dominik/hermes-skills

Fork 0

Files

T

Debian e9cc106625 Initial commit: Hermes Agent Skills collection

2026-07-12 19:02:59 +00:00

10 KiB

Raw Blame History

Frigate NVR Troubleshooting Reference

Infrastructure

Component	Value
CT ID	120
Node	n5pro (10.0.20.91) — migrated from proxmox4
IP	10.0.30.104
Web UI	http://10.0.30.104:5000
API	http://10.0.30.104:5000/api/
go2RTC	http://10.0.30.104:8554/
Config	/config/config.yml (inside CT)
Service	frigate.service (systemd)
Database	/config/frigate.db (SQLite)
GPU	AMD Radeon (gfx1150) via /dev/dri/card0 + /dev/dri/renderD128
GPU Driver	Mesa Gallium radeonsi, VAAPI 1.22 (H264/HEVC decode+encode)
onboot	yes

Version & Features

Frigate 0.16.4 (also tested 0.17-0)
Detector: OpenVINO with yolov9s Plus model (320x320), device: CPU (OpenVINO is Intel-only, AMD GPU unsupported)
Tracked objects: person, face, license_plate, dog, cat, car, amazon, ups, package (+ extended postal labels)
Face recognition: enabled (large model) — runs on CPU (OpenVINO GPU fails, auto-fallback to ONNX Runtime CPU)
License plate recognition: disabled (was device: GPU, OpenVINO doesn't support AMD)
Semantic search: enabled (jinav1 large) — runs on CPU (same fallback)
Hardware accel: none (commented out — see GPU Migration section)

Cameras

Einfahrt (Driveway)

IP: 10.0.50.102
RTSP main: rtsp://admin:<password>@10.0.50.102:554/h264Preview_01_main
RTSP sub: rtsp://admin:<password>@10.0.50.102:554/h264Preview_01_sub
Detection: 640x360 @ 5fps
Motion threshold: 47, contour area: 19
LPR mask zones configured

Terrasse (Patio)

IP: 10.0.50.103
Same RTSP pattern as Einfahrt
Motion threshold: 40, contour area: 27

MQTT Configuration

mqtt:
  enabled: true
  host: 10.0.30.10
  port: 1883
  user: frigate
  password: <see /config/config.yml or 1Password>
  topic_prefix: frigate
  client_id: frigate
  stats_interval: 60
  qos: 0

Verification Commands

# Check CT status (n5pro, NOT proxmox4)
ssh -i ~/.ssh/id_ed25519_proxmox root@10.0.20.91 "pct status 120"

# Start if stopped
ssh -i ~/.ssh/id_ed25519_proxmox root@10.0.20.91 "pct start 120"

# Check Frigate API
curl -s http://10.0.30.104:5000/api/version
curl -s http://10.0.30.104:5000/api/config | python3 -m json.tool | head -50

# Check service logs
ssh -i ~/.ssh/id_ed25519_proxmox root@10.0.20.91 "pct exec 120 -- cat /dev/shm/logs/frigate/current | tail -30"

# Reload HA Frigate integration (does NOT require full HA restart)
TOKEN=$(cat /tmp/ha_token.txt)
curl -s -X POST "http://10.0.30.10:8123/api/config/config_entries/entry/01KKSHKS4HXBAW1JAKDNHMYHKZ/reload" \
  -H "Authorization: Bearer $TOKEN"

GPU Migration: Intel → AMD (2026-07-12)

Frigate CT 120 was migrated from proxmox4 (Intel iGPU) to n5pro (AMD Radeon gfx1150). Three issues arose:

Issue 1: CT Device Mapping

Symptom: pct start 120 fails with "Device /dev/dri/card1 does not exist" Cause: CT config referenced /dev/dri/card1 (Intel) and renderD128 with gid=104 (Intel video group). n5pro has AMD GPU: /dev/dri/card0 (gid=44=video) and /dev/dri/renderD128 (gid=993=render). Fix: Update CT config device mappings:

dev0: /dev/dri/renderD128,gid=993
dev1: /dev/dri/card0,gid=44

Check gids on the target node first: stat -c '%g %G' /dev/dri/renderD128

Issue 2: OpenVINO EP Crash on AMD GPU

Symptom: Frigate crashes during startup with:

onnxruntime.capi.onnxruntime_pybind11_state.RuntimeException: RUNTIME_EXCEPTION:
Exception during initialization: ... OpenVINO-EP ... 
Check '!m_device_map.empty()' failed ... 
[GPU] Can't get OPTIMIZATION_CAPABILITIES property as no supported devices found

Cause: OpenVINO Execution Provider is Intel-only. It calls ov.compile_model(device_name="GPU") which fails on AMD. Affects: detectors (if device: GPU), face_recognition embeddings, semantic_search embeddings, LPR. Fix: Set device: CPU for detectors and LPR in config.yml. Face recognition and semantic search auto-fallback to CPU — the runner code catches the OpenVINO exception and logs "OpenVINO failed to build model, using CPU instead" then uses ONNX Runtime CPUExecutionProvider. Can face_recognition/semantic_search run on AMD GPU? Not with stock Frigate. Would need ONNX Runtime built with ROCm/HIP Execution Provider, which is not shipped with Frigate. The OpenVINO code path is hardcoded for Intel GPU.

Issue 3: FFmpeg Hardware Acceleration

Symptom: preset-intel-qsv-h264 causes ffmpeg crash. preset-vaapi-h264 also fails with "Unable to choose an output format". Cause: Intel QSV is Intel-only. preset-vaapi-h264 is NOT a valid Frigate FFmpeg preset name (Frigate interprets it as an output format, not a hwaccel arg). Fix: Comment out hwaccel_args entirely. CPU decode for 2 cameras is sufficient. AMD VAAPI could theoretically work with explicit ffmpeg args (e.g. -hwaccel vaapi -hwaccel_device /dev/dri/renderD128) but Frigate's preset system doesn't expose this cleanly. Note: vainfo on n5pro confirms AMD VAAPI works for H264/HEVC decode+encode. Installing mesa-va-drivers and vainfo packages on the Proxmox host enables it.

Config Changes Applied

# BEFORE (Intel)
ffmpeg:
  hwaccel_args: preset-intel-qsv-h264
detectors:
  ov:
    type: openvino
    device: GPU
lpr:
  enabled: true
  device: GPU

# AFTER (AMD)
ffmpeg:
  #hwaccel_args: preset-vaapi-h264  # commented out — CPU decode
detectors:
  ov:
    type: openvino
    device: CPU
lpr:
  enabled: false  # was device: GPU, OpenVINO doesn't support AMD
# face_recognition and semantic_search remain enabled — auto-fallback to CPU

Conflicting HA Addon

Critical: There is a SECOND Frigate running as an HA addon (ccab4aaf_frigate, repo ccab4aaf). It was in error state and conflicts with the CT-based Frigate. Must be stopped:

# Via HA CLI through qm guest exec
qm guest exec 106 -- sh -c 'ha addons stop ccab4aaf_frigate'

Also stop hm2mqtt addon (1376549a_hm2mqtt) if HAME Energy entities were cleaned up — otherwise it re-discovers them via MQTT:

qm guest exec 106 -- sh -c 'ha addons stop 1376549a_hm2mqtt'

Common Issues

CT Stopped (Most Common)

Symptom: All Frigate entities unavailable in HA, 10.0.30.104 unreachable. Fix: pct start 120 on n5pro (10.0.20.91), wait 60s, reload HA Frigate integration.

OOM Killer (FIXED 2026-06-27)

Symptom: Frigate process dies repeatedly, log shows "Out of memory: Killed process". Cause: 4GB RAM was insufficient for 2 cameras + face rec + LPR + embeddings + semantic search. Fix Applied: CT RAM increased 4GB → 8GB via pct set 120 -memory 8192 (requires CT stop/start). Current state: 1.2GB used, 6.9GB free — no OOM risk.

