diff --git "a/rotary/impls/torch_rotary.html" "b/rotary/impls/torch_rotary.html"
--- "a/rotary/impls/torch_rotary.html"
+++ "b/rotary/impls/torch_rotary.html"
@@ -4,6 +4,11 @@
     <meta charset="UTF-8">
     <meta name="viewport" content="width=device-width, initial-scale=1.0">
     <title>torch_rotary</title>
+
+    <link rel="preconnect" href="https://fonts.googleapis.com">
+    <link rel="preconnect" href="https://fonts.gstatic.com" crossorigin>
+    <link href="https://fonts.googleapis.com/css2?family=JetBrains+Mono:ital,wght@0,100..800;1,100..800&display=swap" rel="stylesheet">
+
     <script>
 // Iframe-friendly navigation router
         (function() {
@@ -87,1320 +92,1542 @@
     </script>
     <style>
 :root[data-theme="light"] {
-            --bg-primary: #ffffff;
-            --bg-secondary: #f6f8fa;
-            --bg-tertiary: #f8f9fa;
-            --bg-code: #f8f9fa;
-            --bg-error: #fdf2f2;
-            --bg-artifact: #e6f3ff;
-            --bg-artifact-hover: #d0e7ff;
-            
-            --text-primary: #333;
-            --text-secondary: #656d76;
-            --text-error: #c53030;
-            --text-link: #0969da;
-            
-            --border-primary: #e1e5e9;
-            --border-error: #e53e3e;
-            --border-cell-failed: #d73a49;
-            
-            --shadow: rgba(0, 0, 0, 0.1);
-        }
-
-        :root[data-theme="dark"] {
-            --bg-primary: #0a0a0a;
-            --bg-secondary: #121212;
-            --bg-tertiary: #181818;
-            --bg-code: #0d0d0d;
-            --bg-error: #1a0f0f;
-            --bg-artifact: #151515;
-            --bg-artifact-hover: #1a1a1a;
-            
-            --text-primary: #e0e0e0;
-            --text-secondary: #888888;
-            --text-error: #ff6b6b;
-            --text-link: #64b5f6;
-            
-            --border-primary: #2a2a2a;
-            --border-error: #ff6b6b;
-            --border-cell-failed: #ff6b6b;
-            
-            --shadow: rgba(255, 255, 255, 0.05);
-        }
-        /* Monocolor UI theme: black/white background, all text/borders single blue */
-        :root[data-ui="monocolor"] { --mono-color: #0a66ff; }
-        :root[data-ui="monocolor"][data-theme="light"] {
-            --bg-primary: #ffffff;
-        }
-        :root[data-ui="monocolor"][data-theme="dark"] {
-            --bg-primary: #000000;
-        }
-        :root[data-ui="monocolor"] {
-            --bg-secondary: var(--bg-primary);
-            --bg-tertiary: var(--bg-primary);
-            --bg-code: var(--bg-primary);
-            --bg-error: var(--bg-primary);
-            --bg-artifact: var(--bg-primary);
-            --bg-artifact-hover: var(--bg-primary);
-
-            --text-primary: var(--mono-color);
-            --text-secondary: var(--mono-color);
-            --text-error: var(--mono-color);
-            --text-link: var(--mono-color);
-
-            --border-primary: var(--mono-color);
-            --border-error: var(--mono-color);
-            --border-cell-failed: var(--mono-color);
-
-            --shadow: none;
-        }
-        :root[data-ui="monocolor"] a { color: var(--mono-color); }
-        :root[data-ui="monocolor"] .menu-button,
-        :root[data-ui="monocolor"] .theme-toggle,
-        :root[data-ui="monocolor"] .reset-toggle,
-        :root[data-ui="monocolor"] .back-button { background: var(--bg-primary); color: var(--mono-color); border-color: var(--mono-color); }
-        :root[data-ui="monocolor"] .menu-button:hover,
-        :root[data-ui="monocolor"] .theme-toggle:hover,
-        :root[data-ui="monocolor"] .reset-toggle:hover,
-        :root[data-ui="monocolor"] .back-button:hover { background: var(--bg-primary); color: var(--mono-color); border-color: var(--mono-color); }
-        :root[data-ui="monocolor"] .menu-dropdown { background: var(--bg-primary); border-color: var(--mono-color); box-shadow: none; }
-        :root[data-ui="monocolor"] .menu-item { color: var(--mono-color); border-bottom-color: var(--mono-color); }
-        :root[data-ui="monocolor"] .system-info { background: var(--bg-primary); border-color: var(--mono-color); }
-        :root[data-ui="monocolor"] .cell { border-color: var(--mono-color); background: var(--bg-primary); }
-        :root[data-ui="monocolor"] .cell-header { background: var(--bg-primary); border-bottom-color: var(--mono-color); }
-        :root[data-ui="monocolor"] .artifact { background: var(--bg-primary); border-color: var(--mono-color); color: var(--mono-color); }
-        :root[data-ui="monocolor"] .artifact:hover { background: var(--bg-primary); }
-        :root[data-ui="monocolor"] .artifact-preview img,
-        :root[data-ui="monocolor"] .artifact-preview svg { border-color: var(--mono-color); }
-        :root[data-ui="monocolor"] .status-widget { background: var(--bg-primary); border-color: var(--mono-color); color: var(--mono-color); }
-        :root[data-ui="monocolor"] .minimap,
-        :root[data-ui="monocolor"] .file-explorer,
-        :root[data-ui="monocolor"] .tools-widget {
-            background: var(--bg-primary);
-            border-color: var(--mono-color);
-            color: var(--mono-color);
-        }
-        :root[data-ui="monocolor"] .cell-code {
-            background: var(--bg-primary);
-            border-bottom-color: var(--mono-color);
-        }
-        :root[data-ui="monocolor"] .tools-title,
-        :root[data-ui="monocolor"] .file-explorer-section-title,
-        :root[data-ui="monocolor"] .minimap-title { color: var(--mono-color); border-bottom-color: var(--mono-color); }
-        :root[data-ui="monocolor"] .tool-button { background: var(--bg-primary); border-color: var(--mono-color); color: var(--mono-color); }
-        :root[data-ui="monocolor"] .tool-button.active { border-color: var(--mono-color); }
-        :root[data-ui="monocolor"] .file-explorer-item,
-        :root[data-ui="monocolor"] .minimap-item { color: var(--mono-color); }
-        /* Force Pygments code to mono blue on mono bg */
-        :root[data-ui="monocolor"] .highlight { background: var(--bg-primary) !important; color: var(--mono-color) !important; }
-        :root[data-ui="monocolor"] .highlight *,
-        :root[data-ui="monocolor"] .highlight .hll { color: var(--mono-color) !important; background: transparent !important; border-color: var(--mono-color) !important; }
-        /* Default code font + metrics (overridable via frontmatter) */
-        :root { --code-font-size: 0.95rem; --code-line-height: 1.5; --code-pad-y: 0.75rem; }
-        /* Minimal UI theme overrides base variables for a flatter, 90s look */
-        :root[data-ui="none"] {
-            --bg-primary: #ffffff;
-            --bg-secondary: transparent;
-            --bg-tertiary: transparent;
-            --bg-code: #f9f9f9;
-            --bg-error: #fff0f0;
-            --bg-artifact: #f0f7ff;
-            --bg-artifact-hover: #e5f1ff;
-
-            --text-primary: #000000;
-            --text-secondary: #222222;
-            --text-error: #a00000;
-            --text-link: #0000ee;
-
-            --border-primary: #cccccc;
-            --border-error: #cc0000;
-            --border-cell-failed: #cc0000;
-
-            --shadow: none;
-        }
-        html {
-            overscroll-behavior: none;
-        }
-        body {
-            font-family: 'Cascadia Mono', 'Cascadia Code', 'JetBrains Mono', 'SF Mono', Monaco, 'Consolas', monospace;
-            line-height: 1.4;
-            max-width: 1000px;
-            margin: 0 auto;
-            padding: 15px;
-            color: var(--text-primary);
-            background: var(--bg-primary);
-            transition: background-color 0.2s ease, color 0.2s ease;
-            overscroll-behavior: none;
-        }
-        /* Minimal "none" UI theme overrides */
-        :root[data-ui="none"] body {
-            font-family: 'Times New Roman', Times, serif;
-            line-height: 1.5;
-            max-width: 860px;
-            padding: 12px;
-            background: #ffffff;
-            color: #000000;
-            transition: none;
-        }
-        
-        /* Two panel layout removed */
-        
-        .controls {
-            position: fixed;
-            top: 20px;
-            right: 20px;
-            display: flex;
-            flex-direction: column;
-            align-items: flex-end;
-            gap: 0.25rem;
-            z-index: 1000;
-        }
-        .controls-buttons { display: flex; gap: 0.5rem; }
-        
-        .menu-button {
-            position: relative;
-            background: var(--bg-secondary);
-            border: 1px solid var(--border-primary);
-            padding: 8px 12px;
-            border-radius: 2px;
-            color: var(--text-secondary);
-            cursor: pointer;
-            font-family: inherit;
-            font-size: 0.9rem;
-            user-select: none;
-        }
-        /* Keep default control styling when widgets are enabled, even in minimal UI */
-        :root[data-ui="none"][data-widgets="on"] .menu-button,
-        :root[data-ui="none"][data-widgets="on"] .theme-toggle,
-        :root[data-ui="none"][data-widgets="on"] .reset-toggle,
-        :root[data-ui="none"][data-widgets="on"] .back-button {
-            background: #f6f6f6;
-            border: 1px solid #cccccc;
-            color: #222222;
-        }
-        
-        .menu-button:hover {
-            color: var(--text-primary);
-            background: var(--bg-tertiary);
-        }
-        /* Controls state indicator (top-right) */
-        /* Status widget (bottom-right) */
-        .status-widget {
-            position: fixed;
-            right: 20px;
-            bottom: 20px;
-            width: auto;
-            max-width: 260px;
-            background: var(--bg-secondary);
-            border: 1px solid var(--border-primary);
-            border-radius: 2px;
-            padding: 6px 8px;
-            font-size: 0.8rem;
-            color: var(--text-secondary);
-            z-index: 100;
-        }
-        .status-widget strong { color: var(--text-primary); }
-        :root[data-ui="none"][data-widgets="on"] .status-widget { background: #f6f6f6; border-color: #ccc; color: #222; }
-        :root[data-ui="none"][data-widgets="on"] .menu-button:hover,
-        :root[data-ui="none"][data-widgets="on"] .theme-toggle:hover,
-        :root[data-ui="none"][data-widgets="on"] .reset-toggle:hover,
-        :root[data-ui="none"][data-widgets="on"] .back-button:hover {
-            background: #ededed;
-            border-color: #bbbbbb;
-            color: #000000;
-        }
-        
-        .menu-dropdown {
-            position: absolute;
-            top: 100%;
-            right: 0;
-            background: var(--bg-secondary);
-            border: 1px solid var(--border-primary);
-            border-radius: 4px;
-            box-shadow: 0 4px 12px var(--shadow);
-            min-width: 160px;
-            opacity: 0;
-            visibility: hidden;
-            transform: translateY(-8px);
-            transition: all 0.2s ease;
-            z-index: 1001;
-            margin-top: 4px;
-        }
-        :root[data-ui="none"][data-widgets="on"] .menu-dropdown { background: #ffffff; border: 1px solid #cccccc; box-shadow: none; }
-        
-        .menu-button.active .menu-dropdown {
-            opacity: 1;
-            visibility: visible;
-            transform: translateY(0);
-        }
-        
-        .menu-item {
-            display: block;
-            padding: 8px 12px;
-            color: var(--text-secondary);
-            text-decoration: none;
-            font-size: 0.85rem;
-            border-bottom: 1px solid var(--border-primary);
-            cursor: pointer;
-        }
-        :root[data-ui="none"] .menu-item { color: #000; border-bottom: 1px solid #eee; }
-        
-        .menu-item:last-child {
-            border-bottom: none;
-        }
-        
-        .menu-item:hover {
-            background: var(--bg-tertiary);
-            color: var(--text-primary);
-        }
-        
-        .menu-checkbox {
-            display: inline-block;
-            width: 16px;
-            font-family: monospace;
-            color: var(--text-link);
-        }
-        
-        .theme-toggle,
-        .reset-toggle,
-        .back-button {
-            background: var(--bg-secondary);
-            border: 1px solid var(--border-primary);
-            padding: 8px 12px;
-            border-radius: 4px;
-            color: var(--text-secondary);
-            cursor: pointer;
-            font-family: inherit;
-            font-size: 0.9rem;
-            user-select: none;
-        }
-
-        .back-button {
-            text-decoration: none;
-            display: inline-block;
-        }
-
-        .theme-toggle:hover,
-        .reset-toggle:hover,
-        .back-button:hover {
-            color: var(--text-primary);
-            background: var(--bg-tertiary);
-        }
-        
-        .system-info {
-            background: var(--bg-secondary);
-            border: 1px solid var(--border-primary);
-            border-radius: 4px;
-            padding: 8px 12px;
-            margin-bottom: 16px;
-            font-size: 0.85em;
-            color: var(--text-secondary);
-        }
-        
-        .system-info-header {
-            font-weight: 600;
-            color: var(--text-primary);
-            margin-bottom: 2px;
-        }
-        
-        .system-info-content {
-            font-family: monospace;
-        }
-        
-        .theme-toggle, .reset-toggle {
-            background: var(--bg-secondary);
-            border: 1px solid var(--border-primary);
-            border-radius: 2px;
-            /* padding: 0.4rem 0.6rem; */
-            cursor: pointer;
-            font-family: inherit;
-            font-size: 0.8rem;
-            color: var(--text-secondary);
-            user-select: none;
-            transition: all 0.2s ease;
-            text-transform: lowercase;
-            letter-spacing: 0;
-        }
-        
-        .theme-toggle:hover, .reset-toggle:hover {
-            background: var(--bg-tertiary);
-            border-color: var(--text-secondary);
-            color: var(--text-primary);
-        }
-        
-        .minimap {
-            position: fixed;
-            bottom: 20px;
-            right: 20px;
-            width: 220px;
-            max-height: 400px;
-            background: var(--bg-secondary);
-            border: 1px solid var(--border-primary);
-            border-radius: 2px;
-            padding: 0.5rem;
-            font-size: 0.7rem;
-            overflow-y: auto;
-            z-index: 100;
-            opacity: 0.9;
-            transition: opacity 0.2s ease;
-        }
-        /* Hide widgets and controls when disabled via frontmatter */
-        :root[data-widgets="off"] .controls,
-        :root[data-widgets="off"] .minimap,
-        :root[data-widgets="off"] .file-explorer,
-        :root[data-widgets="off"] .tools-widget,
-        :root[data-widgets="off"] .status-widget { display: none !important; }
-        
-        .file-explorer {
-            position: fixed;
-            bottom: 20px; /* default; JS will stack */
-            right: 20px;
-            left: auto;
-            top: auto;
-            width: 220px;
-            max-height: 400px;
-            background: var(--bg-secondary);
-            border: 1px solid var(--border-primary);
-            border-radius: 2px;
-            padding: 0.5rem;
-            font-size: 0.7rem;
-            overflow-y: auto;
-            z-index: 100;
-            opacity: 0.9;
-            transition: opacity 0.2s ease;
-        }
-
-        /* Drawing overlay */
-        .draw-overlay {
-            position: fixed;
-            top: 0;
-            left: 0;
-            width: 100vw;
-            height: 100vh;
-            z-index: 80; /* under widgets (100) and controls (1000) */
-            display: block;
-            pointer-events: none; /* enabled only when a tool is active */
-        }
-
-        /* Tools widget */
-        .tools-widget {
-            position: fixed;
-            bottom: 20px; /* default; JS will stack */
-            right: 20px;
-            left: auto;
-            top: auto;
-            width: 220px;
-            background: var(--bg-secondary);
-            border: 1px solid var(--border-primary);
-            border-radius: 2px;
-            padding: 0.5rem;
-            font-size: 0.7rem;
-            z-index: 100;
-            opacity: 0.95;
-        }
-        .tools-title {
-            font-weight: bold;
-            color: var(--text-secondary);
-            margin-bottom: 0.5rem;
-            padding-bottom: 0.25rem;
-            border-bottom: 1px solid var(--border-primary);
-            cursor: grab;
-            user-select: none;
-        }
-        .tools-row { display: flex; gap: 0.4rem; flex-wrap: wrap; }
-        .tool-button {
-            background: var(--bg-tertiary);
-            border: 1px solid var(--border-primary);
-            border-radius: 2px;
-            padding: 0.25rem 0.4rem;
-            cursor: pointer;
-            color: var(--text-secondary);
-            font-family: inherit;
-            font-size: 0.75rem;
-            user-select: none;
-        }
-        .tool-button:hover { color: var(--text-primary); }
-        .tool-button.active { color: var(--text-primary); border-color: var(--text-secondary); background: var(--bg-secondary); }
-        
-        .minimap:hover, .file-explorer:hover {
-            opacity: 1;
-        }
-        
-        .minimap-title {
-            font-weight: bold;
-            color: var(--text-secondary);
-            margin-bottom: 0.5rem;
-            padding-bottom: 0.25rem;
-            border-bottom: 1px solid var(--border-primary);
-            cursor: grab; /* drag handle */
-            user-select: none;
-        }
-        
-        .minimap-item {
-            display: block;
-            color: var(--text-secondary);
-            text-decoration: none;
-            padding: 0.15rem 0;
-            border-left: 2px solid transparent;
-            padding-left: 0.5rem;
-            transition: all 0.2s ease;
-            cursor: pointer;
-        }
-        
-        .minimap-item:hover {
-            color: var(--text-primary);
-            border-left-color: var(--text-secondary);
-        }
-        
-        .minimap-item.active {
-            color: var(--text-primary);
-            border-left-color: var(--text-link);
-        }
-        
-        .minimap-heading {
-            font-weight: normal;
-        }
-        
-        .minimap-heading.h1 { padding-left: 0.5rem; }
-        .minimap-heading.h2 { padding-left: 1rem; }
-        .minimap-heading.h3 { padding-left: 1.5rem; }
-        .minimap-heading.h4 { padding-left: 2rem; }
-        .minimap-heading.h5 { padding-left: 2.5rem; }
-        .minimap-heading.h6 { padding-left: 3rem; }
-        
-        .minimap-cell {
-            color: var(--text-link);
-            opacity: 0.8;
-            font-style: italic;
-        }
-        
-        .minimap-cell:hover {
-            opacity: 1;
-        }
-        
-        .file-explorer-title {
-            font-weight: bold;
-            color: var(--text-secondary);
-            margin-bottom: 0.5rem;
-            padding-bottom: 0.25rem;
-            border-bottom: 1px solid var(--border-primary);
-            cursor: grab; /* drag handle */
-            user-select: none;
-        }
-        
-        .file-explorer-section {
-            margin-bottom: 0.75rem;
-        }
-        
-        .file-explorer-section-title {
-            font-weight: bold;
-            color: var(--text-secondary);
-            font-size: 0.65rem;
-            margin-bottom: 0.25rem;
-            text-transform: uppercase;
-            letter-spacing: 0.5px;
-        }
-        
-        .file-explorer-item {
-            display: block;
-            color: var(--text-secondary);
-            text-decoration: none;
-            padding: 0.1rem 0;
-            margin-left: 0.5rem;
-            transition: color 0.2s ease;
-            cursor: pointer;
-            font-family: monospace;
-        }
-        
-        .file-explorer-item:hover {
-            color: var(--text-primary);
-        }
-        
-        .file-explorer-item.script {
-            color: var(--text-link);
-        }
-        
-        .file-explorer-item.artifact {
-            color: var(--text-secondary);
-            opacity: 0.8;
-        }
-        
-
-        /* Hide widgets on smaller screens */
-        @media (max-width: 768px) {
-            .minimap, .file-explorer, .tools-widget {
-                display: none;
-            }
-        }
-        
-        .cell {
-            margin: 1rem 0;
-            border: 1px solid var(--border-primary);
-            border-radius: 2px;
-            overflow: hidden;
-            background: var(--bg-secondary);
-        }
-        :root[data-ui="none"] .cell { margin: 1em 0; border: none; background: transparent; }
-        .cell-header {
-            background: var(--bg-secondary);
-            padding: 0.5rem 1rem;
-            border-bottom: 1px solid var(--border-primary);
-            font-family: inherit;
-            font-size: 0.85rem;
-        }
-        :root[data-ui="none"] .cell-header { background: transparent; border: none; padding: 0; font-weight: bold; }
-        :root[data-ui="none"] .cell-content { padding: 0; }
-        :root[data-ui="none"] .copy-button,
-        :root[data-ui="none"] .collapse-indicators,
-        :root[data-ui="none"] .cell-meta,
-        :root[data-ui="none"] .cell-outputs-header { display: none !important; }
-        :root[data-ui="none"] pre, :root[data-ui="none"] code { font-family: Menlo, Monaco, 'Courier New', monospace; }
-        :root[data-ui="none"] .code-content pre { background: #f9f9f9; border: 1px solid #ddd; padding: 8px; }
-        :root[data-ui="none"] .output { background: transparent; border: none; padding: 0.25em 0; }
-            color: var(--text-secondary);
-            cursor: pointer;
-            user-select: none;
-            transition: background-color 0.2s ease;
-        }
-        .cell-header:hover {
-            background: var(--bg-tertiary);
-        }
-        .collapse-indicators {
-            color: var(--text-secondary);
-            font-size: 0.8rem;
-            opacity: 0.7;
-        }
-        .collapse-indicators span:hover {
-            color: var(--text-primary);
-            opacity: 1;
-        }
-        .cell-code {
-            display: block;
-            background: var(--bg-code);
-        }
-        .cell-code.collapsed {
-            display: none;
-        }
-        .cell-code pre {
-            margin: 0;
-            padding: 0.75rem;
-            background: var(--bg-code);
-            overflow-x: auto;
-            color: var(--text-primary);
-        }
-        .cell-output {
-            padding: 0.75rem;
-            /* background: var(--bg-primary); */
-            background: var(--bg-secondary);
-        }
-        .cell-output.collapsed {
-            display: none;
-        }
-        .cell-stdout {
-            background: var(--bg-tertiary);
-            padding: 0.75rem;
-            border-radius: 1px;
-            /* margin: 0.25rem 0; */
-            font-family: inherit;
-            font-size: 0.9rem;
-            white-space: pre-wrap;
-            color: var(--text-primary);
-        }
-
-        .cell-stdout {
-            background: var(--bg-tertiary);
-            padding: 0.75rem;
-            border-radius: 1px;
-            font-family: inherit;
-            font-size: 0.9rem;
-            color: var(--text-primary);
-
-            /* key bits */
-            overflow: auto;          /* show scrollbars when needed */
-            max-width: 100%;         /* respects whatever layout width you give it */
-        }
-
-        .cell-stdout .stdout-text {
-            margin: 0;               /* reset pre default margin */
-            white-space: pre;        /* keep line breaks, NO wrapping */
-            display: inline-block;   /* shrink-to-content */
-            min-width: max-content;  /* allow very long lines to define intrinsic width */
-            font-family: ui-monospace, SFMono-Regular, Menlo, Consolas, monospace;
-            tab-size: 2;
-        }
-
-        .cell-stderr {
-            background: var(--bg-error);
-            border-left: 2px solid var(--border-error);
-            padding: 1rem;
-            margin: 0.5rem 0;
-            font-family: inherit;
-            font-size: 0.9rem;
-            color: var(--text-error);
-            white-space: pre-wrap;
-        }
-        .uv-install-logs {
-            margin: 0.5rem 0;
-        }
-        .uv-logs-header {
-            cursor: pointer;
-            padding: 0.75rem;
-            border-left: 3px solid var(--border-color);
-            font-family: inherit;
-            font-size: 0.85rem;
-            color: var(--text-secondary);
-            user-select: none;
-        }
-        .uv-logs-content {
-            background: var(--bg-secondary);
-            padding: 1rem;
-            border-left: 3px solid var(--border-color);
-            white-space: pre-wrap;
-            font-family: monospace;
-            font-size: 0.85rem;
-            color: var(--text-secondary);
-            overflow-x: auto;
-        }
-        .cell-artifacts {
-            margin: 1rem 0;
-        }
-        .cell-artifacts h4 {
-            margin: 0 0 0.5rem 0;
-            color: var(--text-secondary);
-            font-size: 0.9rem;
-        }
-        .artifact {
-            display: inline-block;
-            background: var(--bg-artifact);
-            padding: 0.25rem 0.5rem;
-            border-radius: 1px;
-            margin: 0.25rem 0.5rem 0.25rem 0;
-            font-family: inherit;
-            font-size: 0.8rem;
-            color: var(--text-link);
-            text-decoration: none;
-            transition: background-color 0.2s ease;
-            border: 1px solid var(--border-primary);
-        }
-        .artifact:hover {
-            background: var(--bg-artifact-hover);
-        }
-        .artifact-preview {
-            margin-top: 1rem;
-        }
-        .artifact-preview img {
-            max-width: 100%;
-            height: auto;
-            border: 1px solid var(--border-primary);
-            border-radius: 1px;
-        }
-        .artifact-preview svg {
-            max-width: 100%;
-            height: auto;
-            border: 1px solid var(--border-primary);
-            border-radius: 1px;
-            display: block;
-        }
-        /* Style SVG text elements */
-        .artifact-preview svg g {
-            fill: var(--text-primary) !important;
-        }
-        /* Auto-theme SVG elements */
-        .artifact-preview svg {
-            background: transparent;
-        }
-        /* Invert SVG images in dark mode */
-        :root[data-theme="dark"] .artifact-preview img[src$=".svg"] {
-            filter: invert(0.9) hue-rotate(180deg);
-        }
-        /* Keep SVG images readable in monocolor mode */
-        :root[data-ui="monocolor"] .artifact-preview img[src$=".svg"] {
-            filter: none;
-        }
-        /* CSV table styling */
-        .artifact-csv {
-            margin-top: 1rem;
-            overflow-x: auto;
-        }
-        .csv-table {
-            width: 100%;
-            border-collapse: collapse;
-            font-size: 0.9rem;
-            background: var(--bg-secondary);
-            border: 1px solid var(--border-primary);
-            border-radius: 1px;
-        }
-        .csv-table th,
-        .csv-table td {
-            padding: 0.5rem 0.75rem;
-            text-align: left;
-            border: 1px solid var(--border-primary);
-        }
-        .csv-table th {
-            background: var(--bg-tertiary);
-            font-weight: 600;
-            color: var(--text-primary);
-        }
-        .csv-table tbody tr:hover {
-            background: var(--bg-artifact-hover);
-        }
-        .artifact-csv-error {
-            margin-top: 1rem;
-            padding: 1rem;
-            background: var(--bg-error);
-            color: var(--text-error);
-            border: 1px solid var(--border-error);
-            border-radius: 1px;
-        }
-        .cell-failed {
-            border-color: var(--border-cell-failed);
-        }
-        .cell-failed .cell-header {
-            background: var(--bg-error);
-            color: var(--text-error);
-        }
-        .cell-commented {
-            opacity: 0.6;
-            border-style: dashed;
-        }
-        .cell-commented .cell-header {
-            background: var(--bg-secondary);
-            color: var(--text-secondary);
-            font-style: italic;
-        }
-        .run-btn {
-            background: var(--bg-tertiary);
-            border: 1px solid var(--border-primary);
-            padding: 2px 6px;
-            border-radius: 2px;
-            color: var(--text-secondary);
-            cursor: pointer;
-            font-size: 0.75em;
-            font-family: inherit;
-            margin-left: 4px;
-        }
-        .run-btn:hover {
-            color: var(--text-primary);
-            background: var(--bg-primary);
-        }
-        .run-btn:disabled {
-            opacity: 0.6;
-            cursor: not-allowed;
-        }
-        .copy-btn {
-            background: var(--bg-tertiary);
-            border: 1px solid var(--border-primary);
-            padding: 2px 6px;
-            border-radius: 2px;
-            color: var(--text-secondary);
-            cursor: pointer;
-            font-size: 0.75em;
-            font-family: inherit;
-            margin-left: 4px;
-        }
-        .copy-btn:hover {
-            color: var(--text-primary);
-            background: var(--bg-primary);
-        }
-        .copy-btn:disabled {
-            opacity: 0.6;
-            cursor: not-allowed;
-        }
-        .copy-btn.copied {
-            color: #4caf50;
-            background: var(--bg-primary);
-            border-color: #4caf50;
-            transition: all 0.2s ease;
-        }
-        .raw-btn {
-            background: var(--bg-tertiary);
-            border: 1px solid var(--border-primary);
-            padding: 2px 6px;
-            border-radius: 2px;
-            color: var(--text-secondary);
-            cursor: pointer;
-            font-size: 0.75em;
-            font-family: inherit;
-            margin-left: 4px;
-            text-decoration: none;
-            display: inline-block;
-        }
-        .raw-btn:hover {
-            color: var(--text-primary);
-            background: var(--bg-primary);
-            text-decoration: none;
-        }
-        .github-btn {
-            background: var(--bg-tertiary);
-            border: 1px solid var(--border-primary);
-            padding: 2px 6px;
-            border-radius: 2px;
-            color: var(--text-secondary);
-            cursor: pointer;
-            font-size: 0.75em;
-            font-family: inherit;
-            margin-left: 4px;
-            text-decoration: none;
-            display: inline-block;
-        }
-        .github-btn:hover {
-            color: var(--text-primary);
-            background: var(--bg-primary);
-            text-decoration: none;
-        }
-        .hf-btn {
-            background: var(--bg-tertiary);
-            border: 1px solid var(--border-primary);
-            padding: 2px 6px;
-            border-radius: 2px;
-            color: var(--text-secondary);
-            cursor: pointer;
-            font-size: 0.75em;
-            font-family: inherit;
-            margin-left: 4px;
-            text-decoration: none;
-            display: inline-block;
-        }
-        .hf-btn:hover {
-            color: var(--text-primary);
-            background: var(--bg-primary);
-            text-decoration: none;
-        }
-        .output-stale {
-            opacity: 0.5;
-            position: relative;
-        }
-        .output-stale::after {
-            content: '⏳ updating...';
-            position: absolute;
-            top: 8px;
-            right: 8px;
-            background: var(--bg-secondary);
-            padding: 4px 8px;
-            border-radius: 2px;
-            font-size: 0.75em;
-            color: var(--text-secondary);
-            border: 1px solid var(--border-primary);
-        }
-        h1, h2, h3, h4, h5, h6 {
-            margin-top: 1.5rem;
-            margin-bottom: 0.75rem;
-            color: var(--text-primary);
-        }
-        h1 {
-            margin-top: 0;
-            margin-bottom: 1rem;
-        }
-        p {
-            margin: 0.75rem 0;
-            color: var(--text-primary);
-        }
-        a {
-            color: var(--text-link);
-        }
-        img {
-            max-width: 100%;
-            height: auto;
-            border-radius: 1px;
-            box-shadow: none;
-        }
-        pre, code {
-            font-family: 'Cascadia Mono', 'Cascadia Code', 'JetBrains Mono', 'SF Mono', Monaco, 'Consolas', monospace;
-            font-size: var(--code-font-size);
-        }
-        .code-wrap { position: relative; }
-        .code-line-highlight { display: none; position: absolute; left: 0; right: 0; height: 1.5em; background: rgba(255, 235, 170, 0.35); pointer-events: none; border-left: 3px solid #f4c542; }
-        .line-number { cursor: pointer; text-decoration: none; color: var(--text-secondary); padding: 0 0.25rem; }
-        .line-number.selected { background: rgba(255, 235, 170, 0.4); color: var(--text-primary); }
-        
-        /* Line numbers */
-        .highlight-with-lines {
-            display: flex;
-        }
-        .line-numbers {
-            background: var(--bg-tertiary);
-            padding: var(--code-pad-y) 0.5rem;
-            font-family: 'Cascadia Mono', 'Cascadia Code', 'JetBrains Mono', 'SF Mono', Monaco, 'Consolas', monospace;
-            font-size: var(--code-font-size);
-            line-height: var(--code-line-height);
-            color: var(--text-secondary);
-            user-select: none;
-            text-align: right;
-            border-right: 1px solid var(--border-primary);
-        }
-        .line-numbers .line-number {
-            display: block;
-            line-height: var(--code-line-height);
-        }
-        .highlight-with-lines .highlight {
-            flex: 1;
-        }
-        .highlight .hll { background-color: transparent; } /* don't conflict with our highlight */
-        .highlight pre {
-            white-space: pre;
-            margin: 0;
-            padding: var(--code-pad-y) 0.75rem;
-            line-height: var(--code-line-height);
-        }
-        
-        /* Collapsed code styling */
-        .cell-code.collapsed {
-            display: none;
-        }
-        .cell-code.expanded {
-            display: block;
-        }
-        
-        .cell-code {
-            display: block;
-            border-bottom: 1px solid var(--border-primary);
-        }
-        
-        
-        pre { line-height: 125%; }
-td.linenos .normal { color: inherit; background-color: transparent; padding-left: 5px; padding-right: 5px; }
-span.linenos { color: inherit; background-color: transparent; padding-left: 5px; padding-right: 5px; }
-td.linenos .special { color: #000000; background-color: #ffffc0; padding-left: 5px; padding-right: 5px; }
-span.linenos.special { color: #000000; background-color: #ffffc0; padding-left: 5px; padding-right: 5px; }
-[data-theme="light"] .highlight .hll { background-color: #ffffcc }
-[data-theme="light"] .highlight { background: #f8f8f8; }
-[data-theme="light"] .highlight .c { color: #3D7B7B; font-style: italic } /* Comment */
-[data-theme="light"] .highlight .err { border: 1px solid #F00 } /* Error */
-[data-theme="light"] .highlight .k { color: #008000; font-weight: bold } /* Keyword */
-[data-theme="light"] .highlight .o { color: #666 } /* Operator */
-[data-theme="light"] .highlight .ch { color: #3D7B7B; font-style: italic } /* Comment.Hashbang */
-[data-theme="light"] .highlight .cm { color: #3D7B7B; font-style: italic } /* Comment.Multiline */
-[data-theme="light"] .highlight .cp { color: #9C6500 } /* Comment.Preproc */
-[data-theme="light"] .highlight .cpf { color: #3D7B7B; font-style: italic } /* Comment.PreprocFile */
-[data-theme="light"] .highlight .c1 { color: #3D7B7B; font-style: italic } /* Comment.Single */
-[data-theme="light"] .highlight .cs { color: #3D7B7B; font-style: italic } /* Comment.Special */
-[data-theme="light"] .highlight .gd { color: #A00000 } /* Generic.Deleted */
-[data-theme="light"] .highlight .ge { font-style: italic } /* Generic.Emph */
-[data-theme="light"] .highlight .ges { font-weight: bold; font-style: italic } /* Generic.EmphStrong */
-[data-theme="light"] .highlight .gr { color: #E40000 } /* Generic.Error */
-[data-theme="light"] .highlight .gh { color: #000080; font-weight: bold } /* Generic.Heading */
-[data-theme="light"] .highlight .gi { color: #008400 } /* Generic.Inserted */
-[data-theme="light"] .highlight .go { color: #717171 } /* Generic.Output */
-[data-theme="light"] .highlight .gp { color: #000080; font-weight: bold } /* Generic.Prompt */
-[data-theme="light"] .highlight .gs { font-weight: bold } /* Generic.Strong */
-[data-theme="light"] .highlight .gu { color: #800080; font-weight: bold } /* Generic.Subheading */
-[data-theme="light"] .highlight .gt { color: #04D } /* Generic.Traceback */
-[data-theme="light"] .highlight .kc { color: #008000; font-weight: bold } /* Keyword.Constant */
-[data-theme="light"] .highlight .kd { color: #008000; font-weight: bold } /* Keyword.Declaration */
-[data-theme="light"] .highlight .kn { color: #008000; font-weight: bold } /* Keyword.Namespace */
-[data-theme="light"] .highlight .kp { color: #008000 } /* Keyword.Pseudo */
-[data-theme="light"] .highlight .kr { color: #008000; font-weight: bold } /* Keyword.Reserved */
-[data-theme="light"] .highlight .kt { color: #B00040 } /* Keyword.Type */
-[data-theme="light"] .highlight .m { color: #666 } /* Literal.Number */
-[data-theme="light"] .highlight .s { color: #BA2121 } /* Literal.String */
-[data-theme="light"] .highlight .na { color: #687822 } /* Name.Attribute */
-[data-theme="light"] .highlight .nb { color: #008000 } /* Name.Builtin */
-[data-theme="light"] .highlight .nc { color: #00F; font-weight: bold } /* Name.Class */
-[data-theme="light"] .highlight .no { color: #800 } /* Name.Constant */
-[data-theme="light"] .highlight .nd { color: #A2F } /* Name.Decorator */
-[data-theme="light"] .highlight .ni { color: #717171; font-weight: bold } /* Name.Entity */
-[data-theme="light"] .highlight .ne { color: #CB3F38; font-weight: bold } /* Name.Exception */
-[data-theme="light"] .highlight .nf { color: #00F } /* Name.Function */
-[data-theme="light"] .highlight .nl { color: #767600 } /* Name.Label */
-[data-theme="light"] .highlight .nn { color: #00F; font-weight: bold } /* Name.Namespace */
-[data-theme="light"] .highlight .nt { color: #008000; font-weight: bold } /* Name.Tag */
-[data-theme="light"] .highlight .nv { color: #19177C } /* Name.Variable */
-[data-theme="light"] .highlight .ow { color: #A2F; font-weight: bold } /* Operator.Word */
-[data-theme="light"] .highlight .w { color: #BBB } /* Text.Whitespace */
-[data-theme="light"] .highlight .mb { color: #666 } /* Literal.Number.Bin */
-[data-theme="light"] .highlight .mf { color: #666 } /* Literal.Number.Float */
-[data-theme="light"] .highlight .mh { color: #666 } /* Literal.Number.Hex */
-[data-theme="light"] .highlight .mi { color: #666 } /* Literal.Number.Integer */
-[data-theme="light"] .highlight .mo { color: #666 } /* Literal.Number.Oct */
-[data-theme="light"] .highlight .sa { color: #BA2121 } /* Literal.String.Affix */
-[data-theme="light"] .highlight .sb { color: #BA2121 } /* Literal.String.Backtick */
-[data-theme="light"] .highlight .sc { color: #BA2121 } /* Literal.String.Char */
-[data-theme="light"] .highlight .dl { color: #BA2121 } /* Literal.String.Delimiter */
-[data-theme="light"] .highlight .sd { color: #BA2121; font-style: italic } /* Literal.String.Doc */
-[data-theme="light"] .highlight .s2 { color: #BA2121 } /* Literal.String.Double */
-[data-theme="light"] .highlight .se { color: #AA5D1F; font-weight: bold } /* Literal.String.Escape */
-[data-theme="light"] .highlight .sh { color: #BA2121 } /* Literal.String.Heredoc */
-[data-theme="light"] .highlight .si { color: #A45A77; font-weight: bold } /* Literal.String.Interpol */
-[data-theme="light"] .highlight .sx { color: #008000 } /* Literal.String.Other */
-[data-theme="light"] .highlight .sr { color: #A45A77 } /* Literal.String.Regex */
-[data-theme="light"] .highlight .s1 { color: #BA2121 } /* Literal.String.Single */
-[data-theme="light"] .highlight .ss { color: #19177C } /* Literal.String.Symbol */
-[data-theme="light"] .highlight .bp { color: #008000 } /* Name.Builtin.Pseudo */
-[data-theme="light"] .highlight .fm { color: #00F } /* Name.Function.Magic */
-[data-theme="light"] .highlight .vc { color: #19177C } /* Name.Variable.Class */
-[data-theme="light"] .highlight .vg { color: #19177C } /* Name.Variable.Global */
-[data-theme="light"] .highlight .vi { color: #19177C } /* Name.Variable.Instance */
-[data-theme="light"] .highlight .vm { color: #19177C } /* Name.Variable.Magic */
-[data-theme="light"] .highlight .il { color: #666 } /* Literal.Number.Integer.Long */
-
-pre { line-height: 125%; }
-td.linenos .normal { color: inherit; background-color: transparent; padding-left: 5px; padding-right: 5px; }
-span.linenos { color: inherit; background-color: transparent; padding-left: 5px; padding-right: 5px; }
-td.linenos .special { color: #000000; background-color: #ffffc0; padding-left: 5px; padding-right: 5px; }
-span.linenos.special { color: #000000; background-color: #ffffc0; padding-left: 5px; padding-right: 5px; }
-[data-theme="dark"] .highlight .hll { background-color: #49483e }
-[data-theme="dark"] .highlight { background: #272822; color: #F8F8F2 }
-[data-theme="dark"] .highlight .c { color: #959077 } /* Comment */
-[data-theme="dark"] .highlight .err { color: #ED007E; background-color: #1E0010 } /* Error */
-[data-theme="dark"] .highlight .esc { color: #F8F8F2 } /* Escape */
-[data-theme="dark"] .highlight .g { color: #F8F8F2 } /* Generic */
-[data-theme="dark"] .highlight .k { color: #66D9EF } /* Keyword */
-[data-theme="dark"] .highlight .l { color: #AE81FF } /* Literal */
-[data-theme="dark"] .highlight .n { color: #F8F8F2 } /* Name */
-[data-theme="dark"] .highlight .o { color: #FF4689 } /* Operator */
-[data-theme="dark"] .highlight .x { color: #F8F8F2 } /* Other */
-[data-theme="dark"] .highlight .p { color: #F8F8F2 } /* Punctuation */
-[data-theme="dark"] .highlight .ch { color: #959077 } /* Comment.Hashbang */
-[data-theme="dark"] .highlight .cm { color: #959077 } /* Comment.Multiline */
-[data-theme="dark"] .highlight .cp { color: #959077 } /* Comment.Preproc */
-[data-theme="dark"] .highlight .cpf { color: #959077 } /* Comment.PreprocFile */
-[data-theme="dark"] .highlight .c1 { color: #959077 } /* Comment.Single */
-[data-theme="dark"] .highlight .cs { color: #959077 } /* Comment.Special */
-[data-theme="dark"] .highlight .gd { color: #FF4689 } /* Generic.Deleted */
-[data-theme="dark"] .highlight .ge { color: #F8F8F2; font-style: italic } /* Generic.Emph */
-[data-theme="dark"] .highlight .ges { color: #F8F8F2; font-weight: bold; font-style: italic } /* Generic.EmphStrong */
-[data-theme="dark"] .highlight .gr { color: #F8F8F2 } /* Generic.Error */
-[data-theme="dark"] .highlight .gh { color: #F8F8F2 } /* Generic.Heading */
-[data-theme="dark"] .highlight .gi { color: #A6E22E } /* Generic.Inserted */
-[data-theme="dark"] .highlight .go { color: #66D9EF } /* Generic.Output */
-[data-theme="dark"] .highlight .gp { color: #FF4689; font-weight: bold } /* Generic.Prompt */
-[data-theme="dark"] .highlight .gs { color: #F8F8F2; font-weight: bold } /* Generic.Strong */
-[data-theme="dark"] .highlight .gu { color: #959077 } /* Generic.Subheading */
-[data-theme="dark"] .highlight .gt { color: #F8F8F2 } /* Generic.Traceback */
-[data-theme="dark"] .highlight .kc { color: #66D9EF } /* Keyword.Constant */
-[data-theme="dark"] .highlight .kd { color: #66D9EF } /* Keyword.Declaration */
-[data-theme="dark"] .highlight .kn { color: #FF4689 } /* Keyword.Namespace */
-[data-theme="dark"] .highlight .kp { color: #66D9EF } /* Keyword.Pseudo */
-[data-theme="dark"] .highlight .kr { color: #66D9EF } /* Keyword.Reserved */
-[data-theme="dark"] .highlight .kt { color: #66D9EF } /* Keyword.Type */
-[data-theme="dark"] .highlight .ld { color: #E6DB74 } /* Literal.Date */
-[data-theme="dark"] .highlight .m { color: #AE81FF } /* Literal.Number */
-[data-theme="dark"] .highlight .s { color: #E6DB74 } /* Literal.String */
-[data-theme="dark"] .highlight .na { color: #A6E22E } /* Name.Attribute */
-[data-theme="dark"] .highlight .nb { color: #F8F8F2 } /* Name.Builtin */
