diff --git "a/causal_conv1d/impls/torch_causal_conv1d.html" "b/causal_conv1d/impls/torch_causal_conv1d.html"
--- "a/causal_conv1d/impls/torch_causal_conv1d.html"
+++ "b/causal_conv1d/impls/torch_causal_conv1d.html"
@@ -4,6 +4,11 @@
     <meta charset="UTF-8">
     <meta name="viewport" content="width=device-width, initial-scale=1.0">
     <title>torch_causal_conv1d</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 | 0.21s
+Cell: nv | 0.28s
  | <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 | 0.21s
 </div>
 </div>
 <div id="output-nv" class="cell-output">
-<div class="cell-stdout"><pre class="stdout-text">Wed Oct 29 15:50:16 2025       
+<div class="cell-stdout"><pre class="stdout-text">Thu Oct 30 15:51:43 2025       
 +-----------------------------------------------------------------------------------------+
 | NVIDIA-SMI 570.195.03             Driver Version: 570.195.03     CUDA Version: 12.8     |
 |-----------------------------------------+------------------------+----------------------+
@@ -3904,7 +4131,7 @@ Cell: nv | 0.21s
 |                                         |                        |               MIG M. |
 |=========================================+========================+======================|
 |   0  NVIDIA L40S                    On  |   00000000:4D:00.0 Off |                    0 |
-| N/A   29C    P0             78W /  350W |       0MiB /  46068MiB |     18%      Default |
+| N/A   27C    P8             22W /  350W |       0MiB /  46068MiB |      0%      Default |
 |                                         |                        |                  N/A |
 +-----------------------------------------+------------------------+----------------------+
 
@@ -3926,9 +4153,9 @@ Cell: nv | 0.21s
 <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" style="cursor: default; opacity: 0.3;">▶ uv-logs</span>
+ <span id="uv-indicator-benchmark" onclick="toggleUvLogsFromHeader('benchmark')" style="cursor: pointer;">▶ uv-logs</span>
 </span> | 
-Cell: benchmark | 3.67s
+Cell: benchmark | 32.46s
  | <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>
@@ -3990,29 +4217,29 @@ PROFILE TRACE: torch_eager | cuda_B2_D64_S128_W2
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
                                                    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     430.557us      2231.44%     430.557us     430.557us             1  
-                                            torch_eager        10.26%     219.304us        99.68%       2.131ms       2.131ms       0.000us         0.00%      21.630us      21.630us             1  
-                                               aten::to         0.52%      11.162us        80.31%       1.717ms     286.188us       0.000us         0.00%      14.269us       2.378us             6  
-                                         aten::_to_copy         1.64%      35.099us        79.78%       1.706ms     284.328us       0.000us         0.00%      14.269us       2.378us             6  
-                                            aten::copy_         2.90%      61.950us        75.45%       1.613ms     268.888us      11.934us        61.85%      14.269us       2.378us             6  
-                                           aten::conv1d         0.34%       7.309us         7.15%     152.793us      50.931us       0.000us         0.00%       7.361us       2.454us             3  
-                                      aten::convolution         0.74%      15.802us         6.80%     145.484us      48.495us       0.000us         0.00%       7.361us       2.454us             3  
-                                     aten::_convolution         1.46%      31.300us         6.06%     129.682us      43.227us       0.000us         0.00%       7.361us       2.454us             3  
-                                aten::_conv_depthwise2d         1.58%      33.771us         3.83%      81.842us      27.281us       7.361us        38.15%       7.361us       2.454us             3  
-void at::native::(anonymous namespace)::conv_depthwi...         0.00%       0.000us         0.00%       0.000us       0.000us       7.361us        38.15%       7.361us       2.454us             3  
-void at::native::unrolled_elementwise_kernel&lt;at::nat...         0.00%       0.000us         0.00%       0.000us       0.000us       6.271us        32.50%       6.271us       2.090us             3  
-void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us       5.663us        29.35%       5.663us       1.888us             3  
-                                Activity Buffer Request        69.43%       1.484ms        69.43%       1.484ms       1.484ms       2.335us        12.10%       2.335us       2.335us             1  
-                                    aten::empty_strided         2.69%      57.542us         2.69%      57.542us       9.590us       0.000us         0.00%       0.000us       0.000us             6  
-                                       cudaLaunchKernel         4.29%      91.753us         4.29%      91.753us      10.195us       0.000us         0.00%       0.000us       0.000us             9  
-                                        aten::unsqueeze         1.26%      26.879us         1.60%      34.300us       3.811us       0.000us         0.00%       0.000us       0.000us             9  
-                                       aten::as_strided         0.54%      11.581us         0.54%      11.581us       0.772us       0.000us         0.00%       0.000us       0.000us            15  
-                                            aten::empty         0.53%      11.370us         0.53%      11.370us       3.790us       0.000us         0.00%       0.000us       0.000us             3  
-                                          aten::resize_         0.55%      11.861us         0.55%      11.861us       3.954us       0.000us         0.00%       0.000us       0.000us             3  
-                                          aten::squeeze         0.32%       6.790us         0.37%       8.000us       2.667us       0.000us         0.00%       0.000us       0.000us             3  
+                                            torch_eager         0.00%       0.000us         0.00%       0.000us       0.000us     467.230us      2421.38%     467.230us     467.230us             1  
+                                            torch_eager        10.72%     231.062us        99.69%       2.148ms       2.148ms       0.000us         0.00%      21.632us      21.632us             1  
+                                               aten::to         0.58%      12.480us        78.88%       1.700ms     283.277us       0.000us         0.00%      14.336us       2.389us             6  
+                                         aten::_to_copy         2.05%      44.092us        78.31%       1.687ms     281.197us       0.000us         0.00%      14.336us       2.389us             6  
+                                            aten::copy_         3.07%      66.050us        73.46%       1.583ms     263.783us      12.000us        62.19%      14.336us       2.389us             6  
+                                           aten::conv1d         0.49%      10.600us         7.90%     170.164us      56.721us       0.000us         0.00%       7.296us       2.432us             3  
+                                      aten::convolution         0.77%      16.490us         7.41%     159.564us      53.188us       0.000us         0.00%       7.296us       2.432us             3  
+                                     aten::_convolution         1.64%      35.301us         6.64%     143.074us      47.691us       0.000us         0.00%       7.296us       2.432us             3  
+                                aten::_conv_depthwise2d         1.69%      36.381us         4.00%      86.271us      28.757us       7.296us        37.81%       7.296us       2.432us             3  
+void at::native::(anonymous namespace)::conv_depthwi...         0.00%       0.000us         0.00%       0.000us       0.000us       7.296us        37.81%       7.296us       2.432us             3  
+void at::native::unrolled_elementwise_kernel&lt;at::nat...         0.00%       0.000us         0.00%       0.000us       0.000us       6.304us        32.67%       6.304us       2.101us             3  
+void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us       5.696us        29.52%       5.696us       1.899us             3  
+                                Activity Buffer Request        66.85%       1.440ms        66.85%       1.440ms       1.440ms       2.336us        12.11%       2.336us       2.336us             1  
+                                    aten::empty_strided         2.80%      60.390us         2.80%      60.390us      10.065us       0.000us         0.00%       0.000us       0.000us             6  
+                                       cudaLaunchKernel         4.73%     101.823us         4.73%     101.823us      11.314us       0.000us         0.00%       0.000us       0.000us             9  
+                                        aten::unsqueeze         1.46%      31.451us         1.84%      39.731us       4.415us       0.000us         0.00%       0.000us       0.000us             9  
+                                       aten::as_strided         0.62%      13.289us         0.62%      13.289us       0.886us       0.000us         0.00%       0.000us       0.000us            15  
+                                            aten::empty         0.58%      12.560us         0.58%      12.560us       4.187us       0.000us         0.00%       0.000us       0.000us             3  
+                                          aten::resize_         0.54%      11.740us         0.54%      11.740us       3.913us       0.000us         0.00%       0.000us       0.000us             3  
+                                          aten::squeeze         0.42%       8.963us         0.49%      10.602us       3.534us       0.000us         0.00%       0.000us       0.000us             3  
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
-Self CPU time total: 2.138ms
-Self CUDA time total: 19.295us
+Self CPU time total: 2.155ms
+Self CUDA time total: 19.296us
 
 
 
@@ -4022,29 +4249,29 @@ PROFILE TRACE: torch_eager | cuda_B2_D64_S128_W4
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
                                                    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     335.581us      1716.35%     335.581us     335.581us             1  
-                                            torch_eager         7.79%     148.313us        99.72%       1.898ms       1.898ms       0.000us         0.00%      21.664us      21.664us             1  
-                                               aten::to         0.37%       7.132us        84.25%       1.604ms     267.308us       0.000us         0.00%      13.760us       2.293us             6  
-                                         aten::_to_copy         1.28%      24.280us        83.88%       1.597ms     266.119us       0.000us         0.00%      13.760us       2.293us             6  
-                                            aten::copy_         2.64%      50.321us        81.08%       1.543ms     257.239us      11.648us        59.57%      13.760us       2.293us             6  
-                                           aten::conv1d         0.32%       6.130us         6.26%     119.243us      39.748us       0.000us         0.00%       7.904us       2.635us             3  
-                                      aten::convolution         0.58%      11.131us         5.94%     113.113us      37.704us       0.000us         0.00%       7.904us       2.635us             3  
-                                     aten::_convolution         1.18%      22.459us         5.36%     101.982us      33.994us       0.000us         0.00%       7.904us       2.635us             3  
-                                aten::_conv_depthwise2d         1.08%      20.592us         3.32%      63.132us      21.044us       7.904us        40.43%       7.904us       2.635us             3  
-void at::native::(anonymous namespace)::conv_depthwi...         0.00%       0.000us         0.00%       0.000us       0.000us       7.904us        40.43%       7.904us       2.635us             3  
-void at::native::unrolled_elementwise_kernel&lt;at::nat...         0.00%       0.000us         0.00%       0.000us       0.000us       6.080us        31.10%       6.080us       2.027us             3  
-void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us       5.568us        28.48%       5.568us       1.856us             3  
-                                Activity Buffer Request        75.89%       1.445ms        75.89%       1.445ms       1.445ms       2.112us        10.80%       2.112us       2.112us             1  
-                                    aten::empty_strided         1.52%      29.001us         1.52%      29.001us       4.834us       0.000us         0.00%       0.000us       0.000us             6  
-                                       cudaLaunchKernel         3.69%      70.220us         3.69%      70.220us       7.802us       0.000us         0.00%       0.000us       0.000us             9  
-                                        aten::unsqueeze         0.92%      17.542us         1.22%      23.251us       2.583us       0.000us         0.00%       0.000us       0.000us             9  
-                                       aten::as_strided         0.48%       9.139us         0.48%       9.139us       0.609us       0.000us         0.00%       0.000us       0.000us            15  
-                                            aten::empty         0.58%      11.060us         0.58%      11.060us       3.687us       0.000us         0.00%       0.000us       0.000us             3  
-                                          aten::resize_         0.51%       9.700us         0.51%       9.700us       3.233us       0.000us         0.00%       0.000us       0.000us             3  
-                                          aten::squeeze         0.31%       5.860us         0.37%       7.020us       2.340us       0.000us         0.00%       0.000us       0.000us             3  
+                                            torch_eager         0.00%       0.000us         0.00%       0.000us       0.000us     337.566us      1726.42%     337.566us     337.566us             1  
+                                            torch_eager         6.86%     130.161us        99.69%       1.893ms       1.893ms       0.000us         0.00%      21.665us      21.665us             1  
+                                               aten::to         0.32%       6.060us        85.13%       1.616ms     269.375us       0.000us         0.00%      13.729us       2.288us             6  
+                                         aten::_to_copy         1.27%      24.100us        84.81%       1.610ms     268.365us       0.000us         0.00%      13.729us       2.288us             6  
+                                            aten::copy_         2.69%      51.011us        81.95%       1.556ms     259.305us      11.617us        59.41%      13.729us       2.288us             6  
+                                           aten::conv1d         0.30%       5.740us         6.23%     118.253us      39.418us       0.000us         0.00%       7.936us       2.645us             3  
+                                      aten::convolution         0.52%       9.902us         5.93%     112.513us      37.504us       0.000us         0.00%       7.936us       2.645us             3  
+                                     aten::_convolution         1.21%      22.959us         5.40%     102.611us      34.204us       0.000us         0.00%       7.936us       2.645us             3  
+                                aten::_conv_depthwise2d         1.18%      22.461us         3.33%      63.161us      21.054us       7.936us        40.59%       7.936us       2.645us             3  
+void at::native::(anonymous namespace)::conv_depthwi...         0.00%       0.000us         0.00%       0.000us       0.000us       7.936us        40.59%       7.936us       2.645us             3  
+void at::native::unrolled_elementwise_kernel&lt;at::nat...         0.00%       0.000us         0.00%       0.000us       0.000us       6.080us        31.09%       6.080us       2.027us             3  
+void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us       5.537us        28.32%       5.537us       1.846us             3  
+                                Activity Buffer Request        76.56%       1.454ms        76.56%       1.454ms       1.454ms       2.112us        10.80%       2.112us       2.112us             1  
+                                    aten::empty_strided         1.59%      30.260us         1.59%      30.260us       5.043us       0.000us         0.00%       0.000us       0.000us             6  
+                                       cudaLaunchKernel         3.84%      72.993us         3.84%      72.993us       8.110us       0.000us         0.00%       0.000us       0.000us             9  
+                                        aten::unsqueeze         0.96%      18.220us         1.27%      24.051us       2.672us       0.000us         0.00%       0.000us       0.000us             9  
+                                       aten::as_strided         0.50%       9.451us         0.50%       9.451us       0.630us       0.000us         0.00%       0.000us       0.000us            15  
+                                            aten::empty         0.52%       9.960us         0.52%       9.960us       3.320us       0.000us         0.00%       0.000us       0.000us             3  
+                                          aten::resize_         0.48%       9.030us         0.48%       9.030us       3.010us       0.000us         0.00%       0.000us       0.000us             3  
+                                          aten::squeeze         0.31%       5.890us         0.39%       7.340us       2.447us       0.000us         0.00%       0.000us       0.000us             3  
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
-Self CPU time total: 1.904ms
-Self CUDA time total: 19.552us
+Self CPU time total: 1.899ms
+Self CUDA time total: 19.553us
 
 
 
@@ -4054,29 +4281,29 @@ PROFILE TRACE: torch_eager | cuda_B2_D64_S512_W2
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
                                                    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     343.036us      1844.97%     343.036us     343.036us             1  
-                                            torch_eager         7.68%     146.161us        99.72%       1.897ms       1.897ms       0.000us         0.00%      20.481us      20.481us             1  
-                                               aten::to         0.37%       6.953us        83.90%       1.596ms     266.066us       0.000us         0.00%      13.536us       2.256us             6  
-                                         aten::_to_copy         1.25%      23.842us        83.53%       1.589ms     264.907us       0.000us         0.00%      13.536us       2.256us             6  
-                                            aten::copy_         2.67%      50.789us        80.65%       1.535ms     255.762us      11.648us        62.65%      13.536us       2.256us             6  
-                                           aten::conv1d         0.33%       6.290us         6.66%     126.782us      42.261us       0.000us         0.00%       6.945us       2.315us             3  
-                                      aten::convolution         0.51%       9.650us         6.33%     120.492us      40.164us       0.000us         0.00%       6.945us       2.315us             3  
-                                     aten::_convolution         1.22%      23.120us         5.83%     110.842us      36.947us       0.000us         0.00%       6.945us       2.315us             3  
-                                aten::_conv_depthwise2d         1.44%      27.490us         3.78%      71.832us      23.944us       6.945us        37.35%       6.945us       2.315us             3  
-void at::native::(anonymous namespace)::conv_depthwi...         0.00%       0.000us         0.00%       0.000us       0.000us       6.945us        37.35%       6.945us       2.315us             3  
-void at::native::unrolled_elementwise_kernel&lt;at::nat...         0.00%       0.000us         0.00%       0.000us       0.000us       5.952us        32.01%       5.952us       1.984us             3  
-void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us       5.696us        30.64%       5.696us       1.899us             3  
-                                Activity Buffer Request        75.34%       1.433ms        75.34%       1.433ms       1.433ms       1.888us        10.15%       1.888us       1.888us             1  
-                                    aten::empty_strided         1.63%      31.030us         1.63%      31.030us       5.172us       0.000us         0.00%       0.000us       0.000us             6  
-                                       cudaLaunchKernel         3.99%      75.911us         3.99%      75.911us       8.435us       0.000us         0.00%       0.000us       0.000us             9  
-                                        aten::unsqueeze         0.90%      17.191us         1.18%      22.431us       2.492us       0.000us         0.00%       0.000us       0.000us             9  
-                                       aten::as_strided         0.46%       8.831us         0.46%       8.831us       0.589us       0.000us         0.00%       0.000us       0.000us            15  
-                                            aten::empty         0.55%      10.511us         0.55%      10.511us       3.504us       0.000us         0.00%       0.000us       0.000us             3  
-                                          aten::resize_         0.43%       8.261us         0.43%       8.261us       2.754us       0.000us         0.00%       0.000us       0.000us             3  
-                                          aten::squeeze         0.31%       5.930us         0.38%       7.310us       2.437us       0.000us         0.00%       0.000us       0.000us             3  
+                                            torch_eager         0.00%       0.000us         0.00%       0.000us       0.000us     333.726us      1795.19%     333.726us     333.726us             1  
+                                            torch_eager         6.76%     126.472us        99.71%       1.865ms       1.865ms       0.000us         0.00%      20.510us      20.510us             1  
+                                               aten::to         0.32%       5.970us        85.12%       1.592ms     265.378us       0.000us         0.00%      13.598us       2.266us             6  
+                                         aten::_to_copy         1.26%      23.561us        84.80%       1.586ms     264.383us       0.000us         0.00%      13.598us       2.266us             6  
+                                            aten::copy_         2.75%      51.371us        81.92%       1.532ms     255.399us      11.678us        62.82%      13.598us       2.266us             6  
+                                           aten::conv1d         0.31%       5.850us         6.37%     119.083us      39.694us       0.000us         0.00%       6.912us       2.304us             3  
+                                      aten::convolution         0.54%      10.170us         6.05%     113.233us      37.744us       0.000us         0.00%       6.912us       2.304us             3  
+                                     aten::_convolution         1.25%      23.320us         5.51%     103.063us      34.354us       0.000us         0.00%       6.912us       2.304us             3  
+                                aten::_conv_depthwise2d         1.20%      22.402us         3.41%      63.713us      21.238us       6.912us        37.18%       6.912us       2.304us             3  
+void at::native::(anonymous namespace)::conv_depthwi...         0.00%       0.000us         0.00%       0.000us       0.000us       6.912us        37.18%       6.912us       2.304us             3  
+void at::native::unrolled_elementwise_kernel&lt;at::nat...         0.00%       0.000us         0.00%       0.000us       0.000us       5.951us        32.01%       5.951us       1.984us             3  
+void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us       5.727us        30.81%       5.727us       1.909us             3  
+                                Activity Buffer Request        76.63%       1.433ms        76.63%       1.433ms       1.433ms       1.920us        10.33%       1.920us       1.920us             1  
+                                    aten::empty_strided         1.62%      30.340us         1.62%      30.340us       5.057us       0.000us         0.00%       0.000us       0.000us             6  
+                                       cudaLaunchKernel         3.76%      70.302us         3.76%      70.302us       7.811us       0.000us         0.00%       0.000us       0.000us             9  
+                                        aten::unsqueeze         0.94%      17.590us         1.23%      22.950us       2.550us       0.000us         0.00%       0.000us       0.000us             9  
+                                       aten::as_strided         0.48%       8.970us         0.48%       8.970us       0.598us       0.000us         0.00%       0.000us       0.000us            15  
+                                            aten::empty         0.54%      10.051us         0.54%      10.051us       3.350us       0.000us         0.00%       0.000us       0.000us             3  
+                                          aten::resize_         0.46%       8.519us         0.46%       8.519us       2.840us       0.000us         0.00%       0.000us       0.000us             3  
+                                          aten::squeeze         0.32%       5.980us         0.39%       7.380us       2.460us       0.000us         0.00%       0.000us       0.000us             3  
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
-Self CPU time total: 1.903ms
-Self CUDA time total: 18.593us
+Self CPU time total: 1.871ms
+Self CUDA time total: 18.590us
 
