I’m using the ProArt X870E with two AMD R9700.
If I try to suspend the system when llama.cpp is running with tensor split-mode it will crash. It never enters sleep mode, it just hangs, and I have to power cycle the system.
By llama.cpp running I mean the system is idle with application running, it’s not doing any AI workload when I suspend.
Suspend works normally as long as llama.ccp doesn’t use tensor mode, or when using tensor mode with amgpu.pcie_p2p=n.
Would it be possible to get the system to suspend with p2p active?
This sounds more like a PCIe/AMD P2P power-management interaction than a llama.cpp workload issue. Since suspend works when tensor split is disabled or amd...pcie_p2p=n is set, P2P mappings may be remaining active when the system tries to transition to sleep.
I’d test with the latest kernel, AMDGPU firmware, BIOS/AGESA, and llama.cpp build first. Checking dmesg immediately before and after the failed suspend may also reveal whether the GPU or PCIe subsystem is refusing to suspend. If P2P is keeping the GPUs in a state that cannot be safely restored, there may not be a software-only llama.cpp setting that fixes it; a kernel/AMDGPU power-management fix may be required.