">user_quant_config</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">llm_args</span><span class="o">.</span><span class="n">quant_config</span>
|
|
<span class="n">hf_quant_config_path</span> <span class="o">=</span> <span class="n">Path</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">_model_dir</span><span class="p">)</span> <span class="o">/</span> <span class="s2">"hf_quant_config.json"</span>
|
|
<span class="k">if</span> <span class="n">hf_quant_config_path</span><span class="o">.</span><span class="n">exists</span><span class="p">():</span>
|
|
<span class="n">logger</span><span class="o">.</span><span class="n">info</span><span class="p">(</span>
|
|
<span class="sa">f</span><span class="s2">"Found </span><span class="si">{</span><span class="n">hf_quant_config_path</span><span class="si">}</span><span class="s2">, pre-quantized checkpoints are used."</span>
|
|
<span class="p">)</span>
|
|
<span class="n">already_quantized</span> <span class="o">=</span> <span class="kc">True</span>
|
|
<span class="k">with</span> <span class="nb">open</span><span class="p">(</span><span class="n">hf_quant_config_path</span><span class="p">,</span> <span class="s2">"r"</span><span class="p">)</span> <span class="k">as</span> <span class="n">f</span><span class="p">:</span>
|
|
<span class="n">hf_quant_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">hf_quant_algo</span> <span class="o">=</span> <span class="n">hf_quant_config</span><span class="p">[</span><span class="s2">"quantization"</span><span class="p">]</span><span class="o">.</span><span class="n">get</span><span class="p">(</span>
|
|
<span class="s2">"quant_algo"</span><span class="p">)</span>
|
|
<span class="k">if</span> <span class="n">hf_quant_algo</span> <span class="o">==</span> <span class="s2">"FP8"</span> <span class="ow">and</span> <span class="n">user_quant_config</span><span class="o">.</span><span class="n">quant_algo</span> \
|
|
<span class="ow">and</span> <span class="n">user_quant_config</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">FP8</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">"Expecting quant_algo to be FP8, got </span><span class="si">{</span><span class="n">user_quant_config</span><span class="o">.</span><span class="n">quant_algo</span><span class="si">}</span><span class="s2">."</span>
|
|
<span class="p">)</span>
|
|
<span class="n">user_quant_config</span><span class="o">.</span><span class="n">quant_algo</span> <span class="o">=</span> <span class="n">hf_quant_algo</span>
|
|
<span class="n">logger</span><span class="o">.</span><span class="n">info</span><span class="p">(</span><span class="sa">f</span><span class="s2">"quant_algo is set to </span><span class="si">{</span><span class="n">hf_quant_algo</span><span class="si">}</span><span class="s2">"</span><span class="p">)</span>
|
|
|
|
<span class="n">hf_kv_cache_quant_algo</span> <span class="o">=</span> <span class="n">hf_quant_config</span><span class="p">[</span><span class="s2">"quantization"</span><span class="p">]</span><span class="o">.</span><span class="n">get</span><span class="p">(</span>
|
|
<span class="s2">"kv_cache_quant_algo"</span><span class="p">)</span>
|
|
<span class="k">if</span> <span class="n">hf_kv_cache_quant_algo</span> <span class="o">!=</span> <span class="n">user_quant_config</span><span class="o">.</span><span class="n">kv_cache_quant_algo</span><span class="p">:</span>
|
|
<span class="k">if</span> <span class="n">user_quant_config</span><span class="o">.</span><span class="n">kv_cache_quant_algo</span> <span class="ow">is</span> <span class="kc">None</span><span class="p">:</span>
|
|
<span class="n">user_quant_config</span><span class="o">.</span><span class="n">kv_cache_quant_algo</span> <span class="o">=</span> <span class="n">hf_kv_cache_quant_algo</span>
|
|
<span class="n">logger</span><span class="o">.</span><span class="n">info</span><span class="p">(</span>
|
|
<span class="sa">f</span><span class="s2">"kv_cache_quant_algo is set to </span><span class="si">{</span><span class="n">hf_kv_cache_quant_algo</span><span class="si">}</span><span class="s2">"</span>
|
|
<span class="p">)</span>
|
|
<span class="k">elif</span> <span class="n">user_quant_config</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="ow">and</span> <span class="n">hf_kv_cache_quant_algo</span> <span class="ow">is</span> <span class="kc">None</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">"User specified kv_cache_quant_algo </span><span class="si">{</span><span class="n">user_quant_config</span><span class="o">.</span><span class="n">kv_cache_quant_algo</span><span class="si">}</span><span class="s2"> "</span>
|
|
<span class="sa">f</span><span class="s2">"will overwrite </span><span class="si">{</span><span class="n">hf_kv_cache_quant_algo</span><span class="si">}</span><span class="s2"> from </span><span class="si">{</span><span class="n">hf_quant_config_path</span><span class="si">}</span><span class="s2">."</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="s2">"User specified kv_cache_quant_algo </span><span class="si">{</span><span class="n">user_quant_config</span><span class="o">.</span><span class="n">kv_cache_quant_algo</span><span class="si">}</span><span class="s2">, "</span>
|
|
<span class="sa">f</span><span class="s2">"while it's </span><span class="si">{</span><span class="n">hf_kv_cache_quant_algo</span><span class="si">}</span><span class="s2"> in </span><span class="si">{</span><span class="n">hf_quant_config_path</span><span class="si">}</span><span class="s2">."</span>
|
|
<span class="p">)</span>
|
|
<span class="k">else</span><span class="p">:</span>
|
|
<span class="n">already_quantized</span> <span class="o">=</span> <span class="kc">False</span>
|
|
|
|
<span class="c1"># FP4 Gemm force to use plugin.</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">llm_args</span><span class="o">.</span><span class="n">quant_config</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="bp">self</span><span class="o">.</span><span class="n">llm_args</span><span class="o">.</span><span class="n">build_config</span><span class="o">.</span><span class="n">plugin_config</span><span class="o">.</span><span class="n">gemm_plugin</span> <span class="o">=</span> <span class="s2">"nvfp4"</span>
|
|
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">llm_args</span><span class="o">.</span><span class="n">load_format</span> <span class="o">==</span> <span class="s1">'dummy'</span><span class="p">:</span>
|
|
<span class="n">config</span> <span class="o">=</span> <span class="n">model_cls</span><span class="o">.</span><span class="n">config_class</span><span class="o">.</span><span class="n">from_hugging_face</span><span class="p">(</span>
|
|
<span class="nb">str</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">_model_dir</span><span class="p">),</span>
|
|
<span class="n">dtype</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">llm_args</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="bp">self</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="bp">self</span><span class="o">.</span><span class="n">llm_args</span><span class="o">.</span><span class="n">quant_config</span><span class="p">,</span>
|
|