 
 
@@ -4086,29 +4313,29 @@ PROFILE TRACE: torch_eager | cuda_B2_D64_S512_W4
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
                                                    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     329.535us      1679.93%     329.535us     329.535us             1  
-                                            torch_eager         6.84%     137.444us        99.71%       2.005ms       2.005ms       0.000us         0.00%      21.760us      21.760us             1  
-                                               aten::to         0.33%       6.580us        85.73%       1.724ms     287.253us       0.000us         0.00%      14.048us       2.341us             6  
-                                         aten::_to_copy         1.15%      23.069us        85.41%       1.717ms     286.156us       0.000us         0.00%      14.048us       2.341us             6  
-                                            aten::copy_         2.51%      50.362us        82.79%       1.664ms     277.390us      11.904us        60.69%      14.048us       2.341us             6  
-                                           aten::conv1d         0.28%       5.589us         5.76%     115.862us      38.621us       0.000us         0.00%       7.712us       2.571us             3  
-                                      aten::convolution         0.52%      10.381us         5.49%     110.273us      36.758us       0.000us         0.00%       7.712us       2.571us             3  
-                                     aten::_convolution         1.13%      22.651us         4.97%      99.892us      33.297us       0.000us         0.00%       7.712us       2.571us             3  
-                                aten::_conv_depthwise2d         1.01%      20.392us         3.08%      61.981us      20.660us       7.712us        39.31%       7.712us       2.571us             3  
-void at::native::(anonymous namespace)::conv_depthwi...         0.00%       0.000us         0.00%       0.000us       0.000us       7.712us        39.31%       7.712us       2.571us             3  
-void at::native::unrolled_elementwise_kernel&lt;at::nat...         0.00%       0.000us         0.00%       0.000us       0.000us       6.176us        31.48%       6.176us       2.059us             3  
-void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us       5.728us        29.20%       5.728us       1.909us             3  
-                                Activity Buffer Request        70.54%       1.418ms        70.54%       1.418ms       1.418ms       2.144us        10.93%       2.144us       2.144us             1  
-                                    aten::empty_strided         1.47%      29.531us         1.47%      29.531us       4.922us       0.000us         0.00%       0.000us       0.000us             6  
-                                       cudaLaunchKernel        10.85%     218.024us        10.85%     218.024us      24.225us       0.000us         0.00%       0.000us       0.000us             9  
-                                        aten::unsqueeze         0.87%      17.580us         1.13%      22.650us       2.517us       0.000us         0.00%       0.000us       0.000us             9  
-                                       aten::as_strided         0.42%       8.370us         0.42%       8.370us       0.558us       0.000us         0.00%       0.000us       0.000us            15  
-                                            aten::empty         0.48%       9.699us         0.48%       9.699us       3.233us       0.000us         0.00%       0.000us       0.000us             3  
-                                          aten::resize_         0.49%       9.760us         0.49%       9.760us       3.253us       0.000us         0.00%       0.000us       0.000us             3  
-                                          aten::squeeze         0.28%       5.670us         0.34%       6.790us       2.263us       0.000us         0.00%       0.000us       0.000us             3  
+                                            torch_eager         0.00%       0.000us         0.00%       0.000us       0.000us     339.229us      1732.17%     339.229us     339.229us             1  
+                                            torch_eager         6.09%     126.194us        99.75%       2.066ms       2.066ms       0.000us         0.00%      21.729us      21.729us             1  
+                                               aten::to         0.29%       6.100us        86.58%       1.793ms     298.900us       0.000us         0.00%      14.018us       2.336us             6  
+                                         aten::_to_copy         1.16%      23.990us        86.28%       1.787ms     297.883us       0.000us         0.00%      14.018us       2.336us             6  
+                                            aten::copy_         2.58%      53.448us        83.67%       1.733ms     288.850us      11.873us        60.63%      14.018us       2.336us             6  
+                                           aten::conv1d         0.32%       6.580us         5.73%     118.763us      39.588us       0.000us         0.00%       7.711us       2.570us             3  
+                                      aten::convolution         0.48%       9.870us         5.42%     112.183us      37.394us       0.000us         0.00%       7.711us       2.570us             3  
+                                     aten::_convolution         1.09%      22.580us         4.94%     102.313us      34.104us       0.000us         0.00%       7.711us       2.570us             3  
+                                aten::_conv_depthwise2d         1.08%      22.411us         3.09%      64.033us      21.344us       7.711us        39.37%       7.711us       2.570us             3  
+void at::native::(anonymous namespace)::conv_depthwi...         0.00%       0.000us         0.00%       0.000us       0.000us       7.711us        39.37%       7.711us       2.570us             3  
+void at::native::unrolled_elementwise_kernel&lt;at::nat...         0.00%       0.000us         0.00%       0.000us       0.000us       6.145us        31.38%       6.145us       2.048us             3  
+void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us       5.728us        29.25%       5.728us       1.909us             3  
+                                Activity Buffer Request        69.66%       1.443ms        69.66%       1.443ms       1.443ms       2.145us        10.95%       2.145us       2.145us             1  
+                                    aten::empty_strided         1.46%      30.210us         1.46%      30.210us       5.035us       0.000us         0.00%       0.000us       0.000us             6  
+                                       cudaLaunchKernel        12.49%     258.686us        12.49%     258.686us      28.743us       0.000us         0.00%       0.000us       0.000us             9  
+                                        aten::unsqueeze         0.87%      18.050us         1.12%      23.200us       2.578us       0.000us         0.00%       0.000us       0.000us             9  
+                                       aten::as_strided         0.42%       8.720us         0.42%       8.720us       0.581us       0.000us         0.00%       0.000us       0.000us            15  
+                                            aten::empty         0.49%      10.140us         0.49%      10.140us       3.380us       0.000us         0.00%       0.000us       0.000us             3  
+                                          aten::resize_         0.46%       9.442us         0.46%       9.442us       3.147us       0.000us         0.00%       0.000us       0.000us             3  
+                                          aten::squeeze         0.28%       5.830us         0.35%       7.220us       2.407us       0.000us         0.00%       0.000us       0.000us             3  
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
-Self CPU time total: 2.010ms
-Self CUDA time total: 19.616us
+Self CPU time total: 2.071ms
+Self CUDA time total: 19.584us
 
 
 
@@ -4118,29 +4345,29 @@ PROFILE TRACE: torch_eager | cuda_B2_D64_S2048_W2
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
                                                    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     344.351us      1401.22%     344.351us     344.351us             1  
-                                            torch_eager         7.42%     151.244us        99.76%       2.034ms       2.034ms       0.000us         0.00%      26.847us      26.847us             1  
-                                               aten::to         0.33%       6.730us        85.23%       1.738ms     289.583us       0.000us         0.00%      15.264us       2.544us             6  
-                                         aten::_to_copy         1.15%      23.491us        84.90%       1.731ms     288.462us       0.000us         0.00%      15.264us       2.544us             6  
-                                            aten::copy_         2.84%      57.871us        82.24%       1.677ms     279.428us      12.992us        52.87%      15.264us       2.544us             6  
-                                           aten::conv1d         0.31%       6.410us         5.76%     117.443us      39.148us       0.000us         0.00%      11.583us       3.861us             3  
-                                      aten::convolution         0.49%      10.031us         5.45%     111.033us      37.011us       0.000us         0.00%      11.583us       3.861us             3  
-                                     aten::_convolution         1.08%      22.081us         4.95%     101.002us      33.667us       0.000us         0.00%      11.583us       3.861us             3  
-                                aten::_conv_depthwise2d         1.04%      21.239us         3.10%      63.201us      21.067us      11.583us        47.13%      11.583us       3.861us             3  
-void at::native::(anonymous namespace)::conv_depthwi...         0.00%       0.000us         0.00%       0.000us       0.000us      11.583us        47.13%      11.583us       3.861us             3  
-void at::native::unrolled_elementwise_kernel&lt;at::nat...         0.00%       0.000us         0.00%       0.000us       0.000us       6.656us        27.08%       6.656us       2.219us             3  
-void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us       6.336us        25.78%       6.336us       2.112us             3  
-                                Activity Buffer Request        70.08%       1.429ms        70.08%       1.429ms       1.429ms       2.272us         9.25%       2.272us       2.272us             1  
-                                    aten::empty_strided         1.51%      30.710us         1.51%      30.710us       5.118us       0.000us         0.00%       0.000us       0.000us             6  
-                                       cudaLaunchKernel        10.42%     212.467us        10.42%     212.467us      23.607us       0.000us         0.00%       0.000us       0.000us             9  
-                                        aten::unsqueeze         0.89%      18.130us         1.15%      23.350us       2.594us       0.000us         0.00%       0.000us       0.000us             9  
-                                       aten::as_strided         0.43%       8.790us         0.43%       8.790us       0.586us       0.000us         0.00%       0.000us       0.000us            15  
-                                            aten::empty         0.48%       9.800us         0.48%       9.800us       3.267us       0.000us         0.00%       0.000us       0.000us             3  
-                                          aten::resize_         0.47%       9.640us         0.47%       9.640us       3.213us       0.000us         0.00%       0.000us       0.000us             3  
-                                          aten::squeeze         0.28%       5.760us         0.35%       7.050us       2.350us       0.000us         0.00%       0.000us       0.000us             3  
+                                            torch_eager         0.00%       0.000us         0.00%       0.000us       0.000us     342.208us      1399.74%     342.208us     342.208us             1  
+                                            torch_eager         6.21%     125.160us        99.74%       2.012ms       2.012ms       0.000us         0.00%      26.720us      26.720us             1  
+                                               aten::to         0.29%       5.910us        86.35%       1.742ms     290.270us       0.000us         0.00%      15.168us       2.528us             6  
+                                         aten::_to_copy         1.25%      25.122us        86.06%       1.736ms     289.285us       0.000us         0.00%      15.168us       2.528us             6  
+                                            aten::copy_         2.93%      59.190us        83.27%       1.679ms     279.905us      12.896us        52.75%      15.168us       2.528us             6  
+                                           aten::conv1d         0.28%       5.620us         5.81%     117.132us      39.044us       0.000us         0.00%      11.552us       3.851us             3  
+                                      aten::convolution         0.49%       9.910us         5.53%     111.512us      37.171us       0.000us         0.00%      11.552us       3.851us             3  
+                                     aten::_convolution         1.15%      23.280us         5.04%     101.602us      33.867us       0.000us         0.00%      11.552us       3.851us             3  
+                                aten::_conv_depthwise2d         1.09%      21.990us         3.08%      62.201us      20.734us      11.552us        47.25%      11.552us       3.851us             3  
+void at::native::(anonymous namespace)::conv_depthwi...         0.00%       0.000us         0.00%       0.000us       0.000us      11.552us        47.25%      11.552us       3.851us             3  
+void at::native::unrolled_elementwise_kernel&lt;at::nat...         0.00%       0.000us         0.00%       0.000us       0.000us       6.624us        27.09%       6.624us       2.208us             3  
+void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us       6.272us        25.65%       6.272us       2.091us             3  
+                                Activity Buffer Request        71.09%       1.434ms        71.09%       1.434ms       1.434ms       2.272us         9.29%       2.272us       2.272us             1  
+                                    aten::empty_strided         1.55%      31.162us         1.55%      31.162us       5.194us       0.000us         0.00%       0.000us       0.000us             6  
+                                       cudaLaunchKernel        10.29%     207.543us        10.29%     207.543us      23.060us       0.000us         0.00%       0.000us       0.000us             9  
+                                        aten::unsqueeze         0.90%      18.220us         1.17%      23.681us       2.631us       0.000us         0.00%       0.000us       0.000us             9  
+                                       aten::as_strided         0.44%       8.971us         0.44%       8.971us       0.598us       0.000us         0.00%       0.000us       0.000us            15  
+                                            aten::empty         0.49%       9.951us         0.49%       9.951us       3.317us       0.000us         0.00%       0.000us       0.000us             3  
+                                          aten::resize_         0.46%       9.230us         0.46%       9.230us       3.077us       0.000us         0.00%       0.000us       0.000us             3  
+                                          aten::squeeze         0.29%       5.780us         0.35%       7.150us       2.383us       0.000us         0.00%       0.000us       0.000us             3  
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
-Self CPU time total: 2.039ms
-Self CUDA time total: 24.575us
+Self CPU time total: 2.017ms
+Self CUDA time total: 24.448us
 
 
 
@@ -4150,29 +4377,29 @@ PROFILE TRACE: torch_eager | cuda_B2_D64_S2048_W4
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
                                                    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     332.990us      1278.37%     332.990us     332.990us             1  
-                                            torch_eager         7.09%     142.971us        99.76%       2.011ms       2.011ms       0.000us         0.00%      28.288us      28.288us             1  
-                                               aten::to         0.35%       7.062us        85.44%       1.723ms     287.120us       0.000us         0.00%      15.232us       2.539us             6  
-                                         aten::_to_copy         1.18%      23.771us        85.09%       1.716ms     285.943us       0.000us         0.00%      15.232us       2.539us             6  
-                                            aten::copy_         2.51%      50.519us        82.47%       1.663ms     277.136us      12.992us        49.88%      15.232us       2.539us             6  
-                                           aten::conv1d         0.32%       6.541us         5.84%     117.833us      39.278us       0.000us         0.00%      13.056us       4.352us             3  
-                                      aten::convolution         0.52%      10.410us         5.52%     111.292us      37.097us       0.000us         0.00%      13.056us       4.352us             3  
-                                     aten::_convolution         1.19%      24.049us         5.00%     100.882us      33.627us       0.000us         0.00%      13.056us       4.352us             3  
-                                aten::_conv_depthwise2d         1.01%      20.460us         2.98%      60.052us      20.017us      13.056us        50.12%      13.056us       4.352us             3  
-void at::native::(anonymous namespace)::conv_depthwi...         0.00%       0.000us         0.00%       0.000us       0.000us      13.056us        50.12%      13.056us       4.352us             3  
-void at::native::unrolled_elementwise_kernel&lt;at::nat...         0.00%       0.000us         0.00%       0.000us       0.000us       6.624us        25.43%       6.624us       2.208us             3  
-void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us       6.368us        24.45%       6.368us       2.123us             3  
-                                Activity Buffer Request        70.71%       1.426ms        70.71%       1.426ms       1.426ms       2.240us         8.60%       2.240us       2.240us             1  
-                                    aten::empty_strided         1.44%      29.071us         1.44%      29.071us       4.845us       0.000us         0.00%       0.000us       0.000us             6  
-                                       cudaLaunchKernel        10.31%     207.805us        10.31%     207.805us      23.089us       0.000us         0.00%       0.000us       0.000us             9  
-                                        aten::unsqueeze         0.90%      18.081us         1.15%      23.201us       2.578us       0.000us         0.00%       0.000us       0.000us             9  
-                                       aten::as_strided         0.43%       8.650us         0.43%       8.650us       0.577us       0.000us         0.00%       0.000us       0.000us            15  
-                                            aten::empty         0.49%       9.891us         0.49%       9.891us       3.297us       0.000us         0.00%       0.000us       0.000us             3  
-                                          aten::resize_         0.42%       8.561us         0.42%       8.561us       2.854us       0.000us         0.00%       0.000us       0.000us             3  
-                                          aten::squeeze         0.31%       6.350us         0.38%       7.610us       2.537us       0.000us         0.00%       0.000us       0.000us             3  
+                                            torch_eager         0.00%       0.000us         0.00%       0.000us       0.000us     360.702us      1391.60%     360.702us     360.702us             1  
+                                            torch_eager         7.02%     142.940us        99.74%       2.030ms       2.030ms       0.000us         0.00%      28.128us      28.128us             1  
+                                               aten::to         0.30%       6.030us        85.23%       1.734ms     289.050us       0.000us         0.00%      15.136us       2.523us             6  
+                                         aten::_to_copy         1.18%      23.913us        84.93%       1.728ms     288.045us       0.000us         0.00%      15.136us       2.523us             6  
+                                            aten::copy_         2.60%      52.858us        82.24%       1.673ms     278.911us      12.928us        49.88%      15.136us       2.523us             6  
+                                           aten::conv1d         0.29%       5.931us         6.05%     123.062us      41.021us       0.000us         0.00%      12.992us       4.331us             3  
+                                      aten::convolution         0.49%      10.049us         5.76%     117.131us      39.044us       0.000us         0.00%      12.992us       4.331us             3  
+                                     aten::_convolution         1.15%      23.381us         5.26%     107.082us      35.694us       0.000us         0.00%      12.992us       4.331us             3  
+                                aten::_conv_depthwise2d         1.11%      22.652us         3.33%      67.801us      22.600us      12.992us        50.12%      12.992us       4.331us             3  
+void at::native::(anonymous namespace)::conv_depthwi...         0.00%       0.000us         0.00%       0.000us       0.000us      12.992us        50.12%      12.992us       4.331us             3  
+void at::native::unrolled_elementwise_kernel&lt;at::nat...         0.00%       0.000us         0.00%       0.000us       0.000us       6.592us        25.43%       6.592us       2.197us             3  
+void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us       6.336us        24.44%       6.336us       2.112us             3  
+                                Activity Buffer Request        70.88%       1.442ms        70.88%       1.442ms       1.442ms       2.208us         8.52%       2.208us       2.208us             1  
+                                    aten::empty_strided         1.52%      30.891us         1.52%      30.891us       5.148us       0.000us         0.00%       0.000us       0.000us             6  
+                                       cudaLaunchKernel        10.00%     203.394us        10.00%     203.394us      22.599us       0.000us         0.00%       0.000us       0.000us             9  
+                                        aten::unsqueeze         0.92%      18.741us         1.20%      24.361us       2.707us       0.000us         0.00%       0.000us       0.000us             9  
+                                       aten::as_strided         0.46%       9.330us         0.46%       9.330us       0.622us       0.000us         0.00%       0.000us       0.000us            15  
+                                            aten::empty         0.51%      10.450us         0.51%      10.450us       3.483us       0.000us         0.00%       0.000us       0.000us             3  
+                                          aten::resize_         0.47%       9.490us         0.47%       9.490us       3.163us       0.000us         0.00%       0.000us       0.000us             3  
+                                          aten::squeeze         0.29%       5.900us         0.36%       7.380us       2.460us       0.000us         0.00%       0.000us       0.000us             3  
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
-Self CPU time total: 2.016ms
-Self CUDA time total: 26.048us
+Self CPU time total: 2.035ms
+Self CUDA time total: 25.920us
 
