iQ-Foundry

Simplify the Workflow, Accelerate Deployment.

This tool simplifies model quantization and validation for edge AI, reducing friction from preparation to real-device deployment—making workflows repeatable and scalable.

🚀 Bring Your Own Model

Supports custom-trained and pretrained models.

New in v0.0.3: FP32 and mixed-precision deployment support!!

iQ-Foundry helps prepare computer vision models for innodisk Qualcomm solution. The current workflow supports compiling compatible computer vision .pt models into .tflite and .onnx artifacts, validating reference-versus-converted quality with mAP@0.5, and running on-device inference on EXMP-Q911 (Qualcomm QCS9075).

iQ-Foundry supports a Bring Your Own Model workflow. You can use your own compatible yolov10, yolov11, or yolov26 models with the pipeline. If you need pretrained YOLO weights, you can download official pretrained models from Ultralytics.

Note: iQ-Foundry is focused on computer vision model conversion, optimization, and deployment. For more comprehensive deployment information related to hardware and applications, see iQ-Studio.

Workflow At A Glance

1. Choose your host workflow: Ubuntu 22.04 or Windows 11.
2. Run qc to generate the quantized and compiled model.
3. Run mAP to compare source and converted model quality on the same dataset.
4. Run test for on-device inference, either from the host through ADB or directly on the target.

Quick Start

Ubuntu Host Native Ubuntu 22.04 with Docker Engine Open Ubuntu Host Guide

Windows Host Windows 11 with Docker Engine inside WSL Open Windows Host Guide

Platform Capabilities Overview

Workflow Modes

Mode	Stage	Description
qc	Quantize and Compile	Convert a supported model into a deployment-ready artifact for validation and device execution.
mAP	Evaluate Converted Model Quality	Measure detection quality so you can compare the original model against the converted result.
test	On-Device Inference	Run the prepared model on target device and save outputs for quick functional and visual verification.

Model and Deployment Support

Category	Support
Input Format	.pt
Model Families	yolov10, yolov11, yolov26
Quantization	FP32 (float), INT8 (W8A8), W8A16 (INT mixed precision)
Target Device	EXMP-Q911 (Qualcomm QCS9075)
Runtime	LiteRT (TensorFlow Lite) ONNX Runtime
Backend	NPU (Qualcomm HTP), CPU

Runtime and Precision Support Matrix

Runtime	FP32	INT8	W8A16
LiteRT	✓	✓	✗
ONNX Runtime	✓	✗	✓

Explore Other Documentation

Document	Purpose	Use It When
docs/other_model_flow.md	High-level flow guide for unsupported or custom models outside the current native iQ-Foundry path.	You want to explore bring-your-own-model onboarding beyond the built-in workflows.
docker/Docker.md	Fallback guide for building the iQF Docker image locally.	You need to build the container image instead of pulling it from Docker Hub.

Changelog

Please refer to the Changelog for all updates.

License

This project is licensed under the Apache License 2.0. See the LICENSE file for details.