<span class="o">**</span><span class="bp">self</span><span class="o">.</span><span class="n">convert_checkpoint_options</span><span class="p">,</span>
|
|
<span class="p">)</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">model</span> <span class="o">=</span> <span class="n">model_cls</span><span class="p">(</span><span class="n">config</span><span class="p">)</span>
|
|
<span class="k">elif</span> <span class="bp">self</span><span class="o">.</span><span class="n">llm_args</span><span class="o">.</span><span class="n">quant_config</span><span class="o">.</span><span class="n">requires_calibration</span> <span class="ow">and</span> <span class="ow">not</span> <span class="n">already_quantized</span><span class="p">:</span>
|
|
<span class="k">assert</span> <span class="bp">self</span><span class="o">.</span><span class="n">workspace</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span>
|
|
<span class="n">checkpoint_dir</span> <span class="o">=</span> <span class="sa">f</span><span class="s2">"</span><span class="si">{</span><span class="bp">self</span><span class="o">.</span><span class="n">workspace</span><span class="si">}</span><span class="s2">/quantized-checkpoint"</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">rank</span> <span class="o">==</span> <span class="mi">0</span><span class="p">:</span>
|
|
<span class="n">model_cls</span><span class="o">.</span><span class="n">quantize</span><span class="p">(</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">_model_dir</span><span class="p">,</span>
|
|
<span class="n">checkpoint_dir</span><span class="p">,</span>
|
|
<span class="n">dtype</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">llm_args</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="bp">self</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="bp">self</span><span class="o">.</span><span class="n">llm_args</span><span class="o">.</span><span class="n">quant_config</span><span class="p">,</span>
|
|
<span class="o">**</span><span class="bp">self</span><span class="o">.</span><span class="n">llm_args</span><span class="o">.</span><span class="n">calib_config</span><span class="o">.</span><span class="n">to_dict</span><span class="p">(),</span>
|
|
<span class="n">trust_remote_code</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">llm_args</span><span class="o">.</span><span class="n">trust_remote_code</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">llm_args</span><span class="o">.</span><span class="n">parallel_config</span><span class="o">.</span><span class="n">is_multi_gpu</span><span class="p">:</span>
|
|
<span class="n">mpi_barrier</span><span class="p">()</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">model</span> <span class="o">=</span> <span class="n">model_cls</span><span class="o">.</span><span class="n">from_checkpoint</span><span class="p">(</span><span class="n">checkpoint_dir</span><span class="p">,</span>
|
|
<span class="n">rank</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">rank</span><span class="p">)</span>
|
|
<span class="k">else</span><span class="p">:</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">model</span> <span class="o">=</span> <span class="n">model_cls</span><span class="o">.</span><span class="n">from_hugging_face</span><span class="p">(</span>
|
|
<span class="nb">str</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">_model_dir</span><span class="p">),</span>
|
|
<span class="n">dtype</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">llm_args</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="bp">self</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="bp">self</span><span class="o">.</span><span class="n">llm_args</span><span class="o">.</span><span class="n">quant_config</span><span class="p">,</span>
|
|
<span class="n">load_model_on_cpu</span><span class="o">=</span>
|
|
<span class="kc">True</span><span class="p">,</span> <span class="c1"># TODO:TRTLLM-195 to enhance the weights loading memory usage and chose best location</span>
|
|
<span class="n">trust_remote_code</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">llm_args</span><span class="o">.</span><span class="n">trust_remote_code</span><span class="p">,</span>
|
|
<span class="n">speculative_model</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">_speculative_model_dir</span><span class="p">,</span>
|
|
<span class="n">medusa_num_heads</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">llm_args</span><span class="o">.</span><span class="n">speculative_config</span><span class="o">.</span>
|
|
<span class="n">num_medusa_heads</span> <span class="k">if</span> <span class="nb">isinstance</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">llm_args</span><span class="o">.</span><span class="n">speculative_config</span><span class="p">,</span>
|
|
<span class="n">MedusaDecodingConfig</span><span class="p">)</span> <span class="k">else</span> <span class="kc">None</span><span class="p">,</span>
|
|
<span class="o">**</span><span class="bp">self</span><span class="o">.</span><span class="n">convert_checkpoint_options</span><span class="p">,</span>
|
|
<span class="p">)</span>
|
|
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">pretrained_config</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">model</span><span class="o">.</span><span class="n">config</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">_model_info</span> <span class="o">=</span> <span class="n">_ModelInfo</span><span class="o">.</span><span class="n">from_pretrained_config</span><span class="p">(</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">pretrained_config</span><span class="p">)</span>
|
|
|
|
<span class="k">def</span> <span class="nf">_load_model_from_ckpt</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
|
|
<span class="w"> </span><span class="sd">''' Load a TRT-LLM model from checkpoint. '''</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">pretrained_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="bp">self</span><span class="o">.</span><span class="n">_model_dir</span><span class="p">,</span> <span class="s1">'config.json'</span><span class="p">))</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">pretrained_config</span><span class="o">.</span><span class="n">mapping</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">mapping</span>
|
|
|
|
<span class="c1">#TODO: TRTLLM-1091, change the architecture in the checkpoint to TRT-LLM one, not HF one.</span>
|
|
<span class="n">architecture</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">pretrained_config</span><span class="o">.</span><span class="n">architecture</span>
|
|
<span class="k">assert</span> <span class="n">architecture</span> <span class="ow">in</span> <span class="n">MODEL_MAP</span><span class="p">,</span> \
|
|
<span class="sa">f</span><span class="s2">"Unsupported model architecture: </span><span class="si">{</span><span class="n">architecture</span><span class="si">}</span><span class="s2">"</span>
|
|