 
 
@@ -4182,29 +4409,29 @@ PROFILE TRACE: torch_eager | cuda_B2_D2048_S128_W2
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
                                                    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     332.220us       868.07%     332.220us     332.220us             1  
-                                            torch_eager         7.10%     144.065us        99.76%       2.024ms       2.024ms       0.000us         0.00%      40.831us      40.831us             1  
-                                           aten::conv1d         0.30%       6.030us         5.72%     116.102us      38.701us       0.000us         0.00%      22.464us       7.488us             3  
-                                      aten::convolution         0.49%       9.861us         5.42%     110.072us      36.691us       0.000us         0.00%      22.464us       7.488us             3  
-                                     aten::_convolution         1.11%      22.459us         4.94%     100.211us      33.404us       0.000us         0.00%      22.464us       7.488us             3  
-                                aten::_conv_depthwise2d         1.00%      20.252us         3.07%      62.362us      20.787us      22.464us        58.70%      22.464us       7.488us             3  
-void at::native::(anonymous namespace)::conv_depthwi...         0.00%       0.000us         0.00%       0.000us       0.000us      22.464us        58.70%      22.464us       7.488us             3  
-                                               aten::to         0.31%       6.271us        85.57%       1.737ms     289.428us       0.000us         0.00%      18.367us       3.061us             6  
-                                         aten::_to_copy         1.14%      23.180us        85.26%       1.730ms     288.383us       0.000us         0.00%      18.367us       3.061us             6  
-                                            aten::copy_         2.42%      49.061us        82.56%       1.675ms     279.226us      15.807us        41.30%      18.367us       3.061us             6  
-void at::native::unrolled_elementwise_kernel&lt;at::nat...         0.00%       0.000us         0.00%       0.000us       0.000us       8.448us        22.07%       8.448us       2.816us             3  
-void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us       7.359us        19.23%       7.359us       2.453us             3  
-                                Activity Buffer Request        70.88%       1.438ms        70.88%       1.438ms       1.438ms       2.560us         6.69%       2.560us       2.560us             1  
-                                    aten::empty_strided         1.57%      31.760us         1.57%      31.760us       5.293us       0.000us         0.00%       0.000us       0.000us             6  
-                                       cudaLaunchKernel        10.30%     209.084us        10.30%     209.084us      23.232us       0.000us         0.00%       0.000us       0.000us             9  
-                                        aten::unsqueeze         0.88%      17.889us         1.13%      22.980us       2.553us       0.000us         0.00%       0.000us       0.000us             9  
-                                       aten::as_strided         0.43%       8.691us         0.43%       8.691us       0.579us       0.000us         0.00%       0.000us       0.000us            15  
-                                            aten::empty         0.58%      11.680us         0.58%      11.680us       3.893us       0.000us         0.00%       0.000us       0.000us             3  
-                                          aten::resize_         0.45%       9.220us         0.45%       9.220us       3.073us       0.000us         0.00%       0.000us       0.000us             3  
-                                          aten::squeeze         0.29%       5.850us         0.36%       7.240us       2.413us       0.000us         0.00%       0.000us       0.000us             3  
+                                            torch_eager         0.00%       0.000us         0.00%       0.000us       0.000us     369.628us       962.57%     369.628us     369.628us             1  
+                                            torch_eager         7.12%     161.009us        99.76%       2.255ms       2.255ms       0.000us         0.00%      40.960us      40.960us             1  
+                                           aten::conv1d         0.32%       7.222us         5.82%     131.613us      43.871us       0.000us         0.00%      22.528us       7.509us             3  
+                                      aten::convolution         0.54%      12.229us         5.50%     124.391us      41.464us       0.000us         0.00%      22.528us       7.509us             3  
+                                     aten::_convolution         1.15%      26.031us         4.96%     112.162us      37.387us       0.000us         0.00%      22.528us       7.509us             3  
+                                aten::_conv_depthwise2d         1.09%      24.630us         3.00%      67.820us      22.607us      22.528us        58.67%      22.528us       7.509us             3  
+void at::native::(anonymous namespace)::conv_depthwi...         0.00%       0.000us         0.00%       0.000us       0.000us      22.528us        58.67%      22.528us       7.509us             3  
+                                               aten::to         0.34%       7.671us        85.42%       1.931ms     321.787us       0.000us         0.00%      18.432us       3.072us             6  
+                                         aten::_to_copy         1.41%      31.890us        85.08%       1.923ms     320.509us       0.000us         0.00%      18.432us       3.072us             6  
+                                            aten::copy_         2.64%      59.711us        82.13%       1.856ms     309.384us      15.872us        41.33%      18.432us       3.072us             6  
+void at::native::unrolled_elementwise_kernel&lt;at::nat...         0.00%       0.000us         0.00%       0.000us       0.000us       8.544us        22.25%       8.544us       2.848us             3  
+void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us       7.328us        19.08%       7.328us       2.443us             3  
+                                Activity Buffer Request        64.20%       1.451ms        64.20%       1.451ms       1.451ms       2.560us         6.67%       2.560us       2.560us             1  
+                                    aten::empty_strided         1.54%      34.861us         1.54%      34.861us       5.810us       0.000us         0.00%       0.000us       0.000us             6  
+                                       cudaLaunchKernel        16.32%     368.786us        16.32%     368.786us      40.976us       0.000us         0.00%       0.000us       0.000us             9  
+                                        aten::unsqueeze         0.93%      20.991us         1.15%      26.100us       2.900us       0.000us         0.00%       0.000us       0.000us             9  
+                                       aten::as_strided         0.41%       9.319us         0.41%       9.319us       0.621us       0.000us         0.00%       0.000us       0.000us            15  
+                                            aten::empty         0.44%       9.850us         0.44%       9.850us       3.283us       0.000us         0.00%       0.000us       0.000us             3  
+                                          aten::resize_         0.44%       9.970us         0.44%       9.970us       3.323us       0.000us         0.00%       0.000us       0.000us             3  
+                                          aten::squeeze         0.31%       7.041us         0.38%       8.701us       2.900us       0.000us         0.00%       0.000us       0.000us             3  
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
-Self CPU time total: 2.029ms
-Self CUDA time total: 38.271us
+Self CPU time total: 2.260ms
+Self CUDA time total: 38.400us
 
 
 
@@ -4214,29 +4441,29 @@ PROFILE TRACE: torch_eager | cuda_B2_D2048_S128_W4
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
                                                    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     347.901us       847.38%     347.901us     347.901us             1  
-                                            torch_eager         7.21%     147.111us        99.76%       2.035ms       2.035ms       0.000us         0.00%      43.616us      43.616us             1  
-                                           aten::conv1d         0.33%       6.680us         5.94%     121.133us      40.378us       0.000us         0.00%      25.376us       8.459us             3  
-                                      aten::convolution         0.49%      10.011us         5.61%     114.453us      38.151us       0.000us         0.00%      25.376us       8.459us             3  
-                                     aten::_convolution         1.21%      24.739us         5.12%     104.442us      34.814us       0.000us         0.00%      25.376us       8.459us             3  
-                                aten::_conv_depthwise2d         1.05%      21.431us         3.09%      62.981us      20.994us      25.376us        61.81%      25.376us       8.459us             3  
-void at::native::(anonymous namespace)::conv_depthwi...         0.00%       0.000us         0.00%       0.000us       0.000us      25.376us        61.81%      25.376us       8.459us             3  
-                                               aten::to         0.35%       7.210us        85.28%       1.740ms     289.922us       0.000us         0.00%      18.240us       3.040us             6  
-                                         aten::_to_copy         1.21%      24.639us        84.93%       1.732ms     288.720us       0.000us         0.00%      18.240us       3.040us             6  
-                                            aten::copy_         2.56%      52.303us        82.23%       1.677ms     279.542us      15.680us        38.19%      18.240us       3.040us             6  
-void at::native::unrolled_elementwise_kernel&lt;at::nat...         0.00%       0.000us         0.00%       0.000us       0.000us       8.288us        20.19%       8.288us       2.763us             3  
-void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us       7.392us        18.00%       7.392us       2.464us             3  
-                                Activity Buffer Request        70.32%       1.434ms        70.32%       1.434ms       1.434ms       2.560us         6.24%       2.560us       2.560us             1  
-                                    aten::empty_strided         1.49%      30.432us         1.49%      30.432us       5.072us       0.000us         0.00%       0.000us       0.000us             6  
-                                       cudaLaunchKernel        10.49%     213.884us        10.49%     213.884us      23.765us       0.000us         0.00%       0.000us       0.000us             9  
-                                        aten::unsqueeze         0.88%      17.871us         1.14%      23.242us       2.582us       0.000us         0.00%       0.000us       0.000us             9  
-                                       aten::as_strided         0.44%       8.942us         0.44%       8.942us       0.596us       0.000us         0.00%       0.000us       0.000us            15  
-                                            aten::empty         0.48%       9.840us         0.48%       9.840us       3.280us       0.000us         0.00%       0.000us       0.000us             3  
-                                          aten::resize_         0.42%       8.540us         0.42%       8.540us       2.847us       0.000us         0.00%       0.000us       0.000us             3  
-                                          aten::squeeze         0.31%       6.230us         0.37%       7.540us       2.513us       0.000us         0.00%       0.000us       0.000us             3  
+                                            torch_eager         0.00%       0.000us         0.00%       0.000us       0.000us     343.007us       838.09%     343.007us     343.007us             1  
+                                            torch_eager         6.47%     141.163us        99.73%       2.175ms       2.175ms       0.000us         0.00%      43.487us      43.487us             1  
+                                           aten::conv1d         0.27%       5.870us         5.52%     120.313us      40.104us       0.000us         0.00%      25.376us       8.459us             3  
+                                      aten::convolution         0.46%      10.120us         5.25%     114.443us      38.148us       0.000us         0.00%      25.376us       8.459us             3  
+                                     aten::_convolution         1.12%      24.490us         4.78%     104.323us      34.774us       0.000us         0.00%      25.376us       8.459us             3  
+                                aten::_conv_depthwise2d         1.00%      21.702us         2.89%      62.963us      20.988us      25.376us        62.00%      25.376us       8.459us             3  
+void at::native::(anonymous namespace)::conv_depthwi...         0.00%       0.000us         0.00%       0.000us       0.000us      25.376us        62.00%      25.376us       8.459us             3  
+                                               aten::to         0.28%       6.129us        86.46%       1.885ms     314.232us       0.000us         0.00%      18.111us       3.018us             6  
+                                         aten::_to_copy         1.13%      24.640us        86.18%       1.879ms     313.211us       0.000us         0.00%      18.111us       3.018us             6  
+                                            aten::copy_         2.51%      54.672us        83.58%       1.823ms     303.754us      15.551us        38.00%      18.111us       3.018us             6  
+void at::native::unrolled_elementwise_kernel&lt;at::nat...         0.00%       0.000us         0.00%       0.000us       0.000us       8.224us        20.09%       8.224us       2.741us             3  
+void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us       7.327us        17.90%       7.327us       2.442us             3  
+                                Activity Buffer Request        66.59%       1.452ms        66.59%       1.452ms       1.452ms       2.560us         6.26%       2.560us       2.560us             1  
+                                    aten::empty_strided         1.47%      32.100us         1.47%      32.100us       5.350us       0.000us         0.00%       0.000us       0.000us             6  
+                                       cudaLaunchKernel        15.50%     338.007us        15.50%     338.007us      37.556us       0.000us         0.00%       0.000us       0.000us             9  
+                                        aten::unsqueeze         0.84%      18.320us         1.10%      24.070us       2.674us       0.000us         0.00%       0.000us       0.000us             9  
+                                       aten::as_strided         0.43%       9.420us         0.43%       9.420us       0.628us       0.000us         0.00%       0.000us       0.000us            15  
+                                            aten::empty         0.46%      10.080us         0.46%      10.080us       3.360us       0.000us         0.00%       0.000us       0.000us             3  
+                                          aten::resize_         0.42%       9.080us         0.42%       9.080us       3.027us       0.000us         0.00%       0.000us       0.000us             3  
+                                          aten::squeeze         0.27%       5.960us         0.34%       7.390us       2.463us       0.000us         0.00%       0.000us       0.000us             3  
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
-Self CPU time total: 2.040ms
-Self CUDA time total: 41.056us
+Self CPU time total: 2.181ms
+Self CUDA time total: 40.927us
 
 
 
@@ -4246,29 +4473,29 @@ PROFILE TRACE: torch_eager | cuda_B2_D2048_S512_W2
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
                                                    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     334.331us       325.39%     334.331us     334.331us             1  
-                                            torch_eager         7.01%     141.713us        99.76%       2.018ms       2.018ms       0.000us         0.00%     108.764us     108.764us             1  
-                                           aten::conv1d         0.30%       6.090us         5.77%     116.623us      38.874us       0.000us         0.00%      70.528us      23.509us             3  
-                                      aten::convolution         0.56%      11.281us         5.46%     110.533us      36.844us       0.000us         0.00%      70.528us      23.509us             3  
-                                     aten::_convolution         1.11%      22.501us         4.91%      99.252us      33.084us       0.000us         0.00%      70.528us      23.509us             3  
-                                aten::_conv_depthwise2d         1.02%      20.538us         3.03%      61.301us      20.434us      70.528us        68.64%      70.528us      23.509us             3  
-void at::native::(anonymous namespace)::conv_depthwi...         0.00%       0.000us         0.00%       0.000us       0.000us      70.528us        68.64%      70.528us      23.509us             3  
-                                               aten::to         0.31%       6.229us        85.56%       1.731ms     288.457us       0.000us         0.00%      38.236us       6.373us             6  
-                                         aten::_to_copy         1.17%      23.650us        85.25%       1.725ms     287.419us       0.000us         0.00%      38.236us       6.373us             6  
-                                            aten::copy_         2.48%      50.230us        82.63%       1.672ms     278.605us      32.221us        31.36%      38.236us       6.373us             6  
-void at::native::unrolled_elementwise_kernel&lt;at::nat...         0.00%       0.000us         0.00%       0.000us       0.000us      17.598us        17.13%      17.598us       5.866us             3  
-void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      14.623us        14.23%      14.623us       4.874us             3  
-                                Activity Buffer Request        70.91%       1.435ms        70.91%       1.435ms       1.435ms       6.015us         5.85%       6.015us       6.015us             1  
-                                    aten::empty_strided         1.45%      29.232us         1.45%      29.232us       4.872us       0.000us         0.00%       0.000us       0.000us             6  
-                                       cudaLaunchKernel        10.36%     209.517us        10.36%     209.517us      23.280us       0.000us         0.00%       0.000us       0.000us             9  
-                                        aten::unsqueeze         0.88%      17.770us         1.13%      22.940us       2.549us       0.000us         0.00%       0.000us       0.000us             9  
-                                       aten::as_strided         0.42%       8.560us         0.42%       8.560us       0.571us       0.000us         0.00%       0.000us       0.000us            15  
-                                            aten::empty         0.48%       9.771us         0.48%       9.771us       3.257us       0.000us         0.00%       0.000us       0.000us             3  
-                                          aten::resize_         0.41%       8.351us         0.41%       8.351us       2.784us       0.000us         0.00%       0.000us       0.000us             3  
-                                          aten::squeeze         0.28%       5.590us         0.33%       6.770us       2.257us       0.000us         0.00%       0.000us       0.000us             3  
+                                            torch_eager         0.00%       0.000us         0.00%       0.000us       0.000us     367.004us       357.73%     367.004us     367.004us             1  
+                                            torch_eager         6.17%     126.763us        99.73%       2.049ms       2.049ms       0.000us         0.00%     108.512us     108.512us             1  
+                                           aten::conv1d         0.28%       5.761us         5.81%     119.372us      39.791us       0.000us         0.00%      70.432us      23.477us             3  
+                                      aten::convolution         0.48%       9.820us         5.53%     113.611us      37.870us       0.000us         0.00%      70.432us      23.477us             3  
+                                     aten::_convolution         1.11%      22.788us         5.05%     103.791us      34.597us       0.000us         0.00%      70.432us      23.477us             3  
+                                aten::_conv_depthwise2d         1.12%      22.910us         3.14%      64.601us      21.534us      70.432us        68.65%      70.432us      23.477us             3  
+void at::native::(anonymous namespace)::conv_depthwi...         0.00%       0.000us         0.00%       0.000us       0.000us      70.432us        68.65%      70.432us      23.477us             3  
+                                               aten::to         0.30%       6.130us        86.37%       1.774ms     295.680us       0.000us         0.00%      38.080us       6.347us             6  
+                                         aten::_to_copy         2.18%      44.819us        86.07%       1.768ms     294.658us       0.000us         0.00%      38.080us       6.347us             6  
+                                            aten::copy_         2.56%      52.622us        82.32%       1.691ms     281.815us      32.160us        31.35%      38.080us       6.347us             6  
+void at::native::unrolled_elementwise_kernel&lt;at::nat...         0.00%       0.000us         0.00%       0.000us       0.000us      17.504us        17.06%      17.504us       5.835us             3  
+void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      14.656us        14.29%      14.656us       4.885us             3  
+                                Activity Buffer Request        69.77%       1.433ms        69.77%       1.433ms       1.433ms       5.920us         5.77%       5.920us       5.920us             1  
+                                    aten::empty_strided         1.57%      32.241us         1.57%      32.241us       5.373us       0.000us         0.00%       0.000us       0.000us             6  
+                                       cudaLaunchKernel        11.08%     227.645us        11.08%     227.645us      25.294us       0.000us         0.00%       0.000us       0.000us             9  
+                                        aten::unsqueeze         0.87%      17.849us         1.12%      23.070us       2.563us       0.000us         0.00%       0.000us       0.000us             9  
+                                       aten::as_strided         0.44%       9.030us         0.44%       9.030us       0.602us       0.000us         0.00%       0.000us       0.000us            15  
+                                            aten::empty         0.49%      10.050us         0.49%      10.050us       3.350us       0.000us         0.00%       0.000us       0.000us             3  
+                                          aten::resize_         0.44%       9.040us         0.44%       9.040us       3.013us       0.000us         0.00%       0.000us       0.000us             3  
+                                          aten::squeeze         0.30%       6.163us         0.38%       7.782us       2.594us       0.000us         0.00%       0.000us       0.000us             3  
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
-Self CPU time total: 2.023ms
-Self CUDA time total: 102.749us
+Self CPU time total: 2.054ms
+Self CUDA time total: 102.592us
 
 
 
