[Linux] [Arch] [AMD GPU] onnxruntime unable to load ROCm drivers

[torkd]

I have been trying to run FoxyFace on my machine with moderate success. I can run Babble fine on the CPU, but if I try to run it on the GPU it seems that the onnxruntime is unable to find the ROCm drivers, even if they are installed.

For clarity, my setup is as follows:

• OS: Arch Linux x86_64
• Kernel: 6.18.2-arch2-1
• CPU: AMD Ryzen 7 7800X3D
• GPU: AMD ATI Radeon RX 7900 XT

pacman -Qs rocm:

rocm packages installed

local/hipblas 7.1.1-1
  ROCm BLAS marshalling library
local/hsa-rocr 7.1.1-2
  HSA Runtime API and runtime for ROCm
local/rocblas 7.1.1-1
  Next generation BLAS implementation for ROCm platform
local/rocm-core 7.1.1-1
  AMD ROCm core package (version files)
local/rocm-device-libs 2:7.1.1-2
  AMD specific device-side language runtime libraries
local/rocm-llvm 2:7.1.1-2
  Radeon Open Compute - LLVM toolchain (llvm, clang, lld)
local/rocm-opencl-runtime 7.1.1-1
  OpenCL implementation for AMD
local/rocm-smi-lib 7.1.1-1
  ROCm System Management Interface Library
local/rocminfo 7.1.1-1
  ROCm Application for Reporting System Info
local/rocrand 7.1.1-1
  Pseudo-random and quasi-random number generator on ROCm
local/rocsolver 7.1.1-1
  Subset of LAPACK functionality on the ROCm platform
local/rocsparse 7.1.1-1
  BLAS for sparse computation on top of ROCm
local/roctracer 7.1.1-1
  ROCm tracer library for performance tracing

Those are the steps I followed to install FoxyFace (as per the Install Guide):

pyenv install 3.12.9 # Using pyenv to manage different Python versions
eval "$(pyenv init - bash)"
pyenv shell 3.12.9
git clone --recurse-submodules https://github.com/Jeka8833/FoxyFace.git
cd FoxyFace/FoxyFace
python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt
pip install onnxruntime-rocm -f https://repo.radeon.com/rocm/manylinux/rocm-rel-7.1.1/ # Installing version 7.1.1 since that's what I have on my system
python Main.py

Here the program would fail with this error:
/home/dav/VR/FaceTracking/FoxyFace/FoxyFace/venv/lib/python3.12/site-packages/onnxruntime/capi/onnxruntime_pybind11_state.so: cannot enable executable stack as shared object requires: Invalid argument

To fix that I ran:

patchelf --clear-execstack /home/dav/VR/FaceTracking/FoxyFace/FoxyFace/venv/lib/python3.12/site-packages/onnxruntime/capi/onnxruntime_pybind11_state.so

Now, when I run the application I get those logs:
latest.log:

logs

2025-12-30 14:42:00,185 [INFO] src.ui.UiImageUtil - (UiImageUtil.py).allow_change_windows_icon(18) [MainThread]: Failed to set app id, app is not Windows
2025-12-30 14:42:00,190 [INFO] __main__ - (Main.py).__init__(46) [MainThread]: Hello, I'm FoxyFace 1.0.4.3
2025-12-30 14:42:00,199 [INFO] src.config.ConfigManager - (ConfigManager.py).__write_task(120) [Config Manager_0]: Config saved, 0 listeners called
2025-12-30 14:42:00,380 [INFO] src.stream.camera.CameraStream - (CameraStream.py).start_new_camera(53) [MainThread]: Camera started
2025-12-30 14:42:00,405 [ERROR] std - (LoggerManager.py).write(11) [MainThread]: /home/dav/VR/FaceTracking/FoxyFace/FoxyFace/venv/lib/python3.12/site-packages/onnxruntime/capi/onnxruntime_inference_collection.py:121: UserWarning: Specified provider 'DmlExecutionProvider' is not in available provider names.Available providers: 'MIGraphXExecutionProvider, ROCMExecutionProvider, CPUExecutionProvider'
warnings.warn(

2025-12-30 14:42:00,405 [ERROR] std - (LoggerManager.py).write(11) [MainThread]: /home/dav/VR/FaceTracking/FoxyFace/FoxyFace/venv/lib/python3.12/site-packages/onnxruntime/capi/onnxruntime_inference_collection.py:121: UserWarning: Specified provider 'CUDAExecutionProvider' is not in available provider names.Available providers: 'MIGraphXExecutionProvider, ROCMExecutionProvider, CPUExecutionProvider'
warnings.warn(

2025-12-30 14:42:00,405 [ERROR] std - (LoggerManager.py).write(11) [MainThread]: /home/dav/VR/FaceTracking/FoxyFace/FoxyFace/venv/lib/python3.12/site-packages/onnxruntime/capi/onnxruntime_inference_collection.py:121: UserWarning: Specified provider 'CoreMLExecutionProvider' is not in available provider names.Available providers: 'MIGraphXExecutionProvider, ROCMExecutionProvider, CPUExecutionProvider'
warnings.warn(