<span class="n">model_cls</span> <span class="o">=</span> <span class="n">MODEL_MAP</span><span class="p">[</span><span class="n">architecture</span><span class="p">]</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">llm_args</span><span class="o">.</span><span class="n">load_format</span> <span class="o">==</span> <span class="s1">'dummy'</span><span class="p">:</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">model</span> <span class="o">=</span> <span class="n">model_cls</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">pretrained_config</span><span class="p">)</span>
|
|
<span class="k">else</span><span class="p">:</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">model</span> <span class="o">=</span> <span class="n">model_cls</span><span class="o">.</span><span class="n">from_checkpoint</span><span class="p">(</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">_model_dir</span><span class="p">,</span> <span class="n">config</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">pretrained_config</span><span class="p">)</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">_model_info</span> <span class="o">=</span> <span class="n">_ModelInfo</span><span class="o">.</span><span class="n">from_pretrained_config</span><span class="p">(</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">pretrained_config</span><span class="p">)</span>
|
|
|
|
<span class="c1"># load parallel embedding related options</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">convert_checkpoint_options</span><span class="p">[</span>
|
|
<span class="s1">'use_parallel_embedding'</span><span class="p">]</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">pretrained_config</span><span class="o">.</span><span class="n">use_parallel_embedding</span>
|
|
|
|
<span class="k">def</span> <span class="nf">_build_engine_from_inmemory_model</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
|
|
<span class="k">assert</span> <span class="nb">isinstance</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">llm_args</span><span class="o">.</span><span class="n">model</span><span class="p">,</span> <span class="n">Module</span><span class="p">)</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">_model_info</span> <span class="o">=</span> <span class="n">_ModelInfo</span><span class="o">.</span><span class="n">from_module</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">model</span><span class="p">)</span>
|
|
|
|
<span class="k">def</span> <span class="nf">_build_engine</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
|
|
<span class="k">assert</span> <span class="nb">isinstance</span><span class="p">(</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">build_config</span><span class="p">,</span>
|
|
<span class="n">BuildConfig</span><span class="p">),</span> <span class="sa">f</span><span class="s2">"build_config is not set yet: </span><span class="si">{</span><span class="bp">self</span><span class="o">.</span><span class="n">build_config</span><span class="si">}</span><span class="s2">"</span>
|
|
|
|
<span class="c1"># avoid the original build_config is modified, avoid the side effect</span>
|
|
<span class="n">copied_build_config</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="n">build_config</span><span class="p">)</span>
|
|
|
|
<span class="n">copied_build_config</span><span class="o">.</span><span class="n">update</span><span class="p">(</span>
|
|
<span class="n">auto_parallel_config</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">auto_parallel_config</span><span class="p">)</span>
|
|
<span class="n">copied_build_config</span><span class="o">.</span><span class="n">update_kv_cache_type</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">_model_info</span><span class="o">.</span><span class="n">architecture</span><span class="p">)</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">auto_parallel_config</span><span class="o">.</span><span class="n">enabled</span><span class="p">:</span>
|
|
<span class="bp">self</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">rank</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">rank</span>
|
|
<span class="k">assert</span> <span class="bp">self</span><span class="o">.</span><span class="n">model</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">,</span> <span class="s2">"model is loaded yet."</span>
|
|
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">_engine</span> <span class="o">=</span> <span class="n">build</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">model</span><span class="p">,</span> <span class="n">copied_build_config</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="bp">self</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="c1"># delete the model explicitly to free all the build-time resources</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">model</span> <span class="o">=</span> <span class="kc">None</span>
|
|
|
|
<span class="k">def</span> <span class="nf">_save_engine_for_runtime</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
|
|
<span class="w"> </span><span class="sd">'''</span>
|
|
<span class="sd"> Persist the engine to disk for the cpp runtime. Currently, the cpp runtime can accept an engine path,</span>
|
|
<span class="sd"> that requires the engine should always be saved to disk.</span>
|
|
|
|
<span class="sd"> This explicit saving will be removed in the future when the cpp runtime can accept the engine buffer directly.</span>
|
|
<span class="sd"> But this is necessary for a build cache, but it can be optimized to async IO.</span>
|
|
<span class="sd"> '''</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">build_cache_enabled</span><span class="p">:</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">_model_dir</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">engine_cache_stage</span><span class="o">.</span><span class="n">cache_dir</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">_model_format</span> <span class="o">=</span> <span class="n">_ModelFormatKind</span><span class="o">.</span><span class="n">TLLM_ENGINE</span>
|
|
<span class="k">return</span>
|
|
|
|
<span class="k">def</span> <span class="nf">_load_engine_buffer</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
|
|
<span class="c1"># Load engine buffer from disk</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">_engine</span> <span class="o">=</span> <span class="n">Engine</span><span class="o">.</span><span class="n">from_dir</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">_model_dir</span><span class="p">)</span>
|
|
|
|
<span class="nd">@staticmethod</span>
|
|
<span class="k">def</span> <span class="nf">load_extra_build_configs_from_engine</span><span class="p">(</span>
|
|
<span class="n">model_dir</span><span class="p">:</span> <span class="nb">str</span><span class="p">)</span> <span class="o">-></span> <span class="n">Optional</span><span class="p">[</span><span class="n">Namespace</span><span class="p">]:</span>
|
|
<span class="w"> </span><span class="sd">''' Load the extra build configs from the engine directory, return None if model isn't an engine. '''</span>
|
|
<span class="k">if</span> <span class="n">ModelLoader</span><span class="o">.</span><span class="n">get_model_format</span><span class="p">(</span>
|
|
<span class="n">model_dir</span><span class="p">)</span> <span class="ow">is</span> <span class="ow">not</span> <span class="n">_ModelFormatKind</span><span class="o">.</span><span class="n">TLLM_ENGINE</span><span class="p">:</span>
