[HW Accel Support]: NVIDIA QUADRO K2200 GPU - Model / Configuration Need help #21006

Gis70 · 2025-11-22T12:58:33Z

Gis70
Nov 22, 2025

Describe the problem you are having

Hello Folks,

I'm am already using FRIGATE 0.16.2 with a USB CORAL on a HPE Microserver Gen8.
I just got a HPE ML30 Gen9 on which i put a NVIDIA QUADRO K2200.
==> I have realized that it is not easy at all to make detection working with a GPU compared to a CORAL TPU.

I have understood that there is no out of the box configuration / model.

So i have first took the docker image ghcr.io/blakeblackshear/frigate:stable-tensorrt

NVIDIA GPU is found but i have an error when i try to use the model yolox_m.onnx directly downloaded from https://github.com/Megvii-BaseDetection/YOLOX/tree/main/demo/ONNXRuntime according to the documentation in Frigate.

I would first try to use a model already done whitout to having to build something or convert it by myself.

I have installed NVIDA driver : NVIDIA-Linux-x86_64-580.105.08.run et nvidia-smi

Here we don't see the FFmpeg process because the container is stopped but when it is up with the CPU as detector we can see all my camera with FFmpeg process.

I somebody could guide me to make everything working it would be good.

Thank you.

Version

0.16.2-4d58206

Frigate config file

database:
  path: /db/frigate.db

tls:
  enabled: false

auth:
  reset_admin_password: false

mqtt:
  enabled: true
  user: '{FRIGATE_MQTT_USER}'
  password: '{FRIGATE_MQTT_PASSWORD}'
  host: 192.168.0.71
  port: 1883

logger:
  # Optional: Default log verbosity (default: shown below)
  default: info
  # Optional: Component specific logger overrides
  logs:
    frigate.event: info
record:
  sync_recordings: true
  enabled: true

go2rtc:
  ffmpeg:
  hwaccel_args: preset-nvidia
  streams:
    DOORBIRD: 
      ffmpeg:http://xxxx:[email protected]/bha-api/video.cgi#video=h264#hardware=cuda
      #- rtsp://xxxxx:[email protected]:554/mpeg/1080p/media.amp
    CUISINE:
      - rtsp://{FRIGATE_RTSP_USER}:[email protected]:554/1#video=h264#hardware=cuda
    CUISINE_LR:
      - rtsp://{FRIGATE_RTSP_USER}:[email protected]:554/2#video=h264#hardware=cuda
    TERRASSE:
      - rtsp://{FRIGATE_RTSP_USER}:{FRIGATE_RTSP_PASSWORD}@192.168.0.78:554/h264Preview_01_main
    TERRASSE_LR:
      - rtsp://{FRIGATE_RTSP_USER}:{FRIGATE_RTSP_PASSWORD}@192.168.0.78:554/h264Preview_01_sub
    ENTREE:
      - rtsp://{FRIGATE_RTSP_USER}:{FRIGATE_RTSP_PASSWORD}@192.168.0.79:554/h264Preview_01_main
    ENTREE_LR:
      - rtsp://{FRIGATE_RTSP_USER}:{FRIGATE_RTSP_PASSWORD}@192.168.0.79:554/h264Preview_01_sub
    TERRAIN:
      - rtsp://{FRIGATE_RTSP_USER}:{FRIGATE_RTSP_PASSWORD}@192.168.0.98:554/h264Preview_01_main
      - ffmpeg:TERRAIN#video=h264#hardware
    TERRAIN_LR:
      - rtsp://{FRIGATE_RTSP_USER}:{FRIGATE_RTSP_PASSWORD}@192.168.0.98:554/h264Preview_01_sub
    GARAGE:
      - rtsp://{FRIGATE_RTSP_USER}:{FRIGATE_RTSP_PASSWORD}@192.168.0.11:554/h264Preview_01_main
    GARAGE_LR:
      - rtsp://{FRIGATE_RTSP_USER}:{FRIGATE_RTSP_PASSWORD}@192.168.0.11:554/h264Preview_01_sub
    DERRIERE:
      - rtsp://{FRIGATE_RTSP_USER}:{FRIGATE_RTSP_PASSWORD}@192.168.0.85:554/h264Preview_01_main
    DERRIERE_LR:
      - rtsp://{FRIGATE_RTSP_USER}:{FRIGATE_RTSP_PASSWORD}@192.168.0.85:554/h264Preview_01_sub
    COUR:
      - rtsp://{FRIGATE_RTSP_USER}:{FRIGATE_RTSP_PASSWORD}@192.168.0.87:554/h264Preview_01_main
      - ffmpeg:COUR#video=h264#hardware
    COUR_LR:
      - rtsp://{FRIGATE_RTSP_USER}:{FRIGATE_RTSP_PASSWORD}@192.168.0.87:554/h264Preview_01_sub

detectors:
  onnx:
    type: onnx

model:
  model_type: yolo-generic
  width: 640 # <--- should match the imgsize set during model export
  height: 640 # <--- should match the imgsize set during model export
  input_tensor: nchw
  input_dtype: float
  path: /config/model_cache/yolox_m.onnx
  labelmap_path: /config/model_cache/coco_80.txt

cameras:
  DOORBIRD:
    ffmpeg:
      hwaccel_args: preset-nvidia-h264
      # input_args: -avoid_negative_ts make_zero -fflags nobuffer -flags low_delay -strict experimental -fflags +genpts+discardcorrupt -use_wallclock_as_timestamps 1
      #output_args:
         #record: -f segment -segment_time 10 -segment_format mp4 -reset_timestamps 1 -strftime 1 -c:v libx264 -an
      inputs:
        - path: rtsp://127.0.0.1:8554/DOORBIRD
          roles:
           # - detect
            - record
    # detect:
    #   enabled: false
    #   width: 1280
    #   height: 720
    #   fps: 5
    # snapshots:
    #   enabled: false
    #   timestamp: true
    #   bounding_box: true
    #   retain:
    #     default: 15
    record:
      enabled: true
      retain:
        days: 1
    #     mode: active_objects
    #   events:
    #     #pre_capture: 5
    #     #post_capture: 5
    #     retain:
    #       default: 10
    #       mode: motion

