Deployed 2536c93 with MkDocs version: 1.6.1

2026-05-03 11:47:30 +08:00
commit 1ff86b66fc
418 changed files with 617336 additions and 0 deletions
@@ -0,0 +1,68 @@
+<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 800 280" width="800" height="280">
+  <text x="400" y="22" text-anchor="middle" font-family="Arial, sans-serif" font-size="14" font-weight="bold" fill="#333">Quantisation Granularity: Per-Tensor vs Per-Channel vs Per-Group</text>
+
+  <!-- Per-tensor -->
+  <text x="130" y="52" text-anchor="middle" font-family="Arial, sans-serif" font-size="11" font-weight="bold" fill="#e74c3c">Per-Tensor</text>
+  <text x="130" y="66" text-anchor="middle" font-family="Arial, sans-serif" font-size="9" fill="#e74c3c">1 scale for entire matrix</text>
+
+  <rect x="50" y="75" width="160" height="120" rx="4" fill="#e74c3c" fill-opacity="0.12" stroke="#e74c3c" stroke-width="2"/>
+  <text x="130" y="140" text-anchor="middle" font-family="Arial, sans-serif" font-size="10" fill="#e74c3c">s = 0.031</text>
+  <text x="130" y="210" text-anchor="middle" font-family="Arial, sans-serif" font-size="8" fill="#666">one outlier forces large scale</text>
+  <text x="130" y="222" text-anchor="middle" font-family="Arial, sans-serif" font-size="8" fill="#666">→ poor resolution everywhere</text>
+
+  <!-- Outlier dot -->
+  <circle cx="180" cy="90" r="5" fill="#e74c3c" fill-opacity="0.6"/>
+  <text x="192" y="93" font-family="Arial, sans-serif" font-size="7" fill="#e74c3c">outlier!</text>
+
+  <!-- Per-channel -->
+  <text x="400" y="52" text-anchor="middle" font-family="Arial, sans-serif" font-size="11" font-weight="bold" fill="#3498db">Per-Channel</text>
+  <text x="400" y="66" text-anchor="middle" font-family="Arial, sans-serif" font-size="9" fill="#3498db">1 scale per column (output channel)</text>
+
+  <rect x="320" y="75" width="40" height="120" rx="2" fill="#3498db" fill-opacity="0.15" stroke="#3498db" stroke-width="1.5"/>
+  <text x="340" y="140" text-anchor="middle" font-family="Arial, sans-serif" font-size="7" fill="#3498db">s₁</text>
+
+  <rect x="360" y="75" width="40" height="120" rx="2" fill="#27ae60" fill-opacity="0.15" stroke="#27ae60" stroke-width="1.5"/>
+  <text x="380" y="140" text-anchor="middle" font-family="Arial, sans-serif" font-size="7" fill="#27ae60">s₂</text>
+
+  <rect x="400" y="75" width="40" height="120" rx="2" fill="#f39c12" fill-opacity="0.15" stroke="#f39c12" stroke-width="1.5"/>
+  <text x="420" y="140" text-anchor="middle" font-family="Arial, sans-serif" font-size="7" fill="#f39c12">s₃</text>
+
+  <rect x="440" y="75" width="40" height="120" rx="2" fill="#9b59b6" fill-opacity="0.15" stroke="#9b59b6" stroke-width="1.5"/>
+  <text x="460" y="140" text-anchor="middle" font-family="Arial, sans-serif" font-size="7" fill="#9b59b6">s₄</text>
+
+  <text x="400" y="210" text-anchor="middle" font-family="Arial, sans-serif" font-size="8" fill="#666">each channel has its own scale</text>
+  <text x="400" y="222" text-anchor="middle" font-family="Arial, sans-serif" font-size="8" fill="#666">→ outlier affects only its channel</text>
+
+  <!-- Per-group -->
+  <text x="660" y="52" text-anchor="middle" font-family="Arial, sans-serif" font-size="11" font-weight="bold" fill="#27ae60">Per-Group (g=4)</text>
+  <text x="660" y="66" text-anchor="middle" font-family="Arial, sans-serif" font-size="9" fill="#27ae60">1 scale per group of g elements</text>
+
+  <!-- 4x4 grid of groups -->
+  <rect x="580" y="75" width="40" height="28" rx="2" fill="#3498db" fill-opacity="0.15" stroke="#3498db" stroke-width="1"/>
+  <rect x="620" y="75" width="40" height="28" rx="2" fill="#e74c3c" fill-opacity="0.15" stroke="#e74c3c" stroke-width="1"/>
+  <rect x="660" y="75" width="40" height="28" rx="2" fill="#27ae60" fill-opacity="0.15" stroke="#27ae60" stroke-width="1"/>
+  <rect x="700" y="75" width="40" height="28" rx="2" fill="#f39c12" fill-opacity="0.15" stroke="#f39c12" stroke-width="1"/>
+
+  <rect x="580" y="105" width="40" height="28" rx="2" fill="#9b59b6" fill-opacity="0.15" stroke="#9b59b6" stroke-width="1"/>
+  <rect x="620" y="105" width="40" height="28" rx="2" fill="#1abc9c" fill-opacity="0.15" stroke="#1abc9c" stroke-width="1"/>
+  <rect x="660" y="105" width="40" height="28" rx="2" fill="#e67e22" fill-opacity="0.15" stroke="#e67e22" stroke-width="1"/>
+  <rect x="700" y="105" width="40" height="28" rx="2" fill="#2c3e50" fill-opacity="0.15" stroke="#2c3e50" stroke-width="1"/>
+
+  <rect x="580" y="135" width="40" height="28" rx="2" fill="#16a085" fill-opacity="0.15" stroke="#16a085" stroke-width="1"/>
+  <rect x="620" y="135" width="40" height="28" rx="2" fill="#8e44ad" fill-opacity="0.15" stroke="#8e44ad" stroke-width="1"/>
+  <rect x="660" y="135" width="40" height="28" rx="2" fill="#c0392b" fill-opacity="0.15" stroke="#c0392b" stroke-width="1"/>
+  <rect x="700" y="135" width="40" height="28" rx="2" fill="#2980b9" fill-opacity="0.15" stroke="#2980b9" stroke-width="1"/>
+
+  <rect x="580" y="165" width="40" height="28" rx="2" fill="#d35400" fill-opacity="0.15" stroke="#d35400" stroke-width="1"/>
+  <rect x="620" y="165" width="40" height="28" rx="2" fill="#27ae60" fill-opacity="0.15" stroke="#27ae60" stroke-width="1"/>
+  <rect x="660" y="165" width="40" height="28" rx="2" fill="#e74c3c" fill-opacity="0.15" stroke="#e74c3c" stroke-width="1"/>
+  <rect x="700" y="165" width="40" height="28" rx="2" fill="#3498db" fill-opacity="0.15" stroke="#3498db" stroke-width="1"/>
+
+  <text x="660" y="210" text-anchor="middle" font-family="Arial, sans-serif" font-size="8" fill="#666">each small block has its own scale</text>
+  <text x="660" y="222" text-anchor="middle" font-family="Arial, sans-serif" font-size="8" fill="#666">→ best accuracy, more scale overhead</text>
+
+  <!-- Accuracy arrow -->
+  <line x1="100" y1="260" x2="700" y2="260" stroke="#999" stroke-width="1.5"/>
+  <polygon points="705,260 695,255 695,265" fill="#999"/>
+  <text x="400" y="276" text-anchor="middle" font-family="Arial, sans-serif" font-size="10" fill="#999">increasing accuracy (more scale factors) →</text>
+</svg>