[Detector Support]: Issue setting up accelerated detection with rocm

[chrisb123]

Describe the problem you are having

Frigate is working when using CPU. I cant seem to setup hardware accelerated detection
Mini PC CPU 8845HS with 780m 32GB
I have gone through these steps to install rocm and the amdgpu module
https://rocm.docs.amd.com/projects/install-on-linux/en/latest/install/install-methods/package-manager/package-manager-debian.html
I've tried setting alternate versions for HSA_OVERRIDE_GFX_VERSION 11.0.0 and 11.0.2

Version

0.16.4-4131252

Frigate config file

detectors:
  onnx:
    type: onnx

model:
  model_type: yolo-generic
  width: 640
  height: 640
  input_tensor: nchw
  input_dtype: float
  path: /config/model_cache/yolox_s.onnx

docker-compose file or Docker CLI command

services:
  frigate:
    container_name: frigate
    privileged: true # this may not be necessary for all setups
    restart: unless-stopped
    stop_grace_period: 30s # allow enough time to shut down the various services
    image: ghcr.io/blakeblackshear/frigate:stable-rocm
    shm_size: "4gb" # update for your cameras based on calculation above
    devices:
      #- /dev/bus/usb:/dev/bus/usb # Passes the USB Coral, needs to be modified for other versions
      #- /dev/apex_0:/dev/apex_0 # Passes a PCIe Coral, follow driver instructions here https://github.com/jnicolson/gasket-builder
      #- /dev/video11:/dev/video11 # For Raspberry Pi 4B
      #- /dev/dri/renderD128:/dev/dri/renderD128 # For intel hwaccel, needs to be updated for your hardware
      - /dev/dri:/dev/dri
      - /dev/kfd:/dev/kfd
    volumes:
      - /etc/localtime:/etc/localtime:ro
      - /config:/config
      - /media/frigate:/media/frigate
      - type: tmpfs # Optional: 1GB of memory, reduces SSD/SD Card wear
        target: /tmp/cache
        tmpfs:
          size: 1000000000
    ports:
      - "8971:8971"
      - "5000:5000" # Internal unauthenticated access. Expose carefully.
      - "8554:8554" # RTSP feeds
      - "8555:8555/tcp" # WebRTC over tcp
      - "8555:8555/udp" # WebRTC over udp
    environment:
      FRIGATE_RTSP_PASSWORD: "x"
      HSA_OVERRIDE_GFX_VERSION: "11.0.2"
      LIBVA_DRIVER_NAME: "radeonsi"

