deploy: f7528b364e
This commit is contained in:
eric
@@ -8,7 +8,7 @@
|
||||
NVIDIA’s Jetson Orin Nano promises impressive specs: 1024 CUDA cores, 32 Tensor Cores, and 40 TOPS of INT8 compute performance packed into a compact, power-efficient edge device. On paper, it looks like a capable platform for running Large Language Models locally. But there’s a catch—one that reveals a fundamental tension in modern edge AI hardware design.
|
||||
After running 66 inference tests across seven different language models ranging from 0.5B to 5.4B parameters, I discovered something counterintuitive: the device’s computational muscle sits largely idle during single-stream LLM inference. The bottleneck isn’t computation—it’s memory bandwidth. This isn’t just a quirk of one device; it’s a fundamental characteristic of single-user, autoregressive token generation on edge hardware—a reality that shapes how we should approach local LLM deployment."><meta name=keywords content="software engineer,performance engineering,Google engineer,tech blog,software development,performance optimization,Eric Liu,engineering blog,mountain biking,Jeep enthusiast,overlanding,camping,outdoor adventures"><meta name=twitter:card content="summary"><meta name=twitter:title content="Why Your Jetson Orin Nano's 40 TOPS Goes Unused (And What That Means for Edge AI)"><meta name=twitter:description content="Introduction Link to heading NVIDIA’s Jetson Orin Nano promises impressive specs: 1024 CUDA cores, 32 Tensor Cores, and 40 TOPS of INT8 compute performance packed into a compact, power-efficient edge device. On paper, it looks like a capable platform for running Large Language Models locally. But there’s a catch—one that reveals a fundamental tension in modern edge AI hardware design.
|
||||
After running 66 inference tests across seven different language models ranging from 0.5B to 5.4B parameters, I discovered something counterintuitive: the device’s computational muscle sits largely idle during single-stream LLM inference. The bottleneck isn’t computation—it’s memory bandwidth. This isn’t just a quirk of one device; it’s a fundamental characteristic of single-user, autoregressive token generation on edge hardware—a reality that shapes how we should approach local LLM deployment."><meta property="og:url" content="https://ericxliu.me/posts/benchmarking-llms-on-jetson-orin-nano/"><meta property="og:site_name" content="Eric X. Liu's Personal Page"><meta property="og:title" content="Why Your Jetson Orin Nano's 40 TOPS Goes Unused (And What That Means for Edge AI)"><meta property="og:description" content="Introduction Link to heading NVIDIA’s Jetson Orin Nano promises impressive specs: 1024 CUDA cores, 32 Tensor Cores, and 40 TOPS of INT8 compute performance packed into a compact, power-efficient edge device. On paper, it looks like a capable platform for running Large Language Models locally. But there’s a catch—one that reveals a fundamental tension in modern edge AI hardware design.
|
||||
After running 66 inference tests across seven different language models ranging from 0.5B to 5.4B parameters, I discovered something counterintuitive: the device’s computational muscle sits largely idle during single-stream LLM inference. The bottleneck isn’t computation—it’s memory bandwidth. This isn’t just a quirk of one device; it’s a fundamental characteristic of single-user, autoregressive token generation on edge hardware—a reality that shapes how we should approach local LLM deployment."><meta property="og:locale" content="en"><meta property="og:type" content="article"><meta property="article:section" content="posts"><meta property="article:published_time" content="2025-10-04T00:00:00+00:00"><meta property="article:modified_time" content="2025-10-04T20:41:50+00:00"><link rel=preload href=/fonts/fa-solid-900.woff2 as=font type=font/woff2 crossorigin><link rel=preload href=/fonts/fa-brands-400.woff2 as=font type=font/woff2 crossorigin><link rel=canonical href=https://ericxliu.me/posts/benchmarking-llms-on-jetson-orin-nano/><link rel=preload href=/fonts/fa-brands-400.woff2 as=font type=font/woff2 crossorigin><link rel=preload href=/fonts/fa-regular-400.woff2 as=font type=font/woff2 crossorigin><link rel=preload href=/fonts/fa-solid-900.woff2 as=font type=font/woff2 crossorigin><link rel=stylesheet href=/css/coder.min.4b392a85107b91dbdabc528edf014a6ab1a30cd44cafcd5325c8efe796794fca.css integrity="sha256-SzkqhRB7kdvavFKO3wFKarGjDNRMr81TJcjv55Z5T8o=" crossorigin=anonymous media=screen><link rel=stylesheet href=/css/coder-dark.min.a00e6364bacbc8266ad1cc81230774a1397198f8cfb7bcba29b7d6fcb54ce57f.css integrity="sha256-oA5jZLrLyCZq0cyBIwd0oTlxmPjPt7y6KbfW/LVM5X8=" crossorigin=anonymous media=screen><link rel=icon type=image/svg+xml href=/images/favicon.svg sizes=any><link rel=icon type=image/png href=/images/favicon-32x32.png sizes=32x32><link rel=icon type=image/png href=/images/favicon-16x16.png sizes=16x16><link rel=apple-touch-icon href=/images/apple-touch-icon.png><link rel=apple-touch-icon sizes=180x180 href=/images/apple-touch-icon.png><link rel=manifest href=/site.webmanifest><link rel=mask-icon href=/images/safari-pinned-tab.svg color=#5bbad5><script async src="https://pagead2.googlesyndication.com/pagead/js/adsbygoogle.js?client=ca-pub-3972604619956476" crossorigin=anonymous></script><script type=application/ld+json>{"@context":"http://schema.org","@type":"Person","name":"Eric X. Liu","url":"https:\/\/ericxliu.me\/","description":"Software \u0026 Performance Engineer at Google","sameAs":["https:\/\/www.linkedin.com\/in\/eric-x-liu-46648b93\/","https:\/\/git.ericxliu.me\/eric"]}</script><script type=application/ld+json>{"@context":"http://schema.org","@type":"BlogPosting","headline":"Why Your Jetson Orin Nano\u0027s 40 TOPS Goes Unused (And What That Means for Edge AI)","genre":"Blog","wordcount":"1866","url":"https:\/\/ericxliu.me\/posts\/benchmarking-llms-on-jetson-orin-nano\/","datePublished":"2025-10-04T00:00:00\u002b00:00","dateModified":"2025-10-04T20:41:50\u002b00:00","description":"\u003ch2 id=\u0022introduction\u0022\u003e\n Introduction\n \u003ca class=\u0022heading-link\u0022 href=\u0022#introduction\u0022\u003e\n \u003ci class=\u0022fa-solid fa-link\u0022 aria-hidden=\u0022true\u0022 title=\u0022Link to heading\u0022\u003e\u003c\/i\u003e\n \u003cspan class=\u0022sr-only\u0022\u003eLink to heading\u003c\/span\u003e\n \u003c\/a\u003e\n\u003c\/h2\u003e\n\u003cp\u003eNVIDIA\u0026rsquo;s Jetson Orin Nano promises impressive specs: 1024 CUDA cores, 32 Tensor Cores, and 40 TOPS of INT8 compute performance packed into a compact, power-efficient edge device. On paper, it looks like a capable platform for running Large Language Models locally. But there\u0026rsquo;s a catch—one that reveals a fundamental tension in modern edge AI hardware design.\u003c\/p\u003e\n\u003cp\u003eAfter running 66 inference tests across seven different language models ranging from 0.5B to 5.4B parameters, I discovered something counterintuitive: the device\u0026rsquo;s computational muscle sits largely idle during single-stream LLM inference. The bottleneck isn\u0026rsquo;t computation—it\u0026rsquo;s memory bandwidth. This isn\u0026rsquo;t just a quirk of one device; it\u0026rsquo;s a fundamental characteristic of single-user, autoregressive token generation on edge hardware—a reality that shapes how we should approach local LLM deployment.\u003c\/p\u003e","author":{"@type":"Person","name":"Eric X. Liu"}}</script></head><body class="preload-transitions colorscheme-auto"><div class=float-container><a id=dark-mode-toggle class=colorscheme-toggle><i class="fa-solid fa-adjust fa-fw" aria-hidden=true></i></a></div><main class=wrapper><nav class=navigation><section class=container><a class=navigation-title href=https://ericxliu.me/>Eric X. Liu's Personal Page
|
||||
After running 66 inference tests across seven different language models ranging from 0.5B to 5.4B parameters, I discovered something counterintuitive: the device’s computational muscle sits largely idle during single-stream LLM inference. The bottleneck isn’t computation—it’s memory bandwidth. This isn’t just a quirk of one device; it’s a fundamental characteristic of single-user, autoregressive token generation on edge hardware—a reality that shapes how we should approach local LLM deployment."><meta property="og:locale" content="en"><meta property="og:type" content="article"><meta property="article:section" content="posts"><meta property="article:published_time" content="2025-10-04T00:00:00+00:00"><meta property="article:modified_time" content="2026-01-08T18:13:13+00:00"><link rel=preload href=/fonts/fa-solid-900.woff2 as=font type=font/woff2 crossorigin><link rel=preload href=/fonts/fa-brands-400.woff2 as=font type=font/woff2 crossorigin><link rel=canonical href=https://ericxliu.me/posts/benchmarking-llms-on-jetson-orin-nano/><link rel=preload href=/fonts/fa-brands-400.woff2 as=font type=font/woff2 crossorigin><link rel=preload href=/fonts/fa-regular-400.woff2 as=font type=font/woff2 crossorigin><link rel=preload href=/fonts/fa-solid-900.woff2 as=font type=font/woff2 crossorigin><link rel=stylesheet href=/css/coder.min.4b392a85107b91dbdabc528edf014a6ab1a30cd44cafcd5325c8efe796794fca.css integrity="sha256-SzkqhRB7kdvavFKO3wFKarGjDNRMr81TJcjv55Z5T8o=" crossorigin=anonymous media=screen><link rel=stylesheet href=/css/coder-dark.min.a00e6364bacbc8266ad1cc81230774a1397198f8cfb7bcba29b7d6fcb54ce57f.css integrity="sha256-oA5jZLrLyCZq0cyBIwd0oTlxmPjPt7y6KbfW/LVM5X8=" crossorigin=anonymous media=screen><link rel=icon type=image/svg+xml href=/images/favicon.svg sizes=any><link rel=icon type=image/png href=/images/favicon-32x32.png sizes=32x32><link rel=icon type=image/png href=/images/favicon-16x16.png sizes=16x16><link rel=apple-touch-icon href=/images/apple-touch-icon.png><link rel=apple-touch-icon sizes=180x180 href=/images/apple-touch-icon.png><link rel=manifest href=/site.webmanifest><link rel=mask-icon href=/images/safari-pinned-tab.svg color=#5bbad5><script async src="https://pagead2.googlesyndication.com/pagead/js/adsbygoogle.js?client=ca-pub-3972604619956476" crossorigin=anonymous></script><script type=application/ld+json>{"@context":"http://schema.org","@type":"Person","name":"Eric X. Liu","url":"https:\/\/ericxliu.me\/","description":"Software \u0026 Performance Engineer at Google","sameAs":["https:\/\/www.linkedin.com\/in\/eric-x-liu-46648b93\/","https:\/\/git.ericxliu.me\/eric"]}</script><script type=application/ld+json>{"@context":"http://schema.org","@type":"BlogPosting","headline":"Why Your Jetson Orin Nano\u0027s 40 TOPS Goes Unused (And What That Means for Edge AI)","genre":"Blog","wordcount":"1866","url":"https:\/\/ericxliu.me\/posts\/benchmarking-llms-on-jetson-orin-nano\/","datePublished":"2025-10-04T00:00:00\u002b00:00","dateModified":"2026-01-08T18:13:13\u002b00:00","description":"\u003ch2 id=\u0022introduction\u0022\u003e\n Introduction\n \u003ca class=\u0022heading-link\u0022 href=\u0022#introduction\u0022\u003e\n \u003ci class=\u0022fa-solid fa-link\u0022 aria-hidden=\u0022true\u0022 title=\u0022Link to heading\u0022\u003e\u003c\/i\u003e\n \u003cspan class=\u0022sr-only\u0022\u003eLink to heading\u003c\/span\u003e\n \u003c\/a\u003e\n\u003c\/h2\u003e\n\u003cp\u003eNVIDIA\u0026rsquo;s Jetson Orin Nano promises impressive specs: 1024 CUDA cores, 32 Tensor Cores, and 40 TOPS of INT8 compute performance packed into a compact, power-efficient edge device. On paper, it looks like a capable platform for running Large Language Models locally. But there\u0026rsquo;s a catch—one that reveals a fundamental tension in modern edge AI hardware design.\u003c\/p\u003e\n\u003cp\u003eAfter running 66 inference tests across seven different language models ranging from 0.5B to 5.4B parameters, I discovered something counterintuitive: the device\u0026rsquo;s computational muscle sits largely idle during single-stream LLM inference. The bottleneck isn\u0026rsquo;t computation—it\u0026rsquo;s memory bandwidth. This isn\u0026rsquo;t just a quirk of one device; it\u0026rsquo;s a fundamental characteristic of single-user, autoregressive token generation on edge hardware—a reality that shapes how we should approach local LLM deployment.\u003c\/p\u003e","author":{"@type":"Person","name":"Eric X. Liu"}}</script></head><body class="preload-transitions colorscheme-auto"><div class=float-container><a id=dark-mode-toggle class=colorscheme-toggle><i class="fa-solid fa-adjust fa-fw" aria-hidden=true></i></a></div><main class=wrapper><nav class=navigation><section class=container><a class=navigation-title href=https://ericxliu.me/>Eric X. Liu's Personal Page
|
||||
</a><input type=checkbox id=menu-toggle>
|
||||
<label class="menu-button float-right" for=menu-toggle><i class="fa-solid fa-bars fa-fw" aria-hidden=true></i></label><ul class=navigation-list><li class=navigation-item><a class=navigation-link href=/posts/>Posts</a></li><li class=navigation-item><a class=navigation-link href=https://chat.ericxliu.me>Chat</a></li><li class=navigation-item><a class=navigation-link href=https://git.ericxliu.me/user/oauth2/Authenitk>Git</a></li><li class=navigation-item><a class=navigation-link href=https://coder.ericxliu.me/api/v2/users/oidc/callback>Coder</a></li><li class=navigation-item><a class=navigation-link href=/about/>About</a></li><li class=navigation-item><a class=navigation-link href=/>|</a></li><li class=navigation-item><a class=navigation-link href=https://sso.ericxliu.me>Sign in</a></li></ul></section></nav><div class=content><section class="container post"><article><header><div class=post-title><h1 class=title><a class=title-link href=https://ericxliu.me/posts/benchmarking-llms-on-jetson-orin-nano/>Why Your Jetson Orin Nano's 40 TOPS Goes Unused (And What That Means for Edge AI)</a></h1></div><div class=post-meta><div class=date><span class=posted-on><i class="fa-solid fa-calendar" aria-hidden=true></i>
|
||||
<time datetime=2025-10-04T00:00:00Z>October 4, 2025
|
||||
@@ -25,17 +25,17 @@ After running 66 inference tests across seven different language models ranging
|
||||
<a class=heading-link href=#the-testing-process><i class="fa-solid fa-link" aria-hidden=true title="Link to heading"></i>
|
||||
<span class=sr-only>Link to heading</span></a></h3><p>Each model faced 10-12 prompts of varying complexity—from simple arithmetic to technical explanations about LLMs themselves. All tests ran with batch size = 1, simulating a single user interacting with a local chatbot—the typical edge deployment scenario. Out of 84 planned tests, 66 completed successfully (78.6% success rate). The failures? Mostly out-of-memory crashes on larger models and occasional inference engine instability.</p><h3 id=understanding-the-limits-roofline-analysis>Understanding the Limits: Roofline Analysis
|
||||
<a class=heading-link href=#understanding-the-limits-roofline-analysis><i class="fa-solid fa-link" aria-hidden=true title="Link to heading"></i>
|
||||
<span class=sr-only>Link to heading</span></a></h3><p>To understand where performance hits its ceiling, I applied roofline analysis—a method that reveals whether a workload is compute-bound (limited by processing power) or memory-bound (limited by data transfer speed). For each model, I calculated:</p><ul><li><strong>FLOPs per token</strong>: Approximately 2 × total_parameters (accounting for matrix multiplications in forward pass)</li><li><strong>Bytes per token</strong>: model_size × 1.1 (including 10% overhead for activations and KV cache)</li><li><strong>Operational Intensity (OI)</strong>: FLOPs per token / Bytes per token</li><li><strong>Theoretical performance</strong>: min(compute_limit, bandwidth_limit)</li></ul><p>The roofline model works by comparing a workload’s operational intensity (how many calculations you do per byte of data moved) against the device’s balance point. If your operational intensity is too low, you’re bottlenecked by memory bandwidth—and as we’ll see, that’s exactly what happens with LLM inference.</p><p><img src=/images/benchmarking-llms-on-jetson-orin-nano/16d64bdc9cf14b05b7c40c4718b8091b.png alt="S3 File"></p><h2 id=the-results-speed-and-efficiency>The Results: Speed and Efficiency
|
||||
<span class=sr-only>Link to heading</span></a></h3><p>To understand where performance hits its ceiling, I applied roofline analysis—a method that reveals whether a workload is compute-bound (limited by processing power) or memory-bound (limited by data transfer speed). For each model, I calculated:</p><ul><li><strong>FLOPs per token</strong>: Approximately 2 × total_parameters (accounting for matrix multiplications in forward pass)</li><li><strong>Bytes per token</strong>: model_size × 1.1 (including 10% overhead for activations and KV cache)</li><li><strong>Operational Intensity (OI)</strong>: FLOPs per token / Bytes per token</li><li><strong>Theoretical performance</strong>: min(compute_limit, bandwidth_limit)</li></ul><p>The roofline model works by comparing a workload’s operational intensity (how many calculations you do per byte of data moved) against the device’s balance point. If your operational intensity is too low, you’re bottlenecked by memory bandwidth—and as we’ll see, that’s exactly what happens with LLM inference.</p><p><img src="http://localhost:4998/attachments/image-79378d40267258c0d8968238cc62bd197dc894fa.png?client=default&bucket=obsidian" alt="S3 File"></p><h2 id=the-results-speed-and-efficiency>The Results: Speed and Efficiency
|
||||
<a class=heading-link href=#the-results-speed-and-efficiency><i class="fa-solid fa-link" aria-hidden=true title="Link to heading"></i>
|
||||
<span class=sr-only>Link to heading</span></a></h2><h3 id=what-actually-runs-fast>What Actually Runs Fast
|
||||
<a class=heading-link href=#what-actually-runs-fast><i class="fa-solid fa-link" aria-hidden=true title="Link to heading"></i>
|
||||
<span class=sr-only>Link to heading</span></a></h3><p>Here’s how the models ranked by token generation speed:</p><table><thead><tr><th>Rank</th><th>Model</th><th>Backend</th><th>Avg Speed (t/s)</th><th>Std Dev</th><th>Success Rate</th></tr></thead><tbody><tr><td>1</td><td>qwen3:0.6b</td><td>Ollama</td><td>38.84</td><td>1.42</td><td>100%</td></tr><tr><td>2</td><td>qwen2.5:0.5b</td><td>Ollama</td><td>35.24</td><td>2.72</td><td>100%</td></tr><tr><td>3</td><td>gemma3:1b</td><td>Ollama</td><td>26.33</td><td>2.56</td><td>100%</td></tr><tr><td>4</td><td>Qwen/Qwen2.5-0.5B-Instruct</td><td>vLLM</td><td>15.18</td><td>2.15</td><td>100%</td></tr><tr><td>5</td><td>Qwen/Qwen3-0.6B-FP8</td><td>vLLM</td><td>12.81</td><td>0.36</td><td>100%</td></tr><tr><td>6</td><td>gemma3n:e2b</td><td>Ollama</td><td>8.98</td><td>1.22</td><td>100%</td></tr><tr><td>7</td><td>google/gemma-3-1b-it</td><td>vLLM</td><td>4.59</td><td>1.52</td><td>100%</td></tr></tbody></table><p>The standout finding: quantized sub-1B models hit 25-40 tokens/second, with Ollama consistently outperforming vLLM by 2-6× thanks to aggressive quantization and edge-optimized execution. These numbers align well with independent benchmarks from NVIDIA’s Jetson AI Lab (Llama 3.2 3B at 27.7 t/s, SmolLM2 at 41 t/s), confirming this is typical performance for the hardware class.
