Nnsight

NNsight is a Python library for interpreting and manipulating the internals of deep learning models. It gives researchers a Pythonic interface to read activations at any layer during a forward pass, modify those activations to study causal effects, and compute gradients with respect to intermediate values. It was developed by the NDIF team at Northeastern University. The library runs locally on any PyTorch model, with dedicated support for Hugging Face Transformers models and for vLLM. For models too large to run on a single machine, the same code can be executed remotely on infrastructure hosted by the National Deep Inference Fabric (NDIF) by adding a remote flag. Documentation covers batching interventions across many inputs, caching activations in bulk, persistent model editing, early stopping and skipping parts of execution, tensor parallelism across several GPUs, and quantization to 4 or 8 bits. It suits interpretability researchers and engineers who need transparent access to model internals that ordinary commercial inference APIs do not expose. The site includes tutorials on causal tracing, probing, attention, circuits and steering. Installation is a single pip command and the library is developed in the open on GitHub. The site does not publish prices, and it does not list any charges for the library itself or for remote execution on NDIF.