ndif-team/nnsight

Health v2 · maintenance only

Adoption not part of the score

1085 stars · 114 forks observed · 2026-09-03

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-30, confidence not recorded

nnsight is a Python library for interpreting and intervening on the internals of deep learning models, built on PyTorch. It lets researchers access, modify, and save activations and gradients during forward passes, locally on any PyTorch model or remotely on large models via the NDIF infrastructure.

Use cases

• access hidden states and activations from any layer of a neural network
• modify activations to study causal effects in language models
• run activation patching experiments on transformers
• interpret internals of large language models too big to run locally
• compute gradients with respect to intermediate model values
• batch interventions across multiple inputs efficiently
• trace model execution and save intermediate tensors for analysis

When to choose

• you need fine-grained access to model internals like activations and gradients during inference
• you want to run interpretability experiments on both small local models and huge remote models with the same code
• you are doing mechanistic interpretability research on PyTorch or HuggingFace models
• you need to intervene on model execution mid-forward-pass

When to avoid

• you only need standard inference or fine-tuning without inspecting internals
• you work with non-PyTorch frameworks like JAX or TensorFlow
• you need a no-code or GUI interpretability tool rather than a Python API

Facets

library · maturity active

machine-learning deep-learning llm-inference sdk machine-learning deep-learning large-language-models artificial-intelligence developer-tools python cross-platform interpretability activation-patching pytorch model-internals neural-network-intervention ndif remote-execution mechanistic-interpretability gpu

Member repositories

For agents

markdown · JSON · MCP: product_card(name="ndif-team/nnsight")