|
||||
<img src=/images/benchmarking-llms-on-jetson-orin-nano/ee04876d75d247f9b27a647462555777.png alt="S3 File"></p><h3 id=responsiveness-first-token-latency>Responsiveness: First Token Latency
|
||||
<img src="http://localhost:4998/attachments/image-7913a54157c2f4b8d0b7f961640a9c359b2d2a4f.png?client=default&bucket=obsidian" alt="S3 File"></p><h3 id=responsiveness-first-token-latency>Responsiveness: First Token Latency
|
||||
<a class=heading-link href=#responsiveness-first-token-latency><i class="fa-solid fa-link" aria-hidden=true title="Link to heading"></i>
|
||||
<span class=sr-only>Link to heading</span></a></h3><p>The time to generate the first output token—a critical metric for interactive applications—varied significantly:</p><ul><li>qwen3:0.6b (Ollama): 0.522 seconds</li><li>gemma3:1b (Ollama): 1.000 seconds</li><li>qwen2.5:0.5b (Ollama): 1.415 seconds</li><li>gemma3n:e2b (Ollama): 1.998 seconds</li></ul><p>Smaller, quantized models get to that first token faster—exactly what you want for a chatbot or interactive assistant where perceived responsiveness matters as much as raw throughput.</p><h3 id=the-memory-bottleneck-revealed>The Memory Bottleneck Revealed
|
||||
<a class=heading-link href=#the-memory-bottleneck-revealed><i class="fa-solid fa-link" aria-hidden=true title="Link to heading"></i>
|
||||
<span class=sr-only>Link to heading</span></a></h3><p>When I compared actual performance against theoretical limits, the results were striking:</p><table><thead><tr><th>Model</th><th>Theoretical (t/s)</th><th>Actual (t/s)</th><th>Efficiency</th><th>Bottleneck</th><th>OI (FLOPs/byte)</th></tr></thead><tbody><tr><td>gemma3:1b</td><td>109.90</td><td>26.33</td><td>24.0%</td><td>Memory</td><td>3.23</td></tr><tr><td>qwen3:0.6b</td><td>103.03</td><td>38.84</td><td>37.7%</td><td>Memory</td><td>1.82</td></tr><tr><td>qwen2.5:0.5b</td><td>219.80</td><td>35.24</td><td>16.0%</td><td>Memory</td><td>3.23</td></tr><tr><td>gemma3n:e2b</td><td>54.95</td><td>8.98</td><td>16.3%</td><td>Memory</td><td>3.23</td></tr><tr><td>google/gemma-3-1b-it</td><td>30.91</td><td>4.59</td><td>14.9%</td><td>Memory</td><td>0.91</td></tr><tr><td>Qwen/Qwen3-0.6B-FP8</td><td>103.03</td><td>12.81</td><td>12.4%</td><td>Memory</td><td>1.82</td></tr><tr><td>Qwen/Qwen2.5-0.5B-Instruct</td><td>61.82</td><td>15.18</td><td>24.6%</td><td>Memory</td><td>0.91</td></tr></tbody></table><p>Every single model is memory-bound in this single-stream inference scenario. Average hardware efficiency sits at just 20.8%—meaning the computational units spend most of their time waiting for data rather than crunching numbers. That advertised 40 TOPS? Largely untapped when generating one token at a time for a single user.
|
||||
<img src=/images/benchmarking-llms-on-jetson-orin-nano/ee04876d75d247f9b27a647462555777.png alt="S3 File"></p><h2 id=what-this-actually-means>What This Actually Means
|
||||
<img src="http://localhost:4998/attachments/image-7913a54157c2f4b8d0b7f961640a9c359b2d2a4f.png?client=default&bucket=obsidian" alt="S3 File"></p><h2 id=what-this-actually-means>What This Actually Means
|
||||
<a class=heading-link href=#what-this-actually-means><i class="fa-solid fa-link" aria-hidden=true title="Link to heading"></i>
|
||||
<span class=sr-only>Link to heading</span></a></h2><h3 id=why-memory-bandwidth-dominates-in-single-stream-inference>Why Memory Bandwidth Dominates (in Single-Stream Inference)
|
||||
<a class=heading-link href=#why-memory-bandwidth-dominates-in-single-stream-inference><i class="fa-solid fa-link" aria-hidden=true title="Link to heading"></i>
|
||||
@@ -62,4 +62,4 @@ After running 66 inference tests across seven different language models ranging
|
||||
2016 -
|
||||
2026
|
||||
Eric X. Liu
|
||||
<a href="https://git.ericxliu.me/eric/ericxliu-me/commit/3b1396d">[3b1396d]</a></section></footer></main><script src=/js/coder.min.6ae284be93d2d19dad1f02b0039508d9aab3180a12a06dcc71b0b0ef7825a317.js integrity="sha256-auKEvpPS0Z2tHwKwA5UI2aqzGAoSoG3McbCw73gloxc="></script><script defer src=https://static.cloudflareinsights.com/beacon.min.js data-cf-beacon='{"token": "987638e636ce4dbb932d038af74c17d1"}'></script></body></html>
|
||||
<a href="https://git.ericxliu.me/eric/ericxliu-me/commit/f7528b3">[f7528b3]</a></section></footer></main><script src=/js/coder.min.6ae284be93d2d19dad1f02b0039508d9aab3180a12a06dcc71b0b0ef7825a317.js integrity="sha256-auKEvpPS0Z2tHwKwA5UI2aqzGAoSoG3McbCw73gloxc="></script><script defer src=https://static.cloudflareinsights.com/beacon.min.js data-cf-beacon='{"token": "987638e636ce4dbb932d038af74c17d1"}'></script></body></html>
|
||||
@@ -25,4 +25,4 @@ Understanding the Two Primary Maintenance Cycles Link to heading The Breville Ba
|
||||
2016 -
|
||||
2026
|
||||
Eric X. Liu
|
||||
<a href="https://git.ericxliu.me/eric/ericxliu-me/commit/3b1396d">[3b1396d]</a></section></footer></main><script src=/js/coder.min.6ae284be93d2d19dad1f02b0039508d9aab3180a12a06dcc71b0b0ef7825a317.js integrity="sha256-auKEvpPS0Z2tHwKwA5UI2aqzGAoSoG3McbCw73gloxc="></script><script defer src=https://static.cloudflareinsights.com/beacon.min.js data-cf-beacon='{"token": "987638e636ce4dbb932d038af74c17d1"}'></script></body></html>
|
||||
<a href="https://git.ericxliu.me/eric/ericxliu-me/commit/f7528b3">[f7528b3]</a></section></footer></main><script src=/js/coder.min.6ae284be93d2d19dad1f02b0039508d9aab3180a12a06dcc71b0b0ef7825a317.js integrity="sha256-auKEvpPS0Z2tHwKwA5UI2aqzGAoSoG3McbCw73gloxc="></script><script defer src=https://static.cloudflareinsights.com/beacon.min.js data-cf-beacon='{"token": "987638e636ce4dbb932d038af74c17d1"}'></script></body></html>
|
||||
@@ -44,4 +44,4 @@ My detailed Grafana dashboards said everything was fine. But my wife said the SS
|
||||
2016 -
|
||||
2026
|
||||
Eric X. Liu
|
||||
<a href="https://git.ericxliu.me/eric/ericxliu-me/commit/3b1396d">[3b1396d]</a></section></footer></main><script src=/js/coder.min.6ae284be93d2d19dad1f02b0039508d9aab3180a12a06dcc71b0b0ef7825a317.js integrity="sha256-auKEvpPS0Z2tHwKwA5UI2aqzGAoSoG3McbCw73gloxc="></script><script defer src=https://static.cloudflareinsights.com/beacon.min.js data-cf-beacon='{"token": "987638e636ce4dbb932d038af74c17d1"}'></script></body></html>
|
||||
<a href="https://git.ericxliu.me/eric/ericxliu-me/commit/f7528b3">[f7528b3]</a></section></footer></main><script src=/js/coder.min.6ae284be93d2d19dad1f02b0039508d9aab3180a12a06dcc71b0b0ef7825a317.js integrity="sha256-auKEvpPS0Z2tHwKwA5UI2aqzGAoSoG3McbCw73gloxc="></script><script defer src=https://static.cloudflareinsights.com/beacon.min.js data-cf-beacon='{"token": "987638e636ce4dbb932d038af74c17d1"}'></script></body></html>
|
||||
@@ -20,4 +20,4 @@ Our overarching philosophy is simple: isolate and change only one variable at a
|
||||
2016 -
|
||||
2026
|
||||
Eric X. Liu
|
||||
<a href="https://git.ericxliu.me/eric/ericxliu-me/commit/3b1396d">[3b1396d]</a></section></footer></main><script src=/js/coder.min.6ae284be93d2d19dad1f02b0039508d9aab3180a12a06dcc71b0b0ef7825a317.js integrity="sha256-auKEvpPS0Z2tHwKwA5UI2aqzGAoSoG3McbCw73gloxc="></script><script defer src=https://static.cloudflareinsights.com/beacon.min.js data-cf-beacon='{"token": "987638e636ce4dbb932d038af74c17d1"}'></script></body></html>
|
||||
<a href="https://git.ericxliu.me/eric/ericxliu-me/commit/f7528b3">[f7528b3]</a></section></footer></main><script src=/js/coder.min.6ae284be93d2d19dad1f02b0039508d9aab3180a12a06dcc71b0b0ef7825a317.js integrity="sha256-auKEvpPS0Z2tHwKwA5UI2aqzGAoSoG3McbCw73gloxc="></script><script defer src=https://static.cloudflareinsights.com/beacon.min.js data-cf-beacon='{"token": "987638e636ce4dbb932d038af74c17d1"}'></script></body></html>
|
||||
File diff suppressed because one or more lines are too long
@@ -18,4 +18,4 @@ The answer lies in creating a universal language—a bridge between the continuo
|
||||
2016 -
|
||||
2026
|
||||
Eric X. Liu
|
||||
<a href="https://git.ericxliu.me/eric/ericxliu-me/commit/3b1396d">[3b1396d]</a></section></footer></main><script src=/js/coder.min.6ae284be93d2d19dad1f02b0039508d9aab3180a12a06dcc71b0b0ef7825a317.js integrity="sha256-auKEvpPS0Z2tHwKwA5UI2aqzGAoSoG3McbCw73gloxc="></script><script defer src=https://static.cloudflareinsights.com/beacon.min.js data-cf-beacon='{"token": "987638e636ce4dbb932d038af74c17d1"}'></script></body></html>
|
||||
<a href="https://git.ericxliu.me/eric/ericxliu-me/commit/f7528b3">[f7528b3]</a></section></footer></main><script src=/js/coder.min.6ae284be93d2d19dad1f02b0039508d9aab3180a12a06dcc71b0b0ef7825a317.js integrity="sha256-auKEvpPS0Z2tHwKwA5UI2aqzGAoSoG3McbCw73gloxc="></script><script defer src=https://static.cloudflareinsights.com/beacon.min.js data-cf-beacon='{"token": "987638e636ce4dbb932d038af74c17d1"}'></script></body></html>
|
||||
@@ -14,4 +14,4 @@
|
||||
2016 -
|
||||
2026
|
||||
Eric X. Liu
|
||||
<a href="https://git.ericxliu.me/eric/ericxliu-me/commit/3b1396d">[3b1396d]</a></section></footer></main><script src=/js/coder.min.6ae284be93d2d19dad1f02b0039508d9aab3180a12a06dcc71b0b0ef7825a317.js integrity="sha256-auKEvpPS0Z2tHwKwA5UI2aqzGAoSoG3McbCw73gloxc="></script><script defer src=https://static.cloudflareinsights.com/beacon.min.js data-cf-beacon='{"token": "987638e636ce4dbb932d038af74c17d1"}'></script></body></html>
|
||||
<a href="https://git.ericxliu.me/eric/ericxliu-me/commit/f7528b3">[f7528b3]</a></section></footer></main><script src=/js/coder.min.6ae284be93d2d19dad1f02b0039508d9aab3180a12a06dcc71b0b0ef7825a317.js integrity="sha256-auKEvpPS0Z2tHwKwA5UI2aqzGAoSoG3McbCw73gloxc="></script><script defer src=https://static.cloudflareinsights.com/beacon.min.js data-cf-beacon='{"token": "987638e636ce4dbb932d038af74c17d1"}'></script></body></html>
|
||||
@@ -1,4 +1,4 @@
|
||||
<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>Posts on Eric X. Liu's Personal Page</title><link>https://ericxliu.me/posts/</link><description>Recent content in Posts on Eric X. Liu's Personal Page</description><generator>Hugo</generator><language>en</language><lastBuildDate>Thu, 08 Jan 2026 06:02:38 +0000</lastBuildDate><atom:link href="https://ericxliu.me/posts/index.xml" rel="self" type="application/rss+xml"/><item><title>Why I Downgraded Magisk to Root My Pixel 2 XL</title><link>https://ericxliu.me/posts/rooting-pixel-2-xl-for-reverse-engineering/</link><pubDate>Wed, 07 Jan 2026 00:00:00 +0000</pubDate><guid>https://ericxliu.me/posts/rooting-pixel-2-xl-for-reverse-engineering/</guid><description><p>For the past few weeks, I&rsquo;ve been stuck in a stalemate with my EcoFlow Bluetooth Protocol Reverse Engineering Project. I have the hci snoop logs, I have the decompiled APK, and I have a strong suspicion about where the authentication logic is hiding. But suspicion isn&rsquo;t proof.</p>
|
||||
<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>Posts on Eric X. Liu's Personal Page</title><link>https://ericxliu.me/posts/</link><description>Recent content in Posts on Eric X. Liu's Personal Page</description><generator>Hugo</generator><language>en</language><lastBuildDate>Thu, 08 Jan 2026 18:13:13 +0000</lastBuildDate><atom:link href="https://ericxliu.me/posts/index.xml" rel="self" type="application/rss+xml"/><item><title>Why I Downgraded Magisk to Root My Pixel 2 XL</title><link>https://ericxliu.me/posts/rooting-pixel-2-xl-for-reverse-engineering/</link><pubDate>Wed, 07 Jan 2026 00:00:00 +0000</pubDate><guid>https://ericxliu.me/posts/rooting-pixel-2-xl-for-reverse-engineering/</guid><description><p>For the past few weeks, I&rsquo;ve been stuck in a stalemate with my EcoFlow Bluetooth Protocol Reverse Engineering Project. I have the hci snoop logs, I have the decompiled APK, and I have a strong suspicion about where the authentication logic is hiding. But suspicion isn&rsquo;t proof.</p>
|
||||
<p>Static analysis has its limits. I found the &ldquo;smoking gun&rdquo; function—a native method responsible for encrypting the login payload—but understanding <em>how</em> it constructs that payload within a strict 13-byte limit purely from assembly (ARM64) was proving to be a headache.</p></description></item><item><title>Why Your "Resilient" Homelab is Slower Than a Raspberry Pi</title><link>https://ericxliu.me/posts/debugging-authentik-performance/</link><pubDate>Fri, 02 Jan 2026 00:00:00 +0000</pubDate><guid>https://ericxliu.me/posts/debugging-authentik-performance/</guid><description><p>In the world of self-hosting, there are many metrics for success: 99.9% uptime, sub-second latency, or a perfect GitOps pipeline. But for those of us running &ldquo;production&rdquo; at home, there is only one metric that truly matters: <strong>The Wife Acceptance Factor (WAF)</strong>.</p>
