---

📖 Background & Motivation

Mechanistic interpretability aims to reverse-engineer neural networks from weights and activations into human-understandable circuits and computational graphs. However, researchers, students, and engineers frequently encounter significant friction:

• Notebook Fragmentation: Analysis scripts require hundreds of lines of bespoke PyTorch hooking, unembedding matrix projection, and Matplotlib code for every new model.
• Redundant VRAM Duplication: Existing tools often require reloading entire model checkpoints into separate processes, exhausting GPU memory on single-GPU workstations.
• Architecture Fragility: Minor differences between HuggingFace model layer names (e.g. model.layers, gpt_neox.layers, transformer.h) break ad-hoc analysis scripts.

InterpLens bridges this gap by providing an interactive visual debugger and real-time interpretability workbench. Built natively for PyTorch and TransformerLens, InterpLens attaches to pre-loaded models with zero memory duplication, dynamically resolves internal module topologies, and renders real-time telemetry across an ergonomic visual interface.

---

⚡ Key Features & Engines

1. 🔍 Logit Lens & Unembedding Projector

Projects intermediate layer residual stream states $x_l$ directly into vocabulary space via the unembedding matrix $W_U$:

$$\text{logits}l = W_U \cdot \text{LN}{\text{final}}(x_l) \quad \in \mathbb{R}^{V}$$

$$p_l(v) = \text{softmax}(\text{logits}_l) = \frac{\exp(\text{logits}l(v))}{\sum{v'} \exp(\text{logits}_l(v'))}$$

• Prediction Entropy Dynamics: $H(p_l) = -\sum_{v} p_l(v) \log p_l(v)$ tracks the layer where confidence crystallizes.
• KL Divergence Drift: $D_{KL}(p_L \parallel p_l)$ quantifies intermediate layer fidelity to the final output token distribution.

2. 🌊 Residual Stream Inspector

Tracks the geometry and drift of hidden state vectors as information propagates through the network:

• L2 Vector Energy: Monitors $|x_l|_2$ norm expansion across transformer depth.
• Cosine Drift Heatmap: $\cos(x_l, x_{l+1})$ identifies layers responsible for major semantic transitions.
• PCA Layer Trajectories: Visualizes residual stream dynamics projected onto top principal components.

3. 🕸️ Attention Head & Arc Explorer

• $N \times N$ Attention Matrices: Interactive heatmaps of query-key attention patterns across all layers and heads.
• Arc Diagram Visualization: Highlights long-range token-to-token information routing.
• Automated Induction Head Detector: Automatically detects prefix-matching $[A][B] \dots [A] \rightarrow [B]$ copying circuits across all attention heads.

4. ⚡ Neuron Activation & Token Attribution

• Top-K Firing Neurons: Identifies and ranks the highest-activating MLP neurons for any given prompt position.
• Token Activation Strips: Highlights exact prompt tokens that trigger specific polysemantic or monosemantic neurons.

5. 🎯 Causal Patching & ROME Sweeps

• Automated Causal Tracing: Swaps activations between clean and corrupted prompts across layers and token positions to pinpoint exact causal circuits.
• Activation Steering Vectors: Injects scaled direction vectors ($x_l \leftarrow x_l + \alpha \cdot v$) during the forward pass to steer model behavior in real time.

---

🤖 Major Supported Models

InterpLens features native auto-discovery and universal hooking for all major open-weights decoder-only transformers:

Model Family	Checkpoints & Variants	Level	Hook Strategy	Launch Command
Meta Llama 3 / 3.2	Llama-3.2-1B, 3B, Llama-3-8B, 70B, Llama-2-7B	L4	LlamaStrategy (model.layers[i])	interplens launch --model meta-llama/Llama-3.2-1B --hf-token "hf_..."
Alibaba Qwen 2.5	Qwen2.5-0.5B, 1.5B, 3B, 7B, 14B, 32B, Coder-7B	L4	QwenStrategy (model.layers[i])	interplens launch --model Qwen/Qwen2.5-0.5B --device cuda
Mistral & Mixtral	Mistral-7B-v0.1, v0.3, Mistral-7B-Instruct, Mixtral-8x7B	L4	MistralStrategy (model.layers[i])	interplens launch --model mistralai/Mistral-7B-v0.1
Google Gemma / Gemma 2	gemma-2b, gemma-7b, gemma-2-2b, gemma-2-9b, 27b	L4	GemmaStrategy (model.model.layers[i])	interplens launch --model google/gemma-2b
OpenAI GPT-2 Family	gpt2 (124M), gpt2-medium (355M), large (774M), xl (1.5B)	L4	GPT2Strategy / HookedTransformer	interplens launch --model gpt2 --device cuda
EleutherAI Pythia & NeoX	pythia-70m, 160m, 410m, 1b, 1.4b, 2.8b, 6.9b, 12b	L4	PythiaStrategy (gpt_neox.layers[i])	interplens launch --model EleutherAI/pythia-70m
SmolLM & TinyLlama	SmolLM-135M, SmolLM-360M, SmolLM-1.7B, TinyLlama-1.1B	L4	LlamaStrategy (model.layers[i])	interplens launch --model HuggingFaceTB/SmolLM-135M
DeepSeek Family	deepseek-llm-7b-base, deepseek-coder-1.3b, coder-6.7b	L4	LlamaStrategy / GenericAdapter	interplens launch --model deepseek-ai/deepseek-llm-7b-base
Custom PyTorch Modules	Any custom nn.Module or research architecture	L3–L4	PyTorchAutoHooker (auto_hook=True)	il.launch(model=my_model, auto_hook=True)