2025-12-30 14:42:00,421 [INFO] std - (LoggerManager.py).write(11) [MainThread]: *************** EP Error ***************
2025-12-30 14:42:00,421 [INFO] std - (LoggerManager.py).write(11) [MainThread]: EP Error /root/onnx_runtime/onnxruntime/onnxruntime-1.22.2/onnxruntime/python/onnxruntime_pybind_state.cc:1087 std::unique_ptr<onnxruntime::IExecutionProvider> onnxruntime::python::CreateExecutionProviderInstance(const onnxruntime::SessionOptions&, const string&, const ProviderOptionsMap&) ROCM_PATH is set but ROCM wasn't able to be loaded. Please install the correct version of ROCM and MIOpen as mentioned in the GPU requirements page, make sure they're in the PATH, and that your GPU is supported.
when using [('DmlExecutionProvider', {'device_id': '0'}), ('CUDAExecutionProvider', {'device_id': '0'}), ('ROCMExecutionProvider', {'device_id': '0'}), 'CoreMLExecutionProvider', 'CPUExecutionProvider']
2025-12-30 14:42:00,421 [INFO] std - (LoggerManager.py).write(11) [MainThread]: Falling back to ['ROCMExecutionProvider', 'CPUExecutionProvider'] and retrying.
2025-12-30 14:42:00,421 [INFO] std - (LoggerManager.py).write(11) [MainThread]: ****************************************
2025-12-30 14:42:00,435 [WARNING] src.config.ConfigUpdateListener - (ConfigUpdateListener.py).__init__(27) [MainThread]: Failed to call update
Traceback (most recent call last):
File "/home/dav/VR/FaceTracking/FoxyFace/FoxyFace/venv/lib/python3.12/site-packages/onnxruntime/capi/onnxruntime_inference_collection.py", line 472, in __init__
  self._create_inference_session(providers, provider_options, disabled_optimizers)
File "/home/dav/VR/FaceTracking/FoxyFace/FoxyFace/venv/lib/python3.12/site-packages/onnxruntime/capi/onnxruntime_inference_collection.py", line 561, in _create_inference_session
  sess.initialize_session(providers, provider_options, disabled_optimizers)
RuntimeError: /root/onnx_runtime/onnxruntime/onnxruntime-1.22.2/onnxruntime/python/onnxruntime_pybind_state.cc:1087 std::unique_ptr<onnxruntime::IExecutionProvider> onnxruntime::python::CreateExecutionProviderInstance(const onnxruntime::SessionOptions&, const string&, const ProviderOptionsMap&) ROCM_PATH is set but ROCM wasn't able to be loaded. Please install the correct version of ROCM and MIOpen as mentioned in the GPU requirements page, make sure they're in the PATH, and that your GPU is supported.


The above exception was the direct cause of the following exception:

Traceback (most recent call last):
File "/home/dav/VR/FaceTracking/FoxyFace/FoxyFace/src/config/ConfigUpdateListener.py", line 25, in __init__
  self.__update_callback(self.__config_manager)
File "/home/dav/VR/FaceTracking/FoxyFace/FoxyFace/src/pipline/BabblePipeline.py", line 143, in __update_babble_loader_options
  self.__babble_loader.start_new_session(config_manager.config.babble.model_path,
File "/home/dav/VR/FaceTracking/FoxyFace/FoxyFace/src/stream/babble/BabbleModelLoader.py", line 53, in start_new_session
  session = InferenceSession(path, opts, providers=provider)
            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/dav/VR/FaceTracking/FoxyFace/FoxyFace/venv/lib/python3.12/site-packages/onnxruntime/capi/onnxruntime_inference_collection.py", line 485, in __init__
  raise fallback_error from e
File "/home/dav/VR/FaceTracking/FoxyFace/FoxyFace/venv/lib/python3.12/site-packages/onnxruntime/capi/onnxruntime_inference_collection.py", line 480, in __init__
  self._create_inference_session(self._fallback_providers, None)
File "/home/dav/VR/FaceTracking/FoxyFace/FoxyFace/venv/lib/python3.12/site-packages/onnxruntime/capi/onnxruntime_inference_collection.py", line 561, in _create_inference_session
  sess.initialize_session(providers, provider_options, disabled_optimizers)
RuntimeError: /root/onnx_runtime/onnxruntime/onnxruntime-1.22.2/onnxruntime/python/onnxruntime_pybind_state.cc:1087 std::unique_ptr<onnxruntime::IExecutionProvider> onnxruntime::python::CreateExecutionProviderInstance(const onnxruntime::SessionOptions&, const string&, const ProviderOptionsMap&) ROCM_PATH is set but ROCM wasn't able to be loaded. Please install the correct version of ROCM and MIOpen as mentioned in the GPU requirements page, make sure they're in the PATH, and that your GPU is supported.
Stack (most recent call last):
File "/home/dav/VR/FaceTracking/FoxyFace/FoxyFace/Main.py", line 96, in <module>
  with RunMainStream(__splash):
File "/home/dav/VR/FaceTracking/FoxyFace/FoxyFace/Main.py", line 53, in __init__
  self.__babble_pipeline: BabblePipeline = BabblePipeline(self.__config_manager, self.__media_pipe_pipeline)
File "/home/dav/VR/FaceTracking/FoxyFace/FoxyFace/src/pipline/BabblePipeline.py", line 37, in __init__
  self.__babble_loader_options_listener: ConfigUpdateListener = self.__register_change_babble_loader_options()
File "/home/dav/VR/FaceTracking/FoxyFace/FoxyFace/src/pipline/BabblePipeline.py", line 140, in __register_change_babble_loader_options
  return self.__config_manager.create_update_listener(self.__update_babble_loader_options, watch_array, True)
File "/home/dav/VR/FaceTracking/FoxyFace/FoxyFace/src/config/ConfigManager.py", line 60, in create_update_listener
  listener = ConfigUpdateListener(self, update_callback, call_on_create, watched_elements)
File "/home/dav/VR/FaceTracking/FoxyFace/FoxyFace/src/config/ConfigUpdateListener.py", line 27, in __init__
  _logger.warning("Failed to call update", exc_info=True, stack_info=True)
2025-12-30 14:42:00,438 [INFO] src.stream.vrcft.VrcftAutoConnect - (VrcftAutoConnect.py).__start_loop(66) [VRCFT Auto Connect]: Starting VRCFT Auto Connect
2025-12-30 14:42:02,864 [INFO] src.config.ConfigManager - (ConfigManager.py).__write_task(108) [MainThread]: Config did not change
2025-12-30 14:42:04,443 [INFO] src.stream.vrcft.VrcftAutoConnect - (VrcftAutoConnect.py).__start_loop(86) [VRCFT Auto Connect]: Stopped VRCFT Auto Connect
2025-12-30 14:42:04,444 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7f1d5495a930>
2025-12-30 14:42:04,444 [INFO] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7f1d5495a930>
2025-12-30 14:42:04,447 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7f1d5495a930>

