GGUF

import { parseGguf } from '@wetron/gguf';

const graph = parseGguf(bytes);

What is parsed#

• GGUF v2 and v3 headers, including little- and big-endian files
• All metadata scalar, string, and array value types
• Architecture metadata such as context length and embedding size
• General quantization metadata, including a readable general.file_type_name
• Tokenizer configuration, vocabulary arrays, special token IDs, and chat templates
• Every tensor name, shape, and GGML quantization type
• Encoded tensor payload ranges through graph.weights

Metadata appears as attributes on the GGUF vN model node. Standardized tensor names are grouped into embedding, attention, feed-forward, state-space, transformer-block, and output stages. Tensors appear as initializer inputs on those stages and are indexed in graph.initializers and graph.tensorShapes.

Weights#

For little-endian files with weights.kind === "available", graph.weights.source.get(name) returns a zero-copy view into the source GGUF buffer. The property panel decodes GGML scalar types and Q4_0 on demand. Other quantization types retain their descriptors and raw bytes but do not yet have value previews. Big-endian files expose descriptors without weight previews.

An error stating that the file is HTML means the model host’s download page was saved instead of the raw .gguf file. Download the raw file and reopen it.