Cannot compile a custom model for pi AI hat

General Discussion
1

Hello, I am trying to evaluate the hailo 10h on the pi ai hat, but I am having lots of issues compiling and running my own model.

Currently when I apt install hailo-10h-all on the pi it gives me the 5.1.1 tools. My development machine is on Rocky9, so I have no choice but to use the docker image provided. The 5.3.0 one produced an incompatible HEF file, and so I switched to the 5.1 docker image, and it still produces an incompatible HEF file.

I am not doing anything fancy, I am simply trying to use the resnet18 model from the ONNX model zoo to test things out.

I eventually get the error
```hailo_platform.pyhailort.pyhailort.HailoRTException: libhailort failed with error: 7 (HAILO_NOT_IMPLEMENTED)```

All my drivers and packages report version 5.1.1 on the pi, and I compiled directly inside the supplied docker image (But there is only an image supplied for the DFV v5.1.0. I imagine there should not be breaking changes between 5.1.0 and 5.1.1 though?). Am I missing anything? I was able to get the examples to run, but once I deviated things got very messy.

The documentation is extremely confusing and inconsistent, and there are so many utilities and things to install. Version mismatches, massive dependencies, etc etc. I am quite surprised at the friction here, as I imagined a pi hat product would require much better support. There is not (as far as I can tell) a straightforward guide for compiling and running a custom model. Can anyone direct me to one?

2

Welcome to the Hailo Community!

I just installed the Hailo AI Software Suite Docker 2025-10 (DFC 5.1.0) and compiled the yolov11n model using the hailomz CLI. I then ran the model on the Raspberry Pi 5 using the hailortcli run2 set-net yolov11.n.hef command and everything worked as expected.

I recommend using the Hailo AI Software Suite Docker for the model conversion. It simplifies the installation and updates.

The Hailo AI Software Suite Docker includes tutorials covering the model-conversion workflow. Inside the Docker environment, run the following command to start a Jupyter Notebook server with notebooks for each step of the conversion process:

hailo tutorial

Rocky Linux 9 is not officially supported/validated, so there may be compatibility issues.

3

Hi!

For models like resnet everything is available on their Github and on the Developer Zone to install the DFC like @Klaus_Koschinsky said.

I also made 2 GitHub repositories on how to compile models on Hailo chips if I can help :

Getting started with Hailo-8L using MNIST – full training-to-inference example (old account)
I compiled an LLM for the hailo-10h

1 Like
4

Thanks all for the help.

As I said in the post. I did do this but I still ran into issues unfortunately. I eventually used the hailort benchmark command to test that my model did actually work, so I know I have a valid .hef file. It seems some environment issue on the pi is preventing me from integrating it into an application for now.

That is good to know thank you. Is the hailo tutorial command documented anywhere online? I had not encountered it. I am sure there should at least be a step by step guide published online which does not require using a docker image.

Yes there are a few compatibility issues with the docker images, but I managed to work around them by changing some of the volume mounts. It tries to mount eg /etc/timezone which doesn’t exist on my machine. It wasn’t too bad to remove these mounts though. Since only Ubuntu is officially supported, it would be good if the docker file run script did not assume a debian host. It might widen compatibility, but like I said it wasnt too hard to work around this from Rocky9.

Cheers