Merge branch 'nv_main'

docs/features/low_precision_training/introduction/img/mixed_precision_operations.svg

+<?xml version="1.0" encoding="UTF-8" standalone="no"?>
+<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 850 480" width="850" height="480" style="display: block; margin: 0 auto;">
+  <defs>
+    <style>
+      @import url("../_static/css/diagram-colors.css");
+      .arrow { stroke: #616161; stroke-width: 1.5; fill: none; marker-end: url(#arrowhead); }
+      .section-label { font-family: 'Segoe UI', Arial, sans-serif; font-size: 16px; font-weight: 600; fill: #424242; text-anchor: start; }
+    </style>
+    <marker id="arrowhead" markerWidth="3" markerHeight="3" refX="3" refY="1.5" orient="auto">
+      <polygon points="0 0, 3 1.5, 0 3" fill="#616161" />
+    </marker>
+  </defs>
+  
+  <!-- Title -->
+  <text x="425" y="30" class="title" style="text-anchor: middle;">Transformer Layer – default precision of operation in low precision recipe</text>
+  
+  <!-- Row 1: Input → Layer Norm → QKV Linear → QK^T → Softmax -->
+  <rect x="20" y="60" width="115" height="50" rx="5" class="hp"/>
+  <text x="77" y="90" class="text">Input</text>
+  
+  <path d="M 135 85 L 158 85" class="arrow"/>
+  
+  <rect x="158" y="60" width="115" height="50" rx="5" class="hp"/>
+  <text x="215" y="90" class="text">Layer Norm</text>
+  
+  <path d="M 273 85 L 296 85" class="arrow"/>
+  
+  <rect x="296" y="60" width="115" height="50" rx="5" class="gemm"/>
+  <text x="353" y="90" class="text">QKV Linear</text>
+  
+  <path d="M 411 85 L 434 85" class="arrow"/>
+  
+  <rect x="434" y="60" width="115" height="50" rx="5" class="hp"/>
+  <text x="491" y="90" class="text">QK^T</text>
+  
+  <path d="M 549 85 L 572 85" class="arrow"/>
+  
+  <rect x="572" y="60" width="115" height="50" rx="5" class="hp"/>
+  <text x="629" y="90" class="text">Softmax</text>
+  
+  <!-- Row 2: Attn * V → Output Linear → Dropout + Add -->
+  <path d="M 629 110 L 629 145" class="arrow"/>
+  
+  <rect x="572" y="145" width="115" height="50" rx="5" class="hp"/>
+  <text x="629" y="175" class="text">Scores * V</text>
+  
+  <path d="M 572 170 L 549 170" class="arrow"/>
+  
+  <rect x="434" y="145" width="115" height="50" rx="5" class="gemm"/>
+  <text x="491" y="175" class="text">Output Linear</text>
+  
+  <path d="M 434 170 L 273 170" class="arrow"/>
+  
+  <rect x="158" y="145" width="115" height="50" rx="5" class="hp"/>
+  <text x="215" y="175" class="text">Dropout + Add</text>
+  
+  <!-- Row 3: Layer Norm → FFN Linear 1 → GELU → FFN Linear 2 → Output -->
+  <path d="M 215 195 L 215 230" class="arrow"/>
+  
+  <rect x="158" y="230" width="115" height="50" rx="5" class="hp"/>
+  <text x="215" y="260" class="text">Layer Norm</text>
+  
+  <path d="M 273 255 L 296 255" class="arrow"/>
+  
+  <rect x="296" y="230" width="115" height="50" rx="5" class="gemm"/>
+  <text x="353" y="260" class="text">FFN Linear 1</text>
+  
+  <path d="M 411 255 L 434 255" class="arrow"/>
+  
+  <rect x="434" y="230" width="115" height="50" rx="5" class="hp"/>
+  <text x="491" y="260" class="text">GELU</text>
+  
+  <path d="M 549 255 L 572 255" class="arrow"/>
+  
+  <rect x="572" y="230" width="115" height="50" rx="5" class="gemm"/>
+  <text x="629" y="260" class="text">FFN Linear 2</text>
+  
+  <path d="M 687 255 L 710 255" class="arrow"/>
+  
+  <rect x="710" y="230" width="115" height="50" rx="5" class="hp"/>
+  <text x="767" y="260" class="text">Output</text>
+  
+  <!-- Memory State Section -->
+  <text x="20" y="325" class="section-label">Memory State:</text>
+  
+  <!-- Parameters -->
+  <rect x="20" y="340" width="180" height="45" rx="5" class="hp"/>
+  <text x="110" y="365" class="text">Parameters</text>
+  
+  <!-- Gradients -->
+  <rect x="225" y="340" width="140" height="45" rx="5" class="hp"/>
+  <text x="295" y="365" class="text">Gradients</text>
+  
+  <!-- Legend -->
+  <g transform="translate(20, 415)">
+    <!-- High Precision -->
+    <rect x="0" y="0" width="80" height="40" rx="5" class="hp"/>
+    <text x="95" y="23" class="text" style="text-anchor: start;">Higher Precision (FP32/BF16/FP16)</text>
+    
+    <!-- Low Precision -->
+    <rect x="400" y="0" width="80" height="40" rx="5" class="gemm"/>
+    <text x="495" y="23" class="text" style="text-anchor: start;">Lower Precision (FP8, MXFP8 etc.)</text>
+  </g>
+  
+</svg>

