Export Transformed Models
Once a morph is materialized, or a checkpoint has otherwise been transformed within Tensormorph, exporting writes the result out as an ordinary checkpoint - indistinguishable, from the outside, from any other model in that format.
What gets exported
A materialized morph result is a full set of weights, not a diff or a recipe - exporting it produces a complete, standalone checkpoint in your chosen output format (safetensors or ONNX, most commonly), loadable by any tool that reads that format, with no dependency on Tensormorph or on the recipe that produced it.
The recipe that produced an export - the morph DAG itself - is saved separately as recipe.json (see Model Morphing). Keeping both means the exported weights are portable on their own, while the recipe remains available for re-materializing against an updated base or for audit purposes.
Format-specific considerations
- To safetensors - the simpler target; every tensor in the reassembled result is written out flat, with no graph structure to reconcile.
- To ONNX - because ONNX tensors are always dense, with no stride concept, any view relationships present in the working representation are resolved into independent copies on export; there's no ONNX-side equivalent of a Tensormorph view to preserve.
Provenance on export
Where the target format supports it, export carries forward whatever provenance metadata is available - see Minimal vs. rich metadata for ONNX's metadata_props, or a Tensor Card for a richer, structured equivalent - so an exported model doesn't lose the record of what produced it just because it's left Tensormorph's own project format.
Related resources
- Model Morphing - the materialize step this page's export follows.
- Safetensors and ONNX - the two primary export targets.
- File Formats - where
recipe.jsonfits alongside an export.