|
||||
<p>My detailed Grafana dashboards said everything was fine. But my wife said the SSO login was &ldquo;slow sometimes.&rdquo; She was right. Debugging it took me down a rabbit hole of connection pooling, misplaced assumptions, and the harsh reality of running databases on distributed storage.</p></description></item><item><title>How I Got Open WebUI Talking to OpenAI Web Search</title><link>https://ericxliu.me/posts/open-webui-openai-websearch/</link><pubDate>Mon, 29 Dec 2025 00:00:00 +0000</pubDate><guid>https://ericxliu.me/posts/open-webui-openai-websearch/</guid><description><p>OpenAI promised native web search in GPT‑5, but LiteLLM proxy deployments (and by extension Open WebUI) still choke on it—issue <a href="https://github.com/BerriAI/litellm/issues/13042" class="external-link" target="_blank" rel="noopener">#13042</a> tracks the fallout. I needed grounded answers inside Open WebUI anyway, so I built a workaround: route GPT‑5 traffic through the Responses API and mask every <code>web_search_call</code> before the UI ever sees it.</p>
|
||||
<p>This post documents the final setup, the hotfix script that keeps LiteLLM honest, and the tests that prove Open WebUI now streams cited answers without trying to execute the tool itself.</p></description></item><item><title>From Gemini-3-Flash to T5-Gemma-2: A Journey in Distilling a Family Finance LLM</title><link>https://ericxliu.me/posts/technical-deep-dive-llm-categorization/</link><pubDate>Sat, 27 Dec 2025 00:00:00 +0000</pubDate><guid>https://ericxliu.me/posts/technical-deep-dive-llm-categorization/</guid><description><p>Running a family finance system is surprisingly complex. What starts as a simple spreadsheet often evolves into a web of rules, exceptions, and &ldquo;wait, was this dinner or <em>vacation</em> dinner?&rdquo; questions.</p>
|
||||
@@ -66,7 +66,7 @@
|
||||
<p>The answer lies in creating a universal language—a bridge between the continuous, messy world of pixels and audio waves and the discrete, structured world of language tokens. One of the most elegant and powerful tools for building this bridge is <strong>Residual Vector Quantization (RVQ)</strong>.</p></description></item><item><title>Supabase Deep Dive: It's Not Magic, It's Just Postgres</title><link>https://ericxliu.me/posts/supabase-deep-dive/</link><pubDate>Sun, 03 Aug 2025 00:00:00 +0000</pubDate><guid>https://ericxliu.me/posts/supabase-deep-dive/</guid><description><p>In the world of Backend-as-a-Service (BaaS), platforms are often treated as magic boxes. You push data in, you get data out, and you hope the magic inside scales. While this simplicity is powerful, it can obscure the underlying mechanics, leaving developers wondering what&rsquo;s really going on.</p>
|
||||
<p>Supabase enters this space with a radically different philosophy: <strong>transparency</strong>. It provides the convenience of a BaaS, but it’s built on the world&rsquo;s most trusted relational database: PostgreSQL. The &ldquo;magic&rdquo; isn&rsquo;t a proprietary black box; it&rsquo;s a carefully assembled suite of open-source tools that enhance Postgres, not hide it.</p></description></item><item><title>A Deep Dive into PPO for Language Models</title><link>https://ericxliu.me/posts/ppo-for-language-models/</link><pubDate>Sat, 02 Aug 2025 00:00:00 +0000</pubDate><guid>https://ericxliu.me/posts/ppo-for-language-models/</guid><description><p>Large Language Models (LLMs) have demonstrated astonishing capabilities, but out-of-the-box, they are simply powerful text predictors. They don&rsquo;t inherently understand what makes a response helpful, harmless, or aligned with human values. The technique that has proven most effective at bridging this gap is Reinforcement Learning from Human Feedback (RLHF), and at its heart lies a powerful algorithm: Proximal Policy Optimization (PPO).</p>
|
||||
<p>You may have seen diagrams like the one below, which outlines the RLHF training process. It can look intimidating, with a web of interconnected models, losses, and data flows.
|
||||
<img src="https://ericxliu.me/images/ppo-for-language-models/7713bd3ecf27442e939b9190fa08165d.png" alt="S3 File"></p></description></item><item><title>Mixture-of-Experts (MoE) Models Challenges & Solutions in Practice</title><link>https://ericxliu.me/posts/mixture-of-experts-moe-models-challenges-solutions-in-practice/</link><pubDate>Wed, 02 Jul 2025 00:00:00 +0000</pubDate><guid>https://ericxliu.me/posts/mixture-of-experts-moe-models-challenges-solutions-in-practice/</guid><description><p>Mixture-of-Experts (MoEs) are neural network architectures that allow different parts of the model (called &ldquo;experts&rdquo;) to specialize in different types of inputs. A &ldquo;gating network&rdquo; or &ldquo;router&rdquo; learns to dispatch each input (or &ldquo;token&rdquo;) to a subset of these experts. While powerful for scaling models, MoEs introduce several practical challenges.</p>
|
||||
<img src="http://localhost:4998/attachments/image-3632d923eed983f171fba4341825273101f1fc94.png?client=default&amp;bucket=obsidian" alt="S3 File"></p></description></item><item><title>Mixture-of-Experts (MoE) Models Challenges & Solutions in Practice</title><link>https://ericxliu.me/posts/mixture-of-experts-moe-models-challenges-solutions-in-practice/</link><pubDate>Wed, 02 Jul 2025 00:00:00 +0000</pubDate><guid>https://ericxliu.me/posts/mixture-of-experts-moe-models-challenges-solutions-in-practice/</guid><description><p>Mixture-of-Experts (MoEs) are neural network architectures that allow different parts of the model (called &ldquo;experts&rdquo;) to specialize in different types of inputs. A &ldquo;gating network&rdquo; or &ldquo;router&rdquo; learns to dispatch each input (or &ldquo;token&rdquo;) to a subset of these experts. While powerful for scaling models, MoEs introduce several practical challenges.</p>
|
||||
<h3 id="1-challenge-non-differentiability-of-routing-functions">
|
||||
1. Challenge: Non-Differentiability of Routing Functions
|
||||
<a class="heading-link" href="#1-challenge-non-differentiability-of-routing-functions">
|
||||
|
||||
@@ -71,4 +71,4 @@ Do not rely on header forwarding magic. Force the scheme in the plugin configura
|
||||
2016 -
|
||||
2026
|
||||
Eric X. Liu
|
||||
<a href="https://git.ericxliu.me/eric/ericxliu-me/commit/3b1396d">[3b1396d]</a></section></footer></main><script src=/js/coder.min.6ae284be93d2d19dad1f02b0039508d9aab3180a12a06dcc71b0b0ef7825a317.js integrity="sha256-auKEvpPS0Z2tHwKwA5UI2aqzGAoSoG3McbCw73gloxc="></script><script defer src=https://static.cloudflareinsights.com/beacon.min.js data-cf-beacon='{"token": "987638e636ce4dbb932d038af74c17d1"}'></script></body></html>
|
||||
<a href="https://git.ericxliu.me/eric/ericxliu-me/commit/f7528b3">[f7528b3]</a></section></footer></main><script src=/js/coder.min.6ae284be93d2d19dad1f02b0039508d9aab3180a12a06dcc71b0b0ef7825a317.js integrity="sha256-auKEvpPS0Z2tHwKwA5UI2aqzGAoSoG3McbCw73gloxc="></script><script defer src=https://static.cloudflareinsights.com/beacon.min.js data-cf-beacon='{"token": "987638e636ce4dbb932d038af74c17d1"}'></script></body></html>
|
||||
@@ -44,4 +44,4 @@ The <strong>Top-K routing</strong> mechanism, as illustrated in the provided ima
|
||||
2016 -
|
||||
2026
|
||||
Eric X. Liu
|
||||
<a href="https://git.ericxliu.me/eric/ericxliu-me/commit/3b1396d">[3b1396d]</a></section></footer></main><script src=/js/coder.min.6ae284be93d2d19dad1f02b0039508d9aab3180a12a06dcc71b0b0ef7825a317.js integrity="sha256-auKEvpPS0Z2tHwKwA5UI2aqzGAoSoG3McbCw73gloxc="></script><script defer src=https://static.cloudflareinsights.com/beacon.min.js data-cf-beacon='{"token": "987638e636ce4dbb932d038af74c17d1"}'></script></body></html>
|
||||
<a href="https://git.ericxliu.me/eric/ericxliu-me/commit/f7528b3">[f7528b3]</a></section></footer></main><script src=/js/coder.min.6ae284be93d2d19dad1f02b0039508d9aab3180a12a06dcc71b0b0ef7825a317.js integrity="sha256-auKEvpPS0Z2tHwKwA5UI2aqzGAoSoG3McbCw73gloxc="></script><script defer src=https://static.cloudflareinsights.com/beacon.min.js data-cf-beacon='{"token": "987638e636ce4dbb932d038af74c17d1"}'></script></body></html>
|
||||
@@ -86,4 +86,4 @@ This post documents the final setup, the hotfix script that keeps LiteLLM honest
|
||||
2016 -
|
||||
2026
|
||||
Eric X. Liu
|
||||
<a href="https://git.ericxliu.me/eric/ericxliu-me/commit/3b1396d">[3b1396d]</a></section></footer></main><script src=/js/coder.min.6ae284be93d2d19dad1f02b0039508d9aab3180a12a06dcc71b0b0ef7825a317.js integrity="sha256-auKEvpPS0Z2tHwKwA5UI2aqzGAoSoG3McbCw73gloxc="></script><script defer src=https://static.cloudflareinsights.com/beacon.min.js data-cf-beacon='{"token": "987638e636ce4dbb932d038af74c17d1"}'></script></body></html>
|
||||
<a href="https://git.ericxliu.me/eric/ericxliu-me/commit/f7528b3">[f7528b3]</a></section></footer></main><script src=/js/coder.min.6ae284be93d2d19dad1f02b0039508d9aab3180a12a06dcc71b0b0ef7825a317.js integrity="sha256-auKEvpPS0Z2tHwKwA5UI2aqzGAoSoG3McbCw73gloxc="></script><script defer src=https://static.cloudflareinsights.com/beacon.min.js data-cf-beacon='{"token": "987638e636ce4dbb932d038af74c17d1"}'></script></body></html>
|
||||
@@ -98,4 +98,4 @@ When using WireGuard together with MWAN3 on OpenWrt, the tunnel can fail to esta
|
||||
2016 -
|
||||
2026
|
||||
Eric X. Liu
|
||||
<a href="https://git.ericxliu.me/eric/ericxliu-me/commit/3b1396d">[3b1396d]</a></section></footer></main><script src=/js/coder.min.6ae284be93d2d19dad1f02b0039508d9aab3180a12a06dcc71b0b0ef7825a317.js integrity="sha256-auKEvpPS0Z2tHwKwA5UI2aqzGAoSoG3McbCw73gloxc="></script><script defer src=https://static.cloudflareinsights.com/beacon.min.js data-cf-beacon='{"token": "987638e636ce4dbb932d038af74c17d1"}'></script></body></html>
|
||||
<a href="https://git.ericxliu.me/eric/ericxliu-me/commit/f7528b3">[f7528b3]</a></section></footer></main><script src=/js/coder.min.6ae284be93d2d19dad1f02b0039508d9aab3180a12a06dcc71b0b0ef7825a317.js integrity="sha256-auKEvpPS0Z2tHwKwA5UI2aqzGAoSoG3McbCw73gloxc="></script><script defer src=https://static.cloudflareinsights.com/beacon.min.js data-cf-beacon='{"token": "987638e636ce4dbb932d038af74c17d1"}'></script></body></html>
|
||||
@@ -14,4 +14,4 @@
|
||||
2016 -
|
||||
2026
|
||||
Eric X. Liu
|
||||
<a href="https://git.ericxliu.me/eric/ericxliu-me/commit/3b1396d">[3b1396d]</a></section></footer></main><script src=/js/coder.min.6ae284be93d2d19dad1f02b0039508d9aab3180a12a06dcc71b0b0ef7825a317.js integrity="sha256-auKEvpPS0Z2tHwKwA5UI2aqzGAoSoG3McbCw73gloxc="></script><script defer src=https://static.cloudflareinsights.com/beacon.min.js data-cf-beacon='{"token": "987638e636ce4dbb932d038af74c17d1"}'></script></body></html>
|
||||
<a href="https://git.ericxliu.me/eric/ericxliu-me/commit/f7528b3">[f7528b3]</a></section></footer></main><script src=/js/coder.min.6ae284be93d2d19dad1f02b0039508d9aab3180a12a06dcc71b0b0ef7825a317.js integrity="sha256-auKEvpPS0Z2tHwKwA5UI2aqzGAoSoG3McbCw73gloxc="></script><script defer src=https://static.cloudflareinsights.com/beacon.min.js data-cf-beacon='{"token": "987638e636ce4dbb932d038af74c17d1"}'></script></body></html>
|
||||
@@ -6,4 +6,4 @@
|
||||
2016 -
|
||||
2026
|
||||
Eric X. Liu
|
||||
<a href="https://git.ericxliu.me/eric/ericxliu-me/commit/3b1396d">[3b1396d]</a></section></footer></main><script src=/js/coder.min.6ae284be93d2d19dad1f02b0039508d9aab3180a12a06dcc71b0b0ef7825a317.js integrity="sha256-auKEvpPS0Z2tHwKwA5UI2aqzGAoSoG3McbCw73gloxc="></script><script defer src=https://static.cloudflareinsights.com/beacon.min.js data-cf-beacon='{"token": "987638e636ce4dbb932d038af74c17d1"}'></script></body></html>
|
||||
<a href="https://git.ericxliu.me/eric/ericxliu-me/commit/f7528b3">[f7528b3]</a></section></footer></main><script src=/js/coder.min.6ae284be93d2d19dad1f02b0039508d9aab3180a12a06dcc71b0b0ef7825a317.js integrity="sha256-auKEvpPS0Z2tHwKwA5UI2aqzGAoSoG3McbCw73gloxc="></script><script defer src=https://static.cloudflareinsights.com/beacon.min.js data-cf-beacon='{"token": "987638e636ce4dbb932d038af74c17d1"}'></script></body></html>
|
||||
@@ -2,13 +2,13 @@
|
||||
You may have seen diagrams like the one below, which outlines the RLHF training process. It can look intimidating, with a web of interconnected models, losses, and data flows.
