Skip to content

Commit

Permalink
add badge & formatting
Browse files Browse the repository at this point in the history
  • Loading branch information
snyhlxde1 committed Mar 14, 2024
1 parent 5e7a30f commit 1e0345c
Show file tree
Hide file tree
Showing 27 changed files with 94 additions and 68 deletions.
Binary file modified .DS_Store
Binary file not shown.
Binary file modified content/blogs/cllm/.DS_Store
Binary file not shown.
Binary file modified content/blogs/cllm/img/.DS_Store
Binary file not shown.
18 changes: 11 additions & 7 deletions content/blogs/cllm/index.md
Original file line number Diff line number Diff line change
Expand Up @@ -3,27 +3,31 @@ title = "Consistency Large Language Models: A Family of Efficient Parallel Decod
date = 2024-02-21T12:00:00-08:00
authors = ["Siqi Kou*", "Lanxiang Hu*", "Zhezhi He", "Zhijie Deng", "Hao Zhang"]
author = "Siqi Kou*, Lanxiang Hu*, Zhezhi He, Zhijie Deng, Hao Zhang"
draft = false
ShowReadingTime = true
[cover]
image = "img/objective_illustration_global.jpg"
alt = "jacobi trajectory"
caption = "An instance of Jacobi trajectory and an illustration of the global consistency loss learning objective."
draft = false
[socialIcons]
[[socialIcons.icon]]
name = "twitter"
url = "https://twitter.com"
[[socialIcons.icon]]
github = "github"
url = "https://github.com"
name = "github"
url = "https://github.com/hao-ai-lab/Consistency_LLM"
[cover]
image = "img/objective_illustration_global.jpg"
alt = "jacobi trajectory"
caption = "An instance of Jacobi trajectory and an illustration of the global consistency loss learning objective."
+++

{{< socialBadges twitter="Hao-AI-Lab" arxiv-index="2403.00835" github="hao-ai-lab/Consistency_LLM" >}}

{{< justify >}}

**TL;DR:** LLMs have been traditionally regarded as sequential decoders, decoding one token after another. In this blog, we show LLMs can be taught to operate as efficient parallels decoders. We introduce Consistency Large Language Models (CLLMs), a new family of paralel decoders capable of reducing inference latency by efficiently decoding a $n$-token sequence in parallel. Our research shows this process can be effectively learned by pre-trained LLMs, mimicing human cognitive process of forming complete sentences in mind, before articulating word by word. CLLMs are trained with the objective of performing parallel decoding fast and reliably. Specifically, the training objective intends to empower CLLMs with the capability of efficient parallel decoding by mapping any randomly initialized $n$-token sequence to the same result as auto-regressive (AR) decoding in as few steps as possible. This parallel decoding method we employ is called [Jacobi decoding](https://arxiv.org/abs/2305.10427). Experiment results show CLLMs obtained using our proposed method are highly effective, showing $2.4\times$ to $3.4\times$ improvements in generation speed while preserving generation quality in comparison with the pre-trained baselines. In comparison with other SOTA techniques like Medusa2 and Eagle, CLLMs achieve comparable or even better speedup in some tasks with no extra cost and engineering efforts for adaptation, as they require no modifications to the existing model architecture or auxiliary model components.
{{< /justify >}}

{{< image src="img/baseline_vs_cllm_gsm8k_best_acc_demo.gif" alt="cllm-gsm8k-acc-demo" width="120%" title="Figure 1: Demo of speedup by CLLM-ABEL-7B-001 in comparison with baseline [ABEL-7B-001](https://github.com/GAIR-NLP/abel) using Jacobi decoding on GSM8K.">}}


## Background: Jacobi Decoding

{{< justify >}}
Expand Down
4 changes: 2 additions & 2 deletions content/home.md
Original file line number Diff line number Diff line change
Expand Up @@ -5,8 +5,8 @@ url: "/home/"
summary: "home page for Hao Lab @ UCSD"
cover:
image: img/ucsd-geisel.jpg
alt: "Hao Lab @ UCSD"
caption: "Hao Lab @ UCSD"
alt: "Hao AI Lab @ UCSD"
caption: "Hao AI Lab @ UCSD"
---

