TutorialStablesince v0.6.0

Activation Basics & System Telemetry

Understanding how probes capture intermediate layer activations, attention heads, hidden states, tool execution payloads, and telemetry streams.

Audience: AI Engineers & ResearchersRead: 12 minEdit on GitHub

In NeuronScope, an activation or telemetry trace is the captured tensor state, attention weights, prompt context, or tool payload emitted by an intelligent system component during execution.

What is System Telemetry?

Telemetry spans multiple layers of the application hierarchy:

  • Layer Activations: Output tensors from transformer encoder/decoder blocks ([batch, tokens, hidden_dim]).
  • Attention Heads: Multi-head self-attention matrices ([batch, heads, tokens, tokens]).
  • RAG Context & Vectors: Retrieved document chunks and similarity distance scores from vector databases.
  • Agent Tracing: Reasoning step arguments, tool call parameters, and return payloads.

Module & Layer Selectors

Selectors are glob patterns matched against canonical module paths:

select_probes.py
python
probes = [
    ns.probes.Activations("model.layers.*.self_attn.o_proj"),
    ns.probes.Attention("**.self_attn"),
    ns.probes.ToolCalls("agent.tools.*")
]

Tensor Shapes & Native Precision

  • Activations maintain the model's native precision (float16, bfloat16, or float32).
  • Tensors are streaming into lock-free ring buffers without blocking main GPU execution loops.

Common Observation Pitfalls

Related