Spaces:

---

NousResearch

/

cna-refusal-ablation

like 0

Running on L40S

core.py (Dynamic VRAM Safety): Enclosed the core mechanistic interpretability hooks and activation-steering logic within strict try...finally resource execution boundaries. This guarantees that intermediate Multi-Layer Perceptron (MLP) tensors are explicitly deallocated from GPU memory, preventing unhandled runtime interruptions from causing persistent CUDA Out-Of-Memory (OOM) leaks.