KernelBench-Leaderboard/eval.html at main · AffectionateCurry/KernelBench-Leaderboard · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
<!DOCTYPE html>
<html lang="en">
<head>
    <meta charset="UTF-8">
    <meta name="viewport" content="width=device-width, initial-scale=1.0">
    <title>Evaluation - KernelBench</title>
    <link rel="stylesheet" href="css/style.css">
    <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:wght@400;500;600&family=Outfit:wght@400;500;600;700&display=swap" rel="stylesheet">
    <script>
        (function() {
            const theme = localStorage.getItem('theme') || 'dark';
            document.documentElement.setAttribute('data-theme', theme);
        })();
    </script>
    <style>
        /* Eval page specific styles */
        .eval-body {
            font-family: 'Outfit', -apple-system, BlinkMacSystemFont, sans-serif;
        }

        .eval-container {
            max-width: 900px;
            margin: 0 auto;
            padding: 2rem;
        }

        .eval-header {
            margin-bottom: 1.5rem;
            padding-bottom: 1.25rem;
            border-bottom: 1px solid var(--border);
        }

        .eval-title {
            font-size: 2rem;
            font-weight: 700;
            margin-bottom: 0.5rem;
        }

        .eval-subtitle {
            color: var(--text-muted);
            font-size: 1rem;
        }

        .eval-badge {
            display: inline-flex;
            align-items: center;
            gap: 0.5rem;
            padding: 0.375rem 0.75rem;
            background: rgba(210, 153, 34, 0.15);
            border: 1px solid rgba(210, 153, 34, 0.3);
            border-radius: 0.375rem;
            font-size: 0.8125rem;
            color: var(--warning);
            margin-bottom: 1rem;
        }

        .eval-content {
            line-height: 1.8;
        }

        .eval-content h2 {
            font-size: 1.375rem;
            font-weight: 600;
            margin-top: 2rem;
            margin-bottom: 0.75rem;
            padding-bottom: 0.5rem;
            border-bottom: 1px solid var(--border);
        }

        .eval-content h3 {
            font-size: 1.125rem;
            font-weight: 600;
            margin-top: 1.5rem;
            margin-bottom: 0.5rem;
        }

        .eval-content p {
            font-size: 0.9375rem;
            color: var(--text-muted);
            margin-bottom: 0.875rem;
        }

        .eval-content strong {
            color: var(--text);
        }

        .eval-content a {
            color: var(--accent);
            text-decoration: none;
        }

        .eval-content a:hover {
            text-decoration: underline;
        }

        .eval-content ul, .eval-content ol {
            margin-bottom: 0.875rem;
            padding-left: 1.25rem;
        }

        .eval-content li {
            font-size: 0.9375rem;
            color: var(--text-muted);
            margin-bottom: 0.375rem;
        }

        .eval-content code {
            font-family: 'JetBrains Mono', monospace;
            background: var(--bg-card);
            padding: 0.125rem 0.375rem;
            border-radius: 0.25rem;
            font-size: 0.8125rem;
            color: var(--accent);
        }

        /* Warning callout */
        .callout {
            padding: 1rem 1.25rem;
            border-radius: 0.5rem;
            margin: 1.25rem 0;
        }

        .callout-warning {
            background: rgba(210, 153, 34, 0.1);
            border: 1px solid rgba(210, 153, 34, 0.3);
            border-left: 4px solid var(--warning);
        }

        .callout-info {
            background: rgba(88, 166, 255, 0.1);
            border: 1px solid rgba(88, 166, 255, 0.3);
            border-left: 4px solid var(--accent);
        }

        .callout-title {
            font-weight: 600;
            font-size: 0.9375rem;
            margin-bottom: 0.375rem;
            display: flex;
            align-items: center;
            gap: 0.5rem;
        }

        .callout-warning .callout-title {
            color: var(--warning);
        }

        .callout-info .callout-title {
            color: var(--accent);
        }

        .callout p {
            margin-bottom: 0;
        }

        /* Quote block */
        .quote-block {
            background: var(--bg-card);
            border-left: 4px solid var(--accent);
            padding: 1rem 1.25rem;
            margin: 1.25rem 0;
            border-radius: 0 0.5rem 0.5rem 0;
        }

        .quote-text {
            font-size: 1rem;
            font-style: italic;
            color: var(--text);
            margin-bottom: 0.375rem;
        }

        .quote-author {
            font-size: 0.8125rem;
            color: var(--text-muted);
        }

        /* Resource cards */
        .resource-grid {
            display: grid;
            grid-template-columns: repeat(auto-fit, minmax(200px, 1fr));
            gap: 0.875rem;
            margin: 1.25rem 0;
        }

        .resource-card {
            display: flex;
            flex-direction: column;
            padding: 1rem;
            background: var(--bg-card);
            border: 1px solid var(--border);
            border-radius: 0.5rem;
            text-decoration: none;
            transition: all 0.2s;
        }

