Skip to content
Docs
Core Concepts
Level of detail and streaming

Level of Detail and Streaming

Tensormorph is designed to open a checkpoint from a few megabytes to trillions of parameters without a difference in how it feels to use - the mechanism that makes that possible is level-of-detail (LOD) streaming, built directly on the mergeable tile statistics every tensor carries.

The LOD pyramid

A tensor's statistics exist at multiple levels of detail simultaneously - the same way an image pyramid stores an image at multiple resolutions - because coarser-level statistics are just merges of finer-level ones, not separately computed. Zooming out in the Architecture or Tensor Volume view doesn't mean "load less data and hope it's representative"; it means reading a coarser, already-correct level of the same pyramid. This pyramid - the value-level tile data across LOD levels, distinct from the summary statistics in .tmidx - is what .tmtiles stores; see File Formats.

What's resident where

💡

At any moment, only a small fraction of a large model's data is actually resident in fast memory - the tiles you're currently looking at, plus whatever the prefetcher believes you're about to look at. Everything else stays on slower storage until it's needed.

Tensormorph's compute scheduler prioritizes work - and by extension, what stays resident - using a fixed order: the scalar directly under your cursor first, then the current selection, then visible tiles, then near-visible tiles prefetched ahead of a likely pan or zoom, then explicit analysis requests, then queries, then background indexing of whatever's left. This ordering is why panning toward an area you haven't visited yet still feels responsive - the prefetch lane is already working before you arrive.

Streaming from remote sources

The same LOD principle extends to a checkpoint that hasn't been fully downloaded yet: a format like safetensors lets Tensormorph fetch just a header to populate the coarsest level (names, shapes, dtypes) before any weight bytes arrive, with finer levels streaming in only as you actually navigate toward them.

Indexing ahead of time

Interactive LOD streaming assumes an index already exists - tmorph index (optionally with --pyramids to precompute the full LOD pyramid rather than just the base statistics) builds the .tmidx/.tmtiles files a session streams from. See Batch Inspection for indexing a checkpoint ahead of an interactive session, which avoids paying the indexing cost the first time someone opens a very large model.

Related resources

Model morphingArchitecture Editor