|
|
<span class="k">return</span> <span class="kc">None</span>
|
|
|
|
<span class="k">with</span> <span class="nb">open</span><span class="p">(</span><span class="n">Path</span><span class="p">(</span><span class="n">model_dir</span><span class="p">)</span> <span class="o">/</span> <span class="s2">"config.json"</span><span class="p">,</span> <span class="s2">"r"</span><span class="p">)</span> <span class="k">as</span> <span class="n">f</span><span class="p">:</span>
|
|
<span class="n">engine_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">build_config</span> <span class="o">=</span> <span class="n">engine_config</span><span class="p">[</span><span class="s1">'build_config'</span><span class="p">]</span>
|
|
<span class="n">build_config</span><span class="o">.</span><span class="n">pop</span><span class="p">(</span><span class="s2">"plugin_config"</span><span class="p">)</span>
|
|
<span class="k">return</span> <span class="n">Namespace</span><span class="p">(</span><span class="o">**</span><span class="n">build_config</span><span class="p">)</span>
|
|
|
|
<span class="nd">@staticmethod</span>
|
|
<span class="k">def</span> <span class="nf">load_hf_tokenizer</span><span class="p">(</span>
|
|
<span class="n">model_dir</span><span class="p">,</span>
|
|
<span class="n">trust_remote_code</span><span class="p">:</span> <span class="nb">bool</span> <span class="o">=</span> <span class="kc">True</span><span class="p">,</span>
|
|
<span class="n">use_fast</span><span class="p">:</span> <span class="nb">bool</span> <span class="o">=</span> <span class="kc">True</span><span class="p">)</span> <span class="o">-></span> <span class="n">Optional</span><span class="p">[</span><span class="n">TransformersTokenizer</span><span class="p">]:</span>
|
|
<span class="k">try</span><span class="p">:</span>
|
|
|
|
<span class="k">return</span> <span class="n">TransformersTokenizer</span><span class="o">.</span><span class="n">from_pretrained</span><span class="p">(</span>
|
|
<span class="n">model_dir</span><span class="p">,</span>
|
|
<span class="n">legacy</span><span class="o">=</span><span class="kc">False</span><span class="p">,</span>
|
|
<span class="n">padding_side</span><span class="o">=</span><span class="s1">'left'</span><span class="p">,</span>
|
|
<span class="n">truncation_side</span><span class="o">=</span><span class="s1">'left'</span><span class="p">,</span>
|
|
<span class="n">trust_remote_code</span><span class="o">=</span><span class="n">trust_remote_code</span><span class="p">,</span>
|
|
<span class="n">use_fast</span><span class="o">=</span><span class="n">use_fast</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="n">logger</span><span class="o">.</span><span class="n">error</span><span class="p">(</span><span class="sa">f</span><span class="s2">"Failed to load tokenizer from </span><span class="si">{</span><span class="n">model_dir</span><span class="si">}</span><span class="s2">: </span><span class="si">{</span><span class="n">e</span><span class="si">}</span><span class="s2">"</span><span class="p">)</span>
|
|
<span class="k">return</span> <span class="kc">None</span>
|
|
|
|
|
|
<span class="k">class</span> <span class="nc">CachedModelLoader</span><span class="p">:</span>
|
|
<span class="w"> </span><span class="sd">'''</span>
|
|
<span class="sd"> The CacheModelLoader is used to build the model in both single or multi-gpu, with cache might be enabled.</span>
|
|
<span class="sd"> '''</span>
|
|
|
|
<span class="k">def</span> <span class="fm">__init__</span><span class="p">(</span>
|
|
<span class="bp">self</span><span class="p">,</span>
|
|
<span class="n">llm_args</span><span class="p">:</span> <span class="n">LlmArgs</span><span class="p">,</span>
|
|
<span class="n">llm_build_stats</span><span class="p">:</span> <span class="n">weakref</span><span class="o">.</span><span class="n">ReferenceType</span><span class="p">[</span><span class="s2">"LlmBuildStats"</span><span class="p">],</span>
|
|
<span class="n">mpi_session</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="n">MpiSession</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">workspace</span><span class="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="bp">self</span><span class="o">.</span><span class="n">llm_args</span> <span class="o">=</span> <span class="n">llm_args</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">mpi_session</span> <span class="o">=</span> <span class="n">mpi_session</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">_workspace</span> <span class="o">=</span> <span class="n">workspace</span> <span class="ow">or</span> <span class="n">tempfile</span><span class="o">.</span><span class="n">TemporaryDirectory</span><span class="p">()</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">llm_build_stats</span> <span class="o">=</span> <span class="n">llm_build_stats</span>
|
|
|
|
<span class="c1"># This is used for build cache. To compute the cache key, a local HF model is required, it could be download</span>
|
|
<span class="c1"># from HF model hub, so this helps to hold the path.</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">_hf_model_dir</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="n">Path</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</span>
|
|
|
|
<span class="nd">@property</span>
|
|
<span class="k">def</span> <span class="nf">workspace</span><span class="p">(</span><span class="bp">self</span><span class="p">)</span> <span class="o">-></span> <span class="n">Path</span><span class="p">:</span>
|
|
<span class="k">return</span> <span class="n">Path</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">_workspace</span><span class="o">.</span><span class="n">name</span><span class="p">)</span> <span class="k">if</span> <span class="nb">isinstance</span><span class="p">(</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">_workspace</span><span class="p">,</span> <span class="n">tempfile</span><span class="o">.</span><span class="n">TemporaryDirectory</span><span class="p">)</span> <span class="k">else</span> <span class="n">Path</span><span class="p">(</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">_workspace</span><span class="p">)</span>
|
|
|
|
<span class="k">def</span> <span class="fm">__call__</span><span class="p">(</span><span class="bp">self</span><span class="p">)</span> <span class="o">-></span> <span class="n">Tuple</span><span class="p">[</span><span class="n">Path</span><span class="p">,</span> <span class="n">Union</span><span class="p">[</span><span class="n">Path</span><span class="p">,</span> <span class="kc">None</span><span class="p">]]:</span>
|
|
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">llm_args</span><span class="o">.</span><span class="n">model_format</span> <span class="ow">is</span> <span class="n">_ModelFormatKind</span><span class="o">.</span><span class="n">TLLM_ENGINE</span><span class="p">:</span>
|
|
<span class="k">return</span> <span class="bp">self</span><span class="o">.</span><span class="n">llm_args</span><span class="o">.</span><span class="n">model</span><span class="p">,</span> <span class="kc">None</span>