-[data-theme="dark"] .highlight .nc { color: #A6E22E } /* Name.Class */
-[data-theme="dark"] .highlight .no { color: #66D9EF } /* Name.Constant */
-[data-theme="dark"] .highlight .nd { color: #A6E22E } /* Name.Decorator */
-[data-theme="dark"] .highlight .ni { color: #F8F8F2 } /* Name.Entity */
-[data-theme="dark"] .highlight .ne { color: #A6E22E } /* Name.Exception */
-[data-theme="dark"] .highlight .nf { color: #A6E22E } /* Name.Function */
-[data-theme="dark"] .highlight .nl { color: #F8F8F2 } /* Name.Label */
-[data-theme="dark"] .highlight .nn { color: #F8F8F2 } /* Name.Namespace */
-[data-theme="dark"] .highlight .nx { color: #A6E22E } /* Name.Other */
-[data-theme="dark"] .highlight .py { color: #F8F8F2 } /* Name.Property */
-[data-theme="dark"] .highlight .nt { color: #FF4689 } /* Name.Tag */
-[data-theme="dark"] .highlight .nv { color: #F8F8F2 } /* Name.Variable */
-[data-theme="dark"] .highlight .ow { color: #FF4689 } /* Operator.Word */
-[data-theme="dark"] .highlight .pm { color: #F8F8F2 } /* Punctuation.Marker */
-[data-theme="dark"] .highlight .w { color: #F8F8F2 } /* Text.Whitespace */
-[data-theme="dark"] .highlight .mb { color: #AE81FF } /* Literal.Number.Bin */
-[data-theme="dark"] .highlight .mf { color: #AE81FF } /* Literal.Number.Float */
-[data-theme="dark"] .highlight .mh { color: #AE81FF } /* Literal.Number.Hex */
-[data-theme="dark"] .highlight .mi { color: #AE81FF } /* Literal.Number.Integer */
-[data-theme="dark"] .highlight .mo { color: #AE81FF } /* Literal.Number.Oct */
-[data-theme="dark"] .highlight .sa { color: #E6DB74 } /* Literal.String.Affix */
-[data-theme="dark"] .highlight .sb { color: #E6DB74 } /* Literal.String.Backtick */
-[data-theme="dark"] .highlight .sc { color: #E6DB74 } /* Literal.String.Char */
-[data-theme="dark"] .highlight .dl { color: #E6DB74 } /* Literal.String.Delimiter */
-[data-theme="dark"] .highlight .sd { color: #E6DB74 } /* Literal.String.Doc */
-[data-theme="dark"] .highlight .s2 { color: #E6DB74 } /* Literal.String.Double */
-[data-theme="dark"] .highlight .se { color: #AE81FF } /* Literal.String.Escape */
-[data-theme="dark"] .highlight .sh { color: #E6DB74 } /* Literal.String.Heredoc */
-[data-theme="dark"] .highlight .si { color: #E6DB74 } /* Literal.String.Interpol */
-[data-theme="dark"] .highlight .sx { color: #E6DB74 } /* Literal.String.Other */
-[data-theme="dark"] .highlight .sr { color: #E6DB74 } /* Literal.String.Regex */
-[data-theme="dark"] .highlight .s1 { color: #E6DB74 } /* Literal.String.Single */
-[data-theme="dark"] .highlight .ss { color: #E6DB74 } /* Literal.String.Symbol */
-[data-theme="dark"] .highlight .bp { color: #F8F8F2 } /* Name.Builtin.Pseudo */
-[data-theme="dark"] .highlight .fm { color: #A6E22E } /* Name.Function.Magic */
-[data-theme="dark"] .highlight .vc { color: #F8F8F2 } /* Name.Variable.Class */
-[data-theme="dark"] .highlight .vg { color: #F8F8F2 } /* Name.Variable.Global */
-[data-theme="dark"] .highlight .vi { color: #F8F8F2 } /* Name.Variable.Instance */
-[data-theme="dark"] .highlight .vm { color: #F8F8F2 } /* Name.Variable.Magic */
-[data-theme="dark"] .highlight .il { color: #AE81FF } /* Literal.Number.Integer.Long */
-
-        /* Ensure our code metrics override Pygments defaults */
-        .highlight pre {
-            white-space: pre;
-            margin: 0;
-            padding: var(--code-pad-y) 0.75rem !important;
-            line-height: var(--code-line-height) !important;
-            font-size: var(--code-font-size) !important;
-            font-family: 'Cascadia Mono', 'Cascadia Code', 'JetBrains Mono', 'SF Mono', Monaco, 'Consolas', monospace !important;
-            border: none;
-        }
-        .line-numbers { line-height: var(--code-line-height) !important; }
-        .line-numbers .line-number { line-height: var(--code-line-height) !important; }
-
-        /* Custom CSS from frontmatter */
-        
-
-        
-        
-        
-        /* Cursor for tools */
-        body[data-tool="arrow"] .main-content { 
-            cursor: url('data:image/svg+xml;utf8,<svg xmlns="http://www.w3.org/2000/svg" width="24" height="24" viewBox="0 0 24 24" fill="none" stroke="%23e53935" stroke-width="2"><path d="M12 19l7-7 3 3-7 7-3-3z"/><path d="M18 13l-1.5-7.5L2 2l3.5 14.5L13 18l5-5z"/><path d="M2 2l7.586 7.586"/><circle cx="11" cy="11" r="2"/></svg>') 12 12, crosshair;
-        }
-        body[data-tool="pen"] .main-content { 
-            cursor: url('data:image/svg+xml;utf8,<svg xmlns="http://www.w3.org/2000/svg" width="24" height="24" viewBox="0 0 24 24" fill="none" stroke="%23e53935" stroke-width="2"><path d="M12 19l7-7 3 3-7 7-3-3z"/><path d="M18 13l-1.5-7.5L2 2l3.5 14.5L13 18l5-5z"/><circle cx="4" cy="20" r="2" fill="%23e53935"/></svg>') 4 20, pointer;
-        }
-        body[data-tool="eraser"] .main-content { 
-            cursor: url('data:image/svg+xml;utf8,<svg xmlns="http://www.w3.org/2000/svg" width="24" height="24" viewBox="0 0 24 24" fill="none" stroke="%23e53935" stroke-width="2"><path d="M20 20H7l-7-7 7-7h13v14z"/><path d="M13 13l7-7"/><path d="M13 13L9 9"/></svg>') 12 12, auto;
-        }
-
-        /* Color picker styles */
-        .tools-section-title {
-            font-weight: bold;
-            color: var(--text-secondary);
-            font-size: 0.65rem;
-            margin: 0.75rem 0 0.5rem 0;
-            text-transform: uppercase;
-            letter-spacing: 0.5px;
-        }
-        .color-row {
-            display: grid;
-            grid-template-columns: repeat(6, 1fr);
-            gap: 0.25rem;
-            margin-bottom: 0.5rem;
-        }
-        .color-swatch {
-            width: 18px;
-            height: 18px;
-            border: 2px solid var(--border-primary);
-            border-radius: 3px;
-            cursor: pointer;
-            transition: all 0.2s ease;
-            position: relative;
-        }
-        .color-swatch:hover {
-            transform: scale(1.1);
-            border-color: var(--text-secondary);
-        }
-        .color-swatch.selected {
-            border-color: var(--text-primary);
-            box-shadow: 0 0 0 2px var(--text-link);
-        }
-        .color-swatch.selected::after {
-            content: '✓';
-            position: absolute;
-            top: 50%;
-            left: 50%;
-            transform: translate(-50%, -50%);
-            color: white;
-            font-size: 10px;
-            font-weight: bold;
-            text-shadow: 1px 1px 1px black;
-        }
-        .color-input {
-            width: 24px;
-            height: 24px;
-            border: 2px solid var(--border-primary);
-            border-radius: 3px;
-            cursor: pointer;
-            background: none;
-            padding: 0;
-            grid-column: span 2;
-            justify-self: center;
-        }
-        .color-input:hover {
-            border-color: var(--text-secondary);
-        }
-        
-        /* Thickness slider styles */
-        .thickness-row {
-            display: flex;
-            align-items: center;
-            gap: 0.5rem;
-            margin-top: 0.75rem;
-        }
-        .thickness-slider {
-            flex: 1;
-            -webkit-appearance: none;
-            appearance: none;
-            height: 4px;
-            background: var(--border-primary);
-            border-radius: 2px;
-            outline: none;
-            opacity: 0.7;
-            transition: opacity 0.2s;
-        }
-        .thickness-slider:hover {
-            opacity: 1;
-        }
-        .thickness-slider::-webkit-slider-thumb {
-            -webkit-appearance: none;
-            appearance: none;
-            width: 12px;
-            height: 12px;
-            background: var(--text-link);
-            border-radius: 50%;
-            cursor: pointer;
-        }
-        .thickness-slider::-moz-range-thumb {
-            width: 12px;
-            height: 12px;
-            background: var(--text-link);
-            border-radius: 50%;
-            cursor: pointer;
-            border: none;
-        }
-        .thickness-value {
-            font-size: 0.7rem;
-            color: var(--text-secondary);
-            min-width: 20px;
-            text-align: right;
-        }
-
-        .highlight {
-            background: none !important;
-        }
-        
-        /* Loading animations */
-        .loading-spinner {
-            display: inline-block;
-            width: 16px;
-            height: 16px;
-            border: 2px solid var(--border-primary);
-            border-radius: 50%;
-            border-top-color: var(--text-link);
-            animation: spin 1s linear infinite;
-            margin-right: 8px;
-            vertical-align: middle;
-        }
-        
-        @keyframes spin {
-            to { transform: rotate(360deg); }
-        }
-        
-        .loading-skeleton {
-            display: inline-block;
-            background: var(--bg-tertiary);
-            background: linear-gradient(
-                90deg,
-                var(--bg-tertiary) 25%,
-                var(--bg-secondary) 50%,
-                var(--bg-tertiary) 75%
-            );
-            background-size: 200% 100%;
-            animation: loading-shimmer 2s ease-in-out infinite;
-            border-radius: 2px;
-            height: 1em;
-            width: 80px;
-            vertical-align: middle;
-        }
-        
-        @keyframes loading-shimmer {
-            0% { background-position: -200% 0; }
-            100% { background-position: 200% 0; }
-        }
-        
-        /* Loading state for cell output */
-        .cell-output:has(.loading-spinner) {
-            opacity: 0.7;
-            background: var(--bg-secondary);
-            /* border-left: 3px solid var(--text-link); */
-        }
+    --bg-primary: #ffffff;
+    --bg-secondary: #f6f8fa;
+    --bg-tertiary: #f8f9fa;
+    --bg-code: #f8f9fa;
+    --bg-error: #fdf2f2;
+    --bg-artifact: #e6f3ff;
+    --bg-artifact-hover: #d0e7ff;
+
+    --text-primary: #333;
+    --text-secondary: #656d76;
+    --text-error: #c53030;
+    --text-link: #0969da;
+
+    --border-primary: #e1e5e9;
+    --border-error: #e53e3e;
+    --border-cell-failed: #d73a49;
+
+    --shadow: rgba(0, 0, 0, 0.1);
+}
+
+:root[data-theme="dark"] {
+    --bg-primary: #0a0a0a;
+    --bg-secondary: #121212;
+    --bg-tertiary: #181818;
+    --bg-code: #0d0d0d;
+    --bg-error: #1a0f0f;
+    --bg-artifact: #151515;
+    --bg-artifact-hover: #1a1a1a;
+
+    --text-primary: #e0e0e0;
+    --text-secondary: #888888;
+    --text-error: #ff6b6b;
+    --text-link: #64b5f6;
+
+    --border-primary: #2a2a2a;
+    --border-error: #ff6b6b;
+    --border-cell-failed: #ff6b6b;
+
+    --shadow: rgba(255, 255, 255, 0.05);
+}
+
+/* Monocolor UI theme: black/white background, all text/borders single blue */
+:root[data-ui="monocolor"] {
+    --mono-color: #0a66ff;
+}
+
+:root[data-ui="monocolor"][data-theme="light"] {
+    --bg-primary: #ffffff;
+}
+
+:root[data-ui="monocolor"][data-theme="dark"] {
+    --bg-primary: #000000;
+}
+
+:root[data-ui="monocolor"] {
+    --bg-secondary: var(--bg-primary);
+    --bg-tertiary: var(--bg-primary);
+    --bg-code: var(--bg-primary);
+    --bg-error: var(--bg-primary);
+    --bg-artifact: var(--bg-primary);
+    --bg-artifact-hover: var(--bg-primary);
+
+    --text-primary: var(--mono-color);
+    --text-secondary: var(--mono-color);
+    --text-error: var(--mono-color);
+    --text-link: var(--mono-color);
+
+    --border-primary: var(--mono-color);
+    --border-error: var(--mono-color);
+    --border-cell-failed: var(--mono-color);
+
+    --shadow: none;
+}
+
+:root[data-ui="monocolor"] a {
+    color: var(--mono-color);
+}
+
+:root[data-ui="monocolor"] .menu-button,
+:root[data-ui="monocolor"] .theme-toggle,
+:root[data-ui="monocolor"] .reset-toggle,
+:root[data-ui="monocolor"] .back-button {
+    background: var(--bg-primary);
+    color: var(--mono-color);
+    border-color: var(--mono-color);
+}
+
+:root[data-ui="monocolor"] .menu-button:hover,
+:root[data-ui="monocolor"] .theme-toggle:hover,
+:root[data-ui="monocolor"] .reset-toggle:hover,
+:root[data-ui="monocolor"] .back-button:hover {
+    background: var(--bg-primary);
+    color: var(--mono-color);
+    border-color: var(--mono-color);
+}
+
+:root[data-ui="monocolor"] .menu-dropdown {
+    background: var(--bg-primary);
+    border-color: var(--mono-color);
+    box-shadow: none;
+}
+
+:root[data-ui="monocolor"] .menu-item {
+    color: var(--mono-color);
+    border-bottom-color: var(--mono-color);
+}
+
+:root[data-ui="monocolor"] .system-info {
+    background: var(--bg-primary);
+    border-color: var(--mono-color);
+}
+
+:root[data-ui="monocolor"] .cell {
+    border-color: var(--mono-color);
+    background: var(--bg-primary);
+}
+
+:root[data-ui="monocolor"] .cell-header {
+    background: var(--bg-primary);
+    border-bottom-color: var(--mono-color);
+}
+
+:root[data-ui="monocolor"] .artifact {
+    background: var(--bg-primary);
+    border-color: var(--mono-color);
+    color: var(--mono-color);
+}
+
+:root[data-ui="monocolor"] .artifact:hover {
+    background: var(--bg-primary);
+}
+
+:root[data-ui="monocolor"] .artifact-preview img,
+:root[data-ui="monocolor"] .artifact-preview svg {
+    border-color: var(--mono-color);
+}
+
+:root[data-ui="monocolor"] .status-widget {
+    background: var(--bg-primary);
+    border-color: var(--mono-color);
+    color: var(--mono-color);
+}
+
+:root[data-ui="monocolor"] .minimap,
+:root[data-ui="monocolor"] .file-explorer,
+:root[data-ui="monocolor"] .tools-widget {
+    background: var(--bg-primary);
+    border-color: var(--mono-color);
+    color: var(--mono-color);
+}
+
+:root[data-ui="monocolor"] .cell-code {
+    background: var(--bg-primary);
+    border-bottom-color: var(--mono-color);
+}
+
+:root[data-ui="monocolor"] .tools-title,
+:root[data-ui="monocolor"] .file-explorer-section-title,
+:root[data-ui="monocolor"] .minimap-title {
+    color: var(--mono-color);
+    border-bottom-color: var(--mono-color);
+}
+
+:root[data-ui="monocolor"] .tool-button {
+    background: var(--bg-primary);
+    border-color: var(--mono-color);
+    color: var(--mono-color);
+}
+
+:root[data-ui="monocolor"] .tool-button.active {
+    border-color: var(--mono-color);
+}
+
+:root[data-ui="monocolor"] .file-explorer-item,
+:root[data-ui="monocolor"] .minimap-item {
+    color: var(--mono-color);
+}
+
+/* Force Pygments code to mono blue on mono bg */
+:root[data-ui="monocolor"] .highlight {
+    background: var(--bg-primary) !important;
+    color: var(--mono-color) !important;
+}
+
+:root[data-ui="monocolor"] .highlight *,
+:root[data-ui="monocolor"] .highlight .hll {
+    color: var(--mono-color) !important;
+    background: transparent !important;
+    border-color: var(--mono-color) !important;
+}
+
+/* Default code font + metrics (overridable via frontmatter) */
+:root {
+    --code-font-size: 0.95rem;
+    --code-line-height: 1.5;
+    --code-pad-y: 0.75rem;
+}
+
+/* Minimal UI theme overrides base variables for a flatter, 90s look */
+:root[data-ui="none"] {
+    --bg-primary: #ffffff;
+    --bg-secondary: transparent;
+    --bg-tertiary: transparent;
+    --bg-code: #f9f9f9;
+    --bg-error: #fff0f0;
+    --bg-artifact: #f0f7ff;
+    --bg-artifact-hover: #e5f1ff;
+
+    --text-primary: #000000;
+    --text-secondary: #222222;
+    --text-error: #a00000;
+    --text-link: #0000ee;
+
+    --border-primary: #cccccc;
+    --border-error: #cc0000;
+    --border-cell-failed: #cc0000;
+
+    --shadow: none;
+}
+
+html {
+    overscroll-behavior: none;
+}
+
+body {
+    font-family: 'Cascadia Mono', 'Cascadia Code', 'JetBrains Mono', 'SF Mono', Monaco, 'Consolas', monospace;
+    line-height: 1.4;
+    max-width: 1000px;
+    margin: 0 auto;
+    padding: 15px;
+    color: var(--text-primary);
+    background: var(--bg-primary);
+    transition: background-color 0.2s ease, color 0.2s ease;
+    overscroll-behavior: none;
+}
+
+/* Minimal "none" UI theme overrides */
+:root[data-ui="none"] body {
+    font-family: 'Times New Roman', Times, serif;
+    line-height: 1.5;
+    max-width: 860px;
+    padding: 12px;
+    background: #ffffff;
+    color: #000000;
+    transition: none;
+}
+
+/* Two panel layout removed */
+
+.controls {
+    position: fixed;
+    top: 20px;
+    right: 20px;
+    display: flex;
+    flex-direction: column;
+    align-items: flex-end;
+    gap: 0.25rem;
+    z-index: 1000;
+}
+
+.controls-buttons {
+    display: flex;
+    gap: 0.5rem;
+}
+
+.menu-button {
+    position: relative;
+    background: var(--bg-secondary);
+    border: 1px solid var(--border-primary);
+    padding: 8px 12px;
+    border-radius: 2px;
+    color: var(--text-secondary);
+    cursor: pointer;
+    font-family: inherit;
+    font-size: 0.9rem;
+    user-select: none;
+}
+
+/* Keep default control styling when widgets are enabled, even in minimal UI */
+:root[data-ui="none"][data-widgets="on"] .menu-button,
+:root[data-ui="none"][data-widgets="on"] .theme-toggle,
+:root[data-ui="none"][data-widgets="on"] .reset-toggle,
+:root[data-ui="none"][data-widgets="on"] .back-button {
+    background: #f6f6f6;
+    border: 1px solid #cccccc;
+    color: #222222;
+}
+
+.menu-button:hover {
+    color: var(--text-primary);
+    background: var(--bg-tertiary);
+}
+
+/* Controls state indicator (top-right) */
+/* Status widget (bottom-right) */
+.status-widget {
+    position: fixed;
+    right: 20px;
+    bottom: 20px;
+    width: auto;
+    max-width: 260px;
+    background: var(--bg-secondary);
+    border: 1px solid var(--border-primary);
+    border-radius: 2px;
+    padding: 6px 8px;
+    font-size: 0.8rem;
+    color: var(--text-secondary);
+    z-index: 100;
+}
+
+.status-widget strong {
+    color: var(--text-primary);
+}
+
+:root[data-ui="none"][data-widgets="on"] .status-widget {
+    background: #f6f6f6;
+    border-color: #ccc;
+    color: #222;
+}
+
+:root[data-ui="none"][data-widgets="on"] .menu-button:hover,
+:root[data-ui="none"][data-widgets="on"] .theme-toggle:hover,
+:root[data-ui="none"][data-widgets="on"] .reset-toggle:hover,
+:root[data-ui="none"][data-widgets="on"] .back-button:hover {
+    background: #ededed;
+    border-color: #bbbbbb;
+    color: #000000;
+}
+
+.menu-dropdown {
+    position: absolute;
+    top: 100%;
+    right: 0;
+    background: var(--bg-secondary);
+    border: 1px solid var(--border-primary);
+    border-radius: 4px;
+    box-shadow: 0 4px 12px var(--shadow);
+    min-width: 160px;
+    opacity: 0;
+    visibility: hidden;
+    transform: translateY(-8px);
+    transition: all 0.2s ease;
+    z-index: 1001;
+    margin-top: 4px;
+}
+
+:root[data-ui="none"][data-widgets="on"] .menu-dropdown {
+    background: #ffffff;
+    border: 1px solid #cccccc;
+    box-shadow: none;
+}
+
+.menu-button.active .menu-dropdown {
+    opacity: 1;
+    visibility: visible;
+    transform: translateY(0);
+}
+
+.menu-item {
+    display: block;
+    padding: 8px 12px;
+    color: var(--text-secondary);
+    text-decoration: none;
+    font-size: 0.85rem;
+    border-bottom: 1px solid var(--border-primary);
+    cursor: pointer;
+}
+
+:root[data-ui="none"] .menu-item {
+    color: #000;
+    border-bottom: 1px solid #eee;
+}
+
+.menu-item:last-child {
+    border-bottom: none;
+}
+
+.menu-item:hover {
+    background: var(--bg-tertiary);
+    color: var(--text-primary);
+}
+
+.menu-checkbox {
+    display: inline-block;
+    width: 16px;
+    font-family: monospace;
+    color: var(--text-link);
+}
+
+.theme-toggle,
+.reset-toggle,
+.back-button {
+    background: var(--bg-secondary);
+    border: 1px solid var(--border-primary);
+    padding: 8px 12px;
+    border-radius: 4px;
+    color: var(--text-secondary);
+    cursor: pointer;
+    font-family: inherit;
+    font-size: 0.9rem;
+    user-select: none;
+}
+
+.back-button {
+    text-decoration: none;
+    display: inline-block;
+}
+
+.theme-toggle:hover,
+.reset-toggle:hover,
+.back-button:hover {
+    color: var(--text-primary);
+    background: var(--bg-tertiary);
+}
+
+.system-info {
+    background: var(--bg-secondary);
+    border: 1px solid var(--border-primary);
+    border-radius: 4px;
+    padding: 8px 12px;
+    margin-bottom: 16px;
+    font-size: 0.85em;
+    color: var(--text-secondary);
+}
+
+.system-info-header {
+    font-weight: 600;
+    color: var(--text-primary);
+    margin-bottom: 2px;
+}
+
+.system-info-content {
+    font-family: monospace;
+}
+
+.theme-toggle,
+.reset-toggle {
+    background: var(--bg-secondary);
+    border: 1px solid var(--border-primary);
+    border-radius: 2px;
+    /* padding: 0.4rem 0.6rem; */
+    cursor: pointer;
+    font-family: inherit;
+    font-size: 0.8rem;
+    color: var(--text-secondary);
+    user-select: none;
+    transition: all 0.2s ease;
+    text-transform: lowercase;
+    letter-spacing: 0;
+}
+
+.theme-toggle:hover,
+.reset-toggle:hover {
+    background: var(--bg-tertiary);
+    border-color: var(--text-secondary);
+    color: var(--text-primary);
+}
+
+.minimap {
+    position: fixed;
+    bottom: 20px;
+    right: 20px;
+    width: 220px;
+    max-height: 400px;
+    background: var(--bg-secondary);
+    border: 1px solid var(--border-primary);
+    border-radius: 2px;
+    padding: 0.5rem;
+    font-size: 0.7rem;
+    overflow-y: auto;
+    z-index: 100;
+    opacity: 0.9;
+    transition: opacity 0.2s ease;
+}
+
+/* Hide widgets and controls when disabled via frontmatter */
+:root[data-widgets="off"] .controls,
+:root[data-widgets="off"] .minimap,
+:root[data-widgets="off"] .file-explorer,
+:root[data-widgets="off"] .tools-widget,
+:root[data-widgets="off"] .status-widget {
+    display: none !important;
+}
+
+.file-explorer {
+    position: fixed;
+    bottom: 20px;
+    /* default; JS will stack */
+    right: 20px;
+    left: auto;
+    top: auto;
+    width: 220px;
+    max-height: 400px;
+    background: var(--bg-secondary);
+    border: 1px solid var(--border-primary);
+    border-radius: 2px;
+    padding: 0.5rem;
+    font-size: 0.7rem;
+    overflow-y: auto;
+    z-index: 100;
+    opacity: 0.9;
+    transition: opacity 0.2s ease;
+}
+
+/* Drawing overlay */
+.draw-overlay {
+    position: fixed;
+    top: 0;
+    left: 0;
+    width: 100vw;
+    height: 100vh;
+    z-index: 80;
+    /* under widgets (100) and controls (1000) */
+    display: block;
+    pointer-events: none;
+    /* enabled only when a tool is active */
+}
+
+/* Tools widget */
+.tools-widget {
+    position: fixed;
+    bottom: 20px;
+    /* default; JS will stack */
+    right: 20px;
+    left: auto;
+    top: auto;
+    width: 220px;
+    background: var(--bg-secondary);
+    border: 1px solid var(--border-primary);
+    border-radius: 2px;
+    padding: 0.5rem;
+    font-size: 0.7rem;
+    z-index: 100;
+    opacity: 0.95;
+}
+
+.tools-title {
+    font-weight: bold;
+    color: var(--text-secondary);
+    margin-bottom: 0.5rem;
+    padding-bottom: 0.25rem;
+    border-bottom: 1px solid var(--border-primary);
+    cursor: grab;
+    user-select: none;
+}
+
+.tools-row {
+    display: flex;
+    gap: 0.4rem;
+    flex-wrap: wrap;
+}
+
+.tool-button {
+    background: var(--bg-tertiary);
+    border: 1px solid var(--border-primary);
+    border-radius: 2px;
+    padding: 0.25rem 0.4rem;
+    cursor: pointer;
+    color: var(--text-secondary);
+    font-family: inherit;
+    font-size: 0.75rem;
+    user-select: none;
+}
+
+.tool-button:hover {
+    color: var(--text-primary);
+}
+
+.tool-button.active {
+    color: var(--text-primary);
+    border-color: var(--text-secondary);
+    background: var(--bg-secondary);
+}
+
+.minimap:hover,
+.file-explorer:hover {
+    opacity: 1;
+}
+
+.minimap-title {
+    font-weight: bold;
+    color: var(--text-secondary);
+    margin-bottom: 0.5rem;
+    padding-bottom: 0.25rem;
+    border-bottom: 1px solid var(--border-primary);
+    cursor: grab;
+    /* drag handle */
+    user-select: none;
+}
+
+.minimap-item {
+    display: block;
+    color: var(--text-secondary);
+    text-decoration: none;
+    padding: 0.15rem 0;
+    border-left: 2px solid transparent;
+    padding-left: 0.5rem;
+    transition: all 0.2s ease;
+    cursor: pointer;
+}
+
+.minimap-item:hover {
+    color: var(--text-primary);
+    border-left-color: var(--text-secondary);
+}
+
+.minimap-item.active {
+    color: var(--text-primary);
+    border-left-color: var(--text-link);
+}
+
+.minimap-heading {
+    font-weight: normal;
+}
+
+.minimap-heading.h1 {
+    padding-left: 0.5rem;
+}
+
+.minimap-heading.h2 {
+    padding-left: 1rem;
+}
+
+.minimap-heading.h3 {
+    padding-left: 1.5rem;
+}
+
+.minimap-heading.h4 {
+    padding-left: 2rem;
+}
+
+.minimap-heading.h5 {
+    padding-left: 2.5rem;
+}
+
+.minimap-heading.h6 {
+    padding-left: 3rem;
+}
+
+.minimap-cell {
+    color: var(--text-link);
+    opacity: 0.8;
+    font-style: italic;
+}
+
+.minimap-cell:hover {
+    opacity: 1;
+}
+
+.file-explorer-title {
+    font-weight: bold;
+    color: var(--text-secondary);
+    margin-bottom: 0.5rem;
+    padding-bottom: 0.25rem;
+    border-bottom: 1px solid var(--border-primary);
+    cursor: grab;
+    /* drag handle */
+    user-select: none;
+}
+
+.file-explorer-section {
+    margin-bottom: 0.75rem;
+}
+
+.file-explorer-section-title {
+    font-weight: bold;
+    color: var(--text-secondary);
+    font-size: 0.65rem;
+    margin-bottom: 0.25rem;
+    text-transform: uppercase;
+    letter-spacing: 0.5px;
+}
+
+.file-explorer-item {
+    display: block;
+    color: var(--text-secondary);
+    text-decoration: none;
+    padding: 0.1rem 0;
+    margin-left: 0.5rem;
+    transition: color 0.2s ease;
+    cursor: pointer;
+    font-family: monospace;
+}
+
+.file-explorer-item:hover {
+    color: var(--text-primary);
+}
+
+.file-explorer-item.script {
+    color: var(--text-link);
+}
+
+.file-explorer-item.artifact {
+    color: var(--text-secondary);
+    opacity: 0.8;
+}
+
+
+/* Hide widgets on smaller screens */
+@media (max-width: 768px) {
+
+    .minimap,
+    .file-explorer,
+    .tools-widget {
+        display: none;
+    }
+}
+
+.cell {
+    margin: 1rem 0;
+    border: 1px solid var(--border-primary);
+    border-radius: 2px;
+    overflow: hidden;
+    background: var(--bg-secondary);
+}
+
+:root[data-ui="none"] .cell {
+    margin: 1em 0;
+    border: none;
+    background: transparent;
+}
+
+.cell-header {
+    background: var(--bg-secondary);
+    padding: 0.5rem 1rem;
+    border-bottom: 1px solid var(--border-primary);
+    font-family: inherit;
+    font-size: 0.85rem;
+}
+
+:root[data-ui="none"] .cell-header {
+    background: transparent;
+    border: none;
+    padding: 0;
+    font-weight: bold;
+}
+
+:root[data-ui="none"] .cell-content {
+    padding: 0;
+}
+
+:root[data-ui="none"] .copy-button,
+:root[data-ui="none"] .collapse-indicators,
+:root[data-ui="none"] .cell-meta,
+:root[data-ui="none"] .cell-outputs-header {
+    display: none !important;
+}
+
+:root[data-ui="none"] pre,
+:root[data-ui="none"] code {
+    font-family: Menlo, Monaco, 'Courier New', monospace;
+}
+
+:root[data-ui="none"] .code-content pre {
+    background: #f9f9f9;
+    border: 1px solid #ddd;
+    padding: 8px;
+}
+
+:root[data-ui="none"] .output {
+    background: transparent;
+    border: none;
+    padding: 0.25em 0;
+}
+
+color: var(--text-secondary);
+cursor: pointer;
+user-select: none;
+transition: background-color 0.2s ease;
+}
+
+.cell-header:hover {
+    background: var(--bg-tertiary);
+}
+
+.collapse-indicators {
+    color: var(--text-secondary);
+    font-size: 0.8rem;
+    opacity: 0.7;
+}
+
+.collapse-indicators span:hover {
+    color: var(--text-primary);
+    opacity: 1;
+}
+
+.cell-code {
+    display: block;
+    background: var(--bg-code);
+}
+
+.cell-code.collapsed {
+    display: none;
+}
+
+.cell-code pre {
+    margin: 0;
+    padding: 0.75rem;
+    background: var(--bg-code);
+    overflow-x: auto;
+    color: var(--text-primary);
+}
+
+.cell-output {
+    padding: 0.75rem;
+    /* background: var(--bg-primary); */
+    background: var(--bg-secondary);
+}
+
+.cell-output.collapsed {
+    display: none;
+}
+
+.cell-stdout {
+    background: var(--bg-tertiary);
+    padding: 0.75rem;
+    border-radius: 1px;
+    /* margin: 0.25rem 0; */
+    font-family: inherit;
+    font-size: 0.9rem;
+    white-space: pre-wrap;
+    color: var(--text-primary);
+}
+
+.cell-stdout {
+    background: var(--bg-tertiary);
+    padding: 0.75rem;
+    border-radius: 1px;
+    font-family: inherit;
+    font-size: 0.9rem;
+    color: var(--text-primary);
+
+    /* key bits */
+    overflow: auto;
+    /* show scrollbars when needed */
+    max-width: 100%;
+    /* respects whatever layout width you give it */
+}
+
+.cell-stdout .stdout-text {
+    margin: 0;
+    /* reset pre default margin */
+    white-space: pre;
+    /* keep line breaks, NO wrapping */
+    display: inline-block;
+    /* shrink-to-content */
+    min-width: max-content;
+    /* allow very long lines to define intrinsic width */
+    font-family: ui-monospace, SFMono-Regular, Menlo, Consolas, monospace;
+    tab-size: 2;
+}
+
+.cell-stderr {
+    background: var(--bg-error);
+    border-left: 2px solid var(--border-error);
+    padding: 1rem;
+    margin: 0.5rem 0;
+    font-family: inherit;
+    font-size: 0.9rem;
+    color: var(--text-error);
+    white-space: pre-wrap;
+}
+
+.uv-install-logs {
+    margin: 0.5rem 0;
+}
+
+.uv-logs-header {
+    cursor: pointer;
+    padding: 0.75rem;
+    border-left: 3px solid var(--border-color);
+    font-family: inherit;
+    font-size: 0.85rem;
+    color: var(--text-secondary);
+    user-select: none;
+}
+
+.uv-logs-content {
+    background: var(--bg-secondary);
+    padding: 1rem;
+    border-left: 3px solid var(--border-color);
+    white-space: pre-wrap;
+    font-family: monospace;
+    font-size: 0.85rem;
+    color: var(--text-secondary);
+    overflow-x: auto;
+}
+
+.cell-artifacts {
+    margin: 1rem 0;
+}
+
+.cell-artifacts h4 {
+    margin: 0 0 0.5rem 0;
+    color: var(--text-secondary);
+    font-size: 0.9rem;
+}
+
+.artifact {
+    display: inline-block;
+    background: var(--bg-artifact);
+    padding: 0.25rem 0.5rem;
+    border-radius: 1px;
+    margin: 0.25rem 0.5rem 0.25rem 0;
+    font-family: inherit;
+    font-size: 0.8rem;
+    color: var(--text-link);
+    text-decoration: none;
+    transition: background-color 0.2s ease;
+    border: 1px solid var(--border-primary);
+}
+
+.artifact:hover {
+    background: var(--bg-artifact-hover);
+}
+
+.artifact-preview {
+    margin-top: 1rem;
+}
+
+.artifact-preview img {
+    max-width: 100%;
+    height: auto;
+    border: 1px solid var(--border-primary);
+    border-radius: 1px;
+}
+
+.artifact-preview svg {
+    max-width: 100%;
+    height: auto;
+    border: 1px solid var(--border-primary);
+    border-radius: 1px;
+    display: block;
+}
+
+/* Style SVG text elements */
+.artifact-preview svg g {
+    fill: var(--text-primary) !important;
+}
+
+/* Auto-theme SVG elements */
+.artifact-preview svg {
+    background: transparent;
+}
+
+/* Invert SVG images in dark mode */
+:root[data-theme="dark"] .artifact-preview img[src$=".svg"] {
+    filter: invert(0.9) hue-rotate(180deg);
+}
+
+/* Keep SVG images readable in monocolor mode */
+:root[data-ui="monocolor"] .artifact-preview img[src$=".svg"] {
+    filter: none;
+}
+
+/* CSV table styling */
+.artifact-csv {
+    margin-top: 1rem;
+    overflow-x: auto;
+}
+
+.csv-table {
+    width: 100%;
+    border-collapse: collapse;
+    font-size: 0.9rem;
+    background: var(--bg-secondary);
+    border: 1px solid var(--border-primary);
+    border-radius: 1px;
+}
+
+.csv-table th,
+.csv-table td {
+    padding: 0.5rem 0.75rem;
+    text-align: left;
+    border: 1px solid var(--border-primary);
+}
+
+.csv-table th {
+    background: var(--bg-tertiary);
+    font-weight: 600;
+    color: var(--text-primary);
+}
+
+.csv-table tbody tr:hover {
+    background: var(--bg-artifact-hover);
+}
+
+.artifact-csv-error {
+    margin-top: 1rem;
+    padding: 1rem;
+    background: var(--bg-error);
+    color: var(--text-error);
+    border: 1px solid var(--border-error);
+    border-radius: 1px;
+}
+
+.cell-failed {
+    border-color: var(--border-cell-failed);
+}
+
+.cell-failed .cell-header {
+    background: var(--bg-error);
+    color: var(--text-error);
+}
+
+.cell-commented {
+    opacity: 0.6;
+    border-style: dashed;
+}
+
+.cell-commented .cell-header {
+    background: var(--bg-secondary);
+    color: var(--text-secondary);
+    font-style: italic;
+}
+
+.run-btn {
+    background: var(--bg-tertiary);
+    border: 1px solid var(--border-primary);
+    padding: 2px 6px;
+    border-radius: 2px;
+    color: var(--text-secondary);
+    cursor: pointer;
+    font-size: 0.75em;
+    font-family: inherit;
+    margin-left: 4px;
+}
+
+.run-btn:hover {
+    color: var(--text-primary);
+    background: var(--bg-primary);
+}
+
+.run-btn:disabled {
+    opacity: 0.6;
+    cursor: not-allowed;
+}
+
+.copy-btn {
+    background: var(--bg-tertiary);
+    border: 1px solid var(--border-primary);
+    padding: 2px 6px;
+    border-radius: 2px;
+    color: var(--text-secondary);
+    cursor: pointer;
+    font-size: 0.75em;
+    font-family: inherit;
+    margin-left: 4px;
+}
+
+.copy-btn:hover {
+    color: var(--text-primary);
+    background: var(--bg-primary);
+}
+
+.copy-btn:disabled {
+    opacity: 0.6;
+    cursor: not-allowed;
+}
+
+.copy-btn.copied {
+    color: #4caf50;
+    background: var(--bg-primary);
+    border-color: #4caf50;
+    transition: all 0.2s ease;
+}
+
+.raw-btn {
+    background: var(--bg-tertiary);
+    border: 1px solid var(--border-primary);
+    padding: 2px 6px;
+    border-radius: 2px;
+    color: var(--text-secondary);
+    cursor: pointer;
+    font-size: 0.75em;
+    font-family: inherit;
+    margin-left: 4px;
+    text-decoration: none;
+    display: inline-block;
+}
+
+.raw-btn:hover {
+    color: var(--text-primary);
+    background: var(--bg-primary);
+    text-decoration: none;
+}
+
+.github-btn {
+    background: var(--bg-tertiary);
+    border: 1px solid var(--border-primary);
+    padding: 2px 6px;
+    border-radius: 2px;
+    color: var(--text-secondary);
+    cursor: pointer;
+    font-size: 0.75em;
+    font-family: inherit;
+    margin-left: 4px;
+    text-decoration: none;
+    display: inline-block;
+}
+
+.github-btn:hover {
+    color: var(--text-primary);
+    background: var(--bg-primary);
+    text-decoration: none;
+}
+
+.hf-btn {
+    background: var(--bg-tertiary);
+    border: 1px solid var(--border-primary);
+    padding: 2px 6px;
+    border-radius: 2px;
+    color: var(--text-secondary);
+    cursor: pointer;
+    font-size: 0.75em;
+    font-family: inherit;
+    margin-left: 4px;
+    text-decoration: none;
+    display: inline-block;
+}
+
+.hf-btn:hover {
+    color: var(--text-primary);
+    background: var(--bg-primary);
+    text-decoration: none;
+}
+
+.output-stale {
+    opacity: 0.5;
+    position: relative;
+}
+
+.output-stale::after {
+    content: '⏳ updating...';
+    position: absolute;
+    top: 8px;
+    right: 8px;
+    background: var(--bg-secondary);
+    padding: 4px 8px;
+    border-radius: 2px;
+    font-size: 0.75em;
+    color: var(--text-secondary);
+    border: 1px solid var(--border-primary);
+}
+
+h1,
+h2,
+h3,
+h4,
+h5,
+h6 {
+    margin-top: 1.5rem;
+    margin-bottom: 0.75rem;
+    color: var(--text-primary);
+}
+
+h1 {
+    margin-top: 0;
+    margin-bottom: 1rem;
+}
+
+p {
+    margin: 0.75rem 0;
+    color: var(--text-primary);
+}
+
+a {
+    color: var(--text-link);
+}
+
+img {
+    max-width: 100%;
+    height: auto;
+    border-radius: 1px;
+    box-shadow: none;
+}
+
+pre,
+code {
+    font-family: 'Cascadia Mono', 'Cascadia Code', 'JetBrains Mono', 'SF Mono', Monaco, 'Consolas', monospace;
+    font-size: var(--code-font-size);
+}
+
+.code-wrap {
+    position: relative;
+}
+
+.code-line-highlight {
+    display: none;
+    position: absolute;
+    left: 0;
+    right: 0;
+    height: 1.5em;
+    background: rgba(255, 235, 170, 0.35);
+    pointer-events: none;
+    border-left: 3px solid #f4c542;
+}
+
+.line-number {
+    cursor: pointer;
+    text-decoration: none;
+    color: var(--text-secondary);
+    padding: 0 0.25rem;
+}
+
+.line-number.selected {
+    background: rgba(255, 235, 170, 0.4);
+    color: var(--text-primary);
+}
+
+/* Line numbers */
+.highlight-with-lines {
+    display: flex;
+}
+
+.line-numbers {
+    background: var(--bg-tertiary);
+    padding: var(--code-pad-y) 0.5rem;
+    font-family: 'Cascadia Mono', 'Cascadia Code', 'JetBrains Mono', 'SF Mono', Monaco, 'Consolas', monospace;
+    font-size: var(--code-font-size);
+    line-height: var(--code-line-height);
+    color: var(--text-secondary);
+    user-select: none;
+    text-align: right;
+    border-right: 1px solid var(--border-primary);
+}
+
+.line-numbers .line-number {
+    display: block;
+    line-height: var(--code-line-height);
+}
+
+.highlight-with-lines .highlight {
+    flex: 1;
+}
+
+.highlight .hll {
+    background-color: transparent;
+}
+
+/* don't conflict with our highlight */
+.highlight pre {
+    white-space: pre;
+    margin: 0;
+    padding: var(--code-pad-y) 0.75rem;
+    line-height: var(--code-line-height);
+}
+
+/* Collapsed code styling */
+.cell-code.collapsed {
+    display: none;
+}
+
+.cell-code.expanded {
+    display: block;
+}
+
+    {
+    % if config.collapse_code %
+}
+
+.cell-code {
+    display: none;
+}
+
+    {
+    % else %
+}
+
+.cell-code {
+    display: block;
+    border-bottom: 1px solid var(--border-primary);
+}
+
+    {
+    % endif %
+}
+
+    {
+        {
+        pygments_css
+    }
+}
+
+/* Ensure our code metrics override Pygments defaults */
+.highlight pre {
+    white-space: pre;
+    margin: 0;
+    padding: var(--code-pad-y) 0.75rem !important;
+    line-height: var(--code-line-height) !important;
+    font-size: var(--code-font-size) !important;
+    font-family: 'Cascadia Mono', 'Cascadia Code', 'JetBrains Mono', 'SF Mono', Monaco, 'Consolas', monospace !important;
+    border: none;
+}
+
+.line-numbers {
+    line-height: var(--code-line-height) !important;
+}
+
+.line-numbers .line-number {
+    line-height: var(--code-line-height) !important;
+}
+
+/* Custom CSS from frontmatter */
+    {
+        {
+        config.custom_css
+    }
+}
+
+    {
+    # Override code font size from frontmatter (accept number as px) #
+}
+
+    {
+    % if config.code_font_size is not none %
+}
+
+    {
+    % if config.code_font_size is string %
+}
+
+:root {
+    --code-font-size: {
+            {
+            config.code_font_size
+        }
+    }
+
+    ;
+}
+
+    {
+    % else %
+}
+
+:root {
+    --code-font-size: {
+            {
+            config.code_font_size
+        }
+    }
+
+    px;
+}
+
+    {
+    % endif %
+}
+
+    {
+    % endif %
+}
+
+/* Cursor for tools */
+body[data-tool="arrow"] .main-content {
+    cursor: url('data:image/svg+xml;utf8,<svg xmlns="http://www.w3.org/2000/svg" width="24" height="24" viewBox="0 0 24 24" fill="none" stroke="%23e53935" stroke-width="2"><path d="M12 19l7-7 3 3-7 7-3-3z"/><path d="M18 13l-1.5-7.5L2 2l3.5 14.5L13 18l5-5z"/><path d="M2 2l7.586 7.586"/><circle cx="11" cy="11" r="2"/></svg>') 12 12, crosshair;
+}
+
+body[data-tool="pen"] .main-content {
+    cursor: url('data:image/svg+xml;utf8,<svg xmlns="http://www.w3.org/2000/svg" width="24" height="24" viewBox="0 0 24 24" fill="none" stroke="%23e53935" stroke-width="2"><path d="M12 19l7-7 3 3-7 7-3-3z"/><path d="M18 13l-1.5-7.5L2 2l3.5 14.5L13 18l5-5z"/><circle cx="4" cy="20" r="2" fill="%23e53935"/></svg>') 4 20, pointer;
+}
+
+body[data-tool="eraser"] .main-content {
+    cursor: url('data:image/svg+xml;utf8,<svg xmlns="http://www.w3.org/2000/svg" width="24" height="24" viewBox="0 0 24 24" fill="none" stroke="%23e53935" stroke-width="2"><path d="M20 20H7l-7-7 7-7h13v14z"/><path d="M13 13l7-7"/><path d="M13 13L9 9"/></svg>') 12 12, auto;
+}
+
+/* Color picker styles */
+.tools-section-title {
+    font-weight: bold;
+    color: var(--text-secondary);
+    font-size: 0.65rem;
+    margin: 0.75rem 0 0.5rem 0;
+    text-transform: uppercase;
+    letter-spacing: 0.5px;
+}
+
+.color-row {
+    display: grid;
+    grid-template-columns: repeat(6, 1fr);
+    gap: 0.25rem;
+    margin-bottom: 0.5rem;
+}
+
+.color-swatch {
+    width: 18px;
+    height: 18px;
+    border: 2px solid var(--border-primary);
+    border-radius: 3px;
+    cursor: pointer;
+    transition: all 0.2s ease;
+    position: relative;
+}
+
+.color-swatch:hover {
+    transform: scale(1.1);
+    border-color: var(--text-secondary);
+}
+
+.color-swatch.selected {
+    border-color: var(--text-primary);
+    box-shadow: 0 0 0 2px var(--text-link);
+}
+
+.color-swatch.selected::after {
+    content: '✓';
+    position: absolute;
+    top: 50%;
+    left: 50%;
+    transform: translate(-50%, -50%);
+    color: white;
+    font-size: 10px;
+    font-weight: bold;
+    text-shadow: 1px 1px 1px black;
+}
+
+.color-input {
+    width: 24px;
+    height: 24px;
+    border: 2px solid var(--border-primary);
+    border-radius: 3px;
+    cursor: pointer;
+    background: none;
+    padding: 0;
+    grid-column: span 2;
+    justify-self: center;
+}
+
+.color-input:hover {
+    border-color: var(--text-secondary);
+}
+
+/* Thickness slider styles */
+.thickness-row {
+    display: flex;
+    align-items: center;
+    gap: 0.5rem;
+    margin-top: 0.75rem;
+}
+
+.thickness-slider {
+    flex: 1;
+    -webkit-appearance: none;
+    appearance: none;
+    height: 4px;
+    background: var(--border-primary);
+    border-radius: 2px;
+    outline: none;
+    opacity: 0.7;
+    transition: opacity 0.2s;
+}
+
+.thickness-slider:hover {
+    opacity: 1;
+}
+
+.thickness-slider::-webkit-slider-thumb {
+    -webkit-appearance: none;
+    appearance: none;
+    width: 12px;
+    height: 12px;
+    background: var(--text-link);
+    border-radius: 50%;
+    cursor: pointer;
+}
+
+.thickness-slider::-moz-range-thumb {
+    width: 12px;
+    height: 12px;
+    background: var(--text-link);
+    border-radius: 50%;
+    cursor: pointer;
+    border: none;
+}
+
+.thickness-value {
+    font-size: 0.7rem;
+    color: var(--text-secondary);
+    min-width: 20px;
+    text-align: right;
+}
+
+.highlight {
+    background: none !important;
+}
+
+/* Loading animations */
+.loading-spinner {
+    display: inline-block;
+    width: 16px;
+    height: 16px;
+    border: 2px solid var(--border-primary);
+    border-radius: 50%;
+    border-top-color: var(--text-link);
+    animation: spin 1s linear infinite;
+    margin-right: 8px;
+    vertical-align: middle;
+}
+
+@keyframes spin {
+    to {
+        transform: rotate(360deg);
+    }
+}
+
+.loading-skeleton {
+    display: inline-block;
+    background: var(--bg-tertiary);
+    background: linear-gradient(90deg,
+            var(--bg-tertiary) 25%,
+            var(--bg-secondary) 50%,
+            var(--bg-tertiary) 75%);
+    background-size: 200% 100%;
+    animation: loading-shimmer 2s ease-in-out infinite;
+    border-radius: 2px;
+    height: 1em;
+    width: 80px;
+    vertical-align: middle;
+}
+
+@keyframes loading-shimmer {
+    0% {
+        background-position: -200% 0;
+    }
+
+    100% {
+        background-position: 200% 0;
+    }
+}
+
+/* Loading state for cell output */
+.cell-output:has(.loading-spinner) {
+    opacity: 0.7;
+    background: var(--bg-secondary);
+    /* border-left: 3px solid var(--text-link); */
+}
     </style>
     <script>
 // --- Drag utilities ---
@@ -3879,7 +4106,7 @@ span.linenos.special { color: #000000; background-color: #ffffc0; padding-left:
 <span onclick="toggleOutput('nv')" style="cursor: pointer;">▼ output</span>
  <span id="uv-indicator-nv" style="cursor: default; opacity: 0.3;">▶ uv-logs</span>
 </span> | 
-Cell: nv | 0.23s
+Cell: nv | 0.21s
  | <button class="run-btn" onclick="runCell('nv')">▶ run</button>
 <button class="copy-btn" onclick="copyCell('nv')">Copy</button>
 <a href="cells/nv.py" target="_blank" class="raw-btn">Raw</a>
@@ -3895,7 +4122,7 @@ Cell: nv | 0.23s
 </div>
 </div>
 <div id="output-nv" class="cell-output">
-<div class="cell-stdout"><pre class="stdout-text">Wed Oct 29 15:50:24 2025       
+<div class="cell-stdout"><pre class="stdout-text">Thu Oct 30 15:52:23 2025       
 +-----------------------------------------------------------------------------------------+
 | NVIDIA-SMI 570.195.03             Driver Version: 570.195.03     CUDA Version: 12.8     |
 |-----------------------------------------+------------------------+----------------------+
@@ -3904,7 +4131,7 @@ Cell: nv | 0.23s
 |                                         |                        |               MIG M. |
 |=========================================+========================+======================|
 |   0  NVIDIA L40S                    On  |   00000000:4D:00.0 Off |                    0 |
-| N/A   29C    P0             88W /  350W |       0MiB /  46068MiB |      0%      Default |
+| N/A   30C    P0             76W /  350W |       0MiB /  46068MiB |     11%      Default |
 |                                         |                        |                  N/A |
 +-----------------------------------------+------------------------+----------------------+
 
