Implementation:Ollama Ollama WriteGGUF

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

Overview

Concrete tool for writing GGUF binary model files provided by the convert and fs/ggml packages.

Description

The internal writeFile function in convert.go orchestrates the GGUF writing process by calling ggml.WriteGGUF with the complete metadata KV map and tensor list.

splitDim and mergeTensors handle tensor transformations needed before writing: splitting combined weight matrices and merging separate matrices that need to be concatenated in GGUF format.

Usage

Called as the final step of ConvertModel after all tensors and metadata have been prepared.

Code Reference

Source Location

• Repository: ollama
• File: convert/convert.go (writeFile), convert/tensor.go (splitDim, mergeTensors)
• Lines: convert.go:L382-388, tensor.go:L30-89 (splitDim), tensor.go:L91-131 (mergeTensors)

Signature

// Internal to convert package
func writeFile(f *os.File, kv KV, ts []*ggml.Tensor) error

func splitDim(t Tensor, dim int, splits ...split) iter.Seq[*ggml.Tensor]

func mergeTensors(unmatched []Tensor, merges ...merge) (out []*ggml.Tensor, _ []Tensor)

Import

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

I/O Contract

Inputs

Name	Type	Required	Description
f	*os.File	Yes	Output file for GGUF binary
kv	KV	Yes	Complete metadata key-value map
ts	[]*ggml.Tensor	Yes	All model tensors with names, shapes, dtypes, and data

Outputs

Name	Type	Description
error	error	Non-nil if writing fails
Side effect	GGUF file	Complete binary file written to disk

Usage Examples

Internal Usage

// From convert/convert.go ConvertModel
kv := converter.KV(tokenizer)
tensors := converter.Tensors(rawTensors)
err := writeFile(outputFile, kv, tensors)

Related Pages

Implements Principle

• Principle:Ollama_Ollama_GGUF_Assembly

Requires Environment

• Environment:Ollama_Ollama_Go_Runtime