nnsight — анализ внутренних состояний нейросетей

★ 7.7 · ai-tooling

nnsight is a Claude Code skill that provides guidance for interpreting and manipulating neural network internals using the nnsight library, with optional remote execution on massive models via NDIF. The skill covers the full workflow: wrapping models with LanguageModel, saving intermediate activations via `.save()` inside a trace context, patching layer outputs, and analyzing attention patterns — the same code runs locally on GPT-2 or remotely on Llama-3.1-405B by toggling `remote=True`. Dependencies are nnsight>=0.5.0 and torch>=2.0.0; remote NDIF access requires an API key from login.ndif.us. It is built for mechanistic interpretability researchers who need to run experiments on 70B+ models without local GPU resources, and for anyone working with arbitrary PyTorch architectures.