    # motion:
    #   mask: 
    #     0.001,0.66,0.277,0.668,0.78,0.755,0.789,0.403,0.816,0.397,0.842,0.353,0.885,0.207,0.942,0.001,0,0
    # zones:
    #   pain:
    #     coordinates: 0.862,0.849,0.75,0.777,0.715,0.996,0.858,0.999
    #     loitering_time: 0
    #   zone_portier:
    #     coordinates: 0,0.668,0.728,0.758,0.706,0.995,0.003,0.997
    #     loitering_time: 0
    # review:
    #   detections:
    #     required_zones:
    #       - pain
    #       - zone_portier
    motion:
      threshold: 40
      contour_area: 20
      improve_contrast: 'false'
      mask: 0,0,0,0.581,0.774,0.723,0.799,0.435,0.902,0
    zones:
      portier_video:
        coordinates: 
          0.793,0.57,0.775,0.781,0.286,0.704,0,0.725,0,0.982,0.994,0.998,0.996,0.18
        loitering_time: 0
    review:
      alerts:
        required_zones: portier_video
      detections:
        required_zones: portier_video
  CUISINE:
    ffmpeg:
      hwaccel_args: preset-nvidia-h264
      inputs:
        - path: rtsp://127.0.0.1:8554/CUISINE
          roles:
            - record
        - path: rtsp://127.0.0.1:8554/CUISINE_LR
          roles:
            - detect
    onvif:
      host: 192.168.0.38
      port: 8080
      user: admin
      password: Hvalentinik461
    detect:
      enabled: false
      width: 640
      height: 352
      fps: 5
    motion:
  # Optional: The threshold passed to cv2.threshold to determine if a pixel is different enough to be counted as motion. (default: shown below)
  # Increasing this value will make motion detection less sensitive and decreasing it will make motion detection more sensitive.
  # The value should be between 1 and 255.
      threshold: 40 #default 30
        # Optional: Minimum size in pixels in the resized motion image that counts as motion (default: shown below)
        # Increasing this value will prevent smaller areas of motion from being detected. Decreasing will
        # make motion detection more sensitive to smaller moving objects.
        # As a rule of thumb:
        #  - 10 - high sensitivity
        #  - 30 - medium sensitivity
        #  - 50 - low sensitivity
      contour_area: 30
  # Optional: The percentage of the image used to detect lightning or other substantial changes where motion detection
  #           needs to recalibrate. (default: shown below)
  # Increasing this value will make motion detection more likely to consider lightning or ir mode changes as valid motion.
  # Decreasing this value will make motion detection more likely to ignore large amounts of motion such as a person approaching
  # a doorbell camera.
      lightning_threshold: 0.65
      mask:
        - 1,0,1,0.077,0.511,0.088,0.509,0
        - 0.307,0.269,0.406,0.246,0.423,0.438,0.313,0.446
    objects:
      track:
        - person
        - animal
      mask: 0.547,0.577,0.719,0.559,0.919,0.764,0.938,0.803,0.648,0.901
    snapshots:
      enabled: false
      timestamp: false
      bounding_box: true
      retain:
        default: 15
    record:
      enabled: true
      retain:
        days: 1
        mode: all
      alerts:
        retain:
          days: 15
          mode: motion
      detections:
        retain:
          days: 15
          mode: motion
  TERRASSE:
    ffmpeg:
      hwaccel_args: preset-nvidia-h264
      inputs:
        - path: rtsp://127.0.0.1:8554/TERRASSE
          roles:
            - record
        - path: rtsp://127.0.0.1:8554/TERRASSE_LR
          roles:
            - detect
    detect:
      enabled: false
      width: 640
      height: 480
      fps: 4
    motion:
  # Optional: The threshold passed to cv2.threshold to determine if a pixel is different enough to be counted as motion. (default: shown below)
  # Increasing this value will make motion detection less sensitive and decreasing it will make motion detection more sensitive.
  # The value should be between 1 and 255.
      threshold: 60
  # Optional: The percentage of the image used to detect lightning or other substantial changes where motion detection
  #           needs to recalibrate. (default: shown below)
  # Increasing this value will make motion detection more likely to consider lightning or ir mode changes as valid motion.
  # Decreasing this value will make motion detection more likely to ignore large amounts of motion such as a person approaching
  # a doorbell camera.
      #lightning_threshold: 0.8
  # Optional: Minimum size in pixels in the resized motion image that counts as motion (default: shown below)
  # Increasing this value will prevent smaller areas of motion from being detected. Decreasing will
  # make motion detection more sensitive to smaller moving objects.
  # As a rule of thumb:
  #  - 10 - high sensitivity
  #  - 30 - medium sensitivity
  #  - 50 - low sensitivity
      contour_area: 30
  # Optional: Alpha value passed to cv2.accumulateWeighted when averaging frames to determine the background (default: shown below)
  # Higher values mean the current frame impacts the average a lot, and a new object will be averaged into the background faster.
  # Low values will cause things like moving shadows to be detected as motion for longer.
  # https://www.geeksforgeeks.org/background-subtraction-in-an-image-using-concept-of-running-average/
      #frame_alpha: 0.01
  # Optional: Height of the resized motion frame  (default: 100)
  # Higher values will result in more granular motion detection at the expense of higher CPU usage.
  # Lower values result in less CPU, but small changes may not register as motion.
      #frame_height: 100
  # Optional: motion mask
  # NOTE: see docs for more detailed info on creating masks
      #mask: 0,900,1080,900,1080,1920,0,1920
  # Optional: improve contrast (default: shown below)
  # Enables dynamic contrast improvement. This should help improve night detections at the cost of making motion detection more sensitive
  # for daytime.
      #improve_contrast: True
  # Optional: Delay when updating camera motion through MQTT from ON -> OFF (default: shown below).
      #mqtt_off_delay: 30
      mask:
        - 0.519,0.001,0.483,0.301,0.556,0.322,0.599,0.001
        - 0.785,0.15,0.909,1,0.965,0.999,0.999,0.002,0.676,0.005,0.682,0.061
    objects:
      track:
        - person
        - cat
        - dog
        - animal