@@ -4278,29 +4505,29 @@ PROFILE TRACE: torch_eager | cuda_B2_D2048_S512_W4
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
                                                    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     330.969us       293.42%     330.969us     330.969us             1  
-                                            torch_eager        14.99%     119.634us        99.39%     793.059us     793.059us       0.000us         0.00%     118.814us     118.814us             1  
-                                           aten::conv1d         0.68%       5.459us        14.66%     116.982us      38.994us       0.000us         0.00%      80.510us      26.837us             3  
-                                      aten::convolution         1.26%      10.041us        13.98%     111.523us      37.174us       0.000us         0.00%      80.510us      26.837us             3  
-                                     aten::_convolution         2.84%      22.661us        12.72%     101.482us      33.827us       0.000us         0.00%      80.510us      26.837us             3  
-                                aten::_conv_depthwise2d         2.60%      20.719us         7.95%      63.401us      21.134us      80.510us        71.38%      80.510us      26.837us             3  
-void at::native::(anonymous namespace)::conv_depthwi...         0.00%       0.000us         0.00%       0.000us       0.000us      80.510us        71.38%      80.510us      26.837us             3  
-                                               aten::to         0.74%       5.920us        66.51%     530.742us      88.457us       0.000us         0.00%      38.304us       6.384us             6  
-                                         aten::_to_copy         2.94%      23.422us        65.77%     524.822us      87.470us       0.000us         0.00%      38.304us       6.384us             6  
-                                            aten::copy_         6.43%      51.340us        58.99%     470.681us      78.447us      32.288us        28.62%      38.304us       6.384us             6  
-void at::native::unrolled_elementwise_kernel&lt;at::nat...         0.00%       0.000us         0.00%       0.000us       0.000us      17.696us        15.69%      17.696us       5.899us             3  
-void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      14.592us        12.94%      14.592us       4.864us             3  
-                                Activity Buffer Request        29.02%     231.576us        29.02%     231.576us     231.576us       6.016us         5.33%       6.016us       6.016us             1  
-                                    aten::empty_strided         3.85%      30.719us         3.85%      30.719us       5.120us       0.000us         0.00%       0.000us       0.000us             6  
-                                       cudaLaunchKernel        26.56%     211.935us        26.56%     211.935us      23.548us       0.000us         0.00%       0.000us       0.000us             9  
-                                        aten::unsqueeze         2.12%      16.940us         2.72%      21.720us       2.413us       0.000us         0.00%       0.000us       0.000us             9  
-                                       aten::as_strided         1.02%       8.121us         1.02%       8.121us       0.541us       0.000us         0.00%       0.000us       0.000us            15  
-                                            aten::empty         1.20%       9.582us         1.20%       9.582us       3.194us       0.000us         0.00%       0.000us       0.000us             3  
-                                          aten::resize_         1.12%       8.930us         1.12%       8.930us       2.977us       0.000us         0.00%       0.000us       0.000us             3  
-                                          aten::squeeze         0.72%       5.780us         0.87%       6.970us       2.323us       0.000us         0.00%       0.000us       0.000us             3  
+                                            torch_eager         0.00%       0.000us         0.00%       0.000us       0.000us     336.959us       299.49%     336.959us     336.959us             1  
+                                            torch_eager         6.25%     125.522us        99.75%       2.004ms       2.004ms       0.000us         0.00%     118.493us     118.493us             1  
+                                           aten::conv1d         0.38%       7.700us         5.98%     120.223us      40.074us       0.000us         0.00%      80.479us      26.826us             3  
+                                      aten::convolution         0.49%       9.780us         5.60%     112.523us      37.508us       0.000us         0.00%      80.479us      26.826us             3  
+                                     aten::_convolution         1.13%      22.669us         5.11%     102.743us      34.248us       0.000us         0.00%      80.479us      26.826us             3  
+                                aten::_conv_depthwise2d         1.12%      22.452us         3.19%      64.073us      21.358us      80.479us        71.53%      80.479us      26.826us             3  
+void at::native::(anonymous namespace)::conv_depthwi...         0.00%       0.000us         0.00%       0.000us       0.000us      80.479us        71.53%      80.479us      26.826us             3  
+                                               aten::to         0.29%       5.910us        86.14%       1.731ms     288.442us       0.000us         0.00%      38.014us       6.336us             6  
+                                         aten::_to_copy         1.19%      24.001us        85.85%       1.725ms     287.457us       0.000us         0.00%      38.014us       6.336us             6  
+                                            aten::copy_         2.56%      51.481us        83.17%       1.671ms     278.473us      32.031us        28.47%      38.014us       6.336us             6  
+void at::native::unrolled_elementwise_kernel&lt;at::nat...         0.00%       0.000us         0.00%       0.000us       0.000us      17.567us        15.61%      17.567us       5.856us             3  
+void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      14.464us        12.86%      14.464us       4.821us             3  
+                                Activity Buffer Request        71.72%       1.441ms        71.72%       1.441ms       1.441ms       5.983us         5.32%       5.983us       5.983us             1  
+                                    aten::empty_strided         1.49%      29.901us         1.49%      29.901us       4.983us       0.000us         0.00%       0.000us       0.000us             6  
+                                       cudaLaunchKernel        10.00%     200.814us        10.00%     200.814us      22.313us       0.000us         0.00%       0.000us       0.000us             9  
+                                        aten::unsqueeze         0.89%      17.861us         1.15%      23.111us       2.568us       0.000us         0.00%       0.000us       0.000us             9  
+                                       aten::as_strided         0.45%       8.970us         0.45%       8.970us       0.598us       0.000us         0.00%       0.000us       0.000us            15  
+                                            aten::empty         0.50%      10.050us         0.50%      10.050us       3.350us       0.000us         0.00%       0.000us       0.000us             3  
+                                          aten::resize_         0.46%       9.169us         0.46%       9.169us       3.056us       0.000us         0.00%       0.000us       0.000us             3  
+                                          aten::squeeze         0.30%       6.030us         0.38%       7.560us       2.520us       0.000us         0.00%       0.000us       0.000us             3  
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
-Self CPU time total: 797.960us
-Self CUDA time total: 112.798us
+Self CPU time total: 2.009ms
+Self CUDA time total: 112.510us
 
 
 
@@ -4310,29 +4537,29 @@ PROFILE TRACE: torch_eager | cuda_B2_D2048_S2048_W2
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
                                                    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        13.79%     117.896us        93.85%     802.069us     802.069us       0.000us         0.00%     432.858us     432.858us             1  
-                                            torch_eager         0.00%       0.000us         0.00%       0.000us       0.000us     419.356us       106.55%     419.356us     419.356us             1  
-                                           aten::conv1d         0.66%       5.648us        13.24%     113.161us      37.720us       0.000us         0.00%     251.262us      83.754us             3  
-                                      aten::convolution         1.11%       9.481us        12.58%     107.513us      35.838us       0.000us         0.00%     251.262us      83.754us             3  
-                                     aten::_convolution         2.53%      21.627us        11.47%      98.032us      32.677us       0.000us         0.00%     251.262us      83.754us             3  
-                                aten::_conv_depthwise2d         2.35%      20.121us         7.14%      61.002us      20.334us     251.262us        63.84%     251.262us      83.754us             3  
-void at::native::(anonymous namespace)::conv_depthwi...         0.00%       0.000us         0.00%       0.000us       0.000us     251.262us        63.84%     251.262us      83.754us             3  
-                                               aten::to         0.66%       5.670us        63.66%     544.101us      90.683us       0.000us         0.00%     181.596us      30.266us             6  
-                                         aten::_to_copy         2.68%      22.880us        63.00%     538.431us      89.739us       0.000us         0.00%     181.596us      30.266us             6  
-                                            aten::copy_         6.04%      51.591us        56.88%     486.161us      81.027us     142.333us        36.16%     181.596us      30.266us             6  
-void at::native::unrolled_elementwise_kernel&lt;at::nat...         0.00%       0.000us         0.00%       0.000us       0.000us     102.271us        25.98%     102.271us      34.090us             3  
-void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      40.062us        10.18%      40.062us      13.354us             3  
-                                Activity Buffer Request        28.73%     245.556us        28.73%     245.556us     245.556us      39.263us         9.98%      39.263us      39.263us             1  
-                                    aten::empty_strided         3.44%      29.390us         3.44%      29.390us       4.898us       0.000us         0.00%       0.000us       0.000us             6  
-                                       cudaLaunchKernel        24.77%     211.714us        24.77%     211.714us      23.524us       0.000us         0.00%       0.000us       0.000us             9  
-                                        aten::unsqueeze         1.99%      17.042us         2.56%      21.904us       2.434us       0.000us         0.00%       0.000us       0.000us             9  
-                                       aten::as_strided         1.01%       8.621us         1.01%       8.621us       0.575us       0.000us         0.00%       0.000us       0.000us            15  
-                                            aten::empty         1.11%       9.471us         1.11%       9.471us       3.157us       0.000us         0.00%       0.000us       0.000us             3  
-                                          aten::resize_         1.02%       8.710us         1.02%       8.710us       2.903us       0.000us         0.00%       0.000us       0.000us             3  
-                                          aten::squeeze         0.71%       6.031us         0.84%       7.190us       2.397us       0.000us         0.00%       0.000us       0.000us             3  
+                                            torch_eager         5.98%     122.945us        97.76%       2.011ms       2.011ms       0.000us         0.00%     433.437us     433.437us             1  
+                                            torch_eager         0.00%       0.000us         0.00%       0.000us       0.000us     423.709us       107.83%     423.709us     423.709us             1  
+                                           aten::conv1d         0.28%       5.760us         5.73%     117.851us      39.284us       0.000us         0.00%     250.941us      83.647us             3  
+                                      aten::convolution         0.48%       9.830us         5.45%     112.091us      37.364us       0.000us         0.00%     250.941us      83.647us             3  
+                                     aten::_convolution         1.12%      23.111us         4.97%     102.261us      34.087us       0.000us         0.00%     250.941us      83.647us             3  
+                                aten::_conv_depthwise2d         1.03%      21.200us         3.03%      62.360us      20.787us     250.941us        63.86%     250.941us      83.647us             3  
+void at::native::(anonymous namespace)::conv_depthwi...         0.00%       0.000us         0.00%       0.000us       0.000us     250.941us        63.86%     250.941us      83.647us             3  
+                                               aten::to         0.28%       5.851us        84.70%       1.742ms     290.313us       0.000us         0.00%     182.496us      30.416us             6  
+                                         aten::_to_copy         1.16%      23.919us        84.41%       1.736ms     289.338us       0.000us         0.00%     182.496us      30.416us             6  
+                                            aten::copy_         2.53%      51.981us        81.78%       1.682ms     280.333us     142.016us        36.14%     182.496us      30.416us             6  
+void at::native::unrolled_elementwise_kernel&lt;at::nat...         0.00%       0.000us         0.00%       0.000us       0.000us     101.952us        25.94%     101.952us      33.984us             3  
+                                Activity Buffer Request        70.64%       1.453ms        70.64%       1.453ms       1.453ms      40.480us        10.30%      40.480us      40.480us             1  
+void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      40.064us        10.20%      40.064us      13.355us             3  
+                                    aten::empty_strided         1.46%      30.112us         1.46%      30.112us       5.019us       0.000us         0.00%       0.000us       0.000us             6  
+                                       cudaLaunchKernel         9.67%     198.853us         9.67%     198.853us      22.095us       0.000us         0.00%       0.000us       0.000us             9  
+                                        aten::unsqueeze         0.91%      18.669us         1.18%      24.270us       2.697us       0.000us         0.00%       0.000us       0.000us             9  
+                                       aten::as_strided         0.44%       9.151us         0.44%       9.151us       0.610us       0.000us         0.00%       0.000us       0.000us            15  
+                                            aten::empty         0.48%       9.870us         0.48%       9.870us       3.290us       0.000us         0.00%       0.000us       0.000us             3  
+                                          aten::resize_         0.47%       9.710us         0.47%       9.710us       3.237us       0.000us         0.00%       0.000us       0.000us             3  
+                                          aten::squeeze         0.29%       5.960us         0.36%       7.350us       2.450us       0.000us         0.00%       0.000us       0.000us             3  
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
-Self CPU time total: 854.650us
-Self CUDA time total: 393.595us
+Self CPU time total: 2.057ms
+Self CUDA time total: 392.957us
 
 
 
@@ -4342,29 +4569,29 @@ PROFILE TRACE: torch_eager | cuda_B2_D2048_S2048_W4
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
                                                    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        12.73%     119.312us        88.90%     833.220us     833.220us       0.000us         0.00%     487.606us     487.606us             1  
-                                            torch_eager         0.00%       0.000us         0.00%       0.000us       0.000us     476.503us       106.43%     476.503us     476.503us             1  
-                                           aten::conv1d         0.59%       5.550us        12.38%     116.073us      38.691us       0.000us         0.00%     298.682us      99.561us             3  
-                                      aten::convolution         1.01%       9.430us        11.79%     110.523us      36.841us       0.000us         0.00%     298.682us      99.561us             3  
-                                     aten::_convolution         2.32%      21.781us        10.79%     101.093us      33.698us       0.000us         0.00%     298.682us      99.561us             3  
-                                aten::_conv_depthwise2d         2.19%      20.491us         6.88%      64.493us      21.498us     298.682us        66.71%     298.682us      99.561us             3  
-void at::native::(anonymous namespace)::conv_depthwi...         0.00%       0.000us         0.00%       0.000us       0.000us     298.682us        66.71%     298.682us      99.561us             3  
-                                               aten::to         0.60%       5.580us        60.86%     570.404us      95.067us       0.000us         0.00%     188.924us      31.487us             6  
-                                         aten::_to_copy         2.42%      22.662us        60.26%     564.824us      94.137us       0.000us         0.00%     188.924us      31.487us             6  
-                                            aten::copy_         5.33%      49.982us        54.62%     511.981us      85.330us     149.053us        33.29%     188.924us      31.487us             6  
-void at::native::unrolled_elementwise_kernel&lt;at::nat...         0.00%       0.000us         0.00%       0.000us       0.000us     108.926us        24.33%     108.926us      36.309us             3  
-void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      40.127us         8.96%      40.127us      13.376us             3  
-                                Activity Buffer Request        29.44%     275.977us        29.44%     275.977us     275.977us      39.871us         8.91%      39.871us      39.871us             1  
-                                    aten::empty_strided         3.22%      30.181us         3.22%      30.181us       5.030us       0.000us         0.00%       0.000us       0.000us             6  
-                                       cudaLaunchKernel        22.44%     210.343us        22.44%     210.343us      23.371us       0.000us         0.00%       0.000us       0.000us             9  
-                                        aten::unsqueeze         1.80%      16.910us         2.35%      22.009us       2.445us       0.000us         0.00%       0.000us       0.000us             9  
-                                       aten::as_strided         0.91%       8.519us         0.91%       8.519us       0.568us       0.000us         0.00%       0.000us       0.000us            15  
-                                            aten::empty         1.16%      10.891us         1.16%      10.891us       3.630us       0.000us         0.00%       0.000us       0.000us             3  
-                                          aten::resize_         0.94%       8.790us         0.94%       8.790us       2.930us       0.000us         0.00%       0.000us       0.000us             3  
-                                          aten::squeeze         0.59%       5.500us         0.71%       6.690us       2.230us       0.000us         0.00%       0.000us       0.000us             3  
+                                            torch_eager         5.86%     122.119us        95.18%       1.984ms       1.984ms       0.000us         0.00%     485.373us     485.373us             1  
+                                            torch_eager         0.00%       0.000us         0.00%       0.000us       0.000us     475.549us       106.61%     475.549us     475.549us             1  
+                                           aten::conv1d         0.29%       6.020us         5.58%     116.291us      38.764us       0.000us         0.00%     298.429us      99.476us             3  
+                                      aten::convolution         0.46%       9.580us         5.29%     110.271us      36.757us       0.000us         0.00%     298.429us      99.476us             3  
+                                     aten::_convolution         1.07%      22.391us         4.83%     100.691us      33.564us       0.000us         0.00%     298.429us      99.476us             3  
+                                aten::_conv_depthwise2d         1.02%      21.160us         3.01%      62.730us      20.910us     298.429us        66.91%     298.429us      99.476us             3  
+void at::native::(anonymous namespace)::conv_depthwi...         0.00%       0.000us         0.00%       0.000us       0.000us     298.429us        66.91%     298.429us      99.476us             3  
+                                               aten::to         0.28%       5.929us        82.40%       1.718ms     286.300us       0.000us         0.00%     186.944us      31.157us             6  
+                                         aten::_to_copy         1.13%      23.472us        82.12%       1.712ms     285.312us       0.000us         0.00%     186.944us      31.157us             6  
+                                            aten::copy_         2.45%      51.061us        79.57%       1.659ms     276.443us     147.616us        33.09%     186.944us      31.157us             6  
+void at::native::unrolled_elementwise_kernel&lt;at::nat...         0.00%       0.000us         0.00%       0.000us       0.000us     107.712us        24.15%     107.712us      35.904us             3  
+void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      39.904us         8.95%      39.904us      13.301us             3  
+                                Activity Buffer Request        68.65%       1.431ms        68.65%       1.431ms       1.431ms      39.328us         8.82%      39.328us      39.328us             1  
+                                    aten::empty_strided         1.43%      29.742us         1.43%      29.742us       4.957us       0.000us         0.00%       0.000us       0.000us             6  
+                                       cudaLaunchKernel         9.54%     198.903us         9.54%     198.903us      22.100us       0.000us         0.00%       0.000us       0.000us             9  
+                                        aten::unsqueeze         0.85%      17.731us         1.11%      23.210us       2.579us       0.000us         0.00%       0.000us       0.000us             9  
+                                       aten::as_strided         0.44%       9.210us         0.44%       9.210us       0.614us       0.000us         0.00%       0.000us       0.000us            15  
+                                            aten::empty         0.47%       9.850us         0.47%       9.850us       3.283us       0.000us         0.00%       0.000us       0.000us             3  
+                                          aten::resize_         0.45%       9.320us         0.45%       9.320us       3.107us       0.000us         0.00%       0.000us       0.000us             3  
+                                          aten::squeeze         0.28%       5.850us         0.35%       7.270us       2.423us       0.000us         0.00%       0.000us       0.000us             3  
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
-Self CPU time total: 937.282us
-Self CUDA time total: 447.735us
+Self CPU time total: 2.085ms
+Self CUDA time total: 446.045us
 
 
 