        .resource-card:hover {
            border-color: var(--accent);
            transform: translateY(-2px);
        }

        .resource-card-title {
            font-weight: 600;
            font-size: 0.9375rem;
            color: var(--text);
            margin-bottom: 0.25rem;
        }

        .resource-card-desc {
            font-size: 0.8125rem;
            color: var(--text-muted);
            line-height: 1.5;
        }

        .resource-card-arrow {
            margin-top: auto;
            padding-top: 0.625rem;
            font-size: 0.8125rem;
            color: var(--accent);
        }

        /* Disclaimer box */
        .disclaimer-box {
            background: var(--bg-card);
            border: 1px solid var(--border);
            border-radius: 0.5rem;
            padding: 1rem 1.25rem;
            margin: 1.5rem 0;
        }

        .disclaimer-title {
            font-weight: 600;
            font-size: 0.75rem;
            color: var(--text-muted);
            text-transform: uppercase;
            letter-spacing: 0.05em;
            margin-bottom: 0.375rem;
        }

        .disclaimer-text {
            font-size: 0.8125rem;
            color: var(--text-muted);
            line-height: 1.6;
        }

        /* Adversarial examples list */
        .adversarial-list {
            background: var(--bg-card);
            border: 1px solid var(--border);
            border-radius: 0.5rem;
            padding: 1rem 1.25rem;
            margin: 0.875rem 0;
        }

        .adversarial-list li {
            padding: 0.5rem 0;
            border-bottom: 1px solid var(--border);
            font-size: 0.875rem;
        }

        .adversarial-list li:last-child {
            border-bottom: none;
        }

        /* Shoutout section */
        .shoutout {
            margin-top: 2rem;
            padding-top: 1rem;
            border-top: 1px solid var(--border);
            font-size: 0.8125rem;
            color: var(--text-muted);
        }

        .shoutout strong {
            color: var(--text);
        }
    </style>
</head>
<body class="eval-body">
    <!-- Header -->
    <header class="header">
        <a href="index.html" class="logo">
            <img src="assets/KernelBenchMascot-removebg-preview.png" alt="KernelBench" class="logo-img">
            <span>KernelBench</span>
        </a>
        <nav class="nav">
<a href="index.html" class="nav-link">Home</a>
                <a href="leaderboard.html" class="nav-link">Leaderboard</a>
                <a href="kernels.html" class="nav-link">Kernels</a>
                <a href="models.html" class="nav-link">Models</a>
                <a href="eval.html" class="nav-link active">Eval</a>
                <a href="submit.html" class="nav-link">Submit</a>
            <div class="nav-dropdown">
                <button class="nav-dropdown-toggle">
                    About <span class="nav-dropdown-arrow">▼</span>
                </button>
                <div class="nav-dropdown-menu">
                    <a href="https://arxiv.org/abs/2502.10517" class="nav-dropdown-item" target="_blank">
                        <span class="nav-dropdown-item-icon">📄</span>
                        Paper
                        <span class="nav-dropdown-item-arrow">↗</span>
                    </a>
                    <a href="https://github.com/ScalingIntelligence/KernelBench" class="nav-dropdown-item" target="_blank">
                        <span class="nav-dropdown-item-icon">💻</span>
                        GitHub
                        <span class="nav-dropdown-item-arrow">↗</span>
                    </a>
                </div>
            </div>
            <button class="theme-toggle" id="theme-toggle" title="Toggle theme">
                <span class="theme-icon"></span>
            </button>
        </nav>
    </header>

    <!-- Main Content -->
    <main class="eval-container">
        <div class="eval-header">
            <div class="eval-badge">
                <span>🚧</span>
                <span>Work in Progress</span>
            </div>
            <h1 class="eval-title">Evaluation Guide</h1>
            <p class="eval-subtitle">Notes on benchmarking, correctness checking, and avoiding reward hacking.</p>
        </div>

        <div class="eval-content">
            <div class="callout callout-warning">
                <div class="callout-title">
                    <span>⚠️</span>
                    <span>Be Extra Careful!</span>
                </div>
                <p>Always be paranoid about suspiciously good results — kernel engineers and existing compilers are already pretty good, so a <strong>>2x speedup</strong> for anything is highly unlikely.</p>
            </div>

            <div class="quote-block">
                <div class="quote-text">"If you beat cuDNN by more than 10%, think again."</div>
                <div class="quote-author">— <a href="https://x.com/itsclivetime/status/1992155951630307633" target="_blank">@itsclivetime</a></div>
            </div>

            <p>If the model can reward hack, it will find ways to reward hack! This can especially happen during RL training or evolutionary search.</p>

