Quick Start

This guide gets you up and running with nnterp in 5 minutes.

Installation

Basic Installation:

pip install nnterp

For visualization:

pip install nnterp[display]

Your First Model

Load any transformer model with a unified interface:

from nnterp import StandardizedTransformer

model = StandardizedTransformer("Maykeye/TinyLLama-v0")
print("✓ nnterp working!")

Basic Activation Access

Access model internals:

with model.trace("Hello world"):
    # Access layer outputs
    layer_5_output = model.layers_output[5]
    logits = model.logits.save()

# Access attention and MLP components:
with model.trace("Hello world"):
    attention_output = model.attentions_output[3]
    mlp_output = model.mlps_output[3]

Direct Assignment

Set layer outputs directly:

with model.trace("The capital of France is"):
    # Sum layers 3 and 5, set layer 10 to this value
    combined = model.layers_output[3] + model.layers_output[5]
    mlp_8 = model.mlps_output[8]  # collect for later
    model.layers_output[10] = combined

    # Add MLP output from layer 8 to layer 12 output
    model.layers_output[12] = model.layers_output[12] + mlp_8

    result = model.logits.save()

Next Steps

• Basic Usage - Learn the unified interface and module access

• Interventions - Logit lens, patchscope, and steering methods