The line that catches my eye is this one:
RuntimeError: /root/onnx_runtime/onnxruntime/onnxruntime-1.22.2/onnxruntime/python/onnxruntime_pybind_state.cc:1087 std::unique_ptr<onnxruntime::IExecutionProvider> onnxruntime::python::CreateExecutionProviderInstance(const onnxruntime::SessionOptions&, const string&, const ProviderOptionsMap&) ROCM_PATH is set but ROCM wasn't able to be loaded. Please install the correct version of ROCM and MIOpen as mentioned in the GPU requirements page, make sure they're in the PATH, and that your GPU is supported.

Is it supposed to look into the /root folder?
Also, since it mentions MIOpen, I went on and installed that as well (sudo pacman -S miopen-hip), but nothing changed.
ROCm is installed correctly, I've used it successfully with other applications (and rocminfo shows correct information).
I'm at a bit of a loss, I'm pretty sure I've tried everything I could think of/found online.

[Jeka8833]

Solution 1: Try running it using PyTorch (This is a very large library, so be prepared to download about 3 GByte):

1. Perform a clean installation up to the step where you need to select a video card.
2. Instead of the step pip install onnxruntime-rocm -f https://repo.radeon.com/rocm/manylinux/rocm-rel-6.4.3/ run the following commands:
   pip install torch torchvision --index-url https://download.pytorch.org/whl/rocm6.4
pip install onnxruntime-gpu
3. Now, launch FoxyFace:
   python Main.py

Solution 2: You can always run a neural network without a GPU. In the documentation, this is the “No GPU” drop-down list.

Unfortunately, I can't check, as I don't have AMD or Arch Linux, so I'm just guessing.

[Jeka8833]

++ Gemini says that you need to install:
sudo pacman -S rocm-hip-sdk

and then:
HSA_OVERRIDE_GFX_VERSION=11.0.0 python Main.py

But I wouldn't trust Gemini.

[torkd]

Trying those suggestions.

Running with PyTorch doesn't work; I'm guessing the issue here might be that there is no rocm7.1.1 repo for PyTorch. It tries running on CUDA, fails and resorts to run on the CPU. It does run fine on the CPU, so it's not a "blocking" issue.
With or without the Gemini part didn't change anything.

latest.log:

logs

2025-12-30 17:30:40,859 [INFO] src.ui.UiImageUtil - (UiImageUtil.py).allow_change_windows_icon(18) [MainThread]: Failed to set app id, app is not Windows
2025-12-30 17:30:40,863 [INFO] __main__ - (Main.py).__init__(46) [MainThread]: Hello, I'm FoxyFace 1.0.4.3
2025-12-30 17:30:40,872 [INFO] src.config.ConfigManager - (ConfigManager.py).__write_task(120) [Config Manager_0]: Config saved, 0 listeners called
2025-12-30 17:30:41,052 [INFO] src.stream.camera.CameraStream - (CameraStream.py).start_new_camera(53) [MainThread]: Camera started
2025-12-30 17:30:41,875 [ERROR] std - (LoggerManager.py).write(11) [MainThread]: /home/dav/VR/FaceTracking/FoxyFace/FoxyFace/venv/lib/python3.12/site-packages/onnxruntime/capi/onnxruntime_inference_collection.py:123: UserWarning: Specified provider 'DmlExecutionProvider' is not in available provider names.Available providers: 'TensorrtExecutionProvider, CUDAExecutionProvider, CPUExecutionProvider'
  warnings.warn(

2025-12-30 17:30:41,875 [ERROR] std - (LoggerManager.py).write(11) [MainThread]: /home/dav/VR/FaceTracking/FoxyFace/FoxyFace/venv/lib/python3.12/site-packages/onnxruntime/capi/onnxruntime_inference_collection.py:123: UserWarning: Specified provider 'ROCMExecutionProvider' is not in available provider names.Available providers: 'TensorrtExecutionProvider, CUDAExecutionProvider, CPUExecutionProvider'
  warnings.warn(

2025-12-30 17:30:41,875 [ERROR] std - (LoggerManager.py).write(11) [MainThread]: /home/dav/VR/FaceTracking/FoxyFace/FoxyFace/venv/lib/python3.12/site-packages/onnxruntime/capi/onnxruntime_inference_collection.py:123: UserWarning: Specified provider 'CoreMLExecutionProvider' is not in available provider names.Available providers: 'TensorrtExecutionProvider, CUDAExecutionProvider, CPUExecutionProvider'
  warnings.warn(