@@ -3926,9 +4153,9 @@ Cell: nv | 0.23s
 <span class="collapse-indicators">
 <span onclick="toggleCode('benchmark')" style="cursor: pointer;">▼ code</span> 
 <span onclick="toggleOutput('benchmark')" style="cursor: pointer;">▼ output</span>
- <span id="uv-indicator-benchmark" onclick="toggleUvLogsFromHeader('benchmark')" style="cursor: pointer;">▶ uv-logs</span>
+ <span id="uv-indicator-benchmark" style="cursor: default; opacity: 0.3;">▶ uv-logs</span>
 </span> | 
-Cell: benchmark | 7.50s
+Cell: benchmark | 3.86s
  | <button class="run-btn" onclick="runCell('benchmark')">▶ run</button>
 <button class="copy-btn" onclick="copyCell('benchmark')">Copy</button>
 <a href="cells/benchmark.py" target="_blank" class="raw-btn">Raw</a>
@@ -4007,27 +4234,27 @@ PROFILE TRACE: torch_eager | cuda_B1_S128_H8_D64_R32
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
                                                    Name    Self CPU %      Self CPU   CPU total %     CPU total  CPU time avg     Self CUDA   Self CUDA %    CUDA total  CUDA time avg    # of Calls  
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
-                                            torch_eager         0.00%       0.000us         0.00%       0.000us       0.000us       1.033ms      1157.58%       1.033ms       1.033ms             1  
-                                            torch_eager        14.26%     386.998us        99.70%       2.705ms       2.705ms       0.000us         0.00%      90.431us      90.431us             1  
-                                              aten::mul         6.08%     164.867us        10.45%     283.577us      11.816us      46.976us        52.65%      46.976us       1.957us            24  
-void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      46.976us        52.65%      46.976us       1.957us            24  
-                                            aten::copy_         3.96%     107.533us        62.14%       1.686ms      93.665us      28.959us        32.46%      30.175us       1.676us            18  
-void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      22.303us        25.00%      22.303us       1.859us            12  
-void at::native::vectorized_elementwise_kernel&lt;4, at...         0.00%       0.000us         0.00%       0.000us       0.000us      13.280us        14.89%      13.280us       1.107us            12  
-                                            aten::clone         1.58%      42.971us        61.19%       1.660ms     276.703us       0.000us         0.00%       7.872us       1.312us             6  
-                         Memcpy DtoD (Device -&gt; Device)         0.00%       0.000us         0.00%       0.000us       0.000us       6.656us         7.46%       6.656us       1.109us             6  
-                                              aten::sub         1.73%      46.871us         2.69%      72.911us      12.152us       6.656us         7.46%       6.656us       1.109us             6  
-                                              aten::add         1.35%      36.531us         2.16%      58.672us       9.779us       6.624us         7.42%       6.624us       1.104us             6  
-                                Activity Buffer Request        53.14%       1.442ms        53.14%       1.442ms       1.442ms       1.216us         1.36%       1.216us       1.216us             1  
-                                    aten::empty_strided         2.28%      61.772us         2.28%      61.772us      10.295us       0.000us         0.00%       0.000us       0.000us             6  
-                                        cudaMemcpyAsync         2.73%      74.144us         2.73%      74.144us      12.357us       0.000us         0.00%       0.000us       0.000us             6  
-                                            aten::slice         3.20%      86.920us         4.13%     112.081us       4.670us       0.000us         0.00%       0.000us       0.000us            24  
-                                       aten::as_strided         0.93%      25.161us         0.93%      25.161us       1.048us       0.000us         0.00%       0.000us       0.000us            24  
-                                       cudaLaunchKernel         8.45%     229.371us         8.45%     229.371us       4.779us       0.000us         0.00%       0.000us       0.000us            48  
-                                  cudaDeviceSynchronize         0.30%       8.270us         0.30%       8.270us       8.270us       0.000us         0.00%       0.000us       0.000us             1  
+                                            torch_eager         0.00%       0.000us         0.00%       0.000us       0.000us       1.038ms      1165.07%       1.038ms       1.038ms             1  
+                                            torch_eager        14.25%     384.344us        99.73%       2.691ms       2.691ms       0.000us         0.00%      90.272us      90.272us             1  
+                                              aten::mul         6.11%     164.889us        10.39%     280.433us      11.685us      46.752us        52.50%      46.752us       1.948us            24  
+void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      46.752us        52.50%      46.752us       1.948us            24  
+                                            aten::copy_         4.15%     111.919us        62.66%       1.690ms      93.917us      29.025us        32.59%      30.240us       1.680us            18  
+void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      22.306us        25.05%      22.306us       1.859us            12  
+void at::native::vectorized_elementwise_kernel&lt;4, at...         0.00%       0.000us         0.00%       0.000us       0.000us      13.280us        14.91%      13.280us       1.107us            12  
+                                            aten::clone         1.43%      38.559us        61.06%       1.647ms     274.577us       0.000us         0.00%       7.934us       1.322us             6  
+                         Memcpy DtoD (Device -&gt; Device)         0.00%       0.000us         0.00%       0.000us       0.000us       6.719us         7.54%       6.719us       1.120us             6  
+                                              aten::sub         1.59%      42.770us         2.55%      68.721us      11.454us       6.688us         7.51%       6.688us       1.115us             6  
+                                              aten::add         1.63%      44.070us         2.49%      67.170us      11.195us       6.592us         7.40%       6.592us       1.099us             6  
+                                Activity Buffer Request        53.52%       1.444ms        53.52%       1.444ms       1.444ms       1.215us         1.36%       1.215us       1.215us             1  
+                                    aten::empty_strided         2.14%      57.723us         2.14%      57.723us       9.620us       0.000us         0.00%       0.000us       0.000us             6  
+                                        cudaMemcpyAsync         2.62%      70.572us         2.62%      70.572us      11.762us       0.000us         0.00%       0.000us       0.000us             6  
+                                            aten::slice         2.99%      80.691us         3.82%     103.161us       4.298us       0.000us         0.00%       0.000us       0.000us            24  
+                                       aten::as_strided         0.83%      22.470us         0.83%      22.470us       0.936us       0.000us         0.00%       0.000us       0.000us            24  
+                                       cudaLaunchKernel         8.47%     228.526us         8.47%     228.526us       4.761us       0.000us         0.00%       0.000us       0.000us            48  
+                                  cudaDeviceSynchronize         0.27%       7.361us         0.27%       7.361us       7.361us       0.000us         0.00%       0.000us       0.000us             1  
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
-Self CPU time total: 2.713ms
-Self CUDA time total: 89.215us
+Self CPU time total: 2.698ms
+Self CUDA time total: 89.057us
 
