Implementation:Ollama Ollama Create Imagegen

Knowledge Sources	Ollama
Domains	Image Generation, Model Import
Last Updated	2025-02-15 00:00 GMT

Overview

Imports HuggingFace image generation models into Ollama's per-tensor blob storage format with optional quantization.

Description

The imagegen.go file in x/create implements CreateImageGenModel, which converts a HuggingFace diffusers model directory into Ollama's manifest format. It processes three components (text_encoder, transformer, vae), extracting individual tensors from each .safetensors file via TensorExtractor and storing them as per-tensor blobs in safetensors format (88 bytes header overhead) for native mmap loading. Quantization is supported for linear weights in transformer and text_encoder components (int4, int8, nvfp4, mxfp8) while VAE weights are kept in original precision. The function imports config files (model_index.json, tokenizer.json, component configs), normalizes model_index.json to Ollama format with pipeline class detection and total parameter counting, and writes the final manifest via callbacks. ShouldQuantize determines which tensors are eligible for quantization based on naming patterns.

Usage

Used by "ollama create" when importing an image generation model from a local HuggingFace diffusers directory.

Code Reference

Source Location

• Repository: Ollama
• File: x/create/imagegen.go
• Lines: 1-230

Signature

func CreateImageGenModel(
	modelName string,
	modelDir string,
	quantize string,
	createLayer LayerCreator,
	createTensorLayer QuantizingTensorLayerCreator,
	writeManifest ManifestWriter,
	fn func(status string),
) error

func ShouldQuantize(tensorName, component string) bool

Import

import "github.com/ollama/ollama/x/create"

I/O Contract

Inputs

Name	Type	Required	Description
modelName	string	Yes	Target Ollama model name
modelDir	string	Yes	Path to HuggingFace diffusers model directory
quantize	string	No	Quantization type: "", "int4", "int8", "nvfp4", "mxfp8"
createLayer	LayerCreator	Yes	Callback to create config layers
createTensorLayer	QuantizingTensorLayerCreator	Yes	Callback to create tensor layers (with optional quantization)
writeManifest	ManifestWriter	Yes	Callback to write the final manifest
fn	func(string)	Yes	Status callback for progress reporting

Outputs

Name	Type	Description
error	error	Error from tensor extraction, quantization, or manifest writing

Usage Examples

err := create.CreateImageGenModel(
    "my-flux-model",
    "/path/to/flux-klein",
    "int4", // quantize to 4-bit
    myLayerCreator,
    myTensorLayerCreator,
    myManifestWriter,
    func(status string) {
        fmt.Println(status)
    },
)

Related Pages

• Principle:Ollama_Ollama_ImageGeneration