Model Info Utilities

Relevant source files

The following files were used as context for generating this wiki page:

• src/utils/modelInfo.ts
• src/WorkbenchContext.tsx
• src/types.ts
• src/components/OnnxGraphVisualizer.tsx
• src/components/CenterCanvas.tsx

The Model Info Utilities system provides the logic for simulating and analyzing ONNX model characteristics within the ONNX-Station workbench. Its primary purpose is to generate high-fidelity estimations of operator distributions, memory allocation (VRAM), and model architecture statistics based on user-defined export configurations.

This module acts as a bridge between the configuration stage and the visualization stage. It allows the workbench to display technical insights—such as estimated intermediate tensor allocations and operator frequencies—even before a physical model is exported or benchmarked, using heuristics derived from the model's path, modality, and optimization level.

Sources: src/utils/modelInfo.ts:1-7, src/WorkbenchContext.tsx:112-115

Core Data Structures

The system relies on standardized TypeScript interfaces to ensure consistency across the simulation logic, the application state, and the UI components.

Compiled Model Objects

The output of the utility is encapsulated in the CompiledModelInfo interface, which provides a comprehensive summary of the model's structural and resource requirements.

Field	Type	Description
status	idle | exporting | completed	Current state of the model info generation.
operators	CompiledModelOperator[]	A list of specific ONNX operators and their estimated metrics.
totalOps	number	The aggregate count of all operator nodes in the graph.
totalMemoryMb	number	The total estimated VRAM footprint in Megabytes.
isSimulated	boolean	Flag indicating if metrics are heuristic-based or live.

Sources: src/types.ts:88-100

Operator Metrics

Each entry in the operator list contains granular data used for bar and donut chart visualizations.

Field	Type	Description
opType	string	The ONNX operator name (e.g., MatMul, Conv).
count	number	Frequency of this operator in the model graph.
memoryMb	number	Estimated memory consumption for this operator type.
category	string	Classification (e.g., compute, memory, shape).

Sources: src/types.ts:80-86

Simulation Logic Architecture

The central function generateCompiledModelInfo implements a multi-stage pipeline to transform an ExportConfig into technical metrics.

1. Model Identification and Scaling

The utility first parses the modelPath to determine the base architecture and layer count. It identifies specific model families like Whisper, LLaMA, BERT, or ResNet and assigns a base number of layers (e.g., 32 layers for LLaMA-style models).

2. Multiplier Application

Heuristics are adjusted based on the precision and optimization levels selected in the workbench:

• Dtype Scaling: fp16 and bf16 reduce memory estimates by approximately 50%.
• Optimization Scaling: Higher levels (O1-O3) decrease operator counts via fusion, while O4 (Quantization) increases node count due to the introduction of QuantizeLinear/DequantizeLinear nodes but drastically reduces memory footprint.

Sources: src/utils/modelInfo.ts:46-75

3. Modality-Specific Operator Generation

The system populates the operator list based on the chosen modality (text, vision, or audio).

• Text: Focuses on MatMul, Gather (embeddings), and LayerNormalization.
• Vision: Differentiates between ViT (Vision Transformer) and standard CNN (Convolutional) architectures, generating Conv or MatMul nodes accordingly.
• Audio: Simulates complex encoder-decoder stacks like Whisper, involving Conv1D frontends and FusedAttention nodes.

Sources: src/utils/modelInfo.ts:93-195

Integration with Workbench State

The Model Info Utilities are integrated into the WorkbenchContext to provide reactive updates to the UI.

Visualization Components

The results from the model info utilities are consumed by the OnnxGraphVisualizer component, which uses D3.js to render two primary views:

1. Operator Instance Counts: A horizontal bar chart showing the frequency of ONNX nodes, colored by category (Compute, Memory, Activation, Normalization, or Shape).
2. VRAM Map: A donut chart illustrating the estimated memory distribution, highlighting potential bottlenecks where specific operators (like large MatMuls) consume significant VRAM.

Sources: src/components/OnnxGraphVisualizer.tsx:64-118, src/components/OnnxGraphVisualizer.tsx:174-219

Summary of Optimization Impacts

The utilities reflect the technical impact of Optimum CLI flags on the model structure as follows:

Optimization Flag	Impact on CompiledModelInfo
O1 - O3	Fuses nodes (decreases totalOps), reduces totalMemoryMb.
O4	Adds QLinear nodes; reduces memory by ~65% via quantization.
--dtype fp16	Halves the memoryMb for weight-heavy operators.
--modality	Changes the fundamental set of operators (e.g., Conv vs MatMul).

Sources: src/utils/modelInfo.ts:51-75, src/components/CenterCanvas.tsx:136-155