 
 
@@ -4037,27 +4264,27 @@ PROFILE TRACE: torch_eager | cuda_B1_S128_H8_D128_R64
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
                                                    Name    Self CPU %      Self CPU   CPU total %     CPU total  CPU time avg     Self CUDA   Self CUDA %    CUDA total  CUDA time avg    # of Calls  
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
-                                            torch_eager         0.00%       0.000us         0.00%       0.000us       0.000us     920.346us      1019.88%     920.346us     920.346us             1  
-                                            torch_eager        11.67%     287.669us        99.75%       2.459ms       2.459ms       0.000us         0.00%      91.392us      91.392us             1  
-                                              aten::mul         5.97%     147.150us        10.47%     258.131us      10.755us      47.681us        52.84%      47.681us       1.987us            24  
-void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      47.681us        52.84%      47.681us       1.987us            24  
-                                            aten::copy_         4.01%      98.743us        66.94%       1.650ms      91.665us      29.184us        32.34%      30.335us       1.685us            18  
-void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      22.433us        24.86%      22.433us       1.869us            12  
-void at::native::vectorized_elementwise_kernel&lt;4, at...         0.00%       0.000us         0.00%       0.000us       0.000us      13.376us        14.82%      13.376us       1.115us            12  
-                                            aten::clone         0.96%      23.772us        64.13%       1.581ms     263.446us       0.000us         0.00%       7.902us       1.317us             6  
-                         Memcpy DtoD (Device -&gt; Device)         0.00%       0.000us         0.00%       0.000us       0.000us       6.751us         7.48%       6.751us       1.125us             6  
-                                              aten::sub         1.51%      37.314us         2.51%      61.954us      10.326us       6.720us         7.45%       6.720us       1.120us             6  
-                                              aten::add         1.33%      32.821us         2.21%      54.451us       9.075us       6.656us         7.38%       6.656us       1.109us             6  
-                                Activity Buffer Request        58.20%       1.434ms        58.20%       1.434ms       1.434ms       1.151us         1.28%       1.151us       1.151us             1  
-                                    aten::empty_strided         1.33%      32.830us         1.33%      32.830us       5.472us       0.000us         0.00%       0.000us       0.000us             6  
-                                        cudaMemcpyAsync         2.21%      54.420us         2.21%      54.420us       9.070us       0.000us         0.00%       0.000us       0.000us             6  
-                                            aten::slice         2.84%      69.900us         3.65%      89.853us       3.744us       0.000us         0.00%       0.000us       0.000us            24  
-                                       aten::as_strided         0.81%      19.953us         0.81%      19.953us       0.831us       0.000us         0.00%       0.000us       0.000us            24  
-                                       cudaLaunchKernel         8.92%     219.731us         8.92%     219.731us       4.578us       0.000us         0.00%       0.000us       0.000us            48  
-                                  cudaDeviceSynchronize         0.25%       6.050us         0.25%       6.050us       6.050us       0.000us         0.00%       0.000us       0.000us             1  
+                                            torch_eager         0.00%       0.000us         0.00%       0.000us       0.000us     960.319us      1064.55%     960.319us     960.319us             1  
+                                            torch_eager        12.91%     327.841us        99.79%       2.533ms       2.533ms       0.000us         0.00%      91.361us      91.361us             1  
+                                              aten::mul         6.09%     154.573us        10.36%     263.046us      10.960us      47.616us        52.78%      47.616us       1.984us            24  
+void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      47.616us        52.78%      47.616us       1.984us            24  
+                                            aten::copy_         4.38%     111.264us        65.83%       1.671ms      92.839us      29.313us        32.49%      30.465us       1.692us            18  
+void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      22.496us        24.94%      22.496us       1.875us            12  
+void at::native::vectorized_elementwise_kernel&lt;4, at...         0.00%       0.000us         0.00%       0.000us       0.000us      13.280us        14.72%      13.280us       1.107us            12  
+                                            aten::clone         1.07%      27.110us        62.73%       1.592ms     265.408us       0.000us         0.00%       7.969us       1.328us             6  
+                         Memcpy DtoD (Device -&gt; Device)         0.00%       0.000us         0.00%       0.000us       0.000us       6.817us         7.56%       6.817us       1.136us             6  
+                                              aten::sub         1.66%      42.072us         2.63%      66.652us      11.109us       6.688us         7.41%       6.688us       1.115us             6  
+                                              aten::add         1.28%      32.560us         2.18%      55.291us       9.215us       6.592us         7.31%       6.592us       1.099us             6  
+                                Activity Buffer Request        56.87%       1.444ms        56.87%       1.444ms       1.444ms       1.152us         1.28%       1.152us       1.152us             1  
+                                    aten::empty_strided         1.25%      31.671us         1.25%      31.671us       5.278us       0.000us         0.00%       0.000us       0.000us             6  
+                                        cudaMemcpyAsync         2.12%      53.780us         2.12%      53.780us       8.963us       0.000us         0.00%       0.000us       0.000us             6  
+                                            aten::slice         2.76%      70.023us         3.57%      90.653us       3.777us       0.000us         0.00%       0.000us       0.000us            24  
+                                       aten::as_strided         0.81%      20.630us         0.81%      20.630us       0.860us       0.000us         0.00%       0.000us       0.000us            24  
+                                       cudaLaunchKernel         8.59%     218.025us         8.59%     218.025us       4.542us       0.000us         0.00%       0.000us       0.000us            48  
+                                  cudaDeviceSynchronize         0.21%       5.289us         0.21%       5.289us       5.289us       0.000us         0.00%       0.000us       0.000us             1  
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
-Self CPU time total: 2.465ms
-Self CUDA time total: 90.241us
+Self CPU time total: 2.539ms
+Self CUDA time total: 90.209us
 
 
 
@@ -4067,27 +4294,27 @@ PROFILE TRACE: torch_eager | cuda_B1_S128_H32_D64_R32
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
                                                    Name    Self CPU %      Self CPU   CPU total %     CPU total  CPU time avg     Self CUDA   Self CUDA %    CUDA total  CUDA time avg    # of Calls  
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
-                                            torch_eager         0.00%       0.000us         0.00%       0.000us       0.000us     909.568us       966.47%     909.568us     909.568us             1  
-                                            torch_eager        11.23%     276.876us        99.79%       2.460ms       2.460ms       0.000us         0.00%      95.424us      95.424us             1  
-                                              aten::mul         6.27%     154.461us        10.66%     262.794us      10.950us      48.800us        51.85%      48.800us       2.033us            24  
-void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      48.800us        51.85%      48.800us       2.033us            24  
-                                            aten::copy_         4.02%      99.094us        67.67%       1.668ms      92.677us      30.912us        32.85%      32.224us       1.790us            18  
-void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      23.008us        24.45%      23.008us       1.917us            12  
-void at::native::vectorized_elementwise_kernel&lt;4, at...         0.00%       0.000us         0.00%       0.000us       0.000us      14.400us        15.30%      14.400us       1.200us            12  
-                                            aten::clone         0.93%      22.950us        64.64%       1.593ms     265.583us       0.000us         0.00%       9.216us       1.536us             6  
-                         Memcpy DtoD (Device -&gt; Device)         0.00%       0.000us         0.00%       0.000us       0.000us       7.904us         8.40%       7.904us       1.317us             6  
-                                              aten::sub         1.56%      38.564us         2.52%      62.034us      10.339us       7.200us         7.65%       7.200us       1.200us             6  
-                                              aten::add         1.24%      30.660us         2.12%      52.250us       8.708us       7.200us         7.65%       7.200us       1.200us             6  
-                                Activity Buffer Request        58.87%       1.451ms        58.87%       1.451ms       1.451ms       1.312us         1.39%       1.312us       1.312us             1  
-                                    aten::empty_strided         1.24%      30.531us         1.24%      30.531us       5.089us       0.000us         0.00%       0.000us       0.000us             6  
-                                        cudaMemcpyAsync         2.20%      54.240us         2.20%      54.240us       9.040us       0.000us         0.00%       0.000us       0.000us             6  
-                                            aten::slice         2.65%      65.401us         3.42%      84.323us       3.513us       0.000us         0.00%       0.000us       0.000us            24  
-                                       aten::as_strided         0.77%      18.922us         0.77%      18.922us       0.788us       0.000us         0.00%       0.000us       0.000us            24  
-                                       cudaLaunchKernel         8.80%     216.993us         8.80%     216.993us       4.521us       0.000us         0.00%       0.000us       0.000us            48  
-                                  cudaDeviceSynchronize         0.21%       5.190us         0.21%       5.190us       5.190us       0.000us         0.00%       0.000us       0.000us             1  
+                                            torch_eager         0.00%       0.000us         0.00%       0.000us       0.000us     950.812us      1013.41%     950.812us     950.812us             1  
+                                            torch_eager        12.58%     319.124us        99.78%       2.531ms       2.531ms       0.000us         0.00%      95.135us      95.135us             1  
+                                              aten::mul         6.09%     154.550us        10.34%     262.291us      10.929us      48.671us        51.88%      48.671us       2.028us            24  
+void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      48.671us        51.88%      48.671us       2.028us            24  
+                                            aten::copy_         4.10%     104.029us        66.32%       1.682ms      93.470us      30.783us        32.81%      32.095us       1.783us            18  
+void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      22.943us        24.45%      22.943us       1.912us            12  
+void at::native::vectorized_elementwise_kernel&lt;4, at...         0.00%       0.000us         0.00%       0.000us       0.000us      14.369us        15.32%      14.369us       1.197us            12  
+                                            aten::clone         1.04%      26.300us        63.34%       1.607ms     267.803us       0.000us         0.00%       9.152us       1.525us             6  
+                         Memcpy DtoD (Device -&gt; Device)         0.00%       0.000us         0.00%       0.000us       0.000us       7.840us         8.36%       7.840us       1.307us             6  
+                                              aten::sub         1.64%      41.492us         2.64%      66.953us      11.159us       7.199us         7.67%       7.199us       1.200us             6  
+                                              aten::add         1.26%      31.999us         2.14%      54.310us       9.052us       7.170us         7.64%       7.170us       1.195us             6  
+                                Activity Buffer Request        57.64%       1.462ms        57.64%       1.462ms       1.462ms       1.312us         1.40%       1.312us       1.312us             1  
+                                    aten::empty_strided         1.26%      31.840us         1.26%      31.840us       5.307us       0.000us         0.00%       0.000us       0.000us             6  
+                                        cudaMemcpyAsync         2.05%      52.102us         2.05%      52.102us       8.684us       0.000us         0.00%       0.000us       0.000us             6  
+                                            aten::slice         2.68%      67.986us         3.47%      87.958us       3.665us       0.000us         0.00%       0.000us       0.000us            24  
+                                       aten::as_strided         0.79%      19.972us         0.79%      19.972us       0.832us       0.000us         0.00%       0.000us       0.000us            24  
+                                       cudaLaunchKernel         8.65%     219.475us         8.65%     219.475us       4.572us       0.000us         0.00%       0.000us       0.000us            48  
+                                  cudaDeviceSynchronize         0.22%       5.651us         0.22%       5.651us       5.651us       0.000us         0.00%       0.000us       0.000us             1  
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
-Self CPU time total: 2.465ms
-Self CUDA time total: 94.112us
+Self CPU time total: 2.537ms
+Self CUDA time total: 93.823us
 
 
 
@@ -4097,27 +4324,27 @@ PROFILE TRACE: torch_eager | cuda_B1_S128_H32_D128_R64
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
                                                    Name    Self CPU %      Self CPU   CPU total %     CPU total  CPU time avg     Self CUDA   Self CUDA %    CUDA total  CUDA time avg    # of Calls  
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
-                                            torch_eager         0.00%       0.000us         0.00%       0.000us       0.000us     892.572us       880.74%     892.572us     892.572us             1  
-                                            torch_eager        11.35%     283.366us        99.78%       2.492ms       2.492ms       0.000us         0.00%     102.687us     102.687us             1  
-                                              aten::mul         5.93%     148.202us        10.19%     254.513us      10.605us      52.956us        52.25%      52.956us       2.207us            24  
-void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      52.956us        52.25%      52.956us       2.207us            24  
-                                            aten::copy_         3.94%      98.395us        68.27%       1.705ms      94.725us      32.482us        32.05%      33.826us       1.879us            18  
-void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      24.641us        24.31%      24.641us       2.053us            12  
-void at::native::vectorized_elementwise_kernel&lt;4, at...         0.00%       0.000us         0.00%       0.000us       0.000us      15.905us        15.69%      15.905us       1.325us            12  
-                                            aten::clone         0.86%      21.380us        65.50%       1.636ms     272.651us       0.000us         0.00%       9.185us       1.531us             6  
-                                              aten::add         1.24%      31.000us         2.12%      53.041us       8.840us       8.032us         7.93%       8.032us       1.339us             6  
-                                              aten::sub         1.40%      35.052us         2.32%      58.022us       9.670us       7.873us         7.77%       7.873us       1.312us             6  
-                         Memcpy DtoD (Device -&gt; Device)         0.00%       0.000us         0.00%       0.000us       0.000us       7.841us         7.74%       7.841us       1.307us             6  
-                                Activity Buffer Request        52.43%       1.309ms        52.43%       1.309ms       1.309ms       1.344us         1.33%       1.344us       1.344us             1  
-                                    aten::empty_strided         1.32%      33.071us         1.32%      33.071us       5.512us       0.000us         0.00%       0.000us       0.000us             6  
-                                        cudaMemcpyAsync         9.52%     237.764us         9.52%     237.764us      39.627us       0.000us         0.00%       0.000us       0.000us             6  
-                                            aten::slice         2.60%      64.825us         3.35%      83.624us       3.484us       0.000us         0.00%       0.000us       0.000us            24  
-                                       aten::as_strided         0.75%      18.799us         0.75%      18.799us       0.783us       0.000us         0.00%       0.000us       0.000us            24  
-                                       cudaLaunchKernel         8.44%     210.793us         8.44%     210.793us       4.392us       0.000us         0.00%       0.000us       0.000us            48  
-                                  cudaDeviceSynchronize         0.22%       5.611us         0.22%       5.611us       5.611us       0.000us         0.00%       0.000us       0.000us             1  
+                                            torch_eager         0.00%       0.000us         0.00%       0.000us       0.000us     952.670us       942.15%     952.670us     952.670us             1  
+                                            torch_eager        11.55%     312.506us        99.79%       2.701ms       2.701ms       0.000us         0.00%     102.429us     102.429us             1  
+                                              aten::mul         5.68%     153.743us         9.71%     262.695us      10.946us      52.765us        52.18%      52.765us       2.199us            24  
+void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      52.765us        52.18%      52.765us       2.199us            24  
+                                            aten::copy_         3.97%     107.471us        68.61%       1.857ms     103.165us      32.353us        32.00%      33.665us       1.870us            18  
+void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      24.641us        24.37%      24.641us       2.053us            12  
+void at::native::vectorized_elementwise_kernel&lt;4, at...         0.00%       0.000us         0.00%       0.000us       0.000us      15.999us        15.82%      15.999us       1.333us            12  
+                                            aten::clone         1.01%      27.330us        65.76%       1.780ms     296.625us       0.000us         0.00%       9.024us       1.504us             6  
+                                              aten::add         1.21%      32.850us         2.05%      55.600us       9.267us       8.032us         7.94%       8.032us       1.339us             6  
+                                              aten::sub         1.44%      39.082us         2.35%      63.492us      10.582us       7.967us         7.88%       7.967us       1.328us             6  
+                         Memcpy DtoD (Device -&gt; Device)         0.00%       0.000us         0.00%       0.000us       0.000us       7.712us         7.63%       7.712us       1.285us             6  
+                                Activity Buffer Request        52.99%       1.434ms        52.99%       1.434ms       1.434ms       1.312us         1.30%       1.312us       1.312us             1  
+                                    aten::empty_strided         1.20%      32.420us         1.20%      32.420us       5.403us       0.000us         0.00%       0.000us       0.000us             6  
+                                        cudaMemcpyAsync         9.27%     250.924us         9.27%     250.924us      41.821us       0.000us         0.00%       0.000us       0.000us             6  
+                                            aten::slice         2.56%      69.212us         3.32%      89.782us       3.741us       0.000us         0.00%       0.000us       0.000us            24  
+                                       aten::as_strided         0.76%      20.570us         0.76%      20.570us       0.857us       0.000us         0.00%       0.000us       0.000us            24  
+                                       cudaLaunchKernel         8.14%     220.374us         8.14%     220.374us       4.591us       0.000us         0.00%       0.000us       0.000us            48  
+                                  cudaDeviceSynchronize         0.21%       5.791us         0.21%       5.791us       5.791us       0.000us         0.00%       0.000us       0.000us             1  
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
-Self CPU time total: 2.498ms
-Self CUDA time total: 101.343us
+Self CPU time total: 2.707ms
+Self CUDA time total: 101.117us
 