### Mission Statement
Expand Down
2 changes: 1 addition & 1 deletion hugo.yaml
Original file line number Diff line number Diff line change
@@ -1,6 +1,6 @@
bseURL: 'http://hao-ai-lab.github.io'
languageCode: en-us
title: 'Hao Lab @ UCSD'
title: 'Hao AI Lab @ UCSD'
theme: 'PaperMod'

markup:
Expand Down
11 changes: 11 additions & 0 deletions layouts/shortcodes/socialBadges.html
Original file line number Diff line number Diff line change
@@ -0,0 +1,11 @@
<div style="display: flex !important; align-items: center !important; gap: 10px !important;">
<a href="https://twitter.com/{{ .Get "twitter" }}" target="_blank" rel="noopener noreferrer">
<img src="https://img.shields.io/twitter/url?label=Follow%20%40{{ .Get "twitter" }}&style=social&url=https%3A%2F%2Ftwitter.com%2F{{ .Get "twitter" }}">
</a>
<a href="https://arxiv.org/abs/{{ .Get "arxiv-index" }}" target="_blank" rel="noopener noreferrer">
<img src="https://img.shields.io/badge/arXiv-{{ .Get "arxiv-index" }}-white.svg?style=social">
</a>
<a href="https://github.com/{{ .Get "github" }}" target="_blank" rel="noopener noreferrer">
<img src="https://img.shields.io/github/stars/{{ .Get "github" }}?style=social">
</a>
</div>
6 changes: 3 additions & 3 deletions public/404.html
Original file line number Diff line number Diff line change
Expand Up @@ -7,7 +7,7 @@
<meta http-equiv="X-UA-Compatible" content="IE=edge">
<meta name="viewport" content="width=device-width, initial-scale=1, shrink-to-fit=no">
<meta name="robots" content="noindex, nofollow">
<title>404 Page not found | Hao Lab @ UCSD</title>
<title>404 Page not found | Hao AI Lab @ UCSD</title>
<meta name="keywords" content="">
<meta name="description" content="">
<meta name="author" content="">
Expand Down Expand Up @@ -87,7 +87,7 @@
<header class="header">
<nav class="nav">
<div class="logo">
<a href="//localhost:1313/" accesskey="h" title="Hao Lab @ UCSD (Alt + H)">Hao Lab @ UCSD</a>
<a href="//localhost:1313/" accesskey="h" title="Hao AI Lab @ UCSD (Alt + H)">Hao AI Lab @ UCSD</a>
<div class="logo-switches">
<button id="theme-toggle" accesskey="t" title="(Alt + T)">
<svg id="moon" xmlns="http://www.w3.org/2000/svg" width="24" height="18" viewBox="0 0 24 24"
Expand Down Expand Up @@ -158,7 +158,7 @@
</main>

