Hi Felix, > @Philip, do you have time to test this? Do we need to ask someone > else?
Independent datapoint: I've been running this patch (as part of the Sep-30 8-patch set, hand-backported to 7.2.6) on gfx906 / MI50 under exactly the workload it targets - KFD SVM map/unmap/evict/restore churn from ROCm inference (vLLM model staging + llama.cpp) plus memory pressure. Previously this setup produced deterministic MMHUB "PTE not valid" faults and SVM restore-worker wedges on every run. With the series: zero faults across sustained sessions. So at least on GFX9-class hardware with the classic KFD SVM path, this version of the HMM range handling is behaving correctly. Obviously Philip's coverage on the NPA/UALink side is still needed - we have no such hardware. Tested-by: Olivier Kaloudoff <[email protected]> Regards, Olivier