|
|
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">engine_cache_stage</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="n">CachedStage</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</span>
|
|
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">_hf_model_dir</span> <span class="o">=</span> <span class="kc">None</span>
|
|
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">model_loader</span> <span class="o">=</span> <span class="n">ModelLoader</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">llm_args</span><span class="p">)</span>
|
|
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">build_cache_enabled</span><span class="p">:</span>
|
|
<span class="n">print_colored</span><span class="p">(</span><span class="s2">"Build cache is enabled.</span><span class="se">\n</span><span class="s2">"</span><span class="p">,</span> <span class="s1">'yellow'</span><span class="p">)</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">model_loader</span><span class="o">.</span><span class="n">model_obj</span><span class="o">.</span><span class="n">is_hub_model</span><span class="p">:</span>
|
|
<span class="c1"># This will download the config.json from HF model hub, this helps to create a PretrainedConfig for</span>
|
|
<span class="c1"># cache key.</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">_hf_model_dir</span> <span class="o">=</span> <span class="n">download_hf_pretrained_config</span><span class="p">(</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">model_loader</span><span class="o">.</span><span class="n">model_obj</span><span class="o">.</span><span class="n">model_name</span><span class="p">,</span>
|
|
<span class="n">revision</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">llm_args</span><span class="o">.</span><span class="n">revision</span><span class="p">)</span>
|
|
|
|
<span class="k">elif</span> <span class="bp">self</span><span class="o">.</span><span class="n">model_loader</span><span class="o">.</span><span class="n">model_obj</span><span class="o">.</span><span class="n">is_local_model</span><span class="p">:</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">_hf_model_dir</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">model_loader</span><span class="o">.</span><span class="n">model_obj</span><span class="o">.</span><span class="n">model_dir</span> <span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">llm_args</span><span class="o">.</span><span class="n">model_format</span> <span class="ow">is</span> <span class="n">_ModelFormatKind</span><span class="o">.</span><span class="n">HF</span> <span class="k">else</span> <span class="kc">None</span>
|
|
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">engine_cache_stage</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">_get_engine_cache_stage</span><span class="p">()</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">engine_cache_stage</span><span class="o">.</span><span class="n">is_cached</span><span class="p">():</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">llm_build_stats</span><span class="o">.</span><span class="n">cache_hitted</span> <span class="o">=</span> <span class="kc">True</span>
|
|
<span class="n">print_colored</span><span class="p">(</span>
|
|
<span class="sa">f</span><span class="s2">"Reusing cached engine in </span><span class="si">{</span><span class="bp">self</span><span class="o">.</span><span class="n">engine_cache_stage</span><span class="o">.</span><span class="n">get_engine_path</span><span class="p">()</span><span class="si">}</span><span class="se">\n\n</span><span class="s2">"</span><span class="p">,</span>
|
|
<span class="s1">'grey'</span><span class="p">)</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">model_loader</span><span class="o">.</span><span class="n">model_obj</span><span class="o">.</span><span class="n">model_dir</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">engine_cache_stage</span><span class="o">.</span><span class="n">get_engine_path</span><span class="p">(</span>
|
|
<span class="p">)</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">llm_build_stats</span><span class="o">.</span><span class="n">engine_dir</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">model_loader</span><span class="o">.</span><span class="n">model_obj</span><span class="o">.</span><span class="n">model_dir</span>
|
|
<span class="k">return</span> <span class="bp">self</span><span class="o">.</span><span class="n">llm_build_stats</span><span class="o">.</span><span class="n">engine_dir</span><span class="p">,</span> <span class="bp">self</span><span class="o">.</span><span class="n">_hf_model_dir</span>
|
|
<span class="kn">from</span> <span class="nn">tensorrt_llm._torch.pyexecutor.backend_registries.backend_registry</span> <span class="kn">import</span> \
|
|
<span class="n">get_backend_info</span>
|
|
<span class="k">if</span> <span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">llm_args</span><span class="o">.</span><span class="n">backend</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">)</span> <span class="ow">and</span> <span class="n">get_backend_info</span><span class="p">(</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">llm_args</span><span class="o">.</span><span class="n">backend</span><span class="p">,</span> <span class="s1">'need_hf_model'</span><span class="p">):</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">model_loader</span><span class="o">.</span><span class="n">model_obj</span><span class="o">.</span><span class="n">is_hub_model</span><span class="p">:</span>
|
|
<span class="n">hf_folder</span> <span class="o">=</span> <span class="n">download_hf_model</span><span class="p">(</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">model_loader</span><span class="o">.</span><span class="n">model_obj</span><span class="o">.</span><span class="n">model_name</span><span class="p">,</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">llm_args</span><span class="o">.</span><span class="n">revision</span><span class="p">)</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">_hf_model_dir</span> <span class="o">=</span> <span class="n">hf_folder</span>
|
|
<span class="k">else</span><span class="p">:</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">_hf_model_dir</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">model_loader</span><span class="o">.</span><span class="n">model_obj</span><span class="o">.</span><span class="n">model_dir</span>
|
|
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">llm_args</span><span class="o">.</span><span class="n">quant_config</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="n">logger</span><span class="o">.</span><span class="n">warning</span><span class="p">(</span>
|
|
<span class="s2">"QuantConfig for pytorch backend is ignored. You can load"</span>
|
|
<span class="s2">"quantized model with hf_quant_config.json directly."</span><span class="p">)</span>
|
|
<span class="k">return</span> <span class="kc">None</span><span class="p">,</span> <span class="bp">self</span><span class="o">.</span><span class="n">_hf_model_dir</span>
|
|
|
|