2025-12-30 17:30:41,935 [INFO] src.stream.babble.BabbleModelLoader - (BabbleModelLoader.py).start_new_session(68) [MainThread]: Babble started
2025-12-30 17:30:41,936 [INFO] src.stream.vrcft.VrcftAutoConnect - (VrcftAutoConnect.py).__start_loop(66) [VRCFT Auto Connect]: Starting VRCFT Auto Connect
2025-12-30 17:30:48,744 [INFO] src.config.ConfigManager - (ConfigManager.py).__write_task(108) [MainThread]: Config did not change
2025-12-30 17:30:50,945 [INFO] src.stream.vrcft.VrcftAutoConnect - (VrcftAutoConnect.py).__start_loop(86) [VRCFT Auto Connect]: Stopped VRCFT Auto Connect
2025-12-30 17:30:50,946 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fce1a65bce0>
2025-12-30 17:30:50,946 [INFO] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fce1a65bce0>
2025-12-30 17:30:50,954 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fce1a65bce0>
2025-12-30 17:30:50,954 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fce1a65bce0>
2025-12-30 17:30:50,954 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fce1a65bce0>
2025-12-30 17:30:50,954 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fce1a65bce0>
2025-12-30 17:30:50,954 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fce1a65bce0>
2025-12-30 17:30:50,955 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fce1a65bce0>
2025-12-30 17:30:50,955 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fce1a65bce0>
2025-12-30 17:30:50,955 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fce1a65bce0>
2025-12-30 17:30:50,955 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fce1a65bce0>
2025-12-30 17:30:50,955 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fce1a65bce0>
2025-12-30 17:30:50,955 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fce1a65bce0>
2025-12-30 17:30:50,955 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fce1a65bce0>
2025-12-30 17:30:50,955 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fce1a65bce0>
2025-12-30 17:30:50,955 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fce1a65bce0>
2025-12-30 17:30:50,955 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fce1a65bce0>
2025-12-30 17:30:50,955 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fce1a65bce0>
2025-12-30 17:30:50,955 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fce1a65bce0>
2025-12-30 17:30:50,955 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fce1a65bce0>
2025-12-30 17:30:50,955 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fce1a65bce0>
2025-12-30 17:30:50,955 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fce1a65bce0>
2025-12-30 17:30:50,955 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fce1a65bce0>
2025-12-30 17:30:50,955 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fce1a65bce0>
2025-12-30 17:30:50,955 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fce1a65bce0>
2025-12-30 17:30:50,955 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fce1a65bce0>
2025-12-30 17:30:50,955 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fce1a65bce0>
2025-12-30 17:30:50,955 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fce1a65bce0>
2025-12-30 17:30:50,955 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fce1a65bce0>
2025-12-30 17:30:50,955 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fce1a65bce0>
2025-12-30 17:30:50,955 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fce1a65bce0>
2025-12-30 17:30:50,955 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fce1a65bce0>
2025-12-30 17:30:50,955 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fce1a65bce0>
2025-12-30 17:30:50,955 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fce1a65bce0>
2025-12-30 17:30:50,955 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fce1a65bce0>
2025-12-30 17:30:50,955 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fce1a65bce0>
2025-12-30 17:30:50,955 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fce1a65bce0>
2025-12-30 17:30:50,955 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fce1a65bce0>
2025-12-30 17:30:50,955 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fce1a65bce0>

---

Update:
I found PyTorch for ROCm 7.1 in the nightly directory (I'm a dumbass and didn't think of looking there), but the issue remains the same even if I install with:
pip install torch torchvision --index-url https://download.pytorch.org/whl/nightly/rocm7.1/

latest.log:

logs

2025-12-30 18:28:26,818 [INFO] src.ui.UiImageUtil - (UiImageUtil.py).allow_change_windows_icon(18) [MainThread]: Failed to set app id, app is not Windows
2025-12-30 18:28:26,823 [INFO] __main__ - (Main.py).__init__(46) [MainThread]: Hello, I'm FoxyFace 1.0.4.3
2025-12-30 18:28:26,831 [INFO] src.config.ConfigManager - (ConfigManager.py).__write_task(108) [Config Manager_0]: Config did not change
2025-12-30 18:28:27,012 [INFO] src.stream.camera.CameraStream - (CameraStream.py).start_new_camera(53) [MainThread]: Camera started
2025-12-30 18:28:27,840 [ERROR] std - (LoggerManager.py).write(11) [MainThread]: /home/dav/VR/FaceTracking/FoxyFace/FoxyFace/venv/lib/python3.12/site-packages/onnxruntime/capi/onnxruntime_inference_collection.py:123: UserWarning: Specified provider 'DmlExecutionProvider' is not in available provider names.Available providers: 'TensorrtExecutionProvider, CUDAExecutionProvider, CPUExecutionProvider'
  warnings.warn(

2025-12-30 18:28:27,840 [ERROR] std - (LoggerManager.py).write(11) [MainThread]: /home/dav/VR/FaceTracking/FoxyFace/FoxyFace/venv/lib/python3.12/site-packages/onnxruntime/capi/onnxruntime_inference_collection.py:123: UserWarning: Specified provider 'ROCMExecutionProvider' is not in available provider names.Available providers: 'TensorrtExecutionProvider, CUDAExecutionProvider, CPUExecutionProvider'
  warnings.warn(

