Skip to content
Docs
CLI and Automation
Batch inspection

Batch Inspection

Anything you'd do to one checkpoint interactively - index it, query it, check its statistics - often needs doing across many checkpoints at once: every checkpoint in a training run's history, every candidate in a sweep, every model in a shared cache. The tmorph CLI is built for exactly this, since every command operates on a path and produces script-consumable output.

Indexing many checkpoints ahead of time

for ckpt in checkpoints/*.safetensors; do
  tmorph index "$ckpt" --pyramids
done

Pre-indexing a batch this way means opening any of them later - interactively or from another script - skips the indexing cost entirely; see File Formats for why .tmidx/.tmtiles are safe to generate ahead of time and cache.

Querying across a batch

Because tmorph query accepts the same TQL predicates used interactively, a question like "which of these checkpoints has any NaN" or "which layer's absmax grew the most across this sweep" is a loop over tmorph query calls, with results aggregated in whatever scripting language you're already using - not a question that requires opening each checkpoint in the desktop app one at a time.

Turning a batch scan into a report

Combined with Headless Rendering, a batch inspection script can produce a visual report (heatmap thumbnails, distribution plots) for every checkpoint in a batch without any interactive session - useful for a nightly sweep summary, or for attaching evidence to an automated flag rather than just a pass/fail signal.

Related resources

lob CLICI validation