Snapviewer

A PyTorch memory snapshot viewer alternative to https://docs.pytorch.org/memory_viz with rich features. Display large snapshots smoothly!

Tested on Windows and macOS.

Usage:

• You need Rust toolchain and Python installed.

• First you need a virtual environment (via venv or conda). Here we use conda:

  • Activate your environment

    conda activate base

  • Install dependencies

    pip install -r requirements.txt

• Compile binary

  cargo build --release --bin snapviewer-renderer --no-default-features

• Run

  -rr is for --resolution-ratio, used to deal with the rendering pattern of Apple's retina display. You probably need -rr 2.0 if you are using MacBook.

  Option A: Pass the .pickle directly. Preprocessing artifacts are cached at ~/.snapviewer_cache/ and reused on subsequent runs.

  python gui.py --pickle snap/large.pickle --res 1200 500 -rr 2.0

  Option B: Pre-process manually and pass the directory.

  # 1. Convert snapshot — outputs allocations.json and elements.db under the directory
  python convert_snap.py -i snap/large.pickle -o ./large
  
  # 2. Run
  python gui.py --dir ./large --res 1200 500 -rr 2.0

Warning: This software is in pre-alpha stage. Everything including snapshot format, data storing/loading logic is under frequent change.

Controls

• Pan: WASD / Left Mouse Drag
• Zoom: Mouse Wheel
• (Ctrl + Left click) on an allocation for detailed info about it

Troubleshoot

• If you see errors with message like cannot open input file 'sqlite3.lib', enable feature flag --features bundled-sqlite.

Notes

• Minimal dependency is not a goal.
• On macos, TKinter is required to run on main thread; while on all platforms the renderer is also required to run on main thread. This means we need multiple processes if we want to do cross platform.
• todo:
  • test this zmq-based impl on linux