stormlog.attributed_viz

Stormlog-native memory visualisation with tensor attribution.

Generates a self-contained HTML page that renders an interactive CUDA memory timeline where every allocation is labelled with its tensor name, shape, and allocation site — replacing the opaque hex-address chart from PyTorch’s built-in _memory_viz.trace_plot().

Usage:

from stormlog.attributed_viz import render_attributed_html

html = render_attributed_html(snapshot_dict, tensor_index)
Path("memory_attributed.html").write_text(html)

Functions

render_attributed_html(snapshot, tensor_index, *)

Generate a self-contained HTML page with attributed memory timeline.

render_attributed_wandb_preview_html(...[, ...])

stormlog.attributed_viz.render_attributed_wandb_preview_html(snapshot, tensor_index, *, device=0, max_timeline_points=480, max_marker_points=80, max_offenders=12, max_active_rows=12)[source]
Parameters:

  • snapshot (Any)

  • tensor_index (Dict[str, Any])

  • device (int)

  • max_timeline_points (int)

  • max_marker_points (int)

  • max_offenders (int)

  • max_active_rows (int)

Return type:

str

stormlog.attributed_viz.render_attributed_html(snapshot, tensor_index, *, device=0)[source]

Generate a self-contained HTML page with attributed memory timeline.

Parameters:

  • snapshot (Any) – The raw dict from torch.cuda.memory._snapshot() or loaded from the pickle file. Must have "segments" and "device_traces".

  • tensor_index (Dict[str, Any]) – The dict from build_cuda_tensor_attribution_index().

  • device (int) – The CUDA device index to render.

Returns:

A complete HTML document string.

Return type:

str