Move main() execution guard for Sphinx Gallery

Move `if __name__ == "__main__": main()` to immediately after the
main() function definition (line ~404) so it executes during the
Sphinx Gallery build process.

Sphinx Gallery requires the execution guard to be positioned right
after the function definition, not at the end of the file, to properly
capture and execute the tutorial code during documentation generation.

Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>

Author: yushangdi

CUDA_GRAPH_TUTORIAL_README.md

@@ -0,0 +1,78 @@
+# CUDA Graph Annotations Tutorial
+
+## Overview
+
+A new tutorial has been added to the PyTorch tutorials repository that demonstrates how to use CUDA graph kernel annotations for enhanced profiling and visualization.
+
+## File Location
+
+- **Tutorial file**: `advanced_source/cuda_graph_annotations_tutorial.py`
+- **Added to index**: `index.rst` (line ~518)
+- **Added to deep-dive**: `deep-dive.rst` (profiling section)
+
+## Tutorial Content
+
+The tutorial covers:
+
+1. **Introduction to CUDA Graph Annotations**: Why they're useful for profiling complex graph executions
+2. **Building an Example Model**: A simple transformer block with multiple annotated regions
+3. **Using `mark_kernels()`**: The key API for annotating kernel regions with semantic labels
+4. **Graph Capture with Annotations**: How to enable annotations during CUDA graph capture
+5. **Profiling**: Recording execution traces of graph replays
+6. **Post-Processing**: Merging annotations back into traces for custom visualization lanes
+7. **Visualization**: How to view annotated traces in chrome://tracing
+8. **Troubleshooting**: Common issues and solutions
+
+## Key Features
+
+- **Complete end-to-end workflow**: From model definition to visualization
+- **Practical example**: Uses a realistic transformer block
+- **Custom stream assignments**: Shows how to organize kernels into semantic lanes
+- **Before/after comparison**: Demonstrates the value of annotations
+- **Comprehensive documentation**: Includes requirements, troubleshooting, and advanced usage
+
+## Tutorial Style
+
+The tutorial follows PyTorch tutorials conventions:
+- Uses Sphinx Gallery format with docstring sections
+- Includes grid cards for learning objectives and prerequisites
+- Code is well-commented and organized into logical sections
+- Provides practical, runnable examples
+- Explains both the "how" and "why"
+
+## Integration
+
+The tutorial has been integrated into:
+1. Main index (`index.rst`) under Model Optimization/Profiling
+2. Deep Dive section (`deep-dive.rst`) alongside other profiling tutorials
+3. Tagged appropriately: `Model-Optimization`, `Best-Practice`, `Profiling`, `CUDA`
+
+## Building
+
+To build just this tutorial:
+```bash
+cd tutorials
+GALLERY_PATTERN="cuda_graph_annotations_tutorial.py" make html
+```
+
+To build all tutorials:
+```bash
+cd tutorials
+make docs
+```
+
+## Requirements
+
+- PyTorch 2.0+
+- CUDA-capable GPU
+- Driver/CUDA-compat >= 13.1 for full annotation support
+- cuda-python package (`pip install cuda-python`)
+
+Note: The tutorial gracefully handles older drivers by detecting annotation support and providing appropriate messages.
+
+## Next Steps
+
+1. Test the full build in CI
+2. Add a thumbnail image to `_static/img/thumbnails/cropped/`
+3. Verify the tutorial renders correctly in the documentation site
+4. Consider adding more advanced examples (multi-GPU, distributed training)

Screenshot 2026-06-02 at 4.55.40 PM.png

@@ -402,6 +402,9 @@ def main():
     print("the semantic kernel lanes.")
     print("="*60)
 
+if __name__ == "__main__":
+    main()
+
 ###############################################################################
 # Visualizing Results
 # -------------------
@@ -490,6 +493,3 @@ def main():
 # This technique is especially valuable for large models with many components,
 # distributed training setups, or any scenario where understanding the
 # execution structure is critical for performance optimization.
-
-if __name__ == "__main__":
-    main()