<span class="k">return</span> <span class="bp">self</span><span class="o">.</span><span class="n">_build_model</span><span class="p">(),</span> <span class="bp">self</span><span class="o">.</span><span class="n">_hf_model_dir</span>
|
|
|
|
<span class="k">def</span> <span class="nf">get_engine_dir</span><span class="p">(</span><span class="bp">self</span><span class="p">)</span> <span class="o">-></span> <span class="n">Path</span><span class="p">:</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">llm_args</span><span class="o">.</span><span class="n">model_format</span> <span class="ow">is</span> <span class="n">_ModelFormatKind</span><span class="o">.</span><span class="n">TLLM_ENGINE</span><span class="p">:</span>
|
|
<span class="k">return</span> <span class="bp">self</span><span class="o">.</span><span class="n">model_obj</span><span class="o">.</span><span class="n">model_dir</span>
|
|
|
|
<span class="c1"># generate a new path for writing the engine</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">build_cache_enabled</span><span class="p">:</span>
|
|
<span class="n">cache_stage</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">_get_engine_cache_stage</span><span class="p">()</span>
|
|
<span class="k">return</span> <span class="n">cache_stage</span><span class="o">.</span><span class="n">get_engine_path</span><span class="p">()</span>
|
|
|
|
<span class="k">return</span> <span class="bp">self</span><span class="o">.</span><span class="n">workspace</span> <span class="o">/</span> <span class="s2">"tmp.engine"</span>
|
|
|
|
<span class="nd">@property</span>
|
|
<span class="k">def</span> <span class="nf">build_cache_enabled</span><span class="p">(</span><span class="bp">self</span><span class="p">)</span> <span class="o">-></span> <span class="nb">bool</span><span class="p">:</span>
|
|
<span class="n">_enable_build_cache</span><span class="p">,</span> <span class="n">_</span> <span class="o">=</span> <span class="n">get_build_cache_config_from_env</span><span class="p">()</span>
|
|
|
|
<span class="k">return</span> <span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">llm_args</span><span class="o">.</span><span class="n">enable_build_cache</span> <span class="ow">or</span> <span class="n">_enable_build_cache</span><span class="p">)</span> <span class="ow">and</span> <span class="p">(</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">llm_args</span><span class="o">.</span><span class="n">model_format</span> <span class="ow">is</span> <span class="n">_ModelFormatKind</span><span class="o">.</span><span class="n">HF</span>
|
|
<span class="p">)</span> <span class="ow">and</span> <span class="ow">not</span> <span class="bp">self</span><span class="o">.</span><span class="n">llm_args</span><span class="o">.</span><span class="n">parallel_config</span><span class="o">.</span><span class="n">auto_parallel</span>
|
|
|
|
<span class="k">def</span> <span class="nf">_get_engine_cache_stage</span><span class="p">(</span><span class="bp">self</span><span class="p">)</span> <span class="o">-></span> <span class="n">CachedStage</span><span class="p">:</span>
|
|
<span class="w"> </span><span class="sd">''' Get the cache stage for engine building. '''</span>
|
|
<span class="n">build_cache</span> <span class="o">=</span> <span class="n">BuildCache</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">llm_args</span><span class="o">.</span><span class="n">enable_build_cache</span><span class="p">)</span>
|
|
|
|
<span class="k">assert</span> <span class="bp">self</span><span class="o">.</span><span class="n">_hf_model_dir</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">,</span> <span class="s2">"HF model dir is required for cache key."</span>
|
|
|
|
<span class="k">def</span> <span class="nf">serialize</span><span class="p">(</span><span class="n">d</span><span class="p">)</span> <span class="o">-></span> <span class="nb">str</span><span class="p">:</span>
|
|
<span class="n">dic</span> <span class="o">=</span> <span class="n">asdict</span><span class="p">(</span><span class="n">d</span><span class="p">)</span> <span class="k">if</span> <span class="ow">not</span> <span class="nb">isinstance</span><span class="p">(</span>
|
|
<span class="n">d</span><span class="p">,</span> <span class="n">PretrainedConfig</span><span class="p">)</span> <span class="k">else</span> <span class="n">d</span><span class="o">.</span><span class="n">to_dict</span><span class="p">()</span>
|
|
<span class="k">return</span> <span class="n">json</span><span class="o">.</span><span class="n">dumps</span><span class="p">(</span><span class="n">dic</span><span class="p">,</span> <span class="n">sort_keys</span><span class="o">=</span><span class="kc">True</span><span class="p">)</span>
|
|
|
|
<span class="n">parallel_config</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">llm_args</span><span class="o">.</span><span class="n">parallel_config</span>
|
|
|
|
<span class="n">force_rebuild</span> <span class="o">=</span> <span class="kc">False</span>
|
|
<span class="k">if</span> <span class="n">parallel_config</span><span class="o">.</span><span class="n">auto_parallel</span><span class="p">:</span>
|
|
<span class="n">force_rebuild</span> <span class="o">=</span> <span class="kc">True</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">llm_args</span><span class="o">.</span><span class="n">model_format</span> <span class="ow">is</span> <span class="ow">not</span> <span class="n">_ModelFormatKind</span><span class="o">.</span><span class="n">HF</span><span class="p">:</span>
|
|
<span class="n">force_rebuild</span> <span class="o">=</span> <span class="kc">True</span>
|
|
|
|
<span class="k">return</span> <span class="n">build_cache</span><span class="o">.</span><span class="n">get_engine_building_cache_stage</span><span class="p">(</span>
|
|
<span class="n">build_config</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">llm_args</span><span class="o">.</span><span class="n">build_config</span><span class="p">,</span>
|
|
<span class="n">model_path</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">_hf_model_dir</span><span class="p">,</span>
|
|
<span class="n">force_rebuild</span><span class="o">=</span><span class="n">force_rebuild</span><span class="p">,</span>
|
|
<span class="c1"># Other configs affecting the engine building</span>
|
|
<span class="n">parallel_config</span><span class="o">=</span><span class="n">serialize</span><span class="p">(</span><span class="n">parallel_config</span><span class="p">),</span>
|
|
<span class="n">pretrained_config</span><span class="o">=</span><span class="n">serialize</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">get_pretrained_config</span><span class="p">()),</span>
|
|
<span class="n">quant_config</span><span class="o">=</span><span class="n">serialize</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">llm_args</span><span class="o">.</span><span class="n">quant_config</span><span class="p">),</span>
|
|
<span class="p">)</span>
|
|
|
|
<span class="k">def</span> <span class="nf">get_pretrained_config</span><span class="p">(</span><span class="bp">self</span><span class="p">)</span> <span class="o">-></span> <span class="n">PretrainedConfig</span><span class="p">:</span>
|
|