<footer class="footer">
<span>&copy; 2024 <a href="//localhost:1313/">Hao Lab @ UCSD</a></span>
<span>&copy; 2024 <a href="//localhost:1313/">Hao AI Lab @ UCSD</a></span>
<span>
Powered by
<a href="https://gohugo.io/" rel="noopener noreferrer" target="_blank">Hugo</a> &
Expand Down
20 changes: 16 additions & 4 deletions public/blogs/cllm/index.html
Original file line number Diff line number Diff line change
Expand Up @@ -7,7 +7,7 @@
<meta http-equiv="X-UA-Compatible" content="IE=edge">
<meta name="viewport" content="width=device-width, initial-scale=1, shrink-to-fit=no">
<meta name="robots" content="noindex, nofollow">
<title>Consistency Large Language Models: A Family of Efficient Parallel Decoders | Hao Lab @ UCSD</title>
<title>Consistency Large Language Models: A Family of Efficient Parallel Decoders | Hao AI Lab @ UCSD</title>
<meta name="keywords" content="">
<meta name="description" content="TL;DR: LLMs have been traditionally regarded as sequential decoders, decoding one token after another. In this blog, we show LLMs can be taught to operate as efficient parallels decoders. We introduce Consistency Large Language Models (CLLMs), a new family of paralel decoders capable of reducing inference latency by efficiently decoding a $n$-token sequence in parallel. Our research shows this process can be effectively learned by pre-trained LLMs, mimicing human cognitive process of forming complete sentences in mind, before articulating word by word.">
<meta name="author" content="Siqi Kou*, Lanxiang Hu*, Zhezhi He, Zhijie Deng, Hao Zhang">
Expand Down Expand Up @@ -87,7 +87,7 @@
<header class="header">
<nav class="nav">
<div class="logo">
<a href="//localhost:1313/" accesskey="h" title="Hao Lab @ UCSD (Alt + H)">Hao Lab @ UCSD</a>
<a href="//localhost:1313/" accesskey="h" title="Hao AI Lab @ UCSD (Alt + H)">Hao AI Lab @ UCSD</a>
<div class="logo-switches">
<button id="theme-toggle" accesskey="t" title="(Alt + T)">
<svg id="moon" xmlns="http://www.w3.org/2000/svg" width="24" height="18" viewBox="0 0 24 24"
Expand Down Expand Up @@ -168,7 +168,19 @@ <h1 class="post-title entry-hint-parent">
<figure class="entry-cover"><img loading="eager" src="//localhost:1313/img/objective_illustration_global.jpg" alt="jacobi trajectory">
<p>An instance of Jacobi trajectory and an illustration of the global consistency loss learning objective.</p>
</figure>
<div class="post-content"><div style="text-align: justify;">
<div class="post-content"><div style="display: flex !important; align-items: center !important; gap: 10px !important;">
<a href="https://twitter.com/Hao-AI-Lab" target="_blank" rel="noopener noreferrer">
<img src="https://img.shields.io/twitter/url?label=Follow%20%40Hao-AI-Lab&style=social&url=https%3A%2F%2Ftwitter.com%2FHao-AI-Lab">
</a>
<a href="https://arxiv.org/abs/2403.00835" target="_blank" rel="noopener noreferrer">
<img src="https://img.shields.io/badge/arXiv-2403.00835-white.svg?style=social">
</a>
<a href="https://github.com/hao-ai-lab/Consistency_LLM" target="_blank" rel="noopener noreferrer">
<img src="https://img.shields.io/github/stars/hao-ai-lab/Consistency_LLM?style=social">
</a>
</div>

<div style="text-align: justify;">
<strong>TL;DR:</strong> LLMs have been traditionally regarded as sequential decoders, decoding one token after another. In this blog, we show LLMs can be taught to operate as efficient parallels decoders. We introduce Consistency Large Language Models (CLLMs), a new family of paralel decoders capable of reducing inference latency by efficiently decoding a $n$-token sequence in parallel. Our research shows this process can be effectively learned by pre-trained LLMs, mimicing human cognitive process of forming complete sentences in mind, before articulating word by word. CLLMs are trained with the objective of performing parallel decoding fast and reliably. Specifically, the training objective intends to empower CLLMs with the capability of efficient parallel decoding by mapping any randomly initialized $n$-token sequence to the same result as auto-regressive (AR) decoding in as few steps as possible. This parallel decoding method we employ is called <a href="https://arxiv.org/abs/2305.10427">Jacobi decoding</a>. Experiment results show CLLMs obtained using our proposed method are highly effective, showing $2.4\times$ to $3.4\times$ improvements in generation speed while preserving generation quality in comparison with the pre-trained baselines. In comparison with other SOTA techniques like Medusa2 and Eagle, CLLMs achieve comparable or even better speedup in some tasks with no extra cost and engineering efforts for adaptation, as they require no modifications to the existing model architecture or auxiliary model components.
</div>

Expand Down Expand Up @@ -474,7 +486,7 @@ <h2 id="citation">Citation<a hidden class="anchor" aria-hidden="true" href="#cit
</main>

