i use several way to convert my onnx to hef from l...

Below is an end-to-end “copy/paste” guide for doing ONNX → HEF using Hailo’s Docker-based AI Software Suite on Linux, plus the basics of Docker (install, start/enable, run containers).

I’ll assume you’re on Ubuntu 22.04 or 24.04 x86_64, because that’s the most common setup for Hailo’s toolchain. If you’re on another distro, the Docker part is the same idea, but use the install page for your distro. (Docker Documentation)

---

0) What Docker is (in 30 seconds)

• Image = a packaged environment (OS + tools) (e.g., “Hailo AI Software Suite”).
• Container = a running instance of an image.
• The Docker daemon runs in the background and starts containers.

---

1) Install Docker Engine on Ubuntu (recommended method)

1.1 Remove conflicting packages (safe even if nothing is installed)

bash
sudo apt update
sudo apt remove $(dpkg --get-selections docker.io docker-compose docker-compose-v2 docker-doc podman-docker containerd runc | cut -f1)

(These are the packages Docker warns can conflict.) (Docker Documentation)

1.2 Add Docker’s official apt repo + install Docker

bash
sudo apt update
sudo apt install ca-certificates curl
sudo install -m 0755 -d /etc/apt/keyrings
sudo curl -fsSL https://download.docker.com/linux/ubuntu/gpg -o /etc/apt/keyrings/docker.asc
sudo chmod a+r /etc/apt/keyrings/docker.asc

sudo tee /etc/apt/sources.list.d/docker.sources <<EOF
Types: deb
URIs: https://download.docker.com/linux/ubuntu
Suites: $(. /etc/os-release && echo "${UBUNTU_CODENAME:-$VERSION_CODENAME}")
Components: stable
Signed-By: /etc/apt/keyrings/docker.asc
EOF

sudo apt update
sudo apt install docker-ce docker-ce-cli containerd.io docker-buildx-plugin docker-compose-plugin

These commands are straight from Docker’s Ubuntu install guide. (Docker Documentation)

1.3 Start/enable Docker (“activate docker”)

bash
sudo systemctl start docker
sudo systemctl enable docker.service
sudo systemctl enable containerd.service

Enabling on boot is recommended. (Docker Documentation)

1.4 Quick test

bash
sudo docker run hello-world

Docker uses this as the standard verification test. (Docker Documentation)

1.5 (Optional) Run docker without sudo

bash
sudo groupadd docker
sudo usermod -aG docker $USER
newgrp docker
docker run hello-world

Important security note: the docker group effectively grants root-level power. (Docker Documentation)

---

2) Get Hailo’s Docker package (AI Software Suite – Docker)

Hailo typically provides the Docker suite as a downloadable archive from their Developer Zone, often containing:

• a hailo_ai_sw_suite_<version>.tar.gz (Docker image tarball)
• a helper script like hailo_ai_sw_suite_docker_run.sh (Dylan Tintenfich)

A common flow described by vendors/users is:

• Download “Hailo AI Software Suite – Docker”
• (If you need PCIe device access) also download/install the HailoRT PCIe driver deb, then reboot (Macnica)

Example shown:

bash
sudo dpkg --install hailort-pcie-driver_4.17.0_all.deb
sudo reboot

(Macnica)

If you are only compiling HEF, you often don’t need the PCIe driver. You mainly need it if you want hailortcli scan / hardware inference from inside the container.

---

3) Load the Hailo Docker image + enter the container

Option A (simplest): use Hailo’s provided run script

After unzipping:

bash
chmod +x hailo_ai_sw_suite_docker_run.sh
./hailo_ai_sw_suite_docker_run.sh

To re-enter the same container later:

bash
./hailo_ai_sw_suite_docker_run.sh --resume

(Macnica)

Option B (manual): docker load then docker run

If your download includes a *.tar.gz image, load it:

bash
docker load -i hailo_ai_sw_suite_2024-10.tar.gz
docker images

(Dylan Tintenfich)

Then run a container. If you want the container to access devices/drivers on the host, a known-working pattern is:

bash
docker run -it --name hailo_container --privileged --network host \
  -v /dev:/dev \
  -v /lib/firmware:/lib/firmware \
  -v /lib/udev/rules.d:/lib/udev/rules.d \
  -v /lib/modules:/lib/modules \
  -v /home/$USER/hailo_workspace:/workspace \
  hailo_ai_sw_suite_2024-10:1

