Skip to content
Docs
Reference
File formats

File Formats

Tensormorph reads a wide range of checkpoint formats (safetensors, ONNX, sharded checkpoints, and more - see Working with Models), but it also has its own small family of file formats for the data it derives and produces itself. This page is the reference for those.

ExtensionContents
.tmidxThe indexed tensor table for a checkpoint: the Semantic Model Graph plus precomputed tile statistics for every tensor. Produced by tmorph index.
.tmtilesValue-level tile data at multiple levels of detail (mip-style pyramids) - the actual sampled/aggregated values behind level-of-detail streaming, separate from the statistics in .tmidx.
.tmpairPrecomputed statistics for an aligned A/B pair of models - the artifact a weight diff or comparison is built from, so re-opening a comparison doesn't require realigning and re-diffing from scratch.
.tmrunA captured runtime profile - the artifact behind Runtime Traces, recorded from a RuntimeProbe session or replayed from a live run.
.tmprojProject state: bookmarks, notes, saved TQL queries, and workspace layouts. This is the file that makes a Tensormorph session resumable and shareable, independent of the checkpoint(s) it references.

Index files are derived, not authoritative

💡

Every format above except .tmproj is a derived cache, not a source of truth. A .tmidx/.tmtiles/.tmpair/.tmrun file can always be regenerated from the checkpoint (and, for .tmrun, the run) it was built from. Deleting one costs you re-indexing time, not data.

This is why tmorph index (see Lob CLI) is a safe, idempotent operation to re-run, and why these formats are reasonable to exclude from version control - unlike the checkpoint itself, they're not something anyone needs to review a diff of.

Producing these files without the app

Each format is producible from the CLI, not only from an interactive session - see Batch Inspection and CI Validation for indexing and diffing a checkpoint as part of a script or pipeline.

Related resources