|
||||
"><meta name=keywords content="software engineer,performance engineering,Google engineer,tech blog,software development,performance optimization,Eric Liu,engineering blog,mountain biking,Jeep enthusiast,overlanding,camping,outdoor adventures"><meta name=twitter:card content="summary"><meta name=twitter:title content="A Deep Dive into PPO for Language Models"><meta name=twitter:description content="Large Language Models (LLMs) have demonstrated astonishing capabilities, but out-of-the-box, they are simply powerful text predictors. They don’t inherently understand what makes a response helpful, harmless, or aligned with human values. The technique that has proven most effective at bridging this gap is Reinforcement Learning from Human Feedback (RLHF), and at its heart lies a powerful algorithm: Proximal Policy Optimization (PPO).
|
||||
You may have seen diagrams like the one below, which outlines the RLHF training process. It can look intimidating, with a web of interconnected models, losses, and data flows."><meta property="og:url" content="https://ericxliu.me/posts/ppo-for-language-models/"><meta property="og:site_name" content="Eric X. Liu's Personal Page"><meta property="og:title" content="A Deep Dive into PPO for Language Models"><meta property="og:description" content="Large Language Models (LLMs) have demonstrated astonishing capabilities, but out-of-the-box, they are simply powerful text predictors. They don’t inherently understand what makes a response helpful, harmless, or aligned with human values. The technique that has proven most effective at bridging this gap is Reinforcement Learning from Human Feedback (RLHF), and at its heart lies a powerful algorithm: Proximal Policy Optimization (PPO).
|
||||
You may have seen diagrams like the one below, which outlines the RLHF training process. It can look intimidating, with a web of interconnected models, losses, and data flows."><meta property="og:locale" content="en"><meta property="og:type" content="article"><meta property="article:section" content="posts"><meta property="article:published_time" content="2025-08-02T00:00:00+00:00"><meta property="article:modified_time" content="2025-10-02T08:42:39+00:00"><link rel=preload href=/fonts/fa-solid-900.woff2 as=font type=font/woff2 crossorigin><link rel=preload href=/fonts/fa-brands-400.woff2 as=font type=font/woff2 crossorigin><link rel=canonical href=https://ericxliu.me/posts/ppo-for-language-models/><link rel=preload href=/fonts/fa-brands-400.woff2 as=font type=font/woff2 crossorigin><link rel=preload href=/fonts/fa-regular-400.woff2 as=font type=font/woff2 crossorigin><link rel=preload href=/fonts/fa-solid-900.woff2 as=font type=font/woff2 crossorigin><link rel=stylesheet href=/css/coder.min.4b392a85107b91dbdabc528edf014a6ab1a30cd44cafcd5325c8efe796794fca.css integrity="sha256-SzkqhRB7kdvavFKO3wFKarGjDNRMr81TJcjv55Z5T8o=" crossorigin=anonymous media=screen><link rel=stylesheet href=/css/coder-dark.min.a00e6364bacbc8266ad1cc81230774a1397198f8cfb7bcba29b7d6fcb54ce57f.css integrity="sha256-oA5jZLrLyCZq0cyBIwd0oTlxmPjPt7y6KbfW/LVM5X8=" crossorigin=anonymous media=screen><link rel=icon type=image/svg+xml href=/images/favicon.svg sizes=any><link rel=icon type=image/png href=/images/favicon-32x32.png sizes=32x32><link rel=icon type=image/png href=/images/favicon-16x16.png sizes=16x16><link rel=apple-touch-icon href=/images/apple-touch-icon.png><link rel=apple-touch-icon sizes=180x180 href=/images/apple-touch-icon.png><link rel=manifest href=/site.webmanifest><link rel=mask-icon href=/images/safari-pinned-tab.svg color=#5bbad5><script async src="https://pagead2.googlesyndication.com/pagead/js/adsbygoogle.js?client=ca-pub-3972604619956476" crossorigin=anonymous></script><script type=application/ld+json>{"@context":"http://schema.org","@type":"Person","name":"Eric X. Liu","url":"https:\/\/ericxliu.me\/","description":"Software \u0026 Performance Engineer at Google","sameAs":["https:\/\/www.linkedin.com\/in\/eric-x-liu-46648b93\/","https:\/\/git.ericxliu.me\/eric"]}</script><script type=application/ld+json>{"@context":"http://schema.org","@type":"BlogPosting","headline":"A Deep Dive into PPO for Language Models","genre":"Blog","wordcount":"1393","url":"https:\/\/ericxliu.me\/posts\/ppo-for-language-models\/","datePublished":"2025-08-02T00:00:00\u002b00:00","dateModified":"2025-10-02T08:42:39\u002b00:00","description":"\u003cp\u003eLarge Language Models (LLMs) have demonstrated astonishing capabilities, but out-of-the-box, they are simply powerful text predictors. They don\u0026rsquo;t inherently understand what makes a response helpful, harmless, or aligned with human values. The technique that has proven most effective at bridging this gap is Reinforcement Learning from Human Feedback (RLHF), and at its heart lies a powerful algorithm: Proximal Policy Optimization (PPO).\u003c\/p\u003e\n\u003cp\u003eYou may have seen diagrams like the one below, which outlines the RLHF training process. It can look intimidating, with a web of interconnected models, losses, and data flows.\n\u003cimg src=\u0022\/images\/ppo-for-language-models\/7713bd3ecf27442e939b9190fa08165d.png\u0022 alt=\u0022S3 File\u0022\u003e\u003c\/p\u003e","author":{"@type":"Person","name":"Eric X. Liu"}}</script></head><body class="preload-transitions colorscheme-auto"><div class=float-container><a id=dark-mode-toggle class=colorscheme-toggle><i class="fa-solid fa-adjust fa-fw" aria-hidden=true></i></a></div><main class=wrapper><nav class=navigation><section class=container><a class=navigation-title href=https://ericxliu.me/>Eric X. Liu's Personal Page
|
||||
You may have seen diagrams like the one below, which outlines the RLHF training process. It can look intimidating, with a web of interconnected models, losses, and data flows."><meta property="og:locale" content="en"><meta property="og:type" content="article"><meta property="article:section" content="posts"><meta property="article:published_time" content="2025-08-02T00:00:00+00:00"><meta property="article:modified_time" content="2026-01-08T18:13:13+00:00"><link rel=preload href=/fonts/fa-solid-900.woff2 as=font type=font/woff2 crossorigin><link rel=preload href=/fonts/fa-brands-400.woff2 as=font type=font/woff2 crossorigin><link rel=canonical href=https://ericxliu.me/posts/ppo-for-language-models/><link rel=preload href=/fonts/fa-brands-400.woff2 as=font type=font/woff2 crossorigin><link rel=preload href=/fonts/fa-regular-400.woff2 as=font type=font/woff2 crossorigin><link rel=preload href=/fonts/fa-solid-900.woff2 as=font type=font/woff2 crossorigin><link rel=stylesheet href=/css/coder.min.4b392a85107b91dbdabc528edf014a6ab1a30cd44cafcd5325c8efe796794fca.css integrity="sha256-SzkqhRB7kdvavFKO3wFKarGjDNRMr81TJcjv55Z5T8o=" crossorigin=anonymous media=screen><link rel=stylesheet href=/css/coder-dark.min.a00e6364bacbc8266ad1cc81230774a1397198f8cfb7bcba29b7d6fcb54ce57f.css integrity="sha256-oA5jZLrLyCZq0cyBIwd0oTlxmPjPt7y6KbfW/LVM5X8=" crossorigin=anonymous media=screen><link rel=icon type=image/svg+xml href=/images/favicon.svg sizes=any><link rel=icon type=image/png href=/images/favicon-32x32.png sizes=32x32><link rel=icon type=image/png href=/images/favicon-16x16.png sizes=16x16><link rel=apple-touch-icon href=/images/apple-touch-icon.png><link rel=apple-touch-icon sizes=180x180 href=/images/apple-touch-icon.png><link rel=manifest href=/site.webmanifest><link rel=mask-icon href=/images/safari-pinned-tab.svg color=#5bbad5><script async src="https://pagead2.googlesyndication.com/pagead/js/adsbygoogle.js?client=ca-pub-3972604619956476" crossorigin=anonymous></script><script type=application/ld+json>{"@context":"http://schema.org","@type":"Person","name":"Eric X. Liu","url":"https:\/\/ericxliu.me\/","description":"Software \u0026 Performance Engineer at Google","sameAs":["https:\/\/www.linkedin.com\/in\/eric-x-liu-46648b93\/","https:\/\/git.ericxliu.me\/eric"]}</script><script type=application/ld+json>{"@context":"http://schema.org","@type":"BlogPosting","headline":"A Deep Dive into PPO for Language Models","genre":"Blog","wordcount":"1393","url":"https:\/\/ericxliu.me\/posts\/ppo-for-language-models\/","datePublished":"2025-08-02T00:00:00\u002b00:00","dateModified":"2026-01-08T18:13:13\u002b00:00","description":"\u003cp\u003eLarge Language Models (LLMs) have demonstrated astonishing capabilities, but out-of-the-box, they are simply powerful text predictors. They don\u0026rsquo;t inherently understand what makes a response helpful, harmless, or aligned with human values. The technique that has proven most effective at bridging this gap is Reinforcement Learning from Human Feedback (RLHF), and at its heart lies a powerful algorithm: Proximal Policy Optimization (PPO).\u003c\/p\u003e\n\u003cp\u003eYou may have seen diagrams like the one below, which outlines the RLHF training process. It can look intimidating, with a web of interconnected models, losses, and data flows.\n\u003cimg src=\u0022http:\/\/localhost:4998\/attachments\/image-3632d923eed983f171fba4341825273101f1fc94.png?client=default\u0026amp;bucket=obsidian\u0022 alt=\u0022S3 File\u0022\u003e\u003c\/p\u003e","author":{"@type":"Person","name":"Eric X. Liu"}}</script></head><body class="preload-transitions colorscheme-auto"><div class=float-container><a id=dark-mode-toggle class=colorscheme-toggle><i class="fa-solid fa-adjust fa-fw" aria-hidden=true></i></a></div><main class=wrapper><nav class=navigation><section class=container><a class=navigation-title href=https://ericxliu.me/>Eric X. Liu's Personal Page
|
||||
</a><input type=checkbox id=menu-toggle>
|
||||
<label class="menu-button float-right" for=menu-toggle><i class="fa-solid fa-bars fa-fw" aria-hidden=true></i></label><ul class=navigation-list><li class=navigation-item><a class=navigation-link href=/posts/>Posts</a></li><li class=navigation-item><a class=navigation-link href=https://chat.ericxliu.me>Chat</a></li><li class=navigation-item><a class=navigation-link href=https://git.ericxliu.me/user/oauth2/Authenitk>Git</a></li><li class=navigation-item><a class=navigation-link href=https://coder.ericxliu.me/api/v2/users/oidc/callback>Coder</a></li><li class=navigation-item><a class=navigation-link href=/about/>About</a></li><li class=navigation-item><a class=navigation-link href=/>|</a></li><li class=navigation-item><a class=navigation-link href=https://sso.ericxliu.me>Sign in</a></li></ul></section></nav><div class=content><section class="container post"><article><header><div class=post-title><h1 class=title><a class=title-link href=https://ericxliu.me/posts/ppo-for-language-models/>A Deep Dive into PPO for Language Models</a></h1></div><div class=post-meta><div class=date><span class=posted-on><i class="fa-solid fa-calendar" aria-hidden=true></i>
|
||||
<time datetime=2025-08-02T00:00:00Z>August 2, 2025
|
||||
</time></span><span class=reading-time><i class="fa-solid fa-clock" aria-hidden=true></i>
|
||||
7-minute read</span></div></div></header><div class=post-content><p>Large Language Models (LLMs) have demonstrated astonishing capabilities, but out-of-the-box, they are simply powerful text predictors. They don’t inherently understand what makes a response helpful, harmless, or aligned with human values. The technique that has proven most effective at bridging this gap is Reinforcement Learning from Human Feedback (RLHF), and at its heart lies a powerful algorithm: Proximal Policy Optimization (PPO).</p><p>You may have seen diagrams like the one below, which outlines the RLHF training process. It can look intimidating, with a web of interconnected models, losses, and data flows.
|
||||
<img src=/images/ppo-for-language-models/7713bd3ecf27442e939b9190fa08165d.png alt="S3 File"></p><p>This post will decode that diagram, piece by piece. We’ll explore the “why” behind each component, moving from high-level concepts to the deep technical reasoning that makes this process work.</p><h3 id=translating-rl-to-a-conversation>Translating RL to a Conversation
|
||||
<img src="http://localhost:4998/attachments/image-3632d923eed983f171fba4341825273101f1fc94.png?client=default&bucket=obsidian" alt="S3 File"></p><p>This post will decode that diagram, piece by piece. We’ll explore the “why” behind each component, moving from high-level concepts to the deep technical reasoning that makes this process work.</p><h3 id=translating-rl-to-a-conversation>Translating RL to a Conversation
|
||||
<a class=heading-link href=#translating-rl-to-a-conversation><i class="fa-solid fa-link" aria-hidden=true title="Link to heading"></i>
|
||||
<span class=sr-only>Link to heading</span></a></h3><p>The first step is to understand how the traditional language of reinforcement learning maps to the world of text generation.</p><ul><li><strong>State (<code>s_t</code>)</strong>: In a chat setting, the “state” is the context of the conversation so far. It’s the initial prompt (<code>x</code>) plus all the text the model has generated up to the current moment (<code>y₁, ..., y_{t-1}</code>).</li><li><strong>Action (<code>a_t</code>)</strong>: The “action” is the model’s decision at each step. For an LLM, this means generating the very next token (<code>y_t</code>). A full response is a sequence of these actions.blob:https://aistudio.google.com/872e746f-88c1-40ec-8e45-fa0efce97299</li><li><strong>Reward (<code>r</code>)</strong>: The “reward” is a numeric score that tells the model how good its full response (<code>y</code>) was. This score comes from a separate <strong>Reward Model</strong>, which has been trained on a large dataset of human preference comparisons (e.g., humans rating which of two responses is better). This reward is often only awarded at the end of the entire generated sequence.</li></ul><p>Let’s make this concrete. If a user provides the prompt <strong>(x)</strong>: <em>“The best thing about AI is”</em>, and the model generates the response <strong>(y)</strong>: <em>“its potential to solve problems.”</em>, here is how it’s broken down for training:</p><ul><li><strong>State 1</strong>: “The best thing about AI is”<ul><li><strong>Action 1</strong>: “its”</li></ul></li><li><strong>State 2</strong>: “The best thing about AI is its”<ul><li><strong>Action 2</strong>: " potential"</li></ul></li><li><strong>State 3</strong>: “The best thing about AI is its potential”<ul><li><strong>Action 3</strong>: " to"</li></ul></li><li>…and so on for every generated token.</li></ul><p>This breakdown transforms a single prompt-response pair into a rich trajectory of state-action pairs, which becomes the raw data for our learning algorithm.</p><h3 id=the-cast-of-models-an-actor-critic-ensemble>The Cast of Models: An Actor-Critic Ensemble
|
||||
<a class=heading-link href=#the-cast-of-models-an-actor-critic-ensemble><i class="fa-solid fa-link" aria-hidden=true title="Link to heading"></i>
|
||||
@@ -25,4 +25,4 @@ where <code>δ_t = r_t + γV(s_{t+1}) - V(s_t)</code></p><ul><li><strong>γ (gam
|
||||
2016 -
|
||||
2026
|
||||
Eric X. Liu
|
||||
<a href="https://git.ericxliu.me/eric/ericxliu-me/commit/3b1396d">[3b1396d]</a></section></footer></main><script src=/js/coder.min.6ae284be93d2d19dad1f02b0039508d9aab3180a12a06dcc71b0b0ef7825a317.js integrity="sha256-auKEvpPS0Z2tHwKwA5UI2aqzGAoSoG3McbCw73gloxc="></script><script defer src=https://static.cloudflareinsights.com/beacon.min.js data-cf-beacon='{"token": "987638e636ce4dbb932d038af74c17d1"}'></script></body></html>
|
||||
<a href="https://git.ericxliu.me/eric/ericxliu-me/commit/f7528b3">[f7528b3]</a></section></footer></main><script src=/js/coder.min.6ae284be93d2d19dad1f02b0039508d9aab3180a12a06dcc71b0b0ef7825a317.js integrity="sha256-auKEvpPS0Z2tHwKwA5UI2aqzGAoSoG3McbCw73gloxc="></script><script defer src=https://static.cloudflareinsights.com/beacon.min.js data-cf-beacon='{"token": "987638e636ce4dbb932d038af74c17d1"}'></script></body></html>
|
||||
@@ -7,4 +7,4 @@
|
||||
2016 -
|
||||
2026
|
||||
Eric X. Liu
|
||||
<a href="https://git.ericxliu.me/eric/ericxliu-me/commit/3b1396d">[3b1396d]</a></section></footer></main><script src=/js/coder.min.6ae284be93d2d19dad1f02b0039508d9aab3180a12a06dcc71b0b0ef7825a317.js integrity="sha256-auKEvpPS0Z2tHwKwA5UI2aqzGAoSoG3McbCw73gloxc="></script><script defer src=https://static.cloudflareinsights.com/beacon.min.js data-cf-beacon='{"token": "987638e636ce4dbb932d038af74c17d1"}'></script></body></html>
|
||||
<a href="https://git.ericxliu.me/eric/ericxliu-me/commit/f7528b3">[f7528b3]</a></section></footer></main><script src=/js/coder.min.6ae284be93d2d19dad1f02b0039508d9aab3180a12a06dcc71b0b0ef7825a317.js integrity="sha256-auKEvpPS0Z2tHwKwA5UI2aqzGAoSoG3McbCw73gloxc="></script><script defer src=https://static.cloudflareinsights.com/beacon.min.js data-cf-beacon='{"token": "987638e636ce4dbb932d038af74c17d1"}'></script></body></html>
|
||||
@@ -35,4 +35,4 @@ I used <strong>Magisk v30.6</strong> (the latest as of writing). The patch proce
|
||||
2016 -
|
||||
2026
|
||||
Eric X. Liu
|
||||
<a href="https://git.ericxliu.me/eric/ericxliu-me/commit/3b1396d">[3b1396d]</a></section></footer></main><script src=/js/coder.min.6ae284be93d2d19dad1f02b0039508d9aab3180a12a06dcc71b0b0ef7825a317.js integrity="sha256-auKEvpPS0Z2tHwKwA5UI2aqzGAoSoG3McbCw73gloxc="></script><script defer src=https://static.cloudflareinsights.com/beacon.min.js data-cf-beacon='{"token": "987638e636ce4dbb932d038af74c17d1"}'></script></body></html>
|
||||
<a href="https://git.ericxliu.me/eric/ericxliu-me/commit/f7528b3">[f7528b3]</a></section></footer></main><script src=/js/coder.min.6ae284be93d2d19dad1f02b0039508d9aab3180a12a06dcc71b0b0ef7825a317.js integrity="sha256-auKEvpPS0Z2tHwKwA5UI2aqzGAoSoG3McbCw73gloxc="></script><script defer src=https://static.cloudflareinsights.com/beacon.min.js data-cf-beacon='{"token": "987638e636ce4dbb932d038af74c17d1"}'></script></body></html>
|
||||
@@ -59,4 +59,4 @@ nvidia-smi failed to communicate with the NVIDIA driver modprobe nvidia → “K
|
||||
2016 -
|
||||
2026
|
||||
Eric X. Liu
|
||||
<a href="https://git.ericxliu.me/eric/ericxliu-me/commit/3b1396d">[3b1396d]</a></section></footer></main><script src=/js/coder.min.6ae284be93d2d19dad1f02b0039508d9aab3180a12a06dcc71b0b0ef7825a317.js integrity="sha256-auKEvpPS0Z2tHwKwA5UI2aqzGAoSoG3McbCw73gloxc="></script><script defer src=https://static.cloudflareinsights.com/beacon.min.js data-cf-beacon='{"token": "987638e636ce4dbb932d038af74c17d1"}'></script></body></html>
|
||||
<a href="https://git.ericxliu.me/eric/ericxliu-me/commit/f7528b3">[f7528b3]</a></section></footer></main><script src=/js/coder.min.6ae284be93d2d19dad1f02b0039508d9aab3180a12a06dcc71b0b0ef7825a317.js integrity="sha256-auKEvpPS0Z2tHwKwA5UI2aqzGAoSoG3McbCw73gloxc="></script><script defer src=https://static.cloudflareinsights.com/beacon.min.js data-cf-beacon='{"token": "987638e636ce4dbb932d038af74c17d1"}'></script></body></html>
|
||||
@@ -90,4 +90,4 @@ Supabase enters this space with a radically different philosophy: transparency.
