HEF was compiled for Hailo8L device, while the device itself is Hailo8.

[s2d4]

Describe the problem you are having

Currently the inference speed is at ~11ms but it should be better as Hailo8 is 26TOPS vs 13 TOPS of Hailo8l?

It appears the YOLOv6n Model was compiled for Hailo8l and thus capping performance on the Hailo8?

Would it not be better the other way around?

Version

0.16.0-1fa7ce5

Frigate config file

detectors:
  hailo8l:
    type: hailo8l
    device: PCIe

model:
  width: 320
  height: 320
  input_tensor: nhwc
  input_pixel_format: rgb
  input_dtype: int
  model_type: yolo-generic

docker-compose file or Docker CLI command

devices:
      - /dev/hailo0:/dev/hailo0

Relevant Frigate log output

2025-05-15 02:00:45.278765960  [HailoRT] [warning] HEF was compiled for Hailo8L device, while the device itself is Hailo8. This will result in lower performance.

Install method

Docker Compose

Object Detector

Other

Screenshots of the Frigate UI's System metrics pages

Any other information that may be helpful

No response

[NickM-27]

That's only a couple ms slower than the best time we've seen, though the code is supposed to detect the correct hardware

[s2d4]

As per the official communication https://community.hailo.ai/t/hailo-official-integration-with-frigate/13679

Basically matching here with the logs considering it is reported to be running as a hailo8l when it should be running as hailo8.

What can I do to make it run like it should? or will this need to be fixed in the next release?

btw, the install pcie drivers script didn't work if I already have a higher version installed previously. I had to do a full compile to get it back down to 4.20.1

[NickM-27]

The code that the Hailo team has implemented already tries to pull the correct model. You could try deleting it in config/model_cache so it redownloads

[s2d4]

Thanks for pointing in the right direction! I am not sure why it pulled the wrong HEF file, does it pull the file before the driver version check failure?
If so (It looks like it was pulled when I was having issues with being on v4.21.0. ), perhaps it should be pulled after the driver check succeeds?

[arigit]

Just chiming in to say thank-you @NickM-27 for the answer. Deleting the model cache removed the warning message and brought the inference times to the 6ms range, very impressive.

What confused me initially is that to get frigate to initialize I had to use type: hailo8l instead of type: hailo8 as I assumed I had to use.

detectors:
  hailo:
    type: hailo8l
    device: PCIe