    snapshots:
      enabled: false
      timestamp: false
      bounding_box: true
      retain:
        default: 15
    record:
      enabled: true
      retain:
        days: 3
        mode: all
      alerts:
        retain:
          days: 10
          mode: motion
      detections:
        retain:
          days: 10
          mode: motion
    zones:
      zone_terrasse:
        coordinates: 
          0.001,0.993,0.754,0.998,0.746,0.24,0.772,0.148,0.609,0.017,0.566,0.322,0.56,0.339,0.484,0.317,0.466,0.304,0.503,0.004,0.08,0,0.085,0.332,0,0.416
        loitering_time: 0
        inertia: 3
    review:
      detections:
        required_zones: zone_terrasse
      alerts:
        required_zones: zone_terrasse
  COUR:
    ffmpeg:
      hwaccel_args: preset-nvidia-h264
      inputs:
        - path: rtsp://127.0.0.1:8554/COUR
          roles:
            - record
        - path: rtsp://127.0.0.1:8554/COUR_LR
          roles:
            - detect
    #live:
      #stream_name: COUR
    detect:
      enabled: false
      width: 1536
      height: 576
      fps: 4
    motion:
  # Optional: The threshold passed to cv2.threshold to determine if a pixel is different enough to be counted as motion. (default: shown below)
  # Increasing this value will make motion detection less sensitive and decreasing it will make motion detection more sensitive.
  # The value should be between 1 and 255.
      threshold: 60
  # Optional: The percentage of the image used to detect lightning or other substantial changes where motion detection
  #           needs to recalibrate. (default: shown below)
  # Increasing this value will make motion detection more likely to consider lightning or ir mode changes as valid motion.
  # Decreasing this value will make motion detection more likely to ignore large amounts of motion such as a person approaching
  # a doorbell camera.
      #lightning_threshold: 0.8
  # Optional: Minimum size in pixels in the resized motion image that counts as motion (default: shown below)
  # Increasing this value will prevent smaller areas of motion from being detected. Decreasing will
  # make motion detection more sensitive to smaller moving objects.
  # As a rule of thumb:
  #  - 10 - high sensitivity
  #  - 30 - medium sensitivity
  #  - 50 - low sensitivity
      contour_area: 30
  # Optional: Alpha value passed to cv2.accumulateWeighted when averaging frames to determine the background (default: shown below)
  # Higher values mean the current frame impacts the average a lot, and a new object will be averaged into the background faster.
  # Low values will cause things like moving shadows to be detected as motion for longer.
  # https://www.geeksforgeeks.org/background-subtraction-in-an-image-using-concept-of-running-average/
      #frame_alpha: 0.01
  # Optional: Height of the resized motion frame  (default: 100)
  # Higher values will result in more granular motion detection at the expense of higher CPU usage.
  # Lower values result in less CPU, but small changes may not register as motion.
      #frame_height: 100
  # Optional: motion mask
  # NOTE: see docs for more detailed info on creating masks
      #mask: 0,900,1080,900,1080,1920,0,1920
  # Optional: improve contrast (default: shown below)
  # Enables dynamic contrast improvement. This should help improve night detections at the cost of making motion detection more sensitive
  # for daytime.
      #improve_contrast: True
  # Optional: Delay when updating camera motion through MQTT from ON -> OFF (default: shown below).
      #mqtt_off_delay: 30
      mask:
        - 0.001,0,1,0,1,0.762,0.77,0.302,0.694,0.182,0.584,0.071,0.544,0.051,0.479,0.031,0.453,0.051,0.409,0.091,0.393,0.109,0.309,0.134,0.295,0.142,0.27,0.182,0.052,0.632,0,0.716
        - 0.574,0.046,0.6,0.299,0.513,0.549,0.399,0.555,0.363,0.374,0.379,0.222,0.399,0.103,0.435,0.06,0.542,0.048
        - 0.947,0.644,0.929,0.75,1,0.835,1,0.758
        - 0.486,0.58,0.489,0.994,0.598,0.998,0.582,0.598
    objects:
      track:
        - person
        - cat
        - car
        - motorcycle
        - bicycle
        - dog
        - animal
      # filters:
      #   car:
      #     mask:
      #       - 359,30,499,45,640,21,640,0,249,0,247,57
    snapshots:
      enabled: false
      timestamp: false
      bounding_box: true
      retain:
        default: 15
    record:
      enabled: true
      retain:
        days: 2
        mode: all
      alerts:
        retain:
          days: 10
          mode: motion
      detections:
        retain:
          days: 10
          mode: motion
    zones:
      zone_cour:
        coordinates: 
          0.033,0.665,0.299,0.147,0.391,0.11,0.401,0.139,0.384,0.238,0.386,0.354,0.408,0.526,0.489,0.553,0.496,0.992,0.399,0.998,0.388,1,0.368,1,0.339,0.992,0.153,0.986,0.054,0.868,0.117,0.791,0.043,0.706
        loitering_time: 1
        inertia: 3
      zone_puits:
        coordinates: 
          0.594,0.323,0.936,0.863,0.948,0.793,0.975,0.809,0.996,0.832,0.999,0.983,0.604,1,0.591,0.681,0.587,0.586,0.536,0.557
        loitering_time: 1
        inertia: 3
        objects:
          - animal
          - bicycle
          - cat
          - dog
          - motorcycle
          - person
    review:
      detections:
        required_zones:
          - zone_cour
          - zone_puits
      alerts:
        required_zones:
          - zone_cour
          - zone_puits
  ENTREE:
    ffmpeg:
      hwaccel_args: preset-nvidia-h264
      inputs:
        - path: rtsp://127.0.0.1:8554/ENTREE
          roles:
            - record
        - path: rtsp://127.0.0.1:8554/ENTREE_LR
          roles:
            - detect
    detect:
      enabled: false
      width: 640
      height: 480
      fps: 4
    objects:
      track:
        - person
        - cat
        - motorcycle
        - bicycle
        - dog
      filters:
        person: {}
    snapshots:
      enabled: false
      timestamp: false
      bounding_box: true
      retain:
        default: 15
    record:
      enabled: true
      retain:
        days: 2
        mode: all
      alerts:
        retain:
          days: 10
          mode: motion
      detections:
        retain:
          days: 10
          mode: motion
    zones:
      zone_entree:
        coordinates: 
          0,0.139,0.162,0.087,0.349,0.082,0.424,0.001,0.472,0.004,0.501,0.285,0.53,0.35,0.742,0.366,0.803,0.422,0.816,0.494,0.901,0.486,0.99,0.51,0.998,0.675,0.85,0.675,0.938,0.999,0.002,0.998
        inertia: 3
        loitering_time: 0
        objects:
          - person
          - dog
          - cat
    motion:
      contour_area: 30