docs/features/low_precision_training/mxfp8/img/mxfp8_tensor_scaling_layout.svg

+<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 700 300" width="100%" style="max-width: 700px;">
+  <style>
+    .tensor-fill { fill: #87CEEB; stroke: #444; stroke-width: 2; }
+    .scale-fill { fill: #FFA500; stroke: #444; stroke-width: 2; }
+    .grid-line { stroke: #444; stroke-width: 2; fill: none; }
+    .dots-text { font: bold 24px sans-serif; fill: #333; text-anchor: middle; }
+    .label { font: 12px sans-serif; fill: #333; text-anchor: middle; }
+  </style>
+  
+  <!-- Left tensor (128x128 blocks) - FP8 tensor -->
+  <!-- Main rectangle with white background -->
+  <rect x="60" y="40" width="260" height="240" fill="#FFFFFF" stroke="#444" stroke-width="2"/>
+  
+  <!-- Blue grid area (upper-left) -->
+  <rect x="60" y="40" width="180" height="180" fill="#87CEEB" stroke="#444" stroke-width="2"/>
+  
+  <!-- Grid lines for 3x3 blocks -->
+  <line x1="120" y1="40" x2="120" y2="220" class="grid-line"/>
+  <line x1="180" y1="40" x2="180" y2="220" class="grid-line"/>
+  <line x1="60" y1="100" x2="240" y2="100" class="grid-line"/>
+  <line x1="60" y1="160" x2="240" y2="160" class="grid-line"/>
+  
+  <!-- Dots in white area (right side) -->
+  <text x="280" y="90" class="dots-text">…</text>
+  <text x="280" y="150" class="dots-text">…</text>
+  <text x="280" y="210" class="dots-text">…</text>
+  
+  <!-- Dots in white area (bottom) -->
+  <text x="90" y="260" class="dots-text">⋮</text>
+  <text x="150" y="260" class="dots-text">⋮</text>
+  <text x="210" y="260" class="dots-text">⋮</text>
+  <text x="280" y="260" class="dots-text">⋱</text>
+  
+  <!-- Label -->
+  <text x="190" y="20" class="label">FP8 Tensor (128×128 blocks)</text>
+  
+  
+  <!-- Right tensor (128x4 blocks) - Scaling factors (orange) -->
+  <!-- Main rectangle with white background -->
+  <rect x="480" y="40" width="120" height="240" fill="#FFFFFF" stroke="#444" stroke-width="2"/>
+  
+  <!-- Orange grid area (upper-left) -->
+  <rect x="480" y="40" width="60" height="180" fill="#FFA500" stroke="#444" stroke-width="2"/>
+  
+  <!-- Grid lines for narrow blocks (3 columns x 3 rows) -->
+  <line x1="500" y1="40" x2="500" y2="220" class="grid-line"/>
+  <line x1="520" y1="40" x2="520" y2="220" class="grid-line"/>
+  <line x1="480" y1="100" x2="540" y2="100" class="grid-line"/>
+  <line x1="480" y1="160" x2="540" y2="160" class="grid-line"/>
+  
+  <!-- Dots in white area (right side) -->
+  <text x="565" y="90" class="dots-text">…</text>
+  <text x="565" y="150" class="dots-text">…</text>
+  <text x="565" y="210" class="dots-text">…</text>
+  
+  <!-- Dots in white area (bottom) -->
+  <text x="500" y="260" class="dots-text">⋮</text>
+  <text x="530" y="260" class="dots-text">⋮</text>
+  <text x="565" y="260" class="dots-text">⋱</text>
+  
+  <!-- Label -->
+  <text x="540" y="20" class="label">Scaling Factors (128×4 blocks)</text>
+</svg>