2025-12-30 18:28:27,840 [ERROR] std - (LoggerManager.py).write(11) [MainThread]: /home/dav/VR/FaceTracking/FoxyFace/FoxyFace/venv/lib/python3.12/site-packages/onnxruntime/capi/onnxruntime_inference_collection.py:123: UserWarning: Specified provider 'CoreMLExecutionProvider' is not in available provider names.Available providers: 'TensorrtExecutionProvider, CUDAExecutionProvider, CPUExecutionProvider'
  warnings.warn(

2025-12-30 18:28:27,899 [INFO] src.stream.babble.BabbleModelLoader - (BabbleModelLoader.py).start_new_session(68) [MainThread]: Babble started
2025-12-30 18:28:27,900 [INFO] src.stream.vrcft.VrcftAutoConnect - (VrcftAutoConnect.py).__start_loop(66) [VRCFT Auto Connect]: Starting VRCFT Auto Connect
2025-12-30 18:28:46,165 [INFO] src.config.ConfigManager - (ConfigManager.py).__write_task(108) [MainThread]: Config did not change
2025-12-30 18:28:48,934 [INFO] src.stream.vrcft.VrcftAutoConnect - (VrcftAutoConnect.py).__start_loop(86) [VRCFT Auto Connect]: Stopped VRCFT Auto Connect
2025-12-30 18:28:48,935 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fb3c185b9b0>
2025-12-30 18:28:48,935 [INFO] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fb3c185b9b0>
2025-12-30 18:28:48,943 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fb3c185b9b0>
2025-12-30 18:28:48,943 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fb3c185b9b0>
2025-12-30 18:28:48,943 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fb3c185b9b0>
2025-12-30 18:28:48,943 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fb3c185b9b0>
2025-12-30 18:28:48,943 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fb3c185b9b0>
2025-12-30 18:28:48,943 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fb3c185b9b0>
2025-12-30 18:28:48,943 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fb3c185b9b0>
2025-12-30 18:28:48,943 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fb3c185b9b0>
2025-12-30 18:28:48,943 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fb3c185b9b0>
2025-12-30 18:28:48,943 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fb3c185b9b0>
2025-12-30 18:28:48,943 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fb3c185b9b0>
2025-12-30 18:28:48,943 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fb3c185b9b0>
2025-12-30 18:28:48,943 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fb3c185b9b0>
2025-12-30 18:28:48,943 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fb3c185b9b0>
2025-12-30 18:28:48,943 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fb3c185b9b0>
2025-12-30 18:28:48,943 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fb3c185b9b0>
2025-12-30 18:28:48,943 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fb3c185b9b0>
2025-12-30 18:28:48,943 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fb3c185b9b0>
2025-12-30 18:28:48,943 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fb3c185b9b0>
2025-12-30 18:28:48,943 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fb3c185b9b0>
2025-12-30 18:28:48,943 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fb3c185b9b0>
2025-12-30 18:28:48,943 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fb3c185b9b0>
2025-12-30 18:28:48,943 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fb3c185b9b0>
2025-12-30 18:28:48,943 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fb3c185b9b0>
2025-12-30 18:28:48,944 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fb3c185b9b0>
2025-12-30 18:28:48,944 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fb3c185b9b0>
2025-12-30 18:28:48,944 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fb3c185b9b0>
2025-12-30 18:28:48,944 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fb3c185b9b0>
2025-12-30 18:28:48,944 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fb3c185b9b0>
2025-12-30 18:28:48,944 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fb3c185b9b0>
2025-12-30 18:28:48,944 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fb3c185b9b0>
2025-12-30 18:28:48,944 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fb3c185b9b0>
2025-12-30 18:28:48,944 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fb3c185b9b0>
2025-12-30 18:28:48,944 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fb3c185b9b0>
2025-12-30 18:28:48,944 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fb3c185b9b0>
2025-12-30 18:28:48,944 [ERROR] std - (LoggerManager.py).flush(14) [MainThread]: <src.LoggerManager.LoggerWriter object at 0x7fb3c185b9b0>

Details

[Jeka8833]

I have no idea, I tried to install using Dockerfile but couldn't get Qt and Docker to work together.

For me, Linux is something complicated. I saw that AMD compiles its builds only for Ubuntu, and how to make it work on Arch Linux is a mystery to me.

[ah9142]

Hello.

Not sure if this the issue, but noticed that:
latest ubuntu repositories have libonnxruntime 1.21.0
arch linux has onnxruntime 1.23.2

Looking at onnx docs ( https://onnxruntime.ai/docs/execution-providers/ROCm-ExecutionProvider.html ) we can find:

ROCm 7.0 Will be the last officially supported AMD Release which includes ROCm Execution Provider Please use MIGraphX Execution Provider For your applications instead.
and that the last onnxruntime that has ROCm is 1.22.1.

@torkd you could try downgrading your onnxruntime to 1.22.1 and see if anything changes:

sudo pacman -U https://archive.archlinux.org/packages/o/onnxruntime-opt-rocm/onnxruntime-opt-rocm-1.22.1-1-x86_64.pkg.tar.zst

[ah9142]

And if downgrading the onnx version resolves the issue, then the long term solution would be to add support for MIGraphX Execution Provider. I believe that should be possible by adding it into:

        if use_gpu:
            provider = [("DmlExecutionProvider", {"device_id": device_id_str}),
                        ("CUDAExecutionProvider", {"device_id": device_id_str}),
                        ("ROCMExecutionProvider", {"device_id": device_id_str}), "CoreMLExecutionProvider",
                        "CPUExecutionProvider"]

https://github.com/Jeka8833/FoxyFace/blob/main/FoxyFace/src/stream/babble/BabbleModelLoader.py

[Jeka8833]

And if downgrading the onnx version resolves the issue, then the long term solution would be to add support for MIGraphX Execution Provider. I believe that should be possible by adding it into:

        if use_gpu:
            provider = [("DmlExecutionProvider", {"device_id": device_id_str}),
                        ("CUDAExecutionProvider", {"device_id": device_id_str}),
                        ("ROCMExecutionProvider", {"device_id": device_id_str}), "CoreMLExecutionProvider",
                        "CPUExecutionProvider"]

https://github.com/Jeka8833/FoxyFace/blob/main/FoxyFace/src/stream/babble/BabbleModelLoader.py

If you have verified that MIGraphX works on Arch Linux, you can create a Pull Request, as you have found a solution.
If you do not wish to make a Pull Request, I will make the changes myself.