      mask:
        - 0.682,0.085,0.741,0.087,0.747,0.171,0.707,0.17,0.707,0.111
        - 0.483,0,0.491,0.151,0.515,0.157,0.501,0.001
        - 0.106,0.677,0.233,0.742,0.193,0.882,0.06,0.833
    review:
      detections:
        required_zones: zone_entree
      alerts:
        required_zones: zone_entree
  TERRAIN:
    ffmpeg:
      hwaccel_args: preset-nvidia-h264
      inputs:
        - path: rtsp://127.0.0.1:8554/TERRAIN
          roles:
            - record
        - path: rtsp://127.0.0.1:8554/TERRAIN_LR
          roles:
            - detect
    #live:
      #stream_name: TERRAIN
    onvif:
      host: 192.168.0.16
      port: 8000
      user: admin
      password: Rvalentinik46K
      autotracking:
        # Optional: enable/disable object autotracking. (default: shown below)
        enabled: false
    detect:
      enabled: false
      width: 640
      height: 360
      fps: 4
    motion:
        # Optional: Minimum size in pixels in the resized motion image that counts as motion (default: shown below)
  # Increasing this value will prevent smaller areas of motion from being detected. Decreasing will
  # make motion detection more sensitive to smaller moving objects.
  # As a rule of thumb:
  #  - 10 - high sensitivity
  #  - 30 - medium sensitivity
  #  - 50 - low sensitivity
      contour_area: 50
      mask:
        - 0.001,0.001,0.304,0,0.305,0.064,0.003,0.062
        - 0.586,0.001,0.596,0.142,0.649,0.143,0.654,0.001
        - 0.706,0.041,0.65,0.312,0.753,0.45,0.887,0.486,0.95,0.375,0.983,0.238,0.999,0.171,0.995,0.005
      threshold: 50
      improve_contrast: true
    objects:
      track:
        - person
        - cat
        - motorcycle
        - bicycle
        - dog
      filters:
        person: {}
    snapshots:
      enabled: false
      timestamp: false
      bounding_box: true
      retain:
        default: 30
    record:
      enabled: true
      retain:
        days: 2
        mode: all
      alerts:
        retain:
          days: 10
          mode: motion
      detections:
        retain:
          days: 10
          mode: motion
    zones:
      zone_terrain:
        coordinates: 
          0,0.383,0.103,0.336,0.189,0.299,0.211,0.263,0.151,0.274,0.143,0.24,0.215,0.244,0.258,0.263,0.233,0.307,0.278,0.329,0.341,0.281,0.365,0.245,0.374,0.184,0.429,0.178,0.477,0.16,0.537,0.146,0.586,0.149,0.644,0.159,0.672,0.175,0.65,0.307,0.745,0.451,0.883,0.497,0.945,0.393,0.997,0.195,0.997,0.995,0.002,0.999
        loitering_time: 0
    review:
      detections:
        required_zones: zone_terrain
      alerts:
        required_zones: zone_terrain
  GARAGE:
    ffmpeg:
      hwaccel_args: preset-nvidia-h264
      inputs:
        - path: rtsp://127.0.0.1:8554/GARAGE
          roles:
            - record
        - path: rtsp://127.0.0.1:8554/GARAGE_LR
          roles:
            - detect
    detect:
      enabled: false
      width: 640
      height: 480
      fps: 4
      min_initialized: 10
      max_disappeared: 35
    motion:
      threshold: 80
      lightning_threshold: 0.8
      contour_area: 50
      #frame_alpha: 0.02
      #frame_height: 50
      #improve_contrast: False
      mask:
        - 219,43,406,40,640,67,640,0,0,0,0,76
      improve_contrast: true
    objects:
      track:
        - person
        - cat
        - dog
        - animal
        #- car
        #- motorcycle
      filters:
        person:
      #     # Optional: minimum width*height of the bounding box for the detected object (default: 0)
      #     min_area: 5000
      #     # Optional: maximum width*height of the bounding box for the detected object (default: 24000000)
      #     max_area: 100000
      #     # Optional: minimum width/height of the bounding box for the detected object (default: 0)
      #     min_ratio: 0.5
      #     # Optional: maximum width/height of the bounding box for the detected object (default: 24000000)
      #     max_ratio: 2.0
      #     # Optional: minimum score for the object to initiate tracking (default: shown below)
      #     min_score: 0.5
      #     # Optional: minimum decimal percentage for tracked object's computed score to be considered a true positive (default: shown below)
          threshold: 0.75
      #     # Optional: mask to prevent this object type from being detected in certain areas (default: no mask)
      #     # Checks based on the bottom center of the bounding box of the object
      #     #mask: 0.000,0.000,0.781,0.000,0.781,0.278,0.000,0.278
    snapshots:
      enabled: false
      timestamp: false
      bounding_box: true
      retain:
        default: 10
    record:
      enabled: true
      retain:
        days: 3
        mode: motion
      alerts:
        retain:
          days: 10
          mode: motion
      detections:
        retain:
          days: 10
          mode: motion
    zones:
      zone_0:
        coordinates: 
          0.639,0.095,0.822,0.152,0.846,0.391,0.827,0.676,0.232,0.625,0.214,0.291,0.17,0.136,0.347,0.098