@@ -4374,29 +4601,29 @@ PROFILE TRACE: torch_eager | cuda_B4_D64_S128_W2
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
                                                    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     323.326us      1733.09%     323.326us     323.326us             1  
-                                            torch_eager        14.06%     116.944us        99.42%     826.859us     826.859us       0.000us         0.00%      20.544us      20.544us             1  
-                                               aten::to         0.72%       5.971us        68.57%     570.283us      95.047us       0.000us         0.00%      13.344us       2.224us             6  
-                                         aten::_to_copy         2.68%      22.330us        67.85%     564.312us      94.052us       0.000us         0.00%      13.344us       2.224us             6  
-                                            aten::copy_         6.25%      51.969us        61.73%     513.371us      85.562us      11.456us        61.41%      13.344us       2.224us             6  
-                                           aten::conv1d         0.66%       5.530us        13.54%     112.622us      37.541us       0.000us         0.00%       7.200us       2.400us             3  
-                                      aten::convolution         1.25%      10.420us        12.88%     107.092us      35.697us       0.000us         0.00%       7.200us       2.400us             3  
-                                     aten::_convolution         2.52%      20.950us        11.62%      96.672us      32.224us       0.000us         0.00%       7.200us       2.400us             3  
-                                aten::_conv_depthwise2d         2.43%      20.241us         7.30%      60.692us      20.231us       7.200us        38.59%       7.200us       2.400us             3  
-void at::native::(anonymous namespace)::conv_depthwi...         0.00%       0.000us         0.00%       0.000us       0.000us       7.200us        38.59%       7.200us       2.400us             3  
-void at::native::unrolled_elementwise_kernel&lt;at::nat...         0.00%       0.000us         0.00%       0.000us       0.000us       5.888us        31.56%       5.888us       1.963us             3  
-void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us       5.568us        29.85%       5.568us       1.856us             3  
-                                Activity Buffer Request        31.20%     259.516us        31.20%     259.516us     259.516us       1.888us        10.12%       1.888us       1.888us             1  
-                                    aten::empty_strided         3.44%      28.611us         3.44%      28.611us       4.768us       0.000us         0.00%       0.000us       0.000us             6  
-                                       cudaLaunchKernel        26.78%     222.677us        26.78%     222.677us      24.742us       0.000us         0.00%       0.000us       0.000us             9  
-                                        aten::unsqueeze         2.11%      17.509us         2.72%      22.620us       2.513us       0.000us         0.00%       0.000us       0.000us             9  
-                                       aten::as_strided         1.03%       8.541us         1.03%       8.541us       0.569us       0.000us         0.00%       0.000us       0.000us            15  
-                                            aten::empty         1.40%      11.610us         1.40%      11.610us       3.870us       0.000us         0.00%       0.000us       0.000us             3  
-                                          aten::resize_         0.97%       8.050us         0.97%       8.050us       2.683us       0.000us         0.00%       0.000us       0.000us             3  
-                                          aten::squeeze         0.68%       5.660us         0.82%       6.830us       2.277us       0.000us         0.00%       0.000us       0.000us             3  
+                                            torch_eager         0.00%       0.000us         0.00%       0.000us       0.000us     323.833us      1729.88%     323.833us     323.833us             1  
+                                            torch_eager        14.51%     116.191us        99.37%     795.884us     795.884us       0.000us         0.00%      20.608us      20.608us             1  
+                                               aten::to         0.75%       6.009us        67.15%     537.870us      89.645us       0.000us         0.00%      13.376us       2.229us             6  
+                                         aten::_to_copy         2.93%      23.471us        66.40%     531.861us      88.644us       0.000us         0.00%      13.376us       2.229us             6  
+                                            aten::copy_         6.32%      50.599us        59.65%     477.769us      79.628us      11.488us        61.37%      13.376us       2.229us             6  
+                                           aten::conv1d         0.81%       6.510us        14.38%     115.173us      38.391us       0.000us         0.00%       7.232us       2.411us             3  
+                                      aten::convolution         1.28%      10.221us        13.57%     108.663us      36.221us       0.000us         0.00%       7.232us       2.411us             3  
+                                     aten::_convolution         2.73%      21.890us        12.29%      98.442us      32.814us       0.000us         0.00%       7.232us       2.411us             3  
+                                aten::_conv_depthwise2d         2.76%      22.080us         7.70%      61.700us      20.567us       7.232us        38.63%       7.232us       2.411us             3  
+void at::native::(anonymous namespace)::conv_depthwi...         0.00%       0.000us         0.00%       0.000us       0.000us       7.232us        38.63%       7.232us       2.411us             3  
+void at::native::unrolled_elementwise_kernel&lt;at::nat...         0.00%       0.000us         0.00%       0.000us       0.000us       5.888us        31.45%       5.888us       1.963us             3  
+void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us       5.600us        29.91%       5.600us       1.867us             3  
+                                Activity Buffer Request        31.20%     249.924us        31.20%     249.924us     249.924us       1.888us        10.09%       1.888us       1.888us             1  
+                                    aten::empty_strided         3.82%      30.621us         3.82%      30.621us       5.103us       0.000us         0.00%       0.000us       0.000us             6  
+                                       cudaLaunchKernel        24.75%     198.236us        24.75%     198.236us      22.026us       0.000us         0.00%       0.000us       0.000us             9  
+                                        aten::unsqueeze         2.09%      16.762us         2.71%      21.692us       2.410us       0.000us         0.00%       0.000us       0.000us             9  
+                                       aten::as_strided         1.04%       8.330us         1.04%       8.330us       0.555us       0.000us         0.00%       0.000us       0.000us            15  
+                                            aten::empty         1.15%       9.220us         1.15%       9.220us       3.073us       0.000us         0.00%       0.000us       0.000us             3  
+                                          aten::resize_         1.17%       9.410us         1.17%       9.410us       3.137us       0.000us         0.00%       0.000us       0.000us             3  
+                                          aten::squeeze         0.72%       5.800us         0.89%       7.160us       2.387us       0.000us         0.00%       0.000us       0.000us             3  
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
-Self CPU time total: 831.660us
-Self CUDA time total: 18.656us
+Self CPU time total: 800.944us
+Self CUDA time total: 18.720us
 
 
 
@@ -4406,29 +4633,29 @@ PROFILE TRACE: torch_eager | cuda_B4_D64_S128_W4
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
                                                    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     351.835us      1817.33%     351.835us     351.835us             1  
-                                            torch_eager        14.08%     121.071us        99.43%     854.999us     854.999us       0.000us         0.00%      21.248us      21.248us             1  
-                                               aten::to         0.71%       6.141us        68.62%     590.084us      98.347us       0.000us         0.00%      13.312us       2.219us             6  
-                                         aten::_to_copy         2.73%      23.503us        67.91%     583.943us      97.324us       0.000us         0.00%      13.312us       2.219us             6  
-                                            aten::copy_         6.25%      53.711us        59.45%     511.250us      85.208us      11.424us        59.01%      13.312us       2.219us             6  
-                                           aten::conv1d         0.65%       5.630us        13.53%     116.322us      38.774us       0.000us         0.00%       7.936us       2.645us             3  
-                                      aten::convolution         1.12%       9.630us        12.87%     110.692us      36.897us       0.000us         0.00%       7.936us       2.645us             3  
-                                     aten::_convolution         2.70%      23.181us        11.75%     101.062us      33.687us       0.000us         0.00%       7.936us       2.645us             3  
-                                aten::_conv_depthwise2d         2.42%      20.779us         7.31%      62.821us      20.940us       7.936us        40.99%       7.936us       2.645us             3  
-void at::native::(anonymous namespace)::conv_depthwi...         0.00%       0.000us         0.00%       0.000us       0.000us       7.936us        40.99%       7.936us       2.645us             3  
-void at::native::unrolled_elementwise_kernel&lt;at::nat...         0.00%       0.000us         0.00%       0.000us       0.000us       5.856us        30.25%       5.856us       1.952us             3  
-void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us       5.568us        28.76%       5.568us       1.856us             3  
-                                Activity Buffer Request        31.74%     272.946us        31.74%     272.946us     272.946us       1.888us         9.75%       1.888us       1.888us             1  
-                                    aten::empty_strided         5.72%      49.190us         5.72%      49.190us       8.198us       0.000us         0.00%       0.000us       0.000us             6  
-                                       cudaLaunchKernel        24.15%     207.684us        24.15%     207.684us      23.076us       0.000us         0.00%       0.000us       0.000us             9  
-                                        aten::unsqueeze         2.01%      17.302us         2.65%      22.752us       2.528us       0.000us         0.00%       0.000us       0.000us             9  
-                                       aten::as_strided         1.04%       8.971us         1.04%       8.971us       0.598us       0.000us         0.00%       0.000us       0.000us            15  
-                                            aten::empty         1.16%       9.970us         1.16%       9.970us       3.323us       0.000us         0.00%       0.000us       0.000us             3  
-                                          aten::resize_         1.04%       8.981us         1.04%       8.981us       2.994us       0.000us         0.00%       0.000us       0.000us             3  
-                                          aten::squeeze         0.67%       5.729us         0.81%       6.930us       2.310us       0.000us         0.00%       0.000us       0.000us             3  
+                                            torch_eager         0.00%       0.000us         0.00%       0.000us       0.000us     324.666us      1676.91%     324.666us     324.666us             1  
+                                            torch_eager        15.17%     119.302us        99.37%     781.483us     781.483us       0.000us         0.00%      21.249us      21.249us             1  
+                                               aten::to         0.72%       5.648us        65.85%     517.928us      86.321us       0.000us         0.00%      13.345us       2.224us             6  
+                                         aten::_to_copy         2.87%      22.611us        65.14%     512.280us      85.380us       0.000us         0.00%      13.345us       2.224us             6  
+                                            aten::copy_         6.22%      48.900us        58.49%     460.037us      76.673us      11.457us        59.18%      13.345us       2.224us             6  
+                                           aten::conv1d         0.87%       6.869us        14.99%     117.911us      39.304us       0.000us         0.00%       7.904us       2.635us             3  
+                                      aten::convolution         1.27%      10.002us        14.12%     111.042us      37.014us       0.000us         0.00%       7.904us       2.635us             3  
+                                     aten::_convolution         2.89%      22.710us        12.85%     101.040us      33.680us       0.000us         0.00%       7.904us       2.635us             3  
+                                aten::_conv_depthwise2d         2.75%      21.590us         8.00%      62.920us      20.973us       7.904us        40.82%       7.904us       2.635us             3  
+void at::native::(anonymous namespace)::conv_depthwi...         0.00%       0.000us         0.00%       0.000us       0.000us       7.904us        40.82%       7.904us       2.635us             3  
+void at::native::unrolled_elementwise_kernel&lt;at::nat...         0.00%       0.000us         0.00%       0.000us       0.000us       5.825us        30.09%       5.825us       1.942us             3  
+void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us       5.632us        29.09%       5.632us       1.877us             3  
+                                Activity Buffer Request        30.25%     237.875us        30.25%     237.875us     237.875us       1.888us         9.75%       1.888us       1.888us             1  
+                                    aten::empty_strided         3.77%      29.632us         3.77%      29.632us       4.939us       0.000us         0.00%       0.000us       0.000us             6  
+                                       cudaLaunchKernel        24.87%     195.612us        24.87%     195.612us      21.735us       0.000us         0.00%       0.000us       0.000us             9  
+                                        aten::unsqueeze         2.14%      16.821us         2.78%      21.881us       2.431us       0.000us         0.00%       0.000us       0.000us             9  
+                                       aten::as_strided         1.08%       8.481us         1.08%       8.481us       0.565us       0.000us         0.00%       0.000us       0.000us            15  
+                                            aten::empty         1.22%       9.600us         1.22%       9.600us       3.200us       0.000us         0.00%       0.000us       0.000us             3  
+                                          aten::resize_         1.19%       9.380us         1.19%       9.380us       3.127us       0.000us         0.00%       0.000us       0.000us             3  
+                                          aten::squeeze         0.75%       5.869us         0.93%       7.280us       2.427us       0.000us         0.00%       0.000us       0.000us             3  
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
-Self CPU time total: 859.928us
-Self CUDA time total: 19.360us
+Self CPU time total: 786.473us
+Self CUDA time total: 19.361us
 
 
 
@@ -4438,29 +4665,29 @@ PROFILE TRACE: torch_eager | cuda_B4_D64_S512_W2
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
                                                    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     336.603us      1730.16%     336.603us     336.603us             1  
-                                            torch_eager         7.08%     144.052us        99.72%       2.028ms       2.028ms       0.000us         0.00%      21.631us      21.631us             1  
-                                               aten::to         0.35%       7.181us        85.48%       1.738ms     289.712us       0.000us         0.00%      14.367us       2.394us             6  
-                                         aten::_to_copy         1.19%      24.130us        85.13%       1.731ms     288.515us       0.000us         0.00%      14.367us       2.394us             6  
-                                            aten::copy_         2.47%      50.222us        82.49%       1.678ms     279.593us      12.191us        62.66%      14.367us       2.394us             6  
-                                           aten::conv1d         0.31%       6.211us         5.80%     117.993us      39.331us       0.000us         0.00%       7.264us       2.421us             3  
-                                      aten::convolution         0.49%       9.910us         5.50%     111.782us      37.261us       0.000us         0.00%       7.264us       2.421us             3  
-                                     aten::_convolution         1.11%      22.590us         5.01%     101.872us      33.957us       0.000us         0.00%       7.264us       2.421us             3  
-                                aten::_conv_depthwise2d         1.01%      20.531us         3.08%      62.662us      20.887us       7.264us        37.34%       7.264us       2.421us             3  
-void at::native::(anonymous namespace)::conv_depthwi...         0.00%       0.000us         0.00%       0.000us       0.000us       7.264us        37.34%       7.264us       2.421us             3  
-void at::native::unrolled_elementwise_kernel&lt;at::nat...         0.00%       0.000us         0.00%       0.000us       0.000us       6.272us        32.24%       6.272us       2.091us             3  
-void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us       5.919us        30.42%       5.919us       1.973us             3  
-                                Activity Buffer Request        70.94%       1.443ms        70.94%       1.443ms       1.443ms       2.176us        11.18%       2.176us       2.176us             1  
-                                    aten::empty_strided         1.45%      29.401us         1.45%      29.401us       4.900us       0.000us         0.00%       0.000us       0.000us             6  
-                                       cudaLaunchKernel        10.12%     205.814us        10.12%     205.814us      22.868us       0.000us         0.00%       0.000us       0.000us             9  
-                                        aten::unsqueeze         0.93%      19.000us         1.20%      24.310us       2.701us       0.000us         0.00%       0.000us       0.000us             9  
-                                       aten::as_strided         0.43%       8.650us         0.43%       8.650us       0.577us       0.000us         0.00%       0.000us       0.000us            15  
-                                            aten::empty         0.52%      10.541us         0.52%      10.541us       3.514us       0.000us         0.00%       0.000us       0.000us             3  
-                                          aten::resize_         0.51%      10.450us         0.51%      10.450us       3.483us       0.000us         0.00%       0.000us       0.000us             3  
-                                          aten::squeeze         0.30%       6.000us         0.36%       7.230us       2.410us       0.000us         0.00%       0.000us       0.000us             3  
+                                            torch_eager         0.00%       0.000us         0.00%       0.000us       0.000us     328.865us      1704.41%     328.865us     328.865us             1  
+                                            torch_eager        14.92%     117.622us        99.37%     783.184us     783.184us       0.000us         0.00%      21.439us      21.439us             1  
+                                               aten::to         0.74%       5.810us        66.49%     524.079us      87.347us       0.000us         0.00%      14.207us       2.368us             6  
+                                         aten::_to_copy         3.01%      23.701us        65.75%     518.269us      86.378us       0.000us         0.00%      14.207us       2.368us             6  
+                                            aten::copy_         6.49%      51.190us        58.71%     462.718us      77.120us      12.063us        62.52%      14.207us       2.368us             6  
+                                           aten::conv1d         0.75%       5.890us        14.60%     115.093us      38.364us       0.000us         0.00%       7.232us       2.411us             3  
+                                      aten::convolution         1.22%       9.630us        13.86%     109.203us      36.401us       0.000us         0.00%       7.232us       2.411us             3  
+                                     aten::_convolution         2.83%      22.270us        12.63%      99.573us      33.191us       0.000us         0.00%       7.232us       2.411us             3  
+                                aten::_conv_depthwise2d         2.80%      22.070us         7.82%      61.673us      20.558us       7.232us        37.48%       7.232us       2.411us             3  
+void at::native::(anonymous namespace)::conv_depthwi...         0.00%       0.000us         0.00%       0.000us       0.000us       7.232us        37.48%       7.232us       2.411us             3  
+void at::native::unrolled_elementwise_kernel&lt;at::nat...         0.00%       0.000us         0.00%       0.000us       0.000us       6.240us        32.34%       6.240us       2.080us             3  
+void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us       5.823us        30.18%       5.823us       1.941us             3  
+                                Activity Buffer Request        29.70%     234.095us        29.70%     234.095us     234.095us       2.144us        11.11%       2.144us       2.144us             1  
+                                    aten::empty_strided         4.04%      31.850us         4.04%      31.850us       5.308us       0.000us         0.00%       0.000us       0.000us             6  
+                                       cudaLaunchKernel        25.25%     199.015us        25.25%     199.015us      22.113us       0.000us         0.00%       0.000us       0.000us             9  
+                                        aten::unsqueeze         2.15%      16.950us         2.78%      21.920us       2.436us       0.000us         0.00%       0.000us       0.000us             9  
+                                       aten::as_strided         1.05%       8.280us         1.05%       8.280us       0.552us       0.000us         0.00%       0.000us       0.000us            15  
+                                            aten::empty         1.22%       9.600us         1.22%       9.600us       3.200us       0.000us         0.00%       0.000us       0.000us             3  
+                                          aten::resize_         1.07%       8.421us         1.07%       8.421us       2.807us       0.000us         0.00%       0.000us       0.000us             3  
+                                          aten::squeeze         0.76%       5.960us         0.92%       7.270us       2.423us       0.000us         0.00%       0.000us       0.000us             3  
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
-Self CPU time total: 2.034ms
-Self CUDA time total: 19.455us
+Self CPU time total: 788.184us
+Self CUDA time total: 19.295us
 
 
 