And single-line form (useful if your shell hates line breaks):

bash
docker run -it --name hailo_container --privileged --network host -v /dev:/dev -v /lib/firmware:/lib/firmware -v /lib/udev/rules.d:/lib/udev/rules.d -v /lib/modules:/lib/modules -v /home/$USER/hailo_workspace:/workspace hailo_ai_sw_suite_2024-10:1

(Dylan Tintenfich)

---

4) Verify Hailo tools inside the container

Once you’re inside:

bash
hailo -h

If you installed drivers and passed devices through, you can also try:

bash
hailortcli scan

These are commonly used sanity checks. (Macnica)

Also, inside the Docker, Hailo support often points people to:

bash
hailo tutorial

(which starts a tutorial/Jupyter flow for each step of conversion). (Hailo Community)

---

5) Build a HEF inside Docker (two practical workflows)

Workflow 1: Dataflow Compiler CLI (most general for “any ONNX”)

This is the classic 3-step pipeline:

1. parse ONNX → HAR
2. optimize HAR (needs calibration + model script / ALLS)
3. compile optimized HAR → HEF

Hailo community examples show:

bash
# 1) Parse (ONNX -> .har)
hailo parser onnx ./model.onnx --hw-arch hailo8l --end-node-names <...>

# 2) Optimize (.har -> optimized .har)
hailo optimize <har-file> --hw-arch <hailo-board> --model-script <.alls-file> --calib-set-path <calibration-dataset>

# 3) Compile (optimized .har -> .hef)
hailo compiler <optimized-har-file> --hw-arch <hailo-board>

(Hailo Community)

Notes that matter a lot:

• Pick --hw-arch correctly (e.g., hailo8 vs hailo8l).
• Your .alls matters a LOT. If you don’t have one, start from a Model Zoo ALLS for a similar network (Hailo community recommends this). (Hailo Community)
• To discover flags/options:

bash
hailo -h

(Hailo Community)

---

Workflow 2: Hailo Model Zoo (hailomz) (best when your model matches a supported architecture like YOLO)

If you’re doing YOLO-family models, hailomz can be smoother because it already knows lots of configs/scripts.

Model Zoo’s own “Getting Started” shows the commands for:

• parse: hailomz parse …
• optimize: hailomz optimize … --calib-path …
• compile: hailomz compile … --hw-arch … (GitHub)

A commonly used YOLO-style flow (example):

bash
cd /workspace/models

# Parse (custom checkpoint)
hailomz parse --hw-arch hailo8l --ckpt model.onnx yolov8n

# Optimize (calibration)
hailomz optimize --hw-arch hailo8l --calib-path /workspace/calib --resize 640 640 --classes 80 --har yolov8n.har yolov8n --performance

# Compile (HEF)
hailomz compile yolov8n --hw-arch hailo8l --calib-path /workspace/calib --resize 640 640 --classes 80 --har yolov8n.har --performance

(Those exact flags/structure are shown in a Docker-based ONNX→HEF walkthrough.) (Dylan Tintenfich)

Very important compatibility note (Hailo-8 / Hailo-8L users):
The official Model Zoo repo warns that Hailo-8/8L use the v2.x branch with DFC v3.x, while master targets newer devices. So match your suite/tools versions carefully. (GitHub)

---

6) Fix the “my .alls only works in one line / parsing errors” problem (common causes)

Two very common causes when moving scripts between editors/systems:

A) Windows CRLF line endings inside .alls

If your .alls was edited on Windows, convert it:

bash
sudo apt install dos2unix
dos2unix your_script.alls

Or:

bash
sed -i 's/\r$//' your_script.alls

B) Hidden non-ASCII quotes

If you copied flags like “hailo8l” (curly quotes) from web pages, replace them with normal quotes "hailo8l".

This single issue causes tons of “why won’t it parse” pain.

---

7) The “recommended minimal routine” (what I’d actually do)

1. Install Docker (Section 1).
2. Download Hailo AI Software Suite – Docker (Section 2).
3. Run ./hailo_ai_sw_suite_docker_run.sh (Section 3 Option A). (Macnica)
4. Inside Docker, run hailo tutorial once and follow the notebooks to generate a working HEF pipeline. (Hailo Community)
5. When your pipeline works, copy your commands into a script so you can reproduce it.

---

If you paste (a) your Hailo device type (Hailo-8 vs Hailo-8L), and (b) the exact ONNX model type (YOLOv5/8/9? classification? segmentation?), I can give you the exact hailo parser … end-node-names … and a sane starting .alls strategy.