detach()

Remove MemScale optimization from a model, restoring it to its original state.

Signature

memscale.detach(model)

Parameters

Returns

None.

Behavior

detach() looks for the executor handle that wrap() stored on the model (model._memscale_executor):

• If present — the executor detaches all hooks, the handle is removed, and the model returns to its unoptimized state. An info message is logged.
• If absent — the model was never wrapped. detach() does nothing and logs a warning. It does not raise.

Example

import memscale
 
model = memscale.wrap(model)
# ... train ...
 
memscale.detach(model)   # model is now back to its original form

When you do not need it

If you used the optimize() context manager, detachment happens automatically when the with block exits — calling detach() yourself is unnecessary in that case.

detach() is useful when you wrapped a model with wrap() and later want to run it — for inference or evaluation — without the training-time memory hooks attached.