Implementation:Ollama Ollama LlamaModel KV

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

Overview

Concrete tool for extracting LLaMA architecture hyperparameters into GGUF key-value format provided by the convert package.

Description

llamaModel.KV maps LLaMA-specific config.json fields to GGUF metadata keys. It handles special cases like RoPE scaling parameters, GQA head counts, and vocabulary size adjustments.

The base ModelParameters.KV method handles architecture-independent metadata: file type, quantization version, and tokenizer entries (vocabulary tokens, scores, types, merges, special tokens, chat template).

Usage

Called internally during model conversion to populate the GGUF metadata section.

Code Reference

Source Location

• Repository: ollama
• File: convert/convert_llama.go (llamaModel.KV), convert/convert.go (ModelParameters.KV)
• Lines: convert_llama.go:L51-127 (llamaModel.KV), convert.go:L130-158 (ModelParameters.KV)

Signature

func (p *llamaModel) KV(t *Tokenizer) KV

func (ModelParameters) KV(t *Tokenizer) KV

Import

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

I/O Contract

Inputs

Name	Type	Required	Description
t	*Tokenizer	Yes	Parsed tokenizer with vocabulary, merges, and special tokens

Outputs

Name	Type	Description
KV	KV (map)	GGUF key-value metadata map with architecture and tokenizer entries

Usage Examples

Internal Usage

// From convert/convert.go
kv := converter.KV(tokenizer)
// kv now contains entries like:
// "llama.embedding_length": 4096
// "llama.block_count": 32
// "llama.attention.head_count": 32
// "tokenizer.ggml.tokens": [...]

Related Pages

Implements Principle

• Principle:Ollama_Ollama_Hyperparameter_Extraction

Requires Environment

• Environment:Ollama_Ollama_Go_Runtime