Also, share which version you installed in pacman to run onnxruntime-migraphx. Because I don't have AMD at the moment and want to give people in FoxyFace documentation some kind of hint.

[ah9142]

I have not verified this with FoxyFace. What I said might be totally wrong and irrelevant to this issue.
Somebody else needs to verify this. Or I might give it a try later and also attemp a fix, but not now as it has been a long day.

The reason I commented was because I was having a similar issue with baballonia software where an older version of baballonia was not able to load the ROCm execution provider.

The package I have installed is onnxruntime-opt-rocm 1.23.2-3.
The newer rocm onnxruntimes do not have a ROCm Execution Provider.

In the arch linux baballonia AUR package comments ( https://aur.archlinux.org/packages/baballonia -> scroll down ) I found this useful command which allowed me to check whether the ROCm execution provider entrypoint exists in the onnxruntime that I have installed.
The output with onnxruntime-opt-rocm 1.23.2-3 is:

$ readelf -Ws /usr/lib/libonnxruntime*|grep ExecutionProvider_|sort -u
   505: 0000000000362b10   296 FUNC    GLOBAL DEFAULT   12 OrtSessionOptionsAppendExecutionProvider_MIGraphX@@VERS_1.23.2
   506: 00000000003629e0   296 FUNC    GLOBAL DEFAULT   12 OrtSessionOptionsAppendExecutionProvider_Dnnl@@VERS_1.23.2
   509: 00000000005af980   264 FUNC    GLOBAL DEFAULT   12 OrtSessionOptionsAppendExecutionProvider_CPU@@VERS_1.23.2

E.g. there is no ROCm runtime entrypoint symbol in the binary. With an older version of the package ( onnxruntime-opt-rocm 1.22.1 ) an entrypoint symbol could be found and the old version of baballonia worked too.

After that I found the onnxruntime docs ( https://onnxruntime.ai/docs/execution-providers/ROCm-ExecutionProvider.html ) and the mention about the ROCm runtime execution provider being switched into MIGraphX

But take everything I said with a grain of salt as I did not try out this particular software.
Let's see if I get to testing this later.

[torkd]

I have not verified this with FoxyFace. What I said might be totally wrong and irrelevant to this issue. Somebody else needs to verify this. Or I might give it a try later and also attemp a fix, but not now as it has been a long day.

The reason I commented was because I was having a similar issue with baballonia software where an older version of baballonia was not able to load the ROCm execution provider.

The package I have installed is onnxruntime-opt-rocm 1.23.2-3. The newer rocm onnxruntimes do not have a ROCm Execution Provider.

In the arch linux baballonia AUR package comments ( https://aur.archlinux.org/packages/baballonia -> scroll down ) I found this useful command which allowed me to check whether the ROCm execution provider entrypoint exists in the onnxruntime that I have installed. The output with onnxruntime-opt-rocm 1.23.2-3 is:

$ readelf -Ws /usr/lib/libonnxruntime*|grep ExecutionProvider_|sort -u
   505: 0000000000362b10   296 FUNC    GLOBAL DEFAULT   12 OrtSessionOptionsAppendExecutionProvider_MIGraphX@@VERS_1.23.2
   506: 00000000003629e0   296 FUNC    GLOBAL DEFAULT   12 OrtSessionOptionsAppendExecutionProvider_Dnnl@@VERS_1.23.2
   509: 00000000005af980   264 FUNC    GLOBAL DEFAULT   12 OrtSessionOptionsAppendExecutionProvider_CPU@@VERS_1.23.2

E.g. there is no ROCm runtime entrypoint symbol in the binary. With an older version of the package ( onnxruntime-opt-rocm 1.22.1 ) an entrypoint symbol could be found and the old version of baballonia worked too.

After that I found the onnxruntime docs ( https://onnxruntime.ai/docs/execution-providers/ROCm-ExecutionProvider.html ) and the mention about the ROCm runtime execution provider being switched into MIGraphX

But take everything I said with a grain of salt as I did not try out this particular software. Let's see if I get to testing this later.

I have the same package installed, and even trying to downgrade it with an older version it still fails. As reported in the first post, the application fails when trying to load certain libraries:

2026-01-02 22:27:08.573601473 [E:onnxruntime:Default, provider_bridge_ort.cc:2251 TryGetProviderInfo_ROCM] /root/onnx_runtime/onnxruntime/onnxruntime-1.22.2/onnxruntime/core/session/provider_bridge_ort.cc:1778 onnxruntime::Provider& onnxruntime::ProviderLibrary::Get() [ONNXRuntimeError] : 1 : FAIL : Failed to load library libonnxruntime_providers_rocm.so with error: libhipblas.so.2: cannot open shared object file: No such file or directory

I didn't realise what the error was really saying at the time, but now I have checked with ldconfig -p | grep hipblas:

libhipblaslt.so.1 (libc6,x86-64) => /opt/rocm/lib/libhipblaslt.so.1
libhipblaslt.so (libc6,x86-64) => /opt/rocm/lib/libhipblaslt.so
libhipblas.so.3 (libc6,x86-64) => /opt/rocm/lib/libhipblas.so.3
libhipblas.so (libc6,x86-64) => /opt/rocm/lib/libhipblas.so

For some reason the SONAME of those libraries don't match. Even trying a quick dirty hack of adding a symlink for libhipblas.so.2 => libhipblas.so.3 just results in other libraries having a different SONAME. Not sure what it is caused by, since the pip package is installed with the correct ROCm version.

