Reported by @doc62fr on #1771 on 2026-07-11 and not tracked separately until now, which is our miss. Raising it as its own issue because it blocks the Hailo slice for the one person who can actually test it.
Report
"Lors de l'installation il ne reconnait pas l'Accelerateur AI HAT+2 hailo 10H"
The installer does not detect the AI HAT+2 (Hailo-10H) during install on a Raspberry Pi 5. The tester's board is confirmed working: he runs LLMs on it through hailo-ollama, firmware 5.1.0, and the device is present.
Why it matters
Detection is the gate for everything downstream in #1771. If the installer does not see the accelerator, it will not chain into scripts/install-hailo.sh, so checklist step 4 (installer chain) cannot pass and the backend never gets configured. The tester worked around it by installing hailo-ollama manually via Hailo's own route, which is how we ended up with a server on port 8000 outside taOS management (see #2083).
What to check
- What the detection path actually probes. The install script logs
hailortcli fw-control identify when hailortcli is on PATH, but the earlier hardware-detection step in the main installer is the one that failed here.
- Whether detection depends on
/dev/hailo0 existing at install time, and whether that node is present before HailoRT is installed. There is a TAOS_FORCE_HAILO override precisely because bench boxes lack the node, which suggests the node is the primary signal.
- Ordering: if HailoRT provides both the node and
hailortcli, then detection running before HailoRT is installed can never succeed on a clean machine, and the override is masking a sequencing bug rather than a hardware quirk.
- Whether the 40 TOPS 10H is distinguished correctly from the 13 and 26 TOPS Hailo-8 parts, since the capability map treats them differently.
Next step
Ask the tester to run the detection path on his current setup and paste the output, since his environment has changed since 11 July. He also has error captures from install attempts 4 and 5 that were offered and never collected.
Related: #1771 (slice), #2083 (installer assumes a clean machine), #1535 (setup wizard never detects the NPU, same class for Rockchip).
Reported by @doc62fr on #1771 on 2026-07-11 and not tracked separately until now, which is our miss. Raising it as its own issue because it blocks the Hailo slice for the one person who can actually test it.
Report
The installer does not detect the AI HAT+2 (Hailo-10H) during install on a Raspberry Pi 5. The tester's board is confirmed working: he runs LLMs on it through
hailo-ollama, firmware 5.1.0, and the device is present.Why it matters
Detection is the gate for everything downstream in #1771. If the installer does not see the accelerator, it will not chain into
scripts/install-hailo.sh, so checklist step 4 (installer chain) cannot pass and the backend never gets configured. The tester worked around it by installing hailo-ollama manually via Hailo's own route, which is how we ended up with a server on port 8000 outside taOS management (see #2083).What to check
hailortcli fw-control identifywhenhailortcliis on PATH, but the earlier hardware-detection step in the main installer is the one that failed here./dev/hailo0existing at install time, and whether that node is present before HailoRT is installed. There is aTAOS_FORCE_HAILOoverride precisely because bench boxes lack the node, which suggests the node is the primary signal.hailortcli, then detection running before HailoRT is installed can never succeed on a clean machine, and the override is masking a sequencing bug rather than a hardware quirk.Next step
Ask the tester to run the detection path on his current setup and paste the output, since his environment has changed since 11 July. He also has error captures from install attempts 4 and 5 that were offered and never collected.
Related: #1771 (slice), #2083 (installer assumes a clean machine), #1535 (setup wizard never detects the NPU, same class for Rockchip).