nnsight-remote-interpretability

Inspect PyTorch model activations by using the LanguageModel tracing API to start a deferred trace, access layer outputs via proxy objects, and save the captured activation data for analysis.

Yes, you can run interpretability tracing on remote 70B+ models by enabling remote execution with NDIF, which allows you to capture activations and gradients from large-scale deployments without local hardware.

Cross-prompt activation patching is a causal analysis technique that shares and modifies captured layer activations across different inputs to determine specific component influence on model behavior.

Yes, nnsight supports saving gradients alongside layer activations for causal analysis, allowing researchers to inspect proxy objects and trace gradient flow across local or remote PyTorch models.

Trace neural network internals locally by initializing the LanguageModel tracing API on your PyTorch model, applying deferred tracing to capture layer outputs, and executing the intervention directly on your hardware.

No, you do not need NDIF to use the LanguageModel tracing API; remote execution with NDIF is optional and primarily used for running interpretability interventions on large-scale 70B+ models.