MI100 multi-GPU segfaults on bare-metal ROCm; needs VM passthrough + HSA_AMD_SVM
[blocker] framework
Fields
id
gotchas/gotcha-amd-mi100-multigpu-segfaults-without-vm-passthroughClass
framework
Severity
blocker
Concepts
rocmgfx908multi-gpupcie-passthroughvm
Symptom
More than one MI100 segfaults llama.cpp on bare-metal Arch/ROCm; single card works.
Root cause
gfx908 multi-GPU handling on the ROCm stack of that era is unstable; the working recipe is PCIe passthrough into an Ubuntu VM plus the HSA_AMD_SVM kernel option.
Workaround
Read llama.cpp discussion #2824 BEFORE buying a second card: run the cards under PCIe passthrough in an Ubuntu VM and set HSA_AMD_SVM. gfx908 is on the ROCm 10 compat matrix but maintenance-tier.