CHANGELOG.md

Changelog

0.3.0

released: 2024-08-29

We are excited to announce the release of nnsight0.3.

This version significantly enhances the library's remote execution capabilities. It improves the integration experience with the NDIF backend and allows users to define and execute optimized training loop workflows directly on the remote server, including LoRA and other PEFT methods.

Breaking Changes

• Module input access has a syntactic change:

  • Old: nnsight.Envoy.input
  • New: nnsight.Envoy.inputs
  • Note: nnsight.Envoy.input now provides access to the first positional argument of the module's input.

• scan & validate are set to False by default in the Tracer context.

New Features

• Session context: efficiently package multi-tracing experiments into a single request, enabling faster, more scalable remote experimentation.

• Iterator context: define an intervention loop for iterative execution.

• Model editing: alter a model by setting default edits and interventions in an editing context, applied before each forward pass.

• Early stopping: interrup a model's forward pass at a chosen module before execution completes.

• Conditional context: define interventions within a Conditional context, executed only when the specified condition evaluates to be True.

• Scanning context: perform exclusive model scanning to gather important insights.

• nnsight builtins: define traceable Python builtins as part of the intervention graph.

• Proxy update: assign new values to existing proxies.

• In-Trace logging: add log statements to be called during the intervention graph execution.

• Traceable function calls: make unsupported functions traceable by the intervention graph. Note that all pytorch functions are now traceable by nnsight by default.