llm-viewer

llm-viewer is a local visualizer for Hugging Face text generation model configs.

Drop a config.json into the browser. Backend computes a normalized graph bundle. Frontend renders:

• model structure
• data flow
• tensor shapes
• prefill / decode differences
• block-level detail view

Goal: make LLM architecture inspection feel closer to Netron, but for Hugging Face model_config.json plus runtime shape profiles instead of exported ONNX graphs.

Features

• local browser app powered by FastAPI
• drag-and-drop config.json
• prefill and decode graph profiles
  • symbolic shape inference from transformers config semantics
• model graph and transformer block graph
• shape-aware edges and node summaries
• model summary sidebar from model_config.json
• double-click repeated block to jump into block detail
• pan, zoom, fit, reset

Supported Models

Current decoder-family support:

• code_llama
• diffllama
• doge
• granite
• helium
• llama
• ministral
• mistral
• olmo
• qwen2
• qwen3
• gemma
• gemma2
• gemma3_text
• gemma4_text
• seed_oss
• smollm3
• stablelm

Quick Start

Install in editable mode:

python -m pip install -e .

Core runtime dependencies include fastapi, uvicorn, and transformers.

Start the app:

llm_viewer

This starts a local server and opens a browser tab.

You can also run without install:

PYTHONPATH=src python -m llm_viewer

CLI

Start browser app:

llm_viewer

Export graph JSON:

llm_viewer extract path/to/config.json --profile prefill

Choose profile explicitly:

llm_viewer extract path/to/config.json --profile decode --output graph.json

Example Configs

Sample configs live in examples/:

• examples/llama-config.json
• examples/qwen3.json

How It Works

llm-viewer has two layers:

1. Backend:
   • parses config.json
   • selects an adapter by model_type
   • builds a normalized graph bundle for a runtime profile
2. Frontend:
   • uploads config to backend
   • renders graph, edges, labels, and side panels

Current implementation focuses on a normalized graph abstraction, not exact kernel-level execution tracing. It does not run the model for shape inference. Backend resolves the Hugging Face config class through transformers and derives symbolic shapes from that config semantics.

Project Layout

src/llm_viewer/
  adapters/        model-family graph builders
  static/          browser frontend
  cli.py           CLI entrypoint
  server.py        FastAPI app
  schema.py        graph bundle schema
tests/
examples/

Development

Run tests:

pytest -q

Run app in-place:

PYTHONPATH=src python -m llm_viewer

Acknowledgements

Thanks to:

• Netron for excellent model visualization ideas and high-quality open source code references
• Hugging Face Transformers for the model implementations and config ecosystem this project builds on
• the broader open source LLM tooling community for making model inspection and understanding easier

Status

This project is still MVP-stage.

Good now:

• graph extraction for supported decoder families
• browser visualization flow
• model/block graph switching
• residual path rendering

Still rough:

• layout quality for dense graphs
• richer source/code mapping
• more model families beyond current decoder set
• better graph semantics for non-transformer architectures

License

MIT. See LICENSE.