bbycroft/llm-viz
Llm Viz
llm-viz is a 3D architecture visualizer and inference simulator for large language models. It provides a visual representation of network topology and the mathematical operations used during the process of generating a response.
The tool enables the exploration of internal weight distributions and the layout of layers within a neural network. It facilitates model interpretability and inference debugging by tracking the step-by-step movement of data through the architecture.
The system utilizes GPU-accelerated 3D rendering to visualize tensor flow and spatial mappings of weights. It includes capabilities for static model analysis and instruction-step playback to synchronize the visual state of the network with sequential execution.
Features
- Architecture Visualizations - Offers interactive 3D graphical representations of the internal components and operations of LLM architectures.
- Model Interpretability - Enables visual understanding of complex transformer architectures to make internal workings more transparent.
- Architecture Visualizers - Provides a 3D visualizer for the internal structure and data flow of large language models.
- Neural Network Explorers - Provides an interactive tool for exploring the layout and data flow of weights and layers in neural networks.
- Inference Step Executions - Implements a visual playback system to step through individual inference forward passes for a single token.
- Inference Simulators - Offers a visual walkthrough of the step-by-step process used by a large language model to generate a response.
- Inference Visualizations - Renders the network topology and mathematical operations to show the step-by-step process of generating a response.
- Topology Visualizers - Provides a 3D model of the network architecture to illustrate how data flows during inference.
- Model Weight Visualizations - Maps individual model parameters to 3D coordinates to visualize the physical distribution of weights.
- Neural Network Debugging - Facilitates debugging by tracking the step-by-step movement of data through the network during generation.
- Neural Network Visualizations - Generates graphical representations of neural network architectures and layers to study weight distributions.
- Tensor Flow Animations - Visualizes the movement and transformation of data tensors as they pass through different model layers during inference.
- 3D Scene Renderers - Provides a hardware-accelerated 3D renderer for real-time visualization of complex neural network topologies.
- Model Architecture Analysis - Parses pre-trained weight files to extract and render the fixed structural layout of the model architecture.
- Educational Resources - Interactive visualization of transformer architecture.
