Spaces:
Running
Running
Update index.html
Browse files- index.html +50 -1
index.html
CHANGED
|
@@ -18,6 +18,54 @@
|
|
| 18 |
<p class="subtitle">GPU Placement Advisor for PyTorch/NCCL Workloads</p>
|
| 19 |
</header>
|
| 20 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 21 |
<!-- Step 1: Authentication -->
|
| 22 |
<section class="card">
|
| 23 |
<div class="card-title">
|
|
@@ -141,4 +189,5 @@
|
|
| 141 |
|
| 142 |
<script src="app.js"></script>
|
| 143 |
</body>
|
| 144 |
-
</html>
|
|
|
|
|
|
| 18 |
<p class="subtitle">GPU Placement Advisor for PyTorch/NCCL Workloads</p>
|
| 19 |
</header>
|
| 20 |
|
| 21 |
+
<!-- How It Works (collapsible) -->
|
| 22 |
+
<details class="help-section">
|
| 23 |
+
<summary class="help-toggle">📖 How It Works</summary>
|
| 24 |
+
<div class="help-content">
|
| 25 |
+
<div class="help-block">
|
| 26 |
+
<h3>What is Dystrio?</h3>
|
| 27 |
+
<p>Dystrio analyzes your PyTorch distributed training communication patterns and generates
|
| 28 |
+
Kubernetes pod affinity rules to co-locate GPUs that talk the most.</p>
|
| 29 |
+
</div>
|
| 30 |
+
|
| 31 |
+
<div class="help-block">
|
| 32 |
+
<h3>How do I get a PyTorch trace?</h3>
|
| 33 |
+
<p>Add this to your training script:</p>
|
| 34 |
+
<pre><code>from torch.profiler import profile, ProfilerActivity
|
| 35 |
+
|
| 36 |
+
with profile(
|
| 37 |
+
activities=[ProfilerActivity.CPU, ProfilerActivity.CUDA],
|
| 38 |
+
record_shapes=True,
|
| 39 |
+
with_stack=True
|
| 40 |
+
) as prof:
|
| 41 |
+
# Your training step here
|
| 42 |
+
model(inputs)
|
| 43 |
+
|
| 44 |
+
prof.export_chrome_trace("trace.json")</code></pre>
|
| 45 |
+
<p>Upload the resulting <code>trace.json</code> file.</p>
|
| 46 |
+
</div>
|
| 47 |
+
|
| 48 |
+
<div class="help-block">
|
| 49 |
+
<h3>What is Session ID / Multi-Run?</h3>
|
| 50 |
+
<p><strong>Single run:</strong> Leave Session ID empty. You'll get recommendations based on one trace.</p>
|
| 51 |
+
<p><strong>Multi-run (recommended):</strong> Use the same Session ID across multiple uploads.
|
| 52 |
+
Dystrio tracks which communication patterns are <em>stable</em> vs <em>noisy</em>,
|
| 53 |
+
giving you higher-confidence recommendations.</p>
|
| 54 |
+
<p>Example: Upload 3 traces from different training runs with Session ID "llama-70b-training"
|
| 55 |
+
→ Dystrio identifies consistent patterns and escalates confidence from LOW → HIGH.</p>
|
| 56 |
+
</div>
|
| 57 |
+
|
| 58 |
+
<div class="help-block">
|
| 59 |
+
<h3>How do I use the output?</h3>
|
| 60 |
+
<ol>
|
| 61 |
+
<li>Copy the generated Kubernetes YAML</li>
|
| 62 |
+
<li>Add the <code>affinity:</code> block to your Pod spec</li>
|
| 63 |
+
<li>Deploy – Kubernetes will schedule communicating pods together</li>
|
| 64 |
+
</ol>
|
| 65 |
+
</div>
|
| 66 |
+
</div>
|
| 67 |
+
</details>
|
| 68 |
+
|
| 69 |
<!-- Step 1: Authentication -->
|
| 70 |
<section class="card">
|
| 71 |
<div class="card-title">
|
|
|
|
| 189 |
|
| 190 |
<script src="app.js"></script>
|
| 191 |
</body>
|
| 192 |
+
</html>
|
| 193 |
+
|