AMD APU 780M ML
🗄️Server☑️Ready
Hi everyone,
I’m trying to run Immich machine learning with GPU acceleration on an AMD Ryzen H 255 APU with the integrated Radeon 780M, but GPU inference does not seem to work correctly.
Environment:
• Host: Proxmox
• Immich running inside an LXC container
• ROCm / amdgpu devices passed to the container (/dev/kfd and /dev/dri/renderD128)
• Devices are accessible inside the container and can be opened successfully
However, Immich ML still appears to run inference on the CPU, and the GPU does not seem to be used reliably.
Additional observations:
• /dev/kfd and /dev/dri/renderD128 are present and accessible
• No obvious errors from ROCm when opening the devices
• I consistently see amdgpu reset messages in the host logs every 5-10–20-30 images during ML processing
What I have already tried:
• Verified device access (/dev/kfd and /dev/dri/renderD128) inside the container
• Checked permissions and added the container user to video and render groups
• Confirmed that the devices can be opened successfully from inside the container
• Restarted containers and the host multiple times
• Tested with different numbers of ML workers
• Tried setting environment variables:
HSA_OVERRIDE_GFX_VERSION=11.0.0
HSA_USE_SVM=0
Questions:
Is Radeon 780M (RDNA3 iGPU) currently supported for ML acceleration in Immich?
Does ROCm support this APU yet?
Has anyone successfully used 780M or similar Phoenix / Hawk Point iGPU with Immich ML?
Any guidance or configuration examples would be greatly appreciated.
Thanks!
I’m trying to run Immich machine learning with GPU acceleration on an AMD Ryzen H 255 APU with the integrated Radeon 780M, but GPU inference does not seem to work correctly.
Environment:
• Host: Proxmox
• Immich running inside an LXC container
• ROCm / amdgpu devices passed to the container (/dev/kfd and /dev/dri/renderD128)
• Devices are accessible inside the container and can be opened successfully
However, Immich ML still appears to run inference on the CPU, and the GPU does not seem to be used reliably.
Additional observations:
• /dev/kfd and /dev/dri/renderD128 are present and accessible
• No obvious errors from ROCm when opening the devices
• I consistently see amdgpu reset messages in the host logs every 5-10–20-30 images during ML processing
What I have already tried:
• Verified device access (/dev/kfd and /dev/dri/renderD128) inside the container
• Checked permissions and added the container user to video and render groups
• Confirmed that the devices can be opened successfully from inside the container
• Restarted containers and the host multiple times
• Tested with different numbers of ML workers
• Tried setting environment variables:
HSA_OVERRIDE_GFX_VERSION=11.0.0
HSA_USE_SVM=0
Questions:
Is Radeon 780M (RDNA3 iGPU) currently supported for ML acceleration in Immich?
Does ROCm support this APU yet?
Has anyone successfully used 780M or similar Phoenix / Hawk Point iGPU with Immich ML?
Any guidance or configuration examples would be greatly appreciated.
Thanks!
