[tivris]

Description

Fixes #1927.

GatedMLPBridge.set_processed_weights stored the processed weights and biases twice: in the gate, in and out projections, and in six plain attributes that the processed forward read. Conversions and copy.deepcopy update only the projections' parameters, so the forward read stale tensors.

The processed forward now reads each projection's current weight and bias, detached as before, and the six attributes are gone. JointGateUpMLPBridge now checks only the processed-weights flag. Outputs, hook order and gradients are unchanged when nothing is converted. The hook test's fixture now installs its weights through set_processed_weights instead of setting the attributes.

Type of change

• Bug fix (non-breaking change which fixes an issue)

Checklist:

• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes
• I have not rewritten tests relating to key interfaces which would affect backward compatibility

[jlarson4]

This PR is approved, just waiting for CI to complete before I can merge. Going to rerun BackwardLens once the job is done, I have a fix for that pending but this PR should be ready before that is merged.