Found while a community tester (@doc62fr) worked through #1771 on real hardware, a Raspberry Pi 5 with an AI HAT+2. Nobody on the team has this hardware, so this only surfaced because he is testing it.
The gap
scripts/install-hailo.sh assumes a clean machine. It does not check whether hailo-ollama is already installed and running from Hailo's own route (their installer, a desktop shortcut, or a manual build).
In the reported case the tester had upstream hailo-ollama serving on 0.0.0.0:8000, which is its default. Our script remaps to 7836 by setting Environment=OLLAMA_HOST=127.0.0.1:7836 in the unit it installs, so the two coexist invisibly: a working server on 8000 that taOS cannot see, no hailo-ollama.service, and nothing on 7836. From the user's side taOS simply appears broken while their Hailo setup works fine, which is a confusing place to leave someone.
What it should do
Before installing, detect a pre-existing upstream instance (something listening on 8000 that answers /api/tags, or an existing binary/service not installed by us) and then either:
- Adopt it: reconfigure the existing install to the taOS port and manage it under
hailo-ollama.service, or
- Say so plainly: report what was found, what will change, and what will be left alone, before touching anything.
Silently building a second server alongside the first is the one behaviour to avoid, since both then compete for the accelerator.
Notes
Found while a community tester (@doc62fr) worked through #1771 on real hardware, a Raspberry Pi 5 with an AI HAT+2. Nobody on the team has this hardware, so this only surfaced because he is testing it.
The gap
scripts/install-hailo.shassumes a clean machine. It does not check whether hailo-ollama is already installed and running from Hailo's own route (their installer, a desktop shortcut, or a manual build).In the reported case the tester had upstream hailo-ollama serving on
0.0.0.0:8000, which is its default. Our script remaps to 7836 by settingEnvironment=OLLAMA_HOST=127.0.0.1:7836in the unit it installs, so the two coexist invisibly: a working server on 8000 that taOS cannot see, nohailo-ollama.service, and nothing on 7836. From the user's side taOS simply appears broken while their Hailo setup works fine, which is a confusing place to leave someone.What it should do
Before installing, detect a pre-existing upstream instance (something listening on 8000 that answers
/api/tags, or an existing binary/service not installed by us) and then either:hailo-ollama.service, orSilently building a second server alongside the first is the one behaviour to avoid, since both then compete for the accelerator.
Notes