Relevant Frigate log output

frigate  | 2026-02-05 17:16:10.157487849  [2026-02-05 17:16:10] frigate.app                    INFO    : Starting Frigate (0.16.4-4131252)
frigate  | 2026-02-05 17:16:10.160160329  [2026-02-05 17:16:10] peewee_migrate.logs            INFO    : Starting migrations
frigate  | 2026-02-05 17:16:10.160603694  [2026-02-05 17:16:10] peewee_migrate.logs            INFO    : There is nothing to migrate
frigate  | 2026-02-05 17:16:10.171484590  [2026-02-05 17:16:10] frigate.app                    INFO    : Recording process started: 503
frigate  | 2026-02-05 17:16:10.171673351  [2026-02-05 17:16:10] frigate.app                    INFO    : Review process started: 515
frigate  | 2026-02-05 17:16:10.172971406  [2026-02-05 17:16:10] frigate.app                    INFO    : go2rtc process pid: 119
frigate  | 2026-02-05 17:16:10.183953901  [2026-02-05 17:16:10] detector.onnx                  INFO    : Starting detection process: 528
frigate  | 2026-02-05 17:16:10.186721468  [2026-02-05 17:16:10] frigate.detectors.plugins.onnx INFO    : ONNX: loaded onnxruntime module
frigate  | 2026-02-05 17:16:10.187738811  [2026-02-05 17:16:10] frigate.detectors.plugins.onnx INFO    : ONNX: loading /config/model_cache/yolox_s.onnx
frigate  | 2026-02-05 17:16:10.187855648  [2026-02-05 17:16:10] frigate.app                    INFO    : Embedding process started: 531
frigate  | 2026-02-05 17:16:10.197977212  [2026-02-05 17:16:10] frigate.app                    INFO    : Output process started: 559
frigate  | 2026-02-05 17:16:10.210969946  [2026-02-05 17:16:10] frigate.app                    INFO    : Camera processor started for Undercover: 580
frigate  | 2026-02-05 17:16:10.218887952  [2026-02-05 17:16:10] frigate.app                    INFO    : Camera processor started for Inside: 587
frigate  | 2026-02-05 17:16:10.226555283  [2026-02-05 17:16:10] frigate.app                    INFO    : Capture process started for Undercover: 612
frigate  | 2026-02-05 17:16:10.261840545  [2026-02-05 17:16:10] frigate.app                    INFO    : Capture process started for Inside: 634
frigate  | 2026-02-05 17:16:10.388933134  [2026-02-05 17:16:10] frigate.api.fastapi_app        INFO    : Starting FastAPI app
frigate  | 2026-02-05 17:16:10.534106839  [2026-02-05 17:16:10] frigate.api.fastapi_app        INFO    : FastAPI started
frigate  | 2026-02-05 17:16:10.980409797  loading data from : /config/model_cache/facedet/landmarkdet.yaml
frigate  | 2026-02-05 17:16:11.133281098  [INFO] Starting go2rtc healthcheck service...
frigate  | 2026-02-05 17:16:11.268507692  INFO: Created TensorFlow Lite XNNPACK delegate for CPU.
frigate  | 2026-02-05 17:16:11.760488477  Memory access fault by GPU node-1 (Agent handle: 0x2801cc30) on address 0x7f3acde84000. Reason: Page not present or supervisor privilege.
frigate  | 2026-02-05 17:16:11.760494207  Fatal Python error: Aborted
frigate  | 2026-02-05 17:16:11.760496662
frigate  | 2026-02-05 17:16:11.760498205  Thread 0x00007f3ad4ee3040 (most recent call first):
frigate  | 2026-02-05 17:16:11.760522490    File "/usr/local/lib/python3.11/dist-packages/onnxruntime/capi/onnxruntime_inference_collection.py", line 537 in _create_inference_session
frigate  | 2026-02-05 17:16:11.760524313    File "/usr/local/lib/python3.11/dist-packages/onnxruntime/capi/onnxruntime_inference_collection.py", line 465 in __init__
frigate  | 2026-02-05 17:16:11.760525576    File "/opt/frigate/frigate/detectors/plugins/onnx.py", line 53 in __init__
frigate  | 2026-02-05 17:16:11.760567994    File "/opt/frigate/frigate/detectors/__init__.py", line 18 in create_detector
frigate  | 2026-02-05 17:16:11.760570189    File "/opt/frigate/frigate/object_detection/base.py", line 57 in __init__
frigate  | 2026-02-05 17:16:11.760572102    File "/opt/frigate/frigate/object_detection/base.py", line 112 in run_detector
frigate  | 2026-02-05 17:16:11.760573946    File "/usr/lib/python3.11/multiprocessing/process.py", line 108 in run
frigate  | 2026-02-05 17:16:11.760575749    File "/opt/frigate/frigate/util/process.py", line 41 in run_wrapper
frigate  | 2026-02-05 17:16:11.760578063    File "/usr/lib/python3.11/multiprocessing/process.py", line 314 in _bootstrap
frigate  | 2026-02-05 17:16:11.760598141    File "/usr/lib/python3.11/multiprocessing/popen_fork.py", line 71 in _launch
frigate  | 2026-02-05 17:16:11.760599483    File "/usr/lib/python3.11/multiprocessing/popen_fork.py", line 19 in __init__
frigate  | 2026-02-05 17:16:11.760600846    File "/usr/lib/python3.11/multiprocessing/context.py", line 281 in _Popen
frigate  | 2026-02-05 17:16:11.760610544    File "/usr/lib/python3.11/multiprocessing/context.py", line 224 in _Popen
frigate  | 2026-02-05 17:16:11.760611826    File "/usr/lib/python3.11/multiprocessing/process.py", line 121 in start
frigate  | 2026-02-05 17:16:11.760625862    File "/opt/frigate/frigate/util/process.py", line 30 in start
frigate  | 2026-02-05 17:16:11.760635039    File "/opt/frigate/frigate/object_detection/base.py", line 195 in start_or_restart
frigate  | 2026-02-05 17:16:11.760636382    File "/opt/frigate/frigate/object_detection/base.py", line 163 in __init__
frigate  | 2026-02-05 17:16:11.760644847    File "/opt/frigate/frigate/app.py", line 391 in start_detectors
frigate  | 2026-02-05 17:16:11.760655998    File "/opt/frigate/frigate/app.py", line 648 in start
frigate  | 2026-02-05 17:16:11.760667760    File "/opt/frigate/frigate/__main__.py", line 104 in main
frigate  | 2026-02-05 17:16:11.760678831    File "/opt/frigate/frigate/__main__.py", line 108 in <module>
frigate  | 2026-02-05 17:16:11.760689921    File "<frozen runpy>", line 88 in _run_code
frigate  | 2026-02-05 17:16:11.760698407    File "<frozen runpy>", line 198 in _run_module_as_main
frigate  | 2026-02-05 17:16:11.760827667