|
||||
2016 -
|
||||
2026
|
||||
Eric X. Liu
|
||||
<a href="https://git.ericxliu.me/eric/ericxliu-me/commit/3b1396d">[3b1396d]</a></section></footer></main><script src=/js/coder.min.6ae284be93d2d19dad1f02b0039508d9aab3180a12a06dcc71b0b0ef7825a317.js integrity="sha256-auKEvpPS0Z2tHwKwA5UI2aqzGAoSoG3McbCw73gloxc="></script><script defer src=https://static.cloudflareinsights.com/beacon.min.js data-cf-beacon='{"token": "987638e636ce4dbb932d038af74c17d1"}'></script></body></html>
|
||||
<a href="https://git.ericxliu.me/eric/ericxliu-me/commit/f7528b3">[f7528b3]</a></section></footer></main><script src=/js/coder.min.6ae284be93d2d19dad1f02b0039508d9aab3180a12a06dcc71b0b0ef7825a317.js integrity="sha256-auKEvpPS0Z2tHwKwA5UI2aqzGAoSoG3McbCw73gloxc="></script><script defer src=https://static.cloudflareinsights.com/beacon.min.js data-cf-beacon='{"token": "987638e636ce4dbb932d038af74c17d1"}'></script></body></html>
|
||||
@@ -30,4 +30,4 @@ But to truly understand the field, we must look at the pivotal models that explo
|
||||
2016 -
|
||||
2026
|
||||
Eric X. Liu
|
||||
<a href="https://git.ericxliu.me/eric/ericxliu-me/commit/3b1396d">[3b1396d]</a></section></footer></main><script src=/js/coder.min.6ae284be93d2d19dad1f02b0039508d9aab3180a12a06dcc71b0b0ef7825a317.js integrity="sha256-auKEvpPS0Z2tHwKwA5UI2aqzGAoSoG3McbCw73gloxc="></script><script defer src=https://static.cloudflareinsights.com/beacon.min.js data-cf-beacon='{"token": "987638e636ce4dbb932d038af74c17d1"}'></script></body></html>
|
||||
<a href="https://git.ericxliu.me/eric/ericxliu-me/commit/f7528b3">[f7528b3]</a></section></footer></main><script src=/js/coder.min.6ae284be93d2d19dad1f02b0039508d9aab3180a12a06dcc71b0b0ef7825a317.js integrity="sha256-auKEvpPS0Z2tHwKwA5UI2aqzGAoSoG3McbCw73gloxc="></script><script defer src=https://static.cloudflareinsights.com/beacon.min.js data-cf-beacon='{"token": "987638e636ce4dbb932d038af74c17d1"}'></script></body></html>
|
||||
@@ -1,7 +1,7 @@
|
||||
<!doctype html><html lang=en><head><title>From Gemini-3-Flash to T5-Gemma-2: A Journey in Distilling a Family Finance LLM · Eric X. Liu's Personal Page</title><meta charset=utf-8><meta name=viewport content="width=device-width,initial-scale=1"><meta name=color-scheme content="light dark"><meta http-equiv=Content-Security-Policy content="upgrade-insecure-requests; block-all-mixed-content; default-src 'self'; child-src 'self'; font-src 'self' https://fonts.gstatic.com https://cdn.jsdelivr.net/; form-action 'self'; frame-src 'self' https://www.youtube.com https://disqus.com; img-src 'self' https://referrer.disqus.com https://c.disquscdn.com https://*.disqus.com; object-src 'none'; style-src 'self' 'unsafe-inline' https://fonts.googleapis.com/ https://cdn.jsdelivr.net/; script-src 'self' 'unsafe-inline' https://www.google-analytics.com https://cdn.jsdelivr.net/ https://pagead2.googlesyndication.com https://static.cloudflareinsights.com https://unpkg.com https://ericxliu-me.disqus.com https://disqus.com https://*.disqus.com https://*.disquscdn.com https://unpkg.com; connect-src 'self' https://www.google-analytics.com https://pagead2.googlesyndication.com https://cloudflareinsights.com ws://localhost:1313 ws://localhost:* wss://localhost:* https://links.services.disqus.com https://*.disqus.com;"><meta name=author content="Eric X. Liu"><meta name=description content='Running a family finance system is surprisingly complex. What starts as a simple spreadsheet often evolves into a web of rules, exceptions, and “wait, was this dinner or vacation dinner?” questions.
|
||||
For years, I relied on a rule-based system to categorize our credit card transactions. It worked… mostly. But maintaining if "UBER" in description and amount > 50 style rules is a never-ending battle against the entropy of merchant names and changing habits.'><meta name=keywords content="software engineer,performance engineering,Google engineer,tech blog,software development,performance optimization,Eric Liu,engineering blog,mountain biking,Jeep enthusiast,overlanding,camping,outdoor adventures"><meta name=twitter:card content="summary"><meta name=twitter:title content="From Gemini-3-Flash to T5-Gemma-2: A Journey in Distilling a Family Finance LLM"><meta name=twitter:description content='Running a family finance system is surprisingly complex. What starts as a simple spreadsheet often evolves into a web of rules, exceptions, and “wait, was this dinner or vacation dinner?” questions.
|
||||
For years, I relied on a rule-based system to categorize our credit card transactions. It worked… mostly. But maintaining if "UBER" in description and amount > 50 style rules is a never-ending battle against the entropy of merchant names and changing habits.'><meta property="og:url" content="https://ericxliu.me/posts/technical-deep-dive-llm-categorization/"><meta property="og:site_name" content="Eric X. Liu's Personal Page"><meta property="og:title" content="From Gemini-3-Flash to T5-Gemma-2: A Journey in Distilling a Family Finance LLM"><meta property="og:description" content='Running a family finance system is surprisingly complex. What starts as a simple spreadsheet often evolves into a web of rules, exceptions, and “wait, was this dinner or vacation dinner?” questions.
|
||||
For years, I relied on a rule-based system to categorize our credit card transactions. It worked… mostly. But maintaining if "UBER" in description and amount > 50 style rules is a never-ending battle against the entropy of merchant names and changing habits.'><meta property="og:locale" content="en"><meta property="og:type" content="article"><meta property="article:section" content="posts"><meta property="article:published_time" content="2025-12-27T00:00:00+00:00"><meta property="article:modified_time" content="2026-01-03T06:57:12+00:00"><link rel=preload href=/fonts/fa-solid-900.woff2 as=font type=font/woff2 crossorigin><link rel=preload href=/fonts/fa-brands-400.woff2 as=font type=font/woff2 crossorigin><link rel=canonical href=https://ericxliu.me/posts/technical-deep-dive-llm-categorization/><link rel=preload href=/fonts/fa-brands-400.woff2 as=font type=font/woff2 crossorigin><link rel=preload href=/fonts/fa-regular-400.woff2 as=font type=font/woff2 crossorigin><link rel=preload href=/fonts/fa-solid-900.woff2 as=font type=font/woff2 crossorigin><link rel=stylesheet href=/css/coder.min.4b392a85107b91dbdabc528edf014a6ab1a30cd44cafcd5325c8efe796794fca.css integrity="sha256-SzkqhRB7kdvavFKO3wFKarGjDNRMr81TJcjv55Z5T8o=" crossorigin=anonymous media=screen><link rel=stylesheet href=/css/coder-dark.min.a00e6364bacbc8266ad1cc81230774a1397198f8cfb7bcba29b7d6fcb54ce57f.css integrity="sha256-oA5jZLrLyCZq0cyBIwd0oTlxmPjPt7y6KbfW/LVM5X8=" crossorigin=anonymous media=screen><link rel=icon type=image/svg+xml href=/images/favicon.svg sizes=any><link rel=icon type=image/png href=/images/favicon-32x32.png sizes=32x32><link rel=icon type=image/png href=/images/favicon-16x16.png sizes=16x16><link rel=apple-touch-icon href=/images/apple-touch-icon.png><link rel=apple-touch-icon sizes=180x180 href=/images/apple-touch-icon.png><link rel=manifest href=/site.webmanifest><link rel=mask-icon href=/images/safari-pinned-tab.svg color=#5bbad5><script async src="https://pagead2.googlesyndication.com/pagead/js/adsbygoogle.js?client=ca-pub-3972604619956476" crossorigin=anonymous></script><script type=application/ld+json>{"@context":"http://schema.org","@type":"Person","name":"Eric X. Liu","url":"https:\/\/ericxliu.me\/","description":"Software \u0026 Performance Engineer at Google","sameAs":["https:\/\/www.linkedin.com\/in\/eric-x-liu-46648b93\/","https:\/\/git.ericxliu.me\/eric"]}</script><script type=application/ld+json>{"@context":"http://schema.org","@type":"BlogPosting","headline":"From Gemini-3-Flash to T5-Gemma-2: A Journey in Distilling a Family Finance LLM","genre":"Blog","wordcount":"1355","url":"https:\/\/ericxliu.me\/posts\/technical-deep-dive-llm-categorization\/","datePublished":"2025-12-27T00:00:00\u002b00:00","dateModified":"2026-01-03T06:57:12\u002b00:00","description":"\u003cp\u003eRunning a family finance system is surprisingly complex. What starts as a simple spreadsheet often evolves into a web of rules, exceptions, and \u0026ldquo;wait, was this dinner or \u003cem\u003evacation\u003c\/em\u003e dinner?\u0026rdquo; questions.\u003c\/p\u003e\n\u003cp\u003eFor years, I relied on a rule-based system to categorize our credit card transactions. It worked\u0026hellip; mostly. But maintaining \u003ccode\u003eif \u0026quot;UBER\u0026quot; in description and amount \u0026gt; 50\u003c\/code\u003e style rules is a never-ending battle against the entropy of merchant names and changing habits.\u003c\/p\u003e","author":{"@type":"Person","name":"Eric X. Liu"}}</script></head><body class="preload-transitions colorscheme-auto"><div class=float-container><a id=dark-mode-toggle class=colorscheme-toggle><i class="fa-solid fa-adjust fa-fw" aria-hidden=true></i></a></div><main class=wrapper><nav class=navigation><section class=container><a class=navigation-title href=https://ericxliu.me/>Eric X. Liu's Personal Page
|
||||
For years, I relied on a rule-based system to categorize our credit card transactions. It worked… mostly. But maintaining if "UBER" in description and amount > 50 style rules is a never-ending battle against the entropy of merchant names and changing habits.'><meta property="og:locale" content="en"><meta property="og:type" content="article"><meta property="article:section" content="posts"><meta property="article:published_time" content="2025-12-27T00:00:00+00:00"><meta property="article:modified_time" content="2026-01-08T18:13:13+00:00"><link rel=preload href=/fonts/fa-solid-900.woff2 as=font type=font/woff2 crossorigin><link rel=preload href=/fonts/fa-brands-400.woff2 as=font type=font/woff2 crossorigin><link rel=canonical href=https://ericxliu.me/posts/technical-deep-dive-llm-categorization/><link rel=preload href=/fonts/fa-brands-400.woff2 as=font type=font/woff2 crossorigin><link rel=preload href=/fonts/fa-regular-400.woff2 as=font type=font/woff2 crossorigin><link rel=preload href=/fonts/fa-solid-900.woff2 as=font type=font/woff2 crossorigin><link rel=stylesheet href=/css/coder.min.4b392a85107b91dbdabc528edf014a6ab1a30cd44cafcd5325c8efe796794fca.css integrity="sha256-SzkqhRB7kdvavFKO3wFKarGjDNRMr81TJcjv55Z5T8o=" crossorigin=anonymous media=screen><link rel=stylesheet href=/css/coder-dark.min.a00e6364bacbc8266ad1cc81230774a1397198f8cfb7bcba29b7d6fcb54ce57f.css integrity="sha256-oA5jZLrLyCZq0cyBIwd0oTlxmPjPt7y6KbfW/LVM5X8=" crossorigin=anonymous media=screen><link rel=icon type=image/svg+xml href=/images/favicon.svg sizes=any><link rel=icon type=image/png href=/images/favicon-32x32.png sizes=32x32><link rel=icon type=image/png href=/images/favicon-16x16.png sizes=16x16><link rel=apple-touch-icon href=/images/apple-touch-icon.png><link rel=apple-touch-icon sizes=180x180 href=/images/apple-touch-icon.png><link rel=manifest href=/site.webmanifest><link rel=mask-icon href=/images/safari-pinned-tab.svg color=#5bbad5><script async src="https://pagead2.googlesyndication.com/pagead/js/adsbygoogle.js?client=ca-pub-3972604619956476" crossorigin=anonymous></script><script type=application/ld+json>{"@context":"http://schema.org","@type":"Person","name":"Eric X. Liu","url":"https:\/\/ericxliu.me\/","description":"Software \u0026 Performance Engineer at Google","sameAs":["https:\/\/www.linkedin.com\/in\/eric-x-liu-46648b93\/","https:\/\/git.ericxliu.me\/eric"]}</script><script type=application/ld+json>{"@context":"http://schema.org","@type":"BlogPosting","headline":"From Gemini-3-Flash to T5-Gemma-2: A Journey in Distilling a Family Finance LLM","genre":"Blog","wordcount":"1355","url":"https:\/\/ericxliu.me\/posts\/technical-deep-dive-llm-categorization\/","datePublished":"2025-12-27T00:00:00\u002b00:00","dateModified":"2026-01-08T18:13:13\u002b00:00","description":"\u003cp\u003eRunning a family finance system is surprisingly complex. What starts as a simple spreadsheet often evolves into a web of rules, exceptions, and \u0026ldquo;wait, was this dinner or \u003cem\u003evacation\u003c\/em\u003e dinner?\u0026rdquo; questions.\u003c\/p\u003e\n\u003cp\u003eFor years, I relied on a rule-based system to categorize our credit card transactions. It worked\u0026hellip; mostly. But maintaining \u003ccode\u003eif \u0026quot;UBER\u0026quot; in description and amount \u0026gt; 50\u003c\/code\u003e style rules is a never-ending battle against the entropy of merchant names and changing habits.\u003c\/p\u003e","author":{"@type":"Person","name":"Eric X. Liu"}}</script></head><body class="preload-transitions colorscheme-auto"><div class=float-container><a id=dark-mode-toggle class=colorscheme-toggle><i class="fa-solid fa-adjust fa-fw" aria-hidden=true></i></a></div><main class=wrapper><nav class=navigation><section class=container><a class=navigation-title href=https://ericxliu.me/>Eric X. Liu's Personal Page
|
||||
</a><input type=checkbox id=menu-toggle>
|
||||
<label class="menu-button float-right" for=menu-toggle><i class="fa-solid fa-bars fa-fw" aria-hidden=true></i></label><ul class=navigation-list><li class=navigation-item><a class=navigation-link href=/posts/>Posts</a></li><li class=navigation-item><a class=navigation-link href=https://chat.ericxliu.me>Chat</a></li><li class=navigation-item><a class=navigation-link href=https://git.ericxliu.me/user/oauth2/Authenitk>Git</a></li><li class=navigation-item><a class=navigation-link href=https://coder.ericxliu.me/api/v2/users/oidc/callback>Coder</a></li><li class=navigation-item><a class=navigation-link href=/about/>About</a></li><li class=navigation-item><a class=navigation-link href=/>|</a></li><li class=navigation-item><a class=navigation-link href=https://sso.ericxliu.me>Sign in</a></li></ul></section></nav><div class=content><section class="container post"><article><header><div class=post-title><h1 class=title><a class=title-link href=https://ericxliu.me/posts/technical-deep-dive-llm-categorization/>From Gemini-3-Flash to T5-Gemma-2: A Journey in Distilling a Family Finance LLM</a></h1></div><div class=post-meta><div class=date><span class=posted-on><i class="fa-solid fa-calendar" aria-hidden=true></i>
|
||||
<time datetime=2025-12-27T00:00:00Z>December 27, 2025
|
||||
@@ -43,7 +43,7 @@ It turned out to be a syntax error in my arguments passed to the <code>Trainer</
|
||||
<a class=heading-link href=#pitfall-2-stability-vs-noise><i class="fa-solid fa-link" aria-hidden=true title="Link to heading"></i>
|
||||
<span class=sr-only>Link to heading</span></a></h3><p>The loss curve was initially extremely erratic. The batch size on my GPU was limited (Physical Batch Size = 4).