        inertia: 4
        loitering_time: 1
    review:
      detections:
        required_zones: zone_0
      alerts:
        required_zones: zone_0
  DERRIERE:
    ffmpeg:
      hwaccel_args: preset-nvidia-h264
      inputs:
        - path: rtsp://127.0.0.1:8554/DERRIERE
          roles:
            - record
        - path: rtsp://127.0.0.1:8554/DERRIERE_LR
          roles:
            - detect
    onvif:
      host: 192.168.0.85
      port: 8000
      user: admin
      password: Rvalentinik46K
      autotracking:
        # Optional: enable/disable object autotracking. (default: shown below)
        enabled: false
    detect:
      enabled: false
      width: 640
      height: 480
      fps: 4
    motion:
        # Optional: Minimum size in pixels in the resized motion image that counts as motion (default: shown below)
  # Increasing this value will prevent smaller areas of motion from being detected. Decreasing will
  # make motion detection more sensitive to smaller moving objects.
  # As a rule of thumb:
  #  - 10 - high sensitivity
  #  - 30 - medium sensitivity
  #  - 50 - low sensitivity
      contour_area: 40
      mask:
        - 0.002,0.001,0.394,0.001,0,0.407
        - 0.616,1,1,1,1,0,0.68,0
        - 0.434,0.001,0.455,0.098,0.516,0.092,0.53,0.114,0.619,0.074,0.622,0
        - 0,0.744,0.107,0.801,0.096,0.998,0,0.999
    objects:
      track:
        - person
        - cat
        - dog
        - animal

    snapshots:
      enabled: false
      timestamp: false
      bounding_box: true
      retain:
        default: 15
    record:
      enabled: true
      retain:
        days: 2
        mode: all
      alerts:
        retain:
          days: 10
          mode: motion
      detections:
        retain:
          days: 10
          mode: motion
    zones:
      zone_derriere:
        coordinates: 
          0.514,0.097,0.444,0.119,0.437,0.188,0.002,0.449,0.015,0.737,0.115,0.793,0.101,0.988,0.61,0.993,0.645,0.45,0.635,0.364,0.638,0.269,0.59,0.237,0.585,0.153,0.617,0.15,0.623,0.08,0.53,0.122
        loitering_time: 0
    review:
      detections:
        required_zones: zone_derriere
      alerts:
        required_zones: zone_derriere
version: 0.16-0
detect:
  enabled: true
semantic_search:
  enabled: true
  model_size: large
face_recognition:
  enabled: false
  model_size: small
lpr:
  enabled: true
classification:
  bird:
    enabled: false

docker-compose file or Docker CLI command

version: "3.9"
services:
  frigate:
    container_name: frigate
    privileged: true # this may not be necessary for all setups
    restart: always
    image: ghcr.io/blakeblackshear/frigate:stable-tensorrt
    deploy:    # <------------- Add this section
      resources:
        reservations:
          devices:
            - driver: nvidia
              #device_ids: ['0'] # this is only needed when using multiple GPUs
              count: 1 # number of GPUs
              capabilities: [gpu]
    shm_size: "256MB"
    devices:
      #- /dev/bus/usb:/dev/bus/usb #Used for Coral if Available
      #- /dev/dri/renderD128
      - /dev/dri/renderD129

    volumes:
      - /etc/localtime:/etc/localtime:ro
      - /home/ghislain/frigate/config:/config
      #- /home/ghislain/frigate/config.yml:/config/config.yml
      - /home/ghislain/frigate/db:/db
      - /home/ghislain/frigate/storage:/media/frigate
      - type: tmpfs # Optional: 1GB of memory, reduces SSD/SD Card wear
        target: /tmp/cache
        tmpfs:
          size: 1000000000
    ports:
      - "5000:5000"
      - "8554:8554" # RTSP feeds
      - "8555:8555/tcp" # WebRTC over tcp
      - "8555:8555/udp" # WebRTC over udp
      - "8971:8971"
      - "1984:1984/tcp"
    environment:
        FRIGATE_MQTT_USER: "xxxx"
        FRIGATE_MQTT_PASSWORD: "xxxxx"
        FRIGATE_RTSP_PASSWORD: "xxxx"
        FRIGATE_RTSP_USER: "admin"
        #LIBVA_DRIVER_NAME: i965
        PLUS_API_KEY: "XXXXXXXXXX"