Install method

Docker Compose

Object Detector

Other

Screenshots of the Frigate UI's System metrics pages

n/a

Any other information that may be helpful

No response

[inkeep-ai-assistant[bot]]

The error you're encountering (Memory access fault by GPU node-1) indicates ROCm is attempting to use your AMD 780M integrated GPU but failing during model initialization.(1)

For AMD GPUs with ROCm, you need to configure the HSA_OVERRIDE_GFX_VERSION environment variable correctly.(1) The AMD 780M (gfx1103) requires HSA_OVERRIDE_GFX_VERSION: "11.0.0" according to the automatic detection in Frigate's ROCm build.(1)

Try changing your environment variable from 11.0.2 to 11.0.0:

environment:
  FRIGATE_RTSP_PASSWORD: "x"
  HSA_OVERRIDE_GFX_VERSION: "11.0.0"
  LIBVA_DRIVER_NAME: "radeonsi"

(1)

You should also verify ROCm is working properly by running /opt/rocm/bin/rocminfo inside the Frigate container - it should list both the CPU and GPU with their properties.(1)

To check if ROCm found your GPU:

docker exec -it frigate /opt/rocm/bin/rocminfo

(1)

To verify your GPU chipset version:

docker exec -it frigate /bin/bash -c '(unset HSA_OVERRIDE_GFX_VERSION && /opt/rocm/bin/rocminfo |grep gfx)'

(1)

Important note about AMD integrated GPUs: AMD/ROCm does not officially support integrated GPUs, though they often work with the proper HSA_OVERRIDE_GFX_VERSION configuration.(1) Additionally, YOLO-NAS models are known to not run well on integrated GPUs when using ROCm.(1)

If issues persist after setting HSA_OVERRIDE_GFX_VERSION: "11.0.0", check the Frigate Docker logs for additional error details.(1)

📚 Sources:

1. Supported Hardware

---

Have another question? Just tag @inkeep.

[chrisb123]

Tried HSA_OVERRIDE_GFX_VERSION 11.0.0 same result
Tried the rocminfo command does appear to have found the GPU and show gfx1103

$ docker exec -it frigate /bin/bash -c '(unset HSA_OVERRIDE_GFX_VERSION && /opt/rocm/bin/rocminfo |grep gfx)'
Name: gfx1103
Name: amdgcn-amd-amdhsa--gfx1103

Did I miss some steps?
What else can I check?
Is my config correct?
Do you need more information?

[NickM-27]

YOLOx has not been tested with ROCm fully, your best bet is YOLOv9

[chrisb123]

I tried with yolov9 and got the same error:
Memory access fault by GPU node-1 (Agent handle: 0x2801cc30) on address 0x7f3acde84000. Reason: Page not present or supervisor privilege.
Fatal Python error: Aborted

Followed the instructions here: https://docs.frigate.video/configuration/object_detectors#downloading-yolo-models
It created the docker downloaded files and outputted a yolov9-7-320.onnx file size 7.8Mb
put that file into /config/model_cache
updated the config.yaml

detectors:
onnx:
type: onnx

model:
model_type: yolo-generic
width: 320
height: 320
input_tensor: nchw
input_dtype: float
path: /config/model_cache/yolov9-t-320.onnx
labelmap_path: /labelmap/coco-80.txt

[chrisb123]

I tried increasing the UMA to 8GB, this didnt make any difference to the problem
Any ideas?
What else can I check, re/install, re/config?

I tried adding the module option amdgpu.cwsr_enable=0
This is suppose to help with this kind of error

Apparently they have included support for gfx1103 in v7.1
ROCm/rocm-libraries#1320

How are other people managing to run frigate with 780m?
gmktec AMD Ryzen 7 8845HS K8 Plus 32GB RAM
What version of romc?
7.2
What amdgpu-dkms version?
30.30
What kernel version?
Linux debian 6.12.63+deb13-amd64

[NickM-27]

Looks like we have similar setups, runs correctly for me with 780M, memory access fault suggests it is a memory issue or driver issue.

[chrisb123]

It might be something specific to the gmktec k8 plus, it's a few weeks old, bought it specifically to run frigate, popular brand and not a cheap system
Since it uses system ram I'll try running a memtest

[chrisb123]

I found the solution

Do not install amdgpu-dkms.

Now I have 780m rocm and hailo-8

[msravi]

Hi,

How did you fix this? Which docker image are you using? Can you please post your docker-compose and config files?

[dsimoes]

I also update my amdgpu-dkms driver from 30.1 to 30.3 and now Frigate simply crashes :(
Nothing else changed.