RTSP 461 Unsupported Transport

Symptom: Log shows method SETUP failed: 461 Unsupported transport. Impact: Transient, doesn't prevent operation. go2rtc falls back to TCP transport. Action: Ignore — recordings and detection still work.

Config File Confusion

/opt/frigate/config/config.yml — DEMO config (plays a test video). IGNORE.
/config/config.yml — REAL config with MQTT creds and camera RTSP URLs. THIS is loaded.
Frigate API /api/config shows the effective merged config (file + DB overrides + defaults).

HA Integration Not Reconnecting

Symptom: Frigate is running, MQTT is publishing, but HA entities stay unavailable. Cause: HA Frigate integration doesn't auto-reconnect when backend comes back. Fix: Reload via REST API POST /api/config/config_entries/entry/{ENTRY_ID}/reload. Full HA restart also works but is slower.

HA Integration Details

Config entry ID: 01KKSHKS4HXBAW1JAKDNHMYHKZ
Domain: frigate
Title: 10.0.30.104:5000
~96 entities (cameras, switches, sensors, binary_sensors, images)
Entity naming: <type>.einfahrt_<feature>_2, <type>.terrasse_<feature>_2

Frigate Update Path (0.16.4 → 0.17.2)

Why Not In-Place Update

The CT 120 Frigate was installed via community-scripts/ProxmoxVE (/bin/update exists). The community script's update command redirects to the full CT creation script — it does not support in-place updates. It creates a new CT from scratch.

Frigate itself distributes Docker images only (no .deb, no pip wheel since 0.8+). The bare-metal source build at /opt/frigate cannot be upgraded by simply pulling new source — it requires a full rebuild with correct deps.

Docker-in-LXC Dead End (Tested 2026-07-12)

Attempting to install Docker inside CT 120 (LXC) to use the official Docker image fails systematically:

Failure	Cause
`dockerd: iptables not found`	iptables binary missing, symlinks broken
`docker.socket: Failed to resolve group docker`	docker group not created during install
`failed to create NAT chain DOCKER: iptables not found`	Even after fixing iptables, bridge driver fails
dpkg database corruption	Interrupted apt-get leaves `/var/lib/dpkg/updates/` in bad state, `dpkg --configure -a` loops
`modprobe overlay` fails (ExecStartPre)	LXC lacks kernel module loading capability

Conclusion: Docker in Proxmox LXC is not viable. Frigate docs explicitly warn: "If you choose to run Frigate via LXC the setup can be problematic." Use a QEMU VM instead.

Recommended Migration Path

Create a new QEMU VM on n5pro (not LXC CT)
Install Debian 12 + Docker
Use ghcr.io/blakeblankshear/frigate:stable-rocm (AMD GPU ROCm build)
Copy /config/config.yml and /config/frigate.db from CT 120
Map /dev/dri/renderD128 for AMD GPU access
Update HA Frigate integration URL to new VM IP
Stop CT 120 Frigate once VM is verified

Docker Image Tags Reference

Tag	Use Case
`stable`	Standard amd64/arm64 (Intel iGPU, Coral)
`stable-rocm`	AMD GPU (our n5pro case)
`stable-tensorrt`	Nvidia GPU
`stable-rk`	Rockchip SBCs

Current State (Post-Cleanup 2026-07-12)

CT 120: Frigate 0.16.4 still active (restored after failed Docker attempt)
Docker packages purged from CT 120, cores reverted from 8 back to 4
CT 120 dpkg repaired (dpkg --configure -a --force-all)
HA addon ccab4aaf_frigate stopped (conflict)
HA addon 1376549a_hm2mqtt stopped (HAME energy bridge)

Double Take (Defunct)

Double Take (face recognition companion) was completely removed — no service, container, addon, or config remains. 13 orphaned entities linger in HA states but have no entity registry entries. They will disappear on HA restart. Do NOT attempt to restore Double Take unless reinstalling from scratch.

10 KiB Raw Blame History