Relevant Frigate log output

2025-11-21 23:00:55.221784469    File "/usr/local/lib/python3.11/dist-packages/onnxruntime/capi/onnxruntime_inference_collection.py", line 266, in run
2025-11-21 23:00:55.221790286      return self._sess.run(output_names, input_feed, run_options)
2025-11-21 23:00:55.221839095             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2025-11-21 23:00:55.221851041  onnxruntime.capi.onnxruntime_pybind11_state.Fail: [ONNXRuntimeError] : 1 : FAIL : Non-zero status code returned while running Slice node. Name:'Slice_34' Status Message: CUDA error cudaErrorNoKernelImageForDevice:no kernel image is available for execution on the device
2025-11-21 23:01:10.400661085  [2025-11-21 23:01:10] frigate.watchdog               INFO    : Detection appears to be stuck. Restarting detection process...
2025-11-21 23:01:10.467300678  [2025-11-21 23:01:10] detector.onnx                  INFO    : Starting detection process: 1144
2025-11-21 23:01:10.469339386  [2025-11-21 23:01:10] frigate.detectors.plugins.onnx INFO    : ONNX: loaded onnxruntime module
2025-11-21 23:01:10.470616199  [2025-11-21 23:01:10] frigate.detectors.plugins.onnx INFO    : ONNX: loading /config/model_cache/yolox_m.onnx
2025-11-21 23:01:10.782856731  [2025-11-21 23:01:10] frigate.detectors.plugins.onnx INFO    : ONNX: /config/model_cache/yolox_m.onnx loaded
2025-11-21 23:01:10.792225537  2025-11-21 23:01:10.792177366 [E:onnxruntime:, sequential_executor.cc:516 ExecuteKernel] Non-zero status code returned while running Slice node. Name:'Slice_34' Status Message: CUDA error cudaErrorNoKernelImageForDevice:no kernel image is available for execution on the device
2025-11-21 23:01:10.810555725  Process detector:onnx:
2025-11-21 23:01:10.810559215  Traceback (most recent call last):
2025-11-21 23:01:10.810560694    File "/usr/lib/python3.11/multiprocessing/process.py", line 314, in _bootstrap
2025-11-21 23:01:10.810561712      self.run()
2025-11-21 23:01:10.810562939    File "/opt/frigate/frigate/util/process.py", line 41, in run_wrapper
2025-11-21 23:01:10.810563911      return run(*args, **kwargs)
2025-11-21 23:01:10.810564933             ^^^^^^^^^^^^^^^^^^^^
2025-11-21 23:01:10.810566126    File "/usr/lib/python3.11/multiprocessing/process.py", line 108, in run
2025-11-21 23:01:10.810567195      self._target(*self._args, **self._kwargs)
2025-11-21 23:01:10.810568495    File "/opt/frigate/frigate/object_detection/base.py", line 136, in run_detector
2025-11-21 23:01:10.810569621      detections = object_detector.detect_raw(input_frame)
2025-11-21 23:01:10.810573963                   ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2025-11-21 23:01:10.810575186    File "/opt/frigate/frigate/object_detection/base.py", line 86, in detect_raw
2025-11-21 23:01:10.810576341      return self.detect_api.detect_raw(tensor_input=tensor_input)
2025-11-21 23:01:10.810577562             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2025-11-21 23:01:10.810578781    File "/opt/frigate/frigate/detectors/plugins/onnx.py", line 81, in detect_raw
2025-11-21 23:01:10.810580015      tensor_output = self.model.run(None, {model_input_name: tensor_input})
2025-11-21 23:01:10.810581275                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2025-11-21 23:01:10.810582829    File "/usr/local/lib/python3.11/dist-packages/onnxruntime/capi/onnxruntime_inference_collection.py", line 266, in run
2025-11-21 23:01:10.810584052      return self._sess.run(output_names, input_feed, run_options)
2025-11-21 23:01:10.810585209             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2025-11-21 23:01:10.810587579  onnxruntime.capi.onnxruntime_pybind11_state.Fail: [ONNXRuntimeError] : 1 : FAIL : Non-zero status code returned while running Slice node. Name:'Slice_34' Status Message: CUDA error cudaErrorNoKernelImageForDevice:no kernel image is available for execution on the device
s6-rc: info: service legacy-services: stopping
s6-rc: info: service legacy-services successfully stopped
s6-rc: info: service go2rtc-healthcheck: stopping
s6-rc: info: service certsync: stopping
2025-11-21 23:01:18.798058389  [INFO] The go2rtc-healthcheck service exited with code 256 (by signal 15)
s6-rc: info: service go2rtc-healthcheck successfully stopped
2025-11-21 23:01:18.807252278  [INFO] Service CERTSYNC exited with code 256 (by signal 15)
s6-rc: info: service certsync successfully stopped
s6-rc: info: service nginx: stopping
s6-rc: info: service certsync-log: stopping
s6-rc: info: service certsync-log successfully stopped
2025-11-21 23:01:18.927990900  [INFO] Service NGINX exited with code 0 (by signal 0)
s6-rc: info: service nginx successfully stopped
s6-rc: info: service nginx-log: stopping
s6-rc: info: service frigate: stopping
s6-rc: info: service nginx-log successfully stopped

Relevant go2rtc log output

FFprobe output from your camera

Install method

Docker Compose

Object Detector

TensorRT

Network connection

Wired

Camera make and model

REOLINK

Screenshots of the Frigate UI's System metrics pages

/

Any other information that may be helpful

No response

hawkeye217 · 2025-11-22T12:59:28Z

inkeep-ai-assistant[bot]
bot Nov 22, 2025

The error CUDA error cudaErrorNoKernelImageForDevice:no kernel image is available for execution on the device indicates your NVIDIA Quadro K2200 is encountering compatibility issues with the CUDA runtime in the Frigate container(1)(2).

Your Quadro K2200 has compute capability 5.0, which technically meets the minimum requirement for TensorRT (compute capability 5.0 or greater)(3). However, the specific error suggests the CUDA kernel binaries in the container aren't compatible with your older Maxwell architecture GPU(2).

Configuration Issues

Your ONNX detector configuration is missing the device specification. Add this to your config(4):

detectors:
  onnx:
    type: onnx
    device: '0'  # Add this line

(4)

Model Configuration Problem

The YOLOx model configuration needs adjustments. Based on the documentation, here's the correct configuration for YOLOx(4):

model:
  model_type: yolox
  width: 416  # <--- should match the imgsize set during model export
  height: 416  # <--- should match the imgsize set during model export
  input_tensor: nchw
  input_dtype: float_denorm
  path: /config/model_cache/yolox_m.onnx
  labelmap_path: /labelmap/coco-80.txt

(4)

Note that you had model_type: yolo-generic but for YOLOx models you need model_type: yolox(4). Also, the input_dtype should be float_denorm for YOLOx, not float(4).