---

📦 Installation

Standard Installation

Installs core PyTorch engine, FastAPI telemetry server, and the visual debugger UI:

pip install interplens

Full Research Bundle (Recommended)

Includes TransformerLens native engine, HuggingFace transformers, and Scikit-Learn PCA modules:

pip install interplens[all]

Development Installation

git clone https://github.com/sarimahsan/interplens.git
cd interplens
pip install -e .[all,dev]
pytest

---

🚀 Quickstart Recipes

Recipe 1: Zero-Copy GPU Attach (Researcher Workflow)

Attach the visual debugger UI directly to a model already resident in your GPU VRAM:

import interplens as il
from transformer_lens import HookedTransformer

# Model is already active in GPU VRAM (e.g. inside a research script or notebook)
model = HookedTransformer.from_pretrained("gpt2-small", device="cuda")

# Launch visual workbench without duplicating memory
server = il.launch(model=model, port=8000)

Recipe 2: HuggingFace Pretrained Auto-Load

import interplens as il

# Automatically downloads weights, resolves tokenizers, and hooks internal layers
server = il.launch(
    model_name="Qwen/Qwen2.5-0.5B",
    device="cuda",
    dtype="bfloat16"
)

Recipe 3: Gated Models with Authentication

import interplens as il

# Load gated Llama-3 checkpoints securely
server = il.launch(
    model_name="meta-llama/Llama-3.2-1B",
    hf_token="hf_YourHuggingFaceTokenHere",
    device="cuda"
)

Recipe 4: Arbitrary Custom PyTorch nn.Module

import torch
import torch.nn as nn
import interplens as il

class ResearchTransformer(nn.Module):
    def __init__(self, vocab_size=1000, d_model=128, n_layers=4):
        super().__init__()
        self.embed = nn.Embedding(vocab_size, d_model)
        self.layers = nn.ModuleList([
            nn.TransformerEncoderLayer(d_model=d_model, nhead=4, batch_first=True)
            for _ in range(n_layers)
        ])
        self.unembed = nn.Linear(d_model, vocab_size, bias=False)

    def forward(self, x):
        h = self.embed(x)
        for layer in self.layers:
            h = layer(h)
        return self.unembed(h)

model = ResearchTransformer()
# Automatic forward hook discovery & unembedding binding
server = il.launch(model=model, auto_hook=True, port=8000)

Recipe 5: CLI One-Liner Launch

# Launch interactive studio from terminal
interplens launch --model gpt2 --device cuda --port 8000

---

🛡️ Capability Matrix (L0 – L4)

InterpLens dynamically evaluates and adapts available analysis engines according to model features:

┌────────────────────────────────────────────────────────────────────────┐
│  Level 4: Full Interventional (Causal Sweeps, Steering Vectors)        │
│  Level 3: Component Hooked (Attention Matrices, MLP Neurons)           │
│  Level 2: Unembed Ready (Logit Lens, Entropy, KL Divergence)           │
│  Level 1: Residual Only (Energy Norms, Cosine Drift Heatmaps)          │
│  Level 0: Black-Box (Top-K Vocabulary Output Predictions)              │
└────────────────────────────────────────────────────────────────────────┘

If a custom model lacks attention extraction or an unembedding matrix $W_U$, InterpLens gracefully degrades rather than crashing, keeping all valid diagnostic tools functional.

---

⚙️ Environment Configuration

Variable	Default	Description
INTERPLENS_PORT	8000	HTTP & WebSocket server port.
INTERPLENS_HOST	127.0.0.1	Host binding address (0.0.0.0 for remote cloud VMs).
INTERPLENS_DEVICE	auto	Primary compute device (cuda, mps, cpu).
INTERPLENS_MAX_CACHE_SESSIONS	32	LRU in-memory session cache capacity.
INTERPLENS_GPU_POLL_INTERVAL_MS	1000	Live GPU telemetry polling interval in milliseconds.

---

🛑 Exception Hierarchy

Structured exceptions under interplens.exceptions for programmatic error handling:

import interplens as il
from interplens.exceptions import (
    InterpLensError,
    ModelLoadError,
    AdapterNotFoundError,
    UnembeddingNotFoundError,
    CapabilityError
)

try:
    adapter = il.launch(model_name="meta-llama/Llama-3.2-1B")
except ModelLoadError as e:
    print(f"Failed to load weights/tokenizer: {e}")
except UnembeddingNotFoundError as e:
    print(f"Model lacks W_U projection: {e}")
except InterpLensError as e:
    print(f"InterpLens runtime exception: {e}")

---

👤 Author

Syed Sarim Ahsan
Undergrad AI Researcher

• GitHub: @sarimahsan
• Project: InterpLens on GitHub

---

📚 Citation

If you use InterpLens in your academic research, interpretability experiments, or course work, please cite:

@software{ahsan2026interplens,
  author = {Syed Sarim Ahsan},
  title = {InterpLens: An Interactive Mechanistic Interpretability Toolkit and Visual Debugger for LLMs},
  year = {2026},
  url = {https://github.com/sarimahsan/interplens},
  version = {0.1.0}
}

---

📜 License

Distributed under the MIT License. See LICENSE for details.