Skip to content

Activation Transform Persistence

TorchLens persistence distinguishes portable recipe state from executable Python callables.

Runtime Behavior

  • activation_transform and gradient_postfunc can be ordinary Python callables during capture.
  • Callable outputs can be saved as tensor payloads when the selected save level stores them.
  • The callable object itself is not made portable unless it is represented by a supported built-in helper or import reference.

Portable Save Policy

For portable .tlspec/ saves:

Item Portable behavior
Built-in helpers Persisted as structured recipe data.
Tensor values produced by transforms Persisted when included by the save level.
Opaque Python callables Dropped from portable recipe execution.
Callable display text repr(callable) is retained for audit/readback context when available.

This is the callable-drop / repr-keep behavior: a portable artifact should not execute unknown Python code, but it should still explain which callable was present at capture time.

Audit or Local Execution

Use an audit or executable-with-callables level only when the receiving environment is trusted and has the same callable definitions importable. Portable artifacts are the default for sharing across machines or repositories.