Developer Guide Adding a New Model

Adding a New Model

Overview

TokenPrint currently defaults to Qwen2.5. This guide explains how to add explicit support for a new model architecture (e.g., DeepSeek) to the platform.

Why it matters

Because TokenPrint is architecture-aware, simply changing the MODEL_NAME string in the backend might result in incorrect visuals if the frontend doesn't know how to render the new model's specific quirks (like new normalization layers or attention variants).

How TokenPrint implements it

1. Update the Backend (model.py)

Ensure the ModelEngine can load the model via transformers. Crucially, verify that attn_implementation="eager" actually returns the attention tensors. Some custom model implementations on HuggingFace require patches to expose intermediate variables.

2. Update the Architecture Detector (formulas.ts)

In the frontend, locate detectArch() in lib/formulas.ts. Add a condition to check for the new architecture string (e.g., config.architecture === 'deepseek'). Define the LaTeX formula set for this architecture.

3. Update the Geometry (TransformerStack.tsx)

If the model uses a standard block (like Llama), no changes are needed. If the model uses a novel block (like Phi-2's parallel structure), you must create a new branch in TransformerStack.tsx that alters the 3D layout when that architecture is detected.

4. Test GGUF Compatibility

Download a quantized version of the model. Drag it into the frontend and verify the client-side parser in lib/gguf/ correctly reads the header and maps the tensors.

Diagram

flowchart LR
    HF[HuggingFace Model] --> Backend[Verify Eager Attention]
    Backend --> Formulas[Update detectArch in formulas.ts]
    Formulas --> Geo[Update TransformerStack.tsx]
    Geo --> GGUF[Test Client Parser Compatibility]

Loading

Related pages

• Supported Models
• Architecture

Further reading

• Contributing Ideas

Navigation

Previous	Home	Next
Adding a New Visualization	Home	Creating UI Components