Implementation:Ollama Ollama ConvertModel

Knowledge Sources	Ollama
Domains	Model_Architecture, Format_Conversion
Last Updated	2026-02-14 00:00 GMT

Overview

Concrete tool for converting HuggingFace models to GGUF format with automatic architecture detection provided by the convert package.

Description

ConvertModel is the top-level conversion entry point. It reads config.json from the model directory filesystem, detects the architecture, dispatches to the appropriate converter, processes all tensors and metadata, and writes the complete GGUF file.

LoadModelMetadata performs the architecture detection step, parsing config.json and returning the architecture-specific converter along with the parsed tokenizer.

Usage

Called by CreateHandler when a FROM directive points to a SafeTensors model directory rather than an existing Ollama model.

Code Reference

Source Location

• Repository: ollama
• File: convert/convert.go
• Lines: L367-388 (ConvertModel), L255-361 (LoadModelMetadata)

Signature

func ConvertModel(fsys fs.FS, f *os.File) error

func LoadModelMetadata(fsys fs.FS) (ModelKV, *Tokenizer, error)

Import

import "github.com/ollama/ollama/convert"

I/O Contract

Inputs

Name	Type	Required	Description
fsys	fs.FS	Yes	Filesystem rooted at model directory (config.json, safetensors, tokenizer files)
f	*os.File	Yes	Output file for GGUF binary

Outputs

Name	Type	Description
error	error	Non-nil if architecture unsupported, files missing, or conversion fails
Side effect	GGUF file	Complete GGUF binary written to output file

Usage Examples

Converting a HuggingFace Model

import (
    "os"
    "github.com/ollama/ollama/convert"
)

modelDir := os.DirFS("/path/to/huggingface/model")
outFile, _ := os.Create("model.gguf")
defer outFile.Close()

err := convert.ConvertModel(modelDir, outFile)
if err != nil {
    // handle unsupported architecture or conversion error
}

Related Pages

Implements Principle

• Principle:Ollama_Ollama_Architecture_Detection

Requires Environment

• Environment:Ollama_Ollama_Go_Runtime