@@ -4470,29 +4697,29 @@ PROFILE TRACE: torch_eager | cuda_B4_D64_S512_W4
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
                                                    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     330.813us      1648.79%     330.813us     330.813us             1  
-                                            torch_eager        15.21%     119.271us        99.33%     778.918us     778.918us       0.000us         0.00%      22.208us      22.208us             1  
-                                               aten::to         0.72%       5.611us        65.71%     515.312us      85.885us       0.000us         0.00%      14.305us       2.384us             6  
-                                         aten::_to_copy         2.87%      22.510us        65.00%     509.701us      84.950us       0.000us         0.00%      14.305us       2.384us             6  
-                                            aten::copy_         6.38%      50.021us        58.03%     455.090us      75.848us      12.161us        60.61%      14.305us       2.384us             6  
-                                           aten::conv1d         0.69%       5.380us        15.11%     118.473us      39.491us       0.000us         0.00%       7.903us       2.634us             3  
-                                      aten::convolution         1.31%      10.292us        14.42%     113.093us      37.698us       0.000us         0.00%       7.903us       2.634us             3  
-                                     aten::_convolution         2.98%      23.360us        13.11%     102.801us      34.267us       0.000us         0.00%       7.903us       2.634us             3  
-                                aten::_conv_depthwise2d         2.80%      21.952us         8.17%      64.041us      21.347us       7.903us        39.39%       7.903us       2.634us             3  
-void at::native::(anonymous namespace)::conv_depthwi...         0.00%       0.000us         0.00%       0.000us       0.000us       7.903us        39.39%       7.903us       2.634us             3  
-void at::native::unrolled_elementwise_kernel&lt;at::nat...         0.00%       0.000us         0.00%       0.000us       0.000us       6.240us        31.10%       6.240us       2.080us             3  
-void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us       5.921us        29.51%       5.921us       1.974us             3  
-                                Activity Buffer Request        27.43%     215.065us        27.43%     215.065us     215.065us       2.144us        10.69%       2.144us       2.144us             1  
-                                    aten::empty_strided         4.09%      32.101us         4.09%      32.101us       5.350us       0.000us         0.00%       0.000us       0.000us             6  
-                                       cudaLaunchKernel        27.10%     212.544us        27.10%     212.544us      23.616us       0.000us         0.00%       0.000us       0.000us             9  
-                                        aten::unsqueeze         2.14%      16.752us         2.74%      21.481us       2.387us       0.000us         0.00%       0.000us       0.000us             9  
-                                       aten::as_strided         1.03%       8.081us         1.03%       8.081us       0.539us       0.000us         0.00%       0.000us       0.000us            15  
-                                            aten::empty         1.34%      10.539us         1.34%      10.539us       3.513us       0.000us         0.00%       0.000us       0.000us             3  
-                                          aten::resize_         1.15%       9.010us         1.15%       9.010us       3.003us       0.000us         0.00%       0.000us       0.000us             3  
-                                          aten::squeeze         0.74%       5.829us         0.90%       7.070us       2.357us       0.000us         0.00%       0.000us       0.000us             3  
+                                            torch_eager         0.00%       0.000us         0.00%       0.000us       0.000us     334.745us      1665.90%     334.745us     334.745us             1  
+                                            torch_eager        14.26%     118.712us        99.40%     827.395us     827.395us       0.000us         0.00%      22.270us      22.270us             1  
+                                               aten::to         0.70%       5.840us        67.41%     561.119us      93.520us       0.000us         0.00%      14.335us       2.389us             6  
+                                         aten::_to_copy         2.86%      23.780us        66.71%     555.279us      92.546us       0.000us         0.00%      14.335us       2.389us             6  
+                                            aten::copy_         6.22%      51.741us        60.26%     501.588us      83.598us      12.159us        60.51%      14.335us       2.389us             6  
+                                           aten::conv1d         0.81%       6.751us        14.52%     120.873us      40.291us       0.000us         0.00%       7.935us       2.645us             3  
+                                      aten::convolution         1.20%       9.989us        13.71%     114.122us      38.041us       0.000us         0.00%       7.935us       2.645us             3  
+                                     aten::_convolution         2.78%      23.181us        12.51%     104.133us      34.711us       0.000us         0.00%       7.935us       2.645us             3  
+                                aten::_conv_depthwise2d         2.64%      22.000us         7.72%      64.243us      21.414us       7.935us        39.49%       7.935us       2.645us             3  
+void at::native::(anonymous namespace)::conv_depthwi...         0.00%       0.000us         0.00%       0.000us       0.000us       7.935us        39.49%       7.935us       2.645us             3  
+void at::native::unrolled_elementwise_kernel&lt;at::nat...         0.00%       0.000us         0.00%       0.000us       0.000us       6.239us        31.05%       6.239us       2.080us             3  
+void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us       5.920us        29.46%       5.920us       1.973us             3  
+                                Activity Buffer Request        32.59%     271.245us        32.59%     271.245us     271.245us       2.176us        10.83%       2.176us       2.176us             1  
+                                    aten::empty_strided         3.59%      29.911us         3.59%      29.911us       4.985us       0.000us         0.00%       0.000us       0.000us             6  
+                                       cudaLaunchKernel        24.22%     201.614us        24.22%     201.614us      22.402us       0.000us         0.00%       0.000us       0.000us             9  
+                                        aten::unsqueeze         2.06%      17.131us         2.68%      22.291us       2.477us       0.000us         0.00%       0.000us       0.000us             9  
+                                       aten::as_strided         1.07%       8.900us         1.07%       8.900us       0.593us       0.000us         0.00%       0.000us       0.000us            15  
+                                            aten::empty         1.16%       9.640us         1.16%       9.640us       3.213us       0.000us         0.00%       0.000us       0.000us             3  
+                                          aten::resize_         1.15%       9.591us         1.15%       9.591us       3.197us       0.000us         0.00%       0.000us       0.000us             3  
+                                          aten::squeeze         0.79%       6.549us         0.97%       8.109us       2.703us       0.000us         0.00%       0.000us       0.000us             3  
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
-Self CPU time total: 784.179us
-Self CUDA time total: 20.064us
+Self CPU time total: 832.395us
+Self CUDA time total: 20.094us
 
 
 
@@ -4502,29 +4729,29 @@ PROFILE TRACE: torch_eager | cuda_B4_D64_S2048_W2
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
                                                    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     332.020us       920.74%     332.020us     332.020us             1  
-                                            torch_eager        14.90%     117.475us        99.36%     783.438us     783.438us       0.000us         0.00%      38.651us      38.651us             1  
-                                           aten::conv1d         0.68%       5.380us        14.48%     114.172us      38.057us       0.000us         0.00%      20.190us       6.730us             3  
-                                      aten::convolution         1.18%       9.340us        13.80%     108.792us      36.264us       0.000us         0.00%      20.190us       6.730us             3  
-                                     aten::_convolution         2.79%      21.980us        12.61%      99.452us      33.151us       0.000us         0.00%      20.190us       6.730us             3  
-                                aten::_conv_depthwise2d         2.62%      20.631us         7.92%      62.452us      20.817us      20.190us        55.99%      20.190us       6.730us             3  
-void at::native::(anonymous namespace)::conv_depthwi...         0.00%       0.000us         0.00%       0.000us       0.000us      20.190us        55.99%      20.190us       6.730us             3  
-                                               aten::to         0.79%       6.191us        66.58%     524.962us      87.494us       0.000us         0.00%      18.461us       3.077us             6  
-                                         aten::_to_copy         2.94%      23.190us        65.79%     518.771us      86.462us       0.000us         0.00%      18.461us       3.077us             6  
-                                            aten::copy_         6.46%      50.920us        58.97%     464.950us      77.492us      15.870us        44.01%      18.461us       3.077us             6  
-void at::native::unrolled_elementwise_kernel&lt;at::nat...         0.00%       0.000us         0.00%       0.000us       0.000us       8.479us        23.51%       8.479us       2.826us             3  
-void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us       7.391us        20.50%       7.391us       2.464us             3  
-                                Activity Buffer Request        28.77%     226.875us        28.77%     226.875us     226.875us       2.591us         7.19%       2.591us       2.591us             1  
-                                    aten::empty_strided         3.88%      30.631us         3.88%      30.631us       5.105us       0.000us         0.00%       0.000us       0.000us             6  
-                                       cudaLaunchKernel        26.51%     209.015us        26.51%     209.015us      23.224us       0.000us         0.00%       0.000us       0.000us             9  
-                                        aten::unsqueeze         2.16%      17.000us         2.79%      22.010us       2.446us       0.000us         0.00%       0.000us       0.000us             9  
-                                       aten::as_strided         1.07%       8.441us         1.07%       8.441us       0.563us       0.000us         0.00%       0.000us       0.000us            15  
-                                            aten::empty         1.46%      11.521us         1.46%      11.521us       3.840us       0.000us         0.00%       0.000us       0.000us             3  
-                                          aten::resize_         1.07%       8.440us         1.07%       8.440us       2.813us       0.000us         0.00%       0.000us       0.000us             3  
-                                          aten::squeeze         0.73%       5.720us         0.87%       6.850us       2.283us       0.000us         0.00%       0.000us       0.000us             3  
+                                            torch_eager         0.00%       0.000us         0.00%       0.000us       0.000us     330.142us       918.64%     330.142us     330.142us             1  
+                                            torch_eager        14.68%     120.212us        99.34%     813.674us     813.674us       0.000us         0.00%      38.530us      38.530us             1  
+                                           aten::conv1d         0.79%       6.500us        14.15%     115.923us      38.641us       0.000us         0.00%      20.161us       6.720us             3  
+                                      aten::convolution         1.18%       9.650us        13.36%     109.423us      36.474us       0.000us         0.00%      20.161us       6.720us             3  
+                                     aten::_convolution         2.75%      22.509us        12.18%      99.773us      33.258us       0.000us         0.00%      20.161us       6.720us             3  
+                                aten::_conv_depthwise2d         2.55%      20.922us         7.56%      61.883us      20.628us      20.161us        56.10%      20.161us       6.720us             3  
+void at::native::(anonymous namespace)::conv_depthwi...         0.00%       0.000us         0.00%       0.000us       0.000us      20.161us        56.10%      20.161us       6.720us             3  
+                                               aten::to         0.72%       5.880us        67.15%     549.969us      91.661us       0.000us         0.00%      18.369us       3.061us             6  
+                                         aten::_to_copy         2.82%      23.099us        66.43%     544.089us      90.682us       0.000us         0.00%      18.369us       3.061us             6  
+                                            aten::copy_         6.44%      52.723us        59.97%     491.160us      81.860us      15.777us        43.90%      18.369us       3.061us             6  
+void at::native::unrolled_elementwise_kernel&lt;at::nat...         0.00%       0.000us         0.00%       0.000us       0.000us       8.448us        23.51%       8.448us       2.816us             3  
+void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us       7.329us        20.39%       7.329us       2.443us             3  
+                                Activity Buffer Request        32.20%     263.764us        32.20%     263.764us     263.764us       2.592us         7.21%       2.592us       2.592us             1  
+                                    aten::empty_strided         3.64%      29.830us         3.64%      29.830us       4.972us       0.000us         0.00%       0.000us       0.000us             6  
+                                       cudaLaunchKernel        24.00%     196.543us        24.00%     196.543us      21.838us       0.000us         0.00%       0.000us       0.000us             9  
+                                        aten::unsqueeze         2.14%      17.540us         2.77%      22.711us       2.523us       0.000us         0.00%       0.000us       0.000us             9  
+                                       aten::as_strided         1.07%       8.761us         1.07%       8.761us       0.584us       0.000us         0.00%       0.000us       0.000us            15  
+                                            aten::empty         1.21%       9.871us         1.21%       9.871us       3.290us       0.000us         0.00%       0.000us       0.000us             3  
+                                          aten::resize_         1.13%       9.220us         1.13%       9.220us       3.073us       0.000us         0.00%       0.000us       0.000us             3  
+                                          aten::squeeze         0.68%       5.610us         0.85%       7.000us       2.333us       0.000us         0.00%       0.000us       0.000us             3  
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
-Self CPU time total: 788.468us
-Self CUDA time total: 36.060us
+Self CPU time total: 819.054us
+Self CUDA time total: 35.938us
 
 
 
@@ -4534,29 +4761,29 @@ PROFILE TRACE: torch_eager | cuda_B4_D64_S2048_W4
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
                                                    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     323.480us       850.14%     323.480us     323.480us             1  
-                                            torch_eager        14.55%     115.293us        99.34%     787.399us     787.399us       0.000us         0.00%      40.643us      40.643us             1  
-                                           aten::conv1d         0.68%       5.400us        14.40%     114.153us      38.051us       0.000us         0.00%      22.336us       7.445us             3  
-                                      aten::convolution         1.16%       9.210us        13.72%     108.753us      36.251us       0.000us         0.00%      22.336us       7.445us             3  
-                                     aten::_convolution         2.80%      22.227us        12.56%      99.543us      33.181us       0.000us         0.00%      22.336us       7.445us             3  
-                                aten::_conv_depthwise2d         2.52%      20.003us         7.87%      62.343us      20.781us      22.336us        58.70%      22.336us       7.445us             3  
-void at::native::(anonymous namespace)::conv_depthwi...         0.00%       0.000us         0.00%       0.000us       0.000us      22.336us        58.70%      22.336us       7.445us             3  
-                                               aten::to         0.94%       7.450us        67.15%     532.253us      88.709us       0.000us         0.00%      18.307us       3.051us             6  
-                                         aten::_to_copy         2.90%      22.999us        66.21%     524.803us      87.467us       0.000us         0.00%      18.307us       3.051us             6  
-                                            aten::copy_         6.35%      50.294us        59.67%     472.953us      78.825us      15.714us        41.30%      18.307us       3.051us             6  
-void at::native::unrolled_elementwise_kernel&lt;at::nat...         0.00%       0.000us         0.00%       0.000us       0.000us       8.321us        21.87%       8.321us       2.774us             3  
-void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us       7.393us        19.43%       7.393us       2.464us             3  
-                                Activity Buffer Request        30.44%     241.286us        30.44%     241.286us     241.286us       2.593us         6.81%       2.593us       2.593us             1  
-                                    aten::empty_strided         3.64%      28.851us         3.64%      28.851us       4.808us       0.000us         0.00%       0.000us       0.000us             6  
-                                       cudaLaunchKernel        25.80%     204.463us        25.80%     204.463us      22.718us       0.000us         0.00%       0.000us       0.000us             9  
-                                        aten::unsqueeze         2.08%      16.472us         2.71%      21.512us       2.390us       0.000us         0.00%       0.000us       0.000us             9  
-                                       aten::as_strided         1.07%       8.500us         1.07%       8.500us       0.567us       0.000us         0.00%       0.000us       0.000us            15  
-                                            aten::empty         1.34%      10.600us         1.34%      10.600us       3.533us       0.000us         0.00%       0.000us       0.000us             3  
-                                          aten::resize_         1.09%       8.650us         1.09%       8.650us       2.883us       0.000us         0.00%       0.000us       0.000us             3  
-                                          aten::squeeze         0.71%       5.641us         0.87%       6.891us       2.297us       0.000us         0.00%       0.000us       0.000us             3  
+                                            torch_eager         0.00%       0.000us         0.00%       0.000us       0.000us     330.946us       872.79%     330.946us     330.946us             1  
+                                            torch_eager         6.07%     120.841us        99.75%       1.987ms       1.987ms       0.000us         0.00%      40.478us      40.478us             1  
+                                           aten::conv1d         0.33%       6.510us         5.92%     117.833us      39.278us       0.000us         0.00%      22.271us       7.424us             3  
+                                      aten::convolution         0.49%       9.850us         5.59%     111.323us      37.108us       0.000us         0.00%      22.271us       7.424us             3  
+                                     aten::_convolution         1.11%      22.181us         5.10%     101.473us      33.824us       0.000us         0.00%      22.271us       7.424us             3  
+                                aten::_conv_depthwise2d         1.10%      21.811us         3.17%      63.042us      21.014us      22.271us        58.73%      22.271us       7.424us             3  
+void at::native::(anonymous namespace)::conv_depthwi...         0.00%       0.000us         0.00%       0.000us       0.000us      22.271us        58.73%      22.271us       7.424us             3  
+                                               aten::to         0.30%       5.981us        86.38%       1.720ms     286.727us       0.000us         0.00%      18.207us       3.034us             6  
+                                         aten::_to_copy         1.18%      23.522us        86.08%       1.714ms     285.730us       0.000us         0.00%      18.207us       3.034us             6  
+                                            aten::copy_         2.55%      50.829us        83.41%       1.661ms     276.860us      15.647us        41.27%      18.207us       3.034us             6  
+void at::native::unrolled_elementwise_kernel&lt;at::nat...         0.00%       0.000us         0.00%       0.000us       0.000us       8.320us        21.94%       8.320us       2.773us             3  
+void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us       7.327us        19.32%       7.327us       2.442us             3  
+                                Activity Buffer Request        72.02%       1.434ms        72.02%       1.434ms       1.434ms       2.560us         6.75%       2.560us       2.560us             1  
+                                    aten::empty_strided         1.49%      29.700us         1.49%      29.700us       4.950us       0.000us         0.00%       0.000us       0.000us             6  
+                                       cudaLaunchKernel         9.93%     197.835us         9.93%     197.835us      21.982us       0.000us         0.00%       0.000us       0.000us             9  
+                                        aten::unsqueeze         0.90%      17.980us         1.17%      23.390us       2.599us       0.000us         0.00%       0.000us       0.000us             9  
+                                       aten::as_strided         0.44%       8.840us         0.44%       8.840us       0.589us       0.000us         0.00%       0.000us       0.000us            15  
+                                            aten::empty         0.50%       9.970us         0.50%       9.970us       3.323us       0.000us         0.00%       0.000us       0.000us             3  
+                                          aten::resize_         0.47%       9.410us         0.47%       9.410us       3.137us       0.000us         0.00%       0.000us       0.000us             3  
+                                          aten::squeeze         0.31%       6.110us         0.38%       7.490us       2.497us       0.000us         0.00%       0.000us       0.000us             3  
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
-Self CPU time total: 792.609us
-Self CUDA time total: 38.050us
+Self CPU time total: 1.992ms
+Self CUDA time total: 37.918us
 
 
 