<span class="w"> </span><span class="sd">''' Get the PretrainedConfig for cache key.</span>
|
|
<span class="sd"> NOTE, this is not the HF model's config, but the TRT-LLM's config. We use this as a generic information for</span>
|
|
<span class="sd"> HF and other models. '''</span>
|
|
<span class="k">assert</span> <span class="bp">self</span><span class="o">.</span><span class="n">_hf_model_dir</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span>
|
|
<span class="k">return</span> <span class="n">AutoConfig</span><span class="o">.</span><span class="n">from_hugging_face</span><span class="p">(</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">_hf_model_dir</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">llm_args</span><span class="o">.</span><span class="n">parallel_config</span><span class="o">.</span><span class="n">to_mapping</span><span class="p">(),</span>
|
|
<span class="n">quant_config</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">llm_args</span><span class="o">.</span><span class="n">quant_config</span><span class="p">,</span>
|
|
<span class="n">dtype</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">llm_args</span><span class="o">.</span><span class="n">dtype</span><span class="p">)</span>
|
|
|
|
<span class="k">def</span> <span class="nf">_build_model</span><span class="p">(</span><span class="bp">self</span><span class="p">)</span> <span class="o">-></span> <span class="n">Path</span><span class="p">:</span>
|
|
<span class="n">model_format</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">llm_args</span><span class="o">.</span><span class="n">model_format</span>
|
|
|
|
<span class="k">def</span> <span class="nf">build_task</span><span class="p">(</span><span class="n">engine_dir</span><span class="p">:</span> <span class="n">Path</span><span class="p">):</span>
|
|
<span class="k">if</span> <span class="n">model_format</span> <span class="ow">is</span> <span class="ow">not</span> <span class="n">_ModelFormatKind</span><span class="o">.</span><span class="n">TLLM_ENGINE</span><span class="p">:</span>
|
|
<span class="n">model_loader_kwargs</span> <span class="o">=</span> <span class="p">{</span>
|
|
<span class="s1">'llm_args'</span><span class="p">:</span> <span class="bp">self</span><span class="o">.</span><span class="n">llm_args</span><span class="p">,</span>
|
|
<span class="s1">'workspace'</span><span class="p">:</span> <span class="nb">str</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">workspace</span><span class="p">),</span>
|
|
<span class="s1">'llm_build_stats'</span><span class="p">:</span> <span class="bp">self</span><span class="o">.</span><span class="n">llm_build_stats</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">llm_args</span><span class="o">.</span><span class="n">parallel_config</span><span class="o">.</span><span class="n">is_multi_gpu</span><span class="p">:</span>
|
|
<span class="k">assert</span> <span class="bp">self</span><span class="o">.</span><span class="n">mpi_session</span>
|
|
<span class="c1"># The engine_dir:Path will be stored to MPINodeState.state</span>
|
|
<span class="n">build_infos</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">mpi_session</span><span class="o">.</span><span class="n">submit_sync</span><span class="p">(</span>
|
|
<span class="n">CachedModelLoader</span><span class="o">.</span><span class="n">_node_build_task</span><span class="p">,</span>
|
|
<span class="n">engine_dir</span><span class="o">=</span><span class="n">engine_dir</span><span class="p">,</span>
|
|
<span class="o">**</span><span class="n">model_loader_kwargs</span><span class="p">)</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">llm_build_stats</span><span class="o">.</span><span class="n">build_steps_info</span> <span class="o">=</span> <span class="n">build_infos</span><span class="p">[</span><span class="mi">0</span><span class="p">]</span>
|
|
|
|
<span class="k">else</span><span class="p">:</span> <span class="c1"># single-gpu</span>
|
|
<span class="k">with</span> <span class="n">ModelLoader</span><span class="p">(</span><span class="o">**</span><span class="n">model_loader_kwargs</span><span class="p">)</span> <span class="k">as</span> <span class="n">model_loader</span><span class="p">:</span>
|
|
<span class="n">model_loader</span><span class="p">(</span><span class="n">engine_dir</span><span class="o">=</span><span class="n">engine_dir</span><span class="p">)</span>
|
|
|
|
<span class="n">release_gc</span><span class="p">()</span>
|
|
|
|
<span class="n">has_storage</span> <span class="o">=</span> <span class="kc">True</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">build_cache_enabled</span><span class="p">:</span>
|
|
<span class="k">try</span><span class="p">:</span>
|
|
<span class="c1"># TODO[chunweiy]: Cover the case when the model is from HF model hub.</span>
|
|
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">model_loader</span><span class="o">.</span><span class="n">model_obj</span><span class="o">.</span><span class="n">is_local_model</span><span class="p">:</span>
|
|
<span class="c1"># This is not perfect, but will make build-cache much more robust.</span>
|
|
<span class="n">free_storage</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">engine_cache_stage</span><span class="o">.</span><span class="n">parent</span><span class="o">.</span><span class="n">free_storage_in_gb</span><span class="p">(</span>
|
|
<span class="p">)</span>
|
|
<span class="n">model_size</span> <span class="o">=</span> <span class="n">get_directory_size_in_gb</span><span class="p">(</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">model_loader</span><span class="o">.</span><span class="n">model_obj</span><span class="o">.</span><span class="n">model_dir</span><span class="p">)</span>
|
|
<span class="n">require_size</span> <span class="o">=</span> <span class="n">model_size</span> <span class="o">*</span> <span class="mf">1.3</span>
|
|
<span class="n">has_storage</span> <span class="o">=</span> <span class="n">free_storage</span> <span class="o">>=</span> <span class="n">require_size</span>
|
|
|
|
<span class="k">if</span> <span class="ow">not</span> <span class="n">has_storage</span><span class="p">:</span>
|
|
<span class="n">print_colored</span><span class="p">(</span>
|
|
<span class="sa">f</span><span class="s2">"Build cache is disabled since the cache storage is too small.</span><span class="se">\n</span><span class="s2"> "</span><span class="p">,</span>
|
|
<span class="s1">'yellow'</span><span class="p">)</span>
|
|
<span class="n">print_colored</span><span class="p">(</span>
|
|
<span class="sa">f</span><span class="s2">"Free storage: </span><span class="si">{</span><span class="n">free_storage</span><span class="si">}</span><span class="s2">GB, Required storage: </span><span class="si">{</span><span class="n">require_size</span><span class="si">}</span><span class="s2">GB</span><span class="se">\n</span><span class="s2">"</span><span class="p">,</span>
|
|
<span class="s1">'grey'</span><span class="p">)</span>
|
|
<span class="k">except</span> <span class="ne">ValueError</span><span class="p">:</span>
|
|