<footer class="footer">
<span>&copy; 2024 <a href="//localhost:1313/">Hao Lab @ UCSD</a></span>
<span>&copy; 2024 <a href="//localhost:1313/">Hao AI Lab @ UCSD</a></span>
<span>
Powered by
<a href="https://gohugo.io/" rel="noopener noreferrer" target="_blank">Hugo</a> &
Expand Down
8 changes: 4 additions & 4 deletions public/blogs/index.html
Original file line number Diff line number Diff line change
Expand Up @@ -7,9 +7,9 @@
<meta http-equiv="X-UA-Compatible" content="IE=edge">
<meta name="viewport" content="width=device-width, initial-scale=1, shrink-to-fit=no">
<meta name="robots" content="noindex, nofollow">
<title>Blogs | Hao Lab @ UCSD</title>
<title>Blogs | Hao AI Lab @ UCSD</title>
<meta name="keywords" content="">
<meta name="description" content="Blogs - Hao Lab @ UCSD">
<meta name="description" content="Blogs - Hao AI Lab @ UCSD">
<meta name="author" content="">
<link rel="canonical" href="//localhost:1313/blogs/">
<link crossorigin="anonymous" href="/assets/css/stylesheet.b609c58d5c11bb90b1a54e04005d74ad1ddf22165eb79f5533967e57df9c3b50.css" integrity="sha256-tgnFjVwRu5CxpU4EAF10rR3fIhZet59VM5Z&#43;V9&#43;cO1A=" rel="preload stylesheet" as="style">
Expand Down Expand Up @@ -88,7 +88,7 @@
<header class="header">
<nav class="nav">
<div class="logo">
<a href="//localhost:1313/" accesskey="h" title="Hao Lab @ UCSD (Alt + H)">Hao Lab @ UCSD</a>
<a href="//localhost:1313/" accesskey="h" title="Hao AI Lab @ UCSD (Alt + H)">Hao AI Lab @ UCSD</a>
<div class="logo-switches">
<button id="theme-toggle" accesskey="t" title="(Alt + T)">
<svg id="moon" xmlns="http://www.w3.org/2000/svg" width="24" height="18" viewBox="0 0 24 24"
Expand Down Expand Up @@ -165,7 +165,7 @@ <h2><a href="//localhost:1313/blogs/cllm/">Consistency Large Language Models: A
</main>

<footer class="footer">
<span>&copy; 2024 <a href="//localhost:1313/">Hao Lab @ UCSD</a></span>
<span>&copy; 2024 <a href="//localhost:1313/">Hao AI Lab @ UCSD</a></span>
<span>
Powered by
<a href="https://gohugo.io/" rel="noopener noreferrer" target="_blank">Hugo</a> &
Expand Down
4 changes: 2 additions & 2 deletions public/blogs/index.xml
Original file line number Diff line number Diff line change
@@ -1,9 +1,9 @@
<?xml version="1.0" encoding="utf-8" standalone="yes"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:content="http://purl.org/rss/1.0/modules/content/">
<channel>
<title>Blogs on Hao Lab @ UCSD</title>
<title>Blogs on Hao AI Lab @ UCSD</title>
<link>//localhost:1313/blogs/</link>
<description>Recent content in Blogs on Hao Lab @ UCSD</description>
<description>Recent content in Blogs on Hao AI Lab @ UCSD</description>
<generator>Hugo -- gohugo.io</generator>
<language>en-us</language>
<lastBuildDate>Wed, 21 Feb 2024 12:00:00 -0800</lastBuildDate>
Expand Down
6 changes: 3 additions & 3 deletions public/categories/index.html
Original file line number Diff line number Diff line change
Expand Up @@ -7,7 +7,7 @@
<meta http-equiv="X-UA-Compatible" content="IE=edge">
<meta name="viewport" content="width=device-width, initial-scale=1, shrink-to-fit=no">
<meta name="robots" content="noindex, nofollow">
<title>Categories | Hao Lab @ UCSD</title>
<title>Categories | Hao AI Lab @ UCSD</title>
<meta name="keywords" content="">
<meta name="description" content="">
<meta name="author" content="">
Expand Down Expand Up @@ -88,7 +88,7 @@
<header class="header">
<nav class="nav">
<div class="logo">
<a href="//localhost:1313/" accesskey="h" title="Hao Lab @ UCSD (Alt + H)">Hao Lab @ UCSD</a>
<a href="//localhost:1313/" accesskey="h" title="Hao AI Lab @ UCSD (Alt + H)">Hao AI Lab @ UCSD</a>
<div class="logo-switches">
<button id="theme-toggle" accesskey="t" title="(Alt + T)">
<svg id="moon" xmlns="http://www.w3.org/2000/svg" width="24" height="18" viewBox="0 0 24 24"
Expand Down Expand Up @@ -164,7 +164,7 @@ <h1>Categories</h1>
</main>

