NeuronDump: Open a local LLM as a 3D graph, cut parts, see what breaks (github.com)

NeuronDump has been launched as a versatile tool that allows users to explore the inner workings of local large language models (LLMs) by reading GGUF model files, which are used by frameworks like Ollama and llama.cpp. This utility not only visualizes model components—such as layers, attention heads, and weight statistics—but also enables users to identify unnecessary parts and simplify the model, creating slimmer versions that maintain functionality. The tool can be executed via a command line or a web interface, providing users with an interactive 3D or 2D graphical representation of the model's structure, allowing for detailed inspections and modifications. This innovation is significant for the AI/ML community as it empowers developers to optimize their models by safely removing components that can lead to smaller file sizes and reduced computational requirements. Its features, such as the ability to cut or analyze specific model parts and simulate the impact of these changes, promote a deeper understanding of model internals. NeuronDump’s capabilities to run simulations alongside the original model and provide instant feedback on the effects of modifications make it a valuable resource for researchers and practitioners looking to streamline model performance while ensuring stability.