MemoryViz is a visualization tool adapted from torch/cuda/_memory_viz.py. An official hosted service is available at https://docs.pytorch.org/memory_viz. For more details about the original tool, please see its blog: Understanding GPU Memory 1: Visualizing All Allocations over Time or the API document Understanding CUDA Memory Usage.
- Provides a simple HTTP server for convenient previewing, especially useful for SSH remote development.
- Interactive HTML-based list view to replace legacy plain text displays.
- Enhanced Python frame visibility by prioritizing user-level code and applying semantic highlighting.
- Support for Matplotlib-style colormaps to improve visual memory differentiation.
- Enabled browser-based source code linking for efficient navigation to file paths.
- Integrated regex-based "Important Frame" tagging to automatically identify core PyTorch operators (e.g., einsum, mul).
- Optimized timeline visualization with binary-aligned memory axis (GiB/MiB) and improved UI padding.