<span class="n">has_storage</span> <span class="o">=</span> <span class="kc">False</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="n">logger</span><span class="o">.</span><span class="n">error</span><span class="p">(</span><span class="n">e</span><span class="p">)</span>
|
|
<span class="n">has_storage</span> <span class="o">=</span> <span class="kc">False</span>
|
|
|
|
<span class="k">if</span> <span class="n">enable_llm_debug</span><span class="p">():</span>
|
|
<span class="n">print_colored</span><span class="p">(</span><span class="sa">f</span><span class="s2">"Has cache storage: </span><span class="si">{</span><span class="n">has_storage</span><span class="si">}</span><span class="se">\n</span><span class="s2">"</span><span class="p">,</span> <span class="s1">'yellow'</span><span class="p">)</span>
|
|
|
|
<span class="k">if</span> <span class="n">has_storage</span><span class="p">:</span>
|
|
<span class="k">with</span> <span class="bp">self</span><span class="o">.</span><span class="n">engine_cache_stage</span><span class="o">.</span><span class="n">write_guard</span><span class="p">()</span> <span class="k">as</span> <span class="n">engine_dir</span><span class="p">:</span>
|
|
<span class="n">build_task</span><span class="p">(</span><span class="n">engine_dir</span><span class="p">)</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">llm_build_stats</span><span class="o">.</span><span class="n">cache_hitted</span> <span class="o">=</span> <span class="kc">True</span>
|
|
|
|
<span class="k">else</span><span class="p">:</span>
|
|
<span class="n">print_colored</span><span class="p">(</span>
|
|
<span class="s2">"The cache directory is too small, build-cache is disabled.</span><span class="se">\n</span><span class="s2">"</span><span class="p">,</span>
|
|
<span class="s1">'grey'</span><span class="p">)</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">llm_build_stats</span><span class="o">.</span><span class="n">cache_hitted</span> <span class="o">=</span> <span class="kc">False</span>
|
|
<span class="bp">self</span><span class="o">.</span><span class="n">llm_build_stats</span><span class="o">.</span><span class="n">cache_info</span> <span class="o">=</span> <span class="s2">"The cache root directory is too small."</span>
|
|
|
|
<span class="k">if</span> <span class="ow">not</span> <span class="p">(</span><span class="n">has_storage</span> <span class="ow">and</span> <span class="bp">self</span><span class="o">.</span><span class="n">build_cache_enabled</span><span class="p">):</span>
|
|
<span class="n">build_task</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">get_engine_dir</span><span class="p">())</span>
|
|
|
|
<span class="k">return</span> <span class="bp">self</span><span class="o">.</span><span class="n">get_engine_dir</span><span class="p">()</span>
|
|
|
|
<span class="nd">@print_traceback_on_error</span>
|
|
<span class="nd">@staticmethod</span>
|
|
<span class="k">def</span> <span class="nf">_node_build_task</span><span class="p">(</span>
|
|
<span class="n">llm_args</span><span class="p">:</span> <span class="n">LlmArgs</span><span class="p">,</span>
|
|
<span class="n">workspace</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="nb">str</span> <span class="o">|</span> <span class="n">tempfile</span><span class="o">.</span><span class="n">TemporaryDirectory</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">llm_build_stats</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="s1">'LlmBuildStats'</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</span><span class="p">,</span>
|
|
<span class="n">engine_dir</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="n">Path</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</span><span class="p">,</span>
|
|
<span class="p">):</span>
|
|
<span class="k">if</span> <span class="n">MPINodeState</span><span class="o">.</span><span class="n">is_initialized</span><span class="p">():</span>
|
|
<span class="k">raise</span> <span class="ne">RuntimeError</span><span class="p">(</span><span class="s2">"The MPI node is already initialized."</span><span class="p">)</span>
|
|
|
|
<span class="k">with</span> <span class="n">ModelLoader</span><span class="p">(</span><span class="n">llm_args</span><span class="p">,</span>
|
|
<span class="n">workspace</span><span class="o">=</span><span class="n">workspace</span><span class="p">,</span>
|
|
<span class="n">llm_build_stats</span><span class="o">=</span><span class="n">llm_build_stats</span><span class="p">)</span> <span class="k">as</span> <span class="n">model_loader</span><span class="p">:</span>
|
|
<span class="n">model_loader</span><span class="p">(</span><span class="n">engine_dir</span><span class="o">=</span><span class="n">engine_dir</span><span class="p">)</span>
|
|
<span class="k">return</span> <span class="n">model_loader</span><span class="o">.</span><span class="n">llm_build_stats</span><span class="o">.</span><span class="n">build_steps_info</span>
|
|
|
|
<span class="k">def</span> <span class="nf">save</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">engine_dir</span><span class="p">:</span> <span class="n">Path</span><span class="p">):</span>
|
|
<span class="c1"># copy the engine directory to the target directory</span>
|
|
<span class="n">shutil</span><span class="o">.</span><span class="n">copytree</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">get_engine_dir</span><span class="p">(),</span> <span class="n">engine_dir</span><span class="p">)</span>
|
|
|
|
|
|
<span class="nd">@dataclass</span>
|
|
<span class="k">class</span> <span class="nc">LlmBuildStats</span><span class="p">:</span>
|
|
<span class="w"> </span><span class="sd">''' LlmBuildStats is the statistics for the LLM model building. '''</span>
|
|
<span class="c1"># Whether the cache is hit for the engine</span>
|
|
<span class="n">cache_hitted</span><span class="p">:</span> <span class="nb">bool</span> <span class="o">=</span> <span class="kc">False</span>
|
|
<span class="n">cache_info</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="n">model_from_hf_hub</span><span class="p">:</span> <span class="nb">bool</span> <span class="o">=</span> <span class="kc">False</span>
|
|
|
|
<span class="n">local_model_dir</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="n">Path</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</span>
|
|
|
|
<span class="c1"># The path to the trt-llm engine</span>
|
|
<span class="n">engine_dir</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="n">Path</span><span class="p">]</span> <span class="o">=</span> <span class="kc">None</span>
|
|
|
|
<span class="c1"># The build steps information, including the step name and the latency in seconds.</span>
|
|
<span class="n">build_steps_info</span><span class="p">:</span> <span class="n">List</span><span class="p">[</span><span class="n">Tuple</span><span class="p">[</span><span class="nb">str</span><span class="p">,</span> <span class="nb">float</span><span class="p">]]</span> <span class="o">=</span> <span class="n">field</span><span class="p">(</span><span class="n">default_factory</span><span class="o">=</span><span class="nb">list</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 © 2024, NVidia.
|
|
<br/>
|
|
|
|
</p>
|
|
</div>
|
|
|
|
</div>
|
|
|
|
|
|
|
|
</div>
|
|
|
|
</footer>
|
|
</body>
|
|
</html> |