Roadmap

Roadmap

Overview

TokenPrint is actively evolving. The ultimate goal is to become the standard interactive visual debugger for Transformer execution—the tool that answers "what is my model doing right now, and why?"

Current Status (Shipped)

• Architecture Explorer: Real tensor inventory with point-cloud rendering.
• Live Inference: Streamed greedy generation with dynamic TransformerStack geometry.
• Walkthrough Mode: Eased camera tours with PCA projections and attention bezier curves.
• Debugger Mode: Breakpoints, anomaly sentinels, logit-lens, and raw-component ablation.
• Trace Record/Replay: Exporting generations to .tokenprint.json for offline viewing.

Upcoming Milestones

Milestone 1: MoE and Advanced Architectures

• Support for Mixture of Experts (MoE) routing visualization (e.g., DeepSeek, Mixtral).
• Support for Multi-Head Latent Attention (MLA).

Milestone 2: Activation Patching

• Moving beyond simple "ablation" (zeroing out a head) to true Activation Patching (swapping a hidden state from a corrupted prompt into a clean prompt) directly via the UI.

Milestone 3: Client-side Inference

• Utilizing WebGPU (via WebLLM or Transformers.js) to run actual inference directly on the GGUF files dropped into the browser, completely removing the need for the Python backend for basic use cases.

Milestone 4: Multi-Model Diffing

• Loading two traces from two different quantized versions of the same model (e.g., Q4 vs FP16) and rendering a visual 3D diff to see exactly where the quantization errors accumulate during the forward pass.

Related pages

• Contributing

Further reading

• Full Roadmap File
• Design Review

Navigation

Previous	Home	Next
Scientific Accuracy	Home	FAQ