@@ -4566,29 +4793,29 @@ PROFILE TRACE: torch_eager | cuda_B4_D2048_S128_W2
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
                                                    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     336.347us       525.30%     336.347us     336.347us             1  
-                                            torch_eager        15.07%     122.512us        99.37%     807.929us     807.929us       0.000us         0.00%      68.125us      68.125us             1  
-                                           aten::conv1d         0.67%       5.471us        14.06%     114.283us      38.094us       0.000us         0.00%      41.663us      13.888us             3  
-                                      aten::convolution         1.12%       9.100us        13.38%     108.812us      36.271us       0.000us         0.00%      41.663us      13.888us             3  
-                                     aten::_convolution         2.67%      21.730us        12.26%      99.712us      33.237us       0.000us         0.00%      41.663us      13.888us             3  
-                                aten::_conv_depthwise2d         2.50%      20.351us         7.64%      62.102us      20.701us      41.663us        65.07%      41.663us      13.888us             3  
-void at::native::(anonymous namespace)::conv_depthwi...         0.00%       0.000us         0.00%       0.000us       0.000us      41.663us        65.07%      41.663us      13.888us             3  
-                                               aten::to         0.85%       6.891us        66.79%     543.023us      90.504us       0.000us         0.00%      26.462us       4.410us             6  
-                                         aten::_to_copy         2.85%      23.179us        65.94%     536.132us      89.355us       0.000us         0.00%      26.462us       4.410us             6  
-                                            aten::copy_         6.23%      50.622us        59.33%     482.372us      80.395us      22.367us        34.93%      26.462us       4.410us             6  
-void at::native::unrolled_elementwise_kernel&lt;at::nat...         0.00%       0.000us         0.00%       0.000us       0.000us      11.935us        18.64%      11.935us       3.978us             3  
-void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      10.432us        16.29%      10.432us       3.477us             3  
-                                Activity Buffer Request        30.75%     250.036us        30.75%     250.036us     250.036us       4.095us         6.40%       4.095us       4.095us             1  
-                                    aten::empty_strided         3.76%      30.581us         3.76%      30.581us       5.097us       0.000us         0.00%       0.000us       0.000us             6  
-                                       cudaLaunchKernel        25.31%     205.766us        25.31%     205.766us      22.863us       0.000us         0.00%       0.000us       0.000us             9  
-                                        aten::unsqueeze         2.22%      18.070us         2.95%      23.970us       2.663us       0.000us         0.00%       0.000us       0.000us             9  
-                                       aten::as_strided         1.13%       9.220us         1.13%       9.220us       0.615us       0.000us         0.00%       0.000us       0.000us            15  
-                                            aten::empty         1.19%       9.690us         1.19%       9.690us       3.230us       0.000us         0.00%       0.000us       0.000us             3  
-                                          aten::resize_         0.99%       8.009us         0.99%       8.009us       2.670us       0.000us         0.00%       0.000us       0.000us             3  
-                                          aten::squeeze         0.69%       5.650us         0.84%       6.800us       2.267us       0.000us         0.00%       0.000us       0.000us             3  
+                                            torch_eager         0.00%       0.000us         0.00%       0.000us       0.000us     385.308us       602.34%     385.308us     385.308us             1  
+                                            torch_eager        14.42%     123.450us        99.41%     851.045us     851.045us       0.000us         0.00%      68.065us      68.065us             1  
+                                           aten::conv1d         0.67%       5.711us        13.49%     115.513us      38.504us       0.000us         0.00%      41.633us      13.878us             3  
+                                      aten::convolution         1.22%      10.470us        12.83%     109.802us      36.601us       0.000us         0.00%      41.633us      13.878us             3  
+                                     aten::_convolution         2.63%      22.491us        11.60%      99.332us      33.111us       0.000us         0.00%      41.633us      13.878us             3  
+                                aten::_conv_depthwise2d         2.49%      21.351us         7.22%      61.852us      20.617us      41.633us        65.08%      41.633us      13.878us             3  
+void at::native::(anonymous namespace)::conv_depthwi...         0.00%       0.000us         0.00%       0.000us       0.000us      41.633us        65.08%      41.633us      13.878us             3  
+                                               aten::to         0.71%       6.120us        68.08%     582.862us      97.144us       0.000us         0.00%      26.432us       4.405us             6  
+                                         aten::_to_copy         2.87%      24.611us        67.37%     576.742us      96.124us       0.000us         0.00%      26.432us       4.405us             6  
+                                            aten::copy_         6.21%      53.173us        60.75%     520.070us      86.678us      22.336us        34.92%      26.432us       4.405us             6  
+void at::native::unrolled_elementwise_kernel&lt;at::nat...         0.00%       0.000us         0.00%       0.000us       0.000us      11.936us        18.66%      11.936us       3.979us             3  
+void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      10.400us        16.26%      10.400us       3.467us             3  
+                                Activity Buffer Request        28.33%     242.554us        28.33%     242.554us     242.554us       4.096us         6.40%       4.096us       4.096us             1  
+                                    aten::empty_strided         3.74%      32.061us         3.74%      32.061us       5.344us       0.000us         0.00%       0.000us       0.000us             6  
+                                       cudaLaunchKernel        28.79%     246.523us        28.79%     246.523us      27.391us       0.000us         0.00%       0.000us       0.000us             9  
+                                        aten::unsqueeze         2.02%      17.269us         2.63%      22.529us       2.503us       0.000us         0.00%       0.000us       0.000us             9  
+                                       aten::as_strided         1.08%       9.240us         1.08%       9.240us       0.616us       0.000us         0.00%       0.000us       0.000us            15  
+                                            aten::empty         1.11%       9.521us         1.11%       9.521us       3.174us       0.000us         0.00%       0.000us       0.000us             3  
+                                          aten::resize_         1.03%       8.800us         1.03%       8.800us       2.933us       0.000us         0.00%       0.000us       0.000us             3  
+                                          aten::squeeze         0.68%       5.830us         0.84%       7.230us       2.410us       0.000us         0.00%       0.000us       0.000us             3  
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
-Self CPU time total: 813.049us
-Self CUDA time total: 64.030us
+Self CPU time total: 856.136us
+Self CUDA time total: 63.969us
 
 
 
@@ -4598,29 +4825,29 @@ PROFILE TRACE: torch_eager | cuda_B4_D2048_S128_W4
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
                                                    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     325.852us       467.34%     325.852us     325.852us             1  
-                                            torch_eager        14.81%     118.946us        99.39%     798.399us     798.399us       0.000us         0.00%      73.789us      73.789us             1  
-                                           aten::conv1d         0.70%       5.610us        14.26%     114.513us      38.171us       0.000us         0.00%      47.294us      15.765us             3  
-                                      aten::convolution         1.17%       9.382us        13.56%     108.903us      36.301us       0.000us         0.00%      47.294us      15.765us             3  
-                                     aten::_convolution         2.75%      22.119us        12.39%      99.521us      33.174us       0.000us         0.00%      47.294us      15.765us             3  
-                                aten::_conv_depthwise2d         2.53%      20.361us         7.64%      61.351us      20.450us      47.294us        67.83%      47.294us      15.765us             3  
-void at::native::(anonymous namespace)::conv_depthwi...         0.00%       0.000us         0.00%       0.000us       0.000us      47.294us        67.83%      47.294us      15.765us             3  
-                                               aten::to         0.79%       6.379us        67.07%     538.781us      89.797us       0.000us         0.00%      26.495us       4.416us             6  
-                                         aten::_to_copy         2.79%      22.401us        66.28%     532.402us      88.734us       0.000us         0.00%      26.495us       4.416us             6  
-                                            aten::copy_         6.32%      50.749us        59.88%     480.970us      80.162us      22.431us        32.17%      26.495us       4.416us             6  
-void at::native::unrolled_elementwise_kernel&lt;at::nat...         0.00%       0.000us         0.00%       0.000us       0.000us      12.031us        17.25%      12.031us       4.010us             3  
-void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      10.400us        14.92%      10.400us       3.467us             3  
-                                Activity Buffer Request        31.04%     249.326us        31.04%     249.326us     249.326us       4.064us         5.83%       4.064us       4.064us             1  
-                                    aten::empty_strided         3.61%      29.031us         3.61%      29.031us       4.839us       0.000us         0.00%       0.000us       0.000us             6  
-                                       cudaLaunchKernel        25.24%     202.725us        25.24%     202.725us      22.525us       0.000us         0.00%       0.000us       0.000us             9  
-                                        aten::unsqueeze         2.08%      16.713us         2.70%      21.680us       2.409us       0.000us         0.00%       0.000us       0.000us             9  
-                                       aten::as_strided         1.07%       8.568us         1.07%       8.568us       0.571us       0.000us         0.00%       0.000us       0.000us            15  
-                                            aten::empty         1.32%      10.569us         1.32%      10.569us       3.523us       0.000us         0.00%       0.000us       0.000us             3  
-                                          aten::resize_         1.07%       8.591us         1.07%       8.591us       2.864us       0.000us         0.00%       0.000us       0.000us             3  
-                                          aten::squeeze         0.82%       6.609us         1.00%       8.010us       2.670us       0.000us         0.00%       0.000us       0.000us             3  
+                                            torch_eager         0.00%       0.000us         0.00%       0.000us       0.000us     357.859us       513.70%     357.859us     357.859us             1  
+                                            torch_eager        20.53%     180.503us        99.40%     873.955us     873.955us       0.000us         0.00%      73.695us      73.695us             1  
+                                           aten::conv1d         0.63%       5.530us        15.78%     138.703us      46.234us       0.000us         0.00%      47.359us      15.786us             3  
+                                      aten::convolution         1.12%       9.840us        15.15%     133.173us      44.391us       0.000us         0.00%      47.359us      15.786us             3  
+                                     aten::_convolution         2.65%      23.331us        14.03%     123.333us      41.111us       0.000us         0.00%      47.359us      15.786us             3  
+                                aten::_conv_depthwise2d         2.63%      23.161us         9.53%      83.782us      27.927us      47.359us        67.98%      47.359us      15.786us             3  
+void at::native::(anonymous namespace)::conv_depthwi...         0.00%       0.000us         0.00%       0.000us       0.000us      47.359us        67.98%      47.359us      15.786us             3  
+                                               aten::to         0.72%       6.308us        59.85%     526.239us      87.707us       0.000us         0.00%      26.336us       4.389us             6  
+                                         aten::_to_copy         2.80%      24.578us        59.14%     519.931us      86.655us       0.000us         0.00%      26.336us       4.389us             6  
+                                            aten::copy_         6.12%      53.792us        52.84%     464.590us      77.432us      22.304us        32.02%      26.336us       4.389us             6  
+void at::native::unrolled_elementwise_kernel&lt;at::nat...         0.00%       0.000us         0.00%       0.000us       0.000us      11.840us        17.00%      11.840us       3.947us             3  
+void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      10.464us        15.02%      10.464us       3.488us             3  
+                                Activity Buffer Request        26.53%     233.244us        26.53%     233.244us     233.244us       4.032us         5.79%       4.032us       4.032us             1  
+                                    aten::empty_strided         3.50%      30.763us         3.50%      30.763us       5.127us       0.000us         0.00%       0.000us       0.000us             6  
+                                       cudaLaunchKernel        22.92%     201.494us        22.92%     201.494us      22.388us       0.000us         0.00%       0.000us       0.000us             9  
+                                        aten::unsqueeze         2.03%      17.891us         2.67%      23.440us       2.604us       0.000us         0.00%       0.000us       0.000us             9  
+                                       aten::as_strided         1.06%       9.339us         1.06%       9.339us       0.623us       0.000us         0.00%       0.000us       0.000us            15  
+                                            aten::empty         2.95%      25.971us         2.95%      25.971us       8.657us       0.000us         0.00%       0.000us       0.000us             3  
+                                          aten::resize_         1.22%      10.710us         1.22%      10.710us       3.570us       0.000us         0.00%       0.000us       0.000us             3  
+                                          aten::squeeze         0.71%       6.240us         0.88%       7.780us       2.593us       0.000us         0.00%       0.000us       0.000us             3  
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
-Self CPU time total: 803.260us
-Self CUDA time total: 69.725us
+Self CPU time total: 879.215us
+Self CUDA time total: 69.663us
 
 
 
@@ -4630,29 +4857,29 @@ PROFILE TRACE: torch_eager | cuda_B4_D2048_S512_W2
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
                                                    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     346.241us       186.20%     346.241us     346.241us             1  
-                                            torch_eager         7.14%     146.093us        99.75%       2.041ms       2.041ms       0.000us         0.00%     196.000us     196.000us             1  
-                                           aten::conv1d         0.31%       6.251us         5.79%     118.533us      39.511us       0.000us         0.00%     133.248us      44.416us             3  
-                                      aten::convolution         0.51%      10.359us         5.49%     112.282us      37.427us       0.000us         0.00%     133.248us      44.416us             3  
-                                     aten::_convolution         1.19%      24.280us         4.98%     101.923us      33.974us       0.000us         0.00%     133.248us      44.416us             3  
-                                aten::_conv_depthwise2d         1.04%      21.191us         3.02%      61.762us      20.587us     133.248us        71.66%     133.248us      44.416us             3  
-void at::native::(anonymous namespace)::conv_depthwi...         0.00%       0.000us         0.00%       0.000us       0.000us     133.248us        71.66%     133.248us      44.416us             3  
-                                               aten::to         0.32%       6.489us        85.48%       1.749ms     291.443us       0.000us         0.00%      62.752us      10.459us             6  
-                                         aten::_to_copy         1.16%      23.832us        85.16%       1.742ms     290.362us       0.000us         0.00%      62.752us      10.459us             6  
-                                            aten::copy_         2.53%      51.751us        82.59%       1.689ms     281.575us      52.704us        28.34%      62.752us      10.459us             6  
-void at::native::unrolled_elementwise_kernel&lt;at::nat...         0.00%       0.000us         0.00%       0.000us       0.000us      29.760us        16.00%      29.760us       9.920us             3  
-void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      22.944us        12.34%      22.944us       7.648us             3  
-                                Activity Buffer Request        71.04%       1.453ms        71.04%       1.453ms       1.453ms      10.048us         5.40%      10.048us      10.048us             1  
-                                    aten::empty_strided         1.41%      28.891us         1.41%      28.891us       4.815us       0.000us         0.00%       0.000us       0.000us             6  
-                                       cudaLaunchKernel        10.04%     205.324us        10.04%     205.324us      22.814us       0.000us         0.00%       0.000us       0.000us             9  
-                                        aten::unsqueeze         0.86%      17.550us         1.11%      22.661us       2.518us       0.000us         0.00%       0.000us       0.000us             9  
-                                       aten::as_strided         0.42%       8.641us         0.42%       8.641us       0.576us       0.000us         0.00%       0.000us       0.000us            15  
-                                            aten::empty         0.53%      10.750us         0.53%      10.750us       3.583us       0.000us         0.00%       0.000us       0.000us             3  
-                                          aten::resize_         0.44%       9.021us         0.44%       9.021us       3.007us       0.000us         0.00%       0.000us       0.000us             3  
-                                          aten::squeeze         0.30%       6.220us         0.37%       7.470us       2.490us       0.000us         0.00%       0.000us       0.000us             3  
+                                            torch_eager         0.00%       0.000us         0.00%       0.000us       0.000us     365.250us       197.10%     365.250us     365.250us             1  
+                                            torch_eager        14.70%     119.032us        99.37%     804.604us     804.604us       0.000us         0.00%     195.299us     195.299us             1  
+                                           aten::conv1d         0.95%       7.700us        17.22%     139.393us      46.464us       0.000us         0.00%     133.056us      44.352us             3  
+                                      aten::convolution         1.24%      10.040us        16.26%     131.693us      43.898us       0.000us         0.00%     133.056us      44.352us             3  
+                                     aten::_convolution         2.91%      23.550us        15.02%     121.653us      40.551us       0.000us         0.00%     133.056us      44.352us             3  
+                                aten::_conv_depthwise2d         2.69%      21.763us        10.08%      81.613us      27.204us     133.056us        71.80%     133.056us      44.352us             3  
+void at::native::(anonymous namespace)::conv_depthwi...         0.00%       0.000us         0.00%       0.000us       0.000us     133.056us        71.80%     133.056us      44.352us             3  
+                                               aten::to         0.75%       6.042us        64.10%     518.999us      86.500us       0.000us         0.00%      62.243us      10.374us             6  
+                                         aten::_to_copy         2.90%      23.470us        63.35%     512.957us      85.493us       0.000us         0.00%      62.243us      10.374us             6  
+                                            aten::copy_         6.35%      51.412us        56.59%     458.237us      76.373us      52.258us        28.20%      62.243us      10.374us             6  
+void at::native::unrolled_elementwise_kernel&lt;at::nat...         0.00%       0.000us         0.00%       0.000us       0.000us      29.250us        15.78%      29.250us       9.750us             3  
+void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      23.008us        12.42%      23.008us       7.669us             3  
+                                Activity Buffer Request        28.43%     230.213us        28.43%     230.213us     230.213us       9.985us         5.39%       9.985us       9.985us             1  
+                                    aten::empty_strided         3.86%      31.250us         3.86%      31.250us       5.208us       0.000us         0.00%       0.000us       0.000us             6  
+                                       cudaLaunchKernel        26.81%     217.052us        26.81%     217.052us      24.117us       0.000us         0.00%       0.000us       0.000us             9  
+                                        aten::unsqueeze         2.10%      17.030us         2.74%      22.170us       2.463us       0.000us         0.00%       0.000us       0.000us             9  
+                                       aten::as_strided         1.13%       9.170us         1.13%       9.170us       0.611us       0.000us         0.00%       0.000us       0.000us            15  
+                                            aten::empty         1.22%       9.870us         1.22%       9.870us       3.290us       0.000us         0.00%       0.000us       0.000us             3  
+                                          aten::resize_         1.18%       9.540us         1.18%       9.540us       3.180us       0.000us         0.00%       0.000us       0.000us             3  
+                                          aten::squeeze         0.78%       6.320us         1.00%       8.100us       2.700us       0.000us         0.00%       0.000us       0.000us             3  
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
-Self CPU time total: 2.046ms
-Self CUDA time total: 185.952us
+Self CPU time total: 809.694us
+Self CUDA time total: 185.314us
 
 
 
@@ -4662,29 +4889,29 @@ PROFILE TRACE: torch_eager | cuda_B4_D2048_S512_W4
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
                                                    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     341.281us       162.73%     341.281us     341.281us             1  
-                                            torch_eager        15.25%     117.693us        99.36%     766.878us     766.878us       0.000us         0.00%     223.168us     223.168us             1  
-                                           aten::conv1d         0.68%       5.279us        14.60%     112.702us      37.567us       0.000us         0.00%     154.016us      51.339us             3  
-                                      aten::convolution         1.24%       9.560us        13.92%     107.423us      35.808us       0.000us         0.00%     154.016us      51.339us             3  
-                                     aten::_convolution         2.67%      20.611us        12.68%      97.863us      32.621us       0.000us         0.00%     154.016us      51.339us             3  
-                                aten::_conv_depthwise2d         2.74%      21.170us         8.14%      62.852us      20.951us     154.016us        73.44%     154.016us      51.339us             3  
-void at::native::(anonymous namespace)::conv_depthwi...         0.00%       0.000us         0.00%       0.000us       0.000us     154.016us        73.44%     154.016us      51.339us             3  
-                                               aten::to         0.75%       5.750us        66.22%     511.121us      85.187us       0.000us         0.00%      69.152us      11.525us             6  
-                                         aten::_to_copy         2.86%      22.060us        65.48%     505.371us      84.228us       0.000us         0.00%      69.152us      11.525us             6  
-                                            aten::copy_         6.81%      52.581us        58.74%     453.391us      75.565us      55.712us        26.56%      69.152us      11.525us             6  
-void at::native::unrolled_elementwise_kernel&lt;at::nat...         0.00%       0.000us         0.00%       0.000us       0.000us      32.705us        15.59%      32.705us      10.902us             3  
-void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      23.007us        10.97%      23.007us       7.669us             3  
-                                Activity Buffer Request        28.38%     219.045us        28.38%     219.045us     219.045us      13.440us         6.41%      13.440us      13.440us             1  
-                                    aten::empty_strided         3.88%      29.920us         3.88%      29.920us       4.987us       0.000us         0.00%       0.000us       0.000us             6  
-                                       cudaLaunchKernel        26.50%     204.546us        26.50%     204.546us      22.727us       0.000us         0.00%       0.000us       0.000us             9  
-                                        aten::unsqueeze         2.10%      16.212us         2.69%      20.781us       2.309us       0.000us         0.00%       0.000us       0.000us             9  
-                                       aten::as_strided         1.01%       7.798us         1.01%       7.798us       0.520us       0.000us         0.00%       0.000us       0.000us            15  
-                                            aten::empty         1.33%      10.250us         1.33%      10.250us       3.417us       0.000us         0.00%       0.000us       0.000us             3  
-                                          aten::resize_         1.12%       8.651us         1.12%       8.651us       2.884us       0.000us         0.00%       0.000us       0.000us             3  
-                                          aten::squeeze         0.71%       5.470us         0.87%       6.730us       2.243us       0.000us         0.00%       0.000us       0.000us             3  
+                                            torch_eager         0.00%       0.000us         0.00%       0.000us       0.000us     352.824us       168.80%     352.824us     352.824us             1  
+                                            torch_eager        14.40%     121.160us        99.40%     836.424us     836.424us       0.000us         0.00%     222.266us     222.266us             1  
+                                           aten::conv1d         0.71%       5.981us        14.17%     119.243us      39.748us       0.000us         0.00%     153.724us      51.241us             3  
+                                      aten::convolution         1.17%       9.810us        13.46%     113.262us      37.754us       0.000us         0.00%     153.724us      51.241us             3  
+                                     aten::_convolution         2.76%      23.250us        12.29%     103.452us      34.484us       0.000us         0.00%     153.724us      51.241us             3  
+                                aten::_conv_depthwise2d         2.65%      22.340us         7.64%      64.321us      21.440us     153.724us        73.55%     153.724us      51.241us             3  
+void at::native::(anonymous namespace)::conv_depthwi...         0.00%       0.000us         0.00%       0.000us       0.000us     153.724us        73.55%     153.724us      51.241us             3  
+                                               aten::to         0.70%       5.880us        67.58%     568.691us      94.782us       0.000us         0.00%      68.542us      11.424us             6  
+                                         aten::_to_copy         2.81%      23.631us        66.88%     562.811us      93.802us       0.000us         0.00%      68.542us      11.424us             6  
+                                            aten::copy_         7.48%      62.921us        60.21%     506.640us      84.440us      55.294us        26.45%      68.542us      11.424us             6  
+void at::native::unrolled_elementwise_kernel&lt;at::nat...         0.00%       0.000us         0.00%       0.000us       0.000us      32.352us        15.48%      32.352us      10.784us             3  
+void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us      22.942us        10.98%      22.942us       7.647us             3  
+                                Activity Buffer Request        31.88%     268.245us        31.88%     268.245us     268.245us      13.248us         6.34%      13.248us      13.248us             1  
+                                    aten::empty_strided         3.87%      32.540us         3.87%      32.540us       5.423us       0.000us         0.00%       0.000us       0.000us             6  
+                                       cudaLaunchKernel        23.51%     197.824us        23.51%     197.824us      21.980us       0.000us         0.00%       0.000us       0.000us             9  
+                                        aten::unsqueeze         2.07%      17.378us         2.68%      22.521us       2.502us       0.000us         0.00%       0.000us       0.000us             9  
+                                       aten::as_strided         1.06%       8.883us         1.06%       8.883us       0.592us       0.000us         0.00%       0.000us       0.000us            15  
+                                            aten::empty         1.19%       9.991us         1.19%       9.991us       3.330us       0.000us         0.00%       0.000us       0.000us             3  
+                                          aten::resize_         1.15%       9.640us         1.15%       9.640us       3.213us       0.000us         0.00%       0.000us       0.000us             3  
+                                          aten::squeeze         0.71%       5.990us         0.89%       7.470us       2.490us       0.000us         0.00%       0.000us       0.000us             3  
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
-Self CPU time total: 771.798us
-Self CUDA time total: 209.728us
+Self CPU time total: 841.495us
+Self CUDA time total: 209.018us
 