Alternative Approach

Given your older GPU architecture, consider using YOLO-NAS instead, which may have better compatibility(4). After downloading the YOLO-NAS model, use this configuration(4):

detectors:
  onnx:
    type: onnx
    device: '0'

model:
  model_type: yolonas
  width: 320
  height: 320
  input_pixel_format: bgr
  input_tensor: nchw
  path: /config/yolo_nas_s.onnx
  labelmap_path: /labelmap/coco-80.txt

(4)

Hardware Acceleration

Your FFmpeg hardware acceleration configuration looks correct for NVIDIA GPUs(5):

ffmpeg:
  hwaccel_args: preset-nvidia

(5)

The Quadro K2200 is quite old (Maxwell generation), and while it meets the minimum compute capability requirement, newer CUDA runtimes may have limited kernel support for this architecture(2). If GPU compatibility issues persist, you may need to fall back to CPU mode or consider upgrading to a more recent GPU.

**

📚 Sources:

Have another question? Just tag @inkeep.

5 replies

hawkeye217 Nov 22, 2025
Collaborator

The AI's response is correct here. Your model configuration is incorrect for a YOLOx model.

https://docs.frigate.video/configuration/object_detectors/#yolox-1

Gis70 Nov 22, 2025
Author

it is a good starting, but even with thi ssettings, pb is still present :

detectors:
  onnx:
    type: onnx
    device: '0'

model:
  model_type: yolox
  width: 416 # <--- should match the imgsize set during model export
  height: 416 # <--- should match the imgsize set during model export
  input_tensor: nchw
  input_dtype: float_denorm
  path: /config/model_cache/yolox_tiny.onnx
  labelmap_path: /config/model_cache/coco_80.txt

I have also tryied to try YOLO-NAS Model with the google collab notebook but first try has failed.

i have the fear that i may have a driver problem with the GPU, i will continue to dig but if somebody with some similar hardware config would give me advice....

thanks in advance.

NickM-27 Nov 22, 2025
Collaborator Sponsor

what errors do you get when you use that config?

Gis70 Nov 22, 2025
Author

The same than before changing the configuration with the AI advices.

2025-11-22 20:10:16.334761988  [2025-11-22 20:10:16] frigate.app                    INFO    : go2rtc process pid: 124
2025-11-22 20:10:16.350624529  [2025-11-22 20:10:16] frigate.app                    INFO    : Embedding process started: 483
2025-11-22 20:10:16.354798119  [2025-11-22 20:10:16] detector.onnx                  INFO    : Starting detection process: 480
2025-11-22 20:10:16.355378701  [2025-11-22 20:10:16] frigate.detectors.plugins.onnx INFO    : ONNX: loaded onnxruntime module
2025-11-22 20:10:16.355652028  [2025-11-22 20:10:16] frigate.detectors.plugins.onnx INFO    : ONNX: loading /config/model_cache/yolox_tiny.onnx
2025-11-22 20:10:16.378003259  [2025-11-22 20:10:16] frigate.app                    INFO    : Output process started: 509
2025-11-22 20:10:16.410887054  [2025-11-22 20:10:16] frigate.app                    INFO    : Camera processor started for DOORBIRD: 530
2025-11-22 20:10:16.422641499  [2025-11-22 20:10:16] frigate.app                    INFO    : Camera processor started for CUISINE: 532
2025-11-22 20:10:16.433688541  [2025-11-22 20:10:16] frigate.app                    INFO    : Camera processor started for TERRASSE: 543
2025-11-22 20:10:16.455153230  [2025-11-22 20:10:16] frigate.app                    INFO    : Camera processor started for COUR: 546
2025-11-22 20:10:16.491661630  [2025-11-22 20:10:16] frigate.app                    INFO    : Camera processor started for ENTREE: 558
2025-11-22 20:10:16.494250124  [2025-11-22 20:10:16] frigate.app                    INFO    : Camera processor started for TERRAIN: 568
2025-11-22 20:10:16.498547244  [2025-11-22 20:10:16] frigate.app                    INFO    : Camera processor started for GARAGE: 589
2025-11-22 20:10:16.499040911  [2025-11-22 20:10:16] frigate.app                    INFO    : Camera processor started for DERRIERE: 595
2025-11-22 20:10:16.534504131  [2025-11-22 20:10:16] frigate.app                    INFO    : Capture process started for DOORBIRD: 621
2025-11-22 20:10:16.544470611  [2025-11-22 20:10:16] frigate.app                    INFO    : Capture process started for CUISINE: 631
2025-11-22 20:10:16.555960928  [2025-11-22 20:10:16] frigate.app                    INFO    : Capture process started for TERRASSE: 639
2025-11-22 20:10:16.575954876  [2025-11-22 20:10:16] frigate.app                    INFO    : Capture process started for COUR: 656
2025-11-22 20:10:16.612780900  [2025-11-22 20:10:16] frigate.app                    INFO    : Capture process started for ENTREE: 666
2025-11-22 20:10:16.622929366  [2025-11-22 20:10:16] frigate.app                    INFO    : Capture process started for TERRAIN: 671
2025-11-22 20:10:16.649123980  [2025-11-22 20:10:16] frigate.app                    INFO    : Capture process started for GARAGE: 683
2025-11-22 20:10:16.653084108  [2025-11-22 20:10:16] frigate.app                    INFO    : Capture process started for DERRIERE: 696
2025-11-22 20:10:16.763269811  [2025-11-22 20:10:16] frigate.detectors.plugins.onnx INFO    : ONNX: /config/model_cache/yolox_tiny.onnx loaded
2025-11-22 20:10:16.804560876  [2025-11-22 20:10:16] frigate.api.fastapi_app        INFO    : Starting FastAPI app
2025-11-22 20:10:16.925711743  [2025-11-22 20:10:16] frigate.api.fastapi_app        INFO    : FastAPI started
2025-11-22 20:10:17.157693333  [2025-11-22 20:10:17] frigate.record.util            INFO    : Deleting 4479 recording DB entries with missing files
2025-11-22 20:10:17.161095797  [2025-11-22 20:10:17] frigate.record.util            WARNING : Deleting 80.03% of recordings DB entries, could be due to configuration error. Aborting...
2025-11-22 20:10:21.839322416  2025-11-22 20:10:21.839132035 [E:onnxruntime:, sequential_executor.cc:516 ExecuteKernel] Non-zero status code returned while running Slice node. Name:'Slice_34' Status Message: CUDA error cudaErrorNoKernelImageForDevice:no kernel image is available for execution on the device
2025-11-22 20:10:21.878619439  Process detector:onnx:
2025-11-22 20:10:21.878630345  Traceback (most recent call last):
2025-11-22 20:10:21.878637758    File "/usr/lib/python3.11/multiprocessing/process.py", line 314, in _bootstrap
2025-11-22 20:10:21.878642796      self.run()
2025-11-22 20:10:21.878648815    File "/opt/frigate/frigate/util/process.py", line 41, in run_wrapper
2025-11-22 20:10:21.878656087      return run(*args, **kwargs)
2025-11-22 20:10:21.878665933             ^^^^^^^^^^^^^^^^^^^^
2025-11-22 20:10:21.878672269    File "/usr/lib/python3.11/multiprocessing/process.py", line 108, in run
2025-11-22 20:10:21.878722691      self._target(*self._args, **self._kwargs)
2025-11-22 20:10:21.878729316    File "/opt/frigate/frigate/object_detection/base.py", line 136, in run_detector
2025-11-22 20:10:21.878735084      detections = object_detector.detect_raw(input_frame)
2025-11-22 20:10:21.878740505                   ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2025-11-22 20:10:21.878803293    File "/opt/frigate/frigate/object_detection/base.py", line 86, in detect_raw
2025-11-22 20:10:21.878809224      return self.detect_api.detect_raw(tensor_input=tensor_input)
2025-11-22 20:10:21.878817079             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2025-11-22 20:10:21.878823140    File "/opt/frigate/frigate/detectors/plugins/onnx.py", line 81, in detect_raw
2025-11-22 20:10:21.878892056      tensor_output = self.model.run(None, {model_input_name: tensor_input})
2025-11-22 20:10:21.878898711                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2025-11-22 20:10:21.878906545    File "/usr/local/lib/python3.11/dist-packages/onnxruntime/capi/onnxruntime_inference_collection.py", line 266, in run
2025-11-22 20:10:21.878912174      return self._sess.run(output_names, input_feed, run_options)
2025-11-22 20:10:21.878965367             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2025-11-22 20:10:21.878978160  onnxruntime.capi.onnxruntime_pybind11_state.Fail: [ONNXRuntimeError] : 1 : FAIL : Non-zero status code returned while running Slice node. Name:'Slice_34' Status Message: CUDA error cudaErrorNoKernelImageForDevice:no kernel image is available for execution on the device

