[fix] Honor explicit CrossEncoder activation overrides when loading

[mottopanikeiku]

I fixed CrossEncoder loading so an explicit activation_fn takes precedence over the saved activation. A caller can now replace a custom activation even when its saved class cannot be reconstructed.

Validation on CPU:

• The saved custom-activation reproduction fails before the fix and passes afterward.
• tests/cross_encoder/test_model.py: 75 passed, 8 skipped.
• Tiny BERT training, save/load, and prediction smoke checks pass.
• Ruff, formatting, and typos hooks pass.

I used AI assistance for the implementation and regression coverage. The checks above were run through that workflow.

[tomaarsen]

Hello!

Thanks for the fix and for the simplification!
I also updated predict() to check activation_fn is None, so callable overrides that evaluate to False, such as an empty nn.Sequential(), are respected. This seems like a nice improvement overall. I'll do another light review and then wait for the tests.

• Tom Aarsen

[Copilot]

🟢 Approval recommended

No unresolved review comments remain, and validation and regression tests pass.

Pull request overview

Updates CrossEncoder loading to honor explicit activation overrides, including falsey callables and unresolvable saved activations.

Changes:

• Preserves explicit activation overrides during config parsing.
• Corrects falsey per-call activation handling.
• Adds regression coverage for loading and prediction behavior.

File summaries

File	Summary
tests/cross_encoder/test_model.py	Adds activation loading and override regression tests.
sentence_transformers/cross_encoder/model.py	Implements activation precedence and override handling.

Review details

• Files reviewed: 2/2 changed files
• Comments generated: 0
• Review effort level: Lite

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.