 
 
@@ -4694,29 +4921,29 @@ PROFILE TRACE: torch_eager | cuda_B4_D2048_S2048_W2
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
                                                    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         6.80%     123.239us        52.01%     942.341us     942.341us       0.000us         0.00%       1.520ms       1.520ms             1  
-                                            torch_eager         0.00%       0.000us         0.00%       0.000us       0.000us       1.422ms       100.42%       1.422ms       1.422ms             1  
-                                               aten::to         0.34%       6.231us        36.98%     669.957us     111.660us       0.000us         0.00%     824.603us     137.434us             6  
-                                         aten::_to_copy         1.55%      28.122us        36.63%     663.726us     110.621us       0.000us         0.00%     824.603us     137.434us             6  
-                                            aten::copy_         2.95%      53.430us        24.71%     447.748us      74.625us     720.572us        50.89%     824.603us     137.434us             6  
-                                           aten::conv1d         0.34%       6.111us         6.66%     120.744us      40.248us       0.000us         0.00%     695.357us     231.786us             3  
-                                      aten::convolution         0.56%      10.201us         6.33%     114.633us      38.211us       0.000us         0.00%     695.357us     231.786us             3  
-                                     aten::_convolution         1.36%      24.689us         5.76%     104.432us      34.811us       0.000us         0.00%     695.357us     231.786us             3  
-                                aten::_conv_depthwise2d         1.22%      22.151us         3.50%      63.431us      21.144us     695.357us        49.11%     695.357us     231.786us             3  
-void at::native::(anonymous namespace)::conv_depthwi...         0.00%       0.000us         0.00%       0.000us       0.000us     695.357us        49.11%     695.357us     231.786us             3  
-void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us     407.263us        28.76%     407.263us     135.754us             3  
-void at::native::unrolled_elementwise_kernel&lt;at::nat...         0.00%       0.000us         0.00%       0.000us       0.000us     313.309us        22.13%     313.309us     104.436us             3  
-                                Activity Buffer Request        11.46%     207.684us        11.46%     207.684us     207.684us     104.031us         7.35%     104.031us     104.031us             1  
-                                    aten::empty_strided         2.02%      36.603us        10.37%     187.856us      31.309us       0.000us         0.00%       0.000us       0.000us             6  
-                                       cudaLaunchKernel        11.47%     207.874us        11.47%     207.874us      23.097us       0.000us         0.00%       0.000us       0.000us             9  
-                                        aten::unsqueeze         0.99%      18.011us         1.30%      23.581us       2.620us       0.000us         0.00%       0.000us       0.000us             9  
-                                       aten::as_strided         0.51%       9.270us         0.51%       9.270us       0.618us       0.000us         0.00%       0.000us       0.000us            15  
-                                            aten::empty         0.60%      10.830us         0.60%      10.830us       3.610us       0.000us         0.00%       0.000us       0.000us             3  
-                                          aten::resize_         0.51%       9.210us         0.51%       9.210us       3.070us       0.000us         0.00%       0.000us       0.000us             3  
-                                          aten::squeeze         0.35%       6.401us         0.43%       7.711us       2.570us       0.000us         0.00%       0.000us       0.000us             3  
+                                            torch_eager         6.78%     125.712us        53.74%     996.387us     996.387us       0.000us         0.00%       1.527ms       1.527ms             1  
+                                            torch_eager         0.00%       0.000us         0.00%       0.000us       0.000us       1.423ms       100.39%       1.423ms       1.423ms             1  
+                                               aten::to         0.35%       6.438us        38.84%     720.182us     120.030us       0.000us         0.00%     832.992us     138.832us             6  
+                                         aten::_to_copy         1.55%      28.691us        38.49%     713.744us     118.957us       0.000us         0.00%     832.992us     138.832us             6  
+                                            aten::copy_         2.90%      53.742us        26.33%     488.279us      81.380us     724.000us        51.06%     832.992us     138.832us             6  
+                                           aten::conv1d         0.38%       6.960us         6.55%     121.533us      40.511us       0.000us         0.00%     693.950us     231.317us             3  
+                                      aten::convolution         0.56%      10.430us         6.18%     114.573us      38.191us       0.000us         0.00%     693.950us     231.317us             3  
+                                     aten::_convolution         1.25%      23.268us         5.62%     104.143us      34.714us       0.000us         0.00%     693.950us     231.317us             3  
+                                aten::_conv_depthwise2d         1.23%      22.830us         3.48%      64.552us      21.517us     693.950us        48.94%     693.950us     231.317us             3  
+void at::native::(anonymous namespace)::conv_depthwi...         0.00%       0.000us         0.00%       0.000us       0.000us     693.950us        48.94%     693.950us     231.317us             3  
+void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us     410.655us        28.96%     410.655us     136.885us             3  
+void at::native::unrolled_elementwise_kernel&lt;at::nat...         0.00%       0.000us         0.00%       0.000us       0.000us     313.345us        22.10%     313.345us     104.448us             3  
+                                Activity Buffer Request        13.73%     254.654us        13.73%     254.654us     254.654us     108.992us         7.69%     108.992us     108.992us             1  
+                                    aten::empty_strided         2.01%      37.271us        10.61%     196.774us      32.796us       0.000us         0.00%       0.000us       0.000us             6  
+                                       cudaLaunchKernel        10.89%     201.884us        10.89%     201.884us      22.432us       0.000us         0.00%       0.000us       0.000us             9  
+                                        aten::unsqueeze         0.98%      18.223us         1.29%      23.933us       2.659us       0.000us         0.00%       0.000us       0.000us             9  
+                                       aten::as_strided         0.51%       9.490us         0.51%       9.490us       0.633us       0.000us         0.00%       0.000us       0.000us            15  
+                                            aten::empty         0.54%      10.101us         0.54%      10.101us       3.367us       0.000us         0.00%       0.000us       0.000us             3  
+                                          aten::resize_         0.52%       9.620us         0.52%       9.620us       3.207us       0.000us         0.00%       0.000us       0.000us             3  
+                                          aten::squeeze         0.34%       6.270us         0.41%       7.680us       2.560us       0.000us         0.00%       0.000us       0.000us             3  
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
-Self CPU time total: 1.812ms
-Self CUDA time total: 1.416ms
+Self CPU time total: 1.854ms
+Self CUDA time total: 1.418ms
 
 
 
@@ -4726,44 +4953,44 @@ PROFILE TRACE: torch_eager | cuda_B4_D2048_S2048_W4
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
                                                    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         6.34%     116.852us        42.22%     778.698us     778.698us       0.000us         0.00%       1.501ms       1.501ms             1  
-                                            torch_eager         0.00%       0.000us         0.00%       0.000us       0.000us       1.431ms       100.41%       1.431ms       1.431ms             1  
-                                               aten::to         0.32%       5.860us        28.27%     521.352us      86.892us       0.000us         0.00%     762.491us     127.082us             6  
-                                         aten::_to_copy         1.22%      22.580us        27.95%     515.492us      85.915us       0.000us         0.00%     762.491us     127.082us             6  
-                                            aten::copy_         2.82%      52.081us        25.06%     462.219us      77.037us     686.779us        48.19%     762.491us     127.082us             6  
-                                           aten::conv1d         0.30%       5.601us         6.22%     114.743us      38.248us       0.000us         0.00%     738.362us     246.121us             3  
-                                      aten::convolution         0.51%       9.380us         5.92%     109.142us      36.381us       0.000us         0.00%     738.362us     246.121us             3  
-                                     aten::_convolution         1.20%      22.202us         5.41%      99.762us      33.254us       0.000us         0.00%     738.362us     246.121us             3  
-                                aten::_conv_depthwise2d         1.10%      20.251us         3.34%      61.660us      20.553us     738.362us        51.81%     738.362us     246.121us             3  
-void at::native::(anonymous namespace)::conv_depthwi...         0.00%       0.000us         0.00%       0.000us       0.000us     738.362us        51.81%     738.362us     246.121us             3  
-void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us     398.333us        27.95%     398.333us     132.778us             3  
-void at::native::unrolled_elementwise_kernel&lt;at::nat...         0.00%       0.000us         0.00%       0.000us       0.000us     288.446us        20.24%     288.446us      96.149us             3  
-                                Activity Buffer Request        12.41%     228.855us        12.41%     228.855us     228.855us      75.712us         5.31%      75.712us      75.712us             1  
-                                    aten::empty_strided         1.66%      30.693us         1.66%      30.693us       5.115us       0.000us         0.00%       0.000us       0.000us             6  
-                                       cudaLaunchKernel        11.01%     202.993us        11.01%     202.993us      22.555us       0.000us         0.00%       0.000us       0.000us             9  
-                                        aten::unsqueeze         0.90%      16.659us         1.18%      21.780us       2.420us       0.000us         0.00%       0.000us       0.000us             9  
-                                       aten::as_strided         0.46%       8.512us         0.46%       8.512us       0.567us       0.000us         0.00%       0.000us       0.000us            15  
-                                            aten::empty         0.53%       9.739us         0.53%       9.739us       3.246us       0.000us         0.00%       0.000us       0.000us             3  
-                                          aten::resize_         0.54%       9.960us         0.54%       9.960us       3.320us       0.000us         0.00%       0.000us       0.000us             3  
-                                          aten::squeeze         0.33%       6.150us         0.40%       7.440us       2.480us       0.000us         0.00%       0.000us       0.000us             3  
+                                            torch_eager         4.03%     122.972us        65.43%       1.999ms       1.999ms       0.000us         0.00%       1.502ms       1.502ms             1  
+                                            torch_eager         0.00%       0.000us         0.00%       0.000us       0.000us       1.433ms       100.41%       1.433ms       1.433ms             1  
+                                               aten::to         0.19%       5.740us        56.63%       1.730ms     288.331us       0.000us         0.00%     766.432us     127.739us             6  
+                                         aten::_to_copy         0.79%      24.119us        56.45%       1.724ms     287.375us       0.000us         0.00%     766.432us     127.739us             6  
+                                            aten::copy_         1.70%      52.020us        54.70%       1.671ms     278.493us     691.168us        48.43%     766.432us     127.739us             6  
+                                           aten::conv1d         0.23%       6.891us         3.86%     118.002us      39.334us       0.000us         0.00%     736.031us     245.344us             3  
+                                      aten::convolution         0.33%       9.930us         3.64%     111.111us      37.037us       0.000us         0.00%     736.031us     245.344us             3  
+                                     aten::_convolution         0.74%      22.558us         3.31%     101.181us      33.727us       0.000us         0.00%     736.031us     245.344us             3  
+                                aten::_conv_depthwise2d         0.70%      21.291us         2.07%      63.232us      21.077us     736.031us        51.57%     736.031us     245.344us             3  
+void at::native::(anonymous namespace)::conv_depthwi...         0.00%       0.000us         0.00%       0.000us       0.000us     736.031us        51.57%     736.031us     245.344us             3  
+void at::native::elementwise_kernel&lt;128, 4, at::nati...         0.00%       0.000us         0.00%       0.000us       0.000us     401.120us        28.11%     401.120us     133.707us             3  
+void at::native::unrolled_elementwise_kernel&lt;at::nat...         0.00%       0.000us         0.00%       0.000us       0.000us     290.048us        20.32%     290.048us      96.683us             3  
+                                Activity Buffer Request        47.17%       1.441ms        47.17%       1.441ms       1.441ms      75.264us         5.27%      75.264us      75.264us             1  
+                                    aten::empty_strided         0.95%      29.171us         0.95%      29.171us       4.862us       0.000us         0.00%       0.000us       0.000us             6  
+                                       cudaLaunchKernel         6.58%     201.084us         6.58%     201.084us      22.343us       0.000us         0.00%       0.000us       0.000us             9  
+                                        aten::unsqueeze         0.57%      17.550us         0.75%      22.971us       2.552us       0.000us         0.00%       0.000us       0.000us             9  
+                                       aten::as_strided         0.30%       9.131us         0.30%       9.131us       0.609us       0.000us         0.00%       0.000us       0.000us            15  
+                                            aten::empty         0.33%       9.960us         0.33%       9.960us       3.320us       0.000us         0.00%       0.000us       0.000us             3  
+                                          aten::resize_         0.30%       9.060us         0.30%       9.060us       3.020us       0.000us         0.00%       0.000us       0.000us             3  
+                                          aten::squeeze         0.18%       5.561us         0.23%       7.041us       2.347us       0.000us         0.00%       0.000us       0.000us             3  
 -------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
-Self CPU time total: 1.844ms
-Self CUDA time total: 1.425ms
+Self CPU time total: 3.055ms
+Self CUDA time total: 1.427ms
 
 
 impl                     wl                  p50(ms)  ok
-torch_eager              cuda_B2_D2048_S128_W2     0.08  True
+torch_eager              cuda_B2_D2048_S128_W2     0.09  True
 torch_eager              cuda_B2_D2048_S128_W4     0.08  True
-torch_eager              cuda_B2_D2048_S2048_W2     0.15  True
+torch_eager              cuda_B2_D2048_S2048_W2     0.14  True
 torch_eager              cuda_B2_D2048_S2048_W4     0.16  True
-torch_eager              cuda_B2_D2048_S512_W2     0.08  True
+torch_eager              cuda_B2_D2048_S512_W2     0.09  True
 torch_eager              cuda_B2_D2048_S512_W4     0.08  True
 torch_eager              cuda_B2_D64_S128_W2     0.07  True
 torch_eager              cuda_B2_D64_S128_W4     0.09  True
-torch_eager              cuda_B2_D64_S2048_W2     0.08  True
+torch_eager              cuda_B2_D64_S2048_W2     0.09  True
 torch_eager              cuda_B2_D64_S2048_W4     0.08  True
 torch_eager              cuda_B2_D64_S512_W2     0.09  True
-torch_eager              cuda_B2_D64_S512_W4     0.08  True
+torch_eager              cuda_B2_D64_S512_W4     0.09  True
 torch_eager              cuda_B4_D2048_S128_W2     0.08  True
 torch_eager              cuda_B4_D2048_S128_W4     0.08  True
 torch_eager              cuda_B4_D2048_S2048_W2     0.49  True
@@ -4777,6 +5004,58 @@ torch_eager              cuda_B4_D64_S2048_W4     0.08  True
 torch_eager              cuda_B4_D64_S512_W2     0.08  True
 torch_eager              cuda_B4_D64_S512_W4     0.08  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;">
+   Building kernels-benchmark-tools @ file:///__w/kernels-benchmarks/kernels-benchmarks/tools
+Downloading pillow (6.7MiB)
+Downloading fonttools (4.7MiB)
+Downloading setuptools (1.1MiB)
+Downloading nvidia-cufft-cu12 (184.2MiB)
+Downloading nvidia-cusparselt-cu12 (273.9MiB)
+Downloading nvidia-cuda-cupti-cu12 (9.8MiB)
+Downloading nvidia-cublas-cu12 (566.8MiB)
+Downloading nvidia-cuda-nvrtc-cu12 (84.0MiB)
+Downloading nvidia-cusparse-cu12 (274.9MiB)
+Downloading nvidia-cudnn-cu12 (674.0MiB)
+Downloading nvidia-cusolver-cu12 (255.1MiB)
+Downloading sympy (6.0MiB)
+Downloading nvidia-cufile-cu12 (1.1MiB)
+Downloading networkx (1.9MiB)
+Downloading numpy (16.2MiB)
+Downloading nvidia-curand-cu12 (60.7MiB)
+Downloading nvidia-nvjitlink-cu12 (37.4MiB)
+Downloading kiwisolver (1.4MiB)
+Downloading matplotlib (8.3MiB)
+Downloading nvidia-nccl-cu12 (307.4MiB)
+Downloading torch (846.9MiB)
+Downloading triton (148.3MiB)
+ Downloading nvidia-cufile-cu12
+ Downloading kiwisolver
+ Downloading setuptools
+ Downloading fonttools
+ Downloading networkx
+ Downloading pillow
+      Built kernels-benchmark-tools @ file:///__w/kernels-benchmarks/kernels-benchmarks/tools
+ Downloading nvidia-cuda-cupti-cu12
+ Downloading matplotlib
+ Downloading numpy
+ Downloading sympy
+ Downloading nvidia-nvjitlink-cu12
+ Downloading nvidia-curand-cu12
+ Downloading nvidia-cuda-nvrtc-cu12
+ Downloading triton
+ Downloading nvidia-cufft-cu12
+ Downloading nvidia-cusolver-cu12
+ Downloading nvidia-cusparse-cu12
+ Downloading nvidia-cusparselt-cu12
+ Downloading nvidia-nccl-cu12
+ Downloading nvidia-cublas-cu12
+ Downloading nvidia-cudnn-cu12
+ Downloading torch
+Installed 37 packages in 218ms
+</div>
+</div>
 <div class="cell-artifacts">
 <h4>Artifacts:</h4>
 <a href="artifacts/benchmark/causal_conv1d.jsonl" class="artifact" target="_blank">causal_conv1d.jsonl</a>