hawkeye217 Nov 22, 2025
Collaborator

The error CUDA error cudaErrorNoKernelImageForDevice: no kernel image is available for execution on the device indicates that the CUDA code being executed lacks the necessary binary code compiled for your specific GPU's architecture. This can occur for several reasons:

1. Incompatible PyTorch/CUDA Installation

The version of PyTorch or other CUDA-dependent libraries you are using may not have been built with support for your GPU's compute capability (e.g. sm_86 for Ampere GPUs, sm_120 for RTX 5090).

2. Outdated Drivers or CUDA Toolkit

Your GPU drivers or the CUDA toolkit installed on your system might be incompatible with the version of CUDA used to compile the binaries you are trying to run.

3. Incorrect Device Selection

The code might be attempting to run on a CPU when it expects a GPU, or vice versa.

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Uh oh!

[HW Accel Support]: NVIDIA QUADRO K2200 GPU - Model / Configuration Need help #21006

Uh oh!

{{title}}

Uh oh!

Uh oh!

{{editor}}'s edit

{{editor}}'s edit

Uh oh!

Replies: 1 comment 5 replies

Uh oh!

{{title}}

Uh oh!

Uh oh!

{{title}}

Uh oh!

Uh oh!

{{title}}

Uh oh!

Uh oh!

{{editor}}'s edit

{{editor}}'s edit

Uh oh!

Uh oh!

{{title}}

Uh oh!

Uh oh!

{{title}}

Uh oh!

Uh oh!

{{title}}

Uh oh!

Select a reply

Uh oh!

Uh oh!

[HW Accel Support]: NVIDIA QUADRO K2200 GPU - Model / Configuration Need help #21006

Uh oh!

Uh oh!

Gis70 Nov 22, 2025

Describe the problem you are having

Version

Frigate config file

docker-compose file or Docker CLI command

Relevant Frigate log output

Relevant go2rtc log output

FFprobe output from your camera

Install method

Object Detector

Network connection

Camera make and model

Screenshots of the Frigate UI's System metrics pages

Any other information that may be helpful

Replies: 1 comment · 5 replies

Uh oh!

inkeep-ai-assistant[bot] bot Nov 22, 2025

Configuration Issues

Model Configuration Problem

Alternative Approach

Hardware Acceleration

Uh oh!

hawkeye217 Nov 22, 2025 Collaborator

Uh oh!

Uh oh!

Gis70 Nov 22, 2025 Author

Uh oh!

NickM-27 Nov 22, 2025 Collaborator Sponsor

Uh oh!

Gis70 Nov 22, 2025 Author

Uh oh!

hawkeye217 Nov 22, 2025 Collaborator

1. Incompatible PyTorch/CUDA Installation

2. Outdated Drivers or CUDA Toolkit

3. Incorrect Device Selection

Gis70
Nov 22, 2025

Replies: 1 comment 5 replies

inkeep-ai-assistant[bot]
bot Nov 22, 2025

hawkeye217 Nov 22, 2025
Collaborator

Gis70 Nov 22, 2025
Author

NickM-27 Nov 22, 2025
Collaborator Sponsor

Gis70 Nov 22, 2025
Author

hawkeye217 Nov 22, 2025
Collaborator