 
 
@@ -4127,27 +4354,27 @@ PROFILE TRACE: torch_eager | cuda_B1_S512_H8_D64_R32
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
                                                    Name    Self CPU %      Self CPU   CPU total %     CPU total  CPU time avg     Self CUDA   Self CUDA %    CUDA total  CUDA time avg    # of Calls  
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
-                                            torch_eager         0.00%       0.000us         0.00%       0.000us       0.000us     907.478us       966.25%     907.478us     907.478us             1  
-                                            torch_eager        11.02%     305.318us        99.81%       2.765ms       2.765ms       0.000us         0.00%      95.230us      95.230us             1  
-                                              aten::mul         5.24%     145.172us         9.20%     254.787us      10.616us      49.023us        52.20%      49.023us       2.043us            24  
-void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      49.023us        52.20%      49.023us       2.043us            24  
-                                            aten::copy_         3.74%     103.536us        70.23%       1.945ms     108.067us      30.719us        32.71%      32.031us       1.779us            18  
-void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      22.912us        24.40%      22.912us       1.909us            12  
-void at::native::vectorized_elementwise_kernel&lt;4, at...         0.00%       0.000us         0.00%       0.000us       0.000us      14.176us        15.09%      14.176us       1.181us            12  
-                                            aten::clone         1.09%      30.110us        67.87%       1.880ms     313.329us       0.000us         0.00%       9.119us       1.520us             6  
-                         Memcpy DtoD (Device -&gt; Device)         0.00%       0.000us         0.00%       0.000us       0.000us       7.807us         8.31%       7.807us       1.301us             6  
-                                              aten::sub         1.24%      34.480us         2.10%      58.270us       9.712us       7.104us         7.56%       7.104us       1.184us             6  
-                                              aten::add         1.09%      30.091us         1.87%      51.880us       8.647us       7.072us         7.53%       7.072us       1.179us             6  
-                                Activity Buffer Request        52.12%       1.444ms        52.12%       1.444ms       1.444ms       1.312us         1.40%       1.312us       1.312us             1  
-                                    aten::empty_strided         1.13%      31.430us         1.13%      31.430us       5.238us       0.000us         0.00%       0.000us       0.000us             6  
-                                        cudaMemcpyAsync        12.15%     336.439us        12.15%     336.439us      56.073us       0.000us         0.00%       0.000us       0.000us             6  
-                                            aten::slice         2.48%      68.768us         3.17%      87.719us       3.655us       0.000us         0.00%       0.000us       0.000us            24  
-                                       aten::as_strided         0.68%      18.951us         0.68%      18.951us       0.790us       0.000us         0.00%       0.000us       0.000us            24  
-                                       cudaLaunchKernel         7.82%     216.674us         7.82%     216.674us       4.514us       0.000us         0.00%       0.000us       0.000us            48  
-                                  cudaDeviceSynchronize         0.19%       5.210us         0.19%       5.210us       5.210us       0.000us         0.00%       0.000us       0.000us             1  
+                                            torch_eager         0.00%       0.000us         0.00%       0.000us       0.000us     987.399us      1051.70%     987.399us     987.399us             1  
+                                            torch_eager        12.37%     335.778us        99.82%       2.710ms       2.710ms       0.000us         0.00%      95.198us      95.198us             1  
+                                              aten::mul         5.74%     155.881us         9.81%     266.305us      11.096us      48.927us        52.11%      48.927us       2.039us            24  
+void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      48.927us        52.11%      48.927us       2.039us            24  
+                                            aten::copy_         3.95%     107.229us        67.43%       1.830ms     101.693us      30.753us        32.76%      32.065us       1.781us            18  
+void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      22.881us        24.37%      22.881us       1.907us            12  
+void at::native::vectorized_elementwise_kernel&lt;4, at...         0.00%       0.000us         0.00%       0.000us       0.000us      14.206us        15.13%      14.206us       1.184us            12  
+                                            aten::clone         0.99%      26.953us        64.69%       1.756ms     292.683us       0.000us         0.00%       9.184us       1.531us             6  
+                         Memcpy DtoD (Device -&gt; Device)         0.00%       0.000us         0.00%       0.000us       0.000us       7.872us         8.38%       7.872us       1.312us             6  
+                                              aten::add         1.25%      33.910us         2.11%      57.361us       9.560us       7.103us         7.57%       7.103us       1.184us             6  
+                                              aten::sub         1.62%      44.010us         2.55%      69.231us      11.538us       7.103us         7.57%       7.103us       1.184us             6  
+                                Activity Buffer Request        53.49%       1.452ms        53.49%       1.452ms       1.452ms       1.312us         1.40%       1.312us       1.312us             1  
+                                    aten::empty_strided         1.24%      33.730us         1.24%      33.730us       5.622us       0.000us         0.00%       0.000us       0.000us             6  
+                                        cudaMemcpyAsync         7.66%     207.874us         7.66%     207.874us      34.646us       0.000us         0.00%       0.000us       0.000us             6  
+                                            aten::slice         2.54%      68.958us         3.31%      89.820us       3.743us       0.000us         0.00%       0.000us       0.000us            24  
+                                       aten::as_strided         0.77%      20.862us         0.77%      20.862us       0.869us       0.000us         0.00%       0.000us       0.000us            24  
+                                       cudaLaunchKernel         8.19%     222.327us         8.19%     222.327us       4.632us       0.000us         0.00%       0.000us       0.000us            48  
+                                  cudaDeviceSynchronize         0.18%       5.000us         0.18%       5.000us       5.000us       0.000us         0.00%       0.000us       0.000us             1  
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
-Self CPU time total: 2.770ms
-Self CUDA time total: 93.918us
+Self CPU time total: 2.715ms
+Self CUDA time total: 93.886us
 
 
 
@@ -4157,27 +4384,27 @@ PROFILE TRACE: torch_eager | cuda_B1_S512_H8_D128_R64
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
                                                    Name    Self CPU %      Self CPU   CPU total %     CPU total  CPU time avg     Self CUDA   Self CUDA %    CUDA total  CUDA time avg    # of Calls  
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
-                                            torch_eager         0.00%       0.000us         0.00%       0.000us       0.000us     917.786us       906.49%     917.786us     917.786us             1  
-                                            torch_eager        10.59%     290.695us        99.81%       2.741ms       2.741ms       0.000us         0.00%     102.558us     102.558us             1  
-                                              aten::mul         5.39%     148.136us         9.30%     255.477us      10.645us      52.735us        52.09%      52.735us       2.197us            24  
-void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      52.735us        52.09%      52.735us       2.197us            24  
-                                            aten::copy_         4.15%     114.085us        70.69%       1.941ms     107.839us      32.512us        32.11%      33.824us       1.879us            18  
-void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      24.704us        24.40%      24.704us       2.059us            12  
-void at::native::vectorized_elementwise_kernel&lt;4, at...         0.00%       0.000us         0.00%       0.000us       0.000us      15.999us        15.80%      15.999us       1.333us            12  
-                                            aten::clone         0.78%      21.500us        67.65%       1.858ms     309.627us       0.000us         0.00%       9.120us       1.520us             6  
-                                              aten::sub         1.39%      38.270us         2.26%      62.070us      10.345us       8.063us         7.96%       8.063us       1.344us             6  
-                                              aten::add         1.13%      31.111us         1.93%      52.881us       8.813us       7.936us         7.84%       7.936us       1.323us             6  
-                         Memcpy DtoD (Device -&gt; Device)         0.00%       0.000us         0.00%       0.000us       0.000us       7.808us         7.71%       7.808us       1.301us             6  
-                                Activity Buffer Request        52.71%       1.447ms        52.71%       1.447ms       1.447ms       1.312us         1.30%       1.312us       1.312us             1  
-                                    aten::empty_strided         1.19%      32.762us         1.19%      32.762us       5.460us       0.000us         0.00%       0.000us       0.000us             6  
-                                        cudaMemcpyAsync        11.56%     317.516us        11.56%     317.516us      52.919us       0.000us         0.00%       0.000us       0.000us             6  
-                                            aten::slice         2.38%      65.270us         3.07%      84.260us       3.511us       0.000us         0.00%       0.000us       0.000us            24  
-                                       aten::as_strided         0.69%      18.990us         0.69%      18.990us       0.791us       0.000us         0.00%       0.000us       0.000us            24  
-                                       cudaLaunchKernel         7.83%     214.935us         7.83%     214.935us       4.478us       0.000us         0.00%       0.000us       0.000us            48  
-                                  cudaDeviceSynchronize         0.19%       5.200us         0.19%       5.200us       5.200us       0.000us         0.00%       0.000us       0.000us             1  
+                                            torch_eager         0.00%       0.000us         0.00%       0.000us       0.000us     939.771us       930.81%     939.771us     939.771us             1  
+                                            torch_eager        11.42%     294.218us        99.78%       2.570ms       2.570ms       0.000us         0.00%     102.276us     102.276us             1  
+                                              aten::mul         5.85%     150.653us        10.08%     259.594us      10.816us      52.609us        52.11%      52.609us       2.192us            24  
+void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      52.609us        52.11%      52.609us       2.192us            24  
+                                            aten::copy_         4.01%     103.273us        68.02%       1.752ms      97.337us      32.450us        32.14%      33.763us       1.876us            18  
+void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      24.640us        24.40%      24.640us       2.053us            12  
+void at::native::vectorized_elementwise_kernel&lt;4, at...         0.00%       0.000us         0.00%       0.000us       0.000us      15.904us        15.75%      15.904us       1.325us            12  
+                                            aten::clone         0.87%      22.360us        64.99%       1.674ms     278.983us       0.000us         0.00%       9.123us       1.520us             6  
+                                              aten::sub         1.58%      40.669us         2.53%      65.240us      10.873us       7.968us         7.89%       7.968us       1.328us             6  
+                                              aten::add         1.32%      33.930us         2.20%      56.580us       9.430us       7.936us         7.86%       7.936us       1.323us             6  
+                         Memcpy DtoD (Device -&gt; Device)         0.00%       0.000us         0.00%       0.000us       0.000us       7.810us         7.74%       7.810us       1.302us             6  
+                                Activity Buffer Request        54.28%       1.398ms        54.28%       1.398ms       1.398ms       1.313us         1.30%       1.313us       1.313us             1  
+                                    aten::empty_strided         1.21%      31.291us         1.21%      31.291us       5.215us       0.000us         0.00%       0.000us       0.000us             6  
+                                        cudaMemcpyAsync         7.34%     188.943us         7.34%     188.943us      31.491us       0.000us         0.00%       0.000us       0.000us             6  
+                                            aten::slice         2.69%      69.330us         3.44%      88.671us       3.695us       0.000us         0.00%       0.000us       0.000us            24  
+                                       aten::as_strided         0.75%      19.341us         0.75%      19.341us       0.806us       0.000us         0.00%       0.000us       0.000us            24  
+                                       cudaLaunchKernel         8.46%     218.003us         8.46%     218.003us       4.542us       0.000us         0.00%       0.000us       0.000us            48  
+                                  cudaDeviceSynchronize         0.22%       5.651us         0.22%       5.651us       5.651us       0.000us         0.00%       0.000us       0.000us             1  
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
-Self CPU time total: 2.746ms
-Self CUDA time total: 101.246us
+Self CPU time total: 2.576ms
+Self CUDA time total: 100.963us
 
 
 
@@ -4187,27 +4414,27 @@ PROFILE TRACE: torch_eager | cuda_B1_S512_H32_D64_R32
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
                                                    Name    Self CPU %      Self CPU   CPU total %     CPU total  CPU time avg     Self CUDA   Self CUDA %    CUDA total  CUDA time avg    # of Calls  
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
-                                            torch_eager         0.00%       0.000us         0.00%       0.000us       0.000us     896.601us       744.17%     896.601us     896.601us             1  
-                                            torch_eager        10.66%     286.835us        99.81%       2.687ms       2.687ms       0.000us         0.00%     122.275us     122.275us             1  
-                                              aten::mul         5.47%     147.118us         9.41%     253.291us      10.554us      61.985us        51.45%      61.985us       2.583us            24  
-void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      61.985us        51.45%      61.985us       2.583us            24  
-                                            aten::copy_         3.72%     100.260us        70.38%       1.894ms     105.246us      39.265us        32.59%      41.057us       2.281us            18  
-void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      28.834us        23.93%      28.834us       2.403us            12  
-void at::native::vectorized_elementwise_kernel&lt;4, at...         0.00%       0.000us         0.00%       0.000us       0.000us      19.233us        15.96%      19.233us       1.603us            12  
-                                            aten::clone         0.83%      22.211us        67.89%       1.827ms     304.542us       0.000us         0.00%      12.223us       2.037us             6  
-                         Memcpy DtoD (Device -&gt; Device)         0.00%       0.000us         0.00%       0.000us       0.000us      10.431us         8.66%      10.431us       1.738us             6  
-                                              aten::add         1.14%      30.799us         1.94%      52.140us       8.690us       9.632us         7.99%       9.632us       1.605us             6  
-                                              aten::sub         1.37%      36.770us         2.23%      59.970us       9.995us       9.601us         7.97%       9.601us       1.600us             6  
-                                Activity Buffer Request        53.18%       1.431ms        53.18%       1.431ms       1.431ms       1.792us         1.49%       1.792us       1.792us             1  
-                                    aten::empty_strided         1.21%      32.491us         1.21%      32.491us       5.415us       0.000us         0.00%       0.000us       0.000us             6  
-                                        cudaMemcpyAsync        11.26%     303.147us        11.26%     303.147us      50.525us       0.000us         0.00%       0.000us       0.000us             6  
-                                            aten::slice         2.49%      66.932us         3.17%      85.280us       3.553us       0.000us         0.00%       0.000us       0.000us            24  
-                                       aten::as_strided         0.68%      18.348us         0.68%      18.348us       0.765us       0.000us         0.00%       0.000us       0.000us            24  
-                                       cudaLaunchKernel         7.81%     210.347us         7.81%     210.347us       4.382us       0.000us         0.00%       0.000us       0.000us            48  
-                                  cudaDeviceSynchronize         0.19%       5.020us         0.19%       5.020us       5.020us       0.000us         0.00%       0.000us       0.000us             1  
+                                            torch_eager         0.00%       0.000us         0.00%       0.000us       0.000us     987.019us       820.52%     987.019us     987.019us             1  
+                                            torch_eager        11.12%     293.915us        99.79%       2.637ms       2.637ms       0.000us         0.00%     122.116us     122.116us             1  
+                                              aten::mul         6.22%     164.251us        10.48%     276.937us      11.539us      61.922us        51.48%      61.922us       2.580us            24  
+void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      61.922us        51.48%      61.922us       2.580us            24  
+                                            aten::copy_         3.96%     104.584us        67.08%       1.772ms      98.461us      39.265us        32.64%      41.089us       2.283us            18  
+void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      28.833us        23.97%      28.833us       2.403us            12  
+void at::native::vectorized_elementwise_kernel&lt;4, at...         0.00%       0.000us         0.00%       0.000us       0.000us      19.105us        15.88%      19.105us       1.592us            12  
+                                            aten::clone         0.81%      21.321us        64.15%       1.695ms     282.483us       0.000us         0.00%      12.256us       2.043us             6  
+                         Memcpy DtoD (Device -&gt; Device)         0.00%       0.000us         0.00%       0.000us       0.000us      10.432us         8.67%      10.432us       1.739us             6  
+                                              aten::sub         1.58%      41.691us         2.56%      67.622us      11.270us       9.569us         7.95%       9.569us       1.595us             6  
+                                              aten::add         1.31%      34.540us         2.17%      57.381us       9.563us       9.536us         7.93%       9.536us       1.589us             6  
+                                Activity Buffer Request        53.87%       1.423ms        53.87%       1.423ms       1.423ms       1.824us         1.52%       1.824us       1.824us             1  
+                                    aten::empty_strided         1.17%      30.940us         1.17%      30.940us       5.157us       0.000us         0.00%       0.000us       0.000us             6  
+                                        cudaMemcpyAsync         6.97%     184.193us         6.97%     184.193us      30.699us       0.000us         0.00%       0.000us       0.000us             6  
+                                            aten::slice         3.59%      94.920us         4.40%     116.150us       4.840us       0.000us         0.00%       0.000us       0.000us            24  
+                                       aten::as_strided         0.80%      21.230us         0.80%      21.230us       0.885us       0.000us         0.00%       0.000us       0.000us            24  
+                                       cudaLaunchKernel         8.38%     221.517us         8.38%     221.517us       4.615us       0.000us         0.00%       0.000us       0.000us            48  
+                                  cudaDeviceSynchronize         0.21%       5.631us         0.21%       5.631us       5.631us       0.000us         0.00%       0.000us       0.000us             1  
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
-Self CPU time total: 2.692ms
-Self CUDA time total: 120.483us
+Self CPU time total: 2.642ms
+Self CUDA time total: 120.292us
 
 
 
@@ -4217,27 +4444,27 @@ PROFILE TRACE: torch_eager | cuda_B1_S512_H32_D128_R64
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
                                                    Name    Self CPU %      Self CPU   CPU total %     CPU total  CPU time avg     Self CUDA   Self CUDA %    CUDA total  CUDA time avg    # of Calls  
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
-                                            torch_eager         0.00%       0.000us         0.00%       0.000us       0.000us     885.202us       514.56%     885.202us     885.202us             1  
-                                            torch_eager        18.81%     279.303us        99.64%       1.480ms       1.480ms       0.000us         0.00%     174.944us     174.944us             1  
-                                              aten::mul         9.70%     144.115us        16.98%     252.116us      10.505us      89.439us        51.99%      89.439us       3.727us            24  
-void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      89.439us        51.99%      89.439us       3.727us            24  
-                                            aten::copy_         6.85%     101.723us        47.28%     702.206us      39.011us      57.632us        33.50%      60.544us       3.364us            18  
-void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      40.608us        23.60%      40.608us       3.384us            12  
-void at::native::vectorized_elementwise_kernel&lt;4, at...         0.00%       0.000us         0.00%       0.000us       0.000us      24.961us        14.51%      24.961us       2.080us            12  
-                                            aten::clone         1.41%      20.892us        42.46%     630.635us     105.106us       0.000us         0.00%      19.936us       3.323us             6  
-                         Memcpy DtoD (Device -&gt; Device)         0.00%       0.000us         0.00%       0.000us       0.000us      17.024us         9.90%      17.024us       2.837us             6  
-                                              aten::add         2.07%      30.702us         3.51%      52.142us       8.690us      12.545us         7.29%      12.545us       2.091us             6  
-                                              aten::sub         2.41%      35.732us         4.00%      59.442us       9.907us      12.416us         7.22%      12.416us       2.069us             6  
-                                Activity Buffer Request        17.15%     254.675us        17.15%     254.675us     254.675us       2.912us         1.69%       2.912us       2.912us             1  
-                                    aten::empty_strided         2.07%      30.780us         2.07%      30.780us       5.130us       0.000us         0.00%       0.000us       0.000us             6  
-                                        cudaMemcpyAsync        19.36%     287.456us        19.36%     287.456us      47.909us       0.000us         0.00%       0.000us       0.000us             6  
-                                            aten::slice         4.32%      64.164us         5.58%      82.803us       3.450us       0.000us         0.00%       0.000us       0.000us            24  
-                                       aten::as_strided         1.26%      18.639us         1.26%      18.639us       0.777us       0.000us         0.00%       0.000us       0.000us            24  
-                                       cudaLaunchKernel        14.24%     211.503us        14.24%     211.503us       4.406us       0.000us         0.00%       0.000us       0.000us            48  
-                                  cudaDeviceSynchronize         0.36%       5.410us         0.36%       5.410us       5.410us       0.000us         0.00%       0.000us       0.000us             1  
+                                            torch_eager         0.00%       0.000us         0.00%       0.000us       0.000us     942.977us       547.62%     942.977us     942.977us             1  
+                                            torch_eager        11.98%     313.186us        99.77%       2.608ms       2.608ms       0.000us         0.00%     175.043us     175.043us             1  
+                                              aten::mul         5.92%     154.664us        10.07%     263.135us      10.964us      89.731us        52.11%      89.731us       3.739us            24  
+void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      89.731us        52.11%      89.731us       3.739us            24  
+                                            aten::copy_         4.21%     110.022us        67.75%       1.771ms      98.397us      57.632us        33.47%      60.480us       3.360us            18  
+void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      40.640us        23.60%      40.640us       3.387us            12  
+void at::native::vectorized_elementwise_kernel&lt;4, at...         0.00%       0.000us         0.00%       0.000us       0.000us      24.832us        14.42%      24.832us       2.069us            12  
+                                            aten::clone         1.00%      26.050us        64.65%       1.690ms     281.685us       0.000us         0.00%      19.840us       3.307us             6  
+                         Memcpy DtoD (Device -&gt; Device)         0.00%       0.000us         0.00%       0.000us       0.000us      16.992us         9.87%      16.992us       2.832us             6  
+                                              aten::add         1.22%      32.012us         2.08%      54.302us       9.050us      12.416us         7.21%      12.416us       2.069us             6  
+                                              aten::sub         1.48%      38.721us         2.41%      62.881us      10.480us      12.416us         7.21%      12.416us       2.069us             6  
+                                Activity Buffer Request        54.20%       1.417ms        54.20%       1.417ms       1.417ms       2.848us         1.65%       2.848us       2.848us             1  
+                                    aten::empty_strided         1.15%      30.180us         1.15%      30.180us       5.030us       0.000us         0.00%       0.000us       0.000us             6  
+                                        cudaMemcpyAsync         6.98%     182.574us         6.98%     182.574us      30.429us       0.000us         0.00%       0.000us       0.000us             6  
+                                            aten::slice         2.56%      66.979us         3.34%      87.351us       3.640us       0.000us         0.00%       0.000us       0.000us            24  
+                                       aten::as_strided         0.78%      20.372us         0.78%      20.372us       0.849us       0.000us         0.00%       0.000us       0.000us            24  
+                                       cudaLaunchKernel         8.28%     216.491us         8.28%     216.491us       4.510us       0.000us         0.00%       0.000us       0.000us            48  
+                                  cudaDeviceSynchronize         0.23%       5.900us         0.23%       5.900us       5.900us       0.000us         0.00%       0.000us       0.000us             1  
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
-Self CPU time total: 1.485ms
-Self CUDA time total: 172.032us
+Self CPU time total: 2.614ms
+Self CUDA time total: 172.195us
 
 
 
@@ -4247,27 +4474,27 @@ PROFILE TRACE: torch_eager | cuda_B1_S2048_H8_D64_R32
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
                                                    Name    Self CPU %      Self CPU   CPU total %     CPU total  CPU time avg     Self CUDA   Self CUDA %    CUDA total  CUDA time avg    # of Calls  
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
-                                            torch_eager         0.00%       0.000us         0.00%       0.000us       0.000us     907.735us       751.64%     907.735us     907.735us             1  
-                                            torch_eager        18.35%     272.536us        99.65%       1.480ms       1.480ms       0.000us         0.00%     122.527us     122.527us             1  
-                                              aten::mul         9.89%     146.883us        17.48%     259.553us      10.815us      62.078us        51.40%      62.078us       2.587us            24  
-void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      62.078us        51.40%      62.078us       2.587us            24  
-                                            aten::copy_         6.65%      98.730us        45.99%     682.885us      37.938us      39.328us        32.57%      41.088us       2.283us            18  
-void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      28.832us        23.87%      28.832us       2.403us            12  
-void at::native::vectorized_elementwise_kernel&lt;4, at...         0.00%       0.000us         0.00%       0.000us       0.000us      19.361us        16.03%      19.361us       1.613us            12  
-                                            aten::clone         2.58%      38.249us        42.54%     631.763us     105.294us       0.000us         0.00%      12.256us       2.043us             6  
-                         Memcpy DtoD (Device -&gt; Device)         0.00%       0.000us         0.00%       0.000us       0.000us      10.496us         8.69%      10.496us       1.749us             6  
-                                              aten::add         2.13%      31.663us         3.60%      53.483us       8.914us       9.728us         8.06%       9.728us       1.621us             6  
-                                              aten::sub         2.35%      34.954us         3.91%      58.043us       9.674us       9.633us         7.98%       9.633us       1.605us             6  
-                                Activity Buffer Request        16.88%     250.706us        16.88%     250.706us     250.706us       1.760us         1.46%       1.760us       1.760us             1  
-                                    aten::empty_strided         2.15%      31.912us         2.15%      31.912us       5.319us       0.000us         0.00%       0.000us       0.000us             6  
-                                        cudaMemcpyAsync        18.48%     274.437us        18.48%     274.437us      45.739us       0.000us         0.00%       0.000us       0.000us             6  
-                                            aten::slice         4.31%      63.964us         5.59%      83.053us       3.461us       0.000us         0.00%       0.000us       0.000us            24  
-                                       aten::as_strided         1.29%      19.089us         1.29%      19.089us       0.795us       0.000us         0.00%       0.000us       0.000us            24  
-                                       cudaLaunchKernel        14.59%     216.591us        14.59%     216.591us       4.512us       0.000us         0.00%       0.000us       0.000us            48  
-                                  cudaDeviceSynchronize         0.35%       5.220us         0.35%       5.220us       5.220us       0.000us         0.00%       0.000us       0.000us             1  
+                                            torch_eager         0.00%       0.000us         0.00%       0.000us       0.000us     954.334us       791.88%     954.334us     954.334us             1  
+                                            torch_eager        21.12%     286.823us        99.60%       1.352ms       1.352ms       0.000us         0.00%     122.339us     122.339us             1  
+                                              aten::mul        11.39%     154.733us        19.43%     263.854us      10.994us      61.889us        51.35%      61.889us       2.579us            24  
+void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      61.889us        51.35%      61.889us       2.579us            24  
+                                            aten::copy_         8.06%     109.392us        38.94%     528.759us      29.376us      39.393us        32.69%      41.217us       2.290us            18  
+void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      28.864us        23.95%      28.864us       2.405us            12  
+void at::native::vectorized_elementwise_kernel&lt;4, at...         0.00%       0.000us         0.00%       0.000us       0.000us      19.233us        15.96%      19.233us       1.603us            12  
+                                            aten::clone         1.54%      20.901us        32.67%     443.638us      73.940us       0.000us         0.00%      12.353us       2.059us             6  
+                         Memcpy DtoD (Device -&gt; Device)         0.00%       0.000us         0.00%       0.000us       0.000us      10.529us         8.74%      10.529us       1.755us             6  
+                                              aten::sub         2.93%      39.731us         4.81%      65.293us      10.882us       9.633us         7.99%       9.633us       1.606us             6  
+                                              aten::add         2.54%      34.552us         4.77%      64.792us      10.799us       9.600us         7.97%       9.600us       1.600us             6  
+                                Activity Buffer Request        12.72%     172.763us        12.72%     172.763us     172.763us       1.824us         1.51%       1.824us       1.824us             1  
+                                    aten::empty_strided         2.32%      31.561us         2.32%      31.561us       5.260us       0.000us         0.00%       0.000us       0.000us             6  
+                                        cudaMemcpyAsync        13.45%     182.623us        13.45%     182.623us      30.437us       0.000us         0.00%       0.000us       0.000us             6  
+                                            aten::slice         5.17%      70.140us         6.66%      90.481us       3.770us       0.000us         0.00%       0.000us       0.000us            24  
+                                       aten::as_strided         1.50%      20.341us         1.50%      20.341us       0.848us       0.000us         0.00%       0.000us       0.000us            24  
+                                       cudaLaunchKernel        16.86%     228.904us        16.86%     228.904us       4.769us       0.000us         0.00%       0.000us       0.000us            48  
+                                  cudaDeviceSynchronize         0.40%       5.490us         0.40%       5.490us       5.490us       0.000us         0.00%       0.000us       0.000us             1  
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
-Self CPU time total: 1.485ms
-Self CUDA time total: 120.767us
+Self CPU time total: 1.358ms
+Self CUDA time total: 120.515us
 