<footer class="footer">
<span>&copy; 2024 <a href="//localhost:1313/">Hao Lab @ UCSD</a></span>
<span>&copy; 2024 <a href="//localhost:1313/">Hao AI Lab @ UCSD</a></span>
<span>
Powered by
<a href="https://gohugo.io/" rel="noopener noreferrer" target="_blank">Hugo</a> &
Expand Down
4 changes: 2 additions & 2 deletions public/categories/index.xml
Original file line number Diff line number Diff line change
@@ -1,9 +1,9 @@
<?xml version="1.0" encoding="utf-8" standalone="yes"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:content="http://purl.org/rss/1.0/modules/content/">
<channel>
<title>Categories on Hao Lab @ UCSD</title>
<title>Categories on Hao AI Lab @ UCSD</title>
<link>//localhost:1313/categories/</link>
<description>Recent content in Categories on Hao Lab @ UCSD</description>
<description>Recent content in Categories on Hao AI Lab @ UCSD</description>
<generator>Hugo -- gohugo.io</generator>
<language>en-us</language>
<atom:link href="//localhost:1313/categories/index.xml" rel="self" type="application/rss+xml" />
Expand Down
6 changes: 3 additions & 3 deletions public/contact/index.html
Original file line number Diff line number Diff line change
Expand Up @@ -7,7 +7,7 @@
<meta http-equiv="X-UA-Compatible" content="IE=edge">
<meta name="viewport" content="width=device-width, initial-scale=1, shrink-to-fit=no">
<meta name="robots" content="noindex, nofollow">
<title>Hao Lab @ UCSD</title>
<title>Hao AI Lab @ UCSD</title>
<meta name="keywords" content="">
<meta name="description" content="contact">
<meta name="author" content="">
Expand Down Expand Up @@ -87,7 +87,7 @@
<header class="header">
<nav class="nav">
<div class="logo">
<a href="//localhost:1313/" accesskey="h" title="Hao Lab @ UCSD (Alt + H)">Hao Lab @ UCSD</a>
<a href="//localhost:1313/" accesskey="h" title="Hao AI Lab @ UCSD (Alt + H)">Hao AI Lab @ UCSD</a>
<div class="logo-switches">
<button id="theme-toggle" accesskey="t" title="(Alt + T)">
<svg id="moon" xmlns="http://www.w3.org/2000/svg" width="24" height="18" viewBox="0 0 24 24"
Expand Down Expand Up @@ -184,7 +184,7 @@ <h1 class="post-title entry-hint-parent">
</main>

<footer class="footer">
<span>&copy; 2024 <a href="//localhost:1313/">Hao Lab @ UCSD</a></span>
<span>&copy; 2024 <a href="//localhost:1313/">Hao AI Lab @ UCSD</a></span>
<span>
Powered by
<a href="https://gohugo.io/" rel="noopener noreferrer" target="_blank">Hugo</a> &
Expand Down
10 changes: 5 additions & 5 deletions public/home/index.html
Original file line number Diff line number Diff line change
Expand Up @@ -7,7 +7,7 @@
<meta http-equiv="X-UA-Compatible" content="IE=edge">
<meta name="viewport" content="width=device-width, initial-scale=1, shrink-to-fit=no">
<meta name="robots" content="noindex, nofollow">
<title>Hao Lab @ UCSD</title>
<title>Hao AI Lab @ UCSD</title>
<meta name="keywords" content="">
<meta name="description" content="home page for Hao Lab @ UCSD">
<meta name="author" content="">
Expand Down Expand Up @@ -87,7 +87,7 @@
<header class="header">
<nav class="nav">
<div class="logo">
<a href="//localhost:1313/" accesskey="h" title="Hao Lab @ UCSD (Alt + H)">Hao Lab @ UCSD</a>
<a href="//localhost:1313/" accesskey="h" title="Hao AI Lab @ UCSD (Alt + H)">Hao AI Lab @ UCSD</a>
<div class="logo-switches">
<button id="theme-toggle" accesskey="t" title="(Alt + T)">
<svg id="moon" xmlns="http://www.w3.org/2000/svg" width="24" height="18" viewBox="0 0 24 24"
Expand Down Expand Up @@ -165,8 +165,8 @@ <h1 class="post-title entry-hint-parent">