Even downgrading the onnxruntime-opt-rocm package didn't solve this.

---

Moving to the other solution of adding the MIGraphX Execution Provider.
I have edited the src/stream/babble/BabbleModelLoader.py file as follows:

        if use_gpu:
            provider = [("DmlExecutionProvider", {"device_id": device_id_str}),
                        ("CUDAExecutionProvider", {"device_id": device_id_str}),
                        ("ROCMExecutionProvider", {"device_id": device_id_str}),
                        "MIGraphXExecutionProvider", "CoreMLExecutionProvider",
                        "CPUExecutionProvider"]

Then I installed the onnxruntime-migraphx library via pip, on top of onnxruntime-rocm (pip install onnxruntime-migraphx), and ran the application.
FoxyFace started, and those are the logs:

shell

[INFO]: Failed to set app id, app is not Windows
[INFO]: Hello, I'm FoxyFace 1.0.4.3
[INFO]: Config saved, 0 listeners called
[INFO]: Camera started
WARNING: All log messages before absl::InitializeLog() is called are written to STDERR
W0000 00:00:1767391409.349806   62141 face_landmarker_graph.cc:174] Sets FaceBlendshapesGraph acceleration to xnnpack by default.
INFO: Created TensorFlow Lite XNNPACK delegate for CPU.
W0000 00:00:1767391409.355614   62144 inference_feedback_manager.cc:114] Feedback manager requires a model with a single signature inference. Disabling support for feedback tensors.
W0000 00:00:1767391409.363160   62147 inference_feedback_manager.cc:114] Feedback manager requires a model with a single signature inference. Disabling support for feedback tensors.
[ERROR]: /home/dav/VR/FaceTracking/FoxyFace/FoxyFace/venv/lib/python3.12/site-packages/onnxruntime/capi/onnxruntime_inference_collection.py:123: UserWarning: Specified provider 'DmlExecutionProvider' is not in available provider names.Available providers: 'MIGraphXExecutionProvider, CPUExecutionProvider'
  warnings.warn(

[ERROR]: /home/dav/VR/FaceTracking/FoxyFace/FoxyFace/venv/lib/python3.12/site-packages/onnxruntime/capi/onnxruntime_inference_collection.py:123: UserWarning: Specified provider 'CUDAExecutionProvider' is not in available provider names.Available providers: 'MIGraphXExecutionProvider, CPUExecutionProvider'
  warnings.warn(

[ERROR]: /home/dav/VR/FaceTracking/FoxyFace/FoxyFace/venv/lib/python3.12/site-packages/onnxruntime/capi/onnxruntime_inference_collection.py:123: UserWarning: Specified provider 'ROCMExecutionProvider' is not in available provider names.Available providers: 'MIGraphXExecutionProvider, CPUExecutionProvider'
  warnings.warn(

[ERROR]: /home/dav/VR/FaceTracking/FoxyFace/FoxyFace/venv/lib/python3.12/site-packages/onnxruntime/capi/onnxruntime_inference_collection.py:123: UserWarning: Specified provider 'CoreMLExecutionProvider' is not in available provider names.Available providers: 'MIGraphXExecutionProvider, CPUExecutionProvider'
  warnings.warn(

2026-01-02 23:03:29.484079532 [W:onnxruntime:Default, migraphx_execution_provider.cc:167 MIGraphXExecutionProvider] [MIGraphX EP] MIGraphX ENV Override Variables Set:
2026-01-02 23:03:29.528953971 [W:onnxruntime:Default, migraphx_execution_provider.cc:1252 compile_program] Model Compile: Begin
2026-01-02 23:04:05.226669961 [W:onnxruntime:Default, migraphx_execution_provider.cc:1257 compile_program] Model Compile: Complete
2026-01-02 23:04:05.246432153 [F:onnxruntime:Default, migraphx_execution_provider.cc:1526 operator()] MIGraphX: param type mismatch
[INFO]: Babble started
[INFO]: Starting VRCFT Auto Connect
2026-01-02 23:04:05.268120358 [F:onnxruntime:Default, migraphx_execution_provider.cc:1526 operator()] MIGraphX: param type mismatch
2026-01-02 23:04:05.276723255 [F:onnxruntime:Default, migraphx_execution_provider.cc:1526 operator()] MIGraphX: param type mismatch
[...]
2026-01-02 23:04:16.026574775 [F:onnxruntime:Default, migraphx_execution_provider.cc:1526 operator()] MIGraphX: param type mismatch

Those are directly from the shell since they didn't get dumped in the log file. That last line keeps getting spammed into the output for as long as the program runs.
Otherwise, the provider seems to work, and I assume it's running on my GPU looking at amdgpu_top:

Name            |  PID  |KFD| VRAM | GTT  |CPU |GFX |COMP|DMA |VCNU|   │
python          |  64230| Y |  341M|    2M| 50%|  0%|  0%|  0%|  0%|   │

I also noticed that the MIGraphX Execution Provider will compile the model every time, unless certain session options are set (as per ONNXRuntime docs).
I'm honestly not sure how to edit the code to add those params conditionally based on what provider is being used. I might try giving it a shot sometime in the next few days.