|
||||
<strong>The Fix</strong>: I implemented <strong>Gradient Accumulation</strong> (accumulating over 8 steps) to simulate a batch size of 32. This smoothed out the optimization landscape significantly.
|
||||
<img src=/images/technical-deep-dive-llm-categorization/eedb3be8259a4a70aa7029b78a029364.png alt="S3 File"></p><h3 id=pitfall-3-overfitting>Pitfall #3: Overfitting
|
||||
<img src="http://localhost:4998/attachments/image-1b23344ea5541d156e5ac20823d12d7c6723b691.png?client=default&bucket=obsidian" alt="S3 File"></p><h3 id=pitfall-3-overfitting>Pitfall #3: Overfitting
|
||||
<a class=heading-link href=#pitfall-3-overfitting><i class="fa-solid fa-link" aria-hidden=true title="Link to heading"></i>
|
||||
<span class=sr-only>Link to heading</span></a></h3><p>With a small dataset (~2k samples), overfitting is a real risk. I employed a multi-layered defense strategy:</p><ol><li><strong>Data Quality First</strong>: The “Consistency Filtering” phase was the most critical step. By discarding ambiguous samples where the teacher model disagreed with history, I prevented the model from memorizing noise.</li><li><strong>Model Regularization</strong>:<ul><li><strong>LoRA Dropout</strong>: I set <code>lora_dropout=0.1</code>, randomly dropping 10% of the trainable adapter connections during training to force robust feature learning.</li><li><strong>Gradient Clipping</strong>: We capped the gradient norm at <code>1.0</code>. This prevents the “exploding gradient” problem and keeps weight updates stable.</li><li><strong>AdamW</strong>: Using the AdamW optimizer adds decoupled weight decay, implicitly penalizing overly complex weights.</li></ul></li></ol><p>I also set up a rigorous evaluation loop (10% validation split, eval every 50 steps) to monitor the <code>Train Loss</code> vs <code>Eval Loss</code> in real-time. The final curves showed them tracking downwards together, confirming generalization.</p><h2 id=phase-4-results-and-the-travel-edge-case>Phase 4: Results and The “Travel” Edge Case
|
||||
<a class=heading-link href=#phase-4-results-and-the-travel-edge-case><i class="fa-solid fa-link" aria-hidden=true title="Link to heading"></i>
|
||||
@@ -73,4 +73,4 @@ It turned out to be a syntax error in my arguments passed to the <code>Trainer</
|
||||
2016 -
|
||||
2026
|
||||
Eric X. Liu
|
||||
<a href="https://git.ericxliu.me/eric/ericxliu-me/commit/3b1396d">[3b1396d]</a></section></footer></main><script src=/js/coder.min.6ae284be93d2d19dad1f02b0039508d9aab3180a12a06dcc71b0b0ef7825a317.js integrity="sha256-auKEvpPS0Z2tHwKwA5UI2aqzGAoSoG3McbCw73gloxc="></script><script defer src=https://static.cloudflareinsights.com/beacon.min.js data-cf-beacon='{"token": "987638e636ce4dbb932d038af74c17d1"}'></script></body></html>
|
||||
<a href="https://git.ericxliu.me/eric/ericxliu-me/commit/f7528b3">[f7528b3]</a></section></footer></main><script src=/js/coder.min.6ae284be93d2d19dad1f02b0039508d9aab3180a12a06dcc71b0b0ef7825a317.js integrity="sha256-auKEvpPS0Z2tHwKwA5UI2aqzGAoSoG3McbCw73gloxc="></script><script defer src=https://static.cloudflareinsights.com/beacon.min.js data-cf-beacon='{"token": "987638e636ce4dbb932d038af74c17d1"}'></script></body></html>
|
||||
@@ -26,4 +26,4 @@ This article explores the mathematical equivalence between Hinton’s concept of
|
||||
2016 -
|
||||
2026
|
||||
Eric X. Liu
|
||||
<a href="https://git.ericxliu.me/eric/ericxliu-me/commit/3b1396d">[3b1396d]</a></section></footer></main><script src=/js/coder.min.6ae284be93d2d19dad1f02b0039508d9aab3180a12a06dcc71b0b0ef7825a317.js integrity="sha256-auKEvpPS0Z2tHwKwA5UI2aqzGAoSoG3McbCw73gloxc="></script><script defer src=https://static.cloudflareinsights.com/beacon.min.js data-cf-beacon='{"token": "987638e636ce4dbb932d038af74c17d1"}'></script></body></html>
|
||||
<a href="https://git.ericxliu.me/eric/ericxliu-me/commit/f7528b3">[f7528b3]</a></section></footer></main><script src=/js/coder.min.6ae284be93d2d19dad1f02b0039508d9aab3180a12a06dcc71b0b0ef7825a317.js integrity="sha256-auKEvpPS0Z2tHwKwA5UI2aqzGAoSoG3McbCw73gloxc="></script><script defer src=https://static.cloudflareinsights.com/beacon.min.js data-cf-beacon='{"token": "987638e636ce4dbb932d038af74c17d1"}'></script></body></html>
|
||||
@@ -8,7 +8,7 @@
|
||||
|
||||
In deep learning, a “channel” can be thought of as a feature dimension. While this term is common in Convolutional Neural Networks for images (e.g., Red, Green, Blue channels), in LLMs, the analogous concept is the model’s primary embedding dimension, commonly referred to as d_model."><meta name=keywords content="software engineer,performance engineering,Google engineer,tech blog,software development,performance optimization,Eric Liu,engineering blog,mountain biking,Jeep enthusiast,overlanding,camping,outdoor adventures"><meta name=twitter:card content="summary"><meta name=twitter:title content="Transformer's Core Mechanics"><meta name=twitter:description content="The Transformer architecture is the bedrock of modern Large Language Models (LLMs). While its high-level success is widely known, a deeper understanding requires dissecting its core components. This article provides a detailed, technical breakdown of the fundamental concepts within a Transformer block, from the notion of “channels” to the intricate workings of the attention mechanism and its relationship with other advanced architectures like Mixture of Experts.
|
||||
1. The “Channel”: A Foundational View of d_model Link to heading In deep learning, a “channel” can be thought of as a feature dimension. While this term is common in Convolutional Neural Networks for images (e.g., Red, Green, Blue channels), in LLMs, the analogous concept is the model’s primary embedding dimension, commonly referred to as d_model."><meta property="og:url" content="https://ericxliu.me/posts/transformer-s-core-mechanics/"><meta property="og:site_name" content="Eric X. Liu's Personal Page"><meta property="og:title" content="Transformer's Core Mechanics"><meta property="og:description" content="The Transformer architecture is the bedrock of modern Large Language Models (LLMs). While its high-level success is widely known, a deeper understanding requires dissecting its core components. This article provides a detailed, technical breakdown of the fundamental concepts within a Transformer block, from the notion of “channels” to the intricate workings of the attention mechanism and its relationship with other advanced architectures like Mixture of Experts.
|
||||
1. The “Channel”: A Foundational View of d_model Link to heading In deep learning, a “channel” can be thought of as a feature dimension. While this term is common in Convolutional Neural Networks for images (e.g., Red, Green, Blue channels), in LLMs, the analogous concept is the model’s primary embedding dimension, commonly referred to as d_model."><meta property="og:locale" content="en"><meta property="og:type" content="article"><meta property="article:section" content="posts"><meta property="article:published_time" content="2025-04-01T00:00:00+00:00"><meta property="article:modified_time" content="2025-10-02T08:42:39+00:00"><link rel=preload href=/fonts/fa-solid-900.woff2 as=font type=font/woff2 crossorigin><link rel=preload href=/fonts/fa-brands-400.woff2 as=font type=font/woff2 crossorigin><link rel=canonical href=https://ericxliu.me/posts/transformer-s-core-mechanics/><link rel=preload href=/fonts/fa-brands-400.woff2 as=font type=font/woff2 crossorigin><link rel=preload href=/fonts/fa-regular-400.woff2 as=font type=font/woff2 crossorigin><link rel=preload href=/fonts/fa-solid-900.woff2 as=font type=font/woff2 crossorigin><link rel=stylesheet href=/css/coder.min.4b392a85107b91dbdabc528edf014a6ab1a30cd44cafcd5325c8efe796794fca.css integrity="sha256-SzkqhRB7kdvavFKO3wFKarGjDNRMr81TJcjv55Z5T8o=" crossorigin=anonymous media=screen><link rel=stylesheet href=/css/coder-dark.min.a00e6364bacbc8266ad1cc81230774a1397198f8cfb7bcba29b7d6fcb54ce57f.css integrity="sha256-oA5jZLrLyCZq0cyBIwd0oTlxmPjPt7y6KbfW/LVM5X8=" crossorigin=anonymous media=screen><link rel=icon type=image/svg+xml href=/images/favicon.svg sizes=any><link rel=icon type=image/png href=/images/favicon-32x32.png sizes=32x32><link rel=icon type=image/png href=/images/favicon-16x16.png sizes=16x16><link rel=apple-touch-icon href=/images/apple-touch-icon.png><link rel=apple-touch-icon sizes=180x180 href=/images/apple-touch-icon.png><link rel=manifest href=/site.webmanifest><link rel=mask-icon href=/images/safari-pinned-tab.svg color=#5bbad5><script async src="https://pagead2.googlesyndication.com/pagead/js/adsbygoogle.js?client=ca-pub-3972604619956476" crossorigin=anonymous></script><script type=application/ld+json>{"@context":"http://schema.org","@type":"Person","name":"Eric X. Liu","url":"https:\/\/ericxliu.me\/","description":"Software \u0026 Performance Engineer at Google","sameAs":["https:\/\/www.linkedin.com\/in\/eric-x-liu-46648b93\/","https:\/\/git.ericxliu.me\/eric"]}</script><script type=application/ld+json>{"@context":"http://schema.org","@type":"BlogPosting","headline":"Transformer\u0027s Core Mechanics","genre":"Blog","wordcount":"1326","url":"https:\/\/ericxliu.me\/posts\/transformer-s-core-mechanics\/","datePublished":"2025-04-01T00:00:00\u002b00:00","dateModified":"2025-10-02T08:42:39\u002b00:00","description":"\u003cp\u003eThe Transformer architecture is the bedrock of modern Large Language Models (LLMs). While its high-level success is widely known, a deeper understanding requires dissecting its core components. This article provides a detailed, technical breakdown of the fundamental concepts within a Transformer block, from the notion of \u0026ldquo;channels\u0026rdquo; to the intricate workings of the attention mechanism and its relationship with other advanced architectures like Mixture of Experts.\u003c\/p\u003e\n\u003ch3 id=\u00221-the-channel-a-foundational-view-of-d_model\u0022\u003e\n 1. The \u0026ldquo;Channel\u0026rdquo;: A Foundational View of \u003ccode\u003ed_model\u003c\/code\u003e\n \u003ca class=\u0022heading-link\u0022 href=\u0022#1-the-channel-a-foundational-view-of-d_model\u0022\u003e\n \u003ci class=\u0022fa-solid fa-link\u0022 aria-hidden=\u0022true\u0022 title=\u0022Link to heading\u0022\u003e\u003c\/i\u003e\n \u003cspan class=\u0022sr-only\u0022\u003eLink to heading\u003c\/span\u003e\n \u003c\/a\u003e\n\u003c\/h3\u003e\n\u003cp\u003eIn deep learning, a \u0026ldquo;channel\u0026rdquo; can be thought of as a feature dimension. While this term is common in Convolutional Neural Networks for images (e.g., Red, Green, Blue channels), in LLMs, the analogous concept is the model\u0026rsquo;s primary embedding dimension, commonly referred to as \u003ccode\u003ed_model\u003c\/code\u003e.\u003c\/p\u003e","author":{"@type":"Person","name":"Eric X. Liu"}}</script></head><body class="preload-transitions colorscheme-auto"><div class=float-container><a id=dark-mode-toggle class=colorscheme-toggle><i class="fa-solid fa-adjust fa-fw" aria-hidden=true></i></a></div><main class=wrapper><nav class=navigation><section class=container><a class=navigation-title href=https://ericxliu.me/>Eric X. Liu's Personal Page
|
||||
1. The “Channel”: A Foundational View of d_model Link to heading In deep learning, a “channel” can be thought of as a feature dimension. While this term is common in Convolutional Neural Networks for images (e.g., Red, Green, Blue channels), in LLMs, the analogous concept is the model’s primary embedding dimension, commonly referred to as d_model."><meta property="og:locale" content="en"><meta property="og:type" content="article"><meta property="article:section" content="posts"><meta property="article:published_time" content="2025-04-01T00:00:00+00:00"><meta property="article:modified_time" content="2026-01-08T18:13:13+00:00"><link rel=preload href=/fonts/fa-solid-900.woff2 as=font type=font/woff2 crossorigin><link rel=preload href=/fonts/fa-brands-400.woff2 as=font type=font/woff2 crossorigin><link rel=canonical href=https://ericxliu.me/posts/transformer-s-core-mechanics/><link rel=preload href=/fonts/fa-brands-400.woff2 as=font type=font/woff2 crossorigin><link rel=preload href=/fonts/fa-regular-400.woff2 as=font type=font/woff2 crossorigin><link rel=preload href=/fonts/fa-solid-900.woff2 as=font type=font/woff2 crossorigin><link rel=stylesheet href=/css/coder.min.4b392a85107b91dbdabc528edf014a6ab1a30cd44cafcd5325c8efe796794fca.css integrity="sha256-SzkqhRB7kdvavFKO3wFKarGjDNRMr81TJcjv55Z5T8o=" crossorigin=anonymous media=screen><link rel=stylesheet href=/css/coder-dark.min.a00e6364bacbc8266ad1cc81230774a1397198f8cfb7bcba29b7d6fcb54ce57f.css integrity="sha256-oA5jZLrLyCZq0cyBIwd0oTlxmPjPt7y6KbfW/LVM5X8=" crossorigin=anonymous media=screen><link rel=icon type=image/svg+xml href=/images/favicon.svg sizes=any><link rel=icon type=image/png href=/images/favicon-32x32.png sizes=32x32><link rel=icon type=image/png href=/images/favicon-16x16.png sizes=16x16><link rel=apple-touch-icon href=/images/apple-touch-icon.png><link rel=apple-touch-icon sizes=180x180 href=/images/apple-touch-icon.png><link rel=manifest href=/site.webmanifest><link rel=mask-icon href=/images/safari-pinned-tab.svg color=#5bbad5><script async src="https://pagead2.googlesyndication.com/pagead/js/adsbygoogle.js?client=ca-pub-3972604619956476" crossorigin=anonymous></script><script type=application/ld+json>{"@context":"http://schema.org","@type":"Person","name":"Eric X. Liu","url":"https:\/\/ericxliu.me\/","description":"Software \u0026 Performance Engineer at Google","sameAs":["https:\/\/www.linkedin.com\/in\/eric-x-liu-46648b93\/","https:\/\/git.ericxliu.me\/eric"]}</script><script type=application/ld+json>{"@context":"http://schema.org","@type":"BlogPosting","headline":"Transformer\u0027s Core Mechanics","genre":"Blog","wordcount":"1326","url":"https:\/\/ericxliu.me\/posts\/transformer-s-core-mechanics\/","datePublished":"2025-04-01T00:00:00\u002b00:00","dateModified":"2026-01-08T18:13:13\u002b00:00","description":"\u003cp\u003eThe Transformer architecture is the bedrock of modern Large Language Models (LLMs). While its high-level success is widely known, a deeper understanding requires dissecting its core components. This article provides a detailed, technical breakdown of the fundamental concepts within a Transformer block, from the notion of \u0026ldquo;channels\u0026rdquo; to the intricate workings of the attention mechanism and its relationship with other advanced architectures like Mixture of Experts.\u003c\/p\u003e\n\u003ch3 id=\u00221-the-channel-a-foundational-view-of-d_model\u0022\u003e\n 1. The \u0026ldquo;Channel\u0026rdquo;: A Foundational View of \u003ccode\u003ed_model\u003c\/code\u003e\n \u003ca class=\u0022heading-link\u0022 href=\u0022#1-the-channel-a-foundational-view-of-d_model\u0022\u003e\n \u003ci class=\u0022fa-solid fa-link\u0022 aria-hidden=\u0022true\u0022 title=\u0022Link to heading\u0022\u003e\u003c\/i\u003e\n \u003cspan class=\u0022sr-only\u0022\u003eLink to heading\u003c\/span\u003e\n \u003c\/a\u003e\n\u003c\/h3\u003e\n\u003cp\u003eIn deep learning, a \u0026ldquo;channel\u0026rdquo; can be thought of as a feature dimension. While this term is common in Convolutional Neural Networks for images (e.g., Red, Green, Blue channels), in LLMs, the analogous concept is the model\u0026rsquo;s primary embedding dimension, commonly referred to as \u003ccode\u003ed_model\u003c\/code\u003e.\u003c\/p\u003e","author":{"@type":"Person","name":"Eric X. Liu"}}</script></head><body class="preload-transitions colorscheme-auto"><div class=float-container><a id=dark-mode-toggle class=colorscheme-toggle><i class="fa-solid fa-adjust fa-fw" aria-hidden=true></i></a></div><main class=wrapper><nav class=navigation><section class=container><a class=navigation-title href=https://ericxliu.me/>Eric X. Liu's Personal Page
|
||||
</a><input type=checkbox id=menu-toggle>
|
||||
<label class="menu-button float-right" for=menu-toggle><i class="fa-solid fa-bars fa-fw" aria-hidden=true></i></label><ul class=navigation-list><li class=navigation-item><a class=navigation-link href=/posts/>Posts</a></li><li class=navigation-item><a class=navigation-link href=https://chat.ericxliu.me>Chat</a></li><li class=navigation-item><a class=navigation-link href=https://git.ericxliu.me/user/oauth2/Authenitk>Git</a></li><li class=navigation-item><a class=navigation-link href=https://coder.ericxliu.me/api/v2/users/oidc/callback>Coder</a></li><li class=navigation-item><a class=navigation-link href=/about/>About</a></li><li class=navigation-item><a class=navigation-link href=/>|</a></li><li class=navigation-item><a class=navigation-link href=https://sso.ericxliu.me>Sign in</a></li></ul></section></nav><div class=content><section class="container post"><article><header><div class=post-title><h1 class=title><a class=title-link href=https://ericxliu.me/posts/transformer-s-core-mechanics/>Transformer's Core Mechanics</a></h1></div><div class=post-meta><div class=date><span class=posted-on><i class="fa-solid fa-calendar" aria-hidden=true></i>
|
||||
<time datetime=2025-04-01T00:00:00Z>April 1, 2025
|
||||
@@ -20,7 +20,7 @@ In deep learning, a “channel” can be thought of as a feature dimensi
|
||||
<span class=sr-only>Link to heading</span></a></h3><p>A Transformer layer is primarily composed of two sub-layers: a Multi-Head Attention block and a position-wise Feed-Forward Network (FFN). The parameters for these are stored in several key weight matrices. Understanding their dimensions is crucial.</p><p>Let’s define our variables:</p><ul><li><code>d_model</code>: The core embedding dimension.</li><li><code>d_ff</code>: The inner dimension of the FFN, typically <code>4 * d_model</code>.</li><li><code>h</code>: The number of attention heads.</li><li><code>d_head</code>: The dimension of each attention head, where <code>d_model = h * d_head</code>.</li></ul><p>The dimensions of the weight matrices are as follows:</p><table><thead><tr><th>Layer</th><th>Weight Matrix</th><th>Input Vector Shape</th><th>Output Vector Shape</th><th><strong>Weight Matrix Dimension</strong></th></tr></thead><tbody><tr><td><strong>Attention Projections</strong></td><td></td><td></td><td></td><td></td></tr><tr><td>Query</td><td><code>W_Q</code></td><td><code>d_model</code></td><td><code>d_model</code></td><td><strong><code>[d_model, d_model]</code></strong></td></tr><tr><td>Key</td><td><code>W_K</code></td><td><code>d_model</code></td><td><code>d_model</code></td><td><strong><code>[d_model, d_model]</code></strong></td></tr><tr><td>Value</td><td><code>W_V</code></td><td><code>d_model</code></td><td><code>d_model</code></td><td><strong><code>[d_model, d_model]</code></strong></td></tr><tr><td>Output</td><td><code>W_O</code></td><td><code>d_model</code></td><td><code>d_model</code></td><td><strong><code>[d_model, d_model]</code></strong></td></tr><tr><td><strong>Feed-Forward Network</strong></td><td></td><td></td><td></td><td></td></tr><tr><td>Layer 1 (Up-projection)</td><td><code>W_ff1</code></td><td><code>d_model</code></td><td><code>d_ff</code></td><td><strong><code>[d_model, d_ff]</code></strong></td></tr><tr><td>Layer 2 (Down-projection)</td><td><code>W_ff2</code></td><td><code>d_ff</code></td><td><code>d_model</code></td><td><strong><code>[d_ff, d_model]</code></strong></td></tr></tbody></table><h3 id=3-deconstructing-multi-head-attention-mha>3. Deconstructing Multi-Head Attention (MHA)
|
||||
<a class=heading-link href=#3-deconstructing-multi-head-attention-mha><i class="fa-solid fa-link" aria-hidden=true title="Link to heading"></i>
|
||||
<span class=sr-only>Link to heading</span></a></h3><p>The core innovation of the Transformer is Multi-Head Attention. It allows the model to weigh the importance of different tokens in the sequence from multiple perspectives simultaneously.
|
||||
<img src=/images/transformer-s-core-mechanics/c7fe4af2633840cfbc81d7c4e3e42d0c.png alt="S3 File"></p><h4 id=31-the-why-beyond-a-single-attention>3.1. The “Why”: Beyond a Single Attention
|
||||
<img src="http://localhost:4998/attachments/image-c64b0f9df1e4981c4ecdb3b60e8bc78c426ffa68.png?client=default&bucket=obsidian" alt="S3 File"></p><h4 id=31-the-why-beyond-a-single-attention>3.1. The “Why”: Beyond a Single Attention
|
||||
<a class=heading-link href=#31-the-why-beyond-a-single-attention><i class="fa-solid fa-link" aria-hidden=true title="Link to heading"></i>
|
||||
<span class=sr-only>Link to heading</span></a></h4><p>A single attention mechanism would force the model to average all types of linguistic relationships into one pattern. MHA avoids this by creating <code>h</code> parallel subspaces. Each “head” can specialize, with one head learning syntactic dependencies, another tracking semantic similarity, and so on. This creates a much richer representation.</p><h4 id=32-an-encodingdecoding-analogy>3.2. An Encoding/Decoding Analogy
|
||||
<a class=heading-link href=#32-an-encodingdecoding-analogy><i class="fa-solid fa-link" aria-hidden=true title="Link to heading"></i>
|
||||
@@ -36,4 +36,4 @@ In deep learning, a “channel” can be thought of as a feature dimensi
|
||||
2016 -
|
||||
2026
|
||||
Eric X. Liu
|
||||
<a href="https://git.ericxliu.me/eric/ericxliu-me/commit/3b1396d">[3b1396d]</a></section></footer></main><script src=/js/coder.min.6ae284be93d2d19dad1f02b0039508d9aab3180a12a06dcc71b0b0ef7825a317.js integrity="sha256-auKEvpPS0Z2tHwKwA5UI2aqzGAoSoG3McbCw73gloxc="></script><script defer src=https://static.cloudflareinsights.com/beacon.min.js data-cf-beacon='{"token": "987638e636ce4dbb932d038af74c17d1"}'></script></body></html>
|
||||
<a href="https://git.ericxliu.me/eric/ericxliu-me/commit/f7528b3">[f7528b3]</a></section></footer></main><script src=/js/coder.min.6ae284be93d2d19dad1f02b0039508d9aab3180a12a06dcc71b0b0ef7825a317.js integrity="sha256-auKEvpPS0Z2tHwKwA5UI2aqzGAoSoG3McbCw73gloxc="></script><script defer src=https://static.cloudflareinsights.com/beacon.min.js data-cf-beacon='{"token": "987638e636ce4dbb932d038af74c17d1"}'></script></body></html>
|
||||
@@ -1,7 +1,7 @@
|
||||
<!doctype html><html lang=en><head><title>UniFi VLAN Migration to Zone-Based Architecture · Eric X. Liu's Personal Page</title><meta charset=utf-8><meta name=viewport content="width=device-width,initial-scale=1"><meta name=color-scheme content="light dark"><meta http-equiv=Content-Security-Policy content="upgrade-insecure-requests; block-all-mixed-content; default-src 'self'; child-src 'self'; font-src 'self' https://fonts.gstatic.com https://cdn.jsdelivr.net/; form-action 'self'; frame-src 'self' https://www.youtube.com https://disqus.com; img-src 'self' https://referrer.disqus.com https://c.disquscdn.com https://*.disqus.com; object-src 'none'; style-src 'self' 'unsafe-inline' https://fonts.googleapis.com/ https://cdn.jsdelivr.net/; script-src 'self' 'unsafe-inline' https://www.google-analytics.com https://cdn.jsdelivr.net/ https://pagead2.googlesyndication.com https://static.cloudflareinsights.com https://unpkg.com https://ericxliu-me.disqus.com https://disqus.com https://*.disqus.com https://*.disquscdn.com https://unpkg.com; connect-src 'self' https://www.google-analytics.com https://pagead2.googlesyndication.com https://cloudflareinsights.com ws://localhost:1313 ws://localhost:* wss://localhost:* https://links.services.disqus.com https://*.disqus.com;"><meta name=author content="Eric X. Liu"><meta name=description content="Embarking on a network migration to a properly segmented VLAN architecture is a rite of passage for any serious home lab or small business operator. The goal is clear: improve security and organization by separating traffic. However, the path from a flat network to a segmented one is often paved with subtle but critical configuration details that can lead to hours of frustrating troubleshooting.
|
||||
This article documents that journey. It details the pitfalls encountered, the core networking concepts that were essential to understand, and the best practices that ultimately led to a stable, secure, and logical network design built on a zone-based firewall model."><meta name=keywords content="software engineer,performance engineering,Google engineer,tech blog,software development,performance optimization,Eric Liu,engineering blog,mountain biking,Jeep enthusiast,overlanding,camping,outdoor adventures"><meta name=twitter:card content="summary"><meta name=twitter:title content="UniFi VLAN Migration to Zone-Based Architecture"><meta name=twitter:description content="Embarking on a network migration to a properly segmented VLAN architecture is a rite of passage for any serious home lab or small business operator. The goal is clear: improve security and organization by separating traffic. However, the path from a flat network to a segmented one is often paved with subtle but critical configuration details that can lead to hours of frustrating troubleshooting.
|
||||
This article documents that journey. It details the pitfalls encountered, the core networking concepts that were essential to understand, and the best practices that ultimately led to a stable, secure, and logical network design built on a zone-based firewall model."><meta property="og:url" content="https://ericxliu.me/posts/unifi-vlan-migration-to-zone-based-architecture/"><meta property="og:site_name" content="Eric X. Liu's Personal Page"><meta property="og:title" content="UniFi VLAN Migration to Zone-Based Architecture"><meta property="og:description" content="Embarking on a network migration to a properly segmented VLAN architecture is a rite of passage for any serious home lab or small business operator. The goal is clear: improve security and organization by separating traffic. However, the path from a flat network to a segmented one is often paved with subtle but critical configuration details that can lead to hours of frustrating troubleshooting.
|
||||
This article documents that journey. It details the pitfalls encountered, the core networking concepts that were essential to understand, and the best practices that ultimately led to a stable, secure, and logical network design built on a zone-based firewall model."><meta property="og:locale" content="en"><meta property="og:type" content="article"><meta property="article:section" content="posts"><meta property="article:published_time" content="2025-09-22T00:00:00+00:00"><meta property="article:modified_time" content="2025-10-02T08:42:39+00:00"><link rel=preload href=/fonts/fa-solid-900.woff2 as=font type=font/woff2 crossorigin><link rel=preload href=/fonts/fa-brands-400.woff2 as=font type=font/woff2 crossorigin><link rel=canonical href=https://ericxliu.me/posts/unifi-vlan-migration-to-zone-based-architecture/><link rel=preload href=/fonts/fa-brands-400.woff2 as=font type=font/woff2 crossorigin><link rel=preload href=/fonts/fa-regular-400.woff2 as=font type=font/woff2 crossorigin><link rel=preload href=/fonts/fa-solid-900.woff2 as=font type=font/woff2 crossorigin><link rel=stylesheet href=/css/coder.min.4b392a85107b91dbdabc528edf014a6ab1a30cd44cafcd5325c8efe796794fca.css integrity="sha256-SzkqhRB7kdvavFKO3wFKarGjDNRMr81TJcjv55Z5T8o=" crossorigin=anonymous media=screen><link rel=stylesheet href=/css/coder-dark.min.a00e6364bacbc8266ad1cc81230774a1397198f8cfb7bcba29b7d6fcb54ce57f.css integrity="sha256-oA5jZLrLyCZq0cyBIwd0oTlxmPjPt7y6KbfW/LVM5X8=" crossorigin=anonymous media=screen><link rel=icon type=image/svg+xml href=/images/favicon.svg sizes=any><link rel=icon type=image/png href=/images/favicon-32x32.png sizes=32x32><link rel=icon type=image/png href=/images/favicon-16x16.png sizes=16x16><link rel=apple-touch-icon href=/images/apple-touch-icon.png><link rel=apple-touch-icon sizes=180x180 href=/images/apple-touch-icon.png><link rel=manifest href=/site.webmanifest><link rel=mask-icon href=/images/safari-pinned-tab.svg color=#5bbad5><script async src="https://pagead2.googlesyndication.com/pagead/js/adsbygoogle.js?client=ca-pub-3972604619956476" crossorigin=anonymous></script><script type=application/ld+json>{"@context":"http://schema.org","@type":"Person","name":"Eric X. Liu","url":"https:\/\/ericxliu.me\/","description":"Software \u0026 Performance Engineer at Google","sameAs":["https:\/\/www.linkedin.com\/in\/eric-x-liu-46648b93\/","https:\/\/git.ericxliu.me\/eric"]}</script><script type=application/ld+json>{"@context":"http://schema.org","@type":"BlogPosting","headline":"UniFi VLAN Migration to Zone-Based Architecture","genre":"Blog","wordcount":"1001","url":"https:\/\/ericxliu.me\/posts\/unifi-vlan-migration-to-zone-based-architecture\/","datePublished":"2025-09-22T00:00:00\u002b00:00","dateModified":"2025-10-02T08:42:39\u002b00:00","description":"\u003cp\u003eEmbarking on a network migration to a properly segmented VLAN architecture is a rite of passage for any serious home lab or small business operator. The goal is clear: improve security and organization by separating traffic. However, the path from a flat network to a segmented one is often paved with subtle but critical configuration details that can lead to hours of frustrating troubleshooting.\u003c\/p\u003e\n\u003cp\u003eThis article documents that journey. It details the pitfalls encountered, the core networking concepts that were essential to understand, and the best practices that ultimately led to a stable, secure, and logical network design built on a zone-based firewall model.\u003c\/p\u003e","author":{"@type":"Person","name":"Eric X. Liu"}}</script></head><body class="preload-transitions colorscheme-auto"><div class=float-container><a id=dark-mode-toggle class=colorscheme-toggle><i class="fa-solid fa-adjust fa-fw" aria-hidden=true></i></a></div><main class=wrapper><nav class=navigation><section class=container><a class=navigation-title href=https://ericxliu.me/>Eric X. Liu's Personal Page
|
||||
This article documents that journey. It details the pitfalls encountered, the core networking concepts that were essential to understand, and the best practices that ultimately led to a stable, secure, and logical network design built on a zone-based firewall model."><meta property="og:locale" content="en"><meta property="og:type" content="article"><meta property="article:section" content="posts"><meta property="article:published_time" content="2025-09-22T00:00:00+00:00"><meta property="article:modified_time" content="2026-01-08T18:13:13+00:00"><link rel=preload href=/fonts/fa-solid-900.woff2 as=font type=font/woff2 crossorigin><link rel=preload href=/fonts/fa-brands-400.woff2 as=font type=font/woff2 crossorigin><link rel=canonical href=https://ericxliu.me/posts/unifi-vlan-migration-to-zone-based-architecture/><link rel=preload href=/fonts/fa-brands-400.woff2 as=font type=font/woff2 crossorigin><link rel=preload href=/fonts/fa-regular-400.woff2 as=font type=font/woff2 crossorigin><link rel=preload href=/fonts/fa-solid-900.woff2 as=font type=font/woff2 crossorigin><link rel=stylesheet href=/css/coder.min.4b392a85107b91dbdabc528edf014a6ab1a30cd44cafcd5325c8efe796794fca.css integrity="sha256-SzkqhRB7kdvavFKO3wFKarGjDNRMr81TJcjv55Z5T8o=" crossorigin=anonymous media=screen><link rel=stylesheet href=/css/coder-dark.min.a00e6364bacbc8266ad1cc81230774a1397198f8cfb7bcba29b7d6fcb54ce57f.css integrity="sha256-oA5jZLrLyCZq0cyBIwd0oTlxmPjPt7y6KbfW/LVM5X8=" crossorigin=anonymous media=screen><link rel=icon type=image/svg+xml href=/images/favicon.svg sizes=any><link rel=icon type=image/png href=/images/favicon-32x32.png sizes=32x32><link rel=icon type=image/png href=/images/favicon-16x16.png sizes=16x16><link rel=apple-touch-icon href=/images/apple-touch-icon.png><link rel=apple-touch-icon sizes=180x180 href=/images/apple-touch-icon.png><link rel=manifest href=/site.webmanifest><link rel=mask-icon href=/images/safari-pinned-tab.svg color=#5bbad5><script async src="https://pagead2.googlesyndication.com/pagead/js/adsbygoogle.js?client=ca-pub-3972604619956476" crossorigin=anonymous></script><script type=application/ld+json>{"@context":"http://schema.org","@type":"Person","name":"Eric X. Liu","url":"https:\/\/ericxliu.me\/","description":"Software \u0026 Performance Engineer at Google","sameAs":["https:\/\/www.linkedin.com\/in\/eric-x-liu-46648b93\/","https:\/\/git.ericxliu.me\/eric"]}</script><script type=application/ld+json>{"@context":"http://schema.org","@type":"BlogPosting","headline":"UniFi VLAN Migration to Zone-Based Architecture","genre":"Blog","wordcount":"1001","url":"https:\/\/ericxliu.me\/posts\/unifi-vlan-migration-to-zone-based-architecture\/","datePublished":"2025-09-22T00:00:00\u002b00:00","dateModified":"2026-01-08T18:13:13\u002b00:00","description":"\u003cp\u003eEmbarking on a network migration to a properly segmented VLAN architecture is a rite of passage for any serious home lab or small business operator. The goal is clear: improve security and organization by separating traffic. However, the path from a flat network to a segmented one is often paved with subtle but critical configuration details that can lead to hours of frustrating troubleshooting.\u003c\/p\u003e\n\u003cp\u003eThis article documents that journey. It details the pitfalls encountered, the core networking concepts that were essential to understand, and the best practices that ultimately led to a stable, secure, and logical network design built on a zone-based firewall model.\u003c\/p\u003e","author":{"@type":"Person","name":"Eric X. Liu"}}</script></head><body class="preload-transitions colorscheme-auto"><div class=float-container><a id=dark-mode-toggle class=colorscheme-toggle><i class="fa-solid fa-adjust fa-fw" aria-hidden=true></i></a></div><main class=wrapper><nav class=navigation><section class=container><a class=navigation-title href=https://ericxliu.me/>Eric X. Liu's Personal Page
|
||||
</a><input type=checkbox id=menu-toggle>
|
||||
<label class="menu-button float-right" for=menu-toggle><i class="fa-solid fa-bars fa-fw" aria-hidden=true></i></label><ul class=navigation-list><li class=navigation-item><a class=navigation-link href=/posts/>Posts</a></li><li class=navigation-item><a class=navigation-link href=https://chat.ericxliu.me>Chat</a></li><li class=navigation-item><a class=navigation-link href=https://git.ericxliu.me/user/oauth2/Authenitk>Git</a></li><li class=navigation-item><a class=navigation-link href=https://coder.ericxliu.me/api/v2/users/oidc/callback>Coder</a></li><li class=navigation-item><a class=navigation-link href=/about/>About</a></li><li class=navigation-item><a class=navigation-link href=/>|</a></li><li class=navigation-item><a class=navigation-link href=https://sso.ericxliu.me>Sign in</a></li></ul></section></nav><div class=content><section class="container post"><article><header><div class=post-title><h1 class=title><a class=title-link href=https://ericxliu.me/posts/unifi-vlan-migration-to-zone-based-architecture/>UniFi VLAN Migration to Zone-Based Architecture</a></h1></div><div class=post-meta><div class=date><span class=posted-on><i class="fa-solid fa-calendar" aria-hidden=true></i>
|
||||
<time datetime=2025-09-22T00:00:00Z>September 22, 2025
|
||||
@@ -17,10 +17,10 @@ This article documents that journey. It details the pitfalls encountered, the co
|
||||
<span class=sr-only>Link to heading</span></a></h3><p>The culmination of these lessons is a network architecture defined by clear, logical zones, each with a distinct purpose and trust level. This model simplifies firewall management and provides a robust security posture that is easy to understand at a glance.</p><h4 id=network-zones-and-their-roles>Network Zones and Their Roles
|
||||
<a class=heading-link href=#network-zones-and-their-roles><i class="fa-solid fa-link" aria-hidden=true title="Link to heading"></i>
|
||||
<span class=sr-only>Link to heading</span></a></h4><p>The final configuration groups the individual VLANs into distinct zones, forming the foundation of the security policy.</p><ul><li><strong>Internal:</strong> Contains the <code>corp</code> network. This is the most trusted zone for daily work.</li><li><strong>DMZ:</strong> Contains the <code>dns</code> and <code>prod</code> networks for semi-trusted, exposed services.</li><li><strong>IoT:</strong> Contains the <code>iot</code> network. This is a low-trust zone for smart devices.</li><li><strong>Management:</strong> Contains the <code>management</code> network. This is a highly privileged, isolated zone for network infrastructure.
|
||||
<img src=/images/unifi-vlan-migration-to-zone-based-architecture/472bf0cd504f4cd7ab7a33cd3322a5f1.png alt="S3 File"></li></ul><h4 id=the-security-policy-matrix>The Security Policy Matrix
|
||||
<img src="http://localhost:4998/attachments/image-167d5cef9e79e622fff779f3671492a8a5a343ea.png?client=default&bucket=obsidian" alt="S3 File"></li></ul><h4 id=the-security-policy-matrix>The Security Policy Matrix
|
||||
<a class=heading-link href=#the-security-policy-matrix><i class="fa-solid fa-link" aria-hidden=true title="Link to heading"></i>
|
||||
<span class=sr-only>Link to heading</span></a></h4><p>The true power of this model is realized in the firewall’s zone matrix, which dictates the default traffic flow between each zone.
|
||||
<img src=/images/unifi-vlan-migration-to-zone-based-architecture/663d732d14fc4fa8ad051c6926523efb.png alt="S3 File"></p><p>This matrix enforces the desired security policy with clear, high-level rules:</p><ul><li><strong>Complete IoT Isolation:</strong> The <code>IoT</code> row shows that devices in this zone are blocked from initiating any communication with any other internal zone. Their only allowed path is out to the internet.</li><li><strong>Protected Management Plane:</strong> The <code>management</code> row and column are almost entirely red. The critical network infrastructure is blocked from initiating contact with any user-facing zone, and vice-versa, following the principle of least privilege.</li><li><strong>Controlled DMZ Access:</strong> The <code>DMZ</code> is prevented from initiating connections to the trusted <code>Internal</code> zone, preventing a compromised public-facing server from being used as a pivot point to attack internal devices.</li></ul><h4 id=granular-intra-zone-control>Granular Intra-Zone Control
|
||||
<img src="http://localhost:4998/attachments/image-4b9dbea5f7ceb0446d517305bc281b74e7f22ffc.png?client=default&bucket=obsidian" alt="S3 File"></p><p>This matrix enforces the desired security policy with clear, high-level rules:</p><ul><li><strong>Complete IoT Isolation:</strong> The <code>IoT</code> row shows that devices in this zone are blocked from initiating any communication with any other internal zone. Their only allowed path is out to the internet.</li><li><strong>Protected Management Plane:</strong> The <code>management</code> row and column are almost entirely red. The critical network infrastructure is blocked from initiating contact with any user-facing zone, and vice-versa, following the principle of least privilege.</li><li><strong>Controlled DMZ Access:</strong> The <code>DMZ</code> is prevented from initiating connections to the trusted <code>Internal</code> zone, preventing a compromised public-facing server from being used as a pivot point to attack internal devices.</li></ul><h4 id=granular-intra-zone-control>Granular Intra-Zone Control
|
||||
<a class=heading-link href=#granular-intra-zone-control><i class="fa-solid fa-link" aria-hidden=true title="Link to heading"></i>
|
||||
<span class=sr-only>Link to heading</span></a></h4><p>Beyond the high-level zone policies, the configuration also implements granular rules to control traffic <em>within</em> a single zone, providing defense-in-depth.</p><p>These rules explicitly define the communication paths between services. For instance, rules allow a specific device to access a Kubernetes load balancer while another rule allows general DNS access within the zone. This ensures that even within a semi-trusted zone, services can only communicate in expected and necessary ways, further reducing the potential attack surface.</p><p>By adhering to these principles, what began as a day of frustrating troubleshooting evolved into a robust, layered, and logically segmented network that balances simplicity with strong security practices.</p><hr><h3 id=references>References
|
||||
<a class=heading-link href=#references><i class="fa-solid fa-link" aria-hidden=true title="Link to heading"></i>
|
||||
@@ -28,4 +28,4 @@ This article documents that journey. It details the pitfalls encountered, the co
|
||||
2016 -
|
||||
2026
|
||||
Eric X. Liu
|
||||
<a href="https://git.ericxliu.me/eric/ericxliu-me/commit/3b1396d">[3b1396d]</a></section></footer></main><script src=/js/coder.min.6ae284be93d2d19dad1f02b0039508d9aab3180a12a06dcc71b0b0ef7825a317.js integrity="sha256-auKEvpPS0Z2tHwKwA5UI2aqzGAoSoG3McbCw73gloxc="></script><script defer src=https://static.cloudflareinsights.com/beacon.min.js data-cf-beacon='{"token": "987638e636ce4dbb932d038af74c17d1"}'></script></body></html>
|
||||
<a href="https://git.ericxliu.me/eric/ericxliu-me/commit/f7528b3">[f7528b3]</a></section></footer></main><script src=/js/coder.min.6ae284be93d2d19dad1f02b0039508d9aab3180a12a06dcc71b0b0ef7825a317.js integrity="sha256-auKEvpPS0Z2tHwKwA5UI2aqzGAoSoG3McbCw73gloxc="></script><script defer src=https://static.cloudflareinsights.com/beacon.min.js data-cf-beacon='{"token": "987638e636ce4dbb932d038af74c17d1"}'></script></body></html>
|
||||
@@ -9,4 +9,4 @@ One-minute read</span></div></div></header><div class=post-content><ul><li><a hr
|
||||
2016 -
|
||||
2026
|
||||
Eric X. Liu
|
||||
<a href="https://git.ericxliu.me/eric/ericxliu-me/commit/3b1396d">[3b1396d]</a></section></footer></main><script src=/js/coder.min.6ae284be93d2d19dad1f02b0039508d9aab3180a12a06dcc71b0b0ef7825a317.js integrity="sha256-auKEvpPS0Z2tHwKwA5UI2aqzGAoSoG3McbCw73gloxc="></script><script defer src=https://static.cloudflareinsights.com/beacon.min.js data-cf-beacon='{"token": "987638e636ce4dbb932d038af74c17d1"}'></script></body></html>
|
||||
<a href="https://git.ericxliu.me/eric/ericxliu-me/commit/f7528b3">[f7528b3]</a></section></footer></main><script src=/js/coder.min.6ae284be93d2d19dad1f02b0039508d9aab3180a12a06dcc71b0b0ef7825a317.js integrity="sha256-auKEvpPS0Z2tHwKwA5UI2aqzGAoSoG3McbCw73gloxc="></script><script defer src=https://static.cloudflareinsights.com/beacon.min.js data-cf-beacon='{"token": "987638e636ce4dbb932d038af74c17d1"}'></script></body></html>
|
||||
@@ -31,4 +31,4 @@ The GPU TLB hierarchy is sensitive to page sizes.</p><ul><li><strong>4KB Pages:<
|
||||
2016 -
|
||||
2026
|
||||
Eric X. Liu
|
||||
<a href="https://git.ericxliu.me/eric/ericxliu-me/commit/3b1396d">[3b1396d]</a></section></footer></main><script src=/js/coder.min.6ae284be93d2d19dad1f02b0039508d9aab3180a12a06dcc71b0b0ef7825a317.js integrity="sha256-auKEvpPS0Z2tHwKwA5UI2aqzGAoSoG3McbCw73gloxc="></script><script defer src=https://static.cloudflareinsights.com/beacon.min.js data-cf-beacon='{"token": "987638e636ce4dbb932d038af74c17d1"}'></script></body></html>
|
||||
<a href="https://git.ericxliu.me/eric/ericxliu-me/commit/f7528b3">[f7528b3]</a></section></footer></main><script src=/js/coder.min.6ae284be93d2d19dad1f02b0039508d9aab3180a12a06dcc71b0b0ef7825a317.js integrity="sha256-auKEvpPS0Z2tHwKwA5UI2aqzGAoSoG3McbCw73gloxc="></script><script defer src=https://static.cloudflareinsights.com/beacon.min.js data-cf-beacon='{"token": "987638e636ce4dbb932d038af74c17d1"}'></script></body></html>
|
||||
Reference in New Issue
Block a user