MODEL-MM-GLM53-FLASH W9c-3b: the --device cuda SIGSEGV is a KV-page residency defect, not a mixed-residency one - #2483

Merged

mudler merged 11 commits into

mainfrom

row/MODEL-MM-GLM53-FLASH-DEVSEGV

Sep 2, 2026

Commits on Sep 1, 2026

spec(MODEL-MM-GLM53-FLASH): W9c-3b, because O46's mechanism is an artifact of two once-flags
mudler
committed
17718f5
fix(MODEL-MM-GLM53-FLASH): copy the engine's KV pages, because on a device queue they are not host memory
mudler
committed
b5e4cbf
spec(MODEL-MM-GLM53-FLASH): retract O46's mechanism where it is written, not only where it is replaced
mudler
committed
2e35086
test(MODEL-MM-GLM53-FLASH): say what the CPU-queue case can see, because a mutation proved it sees less
mudler
committed
cd9f502
docs(MODEL-MM-GLM53-FLASH): the two public surfaces repeated O46's mechanism, and it is false
mudler
committed
76a79d3
spec(MODEL-MM-GLM53-FLASH): say why the mixed residency state CANNOT be the fault, not only that it is not
mudler
committed
f8b2689
spec(MODEL-MM-GLM53-FLASH): say what else the C-against-D comparison measures
mudler
committed
91a071d
merge: origin/main into the GLM-5.3 KV-residency branch
mudler
committed
dacf22f
spec(MODEL-MM-GLM53-FLASH): the diagnosis is confirmed on GB10, and the job's own guard hid the fix
mudler
committed
0f0df02
spec(MODEL-MM-GLM53-FLASH): re-cut the evidence table around what run 1 settled and what run 2 must answer
mudler
committed
868b49b

Commits on Sep 2, 2026

spec(MODEL-MM-GLM53-FLASH): record run 2, which fixed the crash and failed a criterion I mis-specified
mudler
committed
154382b