</div>
</header>
<figure class="entry-cover"><img loading="eager" src="//localhost:1313/img/ucsd-geisel.jpg" alt="Hao Lab @ UCSD">
<p>Hao Lab @ UCSD</p>
<figure class="entry-cover"><img loading="eager" src="//localhost:1313/img/ucsd-geisel.jpg" alt="Hao AI Lab @ UCSD">
<p>Hao AI Lab @ UCSD</p>
</figure>
<div class="post-content"><h3 id="mission-statement">Mission Statement<a hidden class="anchor" aria-hidden="true" href="#mission-statement">#</a></h3>
<div style="text-align: justify;">
Expand Down Expand Up @@ -366,7 +366,7 @@ <h3 id="highlights">Highlights<a hidden class="anchor" aria-hidden="true" href="
</main>

<footer class="footer">
<span>&copy; 2024 <a href="//localhost:1313/">Hao Lab @ UCSD</a></span>
<span>&copy; 2024 <a href="//localhost:1313/">Hao AI Lab @ UCSD</a></span>
<span>
Powered by
<a href="https://gohugo.io/" rel="noopener noreferrer" target="_blank">Hugo</a> &
Expand Down
6 changes: 3 additions & 3 deletions public/index.html
Original file line number Diff line number Diff line change
Expand Up @@ -8,7 +8,7 @@
<meta http-equiv="X-UA-Compatible" content="IE=edge">
<meta name="viewport" content="width=device-width, initial-scale=1, shrink-to-fit=no">
<meta name="robots" content="noindex, nofollow">
<title>Hao Lab @ UCSD</title>
<title>Hao AI Lab @ UCSD</title>

<meta name="description" content="">
<meta name="author" content="">
Expand Down Expand Up @@ -89,7 +89,7 @@
<header class="header">
<nav class="nav">
<div class="logo">
<a href="//localhost:1313/" accesskey="h" title="Hao Lab @ UCSD (Alt + H)">Hao Lab @ UCSD</a>
<a href="//localhost:1313/" accesskey="h" title="Hao AI Lab @ UCSD (Alt + H)">Hao AI Lab @ UCSD</a>
<div class="logo-switches">
<button id="theme-toggle" accesskey="t" title="(Alt + T)">
<svg id="moon" xmlns="http://www.w3.org/2000/svg" width="24" height="18" viewBox="0 0 24 24"
Expand Down Expand Up @@ -175,7 +175,7 @@ <h2><a href="//localhost:1313/posts/lookahead_decoding/">Break the Sequential De
</main>

<footer class="footer">
<span>&copy; 2024 <a href="//localhost:1313/">Hao Lab @ UCSD</a></span>
<span>&copy; 2024 <a href="//localhost:1313/">Hao AI Lab @ UCSD</a></span>
<span>
Powered by
<a href="https://gohugo.io/" rel="noopener noreferrer" target="_blank">Hugo</a> &
Expand Down
4 changes: 2 additions & 2 deletions public/index.xml
Original file line number Diff line number Diff line change
@@ -1,9 +1,9 @@
<?xml version="1.0" encoding="utf-8" standalone="yes"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:content="http://purl.org/rss/1.0/modules/content/">
<channel>
<title>Hao Lab @ UCSD</title>
<title>Hao AI Lab @ UCSD</title>
<link>//localhost:1313/</link>
<description>Recent content on Hao Lab @ UCSD</description>
<description>Recent content on Hao AI Lab @ UCSD</description>
<generator>Hugo -- gohugo.io</generator>
<language>en-us</language>
<lastBuildDate>Wed, 21 Feb 2024 12:00:00 -0800</lastBuildDate>
Expand Down
Loading

0 comments on commit 1e0345c

Please sign in to comment.