Implementation:Hiyouga LLaMA Factory WebUI Export Component

Knowledge Sources	Hiyouga_LLaMA_Factory
Domains	WebUI, Model Export, Quantization, Gradio Components
Last Updated	2026-02-06 19:00 GMT

Overview

Model export tab component that enables users to export fine-tuned models with optional GPTQ quantization directly from the WebUI.

Description

The export.py module provides two main pieces of functionality. The save_model generator function validates export inputs (model name, path, export directory, quantization settings), constructs the export arguments dictionary including optional GPTQ quantization at 2, 3, 4, or 8 bits, invokes export_model from the training tuner, and performs GPU memory cleanup via torch_gc. The create_export_tab function assembles the Gradio UI with export size slider, GPTQ quantization bit dropdown, quantization dataset path, device selection (cpu/auto), legacy format checkbox, export directory, Hub model ID, and extra args textbox. A can_quantize helper disables quantization when checkpoint adapters are selected.

Usage

Use this component when building the main WebUI to provide model export capabilities. It is instantiated within create_ui in the interface module as the "Export" tab (hidden in demo mode).

Code Reference

Source Location

• Repository: Hiyouga_LLaMA_Factory
• File: src/llamafactory/webui/components/export.py
• Lines: 1-169

Signature

GPTQ_BITS = ["8", "4", "3", "2"]

def can_quantize(checkpoint_path: str | list[str]) -> "gr.Dropdown": ...

def save_model(
    lang: str,
    model_name: str,
    model_path: str,
    finetuning_type: str,
    checkpoint_path: str | list[str],
    template: str,
    export_size: int,
    export_quantization_bit: str,
    export_quantization_dataset: str,
    export_device: str,
    export_legacy_format: bool,
    export_dir: str,
    export_hub_model_id: str,
    extra_args: str,
) -> Generator[str, None, None]: ...

def create_export_tab(engine: "Engine") -> dict[str, "Component"]: ...

Import

from llamafactory.webui.components.export import create_export_tab, save_model, can_quantize

I/O Contract

Inputs

Name	Type	Required	Description
engine	Engine	Yes	The WebUI engine instance providing access to Manager for retrieving top-level element references
lang	str	Yes (save_model)	UI language code for localized alert messages
model_name	str	Yes (save_model)	Name of the model to export
model_path	str	Yes (save_model)	Filesystem or hub path to the base model
finetuning_type	str	Yes (save_model)	Finetuning method (lora, full, freeze, etc.)
checkpoint_path	str or list[str]	No (save_model)	Path(s) to adapter checkpoint(s)
export_quantization_bit	str	No (save_model)	GPTQ quantization bits ("none", "2", "3", "4", "8")
export_dir	str	Yes (save_model)	Directory to write the exported model

Outputs

Name	Type	Description
elem_dict	dict[str, Component]	Dictionary of named Gradio components: export_size, export_quantization_bit, export_quantization_dataset, export_device, export_legacy_format, export_dir, export_hub_model_id, extra_args, export_btn, info_box
status (save_model)	Generator[str]	Yields progress messages ("Exporting model...", "Model exported successfully.")

Usage Examples

# Adding the export tab to the main UI (non-demo mode only)
from llamafactory.webui.components.export import create_export_tab

if not demo_mode:
    with gr.Tab("Export"):
        engine.manager.add_elems("export", create_export_tab(engine))

Related Pages

• Hiyouga_LLaMA_Factory_WebUI_Interface - Main interface that conditionally includes the export tab
• Hiyouga_LLaMA_Factory_WebUI_Common - Provides get_save_dir and load_config used during export
• Hiyouga_LLaMA_Factory_WebUI_Top_Component - Provides model/checkpoint selection inputs consumed by export