Download

Overview

Interpretability is now production practice: steering behavior, probing activations, feature-level debugging, and comparing checkpoints. All of these need programmable read/write access to internal state, mid-inference, at frontier scale. Serving engines transform execution for throughput (batching, sharding, fused kernels, graph capture), interpretability tools assume one process, one module tree, and one visible forward call, and fixed intervention APIs cover static workloads only.

Model Intervention as a Service is an intervention layer over production engines. It exposes a stable logical view of model internals (activations, logits, generation steps, cross-prompt state) while the engine keeps serving, making them inspectable and programmable. In our design, built on NNsight and vLLM, intervention programs attach to inference requests and their logical reads and writes bind to vLLM’s existing execution path:

  • Engine-driven. Intervention code rides vLLM’s normal scheduling and re-enters at phase boundaries: model execution, logits, sampling, and result collection.
  • Distributed. The layer reconstructs tensor-parallel values for user code and scatters edits back to ranks. Under pipeline parallelism, layers missing on a rank resolve to lazy stubs, and the real tensor is pulled from its owning rank only when dereferenced.
  • Optimized. Continuous batching is preserved by translating between request-level logical tokens and the engine’s flat physical buffers.