Repository navigation
User Guide Timeline
The Timeline, located in the Bottom Bar, controls the playback of Live Inference traces and Walkthrough chapters. It allows you to pause, step, and scrub through the model's execution.
LLM inference happens in milliseconds. To understand it, you must slow it down. The Timeline allows you to decouple the visual playback from the backend computation speed, letting you inspect individual layers and tokens at your own pace.
The PlaybackEngine is a headless component running a setInterval ticker.
It walks through the op_catalog (the ordered list of operations) layer-by-layer. When the timeline is paused, the ticker stops, freezing the store.activeOp state and halting the camera glide.
-
Play/Pause (
Spacebar): Toggles autoplay. When paused, the 3D scene freezes exactly on the current operation. -
Step Back/Forward (
J/K): Manually advance or reverse one operation (e.g., from Pre-LayerNorm to Self-Attention). - Speed Multiplier: A dropdown (e.g., 0.5x, 1x, 2x) that normalizes the pacing of the autoplay ticker. It does not fabricate frames; it simply changes the delay between rendering real operations.
- Token Strip: A horizontal list of tokens. The prompt tokens are displayed first, followed by generated tokens. Clicking a token scrubs the timeline to the exact frame that generated it.
Note Pacing is strictly normalized. TokenPrint never fabricates in-between frames or synthetic data to smooth out the playback.
sequenceDiagram
participant UI as Timeline UI
participant Store as Zustand Store
participant Engine as PlaybackEngine
UI->>Store: setPlaying(true)
Store->>Engine: Reacts to playing state
loop Every Tick
Engine->>Store: setActiveOp(next_op)
Engine->>Store: updateCameraTarget()
end
UI->>Store: setPlaying(false)
Engine-->>Engine: Halts ticker
| Previous | Home | Next |
|---|---|---|
| HUD | Home | Scene Navigation |
TokenPrint Wiki | Home | Getting Started | User Guide | Concepts | Architecture | API | GitHub
TokenPrint is an interactive 3D visualization platform for exploring transformer architectures, tensors, and real-time LLM inference.
- User Guide
- Architecture Explorer
- Live Inference
- Tensor Inspector
- Camera Controls
- HUD
- Timeline
- Scene Navigation
- Transformer Concepts
- Tokenization
- Embeddings
- Positional Encoding
- RoPE
- Multi-Head Attention
- KV Cache
- LayerNorm
- Feed Forward Network
- Residual Connections
- Logits
- Softmax
- Sampling
- Autoregressive Generation
- Visualization System
- Scene Graph
- Geometry
- Materials
- Animation System
- Color Mapping
- Camera System
- Interaction Model
- Architecture
- Overall Architecture
- Frontend
- Backend
- WebSocket Protocol
- Data Pipeline
- Renderer
- Event System