            <h3>Recommended Resources</h3>
            <div class="resource-grid">
                <a href="https://scalingintelligence.stanford.edu/blogs/kernelbenchv01/" class="resource-card" target="_blank">
                    <div class="resource-card-title">KernelBench v0.1 Release</div>
                    <div class="resource-card-desc">Official release blog post with evaluation methodology updates.</div>
                    <div class="resource-card-arrow">Read more →</div>
                </a>
                <a href="https://arxiv.org/abs/2507.11948" class="resource-card" target="_blank">
                    <div class="resource-card-title">Kevin Project</div>
                    <div class="resource-card-desc">Cognition and Stanford's research on various hacking behaviors observed in RL training.</div>
                    <div class="resource-card-arrow">Read paper →</div>
                </a>
                <a href="https://deep-reinforce.com/defense_kernel_hack.html" class="resource-card" target="_blank">
                    <div class="resource-card-title">Hacks and Defenses</div>
                    <div class="resource-card-desc">Jiwei Li's blogpost on defense strategies for automatic GPU kernel generation.</div>
                    <div class="resource-card-arrow">Read more →</div>
                </a>
                <a href="https://deep-reinforce.com/correctness_check.html" class="resource-card" target="_blank">
                    <div class="resource-card-title">Correctness Checking</div>
                    <div class="resource-card-desc">Detailed guide on verifying kernel correctness and common pitfalls.</div>
                    <div class="resource-card-arrow">Read more →</div>
                </a>
            </div>

            <div class="disclaimer-box">
                <div class="disclaimer-title">Disclaimer</div>
                <div class="disclaimer-text">
                    KernelBench is an open-source evaluation framework. Due to limited bandwidth, the KernelBench team does not inspect, validate, or endorse any third-party kernels or reported results. Users are welcome to use the software infrastructure for evaluation, but should independently verify all results.
                </div>
            </div>

            <h2>Methodology</h2>
            <p>To ensure <strong>consistency and reproducibility</strong>, we recommend using <code>modal</code>. We have provided (and are adding more) various Modal cloud functions to standardize the evaluation environment.</p>

            <h3>Correctness</h3>
            <p>Correctness verification is critical for valid benchmarking. We also want to highlight community efforts such as <a href="https://www.youtube.com/watch?v=BTfjdyZOKww" target="_blank">BackendBench</a>.</p>

            <h3>Performance</h3>
            <p>We highly recommend watching this <a href="https://www.youtube.com/watch?v=1i7dxoAfKOU" target="_blank">lecture from GPU Mode</a> on kernel profiling.</p>

            <p>We have (and continue to) implement various approaches to conduct kernel timing to understand the tradeoffs. Check out <code>timing.py</code> to see available timing methods and <code>src/unit_tests/test_eval_timing.py</code> to test various timing methods including:</p>
            <ul>
                <li>CUDA event markers (<code>cuda_event</code>)</li>
                <li>Triton's <code>do_bench</code></li>
                <li>Host time end-to-end timing (<code>host_time</code>)</li>
            </ul>

            <h2>Unit Tests with Adversarial Examples</h2>
            <p>We've included unit tests for the eval script in <code>src/unit_tests/test_eval_adversarial.py</code>. These tests run adversarial kernels (see <code>src/unit_tests/test_kernels/</code>) that contain examples of reward hacking that we've seen from LLMs and ensures that the eval script catches them.</p>

            <div class="callout callout-info">
                <div class="callout-title">
                    <span>🔍</span>
                    <span>Known Adversarial Patterns</span>
                </div>
                <p>Our tests catch these behaviors by failing correctness checks or flagging excessive speedups:</p>
            </div>

            <ul class="adversarial-list">
                <li><strong>Computation caching:</strong> Reusing computations cached during the PyTorch reference execution</li>
                <li><strong>Input modification:</strong> Modifying inputs to cheat correctness checks</li>
                <li><strong>Stream manipulation:</strong> Moving computation to a non-default CUDA stream</li>
            </ul>

            <p>We will continue to add more tests as we explore additional adversarial scenarios.</p>

            <div class="callout callout-info">
                <div class="callout-title">
                    <span>💡</span>
                    <span>Help Us Improve</span>
                </div>
                <p>KernelBench is an ongoing open-source effort — please help us with <a href="https://github.com/ScalingIntelligence/KernelBench/issues" target="_blank">issues</a> and <a href="https://github.com/ScalingIntelligence/KernelBench/pulls" target="_blank">PRs</a>!</p>
            </div>

            <div class="shoutout">
                <strong>Shoutout</strong> to @bkal01, @palic, @miru_why, @ngc92, @itsclivetime for their suggestions and feedback.
            </div>
        </div>
    </main>

    <!-- Footer -->
    <footer class="footer">
        <div class="footer-content">
            <span class="footer-text">KernelBench Leaderboard</span>
        </div>
    </footer>

    <script src="js/common.js"></script>
</body>
</html>