 
 
@@ -4277,27 +4504,27 @@ PROFILE TRACE: torch_eager | cuda_B1_S2048_H8_D128_R64
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
                                                    Name    Self CPU %      Self CPU   CPU total %     CPU total  CPU time avg     Self CUDA   Self CUDA %    CUDA total  CUDA time avg    # of Calls  
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
-                                            torch_eager         0.00%       0.000us         0.00%       0.000us       0.000us     894.399us       519.81%     894.399us     894.399us             1  
-                                            torch_eager        10.51%     278.801us        99.79%       2.648ms       2.648ms       0.000us         0.00%     174.911us     174.911us             1  
-                                              aten::mul         5.47%     145.104us         9.49%     251.734us      10.489us      89.535us        52.04%      89.535us       3.731us            24  
-void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      89.535us        52.04%      89.535us       3.731us            24  
-                                            aten::copy_         3.73%      98.901us        70.34%       1.866ms     103.682us      57.696us        33.53%      60.544us       3.364us            18  
-void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      40.704us        23.66%      40.704us       3.392us            12  
-void at::native::vectorized_elementwise_kernel&lt;4, at...         0.00%       0.000us         0.00%       0.000us       0.000us      24.832us        14.43%      24.832us       2.069us            12  
-                                            aten::clone         0.84%      22.190us        67.69%       1.796ms     299.337us       0.000us         0.00%      19.840us       3.307us             6  
+                                            torch_eager         0.00%       0.000us         0.00%       0.000us       0.000us     961.439us       559.06%     961.439us     961.439us             1  
+                                            torch_eager        21.39%     301.083us        99.65%       1.403ms       1.403ms       0.000us         0.00%     174.821us     174.821us             1  
+                                              aten::mul        10.92%     153.723us        18.79%     264.437us      11.018us      89.541us        52.07%      89.541us       3.731us            24  
+void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      89.541us        52.07%      89.541us       3.731us            24  
+                                            aten::copy_         8.57%     120.662us        41.11%     578.630us      32.146us      57.631us        33.51%      60.479us       3.360us            18  
+void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      40.639us        23.63%      40.639us       3.387us            12  
+void at::native::vectorized_elementwise_kernel&lt;4, at...         0.00%       0.000us         0.00%       0.000us       0.000us      24.801us        14.42%      24.801us       2.067us            12  
+                                            aten::clone         1.49%      21.022us        33.99%     478.490us      79.748us       0.000us         0.00%      19.840us       3.307us             6  
                          Memcpy DtoD (Device -&gt; Device)         0.00%       0.000us         0.00%       0.000us       0.000us      16.992us         9.88%      16.992us       2.832us             6  
-                                              aten::sub         1.44%      38.162us         2.33%      61.942us      10.324us      12.448us         7.23%      12.448us       2.075us             6  
-                                              aten::add         1.15%      30.549us         1.97%      52.171us       8.695us      12.384us         7.20%      12.384us       2.064us             6  
-                                Activity Buffer Request        54.02%       1.433ms        54.02%       1.433ms       1.433ms       2.848us         1.66%       2.848us       2.848us             1  
-                                    aten::empty_strided         1.13%      30.052us         1.13%      30.052us       5.009us       0.000us         0.00%       0.000us       0.000us             6  
-                                        cudaMemcpyAsync        10.37%     275.065us        10.37%     275.065us      45.844us       0.000us         0.00%       0.000us       0.000us             6  
-                                            aten::slice         2.49%      65.991us         3.19%      84.601us       3.525us       0.000us         0.00%       0.000us       0.000us            24  
-                                       aten::as_strided         0.70%      18.610us         0.70%      18.610us       0.775us       0.000us         0.00%       0.000us       0.000us            24  
-                                       cudaLaunchKernel         7.95%     211.023us         7.95%     211.023us       4.396us       0.000us         0.00%       0.000us       0.000us            48  
-                                  cudaDeviceSynchronize         0.21%       5.640us         0.21%       5.640us       5.640us       0.000us         0.00%       0.000us       0.000us             1  
+                                              aten::add         2.26%      31.841us         3.85%      54.131us       9.022us      12.481us         7.26%      12.481us       2.080us             6  
+                                              aten::sub         2.79%      39.260us         4.52%      63.691us      10.615us      12.320us         7.16%      12.320us       2.053us             6  
+                                Activity Buffer Request        15.02%     211.404us        15.02%     211.404us     211.404us       2.848us         1.66%       2.848us       2.848us             1  
+                                    aten::empty_strided         2.10%      29.500us         2.10%      29.500us       4.917us       0.000us         0.00%       0.000us       0.000us             6  
+                                        cudaMemcpyAsync        13.01%     183.184us        13.01%     183.184us      30.531us       0.000us         0.00%       0.000us       0.000us             6  
+                                            aten::slice         4.96%      69.812us         6.41%      90.211us       3.759us       0.000us         0.00%       0.000us       0.000us            24  
+                                       aten::as_strided         1.45%      20.399us         1.45%      20.399us       0.850us       0.000us         0.00%       0.000us       0.000us            24  
+                                       cudaLaunchKernel        15.69%     220.815us        15.69%     220.815us       4.600us       0.000us         0.00%       0.000us       0.000us            48  
+                                  cudaDeviceSynchronize         0.35%       4.890us         0.35%       4.890us       4.890us       0.000us         0.00%       0.000us       0.000us             1  
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
-Self CPU time total: 2.653ms
-Self CUDA time total: 172.063us
+Self CPU time total: 1.408ms
+Self CUDA time total: 171.973us
 
 
 
@@ -4307,27 +4534,27 @@ PROFILE TRACE: torch_eager | cuda_B1_S2048_H32_D64_R32
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
                                                    Name    Self CPU %      Self CPU   CPU total %     CPU total  CPU time avg     Self CUDA   Self CUDA %    CUDA total  CUDA time avg    # of Calls  
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
-                                            torch_eager         0.00%       0.000us         0.00%       0.000us       0.000us     888.595us       313.84%     888.595us     888.595us             1  
-                                            torch_eager        18.64%     271.692us        99.64%       1.452ms       1.452ms       0.000us         0.00%     301.536us     301.536us             1  
-                                              aten::mul         9.98%     145.418us        17.29%     252.060us      10.503us     132.896us        46.94%     132.896us       5.537us            24  
-void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us     132.896us        46.94%     132.896us       5.537us            24  
-                                            aten::copy_         6.89%     100.362us        46.38%     676.084us      37.560us     109.376us        38.63%     127.776us       7.099us            18  
-                                            aten::clone         1.48%      21.511us        41.22%     600.853us     100.142us       0.000us         0.00%      70.560us      11.760us             6  
-void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      57.216us        20.21%      57.216us       4.768us            12  
-                         Memcpy DtoD (Device -&gt; Device)         0.00%       0.000us         0.00%       0.000us       0.000us      52.160us        18.42%      52.160us       8.693us             6  
-void at::native::vectorized_elementwise_kernel&lt;4, at...         0.00%       0.000us         0.00%       0.000us       0.000us      40.864us        14.43%      40.864us       3.405us            12  
-                                              aten::sub         2.41%      35.143us         4.02%      58.572us       9.762us      20.512us         7.24%      20.512us       3.419us             6  
-                                              aten::add         2.12%      30.932us         3.62%      52.783us       8.797us      20.352us         7.19%      20.352us       3.392us             6  
-                                Activity Buffer Request        16.97%     247.406us        16.97%     247.406us     247.406us      18.400us         6.50%      18.400us      18.400us             1  
-                                    aten::empty_strided         2.15%      31.370us         2.15%      31.370us       5.228us       0.000us         0.00%       0.000us       0.000us             6  
-                                        cudaMemcpyAsync        18.35%     267.496us        18.35%     267.496us      44.583us       0.000us         0.00%       0.000us       0.000us             6  
-                                            aten::slice         4.85%      70.742us         6.06%      88.302us       3.679us       0.000us         0.00%       0.000us       0.000us            24  
-                                       aten::as_strided         1.20%      17.560us         1.20%      17.560us       0.732us       0.000us         0.00%       0.000us       0.000us            24  
-                                       cudaLaunchKernel        14.59%     212.742us        14.59%     212.742us       4.432us       0.000us         0.00%       0.000us       0.000us            48  
-                                  cudaDeviceSynchronize         0.36%       5.280us         0.36%       5.280us       5.280us       0.000us         0.00%       0.000us       0.000us             1  
+                                            torch_eager         0.00%       0.000us         0.00%       0.000us       0.000us     959.740us       338.81%     959.740us     959.740us             1  
+                                            torch_eager        11.78%     309.495us        99.81%       2.622ms       2.622ms       0.000us         0.00%     301.248us     301.248us             1  
+                                              aten::mul         5.80%     152.430us         9.98%     262.294us      10.929us     133.378us        47.09%     133.378us       5.557us            24  
+void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us     133.378us        47.09%     133.378us       5.557us            24  
+                                            aten::copy_         4.09%     107.511us        67.37%       1.770ms      98.338us     108.832us        38.42%     126.816us       7.045us            18  
+                                            aten::clone         1.07%      28.041us        64.54%       1.696ms     282.603us       0.000us         0.00%      69.600us      11.600us             6  
+void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      57.216us        20.20%      57.216us       4.768us            12  
+                         Memcpy DtoD (Device -&gt; Device)         0.00%       0.000us         0.00%       0.000us       0.000us      51.616us        18.22%      51.616us       8.603us             6  
+void at::native::vectorized_elementwise_kernel&lt;4, at...         0.00%       0.000us         0.00%       0.000us       0.000us      41.054us        14.49%      41.054us       3.421us            12  
+                                              aten::sub         1.57%      41.190us         2.52%      66.080us      11.013us      20.607us         7.27%      20.607us       3.434us             6  
+                                              aten::add         1.56%      40.972us         2.46%      64.512us      10.752us      20.447us         7.22%      20.447us       3.408us             6  
+                                Activity Buffer Request        53.79%       1.413ms        53.79%       1.413ms       1.413ms      17.984us         6.35%      17.984us      17.984us             1  
+                                    aten::empty_strided         1.19%      31.311us         1.19%      31.311us       5.218us       0.000us         0.00%       0.000us       0.000us             6  
+                                        cudaMemcpyAsync         7.14%     187.713us         7.14%     187.713us      31.285us       0.000us         0.00%       0.000us       0.000us             6  
+                                            aten::slice         2.66%      69.760us         3.44%      90.282us       3.762us       0.000us         0.00%       0.000us       0.000us            24  
+                                       aten::as_strided         0.78%      20.522us         0.78%      20.522us       0.855us       0.000us         0.00%       0.000us       0.000us            24  
+                                       cudaLaunchKernel         8.37%     219.936us         8.37%     219.936us       4.582us       0.000us         0.00%       0.000us       0.000us            48  
+                                  cudaDeviceSynchronize         0.19%       5.111us         0.19%       5.111us       5.111us       0.000us         0.00%       0.000us       0.000us             1  
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
-Self CPU time total: 1.458ms
-Self CUDA time total: 283.136us
+Self CPU time total: 2.627ms
+Self CUDA time total: 283.264us
 
 
 
@@ -4337,27 +4564,27 @@ PROFILE TRACE: torch_eager | cuda_B1_S2048_H32_D128_R64
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
                                                    Name    Self CPU %      Self CPU   CPU total %     CPU total  CPU time avg     Self CUDA   Self CUDA %    CUDA total  CUDA time avg    # of Calls  
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
-                                            torch_eager         0.00%       0.000us         0.00%       0.000us       0.000us     944.856us       167.33%     944.856us     944.856us             1  
-                                            torch_eager        19.10%     286.874us        99.66%       1.497ms       1.497ms       0.000us         0.00%     588.218us     588.218us             1  
-                                            aten::copy_         6.48%      97.352us        44.49%     668.224us      37.124us     273.885us        48.50%     297.437us      16.524us            18  
-                                              aten::mul        11.54%     173.280us        19.20%     288.361us      12.015us     224.990us        39.84%     224.990us       9.375us            24  
-void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us     224.990us        39.84%     224.990us       9.375us            24  
-                                            aten::clone         1.34%      20.121us        39.51%     593.393us      98.899us       0.000us         0.00%     206.910us      34.485us             6  
-                         Memcpy DtoD (Device -&gt; Device)         0.00%       0.000us         0.00%       0.000us       0.000us     183.358us        32.47%     183.358us      30.560us             6  
-void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      90.527us        16.03%      90.527us       7.544us            12  
-void at::native::vectorized_elementwise_kernel&lt;4, at...         0.00%       0.000us         0.00%       0.000us       0.000us      65.791us        11.65%      65.791us       5.483us            12  
-                                              aten::sub         2.45%      36.872us         4.07%      61.073us      10.179us      33.407us         5.92%      33.407us       5.568us             6  
-                                              aten::add         2.13%      32.018us         3.64%      54.631us       9.105us      32.384us         5.74%      32.384us       5.397us             6  
-                                Activity Buffer Request        16.63%     249.816us        16.63%     249.816us     249.816us      23.552us         4.17%      23.552us      23.552us             1  
-                                    aten::empty_strided         2.02%      30.350us         2.02%      30.350us       5.058us       0.000us         0.00%       0.000us       0.000us             6  
-                                        cudaMemcpyAsync        17.28%     259.545us        17.28%     259.545us      43.258us       0.000us         0.00%       0.000us       0.000us             6  
-                                            aten::slice         4.52%      67.913us         5.81%      87.211us       3.634us       0.000us         0.00%       0.000us       0.000us            24  
-                                       aten::as_strided         1.28%      19.298us         1.28%      19.298us       0.804us       0.000us         0.00%       0.000us       0.000us            24  
-                                       cudaLaunchKernel        14.87%     223.406us        14.87%     223.406us       4.654us       0.000us         0.00%       0.000us       0.000us            48  
-                                  cudaDeviceSynchronize         0.34%       5.141us         0.34%       5.141us       5.141us       0.000us         0.00%       0.000us       0.000us             1  
+                                            torch_eager         0.00%       0.000us         0.00%       0.000us       0.000us     964.299us       170.17%     964.299us     964.299us             1  
+                                            torch_eager        21.37%     289.253us        99.58%       1.348ms       1.348ms       0.000us         0.00%     590.419us     590.419us             1  
+                                            aten::copy_         7.69%     104.123us        37.93%     513.450us      28.525us     274.106us        48.37%     297.849us      16.547us            18  
+                                              aten::mul        11.75%     159.118us        20.07%     271.705us      11.321us     226.427us        39.96%     226.427us       9.434us            24  
+void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us     226.427us        39.96%     226.427us       9.434us            24  
+                                            aten::clone         1.55%      21.020us        32.53%     440.358us      73.393us       0.000us         0.00%     206.843us      34.474us             6  
+                         Memcpy DtoD (Device -&gt; Device)         0.00%       0.000us         0.00%       0.000us       0.000us     183.100us        32.31%     183.100us      30.517us             6  
+void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      91.006us        16.06%      91.006us       7.584us            12  
+void at::native::vectorized_elementwise_kernel&lt;4, at...         0.00%       0.000us         0.00%       0.000us       0.000us      66.143us        11.67%      66.143us       5.512us            12  
+                                              aten::sub         3.06%      41.432us         4.99%      67.562us      11.260us      33.664us         5.94%      33.664us       5.611us             6  
+                                              aten::add         2.43%      32.930us         4.17%      56.451us       9.408us      32.479us         5.73%      32.479us       5.413us             6  
+                                Activity Buffer Request        11.95%     161.793us        11.95%     161.793us     161.793us      23.743us         4.19%      23.743us      23.743us             1  
+                                    aten::empty_strided         2.85%      38.611us         2.85%      38.611us       6.435us       0.000us         0.00%       0.000us       0.000us             6  
+                                        cudaMemcpyAsync        13.59%     183.934us        13.59%     183.934us      30.656us       0.000us         0.00%       0.000us       0.000us             6  
+                                            aten::slice         5.13%      69.460us         6.64%      89.941us       3.748us       0.000us         0.00%       0.000us       0.000us            24  
+                                       aten::as_strided         1.51%      20.481us         1.51%      20.481us       0.853us       0.000us         0.00%       0.000us       0.000us            24  
+                                       cudaLaunchKernel        16.68%     225.838us        16.68%     225.838us       4.705us       0.000us         0.00%       0.000us       0.000us            48  
+                                  cudaDeviceSynchronize         0.42%       5.710us         0.42%       5.710us       5.710us       0.000us         0.00%       0.000us       0.000us             1  
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
-Self CPU time total: 1.502ms
-Self CUDA time total: 564.666us
+Self CPU time total: 1.354ms
+Self CUDA time total: 566.676us
 
 
 
@@ -4367,27 +4594,27 @@ PROFILE TRACE: torch_eager | cuda_B2_S128_H8_D64_R32
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
                                                    Name    Self CPU %      Self CPU   CPU total %     CPU total  CPU time avg     Self CUDA   Self CUDA %    CUDA total  CUDA time avg    # of Calls  
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
-                                            torch_eager         0.00%       0.000us         0.00%       0.000us       0.000us     916.448us       990.96%     916.448us     916.448us             1  
-                                            torch_eager        10.63%     281.892us        99.80%       2.647ms       2.647ms       0.000us         0.00%      93.601us      93.601us             1  
-                                              aten::mul         5.58%     148.028us         9.67%     256.571us      10.690us      49.634us        53.67%      49.634us       2.068us            24  
-void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      49.634us        53.67%      49.634us       2.068us            24  
-                                            aten::copy_         3.99%     105.971us        69.88%       1.854ms     102.991us      29.439us        31.83%      30.559us       1.698us            18  
-void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      22.655us        24.50%      22.655us       1.888us            12  
-void at::native::vectorized_elementwise_kernel&lt;4, at...         0.00%       0.000us         0.00%       0.000us       0.000us      13.408us        14.50%      13.408us       1.117us            12  
-                                            aten::clone         0.82%      21.802us        66.79%       1.772ms     295.325us       0.000us         0.00%       7.904us       1.317us             6  
-                         Memcpy DtoD (Device -&gt; Device)         0.00%       0.000us         0.00%       0.000us       0.000us       6.784us         7.34%       6.784us       1.131us             6  
-                                              aten::sub         1.36%      36.061us         2.24%      59.441us       9.907us       6.720us         7.27%       6.720us       1.120us             6  
-                                              aten::add         1.25%      33.260us         2.10%      55.590us       9.265us       6.688us         7.23%       6.688us       1.115us             6  
-                                Activity Buffer Request        54.00%       1.433ms        54.00%       1.433ms       1.433ms       1.120us         1.21%       1.120us       1.120us             1  
-                                    aten::empty_strided         1.13%      29.861us         1.13%      29.861us       4.977us       0.000us         0.00%       0.000us       0.000us             6  
-                                        cudaMemcpyAsync         9.52%     252.488us         9.52%     252.488us      42.081us       0.000us         0.00%       0.000us       0.000us             6  
-                                            aten::slice         2.59%      68.801us         3.33%      88.471us       3.686us       0.000us         0.00%       0.000us       0.000us            24  
-                                       aten::as_strided         0.74%      19.670us         0.74%      19.670us       0.820us       0.000us         0.00%       0.000us       0.000us            24  
-                                       cudaLaunchKernel         8.18%     216.965us         8.18%     216.965us       4.520us       0.000us         0.00%       0.000us       0.000us            48  
-                                  cudaDeviceSynchronize         0.20%       5.410us         0.20%       5.410us       5.410us       0.000us         0.00%       0.000us       0.000us             1  
+                                            torch_eager         0.00%       0.000us         0.00%       0.000us       0.000us     940.757us      1018.68%     940.757us     940.757us             1  
+                                            torch_eager        20.92%     284.932us        99.61%       1.357ms       1.357ms       0.000us         0.00%      93.503us      93.503us             1  
+                                              aten::mul        11.51%     156.743us        19.57%     266.566us      11.107us      49.664us        53.78%      49.664us       2.069us            24  
+void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      49.664us        53.78%      49.664us       2.069us            24  
+                                            aten::copy_         7.76%     105.742us        39.84%     542.619us      30.146us      29.343us        31.77%      30.495us       1.694us            18  
+void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      22.528us        24.39%      22.528us       1.877us            12  
+void at::native::vectorized_elementwise_kernel&lt;4, at...         0.00%       0.000us         0.00%       0.000us       0.000us      13.344us        14.45%      13.344us       1.112us            12  
+                                            aten::clone         1.52%      20.734us        33.85%     461.099us      76.850us       0.000us         0.00%       7.967us       1.328us             6  
+                         Memcpy DtoD (Device -&gt; Device)         0.00%       0.000us         0.00%       0.000us       0.000us       6.815us         7.38%       6.815us       1.136us             6  
+                                              aten::sub         2.96%      40.252us         4.79%      65.263us      10.877us       6.688us         7.24%       6.688us       1.115us             6  
+                                              aten::add         2.34%      31.811us         3.99%      54.311us       9.052us       6.656us         7.21%       6.656us       1.109us             6  
+                                Activity Buffer Request        14.09%     191.853us        14.09%     191.853us     191.853us       1.152us         1.25%       1.152us       1.152us             1  
+                                    aten::empty_strided         2.30%      31.379us         2.30%      31.379us       5.230us       0.000us         0.00%       0.000us       0.000us             6  
+                                        cudaMemcpyAsync        13.46%     183.403us        13.46%     183.403us      30.567us       0.000us         0.00%       0.000us       0.000us             6  
+                                            aten::slice         5.20%      70.859us         6.67%      90.910us       3.788us       0.000us         0.00%       0.000us       0.000us            24  
+                                       aten::as_strided         1.47%      20.051us         1.47%      20.051us       0.835us       0.000us         0.00%       0.000us       0.000us            24  
+                                       cudaLaunchKernel        16.08%     218.955us        16.08%     218.955us       4.562us       0.000us         0.00%       0.000us       0.000us            48  
+                                  cudaDeviceSynchronize         0.39%       5.360us         0.39%       5.360us       5.360us       0.000us         0.00%       0.000us       0.000us             1  
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
-Self CPU time total: 2.653ms
-Self CUDA time total: 92.481us
+Self CPU time total: 1.362ms
+Self CUDA time total: 92.351us
 
 
 
@@ -4397,27 +4624,27 @@ PROFILE TRACE: torch_eager | cuda_B2_S128_H8_D128_R64
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
                                                    Name    Self CPU %      Self CPU   CPU total %     CPU total  CPU time avg     Self CUDA   Self CUDA %    CUDA total  CUDA time avg    # of Calls  
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
-                                            torch_eager         0.00%       0.000us         0.00%       0.000us       0.000us     888.809us       923.35%     888.809us     888.809us             1  
-                                            torch_eager        19.05%     273.129us        99.67%       1.429ms       1.429ms       0.000us         0.00%      97.571us      97.571us             1  
-                                              aten::mul        10.09%     144.695us        17.61%     252.506us      10.521us      51.232us        53.22%      51.232us       2.135us            24  
-void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      51.232us        53.22%      51.232us       2.135us            24  
-                                            aten::copy_         6.72%      96.301us        45.37%     650.385us      36.132us      30.786us        31.98%      32.098us       1.783us            18  
-void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      22.944us        23.84%      22.944us       1.912us            12  
-void at::native::vectorized_elementwise_kernel&lt;4, at...         0.00%       0.000us         0.00%       0.000us       0.000us      14.241us        14.79%      14.241us       1.187us            12  
-                                            aten::clone         1.39%      19.911us        40.43%     579.513us      96.586us       0.000us         0.00%       9.154us       1.526us             6  
-                         Memcpy DtoD (Device -&gt; Device)         0.00%       0.000us         0.00%       0.000us       0.000us       7.842us         8.15%       7.842us       1.307us             6  
-                                              aten::add         2.26%      32.360us         3.79%      54.320us       9.053us       7.136us         7.41%       7.136us       1.189us             6  
-                                              aten::sub         2.55%      36.551us         4.17%      59.791us       9.965us       7.105us         7.38%       7.105us       1.184us             6  
-                                Activity Buffer Request        16.56%     237.415us        16.56%     237.415us     237.415us       1.312us         1.36%       1.312us       1.312us             1  
-                                    aten::empty_strided         2.18%      31.230us         2.18%      31.230us       5.205us       0.000us         0.00%       0.000us       0.000us             6  
-                                        cudaMemcpyAsync        17.96%     257.447us        17.96%     257.447us      42.908us       0.000us         0.00%       0.000us       0.000us             6  
-                                            aten::slice         4.71%      67.539us         6.11%      87.581us       3.649us       0.000us         0.00%       0.000us       0.000us            24  
-                                       aten::as_strided         1.40%      20.042us         1.40%      20.042us       0.835us       0.000us         0.00%       0.000us       0.000us            24  
-                                       cudaLaunchKernel        14.80%     212.233us        14.80%     212.233us       4.422us       0.000us         0.00%       0.000us       0.000us            48  
-                                  cudaDeviceSynchronize         0.33%       4.690us         0.33%       4.690us       4.690us       0.000us         0.00%       0.000us       0.000us             1  
+                                            torch_eager         0.00%       0.000us         0.00%       0.000us       0.000us     945.706us       986.10%     945.706us     945.706us             1  
+                                            torch_eager        12.18%     322.968us        99.79%       2.647ms       2.647ms       0.000us         0.00%      97.216us      97.216us             1  
+                                              aten::mul         5.85%     155.091us         9.99%     264.924us      11.039us      50.947us        53.12%      50.947us       2.123us            24  
+void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      50.947us        53.12%      50.947us       2.123us            24  
+                                            aten::copy_         3.92%     103.931us        67.30%       1.785ms      99.174us      30.783us        32.10%      32.095us       1.783us            18  
+void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      22.976us        23.96%      22.976us       1.915us            12  
+void at::native::vectorized_elementwise_kernel&lt;4, at...         0.00%       0.000us         0.00%       0.000us       0.000us      14.174us        14.78%      14.174us       1.181us            12  
+                                            aten::clone         1.18%      31.280us        64.70%       1.716ms     286.035us       0.000us         0.00%       9.119us       1.520us             6  
+                         Memcpy DtoD (Device -&gt; Device)         0.00%       0.000us         0.00%       0.000us       0.000us       7.807us         8.14%       7.807us       1.301us             6  
+                                              aten::add         1.22%      32.380us         2.09%      55.311us       9.219us       7.102us         7.41%       7.102us       1.184us             6  
+                                              aten::sub         1.50%      39.882us         2.41%      63.892us      10.649us       7.072us         7.37%       7.072us       1.179us             6  
+                                Activity Buffer Request        53.95%       1.431ms        53.95%       1.431ms       1.431ms       1.312us         1.37%       1.312us       1.312us             1  
+                                    aten::empty_strided         1.23%      32.600us         1.23%      32.600us       5.433us       0.000us         0.00%       0.000us       0.000us             6  
+                                        cudaMemcpyAsync         7.05%     187.002us         7.05%     187.002us      31.167us       0.000us         0.00%       0.000us       0.000us             6  
+                                            aten::slice         2.63%      69.642us         3.43%      90.901us       3.788us       0.000us         0.00%       0.000us       0.000us            24  
+                                       aten::as_strided         0.80%      21.259us         0.80%      21.259us       0.886us       0.000us         0.00%       0.000us       0.000us            24  
+                                       cudaLaunchKernel         8.29%     220.006us         8.29%     220.006us       4.583us       0.000us         0.00%       0.000us       0.000us            48  
+                                  cudaDeviceSynchronize         0.21%       5.569us         0.21%       5.569us       5.569us       0.000us         0.00%       0.000us       0.000us             1  
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
-Self CPU time total: 1.434ms
-Self CUDA time total: 96.259us
+Self CPU time total: 2.653ms
+Self CUDA time total: 95.904us
 
 
 
