Hello, thank you for your contribution. To maintain consistency across our repositories, we have reimplemented this in PR #707.

Hello, thank you for your contribution. To maintain consistency across our repositories, we have reimplemented this in PR #707.

Thanks for reviewing #694 and adapting the design into #707. I reviewed the merged implementation, and I understand the direction: #707 generalized the worker-side activation-add hook into EasyEdit2’s apply/inference stage for multiple activation-addition methods, while keeping vector generation in-process.

I would be happy to contribute follow-up improvements on top of #707, especially:

1. Robustness improvements for the merged vLLM hook path:

   • preserve worker-side exception details instead of returning only False
   • make missing layer installation fail more explicitly
   • clarify the distinction between in-process vLLM and worker-hook/multiprocessing vLLM
   • add lightweight regression tests for steer/vllm_hooks.py

2. Documentation/config updates:

   • update the EasyEdit2 vLLM tutorial/config examples to show the new multiprocessing options
   • clarify the recommended generate-then-apply workflow
   • document that vector generation remains in-process while apply/inference can use worker-side hooks

I also want to raise one attribution point. Since #694 was the original prototype/design proposal for the worker-side CAA/vLLM hook, and the maintainers said its logic was manually adapted into #707, I think this contribution should be properly recognized if it materially contributed to the core merged method. I am not asking to rewrite history, but I would appreciate a short acknowledgement in #707, #694, or the relevant vLLM documentation.

Thank you for raising this issue and helping us identify the problem. We appreciate your contribution and will acknowledge your support in the contributor section. You’re also welcome to submit additional code in the future.

For code contributions, our main requirements are that they align with the existing framework, maintain simplicity, and avoid introducing overly AI-specific code.