[ah9142]

For some reason the SONAME of those libraries don't match. Even trying a quick dirty hack of adding a symlink for libhipblas.so.2 => libhipblas.so.3 just results in other libraries having a different SONAME. Not sure what it is caused by, since the pip package is installed with the correct ROCm version.
The number at the end of the file indicates the ABI version of that library. Most likely the python package was compiled against libhipblas that had version 2, but arch linux already has a newer version. Symlinking incompatible ABI versions rarely does any good if there was an actual ABI change e.g. they removed/added/changed some symbols in an ABI backwards incompatible way (and its wasn't just the devs bumping the number for fun). To resolve this you would probably need to rollback all the relevant packages on your system back which can be done ( https://wiki.archlinux.org/title/Arch_Linux_Archive#How_to_restore_all_packages_to_a_specific_date ) but can sometimes be a huge pain, so I would not recommend it if there are other options.

In other news I also tried installing and running FoxyFace and basically reproduced the same errors and results as @torkd :

Details

First I tried with the ROCm executrion provider and got the errors from the original post:

[INFO]: EP Error /root/onnx_runtime/onnxruntime/onnxruntime-1.22.2/onnxruntime/python/onnxruntime_pybind_state.cc:1087 std::unique_ptr<onnxruntime::IExecutionProvider> onnxruntime::python::CreateExecutionProviderInstance(const onnxruntime::SessionOptions&, const string&, const ProviderOptionsMap&) ROCM_PATH is set but ROCM wasn't able to be loaded. Please install the correct version of ROCM and MIOpen as mentioned in the GPU requirements page, make sure they're in the PATH, and that your GPU is supported.
 when using [('DmlExecutionProvider', {'device_id': '0'}), ('CUDAExecutionProvider', {'device_id': '0'}), ('ROCMExecutionProvider', {'device_id': '0'}), 'CoreMLExecutionProvider', 'CPUExecutionProvider']
[INFO]: Falling back to ['ROCMExecutionProvider', 'CPUExecutionProvider'] and retrying.
[INFO]: ****************************************
2026-01-03 11:36:27.686290733 [E:onnxruntime:Default, provider_bridge_ort.cc:2251 TryGetProviderInfo_ROCM] /root/onnx_runtime/onnxruntime/onnxruntime-1.22.2/onnxruntime/core/session/provider_bridge_ort.cc:1778 onnxruntime::Provider& onnxruntime::ProviderLibrary::Get() [ONNXRuntimeError] : 1 : FAIL : Failed to load library libonnxruntime_providers_rocm.so with error: libhipblas.so.2: cannot open shared object file: No such file or directory

Next I followed the linux setup guide ( https://foxyface.jeka8833.pp.ua/docs/FoxyFace/install-update-uninstall/install/Install-FoxyFace-Linux ) again but now installed:

pip install onnxruntime-migraphx -f https://repo.radeon.com/rocm/manylinux/rocm-rel-7.1.1/

Additionally, I added the MIGraphXExecutionProvider as @torkd previously did.
Now the model seemed to be running, but constantly producing the errors:

2026-01-03 12:22:18.050551585 [F:onnxruntime:Default, migraphx_execution_provider.cc:1526 operator()] MIGraphX: param type mismatch

I also noticed that the MIGraphX Execution Provider will compile the model every time, unless certain session options are set (as per ONNXRuntime docs).

So, to elaborate you can cache the compilation results by setting the ORT_MIGRAPHX_MODEL_CACHE_PATH env variable to point to a directory where to store the cached model. E.g.

mkdir -p ~/.cache/my-migraphx-cache
export ORT_MIGRAPHX_MODEL_CACHE_PATH=~/.cache/my-migraphx-cache
python Main.py

With this env variable set a cached model will be used on subsequent launches.

@torkd Would you be able to test that the model actually works and that the errors are not actually something relevant?
EDIT: Nevermind I can see that the parameters in the "Processing -> Settings" section are being tracked correctly and I can see my GPU usage going up. So, it seems to be working at least for me.

I guess in that case we could simply add the model to the list of execution providers and ignore the errors.

[Jeka8833]

FoxyFace/FoxyFace/src/stream/babble/BabbleModelLoader.py

Line 27 in 6e862a2

if "CUDAExecutionProvider" in providers:

I think this is why "torch" does not work for AMD; I simply did not specify that ROCMExecutionProvider requires "torch".

[Jeka8833]

1. Perform a clean installation up to the step where you need to select a video card.
2. Replace this line:

   FoxyFace/FoxyFace/src/stream/babble/BabbleModelLoader.py

   Line 27 in 6e862a2

   if "CUDAExecutionProvider" in providers:

   
   With: if "CUDAExecutionProvider" in providers or "ROCMExecutionProvider" in providers or "MIGraphXExecutionProvider" in providers:
3. Instead of the step pip install onnxruntime-rocm -f https://repo.radeon.com/rocm/manylinux/rocm-rel-6.4.3/ run the following commands:

• pip install torch torchvision --index-url https://download.pytorch.org/whl/rocm6.4
• pip install onnxruntime-gpu or pip install onnxruntime-rocm (I have no idea if there is a ROCMExecutionProvider in onnxruntime-gpu)

4. Now, launch FoxyFace:
   python Main.py