@@ -4427,27 +4654,27 @@ PROFILE TRACE: torch_eager | cuda_B2_S128_H32_D64_R32
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
                                                    Name    Self CPU %      Self CPU   CPU total %     CPU total  CPU time avg     Self CUDA   Self CUDA %    CUDA total  CUDA time avg    # of Calls  
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
-                                            torch_eager         0.00%       0.000us         0.00%       0.000us       0.000us     903.536us       870.95%     903.536us     903.536us             1  
-                                            torch_eager        18.87%     271.956us        99.65%       1.436ms       1.436ms       0.000us         0.00%     105.053us     105.053us             1  
-                                              aten::mul        10.20%     146.935us        17.83%     256.897us      10.704us      55.262us        53.27%      55.262us       2.303us            24  
-void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      55.262us        53.27%      55.262us       2.303us            24  
-                                            aten::copy_         6.83%      98.437us        45.05%     649.198us      36.067us      32.478us        31.31%      33.790us       1.877us            18  
-void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      24.670us        23.78%      24.670us       2.056us            12  
-void at::native::vectorized_elementwise_kernel&lt;4, at...         0.00%       0.000us         0.00%       0.000us       0.000us      16.001us        15.42%      16.001us       1.333us            12  
-                                            aten::clone         1.50%      21.580us        40.06%     577.333us      96.222us       0.000us         0.00%       9.120us       1.520us             6  
-                                              aten::sub         2.49%      35.841us         4.72%      67.992us      11.332us       8.001us         7.71%       8.001us       1.333us             6  
-                                              aten::add         2.31%      33.350us         3.86%      55.670us       9.278us       8.000us         7.71%       8.000us       1.333us             6  
-                         Memcpy DtoD (Device -&gt; Device)         0.00%       0.000us         0.00%       0.000us       0.000us       7.808us         7.53%       7.808us       1.301us             6  
-                                Activity Buffer Request        16.46%     237.265us        16.46%     237.265us     237.265us       1.312us         1.26%       1.312us       1.312us             1  
-                                    aten::empty_strided         2.16%      31.090us         2.16%      31.090us       5.182us       0.000us         0.00%       0.000us       0.000us             6  
-                                        cudaMemcpyAsync        17.50%     252.196us        17.50%     252.196us      42.033us       0.000us         0.00%       0.000us       0.000us             6  
-                                            aten::slice         4.40%      63.461us         5.67%      81.650us       3.402us       0.000us         0.00%       0.000us       0.000us            24  
-                                       aten::as_strided         1.26%      18.189us         1.26%      18.189us       0.758us       0.000us         0.00%       0.000us       0.000us            24  
-                                       cudaLaunchKernel        15.66%     225.733us        15.66%     225.733us       4.703us       0.000us         0.00%       0.000us       0.000us            48  
-                                  cudaDeviceSynchronize         0.35%       5.060us         0.35%       5.060us       5.060us       0.000us         0.00%       0.000us       0.000us             1  
+                                            torch_eager         0.00%       0.000us         0.00%       0.000us       0.000us     963.956us       929.78%     963.956us     963.956us             1  
+                                            torch_eager        11.95%     315.942us        99.78%       2.637ms       2.637ms       0.000us         0.00%     104.988us     104.988us             1  
+                                              aten::mul         6.01%     158.721us        10.21%     269.951us      11.248us      55.295us        53.33%      55.295us       2.304us            24  
+void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      55.295us        53.33%      55.295us       2.304us            24  
+                                            aten::copy_         4.03%     106.403us        67.45%       1.783ms      99.031us      32.417us        31.27%      33.729us       1.874us            18  
+void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      24.607us        23.73%      24.607us       2.051us            12  
+void at::native::vectorized_elementwise_kernel&lt;4, at...         0.00%       0.000us         0.00%       0.000us       0.000us      15.964us        15.40%      15.964us       1.330us            12  
+                                            aten::clone         1.02%      26.870us        64.62%       1.708ms     284.615us       0.000us         0.00%       9.122us       1.520us             6  
+                                              aten::add         1.23%      32.629us         2.10%      55.390us       9.232us       7.997us         7.71%       7.997us       1.333us             6  
+                                              aten::sub         1.44%      38.041us         2.36%      62.260us      10.377us       7.967us         7.68%       7.967us       1.328us             6  
+                         Memcpy DtoD (Device -&gt; Device)         0.00%       0.000us         0.00%       0.000us       0.000us       7.810us         7.53%       7.810us       1.302us             6  
+                                Activity Buffer Request        54.08%       1.429ms        54.08%       1.429ms       1.429ms       1.312us         1.27%       1.312us       1.312us             1  
+                                    aten::empty_strided         1.27%      33.640us         1.27%      33.640us       5.607us       0.000us         0.00%       0.000us       0.000us             6  
+                                        cudaMemcpyAsync         6.95%     183.544us         6.95%     183.544us      30.591us       0.000us         0.00%       0.000us       0.000us             6  
+                                            aten::slice         2.64%      69.789us         3.42%      90.471us       3.770us       0.000us         0.00%       0.000us       0.000us            24  
+                                       aten::as_strided         0.78%      20.682us         0.78%      20.682us       0.862us       0.000us         0.00%       0.000us       0.000us            24  
+                                       cudaLaunchKernel         8.39%     221.610us         8.39%     221.610us       4.617us       0.000us         0.00%       0.000us       0.000us            48  
+                                  cudaDeviceSynchronize         0.22%       5.700us         0.22%       5.700us       5.700us       0.000us         0.00%       0.000us       0.000us             1  
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
-Self CPU time total: 1.441ms
-Self CUDA time total: 103.741us
+Self CPU time total: 2.643ms
+Self CUDA time total: 103.676us
 
 
 
@@ -4457,27 +4684,27 @@ PROFILE TRACE: torch_eager | cuda_B2_S128_H32_D128_R64
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
                                                    Name    Self CPU %      Self CPU   CPU total %     CPU total  CPU time avg     Self CUDA   Self CUDA %    CUDA total  CUDA time avg    # of Calls  
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
-                                            torch_eager         0.00%       0.000us         0.00%       0.000us       0.000us     903.548us       729.80%     903.548us     903.548us             1  
-                                            torch_eager        10.56%     280.674us        99.81%       2.652ms       2.652ms       0.000us         0.00%     125.567us     125.567us             1  
-                                              aten::mul         5.49%     145.805us         9.46%     251.467us      10.478us      65.184us        52.65%      65.184us       2.716us            24  
-void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      65.184us        52.65%      65.184us       2.716us            24  
-                                            aten::copy_         3.75%      99.563us        70.08%       1.862ms     103.468us      39.422us        31.84%      41.182us       2.288us            18  
-void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      28.928us        23.37%      28.928us       2.411us            12  
-void at::native::vectorized_elementwise_kernel&lt;4, at...         0.00%       0.000us         0.00%       0.000us       0.000us      19.201us        15.51%      19.201us       1.600us            12  
-                                            aten::clone         0.92%      24.379us        67.48%       1.793ms     298.872us       0.000us         0.00%      12.254us       2.042us             6  
-                         Memcpy DtoD (Device -&gt; Device)         0.00%       0.000us         0.00%       0.000us       0.000us      10.494us         8.48%      10.494us       1.749us             6  
-                                              aten::add         1.15%      30.622us         1.96%      52.162us       8.694us       9.633us         7.78%       9.633us       1.606us             6  
-                                              aten::sub         1.45%      38.422us         2.36%      62.661us      10.443us       9.568us         7.73%       9.568us       1.595us             6  
-                                Activity Buffer Request        54.94%       1.460ms        54.94%       1.460ms       1.460ms       1.760us         1.42%       1.760us       1.760us             1  
-                                    aten::empty_strided         1.16%      30.801us         1.16%      30.801us       5.133us       0.000us         0.00%       0.000us       0.000us             6  
-                                        cudaMemcpyAsync         9.14%     242.866us         9.14%     242.866us      40.478us       0.000us         0.00%       0.000us       0.000us             6  
-                                            aten::slice         2.56%      67.990us         3.30%      87.783us       3.658us       0.000us         0.00%       0.000us       0.000us            24  
-                                       aten::as_strided         0.74%      19.793us         0.74%      19.793us       0.825us       0.000us         0.00%       0.000us       0.000us            24  
-                                       cudaLaunchKernel         7.96%     211.432us         7.96%     211.432us       4.405us       0.000us         0.00%       0.000us       0.000us            48  
-                                  cudaDeviceSynchronize         0.19%       5.160us         0.19%       5.160us       5.160us       0.000us         0.00%       0.000us       0.000us             1  
+                                            torch_eager         0.00%       0.000us         0.00%       0.000us       0.000us     933.942us       757.68%     933.942us     933.942us             1  
+                                            torch_eager        21.17%     287.829us        99.59%       1.354ms       1.354ms       0.000us         0.00%     125.024us     125.024us             1  
+                                              aten::mul        11.38%     154.770us        19.33%     262.774us      10.949us      64.862us        52.62%      64.862us       2.703us            24  
+void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      64.862us        52.62%      64.862us       2.703us            24  
+                                            aten::copy_         7.76%     105.560us        40.17%     546.058us      30.337us      39.265us        31.85%      41.025us       2.279us            18  
+void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      28.865us        23.42%      28.865us       2.405us            12  
+void at::native::vectorized_elementwise_kernel&lt;4, at...         0.00%       0.000us         0.00%       0.000us       0.000us      19.137us        15.53%      19.137us       1.595us            12  
+                                            aten::clone         1.51%      20.520us        34.08%     463.317us      77.220us       0.000us         0.00%      12.160us       2.027us             6  
+                         Memcpy DtoD (Device -&gt; Device)         0.00%       0.000us         0.00%       0.000us       0.000us      10.400us         8.44%      10.400us       1.733us             6  
+                                              aten::sub         2.90%      39.471us         4.67%      63.511us      10.585us       9.569us         7.76%       9.569us       1.595us             6  
+                                              aten::add         2.50%      34.030us         4.22%      57.431us       9.572us       9.568us         7.76%       9.568us       1.595us             6  
+                                Activity Buffer Request        14.30%     194.363us        14.30%     194.363us     194.363us       1.760us         1.43%       1.760us       1.760us             1  
+                                    aten::empty_strided         2.23%      30.321us         2.23%      30.321us       5.053us       0.000us         0.00%       0.000us       0.000us             6  
+                                        cudaMemcpyAsync        13.45%     182.914us        13.45%     182.914us      30.486us       0.000us         0.00%       0.000us       0.000us             6  
+                                            aten::slice         4.84%      65.748us         6.29%      85.480us       3.562us       0.000us         0.00%       0.000us       0.000us            24  
+                                       aten::as_strided         1.45%      19.732us         1.45%      19.732us       0.822us       0.000us         0.00%       0.000us       0.000us            24  
+                                       cudaLaunchKernel        16.08%     218.666us        16.08%     218.666us       4.556us       0.000us         0.00%       0.000us       0.000us            48  
+                                  cudaDeviceSynchronize         0.41%       5.560us         0.41%       5.560us       5.560us       0.000us         0.00%       0.000us       0.000us             1  
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
-Self CPU time total: 2.658ms
-Self CUDA time total: 123.807us
+Self CPU time total: 1.359ms
+Self CUDA time total: 123.264us
 
 
 
@@ -4487,27 +4714,27 @@ PROFILE TRACE: torch_eager | cuda_B2_S512_H8_D64_R32
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
                                                    Name    Self CPU %      Self CPU   CPU total %     CPU total  CPU time avg     Self CUDA   Self CUDA %    CUDA total  CUDA time avg    # of Calls  
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
-                                            torch_eager         0.00%       0.000us         0.00%       0.000us       0.000us     889.436us       855.74%     889.436us     889.436us             1  
-                                            torch_eager        19.42%     274.045us        99.59%       1.406ms       1.406ms       0.000us         0.00%     105.282us     105.282us             1  
-                                              aten::mul        10.41%     146.921us        18.18%     256.563us      10.690us      55.486us        53.38%      55.486us       2.312us            24  
-void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      55.486us        53.38%      55.486us       2.312us            24  
-                                            aten::copy_         6.82%      96.302us        44.56%     628.895us      34.939us      32.513us        31.28%      33.857us       1.881us            18  
-void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      24.705us        23.77%      24.705us       2.059us            12  
-void at::native::vectorized_elementwise_kernel&lt;4, at...         0.00%       0.000us         0.00%       0.000us       0.000us      15.939us        15.34%      15.939us       1.328us            12  
-                                            aten::clone         1.41%      19.928us        39.46%     556.871us      92.812us       0.000us         0.00%       9.152us       1.525us             6  
-                                              aten::sub         2.56%      36.082us         4.16%      58.744us       9.791us       7.970us         7.67%       7.970us       1.328us             6  
-                                              aten::add         2.23%      31.511us         3.85%      54.282us       9.047us       7.969us         7.67%       7.969us       1.328us             6  
-                         Memcpy DtoD (Device -&gt; Device)         0.00%       0.000us         0.00%       0.000us       0.000us       7.808us         7.51%       7.808us       1.301us             6  
-                                Activity Buffer Request        15.99%     225.676us        15.99%     225.676us     225.676us       1.344us         1.29%       1.344us       1.344us             1  
-                                    aten::empty_strided         2.17%      30.631us         2.17%      30.631us       5.105us       0.000us         0.00%       0.000us       0.000us             6  
-                                        cudaMemcpyAsync        17.52%     247.335us        17.52%     247.335us      41.223us       0.000us         0.00%       0.000us       0.000us             6  
-                                            aten::slice         4.52%      63.850us         5.84%      82.475us       3.436us       0.000us         0.00%       0.000us       0.000us            24  
-                                       aten::as_strided         1.32%      18.625us         1.32%      18.625us       0.776us       0.000us         0.00%       0.000us       0.000us            24  
-                                       cudaLaunchKernel        15.21%     214.657us        15.21%     214.657us       4.472us       0.000us         0.00%       0.000us       0.000us            48  
-                                  cudaDeviceSynchronize         0.41%       5.810us         0.41%       5.810us       5.810us       0.000us         0.00%       0.000us       0.000us             1  
+                                            torch_eager         0.00%       0.000us         0.00%       0.000us       0.000us     934.359us       900.66%     934.359us     934.359us             1  
+                                            torch_eager        21.17%     286.322us        99.59%       1.347ms       1.347ms       0.000us         0.00%     105.086us     105.086us             1  
+                                              aten::mul        11.62%     157.214us        19.66%     265.945us      11.081us      55.327us        53.33%      55.327us       2.305us            24  
+void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      55.327us        53.33%      55.327us       2.305us            24  
+                                            aten::copy_         7.65%     103.495us        39.66%     536.482us      29.805us      32.511us        31.34%      33.855us       1.881us            18  
+void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      24.704us        23.81%      24.704us       2.059us            12  
+void at::native::vectorized_elementwise_kernel&lt;4, at...         0.00%       0.000us         0.00%       0.000us       0.000us      15.904us        15.33%      15.904us       1.325us            12  
+                                            aten::clone         1.57%      21.280us        33.91%     458.650us      76.442us       0.000us         0.00%       9.151us       1.525us             6  
+                                              aten::add         2.43%      32.883us         4.09%      55.372us       9.229us       8.001us         7.71%       8.001us       1.333us             6  
+                                              aten::sub         2.87%      38.810us         4.64%      62.781us      10.463us       7.903us         7.62%       7.903us       1.317us             6  
+                         Memcpy DtoD (Device -&gt; Device)         0.00%       0.000us         0.00%       0.000us       0.000us       7.807us         7.53%       7.807us       1.301us             6  
+                                Activity Buffer Request        14.06%     190.184us        14.06%     190.184us     190.184us       1.344us         1.30%       1.344us       1.344us             1  
+                                    aten::empty_strided         2.22%      30.070us         2.22%      30.070us       5.012us       0.000us         0.00%       0.000us       0.000us             6  
+                                        cudaMemcpyAsync        13.39%     181.103us        13.39%     181.103us      30.184us       0.000us         0.00%       0.000us       0.000us             6  
+                                            aten::slice         5.05%      68.302us         6.56%      88.771us       3.699us       0.000us         0.00%       0.000us       0.000us            24  
+                                       aten::as_strided         1.51%      20.469us         1.51%      20.469us       0.853us       0.000us         0.00%       0.000us       0.000us            24  
+                                       cudaLaunchKernel        16.03%     216.891us        16.03%     216.891us       4.519us       0.000us         0.00%       0.000us       0.000us            48  
+                                  cudaDeviceSynchronize         0.41%       5.591us         0.41%       5.591us       5.591us       0.000us         0.00%       0.000us       0.000us             1  
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
-Self CPU time total: 1.411ms
-Self CUDA time total: 103.938us
+Self CPU time total: 1.353ms
+Self CUDA time total: 103.742us
 
 
 
@@ -4517,27 +4744,27 @@ PROFILE TRACE: torch_eager | cuda_B2_S512_H8_D128_R64
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
                                                    Name    Self CPU %      Self CPU   CPU total %     CPU total  CPU time avg     Self CUDA   Self CUDA %    CUDA total  CUDA time avg    # of Calls  
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
-                                            torch_eager         0.00%       0.000us         0.00%       0.000us       0.000us     888.135us       717.15%     888.135us     888.135us             1  
-                                            torch_eager        18.91%     268.465us        99.65%       1.415ms       1.415ms       0.000us         0.00%     125.666us     125.666us             1  
-                                              aten::mul        10.15%     144.114us        17.70%     251.265us      10.469us      65.346us        52.77%      65.346us       2.723us            24  
-void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      65.346us        52.77%      65.346us       2.723us            24  
-                                            aten::copy_         6.90%      97.992us        45.41%     644.725us      35.818us      39.328us        31.76%      41.152us       2.286us            18  
-void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      28.800us        23.26%      28.800us       2.400us            12  
-void at::native::vectorized_elementwise_kernel&lt;4, at...         0.00%       0.000us         0.00%       0.000us       0.000us      19.168us        15.48%      19.168us       1.597us            12  
-                                            aten::clone         1.46%      20.690us        40.33%     572.532us      95.422us       0.000us         0.00%      12.352us       2.059us             6  
-                         Memcpy DtoD (Device -&gt; Device)         0.00%       0.000us         0.00%       0.000us       0.000us      10.528us         8.50%      10.528us       1.755us             6  
-                                              aten::add         2.19%      31.029us         3.69%      52.390us       8.732us       9.600us         7.75%       9.600us       1.600us             6  
-                                              aten::sub         2.50%      35.469us         4.13%      58.580us       9.763us       9.568us         7.73%       9.568us       1.595us             6  
-                                Activity Buffer Request        15.69%     222.765us        15.69%     222.765us     222.765us       1.824us         1.47%       1.824us       1.824us             1  
-                                    aten::empty_strided         2.29%      32.500us         2.29%      32.500us       5.417us       0.000us         0.00%       0.000us       0.000us             6  
-                                        cudaMemcpyAsync        18.50%     262.716us        18.50%     262.716us      43.786us       0.000us         0.00%       0.000us       0.000us             6  
-                                            aten::slice         4.70%      66.710us         6.07%      86.108us       3.588us       0.000us         0.00%       0.000us       0.000us            24  
-                                       aten::as_strided         1.37%      19.398us         1.37%      19.398us       0.808us       0.000us         0.00%       0.000us       0.000us            24  
-                                       cudaLaunchKernel        14.99%     212.875us        14.99%     212.875us       4.435us       0.000us         0.00%       0.000us       0.000us            48  
-                                  cudaDeviceSynchronize         0.35%       5.010us         0.35%       5.010us       5.010us       0.000us         0.00%       0.000us       0.000us             1  
+                                            torch_eager         0.00%       0.000us         0.00%       0.000us       0.000us     944.694us       764.03%     944.694us     944.694us             1  
+                                            torch_eager        20.48%     287.824us        99.60%       1.400ms       1.400ms       0.000us         0.00%     125.438us     125.438us             1  
+                                              aten::mul        10.91%     153.363us        18.83%     264.625us      11.026us      65.151us        52.69%      65.151us       2.715us            24  
+void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      65.151us        52.69%      65.151us       2.715us            24  
+                                            aten::copy_         7.88%     110.793us        41.73%     586.532us      32.585us      39.328us        31.81%      41.120us       2.284us            18  
+void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      28.895us        23.37%      28.895us       2.408us            12  
+void at::native::vectorized_elementwise_kernel&lt;4, at...         0.00%       0.000us         0.00%       0.000us       0.000us      19.167us        15.50%      19.167us       1.597us            12  
+                                            aten::clone         1.52%      21.310us        35.87%     504.089us      84.015us       0.000us         0.00%      12.225us       2.038us             6  
+                         Memcpy DtoD (Device -&gt; Device)         0.00%       0.000us         0.00%       0.000us       0.000us      10.433us         8.44%      10.433us       1.739us             6  
+                                              aten::sub         2.80%      39.332us         4.57%      64.213us      10.702us       9.632us         7.79%       9.632us       1.605us             6  
+                                              aten::add         2.33%      32.799us         3.97%      55.790us       9.298us       9.535us         7.71%       9.535us       1.589us             6  
+                                Activity Buffer Request        15.08%     211.984us        15.08%     211.984us     211.984us       1.792us         1.45%       1.792us       1.792us             1  
+                                    aten::empty_strided         2.18%      30.690us         2.18%      30.690us       5.115us       0.000us         0.00%       0.000us       0.000us             6  
+                                        cudaMemcpyAsync        14.35%     201.734us        14.35%     201.734us      33.622us       0.000us         0.00%       0.000us       0.000us             6  
+                                            aten::slice         4.89%      68.724us         6.32%      88.851us       3.702us       0.000us         0.00%       0.000us       0.000us            24  
+                                       aten::as_strided         1.43%      20.127us         1.43%      20.127us       0.839us       0.000us         0.00%       0.000us       0.000us            24  
+                                       cudaLaunchKernel        15.74%     221.155us        15.74%     221.155us       4.607us       0.000us         0.00%       0.000us       0.000us            48  
+                                  cudaDeviceSynchronize         0.40%       5.570us         0.40%       5.570us       5.570us       0.000us         0.00%       0.000us       0.000us             1  
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
-Self CPU time total: 1.420ms
-Self CUDA time total: 123.842us
+Self CPU time total: 1.405ms
+Self CUDA time total: 123.646us
 
 
 
@@ -4547,27 +4774,27 @@ PROFILE TRACE: torch_eager | cuda_B2_S512_H32_D64_R32
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
                                                    Name    Self CPU %      Self CPU   CPU total %     CPU total  CPU time avg     Self CUDA   Self CUDA %    CUDA total  CUDA time avg    # of Calls  
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
-                                            torch_eager         0.00%       0.000us         0.00%       0.000us       0.000us     910.045us       513.35%     910.045us     910.045us             1  
-                                            torch_eager         9.66%     280.213us        99.83%       2.894ms       2.894ms       0.000us         0.00%     180.188us     180.188us             1  
-                                              aten::mul         5.18%     150.102us         9.00%     260.863us      10.869us      94.655us        53.39%      94.655us       3.944us            24  
-void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      94.655us        53.39%      94.655us       3.944us            24  
-                                            aten::copy_         3.40%      98.673us        72.45%       2.101ms     116.706us      57.885us        32.65%      60.797us       3.378us            18  
-void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      40.799us        23.01%      40.799us       3.400us            12  
-void at::native::vectorized_elementwise_kernel&lt;4, at...         0.00%       0.000us         0.00%       0.000us       0.000us      24.736us        13.95%      24.736us       2.061us            12  
-                                            aten::clone         0.79%      22.860us        70.00%       2.030ms     338.262us       0.000us         0.00%      19.998us       3.333us             6  
-                         Memcpy DtoD (Device -&gt; Device)         0.00%       0.000us         0.00%       0.000us       0.000us      17.086us         9.64%      17.086us       2.848us             6  
-                                              aten::add         1.13%      32.880us         1.89%      54.761us       9.127us      12.416us         7.00%      12.416us       2.069us             6  
-                                              aten::sub         1.18%      34.239us         1.98%      57.551us       9.592us      12.320us         6.95%      12.320us       2.053us             6  
-                                Activity Buffer Request        58.76%       1.704ms        58.76%       1.704ms       1.704ms       2.912us         1.64%       2.912us       2.912us             1  
-                                    aten::empty_strided         1.11%      32.150us         1.11%      32.150us       5.358us       0.000us         0.00%       0.000us       0.000us             6  
-                                        cudaMemcpyAsync         8.21%     238.144us         8.21%     238.144us      39.691us       0.000us         0.00%       0.000us       0.000us             6  
-                                            aten::slice         2.29%      66.481us         2.94%      85.213us       3.551us       0.000us         0.00%       0.000us       0.000us            24  
-                                       aten::as_strided         0.65%      18.732us         0.65%      18.732us       0.781us       0.000us         0.00%       0.000us       0.000us            24  
-                                       cudaLaunchKernel         7.46%     216.224us         7.46%     216.224us       4.505us       0.000us         0.00%       0.000us       0.000us            48  
-                                  cudaDeviceSynchronize         0.17%       5.070us         0.17%       5.070us       5.070us       0.000us         0.00%       0.000us       0.000us             1  
+                                            torch_eager         0.00%       0.000us         0.00%       0.000us       0.000us     938.077us       529.63%     938.077us     938.077us             1  
+                                            torch_eager        22.00%     288.844us        99.57%       1.307ms       1.307ms       0.000us         0.00%     179.967us     179.967us             1  
+                                              aten::mul        11.92%     156.562us        20.13%     264.245us      11.010us      94.881us        53.57%      94.881us       3.953us            24  
+void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      94.881us        53.57%      94.881us       3.953us            24  
+                                            aten::copy_         8.04%     105.524us        37.72%     495.290us      27.516us      57.663us        32.56%      60.511us       3.362us            18  
+void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      40.672us        22.96%      40.672us       3.389us            12  
+void at::native::vectorized_elementwise_kernel&lt;4, at...         0.00%       0.000us         0.00%       0.000us       0.000us      24.575us        13.87%      24.575us       2.048us            12  
+                                            aten::clone         1.60%      21.071us        31.51%     413.758us      68.960us       0.000us         0.00%      19.839us       3.306us             6  
+                         Memcpy DtoD (Device -&gt; Device)         0.00%       0.000us         0.00%       0.000us       0.000us      16.991us         9.59%      16.991us       2.832us             6  
+                                              aten::add         2.42%      31.800us         4.16%      54.561us       9.093us      12.288us         6.94%      12.288us       2.048us             6  
+                                              aten::sub         3.05%      40.090us         5.01%      65.752us      10.959us      12.287us         6.94%      12.287us       2.048us             6  
+                                Activity Buffer Request        10.75%     141.113us        10.75%     141.113us     141.113us       2.848us         1.61%       2.848us       2.848us             1  
+                                    aten::empty_strided         2.28%      29.940us         2.28%      29.940us       4.990us       0.000us         0.00%       0.000us       0.000us             6  
+                                        cudaMemcpyAsync        14.21%     186.543us        14.21%     186.543us      31.091us       0.000us         0.00%       0.000us       0.000us             6  
+                                            aten::slice         5.18%      67.990us         6.68%      87.660us       3.652us       0.000us         0.00%       0.000us       0.000us            24  
+                                       aten::as_strided         1.50%      19.670us         1.50%      19.670us       0.820us       0.000us         0.00%       0.000us       0.000us            24  
+                                       cudaLaunchKernel        16.62%     218.216us        16.62%     218.216us       4.546us       0.000us         0.00%       0.000us       0.000us            48  
+                                  cudaDeviceSynchronize         0.43%       5.650us         0.43%       5.650us       5.650us       0.000us         0.00%       0.000us       0.000us             1  
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
-Self CPU time total: 2.899ms
-Self CUDA time total: 177.276us
+Self CPU time total: 1.313ms
+Self CUDA time total: 177.119us
 
 
 
@@ -4577,27 +4804,27 @@ PROFILE TRACE: torch_eager | cuda_B2_S512_H32_D128_R64
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
                                                    Name    Self CPU %      Self CPU   CPU total %     CPU total  CPU time avg     Self CUDA   Self CUDA %    CUDA total  CUDA time avg    # of Calls  
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
-                                            torch_eager         0.00%       0.000us         0.00%       0.000us       0.000us     919.612us       310.44%     919.612us     919.612us             1  
-                                            torch_eager        10.49%     286.464us        99.82%       2.726ms       2.726ms       0.000us         0.00%     313.057us     313.057us             1  
-                                              aten::mul         5.34%     145.716us         9.29%     253.789us      10.575us     145.182us        49.01%     145.182us       6.049us            24  
-void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us     145.182us        49.01%     145.182us       6.049us            24  
-                                            aten::copy_         3.69%     100.696us        70.60%       1.928ms     107.115us     109.985us        37.13%     126.817us       7.045us            18  
-                                            aten::clone         0.88%      23.951us        68.02%       1.858ms     309.597us       0.000us         0.00%      69.474us      11.579us             6  
-void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      57.343us        19.36%      57.343us       4.779us            12  
-                         Memcpy DtoD (Device -&gt; Device)         0.00%       0.000us         0.00%       0.000us       0.000us      52.642us        17.77%      52.642us       8.774us             6  
-void at::native::vectorized_elementwise_kernel&lt;4, at...         0.00%       0.000us         0.00%       0.000us       0.000us      41.058us        13.86%      41.058us       3.421us            12  
-                                              aten::sub         1.33%      36.191us         2.18%      59.621us       9.937us      20.609us         6.96%      20.609us       3.435us             6  
-                                              aten::add         1.14%      31.230us         1.95%      53.190us       8.865us      20.449us         6.90%      20.449us       3.408us             6  
-                                Activity Buffer Request        56.07%       1.531ms        56.07%       1.531ms       1.531ms      16.832us         5.68%      16.832us      16.832us             1  
-                                    aten::empty_strided         1.17%      32.070us         1.17%      32.070us       5.345us       0.000us         0.00%       0.000us       0.000us             6  
-                                        cudaMemcpyAsync         8.59%     234.696us         8.59%     234.696us      39.116us       0.000us         0.00%       0.000us       0.000us             6  
-                                            aten::slice         2.53%      69.062us         3.26%      88.922us       3.705us       0.000us         0.00%       0.000us       0.000us            24  
-                                       aten::as_strided         0.73%      19.860us         0.73%      19.860us       0.827us       0.000us         0.00%       0.000us       0.000us            24  
-                                       cudaLaunchKernel         7.86%     214.752us         7.86%     214.752us       4.474us       0.000us         0.00%       0.000us       0.000us            48  
-                                  cudaDeviceSynchronize         0.18%       4.930us         0.18%       4.930us       4.930us       0.000us         0.00%       0.000us       0.000us             1  
+                                            torch_eager         0.00%       0.000us         0.00%       0.000us       0.000us     945.076us       318.26%     945.076us     945.076us             1  
+                                            torch_eager        21.55%     289.808us        99.58%       1.339ms       1.339ms       0.000us         0.00%     314.171us     314.171us             1  
+                                              aten::mul        11.43%     153.633us        19.62%     263.817us      10.992us     145.952us        49.15%     145.952us       6.081us            24  
+void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us     145.952us        49.15%     145.952us       6.081us            24  
+                                            aten::copy_         9.11%     122.489us        38.99%     524.297us      29.128us     110.173us        37.10%     127.389us       7.077us            18  
+                                            aten::clone         1.65%      22.169us        33.13%     445.468us      74.245us       0.000us         0.00%      70.110us      11.685us             6  
+void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      57.279us        19.29%      57.279us       4.773us            12  
+                         Memcpy DtoD (Device -&gt; Device)         0.00%       0.000us         0.00%       0.000us       0.000us      52.894us        17.81%      52.894us       8.816us             6  
+void at::native::vectorized_elementwise_kernel&lt;4, at...         0.00%       0.000us         0.00%       0.000us       0.000us      40.830us        13.75%      40.830us       3.402us            12  
+                                              aten::sub         2.94%      39.549us         4.81%      64.690us      10.782us      20.511us         6.91%      20.511us       3.418us             6  
+                                              aten::add         2.41%      32.411us         4.09%      55.020us       9.170us      20.319us         6.84%      20.319us       3.386us             6  
+                                Activity Buffer Request        11.32%     152.193us        11.32%     152.193us     152.193us      17.216us         5.80%      17.216us      17.216us             1  
+                                    aten::empty_strided         2.31%      31.082us         2.31%      31.082us       5.180us       0.000us         0.00%       0.000us       0.000us             6  
+                                        cudaMemcpyAsync        13.88%     186.593us        13.88%     186.593us      31.099us       0.000us         0.00%       0.000us       0.000us             6  
+                                            aten::slice         5.09%      68.450us         6.56%      88.160us       3.673us       0.000us         0.00%       0.000us       0.000us            24  
+                                       aten::as_strided         1.47%      19.710us         1.47%      19.710us       0.821us       0.000us         0.00%       0.000us       0.000us            24  
+                                       cudaLaunchKernel        16.43%     220.956us        16.43%     220.956us       4.603us       0.000us         0.00%       0.000us       0.000us            48  
+                                  cudaDeviceSynchronize         0.42%       5.661us         0.42%       5.661us       5.661us       0.000us         0.00%       0.000us       0.000us             1  
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
-Self CPU time total: 2.731ms
-Self CUDA time total: 296.225us
+Self CPU time total: 1.345ms
+Self CUDA time total: 296.955us
 
 
 
@@ -4607,27 +4834,27 @@ PROFILE TRACE: torch_eager | cuda_B2_S2048_H8_D64_R32
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
                                                    Name    Self CPU %      Self CPU   CPU total %     CPU total  CPU time avg     Self CUDA   Self CUDA %    CUDA total  CUDA time avg    # of Calls  
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
-                                            torch_eager         0.00%       0.000us         0.00%       0.000us       0.000us     889.394us       501.79%     889.394us     889.394us             1  
-                                            torch_eager        17.97%     266.975us        99.65%       1.481ms       1.481ms       0.000us         0.00%     180.092us     180.092us             1  
-                                              aten::mul         9.80%     145.611us        16.96%     251.937us      10.497us      94.974us        53.58%      94.974us       3.957us            24  
-void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      94.974us        53.58%      94.974us       3.957us            24  
-                                            aten::copy_         6.75%     100.282us        47.98%     712.837us      39.602us      57.694us        32.55%      60.542us       3.363us            18  
-void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      40.736us        22.98%      40.736us       3.395us            12  
-void at::native::vectorized_elementwise_kernel&lt;4, at...         0.00%       0.000us         0.00%       0.000us       0.000us      24.576us        13.87%      24.576us       2.048us            12  
-                                            aten::clone         1.38%      20.549us        43.06%     639.725us     106.621us       0.000us         0.00%      19.806us       3.301us             6  
-                         Memcpy DtoD (Device -&gt; Device)         0.00%       0.000us         0.00%       0.000us       0.000us      16.958us         9.57%      16.958us       2.826us             6  
-                                              aten::sub         2.49%      37.040us         4.14%      61.531us      10.255us      12.289us         6.93%      12.289us       2.048us             6  
-                                              aten::add         2.11%      31.282us         3.59%      53.402us       8.900us      12.287us         6.93%      12.287us       2.048us             6  
-                                Activity Buffer Request        19.87%     295.257us        19.87%     295.257us     295.257us       2.848us         1.61%       2.848us       2.848us             1  
-                                    aten::empty_strided         2.04%      30.372us         2.04%      30.372us       5.062us       0.000us         0.00%       0.000us       0.000us             6  
-                                        cudaMemcpyAsync        17.34%     257.637us        17.34%     257.637us      42.940us       0.000us         0.00%       0.000us       0.000us             6  
-                                            aten::slice         4.31%      64.000us         5.58%      82.951us       3.456us       0.000us         0.00%       0.000us       0.000us            24  
-                                       aten::as_strided         1.28%      18.951us         1.28%      18.951us       0.790us       0.000us         0.00%       0.000us       0.000us            24  
-                                       cudaLaunchKernel        14.31%     212.598us        14.31%     212.598us       4.429us       0.000us         0.00%       0.000us       0.000us            48  
-                                  cudaDeviceSynchronize         0.35%       5.130us         0.35%       5.130us       5.130us       0.000us         0.00%       0.000us       0.000us             1  
+                                            torch_eager         0.00%       0.000us         0.00%       0.000us       0.000us     986.080us       556.73%     986.080us     986.080us             1  
+                                            torch_eager        12.52%     336.567us        99.81%       2.683ms       2.683ms       0.000us         0.00%     179.999us     179.999us             1  
+                                              aten::mul         5.82%     156.365us         9.99%     268.575us      11.191us      94.976us        53.62%      94.976us       3.957us            24  
+void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      94.976us        53.62%      94.976us       3.957us            24  
+                                            aten::copy_         3.98%     106.939us        67.04%       1.802ms     100.094us      57.535us        32.48%      60.415us       3.356us            18  
+void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      40.703us        22.98%      40.703us       3.392us            12  
+void at::native::vectorized_elementwise_kernel&lt;4, at...         0.00%       0.000us         0.00%       0.000us       0.000us      24.608us        13.89%      24.608us       2.051us            12  
+                                            aten::clone         1.08%      29.091us        64.22%       1.726ms     287.668us       0.000us         0.00%      19.712us       3.285us             6  
+                         Memcpy DtoD (Device -&gt; Device)         0.00%       0.000us         0.00%       0.000us       0.000us      16.832us         9.50%      16.832us       2.805us             6  
+                                              aten::add         1.21%      32.499us         2.06%      55.240us       9.207us      12.320us         6.96%      12.320us       2.053us             6  
+                                              aten::sub         1.59%      42.650us         2.57%      69.041us      11.507us      12.288us         6.94%      12.288us       2.048us             6  
+                                Activity Buffer Request        53.52%       1.438ms        53.52%       1.438ms       1.438ms       2.880us         1.63%       2.880us       2.880us             1  
+                                    aten::empty_strided         1.16%      31.221us         1.16%      31.221us       5.204us       0.000us         0.00%       0.000us       0.000us             6  
+                                        cudaMemcpyAsync         7.20%     193.473us         7.20%     193.473us      32.245us       0.000us         0.00%       0.000us       0.000us             6  
+                                            aten::slice         2.61%      70.195us         3.39%      91.232us       3.801us       0.000us         0.00%       0.000us       0.000us            24  
+                                       aten::as_strided         0.78%      21.037us         0.78%      21.037us       0.877us       0.000us         0.00%       0.000us       0.000us            24  
+                                       cudaLaunchKernel         8.35%     224.324us         8.35%     224.324us       4.673us       0.000us         0.00%       0.000us       0.000us            48  
+                                  cudaDeviceSynchronize         0.19%       4.980us         0.19%       4.980us       4.980us       0.000us         0.00%       0.000us       0.000us             1  
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
-Self CPU time total: 1.486ms
-Self CUDA time total: 177.244us
+Self CPU time total: 2.688ms
+Self CUDA time total: 177.119us
 
 
 
@@ -4637,27 +4864,27 @@ PROFILE TRACE: torch_eager | cuda_B2_S2048_H8_D128_R64
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
                                                    Name    Self CPU %      Self CPU   CPU total %     CPU total  CPU time avg     Self CUDA   Self CUDA %    CUDA total  CUDA time avg    # of Calls  
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
-                                            torch_eager         0.00%       0.000us         0.00%       0.000us       0.000us     909.016us       306.24%     909.016us     909.016us             1  
-                                            torch_eager        19.05%     269.264us        99.66%       1.409ms       1.409ms       0.000us         0.00%     314.684us     314.684us             1  
-                                              aten::mul        10.56%     149.323us        19.02%     268.875us      11.203us     145.440us        49.00%     145.440us       6.060us            24  
-void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us     145.440us        49.00%     145.440us       6.060us            24  
-                                            aten::copy_         6.96%      98.305us        44.09%     623.125us      34.618us     110.751us        37.31%     128.606us       7.145us            18  
-                                            aten::clone         1.45%      20.520us        38.80%     548.422us      91.404us       0.000us         0.00%      71.453us      11.909us             6  
-void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      57.153us        19.25%      57.153us       4.763us            12  
-                         Memcpy DtoD (Device -&gt; Device)         0.00%       0.000us         0.00%       0.000us       0.000us      53.598us        18.06%      53.598us       8.933us             6  
-void at::native::vectorized_elementwise_kernel&lt;4, at...         0.00%       0.000us         0.00%       0.000us       0.000us      40.638us        13.69%      40.638us       3.387us            12  
-                                              aten::add         2.27%      32.070us         3.85%      54.390us       9.065us      20.352us         6.86%      20.352us       3.392us             6  
-                                              aten::sub         2.35%      33.277us         4.05%      57.282us       9.547us      20.286us         6.83%      20.286us       3.381us             6  
-                                Activity Buffer Request        15.96%     225.655us        15.96%     225.655us     225.655us      17.855us         6.02%      17.855us      17.855us             1  
-                                    aten::empty_strided         2.15%      30.350us         2.15%      30.350us       5.058us       0.000us         0.00%       0.000us       0.000us             6  
-                                        cudaMemcpyAsync        16.79%     237.294us        16.79%     237.294us      39.549us       0.000us         0.00%       0.000us       0.000us             6  
-                                            aten::slice         4.69%      66.249us         6.00%      84.797us       3.533us       0.000us         0.00%       0.000us       0.000us            24  
-                                       aten::as_strided         1.31%      18.548us         1.31%      18.548us       0.773us       0.000us         0.00%       0.000us       0.000us            24  
-                                       cudaLaunchKernel        16.11%     227.748us        16.11%     227.748us       4.745us       0.000us         0.00%       0.000us       0.000us            48  
-                                  cudaDeviceSynchronize         0.34%       4.840us         0.34%       4.840us       4.840us       0.000us         0.00%       0.000us       0.000us             1  
+                                            torch_eager         0.00%       0.000us         0.00%       0.000us       0.000us     955.007us       321.87%     955.007us     955.007us             1  
+                                            torch_eager        21.61%     290.382us        99.58%       1.338ms       1.338ms       0.000us         0.00%     314.050us     314.050us             1  
+                                              aten::mul        12.35%     165.965us        20.49%     275.388us      11.475us     146.274us        49.30%     146.274us       6.095us            24  
+void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us     146.274us        49.30%     146.274us       6.095us            24  
+                                            aten::copy_         7.99%     107.375us        38.18%     513.111us      28.506us     109.984us        37.07%     127.328us       7.074us            18  
+                                            aten::clone         1.53%      20.570us        31.98%     429.868us      71.645us       0.000us         0.00%      70.048us      11.675us             6  
+void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      57.280us        19.31%      57.280us       4.773us            12  
+                         Memcpy DtoD (Device -&gt; Device)         0.00%       0.000us         0.00%       0.000us       0.000us      52.704us        17.76%      52.704us       8.784us             6  
+void at::native::vectorized_elementwise_kernel&lt;4, at...         0.00%       0.000us         0.00%       0.000us       0.000us      40.448us        13.63%      40.448us       3.371us            12  
+                                              aten::sub         2.99%      40.150us         4.79%      64.400us      10.733us      20.288us         6.84%      20.288us       3.381us             6  
+                                              aten::add         2.45%      32.907us         4.13%      55.499us       9.250us      20.160us         6.79%      20.160us       3.360us             6  
+                                Activity Buffer Request        11.77%     158.223us        11.77%     158.223us     158.223us      17.344us         5.85%      17.344us      17.344us             1  
+                                    aten::empty_strided         2.28%      30.711us         2.28%      30.711us       5.118us       0.000us         0.00%       0.000us       0.000us             6  
+                                        cudaMemcpyAsync        13.78%     185.224us        13.78%     185.224us      30.871us       0.000us         0.00%       0.000us       0.000us             6  
+                                            aten::slice         5.13%      68.942us         6.58%      88.372us       3.682us       0.000us         0.00%       0.000us       0.000us            24  
+                                       aten::as_strided         1.45%      19.430us         1.45%      19.430us       0.810us       0.000us         0.00%       0.000us       0.000us            24  
+                                       cudaLaunchKernel        16.26%     218.554us        16.26%     218.554us       4.553us       0.000us         0.00%       0.000us       0.000us            48  
+                                  cudaDeviceSynchronize         0.42%       5.611us         0.42%       5.611us       5.611us       0.000us         0.00%       0.000us       0.000us             1  
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
-Self CPU time total: 1.413ms
-Self CUDA time total: 296.829us
+Self CPU time total: 1.344ms
+Self CUDA time total: 296.706us
 
 
 
@@ -4667,27 +4894,27 @@ PROFILE TRACE: torch_eager | cuda_B2_S2048_H32_D64_R32
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
                                                    Name    Self CPU %      Self CPU   CPU total %     CPU total  CPU time avg     Self CUDA   Self CUDA %    CUDA total  CUDA time avg    # of Calls  
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
-                                            torch_eager         0.00%       0.000us         0.00%       0.000us       0.000us     916.757us       157.09%     916.757us     916.757us             1  
-                                            torch_eager        19.46%     274.242us        99.65%       1.404ms       1.404ms       0.000us         0.00%     607.350us     607.350us             1  
-                                            aten::copy_         7.01%      98.793us        43.42%     611.905us      33.995us     268.603us        46.03%     292.379us      16.243us            18  
-                                              aten::mul        10.57%     148.926us        18.84%     265.480us      11.062us     249.086us        42.68%     249.086us      10.379us            24  
-void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us     249.086us        42.68%     249.086us      10.379us            24  
-                                            aten::clone         1.44%      20.340us        38.12%     537.253us      89.542us       0.000us         0.00%     202.173us      33.696us             6  
-                         Memcpy DtoD (Device -&gt; Device)         0.00%       0.000us         0.00%       0.000us       0.000us     178.397us        30.57%     178.397us      29.733us             6  
-void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      90.206us        15.46%      90.206us       7.517us            12  
-void at::native::vectorized_elementwise_kernel&lt;4, at...         0.00%       0.000us         0.00%       0.000us       0.000us      65.885us        11.29%      65.885us       5.490us            12  
-                                              aten::sub         2.63%      37.022us         4.37%      61.602us      10.267us      33.151us         5.68%      33.151us       5.525us             6  
-                                              aten::add         2.33%      32.810us         3.92%      55.180us       9.197us      32.734us         5.61%      32.734us       5.456us             6  
-                                Activity Buffer Request        15.58%     219.605us        15.58%     219.605us     219.605us      23.776us         4.07%      23.776us      23.776us             1  
-                                    aten::empty_strided         2.10%      29.631us         2.10%      29.631us       4.938us       0.000us         0.00%       0.000us       0.000us             6  
-                                        cudaMemcpyAsync        16.49%     232.396us        16.49%     232.396us      38.733us       0.000us         0.00%       0.000us       0.000us             6  
-                                            aten::slice         4.73%      66.612us         6.10%      85.953us       3.581us       0.000us         0.00%       0.000us       0.000us            24  
-                                       aten::as_strided         1.37%      19.341us         1.37%      19.341us       0.806us       0.000us         0.00%       0.000us       0.000us            24  
-                                       cudaLaunchKernel        15.94%     224.615us        15.94%     224.615us       4.679us       0.000us         0.00%       0.000us       0.000us            48  
-                                  cudaDeviceSynchronize         0.35%       4.910us         0.35%       4.910us       4.910us       0.000us         0.00%       0.000us       0.000us             1  
+                                            torch_eager         0.00%       0.000us         0.00%       0.000us       0.000us     962.939us       164.48%     962.939us     962.939us             1  
+                                            torch_eager        21.30%     292.019us        99.59%       1.365ms       1.365ms       0.000us         0.00%     609.117us     609.117us             1  
+                                            aten::copy_         7.59%     104.052us        39.10%     536.059us      29.781us     268.735us        45.90%     292.415us      16.245us            18  
+                                              aten::mul        11.61%     159.130us        19.77%     271.083us      11.295us     251.454us        42.95%     251.454us      10.477us            24  
+void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us     251.454us        42.95%     251.454us      10.477us            24  
+                                            aten::clone         1.60%      21.919us        33.19%     455.067us      75.844us       0.000us         0.00%     201.504us      33.584us             6  
+                         Memcpy DtoD (Device -&gt; Device)         0.00%       0.000us         0.00%       0.000us       0.000us     177.824us        30.37%     177.824us      29.637us             6  
+void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      90.911us        15.53%      90.911us       7.576us            12  
+void at::native::vectorized_elementwise_kernel&lt;4, at...         0.00%       0.000us         0.00%       0.000us       0.000us      65.248us        11.15%      65.248us       5.437us            12  
+                                              aten::sub         2.98%      40.869us         4.94%      67.700us      11.283us      32.703us         5.59%      32.703us       5.451us             6  
+                                              aten::add         2.40%      32.850us         4.07%      55.841us       9.307us      32.545us         5.56%      32.545us       5.424us             6  
+                                Activity Buffer Request        13.18%     180.724us        13.18%     180.724us     180.724us      23.680us         4.04%      23.680us      23.680us             1  
+                                    aten::empty_strided         2.23%      30.541us         2.23%      30.541us       5.090us       0.000us         0.00%       0.000us       0.000us             6  
+                                        cudaMemcpyAsync        13.71%     188.023us        13.71%     188.023us      31.337us       0.000us         0.00%       0.000us       0.000us             6  
+                                            aten::slice         5.13%      70.322us         6.59%      90.292us       3.762us       0.000us         0.00%       0.000us       0.000us            24  
+                                       aten::as_strided         1.46%      19.970us         1.46%      19.970us       0.832us       0.000us         0.00%       0.000us       0.000us            24  
+                                       cudaLaunchKernel        16.41%     225.035us        16.41%     225.035us       4.688us       0.000us         0.00%       0.000us       0.000us            48  
+                                  cudaDeviceSynchronize         0.41%       5.640us         0.41%       5.640us       5.640us       0.000us         0.00%       0.000us       0.000us             1  
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
-Self CPU time total: 1.409ms
-Self CUDA time total: 583.574us
+Self CPU time total: 1.371ms
+Self CUDA time total: 585.437us
 
 
 
@@ -4697,61 +4924,55 @@ PROFILE TRACE: torch_eager | cuda_B2_S2048_H32_D128_R64
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
                                                    Name    Self CPU %      Self CPU   CPU total %     CPU total  CPU time avg     Self CUDA   Self CUDA %    CUDA total  CUDA time avg    # of Calls  
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
-                                            torch_eager        12.10%     272.127us        61.47%       1.382ms       1.382ms       0.000us         0.00%       1.837ms       1.837ms             1  
-                                            torch_eager         0.00%       0.000us         0.00%       0.000us       0.000us       1.810ms       102.21%       1.810ms       1.810ms             1  
-                                            aten::copy_         4.74%     106.692us        27.02%     607.756us      33.764us     794.110us        44.84%     859.966us      47.776us            18  
-                                              aten::mul         6.35%     142.895us        11.18%     251.386us      10.474us     829.085us        46.82%     829.085us      34.545us            24  
-void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us     829.085us        46.82%     829.085us      34.545us            24  
-                                            aten::clone         0.94%      21.099us        23.42%     526.743us      87.790us       0.000us         0.00%     627.678us     104.613us             6  
-                         Memcpy DtoD (Device -&gt; Device)         0.00%       0.000us         0.00%       0.000us       0.000us     561.822us        31.73%     561.822us      93.637us             6  
-void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us     232.288us        13.12%     232.288us      19.357us            12  
-void at::native::vectorized_elementwise_kernel&lt;4, at...         0.00%       0.000us         0.00%       0.000us       0.000us     147.650us         8.34%     147.650us      12.304us            12  
-                                              aten::sub         1.58%      35.541us         2.61%      58.661us       9.777us      89.538us         5.06%      89.538us      14.923us             6  
-                                Activity Buffer Request         9.29%     208.845us         9.29%     208.845us     208.845us      65.856us         3.72%      65.856us      65.856us             1  
-                                              aten::add         1.43%      32.251us         2.42%      54.461us       9.077us      58.112us         3.28%      58.112us       9.685us             6  
-                                    aten::empty_strided         1.39%      31.342us         1.39%      31.342us       5.224us       0.000us         0.00%       0.000us       0.000us             6  
-                                        cudaMemcpyAsync        10.27%     230.957us        10.27%     230.957us      38.493us       0.000us         0.00%       0.000us       0.000us             6  
-                                            aten::slice         2.99%      67.270us         3.80%      85.550us       3.565us       0.000us         0.00%       0.000us       0.000us            24  
-                                       aten::as_strided         0.81%      18.280us         0.81%      18.280us       0.762us       0.000us         0.00%       0.000us       0.000us            24  
-                                       cudaLaunchKernel         9.56%     215.083us         9.56%     215.083us       4.481us       0.000us         0.00%       0.000us       0.000us            48  
-                                  cudaDeviceSynchronize        38.53%     866.589us        38.53%     866.589us     866.589us       0.000us         0.00%       0.000us       0.000us             1  
+                                            torch_eager         9.18%     318.848us        77.56%       2.693ms       2.693ms       0.000us         0.00%       1.840ms       1.840ms             1  
+                                            torch_eager         0.00%       0.000us         0.00%       0.000us       0.000us       1.811ms       102.06%       1.811ms       1.811ms             1  
+                                            aten::copy_         3.19%     110.682us        53.02%       1.841ms     102.257us     792.737us        44.68%     858.369us      47.687us            18  
+                                              aten::mul         4.39%     152.554us         7.57%     262.845us      10.952us     833.316us        46.97%     833.316us      34.721us            24  
+void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us     833.316us        46.97%     833.316us      34.721us            24  
+                                            aten::clone         0.79%      27.538us        50.82%       1.764ms     294.050us       0.000us         0.00%     624.865us     104.144us             6  
+                         Memcpy DtoD (Device -&gt; Device)         0.00%       0.000us         0.00%       0.000us       0.000us     559.233us        31.52%     559.233us      93.206us             6  
+void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us     233.504us        13.16%     233.504us      19.459us            12  
+void at::native::vectorized_elementwise_kernel&lt;4, at...         0.00%       0.000us         0.00%       0.000us       0.000us     148.032us         8.34%     148.032us      12.336us            12  
+                                              aten::sub         1.13%      39.132us         1.88%      65.111us      10.852us      90.112us         5.08%      90.112us      15.019us             6  
+                                Activity Buffer Request        41.37%       1.436ms        41.37%       1.436ms       1.436ms      65.632us         3.70%      65.632us      65.632us             1  
+                                              aten::add         0.97%      33.650us         1.61%      56.062us       9.344us      57.920us         3.26%      57.920us       9.653us             6  
+                                    aten::empty_strided         0.92%      31.941us         0.92%      31.941us       5.324us       0.000us         0.00%       0.000us       0.000us             6  
+                                        cudaMemcpyAsync         6.62%     229.834us         6.62%     229.834us      38.306us       0.000us         0.00%       0.000us       0.000us             6  
+                                            aten::slice         2.00%      69.363us         2.59%      89.831us       3.743us       0.000us         0.00%       0.000us       0.000us            24  
+                                       aten::as_strided         0.59%      20.468us         0.59%      20.468us       0.853us       0.000us         0.00%       0.000us       0.000us            24  
+                                       cudaLaunchKernel         6.41%     222.613us         6.41%     222.613us       4.638us       0.000us         0.00%       0.000us       0.000us            48  
+                                  cudaDeviceSynchronize        22.44%     778.913us        22.44%     778.913us     778.913us       0.000us         0.00%       0.000us       0.000us             1  
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
-Self CPU time total: 2.249ms
-Self CUDA time total: 1.771ms
+Self CPU time total: 3.472ms
+Self CUDA time total: 1.774ms
 
 
 impl                     wl                  p50(ms)  ok
-torch_eager              cuda_B1_S128_H32_D128_R64     0.21  True
-torch_eager              cuda_B1_S128_H32_D64_R32     0.22  True
-torch_eager              cuda_B1_S128_H8_D128_R64     0.22  True
+torch_eager              cuda_B1_S128_H32_D128_R64     0.22  True
+torch_eager              cuda_B1_S128_H32_D64_R32     0.23  True
+torch_eager              cuda_B1_S128_H8_D128_R64     0.23  True
 torch_eager              cuda_B1_S128_H8_D64_R32     0.17  True
-torch_eager              cuda_B1_S2048_H32_D128_R64     0.22  True
-torch_eager              cuda_B1_S2048_H32_D64_R32     0.21  True
-torch_eager              cuda_B1_S2048_H8_D128_R64     0.21  True
-torch_eager              cuda_B1_S2048_H8_D64_R32     0.21  True
-torch_eager              cuda_B1_S512_H32_D128_R64     0.21  True
-torch_eager              cuda_B1_S512_H32_D64_R32     0.21  True
-torch_eager              cuda_B1_S512_H8_D128_R64     0.21  True
+torch_eager              cuda_B1_S2048_H32_D128_R64     0.23  True
+torch_eager              cuda_B1_S2048_H32_D64_R32     0.22  True
+torch_eager              cuda_B1_S2048_H8_D128_R64     0.22  True
+torch_eager              cuda_B1_S2048_H8_D64_R32     0.22  True
+torch_eager              cuda_B1_S512_H32_D128_R64     0.22  True
+torch_eager              cuda_B1_S512_H32_D64_R32     0.22  True
+torch_eager              cuda_B1_S512_H8_D128_R64     0.22  True
 torch_eager              cuda_B1_S512_H8_D64_R32     0.22  True
-torch_eager              cuda_B2_S128_H32_D128_R64     0.21  True
-torch_eager              cuda_B2_S128_H32_D64_R32     0.21  True
-torch_eager              cuda_B2_S128_H8_D128_R64     0.21  True
-torch_eager              cuda_B2_S128_H8_D64_R32     0.21  True
+torch_eager              cuda_B2_S128_H32_D128_R64     0.22  True
+torch_eager              cuda_B2_S128_H32_D64_R32     0.22  True
+torch_eager              cuda_B2_S128_H8_D128_R64     0.22  True
+torch_eager              cuda_B2_S128_H8_D64_R32     0.22  True
 torch_eager              cuda_B2_S2048_H32_D128_R64     0.64  True
 torch_eager              cuda_B2_S2048_H32_D64_R32     0.23  True
-torch_eager              cuda_B2_S2048_H8_D128_R64     0.22  True
-torch_eager              cuda_B2_S2048_H8_D64_R32     0.21  True
-torch_eager              cuda_B2_S512_H32_D128_R64     0.21  True
-torch_eager              cuda_B2_S512_H32_D64_R32     0.21  True
-torch_eager              cuda_B2_S512_H8_D128_R64     0.21  True
-torch_eager              cuda_B2_S512_H8_D64_R32     0.21  True
+torch_eager              cuda_B2_S2048_H8_D128_R64     0.23  True
+torch_eager              cuda_B2_S2048_H8_D64_R32     0.22  True
+torch_eager              cuda_B2_S512_H32_D128_R64     0.22  True
+torch_eager              cuda_B2_S512_H32_D64_R32     0.23  True
+torch_eager              cuda_B2_S512_H8_D128_R64     0.22  True
+torch_eager              cuda_B2_S512_H8_D64_R32     0.22  True
 </pre></div>
-<div class="uv-install-logs" id="uv-logs-benchmark">
-<div class="uv-logs-header" onclick="toggleUvLogs(this)">▶ UV Install Logs</div>
-<div class="uv-logs-content" style="display: none;">
-Installed 37 packages in 226ms
-</div>
-</div>
 <div class="cell-artifacts">
 <h4>Artifacts:</h4>
 <a href="artifacts/benchmark/rotary.jsonl" class="artifact" target="_blank">rotary.jsonl</a>