Installation

Installation#

System requirements#

We recommend using an NVIDIA GPU for faster compiles with qbcompiler, but it is not necessary. Currently, the CPU-only version of qbcompiler is also supported.

Reference System#

SDK Installation#

We recommend installing the qbcompiler using Mobilint’s official Docker container images. Starting with qbcompiler v0.10.0.0, the first two digits of the compiler version must match the first two digits of the Docker image tag. For example, if you use qbcompiler v1.0.0, choose an image whose tag starts with 1.0. In addition, from qbcompiler Docker 0.11 and later, the images are based on CUDA 12.8.1, which requires an NVIDIA driver from the R570 branch or newer. For best compatibility, we recommend updating your NVIDIA driver to the latest available version.

Building Docker Container#

Run the following commands to build the docker container.

# Download Docker Image
docker pull mobilint/qbcompiler:1.0-cuda12.8.1-ubuntu22.04
# mkdir {WORKING DIRECTORY} (if needed)
cd {WORKING DIRECTORY}
# Build the Docker container
docker run -it --gpus all --ipc=host --name {YOUR_CONTAINER_NAME} -v $(pwd):/workspace mobilint/qbcompiler:1.0-cuda12.8.1-ubuntu22.04 /bin/bash

(Recommended) If the trained models and datasets are stored in different directories, you can mount them to the docker container as follows:

docker run -it --gpus all --ipc=host --name {YOUR_CONTAINER_NAME} -v $(pwd):/workspace -v {PATH TO MODEL DIR}:/models -v {PATH TO DATASET DIR}:/datasets mobilint/qbcompiler:1.0-cuda12.8.1-ubuntu22.04 /bin/bash

To run the NPU inside the Docker container, add --device /dev/aries0:/dev/aries0 when creating the Docker container.

(Optional) By building the Docker container with the CPU-only version image, users may run the Mobilint Compiler on computers without GPUs.

# Download Docker Image
docker pull mobilint/qbcompiler:1.0-cpu-ubuntu22.04
cd {WORKING DIRECTORY}
# Build the Docker container
docker run -it --ipc=host --name {YOUR_CONTAINER_NAME} -v $(pwd):/workspace mobilint/qbcompiler:1.0-cpu-ubuntu22.04 /bin/bash

Installation of qbcompiler#

qbcompiler packages are available in Mobilint® Software Development Kit (SDK). After entering the page, download the qbcompiler package for this release. The package is in the form of a Python wheel file, which can be installed using pip.

Run the following commands to install qbcompiler on the docker container.

# Download qbcompiler-1.0.0+aries-py3-none-any.whl file
# Copy qbcompiler whl file to Docker
docker cp {Path to qbcompiler-1.0.0+aries-py3-none-any.whl} {YOUR_CONTAINER_NAME}:/
# Start Docker
docker start {YOUR_CONTAINER_NAME}
# Attach to Docker container
docker exec -it {YOUR_CONTAINER_NAME} /bin/bash
# Install qbcompiler
cd /
python -m pip install qbcompiler-1.0.0+aries-py3-none-any.whl

In qbcompiler releases before 1.2, wheel filenames may include an NPU name in the local version segment, such as +aries. Install the exact wheel file provided for the release.