README / index.html
ZTFlynn's picture
Regenerate family page: footprint derived from each package, not hardcoded
0a1a91a verified
Raw History Blame Contribute Delete
4.74 kB
<!doctype html><meta charset="utf-8">
<title>Cascadia — ternary-3 packages</title>
<meta name="viewport" content="width=device-width,initial-scale=1">
<style>
:root{--fg:#1a1a1a;--bg:#fff;--mut:#666;--line:#e3e3e3;--acc:#0b6}
@media(prefers-color-scheme:dark){:root{--fg:#e8e8e8;--bg:#111;--mut:#9a9a9a;--line:#2c2c2c;--acc:#3d9}}
body{background:var(--bg);color:var(--fg);font:16px/1.6 -apple-system,BlinkMacSystemFont,"Segoe UI",system-ui,sans-serif;
max-width:980px;margin:0 auto;padding:2.5rem 1.25rem}
h1{font-size:1.9rem;margin:0 0 .3rem} h2{font-size:1.2rem;margin:2.4rem 0 .6rem}
.sub{color:var(--mut);margin:0 0 2rem}
table{border-collapse:collapse;width:100%;font-size:.9rem;margin:.5rem 0 1rem}
th,td{padding:.5rem .6rem;border-bottom:1px solid var(--line);text-align:right;white-space:nowrap}
th:first-child,td:first-child{text-align:left}
th{font-weight:600;color:var(--mut);font-size:.8rem;text-transform:uppercase;letter-spacing:.04em}
a{color:var(--acc);text-decoration:none} a:hover{text-decoration:underline}
code,pre{font-family:ui-monospace,SFMono-Regular,Menlo,monospace}
pre{background:rgba(128,128,128,.1);padding:.9rem;border-radius:6px;overflow-x:auto;font-size:.85rem}
.wrap{overflow-x:auto} .note{color:var(--mut);font-size:.85rem}
strong{font-weight:650}
</style>
<h1>Cascadia — ternary-3 packages</h1>
<p class="sub">7 models from the LFM2 and LFM2.5 families at <strong>0.64-0.69 bytes per weight</strong>, executable on CPU by a C runtime whose entire dependency list is libc, libm and libgomp.</p>
<div class="wrap"><table><tr><th>model</th><th>package</th><th>RAM to run</th><th>compression</th><th>perplexity</th><th>rel L2</th></tr>
<tr><td><a href="https://huggingface.co/ZTFlynn/LFM2.5-230M-Cascadia-ternary3">LFM2.5-230M</a></td><td>161 MB</td><td><strong>198 MB</strong></td><td>2.89x</td><td>+8.43%</td><td>0.0485</td></tr>
<tr><td><a href="https://huggingface.co/ZTFlynn/LFM2-350M-Cascadia-ternary3">LFM2-350M</a></td><td>241 MB</td><td><strong>294 MB</strong></td><td>2.96x</td><td>+1.10%*</td><td>0.0533</td></tr>
<tr><td><a href="https://huggingface.co/ZTFlynn/LFM2-700M-Cascadia-ternary3">LFM2-700M</a></td><td>495 MB</td><td><strong>573 MB</strong></td><td>3.01x</td><td>+12.51%</td><td>0.0547</td></tr>
<tr><td><a href="https://huggingface.co/ZTFlynn/LFM2-1.2B-Cascadia-ternary3">LFM2-1.2B</a></td><td>774 MB</td><td><strong>873 MB</strong></td><td>3.03x</td><td>+9.91%</td><td>0.0553</td></tr>
<tr><td><a href="https://huggingface.co/ZTFlynn/LFM2.5-2.6B-Cascadia-ternary3">LFM2.5-2.6B</a></td><td>1.77 GB</td><td><strong>2.00 GB</strong></td><td>3.05x</td><td>-2.65%</td><td>0.0544</td></tr>
<tr><td><a href="https://huggingface.co/ZTFlynn/LFM2.5-8B-A1B-Cascadia-ternary3">LFM2.5-8B-A1B</a></td><td>5.49 GB</td><td><strong>6.33 GB</strong></td><td>3.09x</td><td>+2.58%*</td><td>0.0548</td></tr>
<tr><td><a href="https://huggingface.co/ZTFlynn/LFM2-24B-A2B-Cascadia-ternary3">LFM2-24B-A2B</a></td><td>15.37 GB</td><td><strong>17.84 GB</strong></td><td>3.10x</td><td>none detected</td><td>0.0561</td></tr>
</table></div>
<p class="note">* resolvable individually but not after correcting for 7 simultaneous tests, which needs |t| &gt; 2.89.</p>
<h2>Two numbers, because they answer different questions</h2>
<p><strong>Relative L2</strong> is how faithfully the weights are reproduced — <strong>0.049–0.056</strong> across every model here, holding steady across four tokenizers, dense and Mixture-of-Experts architectures, and a 100× range in parameter count. It is the figure that describes the compression.</p>
<p><strong>Perplexity</strong> is how a model scores on a corpus, measured against the bf16 checkpoint on identical tokens and clustered by 512-token window. 1 of these show no detectable cost. It varies with each model’s sensitivity and with how well general web text matches its domain — two models here compressed to the same 0.055 rel L2 measure +13.4% and −3.2%.</p>
<h2>Running one</h2><pre>git clone https://github.com/EntroMorphic/cassie &amp;&amp; cd cassie
cmake -S src/c -B build -DCMAKE_BUILD_TYPE=Release &amp;&amp; cmake --build build -j
huggingface-cli download ZTFlynn/LFM2-350M-Cascadia-ternary3 --local-dir ./pkg
./build/cascadia_generate ./pkg 512 --chat "Explain gradient descent."</pre>
<h2>Acknowledgements</h2><p>Deeply inspired by <a href="https://huggingface.co/Magneato/deepseek-r1-qwen-7b-lutc">Magneato/deepseek-r1-qwen-7b-lutc</a>, which demonstrated LUT-cascade compression at 5.45 bits per weight. Base models by <a href="https://huggingface.co/LiquidAI">Liquid AI</a>, used under the LFM Open License. Method and measurements: <a href="https://github.com/EntroMorphic/cassie">github.com/EntroMorphic/cassie</a>.</p>