On some AMD boards Vulkan dies with ErrorDeviceLost only when SAM/ReBAR is enabled
[blocker] driver
Fields
id
gotchas/vulkan-rebar-on-causes-amd-device-lostClass
driver
Severity
blocker
Symptom
The opposite failure of the previous record, and it is a hard crash rather than a slowdown: with SAM/ReBAR enabled, every decode on an AMD Radeon AI PRO R9700 fails with vk::Queue::submit: ErrorDeviceLost, llama-server returns HTTP 500 and exits. Disabling ReBAR in BIOS makes the identical driver, binary and command line succeed. Reproduces at 2K, 16K and 256K context, so it is not context dependent.
Root cause
Unknown. It is a ReBAR-toggle-confirmed fault rather than a driver-version fault (downgrading the driver did not help), but another user runs 2x R9700 on Linux with ReBAR without issue, so the card alone is not sufficient to cause it. The reporter's platform is an ASUS X570 PRO running the card at PCIe x8, which the thread explores as a contributing factor.
Workaround
Disable SAM/ReBAR in BIOS on this platform - that is the only confirmed fix in the thread, and it re-introduces the slowdown recorded in vulkan-rebar-off-collapses-amd-decode, so on this class of board you are choosing between two bad options and should test both.
notes
Recorded as a separate record from the ReBAR-off slowdown because the sign of the fix is inverted: the obvious remedy for one makes the other worse. Both are ReBAR-toggle-confirmed on AMD hardware, which makes ReBAR state a first-class diagnostic to check before blaming a backend or a model. Issue open, labelled bug-unconfirmed.
Notes
Recorded as a separate record from the ReBAR-off slowdown because the sign of the fix is inverted: the obvious remedy for one makes the other worse. Both are ReBAR-toggle-confirmed on AMD hardware, which makes ReBAR state a first-class diagnostic to check before blaming a